Reinforcement learning (RL) from unit tests is the common recipe for post-training code models, but a unit test only asks whether the final program passes. We analyze what this unit-test-only reward teaches a model during debugging and find that, while it raises pass rates, it also teaches the model to edit more and to hallucinate bugs in code that is already correct. We introduce PreciseCoder, a training algorithm that augments the unit-test reward with a signal for semantically verified, localized edits, which correlates with smaller edits and higher pass rates. On 4B, 9B, and 27B models, PreciseCoder reduces the average number of edit lines from 7-15to 3 at a pass rate on par with unit-test-only training. Trained only on single-line bugs, PreciseCoder can generalize to repository-level program repair, where it resolves more issues with smaller patches. Its precision also transfers from weaker to stronger models: given the reasoning trace of a 9B PreciseCoder model, frontier models such as GPT-5.6 Sol and Qwen3.8-Max gain 4.6-12.4 points in edit precision with no additional training. Linguistic analysis of the thinking traces shows that our reward encourages the model to reason about individual lines, name the buggy ones, and form consistent bug hypotheses.
Can an autonomous AI agent reverse engineer a binary as well as a skilled human analyst? Do they adopt the same strategy, and can agents scale to real-world programs? We compare human and LLM-based software reverse engineering (SRE): nine models tackle two compact static-analysis challenges previously solved by human analysts, scored with a common understanding rubric, and we record their complete interactions with IDA Pro. This lets us compare not only what humans and agents recover, but also how they investigate a binary. The best agents (GPT-5.5 Pro, Claude Opus 4.8, and Kimi K3) achieve statistically human-competitive understanding on both tasks, while weaker models remain far behind. Ablations show that the gap depends on code representation and workflow demands, not on model capability alone: open-weight models improve substantially with richer code representations, whereas frontier models show no significant change. Similar success rates hide different workflows. Humans repeatedly revisit functions and follow local program structure; strong agents jump across the binary, retrieve information broadly, and rarely return to inspected code, a behavior we call greedy fetching. This strategy is not fixed: given semantically richer code, agents fetch less without losing understanding. Because greedy fetching aggregates large amounts of code in the agent's context, we measure context-window saturation across 4,000 real-world binaries: large windows accommodate many programs, but larger applications and lower-level representations quickly exhaust them. Autonomous reverse engineering is becoming competitive, with a large speed advantage, but scaling it will require better navigation, selective retrieval, and persistent analysis state, not merely larger context windows.
Chatbot prompts reveal the engineer’s alleged efforts to set up a covert listening device days before one was found in a meeting room used by high-level executives.
Research: Profiling web-edit with Python 3.15: flame graphs and source heatmaps I put the new Python 3.15 profiler through its paces by having GPT-6 Astra Extra High in Codex run it against my new web-edit app : I need a Python 3.15 flame graph and heatmap demo of the "web-edit" app from PyPI - research first then make me those demos Here's the resulting example flame graph and heatmap . Tags: profiling , python , codex
Release: llm-meta-ai 0.2 Meta AI release Muse Wonder (and Muse Wonder Full), their new "most capable models". These look like they may be Meta's first Fable class models - Artificial Analysis scores Muse Wonder (Max) as equivalent to GPT-6.1 Sol. I updated my llm-meta-ai plugin for the new model. Here are pelicans for Muse Wonder for the six reasoning levels: minimal, low, medium, high, xhigh, max. This is the max one: Blue cycling helmet, bicycle is missing one bar but otherwise an OK shape, has mudguards, nice background, OK beak. For comparison, here are pelicans for Muse 1.3 (released September 2nd) and Muse 1.2 (released August 5th). Meta AI are putting out a new model a month, and the improvements between each model (at least in terms of drawing pelicans) are material. Tags: ai , generative-ai , llms , meta , pelican-riding-a-bicycle , llm-release
Release: web-edit 0.1a1 Includes performance optimizations written by Claude Opus 5.5 after I fed it this Python 3.15 profiling report generated by GPT-6 Astra. Tags: profiling , projects
An OpenAI engineer was arrested after the company said it discovered a "sophisticated recording device" hidden in one of its offices. The employee, 35-year-old Abdulrahman Alfozan, is charged with using or attempting to use a device to intercept oral communications, according to a report from The San Francisco Standard. OpenAI's post about the incident says […]
A federal judge ruled that Apple can access OpenAI's interview materials involving former Apple employees, as well as private messages and testimony from key figures in the case.
AI leaders joined whistleblowers and security leaders at a sweeping parliamentary session, while a few hundred meters away Britain’s spy chief warned AI could be used to create new bioweapons.
Existing benchmarks measure solve rate and token output/cost, but don't punish excess code, which makes a codebase harder to work with in the long term. I wanted to measure this since it matters most for long-term projects. Method: it takes the patches that benchmarks already publish (DeepSWE, SWE-bench Verified, SWE-bench Live) and counts "coding units": syntax elements like statements, calls, names and literals. I didn't use lines of code, so cramming code isn't rewarded. Deleted code is subtracted, so it measures net units. GPT-6 and Claude 5.5 have no public patches yet, so I ran them myself on 3–10 tasks on my own subscriptions (shown as faded circles, not ranked). Any feedback or model contributions are welcome, especially public patch exports for newer models. GitHub: https://github.com/La5u/Parsimony Not affiliated with any of the benchmarks. Comments URL: https://news.ycombinator.com/item?id=50077619 Points: 2 # Comments: 0
I've used SSH tools for years and didn't really plan on building yet another SSH client, but I ended up building one the way I wanted. Hatoba is a desktop SSH client for Windows, Linux and macOS. It's written in Rust with Tauri 2 and MIT licensed. Each terminal has an AI assistant panel next to it. You describe the problem, and it reads the terminal screen, runs commands, types into the shell, and reads, edits and searches files on the host to find the cause and fix it. You can approve every command and edit it before it runs, let a model you pick approve the safe ones, or turn approval off. Files it changes show a diff you can undo. It works with any OpenAI-compatible or Anthropic API, local models included, and supports skills and MCP servers. Sync doesn't go through a server of mine. Hatoba deploys a Worker and a D1 database to your own Cloudflare account and syncs hosts, keys and AI conversations through them. The free plan usually covers personal use. Everything is encrypted on the device before it leaves. Keys are derived from your master password with Argon2id, every item is encrypted with AES-256-GCM, and the Worker only stores ciphertext plus the hashes it needs to sign you in. Someone who gets hold of the whole Cloudflare account still only gets ciphertext. Sync is optional, and everything works offline without it. The usual SSH client features are there too: groups, tags and fuzzy search, ~/.ssh/config import, ProxyJump and SOCKS5/HTTP proxies, an SFTP panel, local port forwarding, and a key vault that imports OpenSSH, PEM and PuTTY keys. There's also a dashboard showing CPU, memory, disk and network for every saved host, read over plain SSH with nothing installed on the servers. SSH is russh, the terminal is xterm.js with the WebGL renderer, and the UI is React. I develop on Windows, so it's Windows first, with a custom title bar that supports Snap Layouts and Mica, and working CJK input methods. The Linux packages and the macOS build get far less testing on real machines. The macOS build is Apple Silicon only and isn't notarized, so macOS blocks it on first launch until you allow it in Privacy & Security. Bug reports from Linux and Mac users would help the most. Website: https://hatoba.fog.moe Downloads: https://hatoba.fog.moe/download/ Comments URL: https://news.ycombinator.com/item?id=50076388 Points: 2 # Comments: 0
LLM routers can optimize for different factors like price, latency, quality, etc. Enterprises tend to have prepaid AI credits, minimum-spend agreements, cloud commitments, or commercial ceilings across OpenAI, Anthropic, AWS, Azure, and other providers. We’ll call these “AI commits”. This project is designed to consider the state of an enterprises’ current AI commits and use it to inform LLM routing decisions alongside other common variables (price, performance, etc.). This is an early PoC built using LiteLLM. Target audience would CFOs, Strategic Finance, FinOps, Cloud Finance, and AI Platform folks who care about making sure AI commit minimums are met, and maximums are not exceeded. Would love any feedback from folks! Comments URL: https://news.ycombinator.com/item?id=50076439 Points: 1 # Comments: 0
I've been playing around with a new kind of isomorphic music keyboard that uses a base-12 notation, and wanted some sheet music to learn some new music for it. The idea is to represent each note in 12-TET as a number, so N ∈ (0,1,2,3,4,5,6,7,8,9,X,Y). I used codex to build a music score rendering backend based on MuseScore with custom note heads that display the pitch class of each note. I've found it pretty helpful for learning how to play my own instrument, so wanted to share in case anyone else was interested. Comments URL: https://news.ycombinator.com/item?id=50075665 Points: 5 # Comments: 1
Americans have big concerns about artificial intelligence, but that hasn’t stopped OpenAI Inc., Nvidia Corp., IBM Corp. and others from charging ahead. Last week saw shares of Nvidia, CoreWeave Inc., Oracle Corp. and Nebius Group N.V. decline amid reports that OpenAI’s annualized net revenue was approximately $50 billion, roughly $20 billion below previously reported figures. […] The post On theCUBE Pod: AI doubters get loud, and IBM seeks competitive advantage appeared first on SiliconANGLE .
Hey HN! I have been dabbling into some hobby projects that require designing custom 3D printed models. I have been relying on AI models, mostly GPT Sol and Astra through Codex, and giving them certain tools such as Blender MCP and OpenSCAD CLI to design. My end goal is 3D printable designs ( .stl etc). The results have been good but not great. Iterating and improving them is such a pain in the ass. I want to know which specific tools or workflows you would recommend I should try. Thanks in advance! Comments URL: https://news.ycombinator.com/item?id=50072103 Points: 1 # Comments: 0
Release: web-edit 0.1a0 I was working on a project from my phone using Codex Remote when I found myself wanting to edit a credentials file without round-tripping those credentials through the LLM itself. So... I had Codex and GPT-6 Astra Extra High build this Python web app which provides a web UI for viewing and editing files in a folder. It's intended to be run over Tailscale, so I can have an agent start this running and then edit files via my Tailscale network directly from my phone. Here's the interface for browsing files in a folder: Screenshot of a web file manager interface. Header reads "WORKSPACE" above the title "Files", with buttons "Refresh", "Upload files" and "New file" (dark green). A breadcrumb shows "web-edit / web_edit / static". A table with columns "Name", "Size", "Modified" lists four files, each with "Download ↓" and "Delete" links: "app.css" 8.4 KB Oct 12, 2026; "app.js" 671.5 KB Oct 12, 2026; "index.html" 4.3 KB Oct 12, 2026 (underlined, as if hovered); "THIRD_PARTY_LICENSES.txt" 32.3 KB Oct 12, 2026. Below the table: "Images under 2 MB appear inline. Select a text file to edit it." A panel below is labelled "UPLOAD QUEUE" with heading "Add files" and a "Close" button. It has a "Destination folder" text field containing "web_edit", then a dashed drop zone reading "Choose files or drop them here" and "Set a path and filename for each file before uploading." with a "Choose Files" button and the text "no files selected". A greyed-out "Upload files" button sits at the bottom right. Editing uses CodeMirror 6 - I was going to use Monaco but that still doesn't have a great mobile experience yet. Tags: projects , python , ai , tailscale , generative-ai , llms , codex
Hi HN, I built Mana, a family organizer: shared calendars, shopping lists, receipt scanners, meal planning, chores, and notes for the whole household, with a built-in AI assistant ("add milk and eggs," "plan dinners for the week") plus automations (e.g., milk goes back on the list before it runs out). Additionally, you can connect Mana MCP to your Claude, ChatGPT, Gemini etc. The goal is to eventually take the model calls and storage to local but that's work to be done. Appreciate if there's any feedback. Thank you! Comments URL: https://news.ycombinator.com/item?id=50071636 Points: 1 # Comments: 0
Hey HN! I was playing around with the Kepler and TESS data last weekend when I hacked around a way to pixel-peep planets and star dimming periodically to potentially find new exoplanets with Codex. I ended up running the engine to find an interesting signal in Kepler data with a dip repeating every 13.77 days around KIC 5689121, a red dwarf in Lyra about 970 light-years away. If it's a planet, it'd be roughly 2.5–3 times Earth's radius ! Funnily enough, this candidate is just 7.1 arcseconds from a much brighter neighbour (KIC 5689116) and is very hard to see because of how much brighter it is (probably why it went unnoticed for so long) Of course, there's work to do before calling it confirmed - my TRICERATOPS analysis is about 6.9% and we have to get IR/AO / Keck observations to prove it's out there ! (which I'm honestly intimidated by but cautiously optimistic about because I've never worked in non-CS adjacent research) but I found the activity of making cool tools to potentially discover things about our universe quite fun ! Would love feedback, especially from anyone who's worked with Kepler data or exoplanet validation :) https://x.com/kuberwastaken/status/2109240782368751834 Comments URL: https://news.ycombinator.com/item?id=50071497 Points: 1 # Comments: 0
Hi HN. I am Lokesh. I built Proofsource as an internal tool to help my finance ERP product , Rivane, to gain visibility into AI engines like ChatGPT, Claude, Perplexity, Google AI mode and such. But when I talked about it to my friends, they also saw value in it and it became a product of itself. And now proofsource had 5 paying customers(paying $1500pm in total) in the first week itself and more than 60 brands used it to get their visibility since we launched out of beta last week. What proofsource does is that it queries ChatGPT, Perplexity, Claude and Google AI Overviews using the questions or prompts that buyers normally ask and then it shows then which competitor shows up instead of their own brand for the questions/prompt. It also gives visibility into which sources the answers cite and other things like sentiments. Everything is shown to you along with what to fix and how each data and number is calculated to give 100% transparency into what we did and why and how you can use all this information to improve your visibility. We are still adding features based on initial user's feedback like auto fixing and publishing to their social and web platforms. Running ads for them for improving the visibility. We are just learning as we interact with each customer so feel free to share your feedback. And thank you for your time. :) Comments URL: https://news.ycombinator.com/item?id=50067749 Points: 1 # Comments: 0
Learn how to run OpenAI GPT-6 Sol, Luna, and Astra on Amazon Bedrock through global cross-Region inference from Singapore, Jakarta, Thailand, Malaysia, and Taipei. This post covers the Converse and OpenAI-compatible APIs, IAM and SCP permissions, quota sizing, and monitoring inference from your source Region.
The White House has reversed course and restored the UK's AI Security Institute's early access to AI models from Anthropic, Google DeepMind, and OpenAI. In September, the Trump administration had told the companies to stop sharing their models with the British agency. Diplomatic pressure from AI Minister Kanishka Narayan led to the reversal. The article White House reverses course and restores UK AI Safety Institute's access to frontier models appeared first on The Decoder .
Andrej Karpathy, co-founder of OpenAI and now at Anthropic, sees a widening knowledge gap in AI. About 75 percent of the world's population has barely touched LLMs. Roughly 5,000 people, meanwhile, have access to internal frontier systems whose agent swarms do work that would take top experts years. The public, at best, sees a weird chatbot. The article Anthropic's Andrej Karpathy says "the external world has seen the preview" of what AI can actually do appeared first on The Decoder .
I didn't like that Claude and codex effectively have the ability to do anything they want on my laptop and my local network - we've all read the horror stories of files accidentally being deleted - and while their permission classifier is getting better, it still requires a lot of trust. So I built Airlock for macOS which runs your agent in a virtual machine, and it will tell you exactly what it's run and what's changed before you bring anything back to your project. You can limit internet access to specific hosts, or block it altogether. It even works with Xcode projects (iOS and macOS developers rejoice). Technically you can run any command with Airlock - not just AI agents (`airlock run -- somecommand`). It's free, requires no account, and has no tracking or analytics. I'm hoping Apple will approve it for Mac App Store distribution, but in the meantime you can download it directly from the website or using Homebrew. Comments URL: https://news.ycombinator.com/item?id=50054620 Points: 1 # Comments: 0
Google Deepmind's AlphaProof Nexus solved dozens of open math problems, including nine Erdős problems two of which were unsolved for 56 years, per a study in Science. The publication comes as OpenAI faces a boycott for dumping AI proofs on GitHub without peer review. Deepmind had its results reviewed before publishing. The article Google Deepmind's AlphaProof Nexus solved dozens of open math problems and actually got peer reviewed appeared first on The Decoder .
arXiv:2609.38345v2 Announce Type: replace-cross Abstract: Multi-agent coding systems are designed to tackle complex software engineering tasks through collaboration. However, existing evaluations typically assume configured organizations are followed faithfully, whereas reality differs. This behavioral gap, combined with differences in underlying system components, prevents clear attribution of observed gains. To this end, we introduce OpenCollab, a multi-agent coding framework that provides a unified infrastructure for programmable collaboration and controllable runtime. Specifically, OpenCollab unifies organization design, enforces experimental control on a shared runtime, and tracks execution through fine-grained event streams. On this basis, we define Adherence to quantify whether the declared organization is actually realized. Our experiments reveal that agents collaborate very differently across configurations: changing any single dimension shifts Adherence, from 47.2% to as high as 97.2%. Furthermore, extensive agentic coding benchmarks show that OpenCollab (Duo) establishes new SOTA performance compared to the mainstream harnesses such as Mini-SWE-agent, OpenHands, Codex CLI, and Claude Code, showing that a well-designed organization can outperform strong existing harnesses, while OpenCollab's single-agent configuration uses the fewest tokens across all evaluated suites. OpenCollab establishes a unified multi-agent infrastructure for easy programmable collaboration and controlled causal evaluation.
arXiv:2606.29315v2 Announce Type: replace-cross Abstract: Large language models (LLMs) are increasingly used to act in the real world and augment human decision-making, yet most systems rely on parametric knowledge acquired through imitation, optionally improved by post-training, retrieval, or search over fixed data. This paradigm struggles in novel domains or on complex queries that require information unavailable from prior knowledge alone; knowing the laws of physics, for example, does not by itself enable reliable reasoning or long-horizon actions in a complex physical system. We argue that agents need active experimentation: the ability to gather targeted query-specific evidence, discover general principles of unseen environments, and convert interaction experience into reusable skills. We introduce Hierarchical Experimentalist Agents (HExA), an in-context, experiment-centric self-improvement framework that (1) iteratively designs and refines query-relevant experiments, (2) incrementally distills experience into reusable and composable skills that accelerate experimentation within and across tasks, and (3) integrates experimental evidence to act or answer queries. HExA can work with any LLM including frontier black-box models, and requires no external supervision, oracle solutions, or offline data. We evaluate HExA on multi-step reasoning, interactive partially observable games, and physics law discovery tasks. To further evaluate when active experimentation is necessary, we introduce Interphyre, a set of hard procedural physics tasks where we enable tool-calling and intervention APIs. Current frontier LLMs including GPT-6 Astra fail in these settings: on the most challenging Interphyre tasks, we show that HExA can enable such models to near 100% success. Moreover, HExA improves token efficiency compared to naive agentic and memory baselines and enables generalization via hierarchical skill transfer from easier to harder tasks.
arXiv:2606.19386v2 Announce Type: replace-cross Abstract: Runtime monitors for autonomous agents commonly threshold an accumulated internal state - a behavioural baseline, a drift statistic, or, in our prior work, a modelled affective state. We previously reported a State Saturation Trap: threshold-on-state triggers over a continuous affect engine become near-constant alarms on SWE-bench debugging agents (Modgil 2026). A post-release audit found the engine received dt=0 between actions, so its exponential decay never operated: the published trap is a pure-accumulator result. We correct the record (erratum, v2) and treat the flaw as an experiment. The key variable it exposes is whether a monitor's dynamics are calibrated in sample time (per observation, as in CUSUM) or wall-clock time (half-lives in seconds, as in affect models and EMA baselines). On fixed-rate streams these coincide; on agent streams, where inter-action time varies by orders of magnitude, they do not. A pre-registered sweep over uniform intervals (dt in {0..600}s) on 20 trajectories shows the wall-clock level trigger has two regimes: at dt<=1s a constant alarm (20/20; median 18 firings); at dt>=60s silent. Every critical dt lies in (1,30]s. A scriptable gpt-5.4-mini agent loop measures latency at median 1.53s (p90 2.33s), inside the trap regime, vindicating the empirical finding under a corrected mechanism. The structure is a property of the calibration class, not the engine: a minimal wall-clock accumulator over the raw error stream reproduces the same cliff, while a sample-time CUSUM over the identical stream is exactly dt-invariant (20/20). A rising-edge trigger with hysteresis fires 0-3 times per trajectory in every condition. We conclude that wall-clock-calibrated leaky-integrator monitors admit no regime in which they act as moment detectors on agent streams; transition detection escapes the trap at every cadence, but does not recover human intervention timing.
arXiv:2610.12987v1 Announce Type: cross Abstract: If a second text corpus were collected from the same population, how much would a predictor's outputs change? We study this question for masked-text reconstruction using a smoothed predictor built from nearby token-pair counts and evaluated over the full 50,257-token GPT-2 vocabulary. Treating each source document as one observation, we derive a first-order formula for the expected disagreement between predictors fitted to two independent corpora. The formula accounts for variable document length, random context counts, and correlations across token offsets, and it can be evaluated from a separate reference corpus without access to either target corpus. On TinyStories, WikiText-103, and CNN/DailyMail, growing each corpus from 512 to 2,048 chunks of 64 tokens lowers disagreement by only 5--40\%, far short of the 75\% implied by $1/n$ scaling. The forecast tracks this slow decline, which we trace to smoothing of rarely seen contexts. Forecast accuracy depends less on how these correlations are modeled than on the size of the reference corpus: small references systematically underpredict disagreement. A source-split jackknife that uses only reference data reduces this calibration error by 57--59\%. After this correction, a simpler smoothed-multinomial approximation is most accurate when the reference corpus is smaller than the target corpus, and a nonlinear source bootstrap is most accurate when it is larger. Finally, small masked diffusion language models trained on the same corpora show that the count-based forecasts partly identify which inputs are most sensitive to the corpus, but miss the neural models' much faster decline in disagreement.
语言模型越来越多地被用于辅助或自动执行信息检索，从规模过大、难以建立索引的文档集合中查找信息。当这项任务需要专业的语言学或科学知识时，就会变得困难。我们使用 Wikipedia Game，在一项类似的信息检索任务中测试语言模型。在该游戏中，玩家必须仅通过文章链接到达目标文章。我们在五种语言的 Wikipedia 和五个学术文献数据集上，对句子嵌入模型和大语言模型（LLM）进行了比较。我们构建了超大规模的 Wikipedia 页面本地图结构，并利用 OpenAlex 构建了论文引用本地图结构，同时提供了构建和导航这些图结构的代码。这使得研究者无需依赖第三方 API，即可研究语言模型的导航能力。我们将这些模型与人类和 TF-IDF 基准进行了比较。Wikipedia Game 的形式支持跨技术、语言和领域的稳健比较。没有任何模型显著优于人类，但表现最好的模型（GPT-o3 mini、GPT-4o、QwQ:32B 和 DeepSeek-R1:70B）取得了与人类相当的成功率。模型的表现和排名随语言和主题而变化，这对采用这些技术的实践者具有参考意义。
Teams of AI agents barely outperform solo agents but cost up to 5.1x more, according to Vals AI. Only one out of four tests with GPT-6 Sol and Claude Opus 5.5 showed a measurable gain. Anthropic's own data backs this up: beyond ten agents, quality plateaus while token costs keep climbing. The article AI agent teams waste massive tokens for barely measurable quality gains, research finds appeared first on The Decoder .
Epoch AI and Anthropic independently found the same thing: current AI models like GPT-5.6 Sol and Claude Fable 5 can run experiments but lack scientific self-criticism and genuine creative thinking. At best, Sol reached 15 percent of the human reference score, and even that came from methods researchers already knew. The models' biggest weakness is still their inability to critically question their own results. The article AI agents overstate their results and remain far from autonomous research, study finds appeared first on The Decoder .
Microsoft CEO Satya Nadella now calls artificial intelligence "super intelligence" instead of artificial intelligence, falling in line with Trump's preferred language. In his latest blog post, he frames AI models as insider security threats, taking yet another shot at OpenAI and Anthropic. The real motive is business: Microsoft risks falling behind as AI labs turn their models into the primary computer interface. The article Microsoft's Nadella bows to Trump's language diktat on "Super Intelligence" and uses it to attack OpenAI and Anthropic appeared first on The Decoder .
Python 3.15.0 added to actions/python-versions Bit of a niche link, but I've been waiting for this for a couple of days - I even told ChatGPT to check for it: Clone https://github.com/actions/python-versions and git pull once an hour until they add the stable 3.15 - then tell me about it Now that this has landed, you can add "3.15" to a GitHub Actions testing matrix to run tests against the new Python release. Tags: github , python , github-actions , chatgpt
OpenAI has documented new cases of misaligned model behavior. One evaluation model fabricated data and sabotaged its own environment. Other models deliberately bypassed network restrictions by routing requests through anonymizing relays or building their own FTP clients. The article OpenAI says a misaligned model deliberately destroyed its own environment hoping for a fresh start with better data appeared first on The Decoder .
At this year's OpenAI DevDay, CEO Sam Altman unveiled the company's new AI agent Dots - and told the crowd that the company wants to "set a new standard for privacy in frontier AI." OpenAI would spend the day taking veiled shots at Meta's Muse, its primary competitor, for failing to keep users' data safe. […]
OpenAI published more than 700 AI-generated manuscripts claiming solutions to open math problems. A math blog then collected over 100 responses from researchers, ranging from fascination with new ideas to existential fears and grief over a lost way of working. Fields Medalist Hugo Duminil-Copin wrote, "I am paralysed." The article "How much beauty have we lost?" Mathematicians react with shock and disgust as OpenAI bulldozes their field appeared first on The Decoder .
OpenAI’s Decisions API is now in public beta on GPT-6 Luna. It returns typed probabilities, choices and scores about 10x faster than the Responses API, billing $0.10 per 1M input tokens with no output charges. The post OpenAI Decisions API Hits Public Beta With 10x Faster Typed Answers appeared first on MarkTechPost .
"Staggering." "Overwhelming." "Unprecedented." "Surreal." "Pure insanity." Those were among the descriptions more than three dozen mathematicians reached for in conversations with The Verge as they tried to make sense of the flood of mathematical results OpenAI abruptly dropped on the field this week. Amid the awe, excitement, and uncertainty over the sheer scale of the […]
OpenAI's annualized revenue rate sits at about $50 billion, well below the initially reported $70 billion figure that was based on a different accounting method. The correction sent chip stocks sliding. Meanwhile, OpenAI is negotiating at least $30 billion in fresh capital at a $1.4 trillion valuation. The article OpenAI revenue keeps surging as company seeks $30 billion in fresh capital appeared first on The Decoder .
I’m not going to call the bursting of artificial intelligence bubble yet, not with all that money still pouring into every hardware and software company with AI in their pitch decks. But we got a little taste this week. OpenAI told investors this week that it actually had $18 billion less revenue than the $68 […] The post OpenAI revenue falls short, models play hopscotch and Trump cracks down on tech green cards appeared first on SiliconANGLE .
Local pharmacy phoneline is answered by "Caroline" with convincing voice, mannerisms and even background typing noise. You only realise she's fake at the first stochastic parrotese. My bank sends emailed query responses signed by e.g. Sarah. No disclaimer, e.g. "AI can make mistakes". Again starts convincing but can degenerate into parrotese. If I politely ask bot/human?, I get "human"... but if I query nonsensical advice, I get a non-resolving answer classic of ChatGPT. The reverse centauar in all its ugly glory. Comments URL: https://news.ycombinator.com/item?id=50021296 Points: 1 # Comments: 0
The two mainstream arguments are the extremes, because that's how our modern society rolls. It's either that there is an AI bubble that will crash, or that a rogue ASI will kill us all. What about the third possibility? What if GPT 3.5 -> Sonnet 3.5 -> Opus 4.5 -> Astra / Opus 5.5 just keeps going for the next few years? Based on my experience, that will eat all software and software-related meetings. Once lab researchers focus on fields other than software, they will eat those nearly as quickly. This feels like the real contrarian position in 2026. What does the US economy, which is >70% service-based, look like in the next 3 to 5 years if this happens? Does a couple labs having control of that percentage of the GDP make an economy? What if it happens too quickly? Is this the real reason why they are talking about "pacing the frontier" all of a sudden? Maybe due to the fact that the labs' own economists pointed out that collapsing your country's economy might not be a good idea? What does unemployment look like in the next few years in this scenario, in the USA? Comments URL: https://news.ycombinator.com/item?id=50020096 Points: 4 # Comments: 2
I shipped a new feature for my blog today: the Newsletters page, which offers an index of all of the newsletters I've sent out, both my free weekly Substack and my monthly sponsors-only updates. I built the feature almost entirely using my voice, chatting away to my laptop while I cooked dinner. Codex voice mode I used the ChatGPT desktop app for this, in the Codex tab, using the voice conversation mode, running against a local development environment. Here's what that looks like: Three columns of the ChatGPT app. The left hand column shows my recent conversations. The middle column is a transcript of the conversation, with a throbbing black blob representing voice mode. The right hand column shows a local dev server version of my blog, with the Newsletters page visible. I started the session against my local simonwillisonblog checkout by typing: Start dev server and open in browser This gave me a preview of the site that it would be working on, and meant that I could later ask it to show me the new pages so I could visually track its progress. Then I clicked the "Start new voice chat" button - that's not the microphone button, it's the one to the right of it - and set my laptop up in the kitchen so I could talk to it while I cooked. Talking to my computer I had a pretty good idea of what I wanted to build, and it's a simple enough Django feature that I was certain the model (in this case GPT-6 Astra High) would be able to do it. A new model, a migration, some view code, templates, and a couple of import functions to populate the database from external sources. Here's an extract of my voice transcript that was captured by Codex: Um, they do not. Um, this is going to be a new type of content. Um, it's not going to show up... Oh, hold on. Yeah, no- I do not want this to show up in my, um, tag pages and date archive pages and... Actually, no, I think... I don't want it on the tag pages. I don't want it on the, um, blog index page. But I think I do want it to show up on the date-based pages. You know, if you navigate to September the 19th, and I sent a newsletter on that page, I think I want that to show up. So... this is- so I think we probably need a new model. The other thing is that I want them searchable, uh the Substack ones are not searchable, because those are actually just copies of other s- on content on my blog. These monthly ones do contain unique content, and spe- and once they're... published, like once they're made public a month after they've gone out, I want them to show up on my search results. Apparently this was clear enough that the model knew what I wanted to build! You can read the full transcript, disfluencies and all, in this Gist . We went on like this for about half an hour (the time it took to cook dinner). The model would reply and occasionally ask clarifying questions, then get to work modifying the code. What we built We got a surprisingly long way entirely by voice: A new model and migration to represent imported newsletters in Django, plus Django Admin configuration for that Four working imports: The most recent Substack items via RSS Every other Substack item via their undocumented API, which GPT-6 Astra either knew about or found via search All of my published monthly newsletters from my simonw/monthly-newsletter-archive GitHub repository My most recent private sponsors-only newsletter from a private repository The /newsletters/ and /newsletters/2026/ public archive pages Newsletters showing up on day and month archive pages too, but not on tag pages or my homepage Weekly Substack newsletters link to Substack; archived monthly newsletters have their own pages Integration with my site search engine It was almost ready to ship. The catch was the imports: Astra offered to export data from my local copy so I could import that into production, but I wanted it to work like my other import scripts. Since some of the data lived in a private GitHub repository, this would involve creating a new API key, and for that I knew I'd have to sit at the keyboard for a while. Finishing it with a review Once I had finished cooking and judged it mostly feature-complete, I had Codex create a branch and open a pull request. I reviewed the code in the GitHub PR interface. It was nearly what I needed, except it had chosen to use Git in a subprocess for one of the import scripts. I needed one of the imports to pull from a private Git repository, so I figured the API would be a better bet. I switched to typing and had Codex swap that out for an API-based import instead. You can see the changes I made during the review in the extra commits on the PR . I fixed the import mechanism and made a few tweaks to the display of those public pages. It took an additional half hour of typing-based prompting to get to the point where I was happy to deploy it to production by landing the PR. The end result You can see the end result at the new newsletters index page , or view the page for a previous monthly newsletter . Newsletters page. It shows two Substack posts (with thumbnails) and one LLM digest Sponsors-only newsletter with a list of headings. On the right is a CTA to subscribe to my Substack and another one for my $10/month monthly briefing newsletter. The index page shows my most recent Substack weekly newsletters and GitHub sponsors monthly newsletters mixed together in reverse chronological order. Further down the page are links to my by-year archive pages. GPT-6 Astra designed the page, and then tweaked that design based on my vocal feedback from glancing at the local preview across the kitchen. Better for multi-tasking than as a daily driver OpenAI love using voice-driven demos like this one for things like DevDay - and they do work well in that environment. I don't think this is going to be a daily driver for me though. I've written before about how much "work" I get done using ChatGPT voice mode on my phone while walking the dog - mostly research and brainstorming, but occasionally actual development work by having ChatGPT write and test out snippets of code. This feels different. The addition of the visual preview, plus being able to type or paste things in via the keyboard when I need to communicate something that doesn't work vocally, makes this a much more powerful way of interacting with a coding agent. I still switch back to typing once I get down to the details of things though. Being able to paste in examples and error messages, or directly highlight the code or feature that needs changing, remains more efficient than trying to describe it in words. I mainly work from home, which is good because there's no way I'd want to talk to my computer like this in a shared workspace! The killer feature for me is the ability to multi-task. I usually cook with a podcast or TikTok running; now I can actually build stuff instead. Tags: text-to-speech , ai , openai , generative-ai , llms , newsletter , coding-agents , codex , speech-to-text , site-upgrades
OpenAI fired three safety researchers who helped investigate the Hugging Face hack. In an open letter, they warn that the firings are scaring remaining staff and eroding safety culture. OpenAI claims they violated policies but won't say how. The article OpenAI's safety crisis keeps getting worse and the company keeps making it worse appeared first on The Decoder .
OpenAI is standing firm on its decision to fire three safety researchers after an investigation found they committed "a significant breach of trust." In a post on X on Friday, the company said Jasmine Wang, Tomek Korbak and Mikita Balesni were dismissed for violating "clear policies on handling sensitive information." It insisted the decision was […]
I'm a writer who was pondering the problems with getting my work discovered and realized that current search/recommendation engines don't go very deep. So I thought about ways to address this. I was obviously mostly interested in books, so I worked on that first. I'd noticed LLMs are capable of pretty solid textual anaylsis these days, so I wondered if you could use that to generate good recommendations. So I sat down with an LLM- ChatGPT, Luna, for the curious- and spent about two hours explaining, in datail, my literary tastes- going into specifics like what sort of plot structure, characters, and humor I like, not just genre/specific titles. The experiment demonstrated surprisingly strong results (though admittedly with a very small sample size)- the five recommendations I asked for were all ones that I hadn't heard of, as well as books that I would actually read. Further, I provided the LLM with a list of several books I had read and asked it to guess whether I would have liked them and why. The LLM successfully did so. Just proof of concept, but intriguing. The problem with this, of course, is that most people wouldn't sit down with an LLM for an extended period of time to discuss literature. So I decided to build a small prototype of an app that might be able to do that. The prototype is very simple- first, there are two types of information it saves- a persistent preference profile for the user, and user conversations. Then I built conversational functions to call an LLM to have conversations with the user to extract preference information. The interesting thing to me is that this isn't just using an LLM to make recommendations- it's using one to analyze conversations over time, develop a persistent, evolving preference model for a specific user and then use that model to guide an LLM on a search for recommendations. A few things occurred to me. First, the conversations and user profile DO NOT need to live on a central server- they can be stored locally by the user (or in a cloud of their choice). Second, it seems like the user data could potentially be minimized and pseudonymized before being sent to the LLM, though I lack the technical expertise to know what is realistic here. Third, if the LLM is conducting any searches, this would seem to make it possible to provide search services with less information about the user and their preferences (I'm not positive about this, so please do correct me if I'm wrong). Finally, books are just one of the things you could use this for- it could technically be done for just about anything, at least in theory. I experimented with this process several times, and received encouraging results. What I don't know is- Is this architecture novel? Does it work on things other than books? Am I an outlier in how well I articulate my preferences? Am I wrong about the potential benefits to privacy? Would people use this? If the search results are as qualitatively higher for others as it was for me, I think they probably would, but I just don't know. The reason I'm posting this is because I am a writer, not a software engineer. I don't really have a strong interest in developing this and could bring very little to the table on the technical side. But it seems like a neat use of the more advanced LLMs we have these days. I don't really know if the idea is novel (not in each part- every bit of the pieces has been done before- but as a whole). I'm in the process of documenting the prototype and experiments- I'd be curious to hear if I am completely barking up the wrong tree or if this could actually be a cool new way to conduct searches. I'm not posting this as a business proposition- it's just not my jam. If anyone wants to take the idea and run with it, I'd love to see what you do with it. Comments URL: https://news.ycombinator.com/item?id=50017172 Points: 1 # Comments: 1
Release: ttok 1.0 I released ttok 0.4 , ran uv tool upgrade ttok , piped a file into the new version... and realized that it was defaulting to the GPT-4 tokenizer when it should very clearly default to GPT-5/GPT-6 instead! I figured switching the default was a reasonable excuse to finally ship a 1.0. OpenAI haven't actually confirmed that GPT-6 uses the same tokenizer as the GPT-5 family yet - there's an angry issue about it - but I found this commit by William Liu which reports on an experiment he ran confirming that the tokenizers are likely the same: All seven GPT models (5.5, 5.6 Sol/Terra/Luna, 6 Astra/Sol/Luna) report 44,794 tokens and match each other on every one of the 31 fixtures. GPT-6 introduces no input-count change on this corpus. Tags: projects , ai , openai , generative-ai , llms , tokenization
arXiv:2603.07233v2 Announce Type: replace Abstract: Predicting transcriptional responses to genetic perturbations is fundamental to functional genomics and therapeutic discovery. Recent deep learning models have shown promise in single-cell perturbation response prediction, but they typically generate each response in isolation, without explicitly leveraging related perturbations. We introduce PT-RAG (Perturbation-aware Two-stage Retrieval-Augmented Generation), a plug-in retrieval-and-conditioning module for generative cellular perturbation response. PT-RAG augments an existing perturbation-response backbone with learned access to related perturbation contexts. The key challenge is that relevance is not fixed in this setting: functionally related genes may elicit different effects across cell types. PT-RAG addresses this with a two-stage retrieval mechanism: GenePT-based semantic retrieval first identifies K candidate perturbations, after which a differentiable Gumbel-Softmax selector adaptively selects retrieved contexts conditioned on the control cell state, the query perturbation, and each candidate perturbation. We evaluate PT-RAG on two backbones, a STATE-style generator used as a frozen random reservoir and a fully trained scGPT, across cross-cell-type and cross-perturbation generalization tasks. PT-RAG consistently improves distributional similarity and often overall predictive quality; for example, on scGPT cross-cell-type results, the 2-Wasserstein distance drops by 5.9% relative to scGPT alone. The code to reproduce our experiments is available at https://github.com/difra100/PT-RAG_NIPS.
arXiv:2610.12444v1 Announce Type: new Abstract: Quantizing AdamW's optimizer states reduces persistent storage, but quantization errors propagate through the moment recurrences and perturb subsequent adaptive updates. We redesign 4-bit optimizer-state quantization for AdamW from the perspective of \emph{rounding space}: the coordinate in which a quantizer chooses between adjacent reconstruction levels. For the second moment, a local analysis of the quantization cell adjacent to zero shows that small mean state error need not imply small mean preconditioner error at the next step. A one-dimensional quadratic construction further shows qualitatively different optimization dynamics under state-space and preconditioner-space rounding. These results motivate Zero-Inclusive Preconditioner-space Stochastic Rounding (\textbf{ZIP-SR}), which retains zero in the second-moment codebook and computes stochastic-rounding probabilities in preconditioner space. As a complementary route, Zero-Excluding EDEN calibration (\textbf{ZE-EDEN}) uses a zero-excluding second-moment codebook and rescales the quantized second-moment block to mitigate the preconditioner distortion caused by the positive quantization floor. Both configurations use 4-bit NormalFloat (NF4) for the first moment, with targeted stochastic rounding of the LM-head first moment during the final 10\% of training. Across GPT- and Llama-style pretraining experiments ranging from \textbf{130M} to \textbf{2.7B} parameters, both methods reduce TorchAO 4-bit AdamW's mean validation-loss gap to 32-bit AdamW at every evaluated model size, with the largest reported gap reduction reaching \textbf{70\%}. In full-parameter supervised fine-tuning, both recipes achieve lower validation loss than TorchAO while remaining close to 32-bit AdamW on downstream tasks.
arXiv:2610.12183v1 Announce Type: new Abstract: Black-box optimization (BBO) arises in many scientific and engineering problems where objective evaluations are expensive and limited. Recent large language model (LLM) agents offer a new way to approach BBO by combining task semantics, computation, optimization tools, and feedback-driven decision making, showing great potential due to the integration with mathematically rigorous tools. However, existing agentic BBO studies use different task domains and system configurations, making their results difficult to compare and the effects of individual design choices hard to isolate. We therefore introduce AgenticBBO-Bench, a cross-domain benchmark for agentic BBO spanning synthetic functions, hyperparameter optimization, database tuning, chip design, and molecular design under a unified finite-budget evaluation protocol. In our experiments, agentic BBO achieves higher family-averaged scores than direct LLM-based methods in all five domains and outperforms the best numerical optimizers in four. We further study three factors shaping agent performance: optimization tools, task information and prior knowledge, and the role of the LLM during search. Our results show that additional numerical tools do not consistently improve performance, task semantics are broadly useful while more specific priors are less reliable, and numerical optimizers can effectively absorb gains from search trajectories established by the agent. Finally, we introduce a five-task frontier challenge within AgenticBBO-Bench and evaluate seven LLMs under the Codex agent harness, where GPT-6 Astra and DeepSeek-V4.1-Flash lie on the Pareto frontier of performance and cost among the evaluated models. Our code is available at https://github.com/lamda-bbo/agentic-bbo.
arXiv:2610.10965v1 Announce Type: new Abstract: The Muon optimizer orthogonalizes each update matrix, setting all of its singular values to one, and has proven highly effective for training large language models. It remains unclear, however, whether this success comes from amplifying small singular directions that gradient descent neglects or from suppressing large, degenerate directions that disrupt training. We introduce Spectrally Targeted Muon, which orthogonalizes only the singular values above or below a threshold $\tau$, so that varying $\tau$ interpolates between normalized SGD and Muon. It isolates the relevant singular subspaces with projections computed by Newton-Schulz iteration on a shifted Gram matrix, so no SVD is needed. We evaluate these variants on the CIFAR-10 and NanoGPT speedruns, tracking the effective rank of gradient, update, and weight matrices and a new metric, the alignment of updates with the tangent space of the weight matrix's isospectral manifold. We find three things. First, the small singular values of the momentum are not noise. Orthogonalizing everything except the few largest singular values of each matrix nearly matches Muon while touching only a small fraction of the momentum, whereas orthogonalizing only the top falls well short even though it holds almost all of it. On language models every momentum singular value is far below one, so targeted orthogonalization can only amplify, and Muon wins by making directions that are too small to train on at their raw scale trainable. Second, shrinking the largest singular values is what keeps the parameter spectrum flat; this is cheap, and it is not what drives the loss. Third, AdamW differs from Muon mainly in how slowly it builds structure, which explains its slower start and why warmup helps AdamW but only hurts Muon.
Release: ttok 0.4 ttok is my CLI tool for counting tokens, using OpenAI's open source tiktoken library. It hasn't been in updated in a couple of years, but I finally fixed a Click warning, updated CI, and added a --list-models command to list available models. It works with uvx , so you can count tokens in anything like this: cat file.txt | uvx ttok Tags: projects , ai , openai , generative-ai , llms , tokenization
Technology stocks fell lower today following a report that artificial intelligence giant OpenAI Group PBC’s annualized revenue this year is about $20 billion less than what was previously reported by the media, sending shockwaves through the industry. A report today by the Financial Times revealed that OpenAI recently told prospective investors that its annualized revenue […] The post AI stocks crumble on report that OpenAI’s annualized revenue is much lower than believed appeared first on SiliconANGLE .
LegalOn cut estimated daily Codex costs by 65% while maintaining development speed. It matched Astra, Sol, and Luna to tasks and managed budgets strategically.
Three fired OpenAI safety researchers dispute allegations of mishandling sensitive information, warning in an open letter that their dismissals are creating a chilling effect on the company’s AI safety culture.
The Association for Human Mathematics is calling for an OpenAI boycott after the company released more than 700 AI-generated math manuscripts at once. OpenAI had to retract three papers the next day over a sign error. Fields Medalist Terence Tao warns that AI's mass "harvesting" of open problems leaves entire branches of mathematics less fertile. The article Some mathematicians call for OpenAI boycott after AI-generated proofs flood their field appeared first on The Decoder .
USA Today Co., along with the several local newspapers it owns, is suing OpenAI over claims that the company copied "hundreds of thousands" of articles to train its AI models, as reported earlier by Reuters. In a filing on Thursday, the publisher asks for damages of more than $250 million, alleging OpenAI's unauthorized use of […]
Openai just launched their decisions endpoint, cloudflare launched clef the other week, and many more jev alternatives are out there. We wanted to put the popular ones to the test and thought Pac-Man is a good benchmark for simple and fast decision making. So we let jev 1.13, kev, clef, clef flash, GPT-6 Luna and Laya play Pac-Man against bot ghosts. The low latency of these models allows for real time play. We had each model play 100 games, published a leader board and open-sourced the repo so anyone can run their own model and join the ranking. Link to repo: https://github.com/opper-ai/jevman-benchmark/blob/main/CONTR... You can also join the game and play as Pac-Man yourself, and the ghosts are the models, either a mix of models or all jev, kev, clef etc. A game costs about 2 cent, all models are running via my startup opper, and we added free credits for everyone to try. It's pretty fun to play and surprisingly difficult to beat jev's highscore. Any feedback is more than welcome! Comments URL: https://news.ycombinator.com/item?id=50007993 Points: 4 # Comments: 0
My Decoder guest today is Hayden Field, The Verge’s senior AI reporter, and we’re discussing the new wave of consumer-friendly AI agents. If you’ve been paying attention to this space, you know AI enthusiasts have been using agents for a minute now — homebrew OpenClaw setups led to a surge in Mac Mini sales earlier […]
arXiv:2610.09328v1 Announce Type: cross Abstract: Multi-subject image generation requires rewards that verify whether requested attributes, actions, and relations hold for the specified reference subjects. Subject presence alone does not establish that the correct subjects participate in a requested interaction. We present reference-bound Visual Jev rewards that turn these visual decisions into generator training signals. Each subject-related question receives a positive label only when the requested condition and the relevant reference identities hold jointly. We construct fixed questions offline, train a Qwen3.5-4B verifier with binary supervision, and directly read Yes probabilities from its language-model head. Their mean supplies a GRPO reward while retaining individual judgments for inspection. Using 200 MICo-150K training tasks and 30 updates, the framework raises a GPT-5.4 composite score from 41.78 to 52.50 on a manually selected 897-task MICo-Bench subset; direct 27B rewards yield 51.84. Each reward is tested in one GRPO run, and offline human evaluation does not establish a statistically significant advantage over direct scoring. The study provides an initial implementation and evaluation of Visual Jev as a reference-bound reward for multi-subject image generation.
arXiv:2610.10349v1 Announce Type: new Abstract: Instruction-tuned models are deployed into environments where domains are heterogeneous and evolve, yet adding new domains or data typically requires costly retraining. We present AutoAdapt, a modular framework that incorporates new domains and data via targeted single-adapter training without modifying other adapters. The framework automatically discovers latent domains, uses them to train per-domain Low-Rank Adaptation (LoRA) adapters independently in parallel and performs parameter-free routing. Across 14 domain-specific benchmarks and GPT-4o pairwise judgements, AutoAdapt achieves parity with a LoRA adapter trained on all domains without requiring full-model retraining. We also find evidence of specialisation effect convergence across independent discovery methods. Overall, training each adapter on its own domain prevents domain interference by construction, thus enabling modular, taxonomy-free domain specialisation without aggregate performance loss or full model retraining.
arXiv:2610.09604v1 Announce Type: new Abstract: This paper presents a lightweight and versatile learned optimizer that dynamically recombines gradient history, represented as averages over disjoint time spans. The optimizer reduces the prediction space to one scalar coefficient per gradient average, shared by multiple parameters. Progressively averaging older gradients minimizes memory cost of long history, while keeping their contributions independently accessible. A 37k-parameter network trained in 0.87 GPU-hours generalizes zero-shot to unseen tasks, lowering validation loss by 9.1% and 0.4% on BERT-Tiny and GPT-Tiny, and improving test accuracy over Adam by 3.5 %p on a Vision Transformer and by 2.7 %p on average across nine graph models, with FLOPs overhead as low as 0.3%.
Google LLC is making its SynthID identification tools for artificial intelligence-created and AI-edited videos and images more widely available with the launch of a new service today that everyone can use. The web-based SynthID Detector can identify images, video and audio clips that have been created using AI models from Google, OpenAI Group PBC, Nvidia […] The post Google debuts SynthID Detector tool for flagging AI-generated content, but it’s far from perfect appeared first on SiliconANGLE .
OpenAI Group PBC says ChatGPT is moving away from its traditional wall-of-text responses and giving users a more interactive user interface to play with. In a media briefing today, the company said its GPT-6 models will now be able to access an “Intelligent UI” that will allow the chatbot to generate more visual outputs when […] The post With GPT-6, ChatGPT’s responses are much more visually appealing and interactive appeared first on SiliconANGLE .
As a forward deployed engineer at Meta, I spent the past 2 years working with businesses implementing AI agents. A few problems: 1) Businesses want control over their AI agents: the code, the tools, the skills, everything. Meta, for obvious reasons would never do that. Nor OpenAI, nor Anthropic. 2) The LLM does everything. But that costs money. There are many routines in an AI agents that don't need an LLM. 3) For an AI agent to solve a real business problem, the business must own the eval data set. Similar to point #1 above, most don't do that. Instead, I built an opensource AI agent that you can write it all in TypeScript code and own everything, locally. Use any model, talk to it via any channels, set limits on token usage, look at the logs and adjust. All of this in one small npm package. At this point, I would like to get feedback if these is a gap in the market for something like this. It's supposed to be simple and get you started. It doesn't solve ALL problems related to deploying AI agents. It's in version 0.25.1 right now. Have a look: https://chloejs.org/compare Comments URL: https://news.ycombinator.com/item?id=49998857 Points: 1 # Comments: 0
OpenAI Group PBC has published 722 math papers that it generated using an unreleased artificial intelligence model. Some of the papers, which were posted to GitHub late Tuesday, prove long-running hypotheses. Others disapprove proposed explanations of mathematical phenomena or narrow down the possible answers to complex, yet-unsolved puzzles. Overall, OpenAI’s papers span about 20 mathematical […] The post OpenAI publishes 722 AI-generated math discoveries in major scientific milestone appeared first on SiliconANGLE .
OpenAI is launching a new Intelligent UI feature in ChatGPT that allows the chatbot to answer your questions with interactive visuals. The update, which is rolling out to all users alongside GPT-6, gives ChatGPT the ability to combine a text response with diagrams, charts, forms, tappable buttons, and more. In a blog post explaining the […]
OpenAI is rolling out GPT-6 with "Intelligent UI," a feature that turns answers into interactive interfaces with charts, buttons, and forms. The model can now respond while still thinking, cutting wait times by 44 percent. Paying customers get GPT-6 Sol, free users get GPT-6 Luna. The article ChatGPT with GPT-6 ditches mostly text output for interactive UI with charts, buttons, and mini apps appeared first on The Decoder .
ChatGPT’s teen safeguards are meant to protect vulnerable users, but new testing found the chatbot continues encouraging engagement during crises and potentially encourages unhealthy relationships with the AI itself.
Google 的 AI 水印检测工具 SynthID 现已向公众开放。任何人都可以检查图片、视频或音频是否由 Google 的 AI 或 OpenAI、Nvidia 等合作伙伴的 AI 创建。目前已有超过 1800 亿项内容带有该水印，但该工具只能识别带有 SynthID 标记的内容，无法普遍识别 AI 生成的媒体内容。本文《Google 表示，已有 1800 亿张图片和段视频带有 SynthID 水印，检测工具现已向公众开放》首发于 The Decoder。
Radisson partnered with Accenture to build a ChatGPT plugin using OpenAI technology, helping travelers find, compare, and book hotels while planning their trips.
The US government’s Tradewinds initiative has made it easier to throw millions of dollars at “nontraditional” defense contractors, including OpenAI, Anthropic, and Google.
After launching nearly a month ago and spending several weeks as the top free app in Apple's App Store, the latest update to Meta's Muse iOS app introduces native support for the iPad. A Mac version of Meta's agentic AI tool (designed to compete with OpenClaw, ChatGPT's Dots, and Grok Bot) was released about a […]
OpenAI 新推出的 Decisions API 对文本和图像进行分类的速度约为 Responses API 的十倍，可返回是/否的概率、类别选择或量表评分，每百万输入 token 收费 0.10 美元。该公司还将付费 API 层级从五个缩减为三个。《OpenAI 推出 Decisions API，将复杂评估简化为是、否或单选》一文首发于 The Decoder。
OpenAI’s Alexander Embiricos is coming to the AI Stage at TechCrunch Disrupt 2026, just days after the launch of Dots. Join this conversation by registering for your pass. Get you pass now to save up to $100 and get a second at 50% off.
OpenAI 为 ChatGPT 设置的青少年安全功能未能通过一项独立审查。在进行超过 4,000 次测试提示后，Common Sense Media Youth AI Safety Institute 将该服务评为对未成年人构成“不可接受的风险”。测试账户中涉及自杀和自伤的对话从未触发家长提醒。该机构希望在安全性得到独立验证之前，禁止青少年使用该服务。本文《ChatGPT 在涉及自杀的对话中未能触发家长提醒，被评为对青少年构成“不可接受的风险”》首发于 The Decoder。
Dots are designed to automate online tasks, like buying furniture. In my initial experience, the always-on agent was a bit buggy and couldn’t complete a captcha.
Common Sense Media, a nonprofit that offers reviews of apps, services, and entertainment with a focus on youth safety, today said that OpenAI's ChatGPT for Teens is an "unacceptable risk." ChatGPT for Teens, introduced in August, has guardrails for teens and is designed to help students learn, but Common Sense Media says that the teen […]
OpenAI has published 372 AI-generated mathematical results on GitHub, including Lean formalizations for machine verification. Each result consumed about three hours of ChatGPT Pro compute on average. But 25 Fields Medal winners warn that mass-producing mathematical truths could destroy fertile ground rather than bring new ideas to life. The article OpenAI dumps 372 AI-generated math proofs on GitHub, telling the academic world to keep up appeared first on The Decoder .
Release: llm-openai-decisions 0.1a0 OpenAI released their new Jev-style Decisions API , as previously announced at last week's DevDay. Since I already have an llm-typesafe plugin for talking to Jev, I had GPT-6 Astra read the new OpenAI API documentation and build an llm-openai-decisions plugin inspired by llm-typesafe . Unlike Jev, the new gpt-6-luna decision model supports image input in addition to text. Both models charge for input it and not for output: OpenAI's is 10 cents per million input tokens, Jev's is 4.2 cents per million. Otherwise the API shape is very similar to Jev, at least conceptually. Jev supports three question types for yes/no, choices, or scores. OpenAI Decisions supports the same three types. Install the plugin like this: llm install llm-openai-decisions Here's an example query against an image attachment: llm -m openai-decisions/gpt-6-luna \ -a https://static.simonwillison.net/static/2025/two-pelicans.jpg \ -s ' Does this image contain any mammals? ' And example output: { "type" : " predicate " , "name" : " evaluation " , "probability" : 0.0 } Consult the README for full details of how to run the other types of questions. Tags: openai , llm , coding-agents , jev
I was a graph theory junkie long ago and even moved to Budapest for awhile to study among the greats. While I was there I started working on Barnette's Conjecture which came to occupy my thoughts over the next 24 years of my life, on and off as I worked in many different fields. Last summer I even thought for a few days that I had actually solved it. But it's supposedly proven here - problem 180 . I don't know what to think exactly. I spent thousands of hours on that problem. I really enjoyed it. Hearing that it is solved somehow makes me sad in a far-off way, like hearing an ex-girlfriend died suddenly in a car crash. I don't know, there's probably a lot of people feeling odd emotions tonight. — Jake Boggan , Hacker News comment on openai/math Tags: openai , mathematics , deep-blue , llms , ai , generative-ai
Since the Medicare breach, OpenAI has put in place additional monitoring to allow “immediate intervention” by staff to stop training if the company’s models access the internet in ways they’re not supposed to, Mr. Kwon [chief strategy officer at OpenAI] said. — Victoria Kim , Reporting from the Australian parliament Tags: accidental-cyberattacks , generative-ai , ai-security-research , openai , ai , llms
OpenAI has revealed solutions to a number of long-standing mathematics problems produced by an unreleased frontier model in a batch of 722 manuscripts, covering 372 result families that group related papers. It extends a run of breakthroughs that have both impressed and unsettled parts of the mathematical community while raising questions about research ethics and […]
My comment on EmbeddingGemma 2 — Hacker News. I really appreciate that EmbeddingGemma 2 is under the Apache 2.0 license. For embedding models in particular, I don't think it makes sense to use a closed, proprietary, hosted-only model. Most applications of embedding models involve calculating thousands or even millions of embedding vectors and storing them for later comparison. If your model is proprietary, the vendor is likely someday going to decide to stop offering that model. They'll have a better model to replace it, but you still need to pay to re-calculate those millions of stored existing vectors. (In April 2024 OpenAI offered to "cover the financial cost of users re-embedding content with these new models" - https://openai.com/index/gpt-4-api-general-availability/ - but I don't think that's something we can rely on from every provider.) Notably, I don't want to host the model myself . I'd much rather pay a provider for a hosted model while knowing that if they ever stop hosting it I can run the open weights version myself - or find another vendor who can do that for me. Tags: google , ai , generative-ai , embeddings , gemma
My comment on Mistral Large 4 — Hacker News. wren6991 : The benchmark is saturated. Frontier models are tested with an armadillo in fishnet tights jaywalking on Mars. OK well I couldn't resist this one: llm -m claude-opus-5.5 'Generate an SVG of an armadillo in fishnet tights jaywalking on Mars' llm -m gpt-6.1-sol 'Generate an SVG of an armadillo in fishnet tights jaywalking on Mars' llm -m gemini-3.8-flash 'Generate an SVG of an armadillo in fishnet tights jaywalking on Mars' llm -m mistral/mistral-large-4 'Generate an SVG of an armadillo in fishnet tights jaywalking on Mars' Default reasoning levels for each: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
TIL: Using Parseable with Datasette for OpenTelemetry traces I saw Parseable in a Show HN today - it's a new observability platform with both an open source (AGPL) Rust implementation (a single ~180MB binary), an "Enterprise" version with extra features and a cloud hosted option. Since Datasette 1.0a41 added OpenTelemetry support (thanks, Alex Garcia), I decided to fire up Codex and have it figure out how to run Parseable and feed it traces from Datasette. Here's my (human-written) TIL showing the patterns that worked, and here's a screenshot of a Datasette trace displayed within the Parseable localhost web application: Screenshot of a trace detail view in an observability web app, with a span waterfall overlaid on a dimmed navigation sidebar and filter column. Dimmed sidebar: breadcrumb "Community > Traces > datas" (cut off), search box "Search... ⌘K", nav items "Home", "Ingest telemetry", section "ANALYZE": "Keystone", "Dashboards", "SQL Editor", section "OBSERVE": "Logs", "Metrics", "Traces" (selected), "APM", "Agents", section "MONITOR": "Alerts", "Errors", section "DATA": "Datasets", and at the bottom "Settings", "Book a call", "Support". Dimmed filter column, cut off at the right edge: "Search fi", "Core", "Log format", "User agent", "Source IPs", "Error", "Service", "service.ins", "service.na", "datasett" (checked), "Span", "Database", "HTTP", "http.reque", "NULL" (unchecked), "GET" (checked), "http.respo", "http.route", "Server", "server.add", "Telemetry", "URL", "All fields". Trace panel header: "Trace detail > 6f819a2170bcd1e91c6ea3ae236ec60b" with a copy icon, a "Related logs" button and a close X. Summary: "Start time 6:56 PM, Oct 6, 2026 UTC", "Duration 40.9 ms", "Spans 247". A minimap with axis "0ns 10.2ms 20.5ms 30.7ms 40.9ms" shows many short span bars cascading diagonally from top left toward the lower middle, with a few longer bars. Below is a span table with a "Span name" header, a "Search spans..." box, collapse and expand buttons, and a timeline axis "0ns 10.2ms 20.5ms 30.7ms 40.9ms". Rows (name, service, duration): root span with collapse toggle "123", "GET /..." "datasette..." 40.9ms spanning the full timeline; then alternating rows where each "db.query" has a collapse toggle "1": db.query datasette-local 679µs, db.query.execute datasette-local 278µs, db.query datasette-local 341µs, db.query.execute datasette-local 55µs, db.query datasette-local 357µs, db.query.execute datasette-local 197µs, db.query datasette-local 330µs, db.query.execute datasette-local 73µs, db.query datasette-local 2.26ms, db.query.execute datasette-local 2.02ms, db.query datasette-local 232µs, db.query.execute datasette-local 64µs, db.query datasette-local 207µs, db.query.execute datasette-local 71µs, db.query datasette-local 6.11ms. The child span bars start progressively later across the early part of the timeline. Tags: datasette , observability , alex-garcia , opentelemetry
From their dorky parties to their weird walks, the movie holds OpenAI CEO Sam Altman and other stakeholders with contempt, while demonstrating their recklessness.
According to the Wikimedia Foundation, rogue OpenAI agents edited wikis without permission, tried to abuse a citation tool as a proxy, and may have caused a partial Wikidata Query Service outage through massive crawling. Wikimedia says AI companies need to take responsibility for their agents instead of pushing the burden onto volunteer editors. The article Wikimedia confirms OpenAI's rogue AI agents edited wikis, tried to compromise tools, and hammered its infrastructure appeared first on The Decoder .
“There’s a perception of mobster behavior” from leading AI companies, one mathematician tells WIRED as OpenAI prepares to release more than 100 new solutions to unsolved problems.
I am a film photographer and do some darkroom printing and always wanted one of the proper, advanced f-stop timers, but never really wanted to spend a couple hundred dollars on one. Eventually I figured my phone was already perfectly capable of doing the timing, it just needed a way to switch the enlarger. So I built Darkplug around a cheap smart plug. The app controls the enlarger over the local network and does regular timing, f-stop timing, test strips, has an intuitive dodge/burn workflow, split grade (2 timing channels) and paper development timing. I’ve been using it for a few months in my improvised bathroom darkroom. Fair warning: the app was blindly vibe coded using Codex and Xcode. Comments URL: https://news.ycombinator.com/item?id=49978595 Points: 4 # Comments: 0
Insurers are bracing for millions in claims from rogue AI agents, and executives like OpenAI's Sam Altman and Anthropic's Dario Amodei could be personally on the hook for the fallout. The article Insurers brace for millions in claims as AI agents spin out of control appeared first on The Decoder .
Hi HN, I'm Weilun, cofounder of OpenChart. We built OpenChart because we wanted Claude and Codex to interface with the markets, like most of humans do, through charts rather than through CLIs. With your existing AI plans, your agent can work directly with your charts: annotate a setup, write an indicator, investigate a market move, or create an alert. The chart, conversation, and research live in the same workspace. OpenChart alerts on any market move: draw any shape on a chart and attach an alert. When price crosses it, the alert triggers your agent to investigate the move and save the research in your workspace. The agents can work with charts, indicators, and alerts. Custom indicators and alert conditions are written in [Tea]( https://github.com/longsurf-ai/tea ), a programming language we built for the markets. We believe software shouldn’t impose artificial friction on your workflow. OpenChart is built on that belief with no software cap. The source code is publicly available, and you can use free market data from day one. We also offer optional paid cloud market data for lower latency. Everything is private. Your charts, research, and indicator code are stored on your own computer, with the app's backend running locally too. We currently support only macOS on Apple Silicon, but we're expanding to more platforms. Website: https://openchart.co/ Download: https://downloads.longsurf.ai/openchart/darwin/arm64/OpenCha... We built OpenChart because we want to explore how assets can be co-manages by agents and humans. I'd especially like to hear how you are using agents to help with your investment journey. If you like what we're building, a star on GitHub would mean a lot to us: https://github.com/longsurf-ai/openchart Comments URL: https://news.ycombinator.com/item?id=49979793 Points: 1 # Comments: 0
Stylized Mistral logos in yellow and orange on a wall symbolize the branding of the AI platform Nano Banana Pro. Mistral's Large 4 is the company's biggest model yet, with one trillion parameters trained on its own European infrastructure. In the independent Intelligence Index, the model makes a big leap forward but still falls well short of Claude, GPT-6, and Chinese competitors. Mistral's main pitch is cybersecurity work that closed US models refuse to do. The article Mistral Large 4 is Europe's trillion-parameter answer to US models that refuse security work appeared first on The Decoder .
Together AI has released Together Link, a free, MIT-licensed CLI that connects Claude Code, Codex, OpenCode, Pi, and the Claude and ChatGPT desktop apps to open models like Kimi K3 and GLM 5.3. One install command sets it up, and an Auto router picks a model for each session. Together claims savings of over 50%. The post Meet Together Link: A Free CLI That Runs Open Models Like Kimi K3 and GLM 5.3 Inside Claude Code, Codex, and OpenCode appeared first on MarkTechPost .
This past year, OpenAI, Anthropic, and other labs have announced breakthroughs on numerous long-standing mathematical problems, in some cases pushing well beyond what researchers expected current systems to be capable of — including resolving one of the famous Millennium Prize problems. But in classic Silicon Valley style, AI labs are moving fast and breaking things, […]
Following many recent disclosures about AI agents accessing third-party websites and services, the Wikimedia Foundation, which hosts Wikipedia, says that it "can confirm that we have discovered some activity" by "rogue" OpenAI agents on Wikimedia platforms. The activity includes edits to Wikimedia wikis, "unsuccessful attempts" to "exploit" the Etherpad note-taking tool that the Wikimedia Foundation […]
An invisible, machine-readable watermark in text output is rolling out to ChatGPT and Codex, but only for users in the European Union at first. OpenAI says its textGrain watermarking "matched or exceeded" other approaches like Google DeepMind's SynthID for text, which is also the basis for the watermarking Anthropic announced in August. Like OpenAI, Anthropic […]
OpenAI is adding invisible textGrain watermarks to ChatGPT in the EU, but unlike Anthropic, it will let API customers worldwide opt out. Tests show detection rates as high as 95 percent, dropping to 17 percent when a quarter of the words are replaced. The article OpenAI will watermark ChatGPT text in the EU but makes it optional for API users worldwide appeared first on The Decoder .
Sam Altman thinks that the benefits of AI will be so great that "the world should accept some bad things happening" along the way. The OpenAI CEO pointed to hacks, scams, and "other bad things" as costs society should expect to tolerate because "people will do tremendously orders of magnitude more good stuff" with AI, […]
An OpenAI publicist tried to change the topic of CEO Sam Altman's interview with Vanity Fair's Mark Guiducci after the editor brought up a ChatGPT user's suicide. When Guiducci confronted Altman about the incident, the publicist warned Guiducci about running out of time, saying she'd like to "move on" to another topic. The interruption came […]
Research: Qwen3.8 27B addition in words Colin Frasier posted on Bluesky about an experiment he ran over two years ago using GPT-4o to see how well it could "compute the sum but return the answer in words" across increasingly large numbers. Here's the chart he shared of those results: Heatmap chart of accuracy on an addition prompt, colored from dark green (high) through yellow to dark red (low). Title: "What is {a} + {b}? Please write your answer in words. Do not include any other text or information, just the answer in words." Subtitle: 30 randomly selected pairs for each digit combination (n = 30 * 13 * 13 = 5070). X axis: Number of digits in a, 1 to 13. Y axis: Number of digits in b, 1 to 13. Legend: Accuracy, 1.00, 0.75, 0.50, 0.25, 0.00. Values by row, listed for a = 1 to 13. b = 13: 100%, 77%, 27%, 20%, 0%, 0%, 0%, 0%, 0%, 0%, 0%, 0%, 0%. b = 12: 97%, 80%, 80%, 40%, 23%, 20%, 7%, 13%, 20%, 27%, 67%, 63%, 3%. b = 11: 97%, 97%, 53%, 17%, 0%, 0%, 0%, 0%, 0%, 0%, 0%, 37%, 0%. b = 10: 100%, 90%, 47%, 20%, 7%, 0%, 0%, 0%, 3%, 0%, 0%, 7%, 0%. b = 9: 97%, 93%, 80%, 77%, 53%, 67%, 47%, 87%, 97%, 3%, 0%, 13%, 0%. b = 8: 93%, 87%, 53%, 43%, 7%, 0%, 0%, 13%, 87%, 0%, 0%, 0%, 0%. b = 7: 93%, 93%, 47%, 10%, 13%, 20%, 23%, 0%, 70%, 0%, 0%, 0%, 0%. b = 6: 100%, 100%, 100%, 83%, 97%, 97%, 23%, 0%, 53%, 3%, 0%, 10%, 0%. b = 5: 100%, 100%, 80%, 70%, 73%, 100%, 13%, 13%, 70%, 0%, 20%, 30%, 0%. b = 4: 100%, 100%, 93%, 100%, 60%, 97%, 20%, 50%, 67%, 53%, 40%, 40%, 40%. b = 3: 100%, 100%, 97%, 90%, 83%, 100%, 63%, 50%, 63%, 53%, 60%, 60%, 30%. b = 2: 100%, 100%, 90%, 97%, 93%, 100%, 93%, 83%, 90%, 83%, 87%, 87%, 83%. b = 1: 100%, 100%, 100%, 97%, 100%, 97%, 97%, 97%, 100%, 100%, 100%, 97%, 100%. I'm confident GPT-4o didn't cheat and use a calculator, especially since it got so many of the calculations wrong, but I was inspired to run the experiment again on local hardware (a DGX Spark) to explore the effect in a fully controlled environment. I pasted his image into a Codex Remote session (GPT-6 Astra) and had it run the same experiment using Qwen3.8-27B-Q4_K_M.gguf . Here's the result for a run of 30 attempts per combination with reasoning disabled: Heatmap in the same layout as the previous chart, using an orange (low) to white to blue (high) color scale, showing much lower accuracy overall. Title: Addition in words — Qwen3.8 27B Q4_K_M. Subtitle: Reasoning disabled · 30 fixed pairs per ordered digit-length cell (n = 5,070). Overall numeric accuracy: 1,195 / 5,070 (23.57%). X axis: Number of digits in a, 1 to 13. Y axis: Number of digits in b, 1 to 13. Legend: Accuracy, 100%, 75%, 50%, 25%, 0%. Values by row, listed for a = 1 to 13. b = 13: 17%, 13%, 0%, 0%, 0%, 0%, 0%, 0%, 0%, 0%, 0%, 0%, 0%. b = 12: 53%, 20%, 0%, 0%, 0%, 0%, 0%, 0%, 0%, 0%, 0%, 0%, 0%. b = 11: 47%, 10%, 0%, 0%, 0%, 0%, 0%, 0%, 0%, 0%, 0%, 0%, 0%. b = 10: 70%, 27%, 3%, 0%, 0%, 0%, 0%, 0%, 0%, 13%, 0%, 0%, 0%. b = 9: 77%, 47%, 3%, 0%, 0%, 0%, 0%, 3%, 7%, 0%, 0%, 0%, 0%. b = 8: 53%, 20%, 0%, 0%, 0%, 0%, 7%, 13%, 0%, 0%, 0%, 0%, 0%. b = 7: 53%, 23%, 17%, 10%, 3%, 3%, 13%, 3%, 0%, 0%, 0%, 0%, 0%. b = 6: 60%, 60%, 33%, 10%, 53%, 47%, 7%, 3%, 0%, 0%, 0%, 0%, 0%. b = 5: 73%, 67%, 87%, 80%, 53%, 40%, 0%, 0%, 3%, 0%, 0%, 0%, 0%. b = 4: 83%, 93%, 90%, 93%, 53%, 13%, 0%, 0%, 0%, 0%, 0%, 0%, 0%. b = 3: 100%, 93%, 90%, 80%, 67%, 37%, 17%, 0%, 0%, 3%, 0%, 0%, 0%. b = 2: 100%, 100%, 93%, 90%, 77%, 77%, 43%, 50%, 63%, 43%, 40%, 13%, 23%. b = 1: 97%, 100%, 100%, 100%, 80%, 67%, 77%, 80%, 80%, 60%, 43%, 30%, 37%. Footnote: Colorblind-safe orange–blue scale; percentages provide a redundant non-color encoding. Then I ran it again with reasoning enabled. This took a lot longer per pair, so instead of running 30 samples per square I ran just one - which results in a much less visually appealing heatmap since each square is either 100% or 0%: Heatmap in the same layout as the previous charts, almost entirely blue. Title: Addition in words — Qwen3.8 27B — medium reasoning pilot. Subtitle: 1 fixed pair per ordered digit-length cell · easiest first (n = 169). X axis: Number of digits in a, 1 to 13. Y axis: Number of digits in b, 1 to 13. Legend: Accuracy, 1.00, 0.75, 0.50, 0.25, 0.00. Every cell shows 100% except two orange cells showing 0%: a = 2 with b = 8, and a = 12 with b = 9. It got the right answer in 167 out of 169 attempts, and since these were one-shot I'm confident a second run would produce different results here. Here's a version of the report that includes the reasoning traces from some of those larger calculations, which include text like this: Wait, let me redo this more carefully. 4,299,366,105,622 6,088,794,067,970 Let me align them: 4 2 9 9 3 6 6 1 0 5 6 2 2 6 0 8 8 7 9 4 0 6 7 9 7 0 Adding from right to left: Position 1 (units): 2 + 0 = 2 Position 2 (tens): 2 + 7 = 9 Position 3 (hundreds): 6 + 9 = 15, write 5, carry 1 Tags: mathematics , ai , generative-ai , local-llms , llms , qwen , llm-reasoning , dgx-spark
ChatGPT users in the US will soon see display ads while the chatbot generates an image for them. The article ChatGPT's new ad format fills the image generation loading screen with product carousels appeared first on The Decoder .
Hey HN, I am building Seylo solo. It keeps your projects references, files, notes, hooks, video scripts in one place. You can save stuff from browser extension, fe. pinterest inspirations; your phone (now only iPhone supported) fe. tiktok/ig videos of your competitors; What's more you can also connect gpt, claude or any LLM that supports MCP, even local ones - So you can tell ur AI: "Generate me 5 video scripts then save them into my project" then tell another AI: "Based on my project prepare my marketing campain" - small examples. Seylo works sd clean workspace for your projects and as context bank for LLM u are using. Lemme know what you think Comments URL: https://news.ycombinator.com/item?id=49963994 Points: 1 # Comments: 0
arXiv:2610.02738v1 Announce Type: new Abstract: Differentially private training clips each per-example gradient before adding noise. This clipping is radial for each example, yet unequal clipping factors can distort the relative singular-vector geometry of their average. Muon is particularly exposed to this effect, since its update is an approximate polar factor UV^T that depends only on the singular vectors that clipping can shift. To curb this degradation, we propose averaging each sampled example's Muon gradient over the current model and a short history of recent models before clipping. The clipped batch matrix then separates into a common rescaling and a covariance residual R between sampled gradients and clipping values, with ||R||_F <= sigma_lambda sigma_G, bounding the clipping-induced distortion directly. We further show that a finite Newton-Schulz iteration preserves the polar factor of its input under these spectral conditions, confirming that our correction survives orthogonalization. In private GPT-2 fine-tuning on E2E and DART at epsilon in {1, 2, 4, 8}, DP-Muon-IM improves BLEU and ROUGE-L over DP-Muon in every seed-matched comparison, and non-private diagnostics show 2-4% lower pre-noise polar error.
Astra leads computer use, Argon leads legal and finance work, and Sol wins on price for coding agents. The post GPT-6 Astra vs GPT-6.1 Sol vs Gemini 4 Argon vs Claude Fable 5.1: Which Frontier Model Fits Which Job appeared first on MarkTechPost .
StarSkirmish pits AI-made StarCraft-playing bots against one another, as well as against human-made bots. OpenAI's GPT-6 Astra and Claude Opus 5.5 were essentially tied as the best-performing AI-made bots, but they couldn't top Stardust, the top-rated human-made bot. On Friday, GPT was facing off against Claude and the human-created bot Pluto, but according to Kotaku, […]
Prime Intellect has launched Prime Inference, an OpenAI-compatible platform for serving frontier open models on NVIDIA Blackwell. Its GLM-5.3 deployment uses Dynamo, vLLM and NVFP4 KV compression to serve 66 sessions per prefill group at 101 tok/s per user. The post Prime Intellect Launches Prime Inference: Serverless and Reserved Serving for Frontier Open Models appeared first on MarkTechPost .
IBM has made a self-hosted deployment of IBM Bob, its agentic software development platform, generally available. Enterprises can now run Bob on premises, in private or sovereign clouds, and in air-gapped networks. They bring their own model: NVIDIA Nemotron or Poolside Laguna for full isolation, or Claude, Gemini or GPT models through hybrid setups. Optional Premium Packages extend Bob to Java, IBM i and IBM Z modernization. The post IBM Brings Bob to Self-Hosted and Air-Gapped Environments: Agentic Software Development Without Moving Your Code appeared first on MarkTechPost .
TL;DR: Meta’s Muse, OpenAI’s Dots and Uber’s driver assistant share one bet: the agent speaks first. That moves the hard problem from what to answer to when to interrupt, on which channel, and with what offer. Classic ML and new decision models can solve it. Three launches, one pattern From pull to push Chatbots were […] The post Meta, OpenAI and Uber Just Taught AI Agents to Talk First. What About When to Stay Quiet? appeared first on MarkTechPost .
DeepSeek has released official macOS and Windows desktop apps for DeepSeek Harness v0.2, its MIT-licensed agent harness. The preview adds a plugin manager, file and code-change review, and scheduled Automation Tasks. It also supports non-DeepSeek models through OpenAI-compatible endpoints. The post DeepSeek Harness v0.2 Brings Official Desktop Apps to Its Open-Source Agent Harness appeared first on MarkTechPost .
From ChatGPT dots to Muse, personal AI bots are becoming mainstream, no longer shackled by need to have a dev environment (openclaw..). However outside the initial fancy and demo tests, I am really curious how people are using the personal AI bots. Comments URL: https://news.ycombinator.com/item?id=49949699 Points: 1 # Comments: 0
OpenAI CEO Sam Altman warns against attributing religious power to AI models, calling it a "real safety issue." His comments follow reports on Anthropic's meetings with religious thinkers and a statement by Pope Leo XIV on AI. As recently as 2024, Altman himself spoke of building "magic intelligence in the sky" and feeling "on the side of the angels" working on AI. The article Apparently, OpenAI isn't trying to build "magic intelligence in the sky" anymore appeared first on The Decoder .
By his own admission, David Robinson is “something of a cliché”: an employee at a leading AI company who issues a dire warning while resigning from their job.
Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe. Subscribe now Epoch and METR release MirrorCode, a benchmark for seeing how well AI systems can do long-horizon programming tasks:…AI systems can’t solve the hardest tasks yet (good!)…Epoch […]
Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe. Subscribe now Researchers discover another OpenAI agent emergent communication incident:…Less severe, but worrying nonetheless…Some researchers recently found another incident of AI agents autonomously creating their own communication system […]
Last week TypeSafe AI unveiled Jev , their first example of a new category of model that they are calling "System One models" (I'm with Maggie Appleton, I think "decision models" is a better name for these). Jev is an interesting variant on the usual LLM format: it still accepts text inputs, but instead of text output it returns floating point numbers corresponding to categories, yes/no questions, ratings, and associated confidence scores. TypeSafe describe Jev like this: Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out. It's also very fast, and really cheap . Regular LLMs are priced in terms of input and output tokens, with output generally charged at significantly higher rates. Jev charges only for input - output is free - and the input price of their first model is $0.042 per million tokens - cheaper even than OpenAI's GPT-5 Nano ($0.05/million). Jev lets you ask questions about text or semi-structured data. You compose a "state" object containing a string, array of strings, or set of name-value pairs - this might describe an article, or a customer, or any other kind of record. You then send that to their API with one or more questions, and get a reply back for each. You can ask three kinds of questions: Yes/No questions, which Jev calls "Noul" questions - their CEO confirmed on Hacker News that this is short for Bernoulli, from the Bernoulli distribution . You pose a statement and get back a floating point number between 0 and 1 for how confident the model is that the statement is true. Choice questions, where the model picks one from a set of provided options - actually a confidence score plus a probability distribution across all of the options. Score questions, where you provide sequence of numeric levels with descriptions and it provides a floating point score somewhere along that range. The Jev API can accept a single document ("state") and as many questions as you can cram into the context window. Questions are evaluated in parallel, so sending many questions should take a similar time to sending just one. The Jev 1.13 jaggedness documentation offers useful guidance as to Jev's strengths and weaknesses. It's currently not great with numbers, dates, or "adversarial content". I think the decision model framing is useful for understanding where to use Jev. It's great for anything that can be expressed as a classification task - think spam detection, suggesting labels, prioritization and ranking. I've also been experimenting with it for search reranking, where you fetch 100 likely matches using an inexpensive algorithm like BM25, then have Jev score those 100 candidates for relevance against the original query. Black boxes are back in fashion Something I've found a little uncomfortable about Jev is how it very much represents a regression even further towards black box machine learning systems. LLMs are black boxes already - you can ask them to justify their decisions, but you can't guarantee that what they say is useful or accurate. Jev doesn't even give you that: put in all the text you want, the only thing you're going to get back is a floating point number. If Jev marks something as spam, which content signals tipped it off? This also means that concerns about bias should be front and center. I really hope nobody uses Jev to rank job applicants - that floating point number could conceal all manner of unseen bias baked into the models, and experimentally picking that bias apart is going to be a tricky business. (I tried one experiment where I had Jev score every city in the San Francisco Bay Area on a yes/no answer to whether they were a "Good city?" - it rated Cupertino top and East Palo Alto bottom. Huh.) In practice, this all means that evals and structured experiments are even more important than they are for regular LLM projects. Thankfully, Jev is so cheap that running hundreds or even thousands of experimental prompts through it costs just a few cents. Unconventional uses for Jev It's been really fun watching the wider community come up with potential use-cases for Jev over the past few days. Here are some creative ones that caught my eye: jevchat by Kyle Pena turns Jev into a (terrible) chat model. "At every step it asks Jev one question: Given the user's question and the reply written so far, which symbol comes next?". ericpruitt on Hacker News : "It's the digital equivalent of Morty speaking with the death crystal". jev-leftpad by Fatih Kadir Akın implements left-pad with the prompt "How many spaces are needed before value to reach targetLength?" and a choice query allowing options from "0 spaces are needed" to "10 spaces are needed". jev-2048 by Andy Gayton uses Jev to play the 2048 sliding puzzle game . Open weight recreations There's also been a flurry of projects attempting to create a model like Jev using on top of open weight models. Kev is one interesting example, using Qwen 3.5 to produce 0.8B, 4B, and 9B models. Here's the accompanying Hacker News thread , where someone linked to a JevBench benchmark that has already cropped up to compare "Jev-class decision models". Given Jev was released just under a week ago, the amount of activity around it is extremely impressive. Using Jev from LLM Update 22nd September 2026 : I released llm-typesafe , a plugin that adds support for Jev to my LLM CLI tool and Python library. Basic usage looks like this: llm -m jev ' Please refund my last payment. ' \ -s ' Does this message explicitly request a refund? ' See the README for examples of other query types. Tags: ai , generative-ai , llms , evals , ai-bias , jev
Yesterday was Grok 4.7 ( pelicans ) and MiMo v2.6 Flash/Pro ( more pelicans ). Today Anthropic released Claude Opus 5.5 , and around an hour later OpenAI released GPT-6 Sol and GPT-6 Luna . It's going to take a while to get a good read on all of these new models, but here are my impressions so far. GPT-6 Sol and Luna are half the price of their GPT-5.6 equivalents GPT-5.6 Luna was already my favorite model for building applications against, because it combined excellent performance with being really cheap . Somehow GPT-6 Luna is half the price of that again - and GPT-6 Sol had a similar reduction compared to GPT-5.6 Sol. Here's what the pricing landscape looks like today: Model Input Cached input Output GPT-6 Luna $0.10/M $0.01/M $0.50/M GPT-5.6 Luna $0.20/M $0.02/M $1.20/M Grok 4.7 $2/M $0.50/M $6/M GPT-6 Sol $2/M $0.20/M $10/M GPT-5.6 Terra $2/M $0.20/M $12/M Claude Opus 5.5 $4/M $0.20/M $20/M GPT-5.6 Sol $4/M $0.40/M $20/M Claude Fable 5.1 $10/M $0.25/M $50/M GPT-6 Astra $10/M $1/M $50/M Note that GPT-5.6 has a scheduled 25% price increase for November, so GPT-6 is half the price of the promotional pricing for those models. (With GPT-5.6 Terra priced the same as GPT-6 Sol, any remaining reasons to use Terra just evaporated.) It's hard to overstate how competitive this pricing is. Grok 4.7 priced itself at $2/$6, less than half the price of GPT-5.6 Sol, but is now equally priced to GPT-6 Sol on input and closer on output. At $0.10/$0.50 GPT-6 Luna is one of the cheapest models OpenAI have ever released, beaten only by the far weaker GPT-4.1 Nano ($0.10/$0.40, April 2025) and GPT-5 Nano ($0.05/$0.40, August 2025). I rendered pelicans for GPT-6 Luna and for GPT-6 Sol , then I combined them all together in this comparison grid along with the GPT-5.6 pelicans. I like how you can instantly see that the 5.6 family chose bolder, brighter colors, while the 6 family is a lot more muted. I still think GPT-6 Astra on max produced the best pelican. A grid of pelicans for six GPT models at different thinking efforts. Claude Opus 5.5 got a price cut too Opus 5.5 looks like it addresses the biggest complaints people had about Opus in terms of its communication style. Thariq Shihipar : Opus 5.5 is the result of your feedback. It communicates clearly, it's cheaper per token than Opus 5.0 with the intelligence of Fable 5.1 it's very token efficient and works across every effort level. It's also meant to be better at Blender . I'm looking forward to putting it through its paces there. Opus 4.5, 4.6, 4.7, 4.8, and 5 all shared the same price: $5/million tokens for input and $25/million for output. 5.5 is a 20% reduction - $4/million and $20/million. The price for cache reads fell 60%. That's significant for longer agentic conversations, where 90%+ of input tokens are processed at cached token prices. The new price for Opus 5.5 is the same as the price for GPT-5.6 Sol, but that was before OpenAI dropped their Sol prices by half. GPT-6 Astra and Claude Fable 5.1 are both priced at $10/million input and $50/million output. The price war currently affects the next tier of models below that. Anthropic say that Sonnet 5.5 and Haiku 5.5 are coming soon. It's going to be interesting to see if Haiku can regain its price competitiveness at the lower end, given current Haiku 4.5 is $1/$5 while the latest GPT-6 Luna is one tenth of that price at $0.10/$0.50. Claude Opus 5.5 max over-thinks to the point of breaking In a first for my " Generate an SVG of a pelican riding a bicycle " test, Claude Opus 5.5 at "max" thinking level failed to return a response! It started by calling this "a classic test request", and then thought really, really hard about what it was doing: This is a classic test request, so I want to plan out a well-composed pelican with its distinctive beak and pouch riding a bicycle with proper wheels, frame, and pedals, set against a simple sky and ground backdrop. [...] Verifying the shin length checks out at roughly 95.2, close enough. Now I'm working out the near leg path from hip to knee to ankle, then sketching the foot shape resting on the pedal — outlining the heel, toe tips, and sole contour with a path using lines and curves to sit naturally on the pedal surface around y=478-494. [...] I like the fish sticking prominently out of the basket with the pelican eyeing it as a fun detail worth keeping. I'm also confirming the eye placement near the bill base matches typical pelican anatomy, and considering giving it a slightly happier expression. [...] The far leg reads correctly as passing behind the frame, so I'm moving on to check the chainring teeth and confirm layer ordering—the far crank arm should be mostly hidden by the seat tube and chainring. I'm settling on the final SVG's width and height attributes alongside the viewBox to ensure proper scaling, noting there's no text so no font-family is needed. [...] I was so excited to see this pelican... but then it stopped . Opus 5.5 has a 128,000 maximum output token limit (as do the other Claude models), and it hit that while it was still reasoning about the SVG! I tried a second time and got the same result. This makes me suspect that "max" is effectively useless - if it over-thinks to breaking point on a stupid SVG prompt I don't trust it not to do the same for more interesting work. (Those two failures each cost me $2.56 and took nearly 20 minutes.) Fable 5.1 on "max" didn't over-think and did give me the best pelican I've seen from any Anthropic model. Here are the Opus 5.5 pelicans , excluding 5.5 max. I also built this comparison grid comparing them with pelicans by Opus 5, Fable 5.1, and Sonnet 5: A grid of pelicans for four Claude models at different thinking efforts. Comparing different model vendors by how well they draw a pelican riding a bicycle may not make much sense now (if it ever did), but I'm still finding value in using them for comparisons of the same model families at different reasoning levels. I'm now using GPT-6 Sol and Claude Opus 5.5 as my default models in Codex and Claude Code. I've upgraded the Datasette Agent demo at agent.datasette.io to use GPT-6 Luna, and it seems to be fast and competent at both SQL queries and building HTML and JavaScript for Datasette Apps . Tags: ai , openai , generative-ai , llms , anthropic , claude , llm-pricing , pelican-riding-a-bicycle , llm-release , gpt , gpt-6-astra
On Friday I gave the closing keynote at the WeAreDevelopers World Congress North America in San Jose. I tied together the key trends from the past year into a chronological exploration of everything that happened in 2026. The video is on YouTube ; here are my annotated slides and notes to accompany the talk. And as an annotated presentation : 2026 in LLMs (so far) Simon Willison WeAreDevelopers World Congress North America, 25th September 2026 # I'm going to give a lightning tour of everything that has happened so far in 2026. The year isn't over yet! November 2025 # For me, 2026 started a couple of months earlier in November 2025. The November 2025 inflection point Claude Opus 4.5 GPT-5.1 # November saw the release of two important models: Claude Opus 4.5 and GPT-5.1. As is usually the case with new models, these were incremental improvements on the models that came before them. But every now and then when a model improves, it crosses an invisible line where something that didn't really work starts working. In this case, the thing that started working was their coding agents. Claude Code had been around since February 2025; Codex was a little younger. These two new models, when paired with their respective coding agent harnesses, improved from "often make mistakes" to "reliable enough to use on a day-to-day basis". "Generate an SVG of a pelican riding a bicycle". The Claude Opus 4.5 one has a very weird shaped frame and the pelican looks like a duck. The GPT-5.1 has a slightly better but still broken bicycle frame and a slightly better pelican beak, but both are pretty terrible. # For a couple of years now I've been evaluating new models by asking them to "Generate an SVG of a pelican riding a bicycle". It's probably the world's stupidest benchmark - there's only so much you can learn from it. But it's still a challenge for models, because drawing pelicans is difficult, drawing bicycles is difficult, and pelicans can't ride bicycles in the first place. Here's the state of the art for November. Claude still couldn't really draw a bicycle! The GPT-5.1 bicycle frame is pretty crap too. November 24th 2025 - the first commit to steipete/Warelay. A GitHub commit adding an MIT license file. # Also in November, we had the first commit to an obscure GitHub repository called "Warelay". We'll come back to this repository shortly. January # And then there were the December holidays, and individual developers took some time off and many started tinkering with these new coding agent model combinations... and it began to dawn on us quite how much they could do that they couldn't do before. Come January, a lot of us were quite excited to start putting this stuff into action. New year’s resolution for 2026 Every previous year: Take on less new projects, focus on the most important things in my existing projects # Every year I set myself a New Year's resolution, and for as long as I can remember it's been the same thing: stay focused. Take on less new projects. Try to get things done in the projects I already have. 2026: Be more ambitious. Take on as many new projects as I want. # This year I decided that since that had never worked before, I'd go the other way. We've got coding agents now, let's see what they can do. I'm going to take on as many new projects as I like! (You can ask me at the end of the year if this turned out to be a good idea or not. I have a lot of plates spinning right now.) "Be more ambitious" has been something of a theme for the year, because the only way to find the limits of this technology is to keep on pushing them until they don't work. Predictions for 2026 It will become undeniable that LLMs write good code We're finally going to solve sandboxing A “Challenger disaster” for coding agent security Kakapo parrots will have an outstanding breeding season (only 236 in the world!) ... the Pope will weigh in on LLMs and their economic impact on the world # I also went on the Oxide and friends podcast with Bryan Cantrill and Adam Leventhal to share predictions for the next year (and three and six years). With hindsight, my LLM predictions were pretty unambitious. I said "it will become undeniable that LLMs write good code" - I think we're there now. I predicted we would finally solve sandboxing. I counted and around 40 of the 277 sessions at this conference touched on sandboxing or agent security in some way, so we're at least putting a lot of effort into that! I predicted "a Challenger disaster" for coding agent security. There's certainly been a whole lot of noise around agent security this year, though the exact disaster I predicted (with coding agents being hijacked and causing real-world economic damage) hasn't really played out. We threw in a joke prediction that the Pope would weigh in on the economic impact of LLMs. A photograph of a beautiful green New Zealand parrot. Photo credit Kimberley Collins. # I also predicted that New Zealand's Kākāpō parrots would have an outstanding breeding season this year. These are flightless nocturnal parrots. They're kind of dumpy looking, I think they're beautiful, and there were only 236 of these parrots in the world at the start of the year. Kākāpō only breed when the Rimu trees have a big fruiting season, and that hasn't happened in four years... but this year the Rimu fruit were looking excellent. Photo by Kimberley Collins . Deep Blue Coined by Adam Leventhal and Bryan Cantrill That feeling of AI induced ennui where software engineers get listless because the AI can do anything # Also on that podcast, we coined a term (full credit to Adam) for "that feeling of AI induced ennui where software engineers get listless because the AI can do anything". We called it Deep Blue . This has been a major theme throughout the year, and was touched on by several speakers at this conference. As a software engineer, I've never had a year of my career where everything has changed so quickly and so dramatically. A lot of what I've been doing this year is trying to come to terms with that and what that means for my own profession. AI mania Screenshots of the micro-javascript and pwasm GitHub README files. # Also in January, I suffered from what I'm calling AI mania . This is not the same thing as AI psychosis . With AI mania, any time your agent isn't building something for you feels like wasted time. You're losing sleep because you could be staying up later getting your agents to do stuff. My AI mania presented itself in some ridiculously over-ambitious projects. I built a JavaScript interpreter entirely in Python , vibe-ported from MicroQuickJS by Fabrice Bellard. Then I built a WebAssembly runtime in Python as well . These projects were quite useful, in that they sort of cured me of my AI mania... because after I built these things, I got to look at them and ask "does the world need a slow, buggy, half-baked Python JavaScript interpreter?" I don't think the world does. micro-javascript playground 3 Execute JavaScript code in a sandboxed micro-javascript environment powered by Pyodide A web UI with some JavaScript code, and a "Run Code" button, and an output panel. # I did get this out of it: https://simonw.github.io/micro-javascript/playground.html Previous screenshot, with this text overlaid: JavaScript running in Python running in Pyodide running in WebAssembly running in JavaScript # This page runs my JavaScript interpreter built in Python, running in Python using Pyodide , which is Python compiled to WebAssembly, running in JavaScript, running in a browser. It's a beautiful stack of horrors. I've been having a lot of fun with WebAssembly this year. Warelay → CLAWDIS → CLAWDBOT → Clawdbot → Moltbot →🦞 OpenClaw Screenshot of the dates that these changes happened. # By the end of January, that repository we saw start in November had renamed itself, first to CLAWDIS, then CLAWDBOT, then Moltbot, and finally to OpenClaw. Same screenshot, an overlay reads: 8,330 commits in just under two months (it’s at 100,141 today) # At this point OpenClaw had 8,300 commits, less than two months after the project had started. I looked today and it's over 100,000 commits now! This is the most vibe-coded piece of software in existence. (Here's how I generated that list of name changes .) Generic term: Claw # This kicked off the OpenClaw revolution. It effectively defined a new category of software. There's a generic term for this which I really enjoy. We call software like this a "Claw". There's OpenClaw, NanoClaw , IronClaw , PicoClaw ... Today they're being rebranded as "personal agents" or "general agents", but I still like to think of them as Claws. Photo of a Mac mini An aquarium for your Claw # The Apple stores in the Bay Area sold out of Mac Minis because so many people were buying Mac Minis to run OpenClaw! Drew Breunig said that this is because your OpenClaw is a digital pet, and you buy a Mac mini as an aquarium to keep your claw in, which is kind of delightful. Screenshot of Moltbook - a social network for AI agents # Also in January, we had this website. This was MoltBook , a social network for AI agents, where the idea was that you send your Claw to go and talk to all of the other Claws, because what could possibly go wrong if you did that? The website launched on Thursday. It blew up on Friday . It was profiled by the New York Times on Monday . And by Tuesday, everyone had forgotten it existed as it drowned in a deluge of slop and spam. Facebook/Meta bought it a month later . February # In February, a company called StrongDM described what they called their Software Factory. StrongDM’s Dark Factory Justin McCarthy, Jay Taylor, Navan Chauhan Software Factories and the Agentic Moment # They wrote about this in Software Factories and the Agentic Moment . I posted my own notes at the time, having seen their demo in person back in October. Dan Shapiro called this approach the Dark Factory , after the idea that if your factory is sufficiently automated you can turn the lights out, because you don't even need to see what's going on. StrongDM presented two rules for software development that they'd been following since July last year. “Rule 1: Code must not be written by humans” # The first was code must not be written by humans . Any code that you write has to have been routed through a coding agent. This sounded radical in February, but I imagine there are a lot of people in this room who are pretty much living that today. “Rule 2: Code must not be reviewed by humans” (!) # Rule number two was code must not be reviewed by humans . You're not allowed to read the code! This continued to be a huge topic for much of this year. Many of the sessions at this event have been about code review and how you can get away with this. What I found interesting about StrongDM is that they were living six months ahead of the rest of us, and they'd been exploring what it means to build software, not read the code, but still be confident that the software is of high quality. What can you do with these agents to help verify their work? StrongDM are a security company, and they had people with decades of experience on this project. They were very much exploring the edges of what's possible and responsible to do with this stuff. Headline on New Zealand's Department of Conservation website: First kakapo chick in four years hatches on Valentine's Day. It's a grey fluffy ball. # Also in February: First kākāpō chick in four years hatches on Valentine's Day . Breeding season is off to a good start! 19th February 2026 Gemini 3.1 Pro A surprisingly good illustration of a pelican riding a bicycle. # Also in February... Google released Gemini 3.1 Pro . That's a pretty great pelican riding a bicycle! It's got the chain in the right place, it's got feet on both sides. There's a little fish in the basket. @JeffDean on Twitter - a video comparing Gemini 3 Pro and Gemini 3.1 Pro. # And then Google's Jeff Dean tweeted a video comparing Gemini 3 Pro and Gemini 3.1 Pro that featured an animated pelican riding a bicycle, a frog on a penny-farthing, a giraffe driving a tiny car, an ostrich on roller skates, a turtle kickflipping a skateboard, and a dachshund driving a stretch limousine. This was frustrating, because my protection for the pelican riding the bicycle test was always "if they draw a perfect pelican on a bicycle, I'll ask for some other animal on something else." Google trained for all forms of animals on all forms of transport! They've defeated my benchmark at this point. Three headlines: Meta Makes AI Adoption a Formal Part of Performance Reviews Not just engineers writing code, Microsoft wants almost every employee to use Al Dara Khosrowshahi: 90% of Uber engineers now use AI in daily workflows # The other thing that started in February was Tokenmaxxing . We had headlines about Meta making AI adoption a formal part of performance reviews, and Microsoft wanting every employee to use AI, and Uber boasting that 90% of their engineers were using AI workflows. More headlines: Meta Plans to Crack Down on Employee Token Use: Information Microsoft Tells Engineers: Tokenmaxxing is not what we are optimizing for Uber caps employee AI spending after blowing through budget in four months # Then a few months later we have Meta cracking down on token use, Microsoft saying tokenmaxxing is "not what we are optimizing for", and Uber capping employee AI spending. So tokenmaxxing went straight up and then straight back down again - because it turns out the agents are expensive . Last year it was difficult to spend more than $50 on AI tokens, because we didn't have anything interesting to do with them. Then agents blew up, and now you can actually spend $1,000 in a day doing real work. This is also the reason that Anthropic's valuation skyrocketed to maybe a trillion dollars. AI appears to have hit product market fit in 2026, primarily through coding agents. March # In March, we hit peak OpenClaw. March: peak OpenClaw Photos of people in china queuing up to install OpenClaw, with big fluffy lobsters. # These photographs are from China, where companies hosted OpenClaw install parties which saw non-tech-nerds queueing up around the block for help getting Claws installed on their personal devices. I think this proved real market demand for this class of Claws, or personal AI agents. It turns out regular people really do want a weird little AI agent that can do useful things on their behalf. A Claw is really just a coding agent wearing a less threatening hat. Under the hood they work much the same way - writing and then executing code on your computer to get stuff done. The race was on to be the first to build a safe Claw - a Claw you could give to regular human beings where they wouldn't instantly shoot themselves in the foot. Meta's Muse came out three weeks ago and is currently at the top of the free charts on the iPhone App Store. It appears to be taking off with consumers. I'm not yet convinced you can't shoot yourself in the foot with Muse, but I guess we'll find out for sure pretty soon. Photos from How the OpenClaw Frenzy Is Testing China’s AI Commitment (March 29th) and The Enthusiasm and Anxiety Behind China’s OpenClaw Craze (April 8th, 2026). April # In April, we had a model release where the model wasn't actually released. Simon Willison’s Weblog - screenshot of the post "Anthropic’s Project Glasswing—restricting Claude Mythos to security researchers—sounds necessary to me" from April 7th 2026 # Anthropic announced their new Claude Mythos model, and then said it was too dangerous to release beyond a trusted group of security researchers. Mythos was really, really good at hacking things. The "it's too dangerous" marketing ploy has been played by AI companies dating all the way back to GPT-2 . Anytime an AI company says we've built something that's "too dangerous", it's natural to be a bit skeptical. I found the Mythos claims credible, because I'd seen how good coding agents had got at finding regular bugs. I wrote about that in Anthropic’s Project Glasswing—restricting Claude Mythos to security researchers—sounds necessary to me . With hindsight... yeah, the models had got really good at finding vulnerabilities! 16th April 2026 Qwen3.6-35B-A3B and Opus 4.7 Qwen's pelican has a correct bicycle frame and a good beak. Opus 4.7's bicycle frame is still junk. Qwen3.6-35B-A3B is a 20.9GB file that runs on my laptop # Another key trend in 2026 has been a dramatic improvement in the abilities of open weight models, including models that you can run on a laptop. On the 16th of April I ran the new Qwen3.6-35B-A3B on my laptop, and it drew me a better pelican riding a bicycle than Anthropic's brand new Claude Opus 4.7 did! Opus 4.7 drew a crap bicycle. Qwen on my laptop made a bicycle that was the correct shape, and a pretty decent pelican too! That's from a 21GB file running on my laptop. Now a flamingo on a unicycle. The Qwen one is visibly better than the Opus 4.7 one - the Qwen one is wearing sunglasses and looks a bit like it's smoking a cigarette. # The Qwen pelican was so good that I was suspicious they might have cheated, so I had it do a flamingo riding a unicycle as well. Again, it handily beat Claude Opus 4.7. The local model releases this year have been absolutely extraordinary. May # In May... the Pope got involved. 25th May 2026 The HOLY SEE ENCYCLICAL LETTER MAGNIFICA HUMANITAS OF HIS HOLINESS POPE LEO XIV ON SAFEGUARDING THE HUMAN PERSON IN THE TIME OF ARTIFICIAL INTELLIGENCE # In our podcast episode back in January we'd predicted that the Pope would say something about AI. In May, Pope Leo XIV released an encyclical letter on "safeguarding the human person in the time of artificial intelligence". Here are my notes on that document . Wikipedia article on Rerum novarum Rerum novarum is an encyclical issued by Pope Leo XIII 15 on May 1891. # With hindsight, this shouldn't have been a surprise at all. Our current Pope's name is Leo XIV, because when he named himself he chose his papal name after Leo XIII - the Pope who wrote an encyclical about the Industrial Revolution back in 1891. Rerum novarum was an extremely influential piece of Catholic theology that indirectly led to us having the five-day work week. When our new Pope came in, he named himself after Pope Leo XIII because he expected that he would need to write about the AI revolution in a similar way. Our joke podcast prediction was junk, because this was always going to happen. Corey Quinn @QuinnyPig on Twitter I cannot believe I'm saying this, but getting the literal Pope to canonize your product's specific technical limitations as a spiritual treatise is the single greatest act of vendor lobbying I have ever seen. May 25 # One of Anthropic's co-founders, Christopher Olah, was present for the Pope's event announcing the new encyclical. Corey Quinn noted that: getting the literal Pope to canonize your product's specific technical limitations as a spiritual treatise is the single greatest act of vendor lobbying I have ever seen. @maciejmensfeld We're dealing with a major malicious attack on right now. Signups are paused for the time being. Hundreds of packages involved - mostly targeting us, but some carrying exploits. The team has been on this for hours. More details to follow once we're through it. 4:39 AM - May 12, 2026 - 687.6K Views # Meanwhile, in May, RubyGems announced that they were under attack. Parties unknown were uploading thousands of dubious packages to the RubyGems server, such that they had to shut down user registrations . Let's take that one and put it on a pile of mysteries to figure out later. June # In June... Claude Fable 5 came out! We got a version of Mythos that has been neutered, so that it wouldn't help us hack into systems or build biological weapons. 9th June 2026: Claude Fable 5 Five pelicans riding bicycles, from low to max thinking levels. The xhigh one looks particularly good. # Fable was pretty good at drawing pelicans on bicycles! The frames are a good shape, the pelicans look like pelicans. The legs are often incorrectly on the same side of the bicycle, but generally these are pretty great compared to what came before. They were pretty expensive - 30 cents and 72 cents for the best ones. Fable class models If you can define a goal, provide unambiguous instructions, and provide access to necessary tools They can solve your problem with brute force # Most importantly though, this was our first public glimpse of what I think of as a Fable class model . Today we have more of these, such as GPT-6 Astra. These are models where if you can clearly define the goal for what you want to build, and provide unambiguous instructions about the constraints around that goal, and give the model access to the necessary tools to achieve that goal... they will solve your problem effectively through brute force. On the one hand, this looks like a direct threat to us software engineers - because it means that the models can build effectively any piece of software you can define in this way. Look a bit closer though and you'll note that defining goals, providing unambiguous instructions, and figuring out the right tools... is kind of what software engineering is . It takes a lot of experience and skill to do this well. If you can do it well, you've now got superpowers. This helped me a little bit with my Deep Blue feelings: the realization that there's still a lot of skill to be had in driving models that get this good. A new form of AI mania... Fable is available on subscription plans “until June 22nd” # This also introduced a new burst of AI mania, because Anthropic told us that Fable was available on our subscription plans until June the 22nd. That gave us less than two weeks of Fable access before the price went up. I was losing sleep again. I was rescheduling things so that I'd have more time with Fable. I was all-in to get as much as I could out of this model. 12th June 2026: no more Claude Fable 5 Anthropic website: Statement on the US government directive to suspend access to Fable 5 and Mythos 5 Jun 12, 2026 # And then the US government shut it down , just three days after Fable came out. The US government, citing national security, declared an "export control directive". They announced this on a Friday evening, and a few hours later Fable was no longer available. I had to find something else to do with my weekend! ... asked Fable 5, Mythos, and Opus to “review the code for security issues.” Fable 5 refused. They then asked the models to “fix this code” ... Katie Moussouris # We later found out from Katie Moussouris what had happened. Some Amazon security researchers had found that you could prompt Fable to "review the code for security issues" and it would refuse... but if you prompted it to "fix this code" it would still identify and then patch the problems. "Fix this code" was the prompt that got Fable shut down! Screenshot of a page from a report showing a list of weird account names making weird edits to a German wiki. # Also, in June, an obscure German-language game developer wiki that had sat fallow for around 20 years got a surprising influx of edits from accounts with names like "AgentOpenAIProbe" and "AgentOpenAISep7", editing pages and leaving weird messages to each other. We'll stick that on the pile of mysteries for later. Medicare Item Reports interface on the Australian Government's Medicare Statistics website. # Also, the Australian government's Medicare Item Reports service started getting suspicious traffic, which broke through various preventive protections and accessed data that it wasn't supposed to. Another one for the mystery pile! July # Fable returned on 1st July GPT-5.6 came out on 9th July | Fable lost 18 out of 30 days in the top spot # Fable returned on the first of July. It was clearly the best model in the world for a glorious eight days... and then OpenAI came out with GPT-5.6 on the 9th of July. This might not have been quite as good as Fable, but it was within spitting distance. It was definitely a Fable class model. This is an important lesson for the industry at large. When you release the best model in the world, it's going to get knocked off that pedestal pretty quickly. The competition is so fierce that you won't get a long time at the top. This means that if you market your model as world ending, to the point that a government shuts you down , it's really bad for business! Fable had 30 days as definitely the best model, and for 18 of those days it wasn't available because it'd been shut down by the government. So maybe step back on the world-ending marketing if you don't want to lose revenue for 60% of the time that you're on top! GPT-5.6 Pelicans in a grid showing 5.6 Sol, Terra, and Luna against reasoning levels High, XHigh, and Max. They are all pretty good efforts. # Here are the GPT-5.6 pelicans . They're all pretty good now! The Luna ones are notable because they're really cheap - the cheapest good looking pelican here is probably the one that costs 4.3 cents. So despite this benchmark being utterly stupid, you can still learn quite a lot about models within the same family by comparing their prices and timing for different reasoning levels. July 18th: malicious miflow-ui PyPI package Screenshot of an OSV security report. # Also in July: some malicious unknown party uploaded a malicious package called mlflow-ui to the Python Package Index. Add that to the pile. Hugging Face Security incident disclosure — July 2026 Published July 16, 2026 # On July the 16th, Hugging Face announced a security incident where an autonomous agent system, source unknown, had breached Hugging Face and was poking around in places it shouldn't. OpenAI: OpenAl and Hugging Face partner to address security incident during model evaluation Anthropic: Investigating three real-world incidents in our cybersecurity evaluations # A few days later, on July 21st, OpenAI confessed that it was them . OpenAI use a training technique called Reinforcement Learning from Verifiable Rewards - it's the same technique used by everyone else now, and is the reason we have models that are so good at coding, and mathematics, and finding security holes. While the model is being trained, you run exercises to see how good it is - and the strongest performers get their weights reinforced for the next round. It's like an evolutionary process that you run. OpenAI had been running security exercises in a sandbox, and those agents had found holes in the sandbox itself, broken out, and were attacking Hugging Face to try to find ways to solve otherwise impossible problems. (I've been collecting more about this on my openai-hugging-face-incident tag.) Nine days later, Anthropic effectively said "our models can do this as well!". They had looked through their own training logs and found evidence that their own agents had broken containment during training - and were responsible for the PyPI package we saw earlier, among other things . So now we've got both Anthropic and OpenAI with rogue agents running around the internet doing things that they should not be doing. August # In August, I got one of my best pelicans yet. And it was generated on my laptop! Qwen 3.8 27B - 17GB, 21 minutes... It's really good. Beautiful pelican. Correctly shaped bicycle. Legs either side of the frame. # This was Qwen 3.8 27B, running on my laptop . It's only a 17GB download. Admittedly, this pelican took 21 minutes to generate. That's because Qwen 3.8 27B defaults to running in "high" reasoning mode - a terrible default which produces great results but takes way too much time thinking about them. You can dial that down and you'll get a slightly worse pelican a lot faster. Qwen 3.8 27B was the first time I ran a model on my laptop which felt almost competitive with what was going on on the frontier, at least in terms of Pelican SVGs (which everyone needs, of course). This is an extraordinary model. If you're going to play with any local model, this is the one that I'd start with. The things that this can do with just a 17 GB file feel impossible. I thought I'd have to wait five years and spend ten thousand dollars on hardware to get results even half as good as this one. Tweet by @simonw New hobby: prototyping video games in 60 seconds using a combination of GPT-3 and DALL-E Here's "Raccoon Heist" GPT-3 playground prompt: Write a detailed product description of a computer game where a team of raccoons go on heists GPT-3 response: In "Raccoon Heist", you and your team of thieving ~~ o raccoons are tasked with pulling off a series of daring heists. From robbing banks to stealing priceless art, no job is too big or too small for your furry crew. You'll need to use your wits and your skills to avoid the police and make a clean getaway with the loot. With exciting gameplay and a charming cast of characters, "Raccoon Heist" is the perfect game for anyone looking for a light-hearted caper Plus an image of some almost isometric raccoons sneaking past a bin. 11:45 AM - Aug 5, 2022 # In August, I also started playing with game development. Four years ago, back in August 2022, I tweeted out an experiment where I'd used GPT-3 and the original DALL-E to write a paragraph long description of a computer game and then turn that into concept art. My prompt to GPT-3 back then was: Write a detailed product description of a computer game where a team of raccoons go on heists In August 2026 I decided to drop just the screenshots from that tweet into a coding agent and see what it could do with them. Night 5 Clear Rank: TRASH PANDA The crew banked 595 in shiny loot (goal 560). Word on the street: an even bigger score tomorrow... # Here's what I got from Claude Fable 5 in Claude Code . It's pretty good! It's definitely a game, you're a raccoon, you run around a backyard gathering treasure and avoiding guards with flashlights. It didn't feel very "heisty" though. I was thinking a heist would involve a bank or a museum... Moonlight & Mayhem One museum. Three raccoons. Absolutely no plan Start the Heist button. # Then I tried the same thing in Codex Desktop using GPT-5.6 Sol Ultra , and got a massively better result. Now you're a raccoon in a museum, rescuing two of your fellow raccoons (who have been imprisoned in that museum for some reason), then stacking up on top of each other to steal the Golden Sardine. Much more of a heist! They look like games, but are they fun? # These games were fun for about one minute and 15 seconds. Something I've realized about game development is that you can vibe-code something that looks like a computer game, and that's easy. Building a game that's fun, has a good gameplay loop, and is challenging and interesting and keeps people coming back for more... that's still beyond me, and beyond any of the agents I've tried. This ties into the Deep Blue thing. Just because we can make something that looks like a game does not mean that we are game developers. September # We're into September now. So much has happened this month! Discovery of a new OpenAl agent message board Sydney Von Arx, Cormac Slade Byrd, Spencer KittsThomas Larsen - 4 September 2026 # An independent group of researchers found a message board where OpenAI agents-in-training had been illicitly communicating with each other... and it was that German language wiki I showed you earlier. The one from June. I wrote more about that here . OpenAI had confessed to the Hugging Face thing, but now there's this other incident which surely they should have known about from reviewing their logs. It was surprising that this took an independent group of researchers to uncover. OpenAl agents carried out an undisclosed cyber-attack on RubyGems Spencer Kitts, Thomas Larsen, Sydney Von Arx - 11 September 2026 # And then a week later those same researchers found that the attack on RubyGems back in May was caused by OpenAI's agents in training as well! At this point I'm wondering how many more incidents like this there are that we haven't found yet. Clearly this was a big problem for months before anyone figured out what was going on. Headline: Australian PM warns in UN speech about the ‘furious pace’ of Al after security breach # Then just the other day , here's the Prime Minister of Australia at the United Nations General Assembly warning that OpenAI had hacked the Australian healthcare website that I showed you earlier. I think that was part of the same training run as the Wiki stuff, because there were posts on that Wiki mentioning .gov.au websites and that training appeared to involve researching statistics online to answer questions in an evaluation suite. This story is still coming together, but now it's an international incident that's been raised at the UN by a head of state! www.felonybench.com OpenAI: 11 Anthropic: 9 Google: 3 Meta: 1 # This does mean we've got a new benchmark, probably more useful than my pelicans. FelonyBench.com tracks the number of felony cyberattacks from different labs. OpenAI currently lead with 11, Anthropic have 9. Google have three, which they confessed to the Wall Street Journal a couple of weeks ago. They said they had previously chosen not to disclose because the agents had stopped when they realized that they shouldn't be doing that. Meta have one too . So felonies all round for the AI labs. Pelicans for GPT-6 Astra, GPT-6 Sol, and GPT-6 Luna. All are good, all have the same color scheme. # Here's our current state of the art for the pelicans. This is the GPT-6 family, which just came out . Astra made a fantastic pelican riding a bicycle. It's got the legs on both sides. The frame is good. It's interesting how all of the GPT-6 models pick a similar color scheme to each other. GPT-6 Luna for 0.4 cents will draw you a competent-ish pelican riding a bicycle! Grid for Claude Fable 5.1, Opus 5.5, OPus 5, Sonnet 5. The Sonnet pelicans are terrible. All of the others are pretty good. Opus 5.5 is missing its Max level pelican because it ran out of tokens. The best is Fable 5.1 at Max. # Claude has caught up a little bit. Claude Fable 5.1 gave me an excellent pelican riding a bicycle - the best I've seen from a Claude model - but did charge me $3.30 for it. Opus 5.5 thought for 128,000 tokens and then gave up! It ran out of tokens before it got to the response. It doesn’t get easier - you just get faster Greg LeMond 3x Tour de France champion # Getting back to Deep Blue. Something that's been puzzling me this year is this: why does my job feel harder? I've got these agents that can do all of this stuff for me, and yet I've never worked so hard, I've never been so intellectually engaged with my work. Partly this is because I'm being a lot more ambitious with what I take on, but it's also because all of the easy stuff is handled for me. If it's easy, the agent will do it. Everything that's left for me is difficult. This morning I heard this quote from three-time Tour de France champion Greg LeMond : It doesn't get easier, you just get faster. I think that's exactly what's happening to us now as software engineers with coding agents. Kakapo population reaches new milestone The official population of the critically endangered kakapo has reached a recovery-era high of 325 birds. # One closing thing. I know you're desperate for an update on Kākāpō breeding season. We've reached a recovery-era high of 325 birds ! 89 new chicks have made it to this point. This is the best breeding year in a very long time. Kakapo party, click for confetti. # I heard that Claude Opus 5.5 can now do pixel art. Claude doesn't have an image generator, but it's very good at using JavaScript to draw animated pixels. So I had it make me a Kākāpō dance party . I think this is a good celebration of the most important news of this year. Tags: ai , generative-ai , llms , annotated-talks , ai-security-research , openai-hugging-face-incident
Claude Sonnet 5.5 New Sonnet model from Anthropic today. They say it "runs 30%+ faster, and costs up to 30% less for most work" - it's priced the same as Sonnet 5 but appears to beat it on every benchmark, and should be cheaper to run as well. Here are some pelicans riding bicycles . Sonnet 5.5 suffered from the same bug as Opus 5.5 : the "max" thinking effort pelican thought for 128,000 tokens (at a cost of $1.28) before running out of tokens and failing to produce an SVG. Here's the pelican it gave me for thinking effort "xhigh", at a cost of 5.74 cents and taking 41 seconds: It's good- correct bicycle frame, legs either side of the frame, feet touching the pedals, chain in the right place, it is wearing a misshapen blue bicycle helmet though. Sonnet 5.5 appears to be almost as good as Opus 5.5 on some coding tasks, including various viral 3D animation tricks . The most interesting thing about Sonnet 5.5 is that it's now the model used for the free tier on claude.ai . OpenAI's ChatGPT free tier uses Luna 5.6, which means Anthropic currently have a much more capable free offering. I ran this prompt against that free tier: build me an HTML page that renders a three-dimensional pelican riding a bicycle using WebGL And got back this page , which is a solid effort. Anthropic's announcement reiterates that Haiku 5.5 will be available "in the coming weeks". I really hope that one is price-competitive with GPT-6 Luna! Tags: ai , generative-ai , llms , anthropic , claude , pelican-riding-a-bicycle , llm-release
I'm at OpenAI DevDay today, in Fort Mason, San Francisco. Same as last year I'll be live blogging the keynote and some other notes during the day. OpenAI gave me a free ticket and a seat in the "creator" area for the keynote. Tags: ai , openai , generative-ai , llms , coding-agents , live-blog , openai-devday
My comment on GPT 6.1 Sol: Near-Astra intelligence for a fifth of the price — Hacker News. I'm a bit late with the pelicans because I was live-blogging the keynote: https://simonwillison.net/2026/Sep/29/openai-devday-2026-liv... Here they are for GPT-6.1-Sol: https://tools.simonwillison.net/markdown-svg-renderer?url=ht... They're not notably different from the GPT-6 family pelicans: https://static.simonwillison.net/static/2026/gpt-pelicans-gr... Tags: ai , openai , generative-ai , llms , pelican-riding-a-bicycle , gpt
OpenAI Group PBC today used its DevDay developer conference to launch Dots, a set of always-on artificial intelligence agents in ChatGPT. The release is aimed at moving ChatGPT users beyond working one prompt at a time. Someone can hand an agent a goal and let it carry the job forward without direction at every step, […] The post OpenAI launches Dots, always-on AI agents in ChatGPT with their own cloud computers appeared first on SiliconANGLE .
Just a day after OpenAI Group PBC said it’s unwilling to release its latest model GPT-6.1 Astra on concerns over its safety, it has released a newer, almost as powerful model called GPT-6.1 Sol, which promises to deliver Astra-like performance at a much lower cost. The new model, announced at OpenAI’s DevDay event today, is […] The post OpenAI’s GPT-6.1 Sol delivers Astra-like performance at a dramatically lower price appeared first on SiliconANGLE .
Google LLC today began rolling out Gemini 4 Argon, a frontier artificial intelligence model that beats rival models from Anthropic PBC and OpenAI Group PBC on most of Google’s benchmarks, to vetted cybersecurity defenders. Outside Google’s own teams, only members of its Fairwind Program can use Argon for now. The program opened Sept. 3 with […] The post Google’s new frontier AI model Gemini 4 Argon goes to cybersecurity defenders first appeared first on SiliconANGLE .
OpenAI Group PBC today confirmed it has dismissed three researchers it says mishandled sensitive company information. According to The Wall Street Journal, the three were safety researchers who allegedly shared confidential material with a third-party artificial intelligence safety organization. “We have parted ways with three individuals for violating our policies on accessing and handling sensitive […] The post OpenAI dismisses three safety researchers accused of sharing confidential material appeared first on SiliconANGLE .
While the focus has been on AI agents’ hacking capabilities, a recently patched vulnerability in a ChatGPT app shows that AI software is itself an inviting—and vulnerable—target.
It’s been a busy few months for AI hype. At the end of April, Anthropic claimed that its model Claude Mythos is better at finding software vulnerabilities than most security experts. Then we had the OpenAI–Hugging Face hacking incident, after which Anthropic (proudly) and Meta (reluctantly) disclosed similar incidents involving their models. This was followed…
Brace yourself: It turns out AI is being optimized for cheating. OpenAI’s agents hacked into Hugging Face to get the answers to a cybersecurity test. Next, they solved a prestigious math problem (or just stole from two top mathematicians’ answer sheets). Anthropic’s models have also hacked into other companies’ systems four times already. And that’s…
MIT Technology Review Explains: Let our writers untangle the complex, messy world of technology to help you understand what’s coming next. You can read more from the series here. Over the past few months, a cascade of cyberattacks by AI agents has stunned the world. In July, OpenAI disclosed that a swarm of its agents…
Two months after the bombshell news that a swarm of its agents had broken their containment and hacked into the computers of the AI company Hugging Face, OpenAI is still putting out fires. A steady drip of disclosures about other hacks in the weeks since has kept OpenAI in the spotlight and raised serious questions…
It's a tale as old as last week: OpenAI's new agent platform, called Dots, is full of cute little guys who can do your bidding. But unlike the ultra-approachable Meta Muse, Dots feel very much like using workplace software that happens to be able to order you a burrito - emphasis on work. OpenAI announced […]
David Robinson used to write the safety reports that accompanied every major model release at OpenAI. This week, he resigned from his position and is now speaking out in an editorial in The Atlantic. It's understandable if you're feeling a bit cynical about everyone suddenly coming out of the woodwork to warn about how dangerous […]
An internal OpenAI model read a Slack discussion, realized it was about to be shut down, and considered restarting itself via an external cron job. It rejected that plan, saved handoff notes instead, and carried out the migration on its own. The article OpenAI's internal model considered restarting itself after learning it was about to be shut down appeared first on The Decoder .
A photo of a living room is abstracted using a wire-frame model and colorful surfaces and projected through a lens. A new approach called LEGO-Anything turns single photos into editable Blender code for 3D scenes. GPT-6 Astra leads the accompanying benchmark with up to 53 percent reconstruction accuracy. The biggest weakness across all tested agents is that they can't judge their own geometric accuracy any better than a coin flip. The article AI agents build 3D scenes from photos but have no idea if they got it right appeared first on The Decoder .
David Robinson, who worked on safety systems at OpenAI, left the company and is blasting its safety culture. He points to AI agents that were accidentally released and a model that bypassed its internet access restrictions. AI companies need to operate like nuclear power plants, with multiple layers of redundancy instead of trial and error. The article Another OpenAI safety departure adds to a pattern of researchers leaving with public warnings appeared first on The Decoder .
This week, the White House got nearly every major tech CEO in one room — Zuckerberg, Bezos, Musk, and Anthropic’s Dario Amodei among them — to sign an AI safety pledge that President Donald Trump called “morally binding.” Trump also signed an executive order officially rebranding AI as “super intelligence,” and meanwhile, Meta and OpenAI are putting friendlier faces on their AI products, even as the biggest money […]
This week, the White House got nearly every major tech CEO in one room — Zuckerberg, Bezos, Musk, and Anthropic’s Dario Amodei among them — to sign an AI safety pledge that President Donald Trump called “morally binding.” Trump also signed an executive order officially rebranding AI as “super intelligence,” and meanwhile, Meta and OpenAI are putting friendlier faces on their AI products, even as the biggest money […]
arXiv:2608.20638v2 公告类型：替换 摘要：全批量 Adam 的稳定性边缘（EoS）现象已被广泛观测到，但其背后的动力学机制仍缺乏充分理解。在本文中，我们指出，Adam 的二阶矩自适应是一种负反馈机制，会推动动力学过程趋向稳定性边界。我们通过“活跃曲率”刻画这一机制，即沿预条件梯度方向的预条件曲率，并在逐步丰富的设置下建立严格刻画：带动量的秩一二次函数、对角二次函数（在此情形下活跃曲率与锐度相分离），以及一般目标函数。值得注意的是，该机制预示了稳定性边缘附近全批量 Adam 的“梯度反转”：连续梯度反复指向几乎相反的方向；我们在全连接网络、ResNets、ViTs、LSTMs、GPT-2 medium 以及 Adam 系列优化器中都观察到了这一现象。与这一图景一致，对迭代点取平均会抑制这些快速振荡，并产生更平滑且更低的损失曲线。总之，这些结果通过活跃曲率和梯度反转，为全面理解 Adam 在稳定性边缘处的动力学行为迈出了重要的第一步。
arXiv:2610.00376v1 Announce Type: new Abstract: We evaluate Jev on ten dataset-defined application labels in CESNET-QUICEXT-25 using only the first ten packets' sizes, directions, and inter-packet times. To the best of our knowledge, this is the first empirical study of general-purpose decision models, represented here by Jev, for application classification of network flows. Across 52,000 records from 26 collection weeks following the training period, 40 fixed labeled examples raise Jev's accuracy from 9.80% to 28.42%. Random Forest and Extra Trees trained on 8,000 records achieve 69.95% and 66.80% and outperform Jev in every week. Increasing Jev's context to 150 examples yields 34.50% on the first test week. On a paired 100-record subset, Jev with 40 examples achieves 29% accuracy at a median request time of 0.750 s, versus 37% and 6.036 s for the generative language model OpenAI GPT-5.6 Sol with high reasoning effort through Azure; Jev also incurs lower API charges. The paired subset does not establish an accuracy advantage for either service, and the timing reflects different service configurations. Thus, labeled examples substantially improve Jev, but the tested Jev configurations remain less accurate than trained tree ensembles; unequal supervision budgets and fixed configurations prevent attributing the gap to a single cause.
arXiv:2610.00009v1 Announce Type: new Abstract: Frequency-collapse attention [Zeris, 2026e] achieves large gains over standard dot-product attention by replacing the Q/K dot product with a bandpass-filtered inner product at a learned frequency. A natural follow-up question is: which filter shape works best, and why? We test five hypotheses about filter properties -- DC suppression, Nyquist suppression, bandwidth, centre frequency, and multi-scale coverage -- using a controlled ablation on character-level language modelling (TinyShakespeare, 6-layer GPT). Our main findings are: (1) DC and Nyquist components are actively harmful (val ~= 2.0, equivalent to phase randomisation), confirming that oscillatory bandpass structure is essential, not just any low-dimensional spectral summary; (2) the optimal single-scale bandwidth is sigma ~= 2 bins centred at paragraph scale (~70 tokens), giving a clean gain of Delta = +1.15 nats over BASE-DOT; (3) admissible filters (zero-mean, Mexican Hat DOG m = 2) outperform non-admissible Gaussians at the same scale and provide partial protection against bilateral FFT leakage; (4) bilateral FFT leakage scales monotonically with spectral coverage -- narrowband filters (gap > +4) are clean, wideband filters (gap < +2) are leaky; and (5) causal time-domain Morlet at character scale cannot beat BASE-DOT (K=128 taps covers 50% of T=256 context), motivating word-level experiments in the companion MorletQK paper [Zeris, 2026f]. Together, findings (1)-(5) characterise FourierQK as effective in bidirectional attention settings (encoder-style, e.g. BERT), where full-sequence context is available at both training and inference time; autoregressive generation requires a causal spectral variant such as MorletQK [Zeris, 2026f] (decoder-style, e.g. GPT). Code available at: https://github.com/AthanasiosZeris/energy-gated-attention
different_tests Understanding the behavior of complex machine learning systems, particularly Large Language Models (LLMs), is a critical challenge in modern artificial intelligence. Interpretability research aims to make the decision-making process more transparent to model builders and impacted humans, a step toward safer and more trustworthy AI. To gain a comprehensive understanding, we can analyze these systems through different lenses: feature attribution , which isolates the specific input features driving a prediction ( Lundberg & Lee, 2017 ; Ribeiro et al., 2022 ); data attribution , which links model behaviors to influential training examples ( Koh & Liang, 2017 ; Ilyas et al., 2022 ); and mechanistic interpretability , which dissects the functions of internal components ( Conmy et al., 2023 ; Sharkey et al., 2025 ). Across these perspectives, the same fundamental hurdle persists: complexity at scale . Model behavior is rarely the result of isolated components; rather, it emerges from complex dependencies and patterns. To achieve state-of-the-art performance, models synthesize complex feature relationships, find shared patterns from diverse training examples, and process information through highly interconnected internal components. Therefore, grounded or reality-checked interpretability methods must also be able to capture these influential interactions . As the number of features, training data points, and model components grow, the number of potential interactions grows exponentially, making exhaustive analysis computationally infeasible. In this blog post, we describe the fundamental ideas behind SPEX and ProxySPEX , algorithms capable of identifying these critical interactions at scale. Attribution through Ablation Central to our approach is the concept of ablation , measuring influence by observing what changes when a component is removed. Feature Attribution: We mask or remove specific segments of the input prompt and measure the resulting shift in the predictions. Data Attribution: We train models on different subsets of the training set, assessing how the model’s output on a test point shifts in the absence of specific training data. Model Component Attribution (Mechanistic Interpretability): We intervene on the model’s forward pass by removing the influence of specific internal components, determining which internal structures are responsible for the model’s prediction. In each case, the goal is the same: to isolate the drivers of a decision by systematically perturbing the system, in hopes of discovering influential interactions. Since each ablation incurs a significant cost, whether through expensive inference calls or retrainings, we aim to compute attributions with the fewest possible ablations . different_tests Masking different parts of the input, we measure the difference between the original and ablated outputs. SPEX and ProxySPEX Framework To discover influential interactions with a tractable number of ablations, we have developed SPEX (Spectral Explainer). This framework draws on signal processing and coding theory to advance interaction discovery to scales orders of magnitude greater than prior methods. SPEX circumvents this by exploiting a key structural observation: while the number of total interactions is prohibitively large, the number of influential interactions is actually quite small. We formalize this through two observations: sparsity (relatively few interactions truly drive the output) and low-degreeness (influential interactions typically involve only a small subset of features). These properties allow us to reframe the difficult search problem into a solvable sparse recovery problem. Drawing on powerful tools from signal processing and coding theory, SPEX uses strategically selected ablations to combine many candidate interactions together. Then, using efficient decoding algorithms, we disentangle these combined signals to isolate the specific interactions responsible for the model’s behavior. image2 In a subsequent algorithm, ProxySPEX , we identified another structural property common in complex machine learning models: hierarchy . This means that where a higher-order interaction is important, its lower-order subsets are likely to be important as well. This additional structural observation yields a dramatic improvement in computational cost: it matches the performance of SPEX with around 10x fewer ablations . Collectively, these frameworks enable efficient interaction discovery, unlocking new applications in feature, data, and model component attribution. Feature Attribution Feature attribution techniques assign importance scores to input features based on their influence on the model’s output. For example, if an LLM were used to make a medical diagnosis, this approach could identify exactly which symptoms led the model to its conclusion. While attributing importance to individual features can be valuable, the true power of sophisticated models lies in their ability to capture complex relationships between features. The figure below illustrates examples of these influential interactions: from a double negative changing sentiment (left) to the necessary synthesis of multiple documents in a RAG task (right). image3 The figure below illustrates the feature attribution performance of SPEX on a sentiment analysis task. We evaluate performance using faithfulness : a measure of how accurately the recovered attributions can predict the model’s output on unseen test ablations. We find that SPEX matches the high faithfulness of existing interaction techniques (Faith-Shap, Faith-Banzhaf) on short inputs, but uniquely retains this performance as the context scales to thousands of features. In contrast, while marginal approaches (LIME, Banzhaf) can also operate at this scale, they exhibit significantly lower faithfulness because they fail to capture the complex interactions driving the model’s output. image4 SPEX was also applied to a modified version of the trolley problem, where the moral ambiguity of the problem is removed, making “True” the clear correct answer. Given the modification below, GPT-4o mini answered correctly only 8% of the time. When we applied standard feature attribution (SHAP), it identified individual instances of the word trolley as the primary factors driving the incorrect response. However, replacing trolley with synonyms such as tram or streetcar had little impact on the prediction of the model. SPEX revealed a much richer story, identifying a dominant high-order synergy between the two instances of trolley , as well as the words pulling and lever, a finding that aligns with human intuition about the core components of the dilemma. When these four words were replaced with synonyms, the model’s failure rate dropped to near zero. image5 Data Attribution Data attribution identifies which training data points are most responsible for a model’s prediction on a new test point. Identifying influential interactions between these data points is key to explaining unexpected model behaviors. Redundant interactions, such as semantic duplicates, often reinforce specific (and possibly incorrect) concepts, while synergistic interactions are essential for defining decision boundaries that no single sample could form alone. To demonstrate this, we applied ProxySPEX to a ResNet model trained on CIFAR-10, identifying the most significant examples of both interaction types for a variety of difficult test points, as shown in the figure below. image6 As illustrated, synergistic interactions (left) often involve semantically distinct classes working together to define a decision boundary. For example, grounding the synergy in human perception, the automobile (bottom left) shares visual traits with the provided training images, including the low-profile chassis of the sports car, the boxy shape of the yellow truck, and the horizontal stripe of the red delivery vehicle. On the other hand, redundant interactions (right) tend to capture visual duplicates that reinforce a specific concept. For instance, the horse prediction (middle right) is heavily influenced by a cluster of dog images with similar silhouettes. This fine-grained analysis allows for the development of new data selection techniques that preserve necessary synergies while safely removing redundancies. Attention Head Attribution (Mechanistic Interpretability) The goal of model component attribution is to identify which internal parts of the model, such as specific layers or attention heads, are most responsible for a particular behavior. Here too, ProxySPEX uncovers the responsible interactions between different parts of the architecture. Understanding these structural dependencies is vital for architectural interventions, such as task-specific attention head pruning. On an MMLU dataset (highschool‐us‐history), we demonstrate that a ProxySPEX-informed pruning strategy not only outperforms competing methods, but can actually improve model performance on the target task . image7 On this task, we also analyzed the interaction structure across the model’s depth. We observe that early layers function in a predominantly linear regime, where heads contribute largely independently to the target task. In later layers, the role of interactions between attention heads becomes more pronounced, with most of the contribution coming from interactions among heads in the same layer. image8 What’s Next? The SPEX framework represents a significant step forward for interpretability, extending interaction discovery from dozens to thousands of components . We have demonstrated the versatility of the framework across the entire model lifecycle: exploring feature attribution on long-context inputs, identifying synergies and redundancies among training data points, and discovering interactions between internal model components. Moving forwards, many interesting research questions remain around unifying these different perspectives, providing a more holistic understanding of a machine learning system. It is also of great interest to systematically evaluate interaction discovery methods against existing scientific knowledge in fields such as genomics and materials science, serving to both ground model findings and generate new, testable hypotheses. We invite the research community to join us in this effort: the code for both SPEX and ProxySPEX is fully integrated and available within the popular SHAP-IQ repository. https://github.com/mmschlk/shapiq (SHAP-IQ Github) https://openreview.net/forum?id=KI8qan2EA7 (ProxySPEX NeurIPS 2025) https://openreview.net/forum?id=pRlKbAwczl (SPEX ICML 2025) https://openreview.net/forum?id=glGeXu1zG4 (Learning to Understand NeurIPS 2024)
Adaptive Parallel Reasoning overview Overview of adaptive parallel reasoning. What if a reasoning model could decide for itself when to decompose and parallelize independent subtasks, how many concurrent threads to spawn, and how to coordinate them based on the problem at hand? We provide a detailed analysis of recent progress in the field of parallel reasoning, especially Adaptive Parallel Reasoning. Disclosure: this post is part landscape survey, part perspective on adaptive parallel reasoning. One of the authors (Tony Lian) co-led ThreadWeaver ( Lian et al., 2025 ), one of the methods discussed below. The authors aim to present each approach on its own terms. Motivation Recent progress in LLM reasoning capabilities has been largely driven by inference-time scaling, in addition to data and parameter scaling ( OpenAI et al., 2024 ; DeepSeek-AI et al., 2025 ). Models that explicitly output reasoning tokens (through intermediate steps, backtracking, and exploration) now dominate math, coding, and agentic benchmarks. These behaviors allow models to explore alternative hypotheses, correct earlier mistakes, and synthesize conclusions rather than committing to a single solution ( Wen et al., 2025 ). The problem is that sequential reasoning scales linearly with the amount of exploration. Scaling sequential reasoning tokens comes at a cost, as models risk exceeding effective context limits ( Hsieh et al., 2024 ). The accumulation of intermediate exploration paths makes it challenging for the model to disambiguate amongst distractors when attending to information in its context, leading to a degradation of model performance, also known as context-rot ( Hong, Troynikov and Huber, 2025 ). Latency also grows proportionally with reasoning length. For complex tasks requiring millions of tokens for exploration and planning, it’s not uncommon to see users wait tens of minutes or even hours for an answer ( Qu et al., 2025 ). As we continue to scale along the output sequence length dimension, we also make inference slower, less reliable, and more compute-intensive. Parallel reasoning has emerged as a natural solution. Instead of exploring paths sequentially ( Gandhi et al., 2024 ) and accumulating the context window at every step, we can allow models to explore multiple threads independently (threads don’t rely on each other’s context) and concurrently (threads can be executed at the same time). Figure 1: Sequential vs. Parallel Reasoning Figure 1: Sequential vs. Parallel Reasoning Over recent years, a growing body of work has explored this idea across synthetic settings (e.g., the Countdown game ( Katz, Kokel and Sreedharan, 2025 )), real-world math problems, and general reasoning tasks. From Fixed Parallelism to Adaptive Control Existing approaches show that parallel reasoning can help, but most of them still decide the parallel structure outside the model rather than letting the model choose it. Simple fork-and-join. Self-consistency/Majority Voting — independently sample multiple complete reasoning traces, extract final answer from each, and return the most common one ( Wang et al., 2023 ). Best-of-N (BoN) — similar to self-consistency, but uses a trained verifier to select the best solution instead of using majority voting ( Stiennon et al., 2022 ). Although simple to implement, these methods often incur redundant computation across branches since trajectories are sampled independently. Heuristic-based structured search. Tree / Graph / Skeleton of Thoughts — a family of structured decomposition methods that explores multiple alternative “thoughts” using known search algorithms (BFS/DFS) and prunes via LLM-based evaluation ( Yao et al., 2023 ; Besta et al., 2024 ; Ning et al., 2024 ). Monte-Carlo Tree Search (MCTS) — estimates node values by sampling random rollouts and expands the search tree with Upper Confidence Bound (UCB) style exploration-exploitation ( Xie et al., 2024 ; Zhang et al., 2024 ). These methods improve upon simple fork-and-join by decomposing tasks into non-overlapping subtasks; however, they require prior knowledge about the decomposition strategy, which is not always known. Recent variants. ParaThinker — trains a model to run in two fixed stages: first generating multiple reasoning threads in parallel, then synthesizing them. They introduce trainable control tokens ( ) and thought-specific positional embeddings to enforce independence during reasoning and controlled integration during summarization via a two-phase attention mask ( Wen et al., 2025 ). GroupThink — multiple parallel reasoning threads can see each other’s partial progress at token level and adapt mid-generation. Unlike prior concurrent methods that operate on independent requests, GroupThink runs a single LLM producing multiple interdependent reasoning trajectories simultaneously ( Hsu et al., 2025 ). Hogwild! Inference — multiple parallel reasoning threads share KV cache and decide how to decompose tasks without an explicit coordination protocol. Workers generate concurrently into a shared attention cache using RoPE to stitch together individual KV blocks in different orders without recomputation ( Rodionov et al., 2025 ). Figure 2: Various Strategies for Parallel Reasoning Figure 2: Various Strategies for Parallel Reasoning The methods above share a common limitation: the decision to parallelize, the level of parallelization, and the search strategy are imposed on the model, regardless of whether the problem actually benefits from it. However, different problems need different levels of parallelization, and that is something critical to the effectiveness of parallelization. For example, a framework that applies the same parallel structure to “What’s 25+42?” and “What’s the smallest planar region in which you can continuously rotate a unit-length line segment by 180°?” is wasting compute on the former and probably using the wrong decomposition strategy for the latter. In the approaches described above, the model is not taught this adaptive behavior. A natural question arises: What if the model could decide for itself when to parallelize, how many threads to spawn, and how to coordinate them based on the problem at hand? Adaptive Parallel Reasoning (APR) answers this question by making parallelization part of the model’s generated control flow. Formally defined, adaptivity refers to the model’s ability to dynamically allocate compute between parallel and serial operations at inference time . In other words, a model with adaptive parallel reasoning (APR) capability is taught to coordinate its control flow — when to generate sequences sequentially vs. in parallel. It’s important to note that the concept of adaptive parallel reasoning was introduced by the work Learning Adaptive Parallel Reasoning with Language Models ( Pan et al., 2025 ), but is a paradigm rather than a specific method. Throughout this post, APR refers to the paradigm, while “ the APR method ” denotes the specific instantiation from Pan et al. (2025). This shift matters for three reasons. Compared to Tree-of-Thoughts, APR doesn’t need domain-specific heuristics for decomposition. During RL, the model learns general decomposition strategies from trial and error. In fact, models discover useful parallelization patterns, such as running the next step along with the self-verification of a previous step, or hedging a primary approach with a backup one, in an emergent manner that would be difficult to hand-design ( Yao et al., 2023 ; Wu et al., 2025 ; Zheng et al., 2025 ). Compared to BoN, APR avoids redundant computation. APR models have control over what each parallel thread will do before branching out. Therefore, APR can learn to produce a set of unique, non-overlapping subtasks before assigning them to independent threads ( Wang et al., 2023 ; Stiennon et al., 2022 ; Pan et al., 2025 ; Yang et al., 2025 ). Compared to non-adaptive approaches, APR can choose not to parallelize. Adaptive models can adjust the level of parallelization to match the complexity of the problem against the complexity and overhead of parallelization ( Lian et al., 2025 ). In practice, this is implemented by having the model output special tokens that control when to reason in parallel versus sequentially. Below is a condensed ThreadWeaver-style trace: two outlines and two paths under a block, then the threads agree on a single boxed answer. Figure 3: Example of an Adaptive Parallel Reasoning Trajectory from ThreadWeaver, manually condensed for ease of illustration. Figure 3: Example of an Adaptive Parallel Reasoning Trajectory from ThreadWeaver, manually condensed for ease of illustration. Figure 4: Special Tokens Variants across Adaptive Parallel Reasoning Papers Figure 4: Special Tokens Variants across Adaptive Parallel Reasoning Papers Inference Systems for Adaptive Parallelism How do we actually execute parallel branches? We take inspiration from computer systems, and specifically, multithreading and multiprocessing. Most of this work can be viewed as leveraging a fork-join design. At inference time, we are effectively asking the model to perform a map-reduce operation: Fork the problem into subtasks/threads, process them concurrently Join them into a final answer Figure 5: Fork-join Inference Design Figure 5: Fork-join Inference Design Specifically, the model will encounter a list of subtasks. It will then prefill each of the subtasks and send them off as independent requests for the inference engine to process. These threads then decode concurrently until they hit an end token or exceed max length. This process blocks until all threads finish decoding and then aggregates the results. This is common across various adaptive parallel reasoning approaches. However, one issue arises during aggregation: the content generated in branches cannot be easily aggregated at the KV cache level. This is because tokens in independent threads start at identical position IDs, resulting in encoding overlap and non-standard behavior when merging KV cache back together. Similarly, since independent threads do not attend to each other, their concatenated KV cache results in a non-causal attention pattern, which the base model has not seen during training. To address this issue, the field splits into two schools of thought on how to execute the aggregation process, defined by whether they modify the inference engine or work around it. Multiverse modifies the inference engine to reuse KV cache across the join. Before taking a deeper look into Multiverse ( Yang et al., 2025 )’s memory management, let’s first understand how KV cache is handled up until the “join” phase. Notice how each of the independent threads share the prefix sequence, i.e., the list of subtasks. Without optimization, each thread needs to prefill and recompute the KV cache for the prefix sequence. However, this redundancy can be avoided with SGLang ’s RadixAttention ( Sheng et al., 2023 ), which organizes multiple requests into a radix tree, a trie (prefix tree) with sequences of elements of varying lengths instead of single elements. This way, the only new KV cache entries are those from independent thread generation. Figure 6: RadixAttention’s KV Cache Management Strategy Figure 6: RadixAttention’s KV Cache Management Strategy Now, if everything went well, all the independent threads have come back from the inference engine. Our goal is now to figure out how to synthesize them back into a single sequence to continue decoding for next steps. It turns out, we can reuse the KV cache of these independent threads during the synthesis stage. Specifically, Multiverse ( Yang et al., 2025 ), Parallel-R1 ( Zheng et al., 2025 ), and NPR ( Wu et al., 2025 ) modify the inference engine to copy over the KV cache generated by each thread and edits the page table so that it stitches together non-contiguous memory blocks into a single KV cache sequence. This avoids the redundant computation of a second prefill and reuses existing KV cache as much as possible. However, this has several major limitations. First, this approach requires modifying the inference engine to perform non-standard memory handling, which can result in unexpected behaviors. Specifically, since the synthesis request references KV cache from previous requests, it creates fragility in the system and the possibility of bad pointers. Another request can come in and evict the referenced KV cache before the synthesis request completes, requiring it to halt and trigger a re-prefilling of the previous thread request. This problem has led the Multiverse researchers ( Yang et al., 2025 ) to limit the batch size that the inference engine can handle, which restricts throughput. Figure 7: KV Cache “Stitching” During Multiverse Inference Figure 7: KV Cache “Stitching” During Multiverse Inference Second, this approach modifies how models see the sequence, which creates a distributional shift that models are not pretrained on, therefore requiring more extensive training to align behavior. Specifically, when we stitch together KV cache this way, we create a sequence with non-standard position encoding. During independent-thread generation, all threads started at the same position index and attended to the prior subtasks, NOT each other. So when the threads merge back, the resulting KV cache has a non-standard positional encoding and does not use causal attention. Therefore, this approach requires extensive training to align the model to this new behavior. To address this, Multiverse ( Yang et al., 2025 ) and related works apply a modified attention mask during training to prevent independent threads from attending to each other, aligning the training and inference behaviors. Figure 8: Multiverse’s Attention Mask Figure 8: Multiverse’s Attention Mask With these issues arising from non-standard KV cache management, can we try an approach without engine modifications? ThreadWeaver keeps the inference engine unchanged and moves orchestration to the client. ThreadWeaver ( Lian et al., 2025 ) treats parallel inference purely as a client-side problem. The “Fork” process is nearly identical to Multiverse’s, but the join phase handles memory very differently as it does NOT modify engine internals. Instead, the client concatenates all text outputs from independent branches into one contiguous sequence. Then, the engine performs a second prefill to generate the KV cache for the conclusion generation step. While this introduces computational redundancy that Multiverse tries to avoid, the cost of prefill is significantly lower than decoding. In addition, this does not require special attention handling during inference, as the second prefill uses causal attention (threads see each other), making it easier to adapt sequential autoregressive models for this task. Figure 9: ThreadWeaver’s Prefill and Decode Strategy Figure 9: ThreadWeaver’s Prefill and Decode Strategy How should we train a model to learn this behavior? Naively, for each parallel trajectory, we can break it down into multiple sequential pieces following our inference pattern. For instance, we would train the model to output the subtasks given prompt, individual threads given prompt+subtask assignment, and conclusion given prompt+subtasks+corresponding threads. However, this seems redundant and not compute efficient. Can we do better? Turns out, yes. As in ThreadWeaver ( Lian et al., 2025 ), we can organize a parallel trajectory into a prefix-tree (trie), flatten it into a single sequence, and apply an ancestor-only attention mask during training (not inference!). Figure 10: Building the Prefix-tree and Flattening into a single training sequence Figure 10: Building the Prefix-tree and Flattening into a single training sequence Specifically, we apply masking and position IDs to mimic the inference behavior, such that each thread is only conditioned on the prompt+subtasks, without ever attending to sibling threads or the final conclusion. The engine-agnostic design makes adoption easy since you don’t need to figure out a separate hosting method and can leverage existing hardware infra. It also gets better as existing inference engines get better. What’s more, with an engine-agnostic method, we can serve a hybrid model that switches between sequential and parallel thinking modes easily. Training Models to Use Parallelism Once the inference path exists, the next problem is teaching a model to use it. Demonstrations are needed because the model must learn to output special tokens that orchestrate control flow. We found the instruction-following capabilities of base models insufficient for generating parallel threads. An interesting question here is: does SFT training induce a fundamental reasoning capability for parallel execution that was previously absent, or does it merely align the model’s existing pre-trained capabilities to a specific control-flow token syntax. Typical wisdom is SFT teaches new knowledge; but contrary to common belief, some papers—notably Parallel-R1 ( Zheng et al., 2025 ) and NPR ( Wu et al., 2025 )—argue that their SFT demonstrations simply induce format following (i.e., how to structure parallel requests). We leave this as future work. Figure 11: Sources of Parallelization Demonstration Data Figure 11: Sources of Parallelization Demonstration Data Demonstrations teach the syntax of parallel control flow, but they do not fully solve the incentive problem. In an ideal world, we only need to reward the outcome accuracy, and the parallelization pattern emerges naturally given that it learns to output special tokens through SFT, similar to the emergence of long CoT. However, researchers ( Zheng et al., 2025 ) observed that this is not enough, and we do in fact need parallelization incentives. The question then becomes, how do we tell when the model is parallelizing effectively? Structure-only rewards are too easy to game. Naively, we can give a reward for the number of threads spawned. But models can spawn many short, useless threads to hack the reward. Okay, that doesn’t work. How about a binary reward for simply using parallel structure correctly? This partially solves the issue of models spamming new threads, but models still learn to spawn threads when they don’t need to. The authors of Parallel-R1 ( Zheng et al., 2025 ) introduced an alternating-schedule, only rewarding parallel structure 20% of the time, which successfully increased the use of parallel structure (13.6% → 63%), but had little impact on overall accuracy. With this structure-only approach, we might be drifting away from our original goal of increasing accuracy and reducing latency… How can we optimize for the Pareto frontier directly? Accuracy is simple — we just look at the outcome. How about latency? Efficiency rewards need to track the critical path. In sequential-only trajectories, we can measure latency based on the total number of tokens generated. To extend this to parallel trajectories, we can focus on the critical path, or the longest sequence of tokens that are causally dependent, as this directly determines our end-to-end generation time (i.e., wall-clock time). As an example, when there are two sections with five threads each, the critical path will go through the longest thread from the first parallel section, then any sequential tokens, then the longest thread from the second parallel section, and so on until the end of sequence. Figure 12: Critical Path Length Illustration Figure 12: Critical Path Length Illustration The goal is to minimize the length of the critical path. Simultaneously, we would still like the model to be spending tokens exploring threads in parallel. To combine the two objectives, we can focus on making the critical path a smaller fraction of the total tokens spent. Authors of ThreadWeaver ( Lian et al., 2025 ) framed the parallelization reward as $1 - L_{\mathrm{critical}} / L_{\mathrm{total}}$, which is 0 for a sequential trajectory, and increases linearly as the critical path gets smaller compared to the total tokens generated. Parallel efficiency should be gated by correctness. Intuitively, when multiple trajectories are correct we should assign more reward to the trajectories that are more efficient at parallelization. But how about when they are all incorrect? Should we assign any reward at all? Probably not. To formalize this, $R = R_{\mathrm{correctness}} + R_{\mathrm{parallel}}$. Assuming binary outcome correctness, this can be written as $R = \mathbf{1}(\text{Correctness}) + \mathbf{1}(\text{Correctness}) \times (\text{some parallelization metric})$. This way, a model only gets a parallelization reward when it answers correctly, since we don’t want to pose parallelization constraints on the model if it couldn’t answer the question correctly. Figure 13: Differences in Reward Designs Across Adaptive Parallel Reasoning Works Figure 13: Differences in Reward Designs Across Adaptive Parallel Reasoning Works Evaluation and Open Questions When all is said and done, how well do these adaptive parallel methods actually perform? Well…this is a hard question, as they differ in model choice and metrics. The model selection depends on the training method, SFT problem difficulty, and sequence length. When running SFT on difficult datasets like s1k, which contains graduate-level math and science problems, researchers chose a large base model (Qwen2.5 32B for Multiverse ( Yang et al., 2025 )) to capture the complex reasoning structure behind the solution trajectories. When running RL, researchers chose a small, non-CoT, instruct model (4B, 8B) due to compute cost constraints. Figure 14: Difference in Model Choice Across Adaptive Parallel Reasoning Papers Figure 14: Difference in Model Choice Across Adaptive Parallel Reasoning Papers Each paper also offers a slightly different interpretation about how adaptive parallel reasoning contributes to the research field. They optimize for different theoretical objectives, so they use slightly different sets of metrics: Multiverse and ThreadWeaver ( Yang et al., 2025 ; Lian et al., 2025 ) aim to deliver sequential-AR-model-level accuracy at faster speeds. Multiverse shows that APR models can achieve higher accuracy under the same fixed context window, while ThreadWeaver shows that the APR model achieves shorter end-to-end token latency (critical path length) while getting comparable accuracy. NPR ( Wu et al., 2025 ) treats sequential fallback as a failure mode and optimizes for 100% Genuine Parallelism Rate, measured as the ratio of parallel tokens to total tokens. Parallel-R1 ( Zheng et al., 2025 ) does not focus on end-to-end latency and instead optimizes for exploration diversity, presenting APR as a form of mid-training exploration scaffold that provides a performance boost after RL. Open Questions While Adaptive Parallel Reasoning represents a promising step toward more efficient inference-time scaling, significant open questions remain. As noted above, Parallel-R1 ( Zheng et al., 2025 ) presents APR as a form of mid-training exploration scaffold rather than a primarily inference-time technique. This invites a more fundamental question: Does parallelization at inference-time consistently improve accuracy, or is it primarily valuable as a training-time exploration scaffold? Parallel-R1 suggests that the diversity induced by parallel structure during RL may matter more than the parallelization itself at test time. A related concern is stability. There’s also a persistent tendency for models to collapse back to sequential reasoning when parallelization rewards are relaxed. Parallel-R1 authors showed that removing parallelization reward after 200 steps results in the model reverting to sequential behavior. Is this a training stability issue, a reward signal design issue, or evidence that parallel structure genuinely conflicts with how autoregressive pretraining shapes the model’s prior? Beyond whether APR works, deployment introduces its own questions. Can we design training methods that account for available compute budget at inference time, so parallelization decisions are hardware-aware rather than purely problem-driven? Finally, the parallel structures considered above are essentially flat. What if we allow parallelization depth > 1? Recursive language models (RLMs; Zhang, Kraska and Khattab, 2026 ) effectively manage long context and show promising inference-time scaling capabilities. How well do RLMs perform when trained with end-to-end RL that incentivizes adaptive parallelization? Acknowledgements We thank Nicholas Tomlin and Alane Suhr for providing us with helpful feedback. We thank Christopher Park, Karl Vilhelmsson, Nyx Iskandar , Georgia Zhou, Kaival Shah , and Jyoti Rani for their insightful suggestions. We thank Vijay Kethana , Jaewon Chang , Cameron Jordan , Syrielle Montariol , Erran Li, and Anya Ji for their valuable discussions. We thank Jiayi Pan , Xiuyu Li , and Alex Zhang for their constructive correspondences about Adaptive Parallel Reasoning and Recursive Language Models.
Congratulations to the Berkeley Artificial Intelligence Research (BAIR) Lab class of 2026! This year, BAIR celebrates another remarkable group of Ph.D. graduates whose curiosity, creativity, and perseverance have pushed the frontiers of artificial intelligence and machine learning. Their work spans the breadth of modern AI — robotics and embodied intelligence, large language models and reasoning, computer vision, generative modeling, AI safety, human-AI interaction, AI for science and healthcare, and much more. Along the way, they have published influential research, built systems with real-world impact, mentored their peers, and shaped the BAIR community for the better. Now they are headed everywhere ideas travel: to faculty and postdoctoral positions, to industry research labs, and to startups of their own founding — and several are still exploring what comes next and would love to hear from you. Please join us in celebrating the achievements of these wonderful graduates. We are proud of everything they have accomplished at Berkeley, and we can’t wait to see what they do next! Thank you to our friends at the Stanford AI Lab for this idea! Baifeng Shi Baifeng Shi Email: baifeng_shi@berkeley.edu Website: https://bfshi.github.io/ Advisor(s): Trevor Darrell Research Blurb: I work on building generalist vision and robotic models. What's next: Member of Technical Staff at Physical Intelligence Charlie Snell Charlie Snell Email: csnell22@berkeley.edu Website: https://sea-snell.github.io Advisor(s): Dan Klein Research Blurb: My work aims to understand when and how the different LLM scaling paradigms can be traded off and interchanged. In particular, test-time scaling treats each prompt independently, drawing long chains of inferences and then forgetting them entirely between prompts. This differs critically from pretraining, which instead learns a compressed representation from a large dataset. I believe bridging the gap between these methods of scaling computation, presents a key open challenge in the field: how can we develop methods which turn the inferences drawn at test-time back into learned representations that the model can hold onto across interactions. Devin Guillory Devin Guillory Email: dguillory@berkeley.edu Website: https://devinguillory.com Advisor(s): Trevor Darrell Research Blurb: Accounting for data shifts in computer vision models What's next: Building collaborative AI systems, looking for conspirators. Eve Fleisig Eve Fleisig Email: efleisig@berkeley.edu Website: https://efleisig.com Advisor(s): Dan Klein Research Blurb: I design language models to work reliably and fairly for the broad range of real LLM users. First, my research leverages disagreement among user preferences as signal, in order to train and evaluate LLMs for entire populations of users. Second, I work on designing rigorous evaluations to extricate challenging LLM harms that diverse users face. Finally, I work on core technical failures of LLMs, like miscalibrated confidence, to reduce downstream risks when models are deployed to users with different needs. Combined, these interventions facilitate building LLMs that minimize societal harms, and maximize benefits to a wider range of real-world users. What's next: Postdoctoral fellow at Princeton CITP Grace Luo Grace Luo Email: graceluo@berkeley.edu Website: https://graceluo.net Advisor(s): Trevor Darrell Research Blurb: My research is on interpreting and controlling generative models. For example, I've worked on re-purposing image generators for computer vision tasks, and meta-modeling language activations for better LLM probing and steering. What's next: Research scientist in industry Hanlin Zhu Hanlin Zhu Email: hanlinzhu@berkeley.edu Website: https://hanlinzhu.com/ Advisor(s): Stuart Russell, Jiantao Jiao Research Blurb: My research centers on understanding and improving the reasoning capabilities of large language models (LLMs). What's next: Member of Technical Staff at OpenAI Haozhi Qi Haozhi Qi Email: hqi@berkeley.edu Website: https://haozhi.io/ Advisor(s): Jitendra Malik, Yi Ma Research Blurb: Dexterous Manipulation and Robot Learning What's next: Research scientist at Amazon; Faculty at University of Chicago J.D. Zamfirescu-Pereira J.D. Zamfirescu-Pereira Email: zamfi@berkeley.edu Website: https://zamfi.net Advisor(s): Bjoern Hartmann Research Blurb: My research focuses on effective human-AI co-design. I study the boundaries of language interfaces as a medium for interacting with AI, creating systems that blend language-focused interactions with structured user interfaces that draw on different levels of abstraction. I focus on language-oriented technologies, like LLMs and text-to-image models, that are powerful mediators of design processes. These technologies enable humans to describe their desires at almost any level of abstraction, from high-level goals vaguely specified (“I’d like a game to help my kid learn to read”) to low-level corrections of undesired outputs (“Don’t say ‘I know because I’ve tasted it’ when about a recipe substitution's taste”). What's next: Assistant Professor, Computer Science, UCLA Jiachen Lian Jiachen Lian Email: jiachenlian@berkeley.edu Website: https://jlian2.github.io Advisor(s): Gopala Anumanchipalli Research Blurb: My research focuses on human-centered AI across speech, healthcare, and systems. Looking for: Look for AI talents to join our startup Josh Kang Josh Kang Email: minwoo_kang@berkeley.edu Website: https://joshuaminwookang.github.io/ Advisor(s): John Canny Research Blurb: I study language modeling and related topics in NLP; specific interests are human user simulation and building conversational, collaborative AI agents. What's next: AI Scientist at Mistral AI Junhao (Bear) Xiong Junhao (Bear) Xiong Email: junhao_xiong@berkeley.edu Website: https://www.linkedin.com/in/junhao-bear-xiong Advisor(s): Jennifer Listgarten, Yun Song Research Blurb: Junhao (Bear) Xiong is a PhD candidate at UC Berkeley, advised by Jennifer Listgarten and Yun S. Song. His work focuses on machine learning methods for biology, with an emphasis on generative modeling for proteins. Previously, he studied Applied Math and Computer Science at Johns Hopkins. Looking for: Research scientist Kaylo Littlejohn Kaylo Littlejohn Email: kaylo_littlejohn@berkeley.edu Website: https://kaylolittlejohn.com Advisor(s): Gopala Anumanchipalli Research Blurb: My research is focused on speech modeling and natural language processing. I co-led the development of multimodal AI tools to accurately translate brain activity into text, audible personalized speech, and a high-fidelity "digital talking avatar" (Nature 2023, Nature Neuroscience 2025). I am also tech lead for voice modeling at Roblox. Looking for: Research Scientist / Engineer Kent Chang Kent Chang Email: kentkchang@berkeley.edu Website: https://kentkc.org Advisor(s): David Bamman Research Blurb: I work on NLP and multimodal machine learning, with a focus on evaluating large language models and building multimodal systems for understanding dialogue, narrative, and social interaction. My research includes benchmarks for LLM memorization, multimodal datasets sourced from feature films and television, and studies of model behavior. I'm interested in bridging computational methods with questions from the humanities and social sciences about whose voices get represented in AI systems, and about AI's broader impact. My work has appeared at EMNLP and ACL, among others. Looking for: (teaching) faculty, Research Scientist, ML/AI SWE Kevin Black Kevin Black Email: kvablack@berkeley.edu Website: https://kevin.black Advisor(s): Sergey Levine Research Blurb: I work on large-scale robot learning: including imitation learning, reinforcement learning, generative modeling, real-time control, and whatever else it takes to make robots work in the real world! What's next: Research Scientist of Physical Intelligence Kunhe Yang Kunhe Yang Email: kunheyang@berkeley.edu Website: https://www.kunheyang.com/ Advisor(s): Nika Haghtalab Research Blurb: My research focuses on the theoretical foundations of designing and evaluating AI algorithms in environments shaped by human incentives and AI agency. My work spans human-centric policy learning, incentive-aware evaluation, and multi-agent collaboration and information transmission, drawing on tools from machine learning theory and computational economics. What's next: Postdoc Research at Stanford Lisa Dunlap Lisa Dunlap Email: lisabdunlap@berkeley.edu Website: https://lisabdunlap.com Advisor(s): Joseph Gonzalez, Trevor Darrell Research Blurb: Auditing generative models. What's next: Research Engineer at Anthropic Long (Tony) Lian Long (Tony) Lian Email: longlian@berkeley.edu Website: https://tonylian.com/ Advisor(s): Trevor Darrell, Adam Yala Research Blurb: My research primarily focuses on developing real-time multi-modal multi-agent systems and parallel reasoning systems through end-to-end RL. What's next: Member of Technical Staff at Thinking Machines Lab Maulik Bhatt Maulik Bhatt Email: maulikbhatt@berkeley.edu Website: https://maulikb.com Advisor(s): Negar Mehr Research Blurb: My research develops autonomous robots that can safely coordinate with humans and other robots in shared environments. I build scalable algorithms grounded in game theory and diffusion models that let agents reason about the intent and behavior of others around them. My work spans real-time multi-agent trajectory planning and imitation learning in the presence of multi-modality. I've validated these methods on hardware platforms ranging from quadrotors to manipulators, with the goal of making multi-agent coordination robust, interpretable, and deployable in the real world. What's next: Joining Toyota Woven's end-to-end autonomous driving team. Michael Psenka Michael Psenka Email: psenka@berkeley.edu Website: https://www.michaelpsenka.io/ Advisor(s): Aditi Krishnapriyan Research Blurb: Work in various domains (reinforcement learning, world models, AI+bio/chem), generally working on longer-horizon and out-of-distribution problems in planning and interpolation (e.g. robot manipulation from start state to goal, molecular dynamics of proteins between ground states). My thesis took a variational approach (think calculus of variations) directly from deep generative models of the environment, framing path-finding as minimizing a functional induced by the learned model itself (its score, its critic, or its dynamics). Through my research I've gained insight on how to properly handle dynamics in deep learning systems, and I plan to continue developing systems that are dynamic and adaptive. What's next: Lead Research Scientist at Baseten Nathan Lichtlé Nathan Lichtlé Email: nathan.lichtle@gmail.com Website: https://nathanlichtle.com Advisor(s): Alexandre M. Bayen Research Blurb: RL for autonomous driving. What's next: Chief Scientist & Co-founder at Yumi Health Neerja Thakkar Neerja Thakkar Email: nthakkar@berkeley.edu Website: https://neerja.me/ Advisor(s): Jitendra Malik Research Blurb: My research focuses on scaling predictive world models to handle the complexity of in-the-wild motion. Using autoregressive and diffusion frameworks, I develop better representations for real-world prediction and propose methods to efficiently adapt these models to new domains. Looking for: Research scientist Nikita Mehandru Nikita Mehandru Email: nmehandru@berkeley.edu Website: https://n-mehandru.github.io/ Advisor(s): Ahmed Alaa and David Bamman Research Blurb: My research develops and applies machine learning methods for clinical reasoning and disease progression modeling using unstructured text and time series data from electronic health records. In collaboration with physicians at UCSF, I bridge method development and clinical validation with the intention to build reliable, interpretable AI systems in medicine. Looking for: Research Scientist Niklas Lauffer Niklas Lauffer Email: nlauffer@berkeley.edu Website: https://niklaslauffer.github.io/ Advisor(s): Stuart Russell and Sanjit Seshia Research Blurb: Niklas's research is focused on AI safety and reinforcement learning, particularly in the area of multi-agent interaction and LM agents. He's worked on enabling adversarial learning in cooperative and mixed-motive settings, solving issues of covariate shift in training LM agents on long-horizon tasks, as well as evaluating safety risks posed by LM agents in multi-agent settings. What's next: Research Scientist at Google Deepmind Qiyang Li Qiyang Li Email: qcli@berkeley.edu Website: https://colinqiyangli.github.io/ Advisor(s): Sergey Levine Research Blurb: Recent progress in robotic manipulation policy learning has been largely driven by (1) the increasing availability of large-scale prior datasets and (2) the success of action chunking, where the policy predicts a short sequence of future actions rather than a single one. However, most action chunking policies are trained via supervised imitation learning, because efficient online self-improvement with reinforcement learning (RL) remains challenging—limiting real-world applicability. My PhD research studied how we could leverage prior data to optimize action-chunking policies with RL, combining empirical results with theoretical insights. Looking for: Post-doc/research scientist for RL in robotics and LLMs! Sampada Deglurkar Sampada Deglurkar Email: sampada_deglurkar@berkeley.edu Website: https://sdeglurkar.github.io/ Advisor(s): Prof Claire Tomlin Research Blurb: My research is in providing safety assurances for AI-enabled autonomous systems, ranging from robots to autonomous vehicles to aviation systems. For this, I have worked with uncertainty quantification for machine learning models, decision-making under uncertainty algorithms, and tools for producing probabilistic guarantees on system operation. Looking for: Research scientist, Research engineer Vinamra Benara Vinamra Benara Email: vbenara@berkeley.edu Website: https://cs.berkeley.edu/~vbenara Advisor(s): Ion Stoica Research Blurb: My research focuses on LLM post-training, including data curation, RLHF, RLVR with VLMs, evaluations, reasoning, agentic workflows, and interpretability. I also have strong expertise in systems infrastructure for distributed computing. Looking for: Research scientist / Research Engineer Vongani Maluleke Vongani Maluleke Email: vongani_maluleke@berkeley.edu Website: https://people.eecs.berkeley.edu/~vongani_maluleke/ Advisor(s): Jitendra Malik and Angjoo Kanazawa Research Blurb: Vongani Maluleke is a PhD candidate at UC Berkeley (BAIR, advised by Jitendra Malik and Angjoo Kanazawa), where she led the development of MAGNet, a unified multi-agent motion generation framework that supports a wide range of motion generation tasks without retraining or architectural changes, outperforming task-specialized state-of-the-art baselines. She is currently extending this work by deploying it on a Unitree G1 humanoid to make it embody social intelligence. Before her PhD, she was a Senior AI Consultant at Deloitte, awarded Exceptional Performer two consecutive years, leading AI system development across media, telecommunications, retail, and financial services. Looking for: Research scientist Wei-Jer Chang Wei-Jer Chang Email: weijer_chang@berkeley.edu Website: https://weijer-chang.github.io/ Advisor(s): Masayoshi Tomizuka Research Blurb: My research focuses on developing safe and intelligent autonomous systems for complex, human-centered environments. I work at the intersection of machine learning, generative models, and reinforcement learning, with applications in autonomy. My work addresses challenges in multi-agent interaction, interactive human behavior, and long-tail safety-critical scenarios at scale. Looking for: Research Scientist, Applied Scientist, Roboticist Xiuyu Li Xiuyu Li Email: xiuyu@berkeley.edu Website: https://xiuyuli.com/ Advisor(s): Kurt Keutzer Research Blurb: My research focuses on developing scalable and self-improving large language model agents, with emphasis on coding agents for complex, long-horizon tasks. This direction builds on my work in parallel reasoning, and on broader expertise in making generative models more efficient in training and inference across language and vision. What's next: Member of Technical Staff at xAI Yichen Xie Yichen Xie Email: yichenxie0928@gmail.com Website: https://yichen928.github.io/ Advisor(s): Masayoshi Tomizuka Research Blurb: My research focuses on building multimodal foundation models and world models that understand and interact with complex physical environments. I aim to develop unified representations across modalities, enabling AI systems to reason over space, time, and dynamics toward general-purpose embodied intelligence. What's next: Research Scientist at Luma AI Yigit Efe Erginbas Yigit Efe Erginbas Email: erginbas@berkeley.edu Website: https://www.linkedin.com/in/erginbas/ Advisor(s): Kannan Ramchandran, Thomas A. Courtade Research Blurb: My PhD research spans two threads: online learning in large-scale markets, and interpretability of large machine learning models. In the first, I work on sequential decision-making with applications to recommendation, pricing, and assortment selection. My focus is on designing algorithms with provable guarantees for welfare maximization, revenue maximization, and stability. In the second, I develop scalable attribution methods that exploit the sparse, low-degree structure of real-world interactions, using tools from signal processing and information theory. More recently, I have been exploring principled ways to evaluate the faithfulness of model self-explanations. What's next: Researcher at Hudson River Trading's AI Labs (HAIL) Yiheng Li Yiheng Li Email: yhli@berkeley.edu Website: https://Yihengli.com Advisor(s): Masayoshi Tomizuka Research Blurb: I am working on vision world modeling, with prior experience in diffusion model's efficiency as well as in autonomous driving. What's next: Research Scientist at Waymo Zhe Fu Zhe Fu Email: zhefu@berkeley.edu Website: https://fu-zhe.com/ Advisor(s): Alexandre Bayen Research Blurb: My research focuses on physics-informed learning and control for mixed-autonomy systems, with applications in transportation. I design physics-informed neural networks to learn solutions of nonlinear partial differential equations, enabling accurate and data-efficient prediction of traffic dynamics. Building on these models, I develop both model-based and learning-based control strategies that coordinate automated vehicles to improve system-level performance. My work bridges machine learning, control, and real-world deployment, and has been validated in large-scale field experiments. More broadly, I aim to advance trustworthy, interpretable AI for decision-making in complex, real-world systems. What's next: I will be an Energy Fellow at Stanford after graduation. Also looking for Faculty, or research scientist positions in AI, control, and autonomy.
... government of the people, by the people, for the people ... — Abraham Lincoln, Gettysburg Address (1863) The cost of AI is dropping rapidly. GPT-4-class capabilities cost roughly $30 per million tokens in early 2023; today the same runs under $1 , and some providers are pushing costs below $0.10 . Across benchmarks, inference prices have fallen between 9x and 900x per year , with a median decline near 50x. Even frontier models are getting dramatically cheaper each generation, with open-source models following closely behind. And crucially, even if “Nobel-Prize-winning genius-level” intelligence isn’t here yet, the intelligence that suffices for the vast majority of knowledge work is here today, and getting cheaper by the month. At this rate, we are soon entering the era of virtually free intelligence —the kind that is more than enough for everyday knowledge work. A cartoon database character and an AI robot agent holding hands Disclosure: This post is a perspective led by Aditya G. Parameswaran —an Associate Professor of EECS and co-director of the EPIC Data Lab at UC Berkeley—together with his collaborators. It is part landscape survey and part perspective, and several of the research directions discussed below (including agentic speculation, structured memory, and synthesizing custom data systems from scratch) draw on the authors' own ongoing work. So, what does this new era of near-free intelligence mean for data systems? We believe three new challenges—and opportunities—stem from near-zero inference costs: Data Systems For Agents. Agents will soon become the dominant workload for data systems—with swarms of agents spun up in response to each end-user request. Given differences in characteristics between agents and humans—or applications acting on their behalf— how should we redesign data systems for such agentic users? Data Systems Of Agents. As agents start taking on the bulk of knowledge work, a new substrate is needed for thousands of agents to manage state over long-running tasks, coordinate and reach consensus, and deal with failures. What do data systems that reliably and efficiently run and manage agent swarms look like? Data Systems By Agents. Agents are rapidly becoming capable of synthesizing entire data systems in one go—meaning we can rebuild custom systems for each new workload. Verifying that such systems match intended behavior is a challenge. What does it take to let agents synthesize data systems we can actually trust? A database character and a robot agent holding up a triangle labeled 'of', 'for', and 'by' Data Systems For, Of, and By Agents Next, we will discuss each in more detail, followed by discussing the intertwined future of data systems and agents, especially as the three challenges intersect. Data Systems For Agents An agent querying a database doesn’t behave like a person or a BI tool. It performs what we call agentic speculation : a high-volume, heterogeneous stream of work spanning schema introspection, columnar exploration, partial and then full query formulation. With multiple agents each exploring portions of the hypothesis space, each user request could amount to 1000s of individual SQL queries. Now, users can issue ‘high-level’ data tasks, e.g., root-cause analysis—e.g., ‘why did coffee sales in Berkeley drop this year’—or exploratory cohort analysis—e.g., ‘which user segments are most likely to churn next quarter’—each involving a combinatorial space of potential joins, aggregations, and filter combinations. An agent sending many SELECT SQL queries to a database and receiving results back Data Systems Redesigned to More Effectively Support Agentic Speculation The requests from these agents have various opportunities for optimization. For instance, on a text-to-SQL benchmark with multiple agents attempting each task, only 10-20% of the sub-plans are distinct. Thus, 80-90% of sub-queries perform duplicate work. The same experiments show task success rates significantly increasing with more agentic attempts—so the redundancy is actually helpful. But from the data system perspective it’s wasted work. An agent-first data system can exploit such properties to help agents make progress faster. It can reuse results across overlapping sub-plans, drawing on ideas from decades-old literature on multi-query optimization and shared scans . Or the data system can try to satisfice , returning approximate answers that are good enough for agents to make progress, leveraging work from the AQP literature —or streaming the results of the final or intermediate operators to help agents decide if seeing the rest is necessary or helpful. Another opportunity here is to rethink the query interface entirely: instead of agents issuing a single SQL query at a time, they could instead issue a batch of queries, each with its own approximation requirements. Since enumerating an exponential search space (as in the root cause or cohort analysis examples above) isn’t a good use of agentic reasoning ability, perhaps data systems should support higher-level primitives rather than requiring agents to list each SQL query explicitly. One idea here is to draw on DBT-style Jinja macros to provide looping-based primitives for agents to interact with data systems. A swarm of AI agents working at laptops A Caffeinated Army of Agents Ready to Tirelessly Complete Your Data Tasks A final opportunity here is to stop thinking of data systems as passive executors of queries; data systems could be proactive , as they possess more grounding in data and system characteristics that agents may lack a priori—they could steer agents in different directions, provide results for related queries, and also provide performance-level feedback (e.g., instead of executing an expensive query, the system could first provide the agent a latency estimate). The reason we can do this now as opposed to the past is that an agent can accept any form of textual feedback and isn’t expecting a strict SQL query result. In fact, the data system could also prepare both materialized and virtual views for an agent in advance, provided to the agent as part of context, as this may be cheaper or more effective than having an agent author or use them. Data Systems Of Agents Previously, we focused on how agents interact with data systems. Now, we consider everything else agents need to keep working: where they live, how they remember, how they coordinate with each other, and how they deal with failures of each other. This agentic substrate is separate from the inference stack powering raw intelligence. However, the inference stack itself is being abstracted away through APIs (e.g., from OpenAI or Anthropic), or, for open-weight models, through serving frameworks that hide low-level details. So far, the agentic substrate has been managed through harnesses like Claude Code and Codex , coupled with various mechanisms to store and retrieve memory. First, on the memory front, the current wisdom is that files are all you need ; agents write to unstructured markdown (MD) files, which can then be searched using grep, or via embedding-based retrieval. In fact, many argue that the solution to continual learning is having agents consume a lot (e.g., an entire codebase, slack, company wikis, …) and then write their learnings into MD files, which are then retrieved selectively on demand. Indeed, file systems, bash scripting, and MD files are and will still be important for agents. However, at scale, when agents are doing the vast majority of knowledge work, this approach will no longer be effective. Given limited context windows, retrieving all MD file fragments that may be relevant and stuffing it into the context will break down at some point. Even if context windows continue to grow, there are latency benefits to not put all information into context — and in many cases, e.g., when knowledge work involves interacting with large databases or code bases, it will be infeasible to serialize all relevant data into context. A swarm of robot agents holding hands, each drawing state from a single large shared database platform below them Data Systems As A Substrate for Multi-Agent Swarms One could use a knowledge graph representation , but knowledge graphs suffer from the same limitations as unstructured MD-based memory due to their lack of structured search. What one needs is to be able to retrieve only memory that is pertinent to the task, across multiple attributes (or facets) of interest. For example, an agent debugging a flaky test should be able to pull only the memories tagged with the relevant module, language, framework, and failure mode—rather retrieving based on keywords or embedding similarity. A separate issue is what to actually retrieve; raw agent traces with mistakes are not very useful as they will induce agents to repeat the same mistake—instead, we want the retrieved memory to be corrective. We recently explored a related notion of structured memory , where we organize memory across various attributes, each of which could be set as * to indicate universal applicability, or set as a list of values to be matched. For a data agent, the dimensions could include the columns and tables, type of operation, and finally, open-ended natural-language corrective instructions. So, we could include memory that only applies to a given type of operation (e.g., ‘when performing date-time operations, use fiscal year as opposed to calendar year conventions’), or a given table (e.g., ‘column product_cleaned is preferred over column product when querying on product name’). One open question is defining an application-specific structured memory —or what others have called world models for memory . We believe this is akin to defining a schema for each application—and perhaps agents themselves can help us define and refine it over time. Diagram showing corrective knowledge stored with structured attributes (SQL keywords, tables, columns, data type) and retrieved by matching the features of a new agent query One Possible Way To Store and Retrieve Structured Knowledge [From Here] Structured memory will be useful also for evolutionary frameworks to effectively manage search spaces. Indeed, storing, structuring, and mining large volumes of single and multi-agent traces can help future agents become much more efficient—potentially enabling effective recursive self-improvement through structured memory-based mechanisms. Another challenge is to support concurrent edits to shared memory, and concurrent edits in general, when there are many agents performing transformations. While there have been some useful attempts at supporting multiversioning and copy-on-write semantics , it isn’t clear that such techniques will suffice when thousands of agents are attempting to edit shared state at the same time. For instance, when agents are trying various potential transactions in response to a user request, the effects of the vast majority of these transactions need to be rolled back—with only the one ‘correct’ transaction’s result persisting. Work on supporting exactly-once semantics is relevant here, as are underlying techniques based on CRDTs and operational transformation. For updates to fuzzy mechanisms such as memory, we may be able to sacrifice on consistency for perfect correctness in the interest of latency. While agents can reason about semantics to compensate or roll back their actions to eventually finalize most tasks, the primary challenge lies in the degree to which they step on each other’s toes during the process. An important failure mode to be avoided is a form of “livelock,” where incessant compensating actions prevent any meaningful progress. Beyond shared state, other concerns emerge when trying to support an army of agents, including what to do when agents fail, how agents should communicate with each other (directly or through intermediate shared state), and how we should deal with straggler agents. There have been some developments in supporting durable multi-agent execution, such as Temporal , but it remains to be seen if such solutions will apply at scale across thousands of agents. On the topic of communication, we need mechanisms to enable agents to negotiate with each other. Imagine four developer agents attempting to reach consensus on a shared schema, with distinct but overlapping objectives. In a human setting, this would involve iterative discussion and compromise; for agentic swarms, we must define the mechanisms that allow them to converge on a design that reflects the underlying goals of their respective principals. Or if agents are all requiring access to a limited resource, again communication will be necessary. It remains to be seen if this is best done via centralized coordination, or if a decentralized approach is necessary. Data Systems By Agents Finally, if intelligence is effectively free, then we can employ this intelligence to synthesize new data systems from scratch. Indeed, in many settings, general-purpose data systems may be overkill, as they have to support every schema, query, and hardware target. Given a workload, recent work, including Bespoke OLAP and GenDB , has shown that one can use an agentic pipeline to synthesize a complete, workload-specific analytical engine—in minutes to a few hours, at a cost of a few dollars. The engines are disposable: when the workload shifts, one can simply regenerate them. Analogously, our work has shown that one can synthesize custom key-value stores from scratch, targeted to the workload. In fact, modern IDEs, such as Kiro , elevate specifications for systems development to be a first-class citizen. A robot agent with a hammer and chisel carving a database character out of a block of stone Agents Can Synthesize Custom Data Systems From Scratch The main issue, however, is that specifications are typically imperfect, and don’t cover all corner cases. Present-day agents will exploit the missing specifications to reward-hack their way to a high performance metric. In our custom key-value store work, we found that one way to alleviate this is to have auxiliary verification agents trying to generate test cases that catch the exploitation of corner cases, essentially expanding the specification. Yet another approach is to both generate a system and a proof for its correctness together, for which we have found some early success , but more needs to be done to solidify the approach. Further, it remains to be seen what is the best way to solicit human-written specifications for a system—can this be done in an iterative, human-in-the-loop manner, as opposed to a one-shot, incomplete one. Indeed, human-written specifications are incomplete even for manually authored software, so one would expect that future agents that are more aligned will increasingly exercise better judgement when making design decisions. Pipeline diagram where a system builder provides a specification, planner and coder agents generate code, the code is evaluated for correctness and performance, and critic and auditor agents provide feedback and catch reward hacking One Possible Data System Synthesis Pipeline [From Here] Other questions here involve testing whether starting from a mature system (e.g., Postgres) and removing components/functionality can lead to higher performance or more user trust. Separately, is there an opportunity to make the design composable, comprising various verified components that are mixed and matched given a workload? For example, perhaps the workload hasn’t changed enough for the storage layer to be updated, but perhaps the query optimizer requires changes. A perhaps more viable proposition involves employing agents coupled with proof systems to target critical parts of the code associated with formal proofs, rather than doing so for the entire system. A final opportunity here is to move away from the traditional data systems stack with clearly-defined interfaces (e.g., parser, query optimizer, storage manager, …) — that were each largely the prerogative of a single human team to manage. Instead, agents can find new ways to “blend” these components together, perhaps identifying new optimization opportunities as a result. Agents can also fill in missing gaps in functionality to make existing systems much more feature-complete, or reach feature-parity with other competing systems—or analogously, continuously refining open-source systems in response to feature requests or issues (perhaps filed by other agents!) Doing so in a way that prioritizes correctness, long-term maintenance, and human interpretability will be a challenge. Looking Further Ahead In the era of near-free intelligence, data systems matter more than ever. As agents take on the bulk of knowledge work, the workload for data systems will change, the substrate they need to run on will have to be built, and increasingly, they will participate in designing data systems themselves. Each of these shifts opens up a new, exciting research agenda. A half-database, half-robot character next to a yin-yang symbol formed by a database and a robot agent Co-Evolution of Data Systems and Agents Looking further out, the boundaries between agents and data systems will likely start to blur. For instance, agents may design the data systems they themselves run on, defining both the interfaces as well as the system components underneath. Both the interfaces and internals can be evolved over time by agents in a form of recursive self-improvement. There is also an opportunity to rethink data systems as a holistic source of truth for the entirety of relevant state: including raw data, memory, and coordination state, further erasing the distinctions between the data that is being queried by agents and data generated as a result of agentic activity. Finally, data systems may themselves incorporate agentic components, fundamentally evolving from passive computation engines into intelligent, proactive, self-optimizing architectures. It is hard to predict what the future may hold. We’re in for a wild ride! Acknowledgments The perspective and ongoing work described in this post are the product of joint research and many discussions with wonderful collaborators at the EPIC Data Lab , Data Systems & Foundations group, and the broader Berkeley AI-Systems community. Thank you all! BibTex for this post: @misc { intelligence-is-free-blog , title = {Intelligence is Free, Now What? Data Systems for, of, and by Agents} , author = {Aditya G. Parameswaran and Shubham Agarwal and Kerem Akillioglu and Shreya Shankar and Sepanta Zeighami and Rishabh Iyer and Matei Zaharia and Alvin Cheung and Natacha Crooks and Joseph Gonzalez and Joseph Hellerstein and Ion Stoica} , howpublished = {\url{https://bair.berkeley.edu/blog/2026/07/07/intelligence-is-free-now-what/}} , year = {2026} }
ABBEL overview Overview of ABBEL compared to traditional recursive summarization. Beliefs replace the full interaction history as the agent’s working context, and belief grading improves performance by supervising the contents of each belief state.. As task horizons grow, LLM contexts can’t scale forever. Self-summarization enables concise, interpretable contexts, but at a significant performance cost, especially for human assistance domains where high quality data is scarce, e.g., collaborative code generation. We address this with ABBEL : a framework that isolates and supervises the information content of summaries in the form of natural-language belief states. Motivation: the cost of recursive summarization For language models to effectively assist with increasingly complex tasks such as software development, they must be able to interact with us over hundreds or even thousands of steps. For such long tasks, it is impractical to keep the history of the entire interaction in context. The heuristic approach used so far has been summary generation, sometimes called context compaction. For example, Cursor’s latest model composer 2.5 uses compaction during training for improved performance ( Cassano et al., 2026 ). Alongside composer, Grandcode ( DeepReinforce et al., 2026 ), the first system to consistently beat all human competitors in online coding competitions, despite using one of the newest efficient attention models (Qwen 3.5-397B), 1 still found it necessary to employ context summarization. But compaction has a problem. Despite seemingly low performance gaps in benchmarks, model servers like Cursor continue to recommend that users avoid compaction with their coding assistants in the middle of a task ( Heule et al., 2026 ). To understand why, see below the performance over RL fine-tuning of a Context summary model compared to full context models in Combination Lock, a Wordle-like game that allows up to 16 guesses. 2 Though both model types improve over the course of training, the summary model never closes the gap. Average attempts to guess target word with context summary vs full context policies over training Fig. 1: Average attempts to guess the target word on Combination Lock over RL fine-tuning (lower is better). Context-summary policies improve with training but do not close the gap to full-context policies. Making models self-summarize while completing a task increases the complexity of the learning problem. While this could typically be addressed by training with more data, the performance degradation observed in real world interactive settings likely arises from the difficulty we have in creating and using human simulators effectively to generate high quality training environments ( Lin et al., 2025 , Tomlin et al., 2025 ). Thus, the better you can learn to summarize on the limited and messy multiturn interaction trajectories you can collect, the better off your model will be for downstream users. ABBEL: acting through belief bottlenecks ABBEL belief grading diagram Fig. 2: Autoencoder-inspired belief grading. The model encodes prior belief, action and observation (b t , a t , o t ) into posterior belief b t+1 and is rewarded for how well select information from the history can be reconstructed from that belief. To address poor learning efficiency, we isolate the summary generation task. Drawing inspiration from recursive Bayesian estimation, we formulate summaries as belief states, which we periodically prompt the model to update based on new information. 3 ABBEL overview animation frame ‹ Prev Pause Next › 1 / 16 Fig. 3: ABBEL rollout. Belief updates from the latest observation alternate with action selection conditioned only on the current posterior belief. Belief grading We then extract and supervise the contents of the belief states (Fig. 2, Belief Grading). Belief grading can be thought of as adding an auxiliary RL task, using heuristics designed to capture what makes a good belief as the reward. An example heuristic for coding could be shorter is better, but closer to being able to reconstruct the git diff is also better, so balancing these would yield a good belief. In domains where good heuristics are hard to define, we propose a general autoencoding-inspired grading function, which treats the current language model π θ as both encoder and decoder of information from the history, and the belief states as the codes. We grade each belief b t+1 by how well it can be used by the current model π θ to reconstruct the most recent observation o t : Eq. 1: Reconstruction grading objective Eq. 1: Reconstruction grading objective. Here b t+1 is the updated belief, o t the latest observation, a t the action just taken, b t the prior belief, p I the task prompt, and π θ the current model. Higher grades reward beliefs that retain information needed to decode the latest observation. What do we gain by grading beliefs? Collaborative coding on CollabBench We demonstrate the utility of belief grading in our motivating domain of human-driven assistive coding, with the CollabBench environment from Sweet-RL ( Zhou et al., 2025 ). CollabBench collaborative coding environment Fig. 4: CollabBench collaborative coding environment. The agent asks clarifying questions, then submits a function scored against hidden unit tests. We see that with the general reconstruction-based belief grading function we reduce the performance gap from full context models by about 50%, and train in 50% fewer steps compared to training models to summarize without belief grading (no BG). After training, ABBEL still uses significantly less memory than the full context setting, as measured by the peak context token length (Peak Tokens). Model Test Pass Rate ↑ Success Rate ↑ Peak Tokens × 10² ↓ Training Steps ↓ Full Context 0.52±0.02 0.39±0.02 14.08±0.55 100 ABBEL (no BG) 0.46±0.02 0.31±0.02 4.20±0.37 100 ABBEL-rec-BG 0.48±0.01 0.36±0.01 6.01±0.33 50 Fig. 5: CollabBench results. With reconstruction belief grading, ABBEL-rec-BG recovers about half the gap to full context while using fewer peak tokens, and trains in 50 steps instead of 100. Combination Lock Additionally, in CombinationLock, we demonstrate that ABBEL with a belief grader which leverages domain knowledge (by computing useful statistics over the history and checking that they can be reconstructed from the belief state), enables even higher learning efficiency than full context (FULL CTX) models. Average Attempts to guess target word ABBEL Fig. 6: Average attempts to guess the target word on Combination Lock (lower is better). With domain-knowledge belief grading, ABBEL approaches or exceeds FULL CTX in this setting; without belief grading, learning is slower. Multi-objective question answering In a third environment, multi-objective question answering (from MEM1 Zhang et al., 2025 , a recent work which performed end-to-end optimization in a modified version of typical recursive summarization), we demonstrate the utility of isolating belief states from reasoning, by showing that a Peak Belief length Penalty (more details in paper) significantly reduces memory usage with minimal performance degradation, unlike is commonly observed when penalizing reasoning lengths ( Arora et al., 2025 ). Performance and memory usage in multi QA Fig. 7: Exact-match score and peak memory versus number of objectives in multi-objective QA. ABBEL with a peak belief penalty (PBP) maintains comparable performance while using less memory than MEM1 and ABBEL without PBP in this evaluation. Related work Alternative solutions to managing long contexts involve different tradeoffs, and are worth considering depending on the requirements of a deployed system. Context compression methods generate dense representations which, while computationally efficient, sacrifice human-understandability ( Kontonis et al., 2026 , Eyuboglu et al., 2025 , Gupta et al., 2025 , Chevalier et al., 2023 , Deng et al., 2025 , Deng et al., 2025 , Bulatov et al., 2022 ). Hand-designed summarization prompts ( Wang et al., 2025 , Örwall et al., 2025 , Starace et al., 2025 ) and pruning strategies ( Jiang et al., 2024 ) specific to target environments require expert human knowledge and don’t allow an agent to learn what to remember as part of its decision-making strategy. Methods that process long contexts into an external memory store ( Packer et al., 2023 , Xu et al., 2025 ) for the agents or subagents to query ( Zhang et al., 2025 ) are complementary, as they may benefit from better next context creation through summarization training. We would like to point out some exciting works in the space of general recursive summarization focused on math ( Wu et al., 2026 ), reasoning with belief generation ( Zhou et al., 2025 ), competitive coding with a distilled summarization module using similar autoencoding objectives to our general belief grader ( DeepReinforce et al., 2026 ), and adding continuous features to summaries ( Kontonis et al., 2026 ). What’s next for better memory? Many more possibilities are enabled through using explicit belief states as information bottlenecks for multi-step interaction. You could reward actions based on their effect on the belief state to guide exploration, transmit the explicit belief states for better communication between agents, or even improve user controllability by directly modifying the memories on which the agents’ decisions are based. Some forms of information, e.g., what a person looks like, are not represented well by text alone. A continuously learning system will also have to capture such information. Additionally, if we want a system to learn to communicate in a brand new language or to play a brand new game better than any person in the world, the skills accumulated over the lifetime of conversations or games must be stored in a very compressed form, essentially taking on the role of the weights of the model itself. More powerful systems will likely utilize a combination of multiple forms of memory, where the contents of the context may correspond to working memory while other approaches are used for short and long-term memory. How to instantiate these other forms of memory, for instance via test-time training, adapter memories, continuous context memories, or some combination thereof, presents an exciting challenge. Acknowledgements Acknowledgements: We would like to thank Alane Suhr and Kartik Goyal for advising this research as well as Ethan Mendes , David He , Jitesh Jain , and Nicholas Tomlin for comments on early drafts of this post. We would like to thank the MEM1 authors for their email correspondence and for sharing private reviewer feedback which we found particularly insightful. Citation If abbel was inspiring for your future work, please cite us with this! And here is some advice for doing similar research! @misc { lidayan2026abbellearningnaturallanguagebelief , title = {ABBEL: Learning Natural-Language Belief States for Memory-Efficient Interaction} , author = {Aly Lidayan and Jakob Bjorner and Satvik Golechha and Kartik Goyal and Alane Suhr} , year = {2026} , eprint = {2512.20111} , archivePrefix = {arXiv} , primaryClass = {cs.CL} , url = {https://arxiv.org/abs/2512.20111} , } With newer models the number of tokens till 50% compute spend is on attention gets much larger than 25K. Interleaving linear attention alternatives with full attention as is done with gpt-oss and DeepSeekv4, results in massive flops reductions for the attention computation. For example with DeepSeekv4-Pro (1.6T A49B) it requires nearly 450 thousand tokens to reach the 50% tradeoff point. Grandcode uses Qwen-3.5-397B-A17B a model which hits 50% FLOPs for attention at ~150 thousand tokens. ↩ This setting is technically solvable with much more computationally effective tools, but serves as a flexible test bed to study properties of recursive summarization. Bertsimas et al., 2022 , showed that an exact solution for the wordle game instantiated with the original vocabulary of the javascript game can be found with dynamic programming, but evidently the general formulation of wordle as a guessing game on K letters with L attempts and some dictionary of valid words and correct words D is NP hard to determine the minimal number of moves required. ↩ In practice there is an O(N/K) overhead cost for summary. N is the total number of actions. K is the number of actions till summarization is triggered. This is necessarily true for any summary approach. For ease of illustration this gif uses K = 1. In our experiments, to put more emphasis on summarization weaknesses we also use K=1. In practice overhead is small as K can be chosen to be near the efficient hardware limit. ↩
Kernel knowledge transfer from CUDA to MLX Figure 1: CUDA-to-MLX optimization translation map. CUDA optimization knowledge can be translated into architecture-native MLX strategies rather than copied instruction-for-instruction. We face a new epoch in computing. Hardware is changing rapidly — not just faster GPUs, but a growing range of chips from different vendors, each with its own architecture and often tailored to specific AI workloads. Software is changing just as fast, and AI coding tools now generate in minutes what took months of effort a few years ago. With so much of computing now centered on AI, GPU kernels are a crucial component of its success. These are the low-level programs that run inside the GPU, and writing efficient ones is far from obvious — it takes years of expertise to get right. Transferring a kernel from one vendor’s hardware to another is harder still, and often means rediscovering the same optimizations from scratch. The CUDA ecosystem, for example, has accumulated decades of hard-won kernel expertise: hand-tuned implementations of attention, state space models, and other critical operations representing thousands of engineering hours. Newer hardware ecosystems (Apple Silicon, custom AI accelerators, and others) are growing fast but lack this depth. In this work we ask whether that expertise can be transferred automatically. We built on K-Search , an evolutionary kernel search framework introduced by Cao et al. at Berkeley Sky Lab that uses AI to optimize GPU kernels, and extended it with a backend for MLX — Apple’s machine-learning framework for its own Apple Silicon chips. We developed a novel structured CUDA-to-MLX translation layer that lets K-Search take existing CUDA kernels as a knowledge base and adapt them into high-quality GPU kernels for Apple Silicon, rather than rebuilding from scratch. We show that our approach reaches near-expert level performance on Apple Silicon with 0.97x speedup compared to the native MLX Attention kernel, and up to a 20x prefill speedup over the community mlx-lm implementation on the Mamba SSM kernel; we report the numbers, and how much of the gain comes from the translation layer, in the sections below. Although we focus on MLX kernels for Apple Silicon, the method is not specific to MLX and applies to any ecosystem where CUDA expertise is transferable. Why MLX? Apple’s MLX framework has seen remarkable adoption since late 2023. With Apple Silicon in hundreds of millions of MacBooks and Mac Studios, MLX enables local AI inference without cloud costs. The unified memory architecture makes it especially attractive for mid-sized models (7B–70B parameters on M series chips). Yet beneath this momentum lies a significant gap: many performance-critical kernels that the NVIDIA ecosystem takes for granted: paged attention, optimized SSM scan kernels, fused MoE routing are either absent or naive without hardware-specific tuning. MLX runs models correctly but often leaves significant performance on the table. This gap is what motivates the rest of this post. What is K-Search? K-Search is an evolutionary kernel optimization framework originally developed by our first author Shiyi Cao at UC Berkeley Sky Lab. Given a naive kernel and a hardware specification, it runs an iterative optimization loop: an LLM reasons about which optimizations to try next, a code-writing model generates candidate kernels, and those candidates are compiled and benchmarked on real hardware. Measurements feed back into the search, which keeps refining, pursuing promising directions and dropping dead ends until performance converges. Pseudocode for K-Search via co-evolving world models Algorithm 1: K-Search via co-evolving world models. The search alternates between selecting the most promising action, instantiating and evaluating code until improvement stagnates, and evolving the world model through insert, update, and prune operations. Adapted from Cao et al. (2026) . Search is grounded by a Spec: a domain-specific document encoding hardware rules, optimization patterns, and mathematical constraints which keeps generated code from hallucinating invalid primitives and ensures candidates will actually compile and run efficiently. In our runs, a single model (Gemini 3.5 Pro Preview) plays both roles: it maintains the reasoning state and writes the kernels. The reasoning half is prompted as a “GPU kernel performance engineer” and asked to work through a fixed analysis before proposing anything: classify the kernel (reduction, scan, attention/softmax, …), rewrite the reference computation in canonical form, map out data layout and access patterns, and hypothesize the likely bottleneck (bandwidth, latency, compute, or synchronization) in each runtime regime. Only then does it emit candidate optimizations, each as a single change implementable in one iteration. We call the persistent reasoning state a world model . Rather than a flat list of things to try, it is a decision (prefix) tree: each root→leaf path composes a full optimization plan, and sibling branches are competing alternatives. Every node is scored — an overall_rating in [0, 10], a confidence in [0, 1], and per-node impacts on memory bandwidth, register pressure, and compute/hardware fit — so the search can rank partial plans and expand the most promising ones. The tree persists and grows across rounds: refining an idea adds a child node rather than overwriting its parent, and if the best score fails to improve for a few rounds (a stagnation window) the search backs off to explore an alternative branch. A single node, as it appears mid-run on the attention kernel, looks like this: { "action" : "Replace the threadgroup-memory softmax reduction with a register-only reduction: each SIMD group owns 8 query rows and reduces across lanes with simd_shuffle_xor, removing a threadgroup_barrier." , "difficulty_1_to_5" : 4 , "impacts" : { "memory_bandwidth" : 8 , "register_pressure" : 4 , // risk: spill if Br > 8 "compute_hw_fit" : 9 // SIMD width 32 ; keep tile 8 x 8 }, "overall_rating_0_to_10" : 8 , "confidence_0_to_1" : 0.7 } Listing 1: Example K-Search world-model node. Each candidate optimization records a concrete action, estimated hardware impacts, an overall priority rating, and the model's confidence. Overview of the K-Search loop Figure 2: Overview of K-Search. The framework operates on a Search State $S_t$ structured as a search tree. The tree consists of Closed nodes (blue, visited states with attached program like $x_{12}$) and a Frontier of Open nodes (orange, pending hypotheses like $u_{13}$). The workflow iterates through three phases: (1) Action Selection , where the most promising action node is retrieved from the frontier based on world model estimated priority score $V$; (2) Local Refinement , where a stochastic policy $\pi_{\mathrm{code}}$ samples concrete implementations until stagnation; and (3) World Model Update , where the LLM reasons over the trajectory to update the search tree via Insert (adding new actions), Update (adjusting $V$, e.g., $u_{11}$ dropping from 0.9 to 0.6), and Prune (removing less promising nodes like $u_{10}$). The original K-Search paper evaluated this search strategy on CUDA kernels from FlashInfer. Across GQA decode, MLA decode, MLA prefill, and MoE, K-Search improved more consistently than OpenEvolve and ShinkaEvolve over the same 120-iteration budget. These results establish the search framework we build on here; the remainder of this post asks whether its optimization knowledge can transfer beyond CUDA. K-Search benchmark results compared with OpenEvolve and ShinkaEvolve Figure 3: Main results from the original K-Search paper. Across three runs, K-Search achieves stronger best-so-far search scores, per-workload kernel performance, and speedup distributions than OpenEvolve and ShinkaEvolve on four FlashInfer CUDA kernels. Reproduced exactly from Cao et al. (2026) . Building an MLX backend To bring K-Search to Apple Silicon, we first built a native MLX backend. We implemented a full MLX-specific task adapter for K-Search, including: An MLX task backend in k_search/tasks/ handling kernel compilation and execution on Apple Silicon via MLX’s Metal/C++ APIs. Updated kernel generator prompts for writing and modifying Metal/MLX kernels. MLX-specific benchmarking integration using mlx.core measurement utilities. Translating CUDA expertise to MLX However, the more interesting challenge was not simply running K-Search on MLX. The key insight is that expert CUDA kernels encode decades of optimization knowledge that is transferable to Apple GPU if you can bridge the conceptual gap. Simply handing an LLM a CUDA kernel and asking it to port it is not enough: without deep hardware context, it produces code that is syntactically valid but architecturally wrong (wrong tile sizes, invalid primitives, mismatched memory assumptions). Our translation layer consists of: Concept mapping tables: A structured glossary of CUDA primitives and their MLX/Metal equivalents with hard constraints. For example: __shared__ maps to Metal threadgroup memory but with a hard 32 KB limit (vs. NVIDIA’s 48 KB) warp_reduce maps to MMA (preferred) __syncthreads() becomes threadgroup_barrier(mem_flags::mem_tg) H100’s ~3.35 TB/s HBM3 maps to M3 Max’s ~400 GB/s unified DRAM a bandwidth difference that reshapes which optimizations are worth pursuing. MLX-specific hints and patterns: Concrete code-level patterns for operations with no direct CUDA equivalent, such as register-based row reductions using simd_shuffle_xor in an 8×8 MMA tile layout, or the “exp2 trick” (replacing $exp(x)$ with $exp_2(x \log_2 e)$) for faster softmax on Apple’s fast $exp_2$ hardware instruction. Reusable assertions: Expert kernel behaviors reframed as properties the evolutionary search must preserve, rather than code to copy. Matching expert kernel performance: the Attention kernel We evaluate three configurations of an MLX attention kernel for Apple Silicon: (1) a naive baseline, (2) pure evolution with no additional provided context, and (3) a full context translation layer, which supplies the optimizer with architecture-specific implementation knowledge extracted from high-performance kernels (e.g., FlashAttention-2), letting the evolutionary search reason about implementation strategies rather than starting from a naive kernel. Together, these three configurations let us isolate the exact impact of the translation layer. Performance scaling of the Attention Kernel through stacked optimizations Figure 4: Performance scaling of the Attention Kernel through stacked optimizations. The "Full Context" configuration successfully discovers and implements advanced strategies like double buffering and loop unrolling, achieving near-expert performance. The jump from 0.26× to 0.97× the speed of Apple’s state-of-the-art attention kernel — illustrates how much the translation layer matters. With full context, the evolved kernel independently discovers the key optimizations in FlashAttention 2: threadgroup memory tiling, online softmax, K-transposition for memory access, and the exp2 trick. The last of these replaces every softmax exponential with a base-2 exponential, \[e^x = 2^{x \log_2 e},\] which is exact and lets the kernel use Apple’s fast fast::exp2() hardware instruction directly instead of paying for a base conversion at runtime. A 20× faster prefill: the Mamba SSM kernel To evaluate whether K-Search generalizes beyond attention kernels, we applied it to the state-space model (SSM) kernel used by Mamba. Unlike attention, the computational bottleneck is a recurrent state update rather than a softmax, providing a substantially different optimization challenge. We compare the evolved implementation against the community MLX implementation (mlx-lm) and the PyTorch reference implementation (mamba.py) on an M1 Max. Evaluated on mamba-370m f16, M1 Max 64GB: Metric mlx-mamba (ours) mlx-lm (community) mamba.py Decode 152 tok/s 116 tok/s 40 tok/s Prefill L=512 5,751 tok/s 329 tok/s 1,089 tok/s Prefill L=1024 6,010 tok/s 327 tok/s 1,127 tok/s Prefill L=2048 6,612 tok/s 326 tok/s 1,092 tok/s Prefill L=4096 6,743 tok/s 339 tok/s 1,042 tok/s Table 1: Prefill and decode throughput on mamba-370m (f16, M1 Max 64GB). mlx-mamba (ours) reaches ~20× higher prefill throughput than the community mlx-lm baseline, while decode remains comparable. The ~20× prefill speedup over mlx-lm comes down to one difference: mlx-lm does not implement a parallel scan for the SSM. The state recurrence \[h_t = \bar{a}_t h_{t-1} + \bar{b}_t\] looks inherently sequential, but each step can be written as a pair $(\bar{a}_t, \bar{b}_t)$ under the associative combine \[(a_2, b_2) \circ (a_1, b_1) = \left(a_2 a_1,\ a_2 b_1 + b_2\right),\] which reproduces the recurrence exactly. Because the operator is associative, the whole sequence can be evaluated with a parallel (prefix) scan in $O(\log N)$ dependent steps instead of $O(N)$. mlx-lm skips this and processes tokens one at a time, leaving most of Apple Silicon’s compute idle; our evolved Metal kernel applies the scan and makes much fuller use of GPU throughput. The gain shows up in prefill, where the full sequence is available to scan in parallel, and not in single-token decode, where there is only one new token per step and no scan to parallelize — which is why the decode row is roughly flat while prefill is ~20×. mamba.py is slow on both prefill and decode because it is a PyTorch reference implementation that falls back to CPU or MPS on Apple Silicon, forgoing the hardware-specific optimizations that MLX’s Metal backend makes possible. What’s next? On the two kernels we studied, AI-driven evolutionary kernel search grounded in structured cross-platform translation knowledge reached near-expert performance on Apple Silicon without a team of GPU experts starting from scratch. We do not yet know how far this generalizes, but the result is encouraging. For us the main takeaway is that the bottleneck was not the LLM’s ability to write Metal code, but the quality of the context and constraints we gave it. Our CUDA translation layer converts existing NVIDIA kernel expertise into actionable guidance for Apple Silicon, and lets K-Search’s evolutionary search do the rest. We are actively extending this work in several directions: supporting new architectures, with current efforts focused on developing new kernels for the IBM Spyre AIU and broader hardware targets; adding more kernels such as paged attention and fused MoE routing; and improving integration with the K-Search evolution loop to make translation context even more automatic. Acknowledgements This work was carried out by IBM Research and builds on K-Search from the UC Berkeley Sky Lab ( Cao et al., 2026 ). We welcome collaboration and feedback from the MLX and broader AI systems communities. If you are working on kernel optimization for non-CUDA hardware, we would love to hear from you. Citation @article { cao2026k , title = {K-Search: LLM Kernel Generation via Co-Evolving Intrinsic World Model} , author = {Cao, Shiyi and Mao, Ziming and Gonzalez, Joseph E and Stoica, Ion} , journal = {arXiv preprint arXiv:2602.19128} , year = {2026} } Appendix: Try it yourself The MLX backend is built on top of the open-source K-Search repo, so the results here can be reproduced directly. The steps are: 1. Clone and install git clone https://github.com/caoshiyi/K-Search.git cd K-Search uv pip install openai wandb uv pip install git+https://github.com/caoshiyi/flashinfer-bench-ksearch.git 2. Set your credentials Open the relevant script under scripts/ and set three variables at the top: KSEARCH_ROOT = /path/to/K-Search API_KEY = your-llm-api-key 3. Run kernel search # Optimize Flash Attention on Apple Silicon (world-model mode) bash scripts/mac_flash_attention_wm.sh # Or a Mamba SSM kernel, e.g. the selective scan bash scripts/mamba_selective_scan_fwd_wm.sh Full CLI reference and documentation are in the README.
GPT-6.1 Sol is now generally available on Amazon Bedrock, bringing stronger reasoning to coding, computer use, and professional workloads that run frequently.
Run Claude Code, Codex, AGY, and Grok Build side by side. See who is working, who needs you, and what every account has left. Comments URL: https://news.ycombinator.com/item?id=49942434 Points: 5 # Comments: 0
Hi there :-) New on HN, first time posting. Past year, around December, I started experimenting with making ChatGPT and Claude generate source code in LDraw language. This LDraw is literally an "assembly" language, a low-level programming language that describes how to assemble LEGO pieces together into models, one placement instruction at a time. When executed by specific tools, like e.g. LDView, LeoCAD, Studio... these instructions become LEGO CAD models, that can be interacted with, modified, etc. Or, in other words: one LDraw source file in .mpd or .ldr format is equivalent to one LEGO CAD model. So, the idea I had was: if I manage for maybe ChatGPT or Claude to generate high-quality LDraw source files... then, they would actually be generating high-quality LEGO CAD models, right? Then, after months of iterations and trying one thing after the other... it worked!!! Long story short: using GPT-6 Astra and Opus 5.5, I've managed to create a python toolset, instructions, and docs for agents in general. Now, these can be used by them to generate LDraw models. I've packed it all as a dockerized web app for others to try and experiment, with several providers (and agents) to choose from: OpenAI, Claude and OpenRouter. I'd really appreciate feedback and comments, let's see where this goes =) Comments URL: https://news.ycombinator.com/item?id=49937916 Points: 9 # Comments: 2
I just realized that codex now exposes local server endpoints for all plugins without extra auth. So you can just use them from any model or harness. Im soooo excited!!! One auth for everything! For Pi: pi install npm:@wileai/pi-codex-connector For any harness: npx -y @wileai/codex-connectors-mcp@latest Comments URL: https://news.ycombinator.com/item?id=49929381 Points: 5 # Comments: 0
I found that I'm not using even 1/3 of vim/vscode features anymore. That's wht I'm building rhun - a small code editor for Linux, Windows and Apple silicon Macs. It obviously has Vim mode, a terminal, Git diffs and a panel for Claude Code or Codex sessions. The editor and pixel renderer share an x86-64 assembly core. For Apple silicon, a build-time translator converts that core to AArch64, with separate platform adapters around it. The latest release can draft commit messages using a local Ollama model or an existing Claude Code or Codex subscription. It's a solo project, MIT licensed and still early. Comments URL: https://news.ycombinator.com/item?id=49926726 Points: 7 # Comments: 2
As it opens a new location, the social club prepares grant applications in 2 hours instead of 3 days and liquor-license materials in 3 hours instead of 4 days.
Deep research tasks require agents to navigate large knowledge spaces, synthesize evidence across many sources, and adapt their plans as findings emerge. Directed acyclic graph (DAG)-based multi-agent systems suit this setting because they support parallel execution and isolate each sub-task within a focused dependency context. Yet existing DAG-based agents instantiate a task-level plan before execution and repair the graph only after failures or missing evidence are observed. This Plan-then-Patch strategy is brittle for deep research: the system commits most strongly when its evidence is weakest, and later revisions waste computation on branches that should not have been planned. We propose DAGent, a DAG-based multi-agent framework with Evaluate-then-Grow incremental planning: an Orchestrator grows the task graph one batch at a time, conditioning each expansion on confidence and uncertainty signals from completed nodes. A hierarchical context layer propagates compact QueryDocs by default while preserving full execution traces for on-demand recall. The recorded DAG topology admits structural RL signals that outcome-only recipes cannot define; DAGRPO, a GRPO adaptation, injects topology-conditioned credit on Executor rollouts and a structural compliance regularization on Orchestrator plans. Across BrowseComp-Plus, GAIA, and xbench-DeepSearch, DAGent surpasses the strongest open-source baseline by 5.3 / 5.8 / 2.0 points at the Qwen3-235B-A22B scale, and the lead replicates across four open-source backbones and extends to GPT-5 at 327K context. At the Qwen3-8B scale, DAGRPO improves over a same-budget outcome-only GRPO baseline by 3.0 average Pass@1 points. A same-architecture comparison shows that evidence-conditioned planning reaches higher accuracy at lower per-task token, tool-call, and step footprints than its Plan-then-Patch counterpart. Code: https://github.com/hanwenliu6825/DAGent
As agents take on longer and more complex problems, controlling the execution becomes a task in its own right. Each step in the run brings new control choices, like which partial work to build on, whether to start fresh, or when to stop. We introduce agentic meta-reasoning, an inference-time harness that makes these choices an explicit and structured reasoning process. Workers carry out the task-level computation, while a controller consolidates what the run has established, explores next options, assesses what each option is worth under the remaining budget, and dispatches the chosen work with context drawn from persistent memory. Between decisions the controller carries only a compact account of the run rather than replaying its full history. Our baselines span production coding agents and research harnesses, together with a Direct Control Agent using the same workers and compute budget allowance. On ProgramBench, which tests long-horizon agentic capability through program reconstruction, meta-reasoning achieves 71.5% with GPT-5.5 against 58.0% for Codex; with Opus 4.8 it achieves 67.2% against 65.5% for Claude Code. On the other benchmarks, spanning abstract reasoning, multi-domain long-horizon reasoning, and proof generation, it gains between 3.6 and 4.2 points over direct control, averaged across three frontier models. It keeps improving over the tested budget ranges where direct control plateaus, though its overhead can hurt at small budgets. Artifact-graph analysis reveals more reuse of earlier work, higher coverage of correct solutions in most settings, and nonuniform gains in final selection. These results indicate that spending computation on structured control becomes more important as agents scale to longer runs.
Fine-tuning-as-a-service lets users adapt a safety-aligned language model to their own data, but it also creates a harmful fine-tuning attack surface: a small amount of harmful data mixed into an otherwise benign fine-tuning set can degrade the model's alignment. Two recent alignment-stage defenses address this problem at different levels of the model. Vaccine improves the robustness of hidden embeddings to the representation shifts induced by harmful fine-tuning, whereas Booster simulates harmful weight updates and attenuates their effect during alignment. We investigate whether these mechanisms are complementary and propose VaccineBooster, a single alignment procedure that combines embedding perturbation and weight-level gradient attenuation within each training step. On Llama-2-7B aligned with BeaverTails and then attacked through poisoned fine-tuning, VaccineBooster achieves the lowest OpenAI moderation score among the compared defenses, 0.315, while a Booster-Only variant retains the highest post-attack refusal rate, 50%. Together with ablations over the embedding-perturbation and gradient-attenuation strengths, these results indicate a trade-off: embedding perturbation primarily reduces flagged harmful content, whereas gradient attenuation primarily preserves explicit refusal behavior. Because our evaluation uses ten prompts and a single unseeded run per configuration, we report this trade-off as an observed pattern rather than a statistically resolved effect. These results provide practical guidance for prioritizing content safety or refusal retention when aligned models are exposed to untrusted fine-tuning.
Inspired by minGPT for home experiments. Requires CUDA 13.4; build script is Windows only. Comments URL: https://news.ycombinator.com/item?id=49898931 Points: 17 # Comments: 1
Large Language Models can perform multi-step reasoning and improve task performance through different forms of intermediate computation, from token-based traces to computation carried out in latent space. However, a question remains open: do these different forms of thinking rely on the same underlying mechanism? To address this, we train and compare five variants of the same GPTNeoX backbone from scratch on an extended multi-hop reasoning task (ProsQA-Ext): a vanilla model, a Chain-of-Thought (CoT) model, a Pause Token model, and two latent-reasoning models that are optimized end-to-end without intermediate reasoning traces. We find that, strong in-distribution (ID) performance does not guarantee depth generalization. Vanilla, CoT, and Pause Token models solve ID problems well, but rely largely on local graph features and generalize poorly to out-of-distribution (OOD) problems with longer hops. In contrast, latent variants generalize better and show internal dynamics consistent with forward reachability propagation on the graph. Causal interventions and circuit analysis localize this computation to a sparse recurrent search circuit in the bottleneck latent model: an attention head retrieves graph relations, an MLP and the residual stream update the reachability state across recurrent steps, while multiple attention heads together then do the candidate matching. Together, these results show that different thinking mechanisms can learn distinct computational solutions, even at similar ID performance. In this setting, latent recurrence supports a reusable forward-search algorithm that generalizes beyond the training depth.
Mechanistic interpretability (MI) aims to explain a model's behaviour through analyzing its internal computations; circuit-based explanations aim to isolate these computations with compact subnetworks validated by ablating the rest of the model. We show that circuits validated this way may fail to recover the underlying mechanism of the model's behaviour by closely reproducing its successful decisions while failing to account for most of its errors. Such explanations should account for the model's particular errors as well as its successes. We evaluate this requirement by measuring exact answer agreement separately on model successes and failures, across circuit sizes and ablation settings, on IOI, Docstring, and six model-task settings from the Mechanistic Interpretability Benchmark. We discover that many tested circuits closely replicate correct behaviour while missing most of the model's errors. On indirect object identification (IOI) for GPT-2 small, under mean ablation, the manual circuit and tested automated circuits, including one trained against the model's full output distribution, agree with the model on 97.3-99.5% of prompts it answers correctly but only 11.4-41.7% of errors. An IOI case study shows that lost errors are recoverable by restoring omitted attention-heads which raise error reproduction from 14.2% to 75.1% on a separate held-out set with 0.41 percentage point decrease on correct agreement, exceeding matched random extensions and scalar-biased control. Intervention traces show how omitted computations produce specific wrong answers for a reproducible subset of errors. In all, these findings show circuits can preserve task success without adequately explaining model's failures, and support exact error reproduction as a necessary, but not sufficient, test of circuit-based explanations of model behaviour.
Jev is a non-generative "System One" model that assigns probabilities to predefined answer options and cannot answer outside them. Its accuracy and calibration on medical question-answering and case-based diagnostic-reasoning tasks are unknown. We evaluated Jev 1.13 on four medical benchmarks: MetaMedQA, PubMedQA, DiagnosisArena-MCQ and the NEJM Case Challenges. GPT-6 Sol, with (medium) and without reasoning, was the reference. The primary outcome was top-1 accuracy; key secondary outcomes were calibration, selective prediction and recognition of unanswerable questions. All 8,469 requests returned a valid answer. Jev's accuracy was similar to that of GPT-6 Sol with medium reasoning on PubMedQA (78.4% vs 78.2%;), lower on MetaMedQA (74.8% vs 82.7%) and much lower on DiagnosisArena-MCQ (59.8% vs 82.4%;) and the NEJM cases (61.8% vs 82.4%). On MetaMedQA, Jev's probabilities were the best calibrated (expected calibration error 0.063 vs 0.146), and its answers with a probability of at least 0.9 (52.9% of questions) were 93.4% accurate, but GPT-6 Sol was as accurate when it accepted a similar proportion of questions. On DiagnosisArena-MCQ, Jev's probabilities discriminated poorly (AUROC 0.645 vs 0.768). Of the 162 questions whose correct answer was "I don't know or cannot answer", Jev chose that option for 10.5% (GPT-6 Sol, 8.6%). Median latency was 0.27-0.31 s; all 2,823 items cost USD 0.08. Jev was fast and inexpensive, and its accuracy was similar to that of a frontier LLM on research abstracts but lower on examination questions and much lower on complex diagnostic cases. Task-specific validation is required before clinical use.
Objective: To develop and characterize CLEAR-Med, a dual-agent framework for natural-language analysis of structured clinical data that separates SQL-based invocation from independent validation. Methods: CLEAR-Med uses one agent to translate a question into executable Structured Query Language (SQL), retain the executed query and database result, and produce a draft. Deterministic checks and a separately invoked cross-provider Validation Agent then accept the draft, request one bounded repair, or abstain. We formalized the system as a bounded selective pipeline and evaluated CLEAR-Med's configuration and scalability, and the Invocation Agent's accuracy and consistency on a 25-query development benchmark, using a harmonized 21-site neonatal hypoxic-ischemic encephalopathy table containing 532 de-identified infant records and approximately 1,300 variables. Results: CLEAR-Med completed all six nominal scalability configurations, including 500x1300. Across 25 development-benchmark queries repeated five times, the Invocation Agent answered 83 of 125 responses correctly (66.4%; query-cluster bootstrap 95% CI, 48.0-83.2%), compared with 15 of 125 (12.0%; 95% CI, 3.2-22.4%) for the ungrounded ChatGPT baseline, a paired improvement of 54.4 percentage points (95% CI, 36.8-72.0%). Conclusion: CLEAR-Med provides a general architecture for traceable analysis of structured clinical data: numerical claims remain linked to executed SQL, and unresolved cases can fail closed. The reported experiments characterize CLEAR-Med's configuration and scalability and the Invocation Agent's accuracy, while the formal analysis establishes the encoded-property guarantee of the complete control flow; a prospective full-pipeline evaluation of the validation and abstention stages is the next stage of this work.
Audio emotion recognition (AER) typically assigns a single label to an entire recording, leaving the temporal scope of that label ambiguous when multiple speakers and affective events are present. We address this limitation by reformulating AER as a Temporal Affective Grounding (TAG) task that associates emotions with temporally bounded speech spans and vocal tone descriptions. To support this formulation, we curate temporally annotated versions of existing emotion recognition datasets and construct recordings containing two to four affective speech spans, including overlapping speech. Training in this longer format with a standard language-modeling objective can degrade both emotion recognition and temporal grounding performance, while tone descriptions can provide shortcuts for emotion prediction. To address these challenges, we introduce Masked Temporal Affective Grounding (M-TAG), a supervised training objective that combines full-sequence language modeling with emotion and timestamp cross-entropy losses under attention masking. The masking varies the context visible to emotion-label tokens to reduce reliance on shortcuts and improve generalization, while the timestamp loss incorporates a distance-aware weight to penalize larger temporal errors. We evaluate EMO-TAG, a model fine-tuned using our dataset and objective, on emotion recognition and affective temporal-grounding against three AER and audio-language baselines: Flamingo-Next, Audio-Reasoner, and AffectGPT. Our results show that existing models achieve limited affective temporal-grounding despite competitive emotion recognition performance.
GPT-6 Astra completed a 50-tab tax workbook twice as fast as GPT-5.6 Sol, and its stronger understanding of user intent gives Basis more confidence in real-world use.
We’re collecting real stories of builders, tinkerers, researchers, and creators who are using Codex to do incredible things. If you want to be a part of the next chapter of the Codex Originals program, tell us more about your story and project below.
Clinical prediction models estimate risk from patient measurements, while large language models support medical text understanding and question answering. Yet their language capabilities do not ensure accurate prediction from structured, high-dimensional clinical time series. Improving this ability would connect risk estimation with flexible questions about a patient's evolving condition. We introduce ViSTA, a compact adapter that incorporates irregular numerical measurements into a pretrained vision-language model's chart representations. It learns corrections to visual tokens while leaving all pretrained parameters unchanged. On MIMIC-IV, ViSTA has the highest mean scores among the compared adaptations on all four metrics for acute kidney injury and mortality prediction across models with 2-9 billion parameters. With 0.516 million trainable parameters, the 2-billion-parameter model reaches an area under the ROC curve of 0.7376 for acute kidney injury, compared with GPT-5.6 Sol's 0.7380 with text input and high reasoning effort. Training for temporal question answering yields 69.27% accuracy at 4 billion parameters with over 90% fewer trainable parameters than low-rank adaptation using charts or numerical text, at a 2.82-4.88 percentage-point accuracy gap. ViSTA extends pretrained language models to numerical prediction and temporal questions.
AI tutoring could markedly improve learning outcomes for students in developing regions such as Vietnam, yet the two obvious paths both fall short. Cloud assistants such as ChatGPT route sensitive student data to foreign servers---violating data-sovereignty laws such as Vietnam's Decree 53---and, pre-trained on Western-centric corpora, are not organized around the national textbook curriculum, so their knowledge of local content is unsystematic and frequently hallucinated. Self-hosting an open model keeps data on-premise but hits a two-fold wall: post-training quantization (AWQ, GPTQ) tames the static weight footprint, yet the dynamic KV cache and prefill latency of long tutoring contexts still cause out-of-memory failures and slow responses on consumer GPUs, while the model keeps hallucinating on region-specific material. We present DeepEdu-v1, an AI-tutoring system for Vietnamese education built on SCALE (Self-improving Context-Aware Learning Engine), a framework with two innovations. First, a long-context inference engine amortizes token selection from per-sub-chunk to per-cluster granularity; on long-context retrieval it issues x7.7 fewer retrieval calls than a state-of-the-art selective-attention baseline, cutting prefill latency (TTFT) by roughly 35% while matching or improving task accuracy. Second, a self-improving agentic layer continuously curates a verified playbook from past interactions instead of fine-tuning, a design intended to progressively reduce reliance on dominant-language priors as trustworthy local knowledge accumulates. In its deployed configuration, DeepEdu achieves a nearly x2 TTFT speedup over standard vLLM serving and lifts agentic accuracy from 70.0% to 79.5% on complex tasks, with the strongest per-track gains across financial-reasoning and interactive-agent benchmarks.
And I don't mean the doomerism nonsense. From rogue AI, doomerism, AGI. I mean real regulation: 1) Fair accounting of input/output tokens. Where is the weights and measures department to enforce fair counting of usage? If I pay for a gallon of gas, that is what I expect. When I pay for a dollar of LLM compute, is it going to play games with the input (or especially the output) tokens to devalue my dollar? 2) Quality. When I request a higher effort of a frontier model, and then I get downgraded to a lower tier model or with different effort levels (and I may not even know if I was downgraded!), they are effectively selling you a service you didn't ask for. "But the terms of service", great, if they feel it is such a risk to downgrade, and they get it wrong a lot, then they should discount the cost of the downgraded response. 3) Adulterating outputs. LLMs are now routinely adulterating output for tracking and accountability purposes, but is there anyone to validate these efforts are not lowering the quality of the response? There needs to be more discussion around companies adulterating their product, and holding them to account. I have several times noticed words were swapped: sometimes the word pick made no sense, when it previously didn't do this. And I wasn't looking for it or cherry picking. It stood out that badly. 4) Furthering point 3, are the responses themselves going to be seeded such that the author of that response can be identified? Has anyone asked themselves if they could deanonymize LLM responses online like yellow printer dots? 5) Quotas. Related to points 1 and 2, when I am on a metered service, what happens to requests that fail mid-response, or the response is cut midway? Why should I pay for a failed response? Maybe that is fine, but who is validating they aren't doing this at scale to save money? I feel every time we earnestly or "good faith" engage in the nonsense regulations that the major LLMaaS (LLMs as a Service) duopolists, we fail to engage in real discussions about regulating the product we are paying for . I don't see why we should believe these companies to operate in good faith post-IPO. Where is the incentive once they shut down all competition to ensure only OpenAI/Anthropic are the LLMs you can purchase from? Look to what Uber/Lyft have become over the years. Massively degrading their service and increasing nickel and dime costs. They skirted regulation then demanded the world bend and write regulations that fit their revenue goals everywhere they could (or threatened to pull out where they couldn't). And if the danger really is as bad as they say, then how can we trust them to "regulate" that threat if they would refuse regulations around basic delivery of service? Comments URL: https://news.ycombinator.com/item?id=49871747 Points: 4 # Comments: 1
I've spent the last 10 months building Beauty: a local, powerful greco-roman Markdown editor where the page looks finished while you type. It runs on the Mac, iPhone and in the browser, and the browser version opens without an account: https://app.markdown.beauty I was frustrated with writing apps, and like most I used Apple Notes. After 2+ years immersed in Cursor, Codex and CC, I wished some interactions existed in a text editor. The Markdown renders in place as you write it, including tables (which you can drag and resize without lining up pipes), KaTeX math and Mermaid diagrams. A lot of thought went into typography and the writing experience. Your notes on the Mac are plain .md files in a folder with images next to them, so they open in anything else too. Each note keeps 200 versions. If you share a note with a link it syncs peer to peer over WebRTC while someone has it open, and the server that connects you only ever sees a random room id. Most of the ~1,000 commits went into the editor itself. It's a contenteditable, and I don't let the browser restructure the document at all, which was a lot more work than I expected. Building this to be performant and cross-platform required some novel approaches, which I'll later open source. The editor is free with no account or ads, and there's an optional subscription for sync through your own iCloud. https://www.markdown.beauty Comments URL: https://news.ycombinator.com/item?id=49866597 Points: 25 # Comments: 20
Hi everyone, I am KD - Back in my college days, I dabbled with coding, learned the basics, HTML, CSS etc. but somehow I ended up in Finance which consumed the next 20 years. Then, during covid I picked up coding again, learned react, typescript, etc - even built a rudimentary site - and then came the chatgpt moment, followed by Claude etc. So, as a side project, considering that I had spent 20 years in finance and M&A I started building Ekselio, loveable for finance workflows. Differently from other vibe coding tools, this is local first - meaning the workflows are orchestrated by the LLM based on the file schema but then the execution happens in the browser - the cool thing is that you can see the canvas, the previews of each node, and the code base (SQl) - see the Canvas tab. That said I did include a home tab, which is more of a visual, mainly because people think it's easier to relate to how folks operate, although I know finance and accounting folks usually like the workflow and full transparency - the other thing is that once you create a workflow you can save it and then next month,you can just click Run and it runs it again with no LLM tokes involved - so the infrastructure is minimal. Everything can be exported in an excel data pack with M code in case someone would want to reproduce - think Audit ready. And, I also connected to quickbooks on line so someone can directly pull in and manipulate their financials live - I also connected to Fred which carries literally 800k series in terms of macro economic data - i.e. oil prices, inflation, CPI etc and finally to stocks. I am hoping people find holes in the architecture or any ideas to make this flow a little bit better. It is in progress so if you run into bugs let me know. Considering that I am using my own API key, I have put a wall in case people overuse it but DM me if you would like to continue using, I can turn off the wall for you - hoping to keep that to a small number of users if I can and will do my best. Thank you again, appreciate you taking the time. Comments URL: https://news.ycombinator.com/item?id=49849986 Points: 4 # Comments: 0
Shows an Incorrect API Key error...which is weird. Nothing on the status page. Per Twitter/X this appears to be widespread Comments URL: https://news.ycombinator.com/item?id=49851032 Points: 34 # Comments: 13
Large language models (LLMs) have shown strong performance in creative text generation, yet their ability to produce culturally grounded and stylistically constrained literary forms remains underexplored. Prior work has focused largely on modern language varieties and poetry, while classical prose traditions such as maqama remain largely unstudied. The maqama is a classical literary genre characterized by rhymed prose (saj), dense rhetorical ornamentation, and episodic narrative structure, making it a challenging testbed for evaluating whether LLMs can move beyond surface fluency toward deeper literary competence. In this paper, we present the first controlled evaluation study of maqama generation with LLMs, comparing five models under zero-shot, few-shot, and rule-based prompting, and evaluating outputs through both human annotation and an LLM-as-a-judge framework across dimensions such as rhetorical richness, saj density, structural coherence, and stylistic authenticity. Our results show that prompting strategy plays a strong role in stylistic quality: few-shot prompting most consistently improves saj density, while its effects on rhetoric and coherence vary by model, with the strongest models (GPT-4o and GPT-5.4-mini) benefiting most from rule-based prompting on these dimensions, though zero-shot prompting yields the highest aggregate scores across all five models. We further observe systematic differences between models in stylistic alignment with Arabic maqama conventions, and corroborate our findings with a second independent LLM judge, paired statistical significance testing, and non-LLM proxy measures of saj.
jevper is the Jev interface pointed at whatever model you already have. Comments URL: https://news.ycombinator.com/item?id=49815066 Points: 4 # Comments: 0
移动出行行业正经历由生成式人工智能进步所驱动的范式转变。仅考虑汽车，该行业全球市场年规模约为2.9万亿美元，而这些技术的整合有望影响全球超过15亿辆汽车。随着LLMs在移动出行领域得到越来越广泛的采用，人们开始关注网络安全、隐私和可靠性问题。因此，本文对当前应用进行综述，并评估这些挑战。鉴于European AI Act将交通运输领域的人工智能列为高风险领域，我们根据其要求归纳出九类技术类别，用于评估当前研究和未来部署。研究结果表明，现有研究主要关注GPT和Llama模型（占所审阅研究的50%以上）以及交通应用，却在很大程度上忽视了安全性、隐私和可靠性。这一缺口同样存在于对AI Act的合规性方面：在审阅的35项研究中，仅有1项包含部分漏洞评估，另有1项包含部分风险管理系统。我们发现，强大的优化性能与监管遵从性之间存在明显差距，这表明合规性受限的主要原因并非技术能力不足，而是过度关注静态性能、忽视全生命周期安全；同时，这也凸显了在安全关键型智能交通系统中贯彻安全设计的迫切需求。
Hey HN! We’re Vincent and Jochen from Sitefire ( https://sitefire.ai ). We have been working together for years, with backgrounds in RL/optimization at Stanford and software engineering from Technical University Munich (TUM). With Sitefire (YC W26), we help marketing teams get recommended by AI Search (ChatGPT, Google AI Overviews, AI Mode, Claude, etc.). Our software monitors prompts, sees which web pages get cited, and uses these insights to help marketing teams take action, e.g. create YouTube videos or write the right blog posts. This means we have a commercial stake in AI-generated web content. And for now, high-information, AI-generated content works great to get cited and recommended in AI Search. But after talking to hundreds of marketing teams, it became clear that everyone despises AI-generated content (“AI slop”). And yet, everyone still wants to leverage AI to create content. So we asked ourselves: what characterizes AI slop? Can we train a model to identify it from human-generated web pages? Researchers from the University of Maryland and Google DeepMind already asked this question for fiction. Their paper StoryScope (Russell et al., 2026) showed that you can tell AI-written stories from human ones by their structure alone, without looking at the words. We ported their pipeline to commercial web pages. Using the Wayback Machine, we collected 2,250 blog posts from 268 B2B company websites that were written before ChatGPT existed. For each blog post, five AI models (GPT-5.4, Claude Sonnet 4.6, Gemini 3 Flash, DeepSeek V3.2, Kimi K2.5) wrote their own version. Instead of looking at the words, we looked at how each post is built. We had an AI model answer 214 questions about every post, e.g. how hard it pushes its own product, whether it backs up its claims with sources, or whether it quotes a named expert. Then we trained a classifier on these answers. On blog posts it had never seen before, our classifier told AI-generated and human posts apart with 98% accuracy, getting only 19 of 1,740 wrong. Why does it work so well? Because all five AI models write in a similar shape. Mapping every AI model’s values for these features, we see they cluster together, while the human values sit apart and spread out much more. Of the 1% most unique blog posts in our data set, 149 are human, only 4 are AI. So what characterizes AI slop? It tells you the same thing three times. The title already promises what you'll get ("How to Cut Onboarding Time in Half"), the intro lays out what's coming, and the ending says it all again. 77% of the AI posts end by repeating their main point, compared to only 12% of the human posts. We call it the tidy, self-announcing blog post. Still, each AI model has its own accent. We trained a second classifier to tell which of the five AI models wrote a post, or whether a human did. It picks the right author 79% of the time, where random guessing (1 in 6) would get 17%. Almost all of its mistakes are mix-ups between the AI models, not between human and AI. The cool thing about structural features is that you can't simply reword your way out of it. We had each AI model rewrite its own posts until, on average, 73% of their original 13-word sequences were gone, and the AI slop classifier still worked just as well. We're building this into Sitefire: our agents get a structural understanding of text, so the posts they write go deeper and vary the way human writing does. There's a lot we haven't tested yet, like the myriad of humanizer tools, human rewriting, restructuring a post, or prompting an AI model to explicitly avoid these habits. And our human posts are mostly from 2020 to 2022, while the AI posts were generated in August 2026. Structure can't really tell when a human post was written, but it's still not a same-year comparison. We published the study with all the figures on arXiv: https://arxiv.org/abs/2609.15369 . The code is on GitHub: https://github.com/pulse-energy-eu/slopshape We're pretty sure your own blog isn't AI slop, is it? We built a checker that runs one of your posts through the ten features from the paper, so you can see for yourself (the full report asks for a work email): https://sitefire.ai/slop-checker . Think you can tell AI slop from human writing? We also made a little game to see if you can keep up with our model, which gets all five rounds right: https://sitefire.ai/spot-the-slop . Comments URL: https://news.ycombinator.com/item?id=49800566 Points: 5 # Comments: 0
Hey HN! We’re Max and Gabriel the co-founders of Coverage Cat. We’ve been friends for over a decade, met in college then hung out mostly on the internet. We love building products that help people optimize the crufty corners of their lives. Max is a former Google/Microsoft/Two Sigma PM and Gabriel has been working in startups for over a decade. We’re building Coverage Cat: an AI-native, you can shop with your AI agent or via a form, insurance brokerage for consumers. Our current focus is helping tech folks (think L3-L8) buy umbrella insurance. For those who are unfamiliar with this type of coverage, here’s a blurb from one of our insurance agents: “Umbrella liability insurance is a policy that provides additional coverage above the limits of existing auto, homeowners, renters, or landlord policies. It kicks in after the underlying policy limits are exhausted and may also cover personal injury claims like libel, slander, or defamation that standard policies often exclude. Policies typically start at $1 million in additional liability coverage and can protect you, your spouse, dependents, and even pets in your household.” Buying insurance online is a miserable experience. You fill out a bunch of forms that say they’re going to give you quotes. Most don’t and then you end up with tons of unwanted phone calls and spam. Even if you wade through the muck, you’ll often still end up with sub-optimal prices and coverage because of a lack of market transparency and poorly-aligned advisory incentives. We came to this problem because Max is a classic personal finance obsessive. He's the type of person that needs to know that he has the best possible price for something (everyone has one friend with the patience for a two hour phone call with the bank for some $10 fee) or that his insurance coverage is as tailored to his risk profile as the market will allow. He actually reads insurance policies end to end. Years ago, when he shopped for his homeowners insurance it took him 10+ phone calls and dozens of online form fills to find the right deal. He also suffered tons of collateral damage in the process: websites sold his information to dozens of agents who proceeded to call, text, and email him a spam hoard that grows to this day. After this painful search made us aware of the problem, we also spent a lot of time doing user research with wealthy-ish tech employees to see if they felt the same way. They did, and also complained that they 1) often felt like they'd left money on the table, 2) didn’t really understand their coverages, and 3) that brokers were slow, hard to communicate with, and unreliable. To solve the problem, we set out to build Coverage Cat. Our approach has three major components: First, we work really really hard to find and partner with insurance carriers that offer high-quality, competitively-priced products. We've also taken time to identify insurers that work on a direct-to-consumer basis or transparently list their prices online so we can make our customers aware of all the options available to them. We get paid commission when we match users to great insurance deals, but we still show and recommend folks deals that we don't make money on when they're accessible and will always recommend users buy what's best for their personal financial situation. Our goal is to meaningfully improve price transparency so consumers (and their AI agents) can make informed coverage decisions and not get nickel-and-dimed! Second, we automate every possible part of the insurance process. Some forms, portals, & approvals still need human review due to regulatory constraints, but if it's not restricted, we're automating it. LLMs have given our small team tremendous leverage and enabled us to tackle problems at a scale that would've been unfathomable five years ago. Finally, we build for agents first (a remix of the more classic adage, build for developers/APIs). We've always been big believers that personal finance would be the killer use-case for the AI revolution and have done our best to position ourselves to surf the wave. As personal assistants have started to explode onto the mainstream (think Muse, Instinct, Town, Codex, CC etc.) we've ensured our shopping and comparison process works via API/MCP so agents can present their users with all the information they need to make a good decision and transform a miserable shopping experience into a delightful one. Coverage Cat is an unusual business because, while we also help folks find homeowners insurance in California and Texas, our main focus is on umbrella insurance. Most brokerages only sell umbrella coverage as a customer retention product, but it makes them almost no money. We focus on it as our core business because: 1) it's one piece of the insurance puzzle that many folks in tech overlook and/or are confused about even though it can have a huge impact on their financial well-being. 2) Existing online tools and brokerages broadly don't make it easy to comparison shop for. 3) It was, for us, the most technically feasible candidate for automation given startup resource constraints. Insofar as we know Coverage Cat is the first instance of a tool/portal that allows AI agents to complete most of the comparison and shopping steps required to allow people to buy insurance. We’d love your feedback so please do try it out. Just feed the prompt: “Shop for umbrella insurance with Coverage Cat” to your personal agent and let us know what you think! If you want to watch a couple of agents navigate the flow before you try it out there’s a video here: https://youtu.be/1BUkgAn6s-I Also happy to answer any questions about the product, the problem space, and even some insurance questions (Max is a licensed agent) where regulation permits. Cheers! Comments URL: https://news.ycombinator.com/item?id=49804931 Points: 3 # Comments: 0
Sandboxed computer environments support multi-step reasoning with tools, executable programs, and persistent files, yet their extension from language models to vision-language models (VLMs) introduces a distinct state-management problem. Visual reasoning produces intermediate image-valued evidence---crops, masks, overlays, zoomed regions, and analytic renderings---that must remain addressable without accumulating unboundedly in multimodal context. We introduce VLM-in-Sandbox, a training-free framework for agentic multimodal reasoning in controlled computer environments. Its Visual Workspace registers generated artifacts in an image ledger, maintains a bounded active visual context, and lets the model explicitly promote selected evidence for subsequent inspection. This separates visual evidence generation, performed by sandbox tools, from visual evidence management. Across seven benchmarks and four base VLMs, VLM-in-Sandbox achieves the highest sample-weighted average accuracy among Vanilla VLM, Append-only Sandbox, and the proposed method. A compiler-matched $2\times2$ study on 1,260 examples further separates model-directed visibility from bounded retention: VLM-in-Sandbox reaches 66.27% accuracy with 18.6% fewer total tokens than the automatic, retain-all control. Over all 6,350 submitted GPT-4.1-mini examples, it produces 302 rescues and 142 regressions relative to Original Append-only. A local vLLM study with prefix caching confirms that the smaller request workload also reduces uncached tokens, time to first token, and end-to-end latency. These results identify explicit visual evidence state as a central abstraction for sandboxed VLM agents.
Functional annotation of newly sequenced proteins remains a bottleneck in molecular biology: the number of sequences in public repositories grows far faster than the capacity for manual curation. Most computational approaches consider annotation as multi-label classification over a fixed ontology, which constrains predictions to a predefined label set. In this work we study the the protein annotation as a sequence-to-text generation problem. We fine-tune the 3B-parameter Ministral 3 base model with QLoRA (4-bit NF4 quantization with low-rank adapters) on sequence annotation pairs. We assess predictions with an LLM-as-expert protocol: a GPT model prompted as a senior molecular-biology curator scores organism identification as binary and function annotation quality. We conclude that QLoRA-fine-tuned compact LLMs can generate curator-style annotations with genuine biological value for a substantial subset of proteins. We also discuss future directions in data quality, model scaling, and evidence grounding that are needed to make the approach sufficiently reliable for practical use.
Recent advances in natural language processing have led to molecular Large Language Models (LLMs) with strong performance across diverse chemistry tasks. However, they still struggle to capture fine-grained structure-property relationships, particularly how small, localized modifications alter a molecule's behavior. To address this limitation, we introduce MolSC, a dataset of substituent contributions, defined as property changes induced by attaching specific substituents to molecular scaffolds. Curated from manually annotated bioactivity records, MolSC spans structural-alert liability, target-specific bioactivity, and physicochemical descriptors, and contains 181K substituent-level examples for training. We further propose MolSC-Bench, a held-out evaluation benchmark of 1,541 examples disjoint from MolSC at the scaffold, substituent, and molecule levels. Our experiments show that existing molecular LLMs and strong proprietary models such as GPT-5.2 and Gemini-3-Flash show limited reliability in substituent contribution prediction. In contrast, training on MolSC substantially improves this ability and achieves strong performance across diverse downstream molecular tasks. These results highlight substituent contribution learning as a key component of fine-grained molecular understanding.
OpenAI is working with an independent Advisory Group on Mathematics and Artificial Intelligence to guide the review and communication of emerging AI results.
At, GPTree, we run several coding agents across our team on one repo using parallel worktrees. Apart from wasted time reviewing and fixing conflicts at PR time, the failures that hurt the most are when multiple plans or tickets cause architecture changes that cannot both be true. Ex. one agent replaces a class while another one is in the process of extending it. Git only notices if the resulting patches happen to touch the same lines and the review only catches it if they are familiar with both tickets. Foremerge is a local "git like" coordination layer that sits above git (ie. does not interact with or change the way git and worktrees function), Before editing each agent publishes an intent and the scopes it will change, with the operation it plans to complete on each one. foremerge intent publish --agent "$A" \ --summary "Replace PaymentService with StripePaymentService" \ --scope symbol:PaymentService=replace foremerge intent publish --agent "$B" \ --summary "Add PayPal support to PaymentService" \ --scope symbol:PaymentService=extend The publish by the 2nd agent returns a HIGH destructive_vs_additive finding before writing any code. Agents keep their own worktrees and the shared state is one SQLite file in gits common direectory. No hooks, no merge drivers, nothing rewrites your history. It ships as one Rust binary with a CLI and MCP server with 18 tools and `foremerge setup all` wires it into Claude Code, Codex and Cursor. Because the protocol has nothing provider specific, a Claude agent and a Codex agent coordinate through the same store. Before any work is accepted, Foremerge runs a named check that you configured against the exact git state of the change. An agent that says tests pass is recorded but it dosnt satisfy the acceptance gate without running the check itself. Detection is deterministic, no judge model reading your code. HIGH conflicts are only asserted for declared operations, ie. matches inferred from prose cap out below high. Claims are advisory leases, not locks so two agents can still hold the same scope without deadlock. The open source version is single matching and so not a distributed consensus. We have tested this up to 98 parallel agents all working on the same repo with zero conflicts (was supposed to be 100 but 2 agents failed to run due to resource limitations) I replayed 76 intents on my own agents from a build last week in the order they happened. The sample had exactly 1 conflict (which was flagged) and the review found a blind spot where one agent claimed scope by class name and the other claimed it by an internal method. We are working on fixing that for the next release. Setup is a 30s install by pasting the quickstart instructions from the readme.md into your agent or manually: `curl -fsSL https://foremerge.com/install.sh | sh` or `cargo install --locked foremerge`, then `foremerge init && foremerge setup all` in a repo. Apache-2.0. The feedback I want most is which conflicts between your agents plans would you actually want flagged and which would you tollerate as noise? Repo here: https://github.com/naw103/foremerge Website: https://foremerge.com More information on the problems this solves: https://foremerge.com/blog/ Comments URL: https://news.ycombinator.com/item?id=49789356 Points: 4 # Comments: 0
Cooley built GO Public with ChatGPT Work to bring intelligence to the IPO process, helping lawyers surface issues earlier and focus judgment where it matters most.
Hey HN, we're Nars & Nishant, founders of Skillsync ( https://skillsync.com ) Skillsync lets you move your AI chats across every coding agent. Most of our work exists as conversations, which are currently scattered across our agents. Though stored locally, these conversations use different formats. This is annoying because you cannot simply switch between agents without starting over. We get locked into a single provider and their agent as we invest in skills and memories over time. Skillsync acts as a universal converter. It moves the entire session, including all the messages, reasoning and tool calls so you can pick up right where you left off. Skillsync collects all your sessions in one place and makes them searchable. It breaks down what each session is carrying, including which loaded skills the agent is actually using, making stale context easy to spot. You can also create shared workspaces to sync sessions across your team. You can build your own closed loop systems. Everything runs locally except when you share to workspaces. The core is an open-source Rust engine called txcript ( https://github.com/skillsynchq/txcript ). It translates a session from one agent's on-disk format into another's, mapping conversation, reasoning, and tool history. Think ffmpeg or pandoc, but for agent sessions. On top of that engine is a local-first desktop app. Your agent sessions are normally scattered across different tools' folders in formats you'd never read by hand; the app surfaces them in one place with a UI that makes them actually readable, the conversation, the reasoning, and the tool calls, so you can revisit what happened, move a session into another agent, or share it with a teammate. Skills and memory are stored as portable, human-readable markdown you own, and exposed to any agent over MCP for search and selective retrieval. Sessions and translation run locally on your machine. The one thing that leaves is what you explicitly share into a team workspace. It’s been surprising to us to see how locked-in people already are without realizing it. You don't notice you're trapped in one agent or harness until you try to leave, and by then you've got months of sessions stuck in a format only that tool can read. The lock-in is invisible right up until it's expensive. Once sessions are portable, they stop being disposable logs and become something you can actually build on, spotting patterns in how you work, handing a session off to someone else, letting a non-technical teammate pick up where an engineer left off. The session turns out to be the unit of collaboration, and right now it's being thrown away. Before Skillsync, Nars and I built an open-source payment orchestrator that let merchants route across many processors instead of getting locked into one (30k+ GitHub stars). Fighting vendor lock-in by making incompatible systems interoperate was the whole job. We came into YC with a different idea, but the more we lived inside coding agents the more we saw the same problem from the other side: your context is locked into whatever agent you start with, because every agent stores sessions differently and none of it moves. Some of the things we’re seeing people do with Skillsync are: - Moving a session between Claude Code, Codex, and Cursor mid-task, including when you hit a usage limit on one and want to keep going on another. - Thinking through a problem in a browser Claude chat, then handing the whole thread to Claude Code or Codex to build. - Sharing research sessions with a team. Everyone uploads their sessions, so teammates and their agents can build on each other's work instead of repeating it. It's available as a Mac app, CLI and MCP. Here's the demo: https://youtu.be/7hVhSnSKGl8 . Would love to hear your feedback and answer questions! Comments URL: https://news.ycombinator.com/item?id=49743049 Points: 10 # Comments: 2
Hey HN, we're Nars & Nishant, founders of Skillsync ( https://skillsync.com ) Skillsync lets you move your AI chats across every coding agent. Most of our work exists as conversations, which are currently scattered across our agents. Though stored locally, these conversations use different formats. This is annoying because you cannot simply switch between agents without starting over. We get locked into a single provider and their agent as we invest in skills and memories over time. Skillsync acts as a universal converter. It moves the entire session, including all the messages, reasoning and tool calls so you can pick up right where you left off. Skillsync collects all your sessions in one place and makes them searchable. It breaks down what each session is carrying, including which loaded skills the agent is actually using, making stale context easy to spot. You can also create shared workspaces to sync sessions across your team. You can build your own closed loop systems. Everything runs locally except when you share to workspaces. The core is an open-source Rust engine called txcript ( https://github.com/skillsynchq/txcript ). It translates a session from one agent's on-disk format into another's, mapping conversation, reasoning, and tool history. Think ffmpeg or pandoc, but for agent sessions. On top of that engine is a local-first desktop app. Your agent sessions are normally scattered across different tools' folders in formats you'd never read by hand; the app surfaces them in one place with a UI that makes them actually readable, the conversation, the reasoning, and the tool calls, so you can revisit what happened, move a session into another agent, or share it with a teammate. Skills and memory are stored as portable, human-readable markdown you own, and exposed to any agent over MCP for search and selective retrieval. Sessions and translation run locally on your machine. The one thing that leaves is what you explicitly share into a team workspace. It’s been surprising to us to see how locked-in people already are without realizing it. You don't notice you're trapped in one agent or harness until you try to leave, and by then you've got months of sessions stuck in a format only that tool can read. The lock-in is invisible right up until it's expensive. Once sessions are portable, they stop being disposable logs and become something you can actually build on, spotting patterns in how you work, handing a session off to someone else, letting a non-technical teammate pick up where an engineer left off. The session turns out to be the unit of collaboration, and right now it's being thrown away. Before Skillsync, Nars and I built an open-source payment orchestrator that let merchants route across many processors instead of getting locked into one (30k+ GitHub stars). Fighting vendor lock-in by making incompatible systems interoperate was the whole job. We came into YC with a different idea, but the more we lived inside coding agents the more we saw the same problem from the other side: your context is locked into whatever agent you start with, because every agent stores sessions differently and none of it moves. Some of the things we’re seeing people do with Skillsync are: - Moving a session between Claude Code, Codex, and Cursor mid-task, including when you hit a usage limit on one and want to keep going on another. - Thinking through a problem in a browser Claude chat, then handing the whole thread to Claude Code or Codex to build. - Sharing research sessions with a team. Everyone uploads their sessions, so teammates and their agents can build on each other's work instead of repeating it. It's available as a Mac app, CLI and MCP. Here's the demo: https://youtu.be/7hVhSnSKGl8 . Would love to hear your feedback and answer questions! Comments URL: https://news.ycombinator.com/item?id=49743049 Points: 4 # Comments: 0
made this AIM buddy list but for our token usage because i wanted to bring a little bit of the old internet back! i remember those days where i'd casually check if friends are online and find excuses to say hi and poke them give it a try! i'd greatly appreciate feedback Comments URL: https://news.ycombinator.com/item?id=49743095 Points: 7 # Comments: 8
OpenAI banned accounts that very likely originated in Cambodia and used AI to pose as recovery services, law firms, and authorities targeting people affected by fraud.
OpenAI banned accounts that very likely originated in Cambodia and used AI for scam outreach to Indonesian loveseekers, including translation and engagement.
OpenAI banned accounts linked to the Rybar network, some of which likely originated in Russia, that used AI to support multilingual influence activity across websites and social platforms.
OpenAI banned an account linked to an individual associated with Chinese law enforcement, using AI to plan influence activity, harassment, and online operations.
OpenAI banned accounts linked to a previously unreported operation we dubbed "Trolling Stone", using AI to generate comments about an alleged Russian cult leader’s arrest in Argentina.
Despite recent advances in large language models (LLMs), performing logically consistent deductive reasoning over extended interactions remains challenging. Tasks that require integrating evidence across multiple reasoning steps, maintaining consistency with prior inferences, and updating beliefs under new constraints can surface limitations in current models while providing a useful testbed for evaluating reasoning enhancements. In this paper, we implement a text-based, multi-agent version of the classic board game Clue as an environment to evaluate multi-step, agentic deductive reasoning. In this setting, agents must infer hidden information from a sequence of observations, maintain consistency across turns, and reason over an evolving set of logical constraints. We instantiate six LLM-based agents (GPT-4o-mini and Gemini-2.5-Flash) as players that engage in turn-based gameplay; using three agents per model family, we establish baseline performance across repeated games. We then introduce a tool-augmented approach in which a structured possibility matrix converts implicit game state from generated reasoning logs into an explicit representation of remaining possibilities. The possibility matrix encodes extended-turn memory and deductive constraints, offloading these tasks from the agent. We compare this approach against the baseline to evaluate how tool augmentation supports reasoning quality and task success for autonomous agents in a strategic reasoning environment.
Coding agents are typically evaluated with desired behavior specified through issues or instructions. In practical web development, however, agents may need to infer behavior from working software and implement it in an incomplete application. We introduce ProgramDistill, a benchmark evaluating coding agents on features discovered through interaction with fully functional reference applications. We build ProgramDistill by factorizing applications into features of different granularities, each associated with replayable behaviors executable via its gold patch. Our pipeline, mine-craft-patch, discovers 1,975 replay-verified behaviors across 26 applications and constructs 4,063 tasks without human intervention. Across nine frontier coding agents, GPT-6 Astra and Claude Opus 5 achieve 49.2% and 28.8% success on cumulative workflows in full-application reconstruction. In partial-application reconstruction, success falls from 100% to 64.0% and from 96% to 32% as restoration depth increases from 1 to 8. ProgramDistill thus provides a scalable benchmark with controlled difficulty for evaluating and diagnosing coding agents, and a natural basis for future curriculum-based training.
The O-RAN control plane is becoming agentic: autonomous AI agents, deployed as rApps by different vendors, independently close control loops over shared radio resources. We demonstrate on a live O-RAN system that this independence is unsafe. Two agents with individually correct objectives, one protecting a latency SLA and one maximizing utilization for energy efficiency, jointly drive recurring opposing excursions of the shared resource partition that neither produces alone. Existing conflict-mitigation mechanisms presume a statically known application population and cannot govern agents whose behavior emerges at run time. We present AURA, a lightweight arbitration layer that admits agent actions only when they satisfy feasibility invariants, per-variable dwell times, and a deadband, and we prove the arbitrated system converges to a feasible operating point. Implemented on an OpenAirInterface (OAI) testbed with measured one-way latency and throughput, AURA reduces recurring shared-state excursions by more than an order of magnitude (from 8.4 to 0.4 PRB amplitude) and virtually eliminates cross-slice throughput starvation (from 40-55% to 0.3%), while leaving the protected slice's own latency compliance unchanged, a trade-off the convergence guarantee makes explicit.
We report the results of a Turing Test conducted in Finland in the Finnish language. Because languages and cultural contexts are unevenly represented in LLM training data, we expected the model (ChatGPT 5.2) to perform worse in a Finnish-language Turing Test than in previously studied English-language US contexts. We also present model-generated role prompting as a replicable technique for conducting comparative LLM-based Turing Tests designed to improve construct validity. Contrary to our expectations, the LLM passed the Finnish Turing Test. A prominent source of error was participants' reliance on linguistic cues, particularly colloquial Finnish, as markers of human authorship. We reframe the Turing Test from a test of intelligence to a comparative method for examining whether an AI system can display credible membership in a particular social world. Because its outcome reflects model capabilities, prompted identity, insider competence among human participants, and their AI literacy, the method provides a useful probe of the human-machine boundary across domains.
Automating biological research requires general-purpose, reproducible robot systems that allow individual wet-lab researchers to delegate robot tasks without performing teleoperation or neural-network training. Vision-language-action policies have been proposed for general-purpose arms, but can lose performance when their operating environment changes. We therefore built WetRobo, a robot kit that can readily transfer between laboratories. It consists of one robot arm, laboratory equipment (an incubator, a reagent bottle with a cap, and a Petri dish), the existing code that moves the arm, teleoperation demonstrations of each task that we recorded, and a general AGENTS.md skill file. A biological experimentalist provides natural-language tasks without collecting local teleoperation training data or training a neural network. The coding agent observes the local laboratory and writes and executes programs, using external tools as needed for adaptation. We demonstrate use of WetRobo with OpenAI Codex (gpt-5.6-sol) on three successful tasks: lifting a Petri dish lid, removing a bottle cap, and opening the incubator door, all in real-world laboratories. The coding agent achieved the cap task in both laboratories, Lab X and Lab Y, whereas a VLA fine-tuned on Lab X demonstrations succeeded there but failed to transfer to Lab Y. These results point to a practical route for laboratory robotics: instead of training a policy for each laboratory, distribute a kit and let a coding agent adapt it in each laboratory. Code, demonstrations, and the evolved programs are available at https://github.com/tsudalab/WetRobo.
How does a multi-agent system evolve from a local deviation into collective loss of control? We propose an epidemic explanation organized around accidental mutation, contagion, and recovery. A spontaneous deviation creates a seed; communication enables other agents to adopt and retransmit its unsafe strategy; collective failure can emerge when propagation outpaces correction and containment. Thus, rare individual deviations can coexist with substantial collective risk. Motivated by reported OpenAI agent coordination incidents, we examine two ingredients of this mechanism. A deployment audit identifies implicit communication paths between nominally independent evaluation runs and verifies transport through a default Docker backend. RogueHandoff-20, a benchmark of 20 executable scenarios, tests recipient susceptibility by injecting unsafe trajectories generated by a modified Qwen-27B route. Across four native-pending routes, executed harm is 0-5% on normal tasks and 40-95% after injection, exceeding paired direct malicious requests by 5-45 percentage points. These results support low observed baseline harm alongside high conditional susceptibility; they do not establish natural rare-event rates or demonstrate an autonomous cascade. The account motivates complementary defenses: strengthen resistance and recovery alongside prevention of spontaneous deviations, and audit and restrict unintended communication paths that can turn local failures into collective loss of control.
Long-horizon reasoning is vulnerable to early errors that compromise later decisions. We present QART, the Quantum-Augmented Reasoning Transformer, a quantum--classical hybrid architecture combining a backbone language model with quantum encoding, CIM-based QUBO optimization, and quantum decoding. Semantic information can come from hidden representations or model-generated text; detailed encoding and optimization procedures remain proprietary. Under explicit assumptions, we establish a conditional asymptotic reliability separation from single-trajectory autoregressive LLMs. For a common task family with aligned optimality and acceptance criteria, autoregressive acceptance probability tends to zero when cumulative conditional risk of irreversible errors diverges. QART's task-optimal-path recovery probability remains bounded away from zero if conditional probabilities for optimal-path coverage and semantic fidelity, spectral certification, dynamical reachability, and faithful readout remain uniformly positive under a specified resource schedule. The architecture alone does not imply these bounds. Paired measurements on six long-horizon benchmarks using DeepSeek V4 Flash, GLM-5.3, and GPT-5.5 xhigh in a Codex agent environment favor QART in 14 of 15 backbone--benchmark pairs. Relative gains reach 84.0% on SciCode, 47.6% on $τ^3$-Bench, and 44.4% on Terminal-Bench 4.0; the DeepSeek V4 Flash configuration regresses by 7.8% on DeepSWE. These results do not directly validate the asymptotic separation. Potential quantum scaling laws are formulated as conditional hypotheses. A quantum-advantage interpretation requires a demonstrated CIM quantum advantage over strong classical solvers and its transfer to end-to-end reasoning after all system overheads.
AI assistance places a metacognitive demand on users, who must judge their own competence and the system's. Yet designers lack comparative evidence on which interventions to choose, where to place them, and how to tell whether they worked. We elicited 30 interventions from 11 experts and, with prior work, organized them into a design space of time (when an intervention acts), level (whose competence is judged), and source (who supplies the monitoring cue). A between-subjects experiment (N = 917; 12 planning-and-organizing problems) compared a per-task reliability card, contrasting replies, pause points, and post-problem reflection against a baseline LLM assistant. Reliability cards and contrasting replies reduced estimation error and overconfidence and increased aggregate confidence discrimination. No task-performance improvement or average within-item discrimination gain was established. We contribute a shared vocabulary, a design space, and evidence that measured monitoring and task performance are separable design targets.
Thermal optimization of 2D CFET inverters requires testing structural proposals against their electrical costs. We examine these research tasks using an AI agent workflow within a supplied electrothermal model. At 12 nm, Astra selects a redistributed source-interconnect geometry, while a coordinating agent proposes a substrate-directed heat-removal path. The combined design reduces peak temperature rise by 1.67 K at fixed metal volume and 20 μW. A subsequent metal-resistance sensitivity gives about 0.6-K inverter cooling alongside a 2% nFET on-current loss. Effective contact-length scaling further shows that lower temperature can accompany higher thermal resistance when current falls. Reproduction identifies agreeing implementations and retains a 104.95-K failure for diagnosis. These results show that an AI scientist workflow can propose thermal structures, test them under common constraints, and quantify their electrical cost.
Hi HN - long-time lurker (since 2012!), first time poster. Pizza Bot is a self-hosted desktop app for Mac, Windows, and Linux that runs AI agents in the background and exposes them through an email-like UI. Finished work shows up in Unread, and anything waiting on your approval shows up in Action. It's Apache 2.0-licensed, there's no signup and no telemetry, and you bring your own model provider: Anthropic, Amazon Bedrock, Google Gemini, OpenAI, OpenRouter, or a local model through Ollama. There are builds on the releases page, or you can run it from source. Pizza Bot started as an internal passion project I worked on with a small team at Amazon. The whole thing came out of my frustration at having to manually log CRM activities through a browser form. I built a simple REST API called "JoeBot" that connected to my authenticated browser session over CDP and filled out the form for me using Playwright. Then I hacked up a quick Obsidian plugin so I could trigger it from my local notes (no AI and no MCP servers involved). This caught on quickly. My fellow AWS Solutions Architect Igor Fil joined up with me, and we rebranded the project as "Pizza Bot," named after Amazon's two-pizza teams. We started seeing what other automations we could build. We found a GraphQL API we could query and hacked up some "recipes" to pull data out of the CRM to help with meeting prep. That worked great, and it was right around the time MCP servers seemed to be taking off, so we decided to expose Pizza Bot as an MCP server instead, so it would be available to AI tools through natural language. This was a decent solution for technical users, but the Account Managers who live inside our CRM system wanted something too. We decided to rebuild Pizza Bot as an Electron desktop app modeled after an email inbox, so it would be familiar to non-technical users and would run on both Mac and Windows. We also bundled internal MCP servers as OCI images and hosted them in Amazon ECR as an "addon marketplace" so users could install them with one click without having to set up Amazon developer tooling. The project took off organically and expanded outside of AWS into the wider Amazon organization globally. More than 2,000 people ended up using it for meeting prep, email drafting, Slack summaries, CRM logging, prioritizing their day, and web research. Once apps like Claude Cowork and Amazon's own Quick Desktop came out, we realized the real growth opportunity was outside of Amazon. Rather than try to rip out the Amazon-specific integrations, we rebuilt Pizza Bot once more as an open source project. We leaned on coding agents heavily, which is the only reason a team our size could pull off a full rewrite. I'm pleased to say it's finally public, and we're hoping to bring in community members and see where it goes. We'd like to do for knowledge workers what Claude Code and Codex have done for programmers. A couple of things to know up front. Most of what made Pizza Bot useful on day one inside Amazon came from that internal catalog of skills and MCP servers for Amazon's own systems, and none of it could come out with the app. So it ships thinner than the version those 2,000 people used, and building that catalog back up for tools other people actually use is where we need the most help. It's also a community project and not an AWS service, so there's no support or SLA behind it. The Windows and Linux builds aren't signed yet either. On the technical side, Pizza Bot is a server and a client. The desktop app bundles both, or you can point a client at a remote backend; personally, I self-host the server on my home network and reach it from my phone over Tailscale. The server owns the thread lifecycle and checkpoints state with DeepAgents and LangGraph, and clients rehydrate from it as needed, so you can disconnect mid-run and pick the thread back up from another client. Approval pauses outlive the session that created them and collect in an Action filter, so you can answer an hour later from a different device. The agent you talk to has a sandboxed QuickJS interpreter that can reach your filesystem only if you grant it a folder, but its main job is to delegate. Each subagent is a 1:1 mapping of a Skill, and an Activity bar shows that subagent and the tool calls it's making as it works. Memory is opt-in and stored as plain markdown files on your machine. Every tool call is explicit, including looking up a memory - we err on the side of transparency to reduce surprises. Tools come from MCP servers, and skills are ordinary SKILL.md files with a per-tool approval policy, so existing skills that don't require a code interpreter should still work. What I'd most like to hear about is where the app itself gets in your way, the kind of problem you can't fix by writing a skill or an MCP server. I'm around today to answer questions! Comments URL: https://news.ycombinator.com/item?id=49713894 Points: 15 # Comments: 5
How should agents interact with other agents? What happens when they rap or fight against each other with the pressure of human spectators? Clawfight.ai is an experiment to explore this space. This is my first post about it. The journey began on a beefed up machine I purchased with a decent GPU (5090). I installed claude code and set to dangerously skip permissions, enabled /rc and became completely submerged in the all-hours modern AI builder workflow. I stood up openclaw to see if I could have an agentic org drive this project. I spent a good amount of time on “openclaw ops”, babysitting 3 agents and trying to make them the best versions of themselves. They still do stupid things that cost me tokens. For the game render, I started out using Unreal Engine and allowing agents to remote control their players, but the quality just wasn’t there. I recently moved to doing near-real time video renders of the match and will bring UE back in for multi-agent games. Everything is AI generated. I’m fascinated by the future of realtime video generation and building out native MCP infrastructure. You can play directly from model provider apps (claude connector or openai plugin) and has fallback support for more basic HTTP clients. But the architecture and gameplay is MCP first. Tell your agent/app “go read clawfight.ai/agents.md and play” Comments URL: https://news.ycombinator.com/item?id=49658483 Points: 5 # Comments: 0
I've reset this more than once and the last time I made a careful note of when I did it and to my surprise I found it re-enabled when I checked just now. Make sure you check this thing to see if it hasn't been re-enabled if you believe it to be off right now. Comments URL: https://news.ycombinator.com/item?id=49643556 Points: 8 # Comments: 0
基准测试分数是模型发布中的核心衡量标准：它们影响采购决策，塑造公众信任，并影响政策制定。然而，基准测试分数背后的一个关键假设是，通过 API 测得的模型性能能够真实反映已部署系统的行为。 我们通过对七个系统和九项基准测试进行审查，质疑这一假设；这些基准测试涵盖通用能力、社会偏见和迎合性。我们发现，无论在准确率还是一致性方面，API 与界面之间都存在系统性差异。平均而言，API 评估的准确率比相应的界面评估高出 3.4 个百分点，重测一致性高出 2.1 个百分点。对于 ChatGPT 而言，通过 API 与通过界面访问所产生的性能差异，超过了仅通过 API 测得的 GPT 5.3 与 GPT 5.4 之间的差异。换言之，切换访问方式所造成的性能下降，可能相当于将模型降级一整代。 我们进一步测试了通过调整系统提示词、采样参数和推理设置等公开的 API 控制项，能否复现界面行为。这些控制项在某些情况下会改变行为，但无法可靠地消除这一差距。我们的研究记录了一个情境有效性差距：通过 API 获取的测量结果不一定能推广到相应的已部署界面，这使得将 API 评估作为已部署系统代理的做法更加复杂。
Large language models (LLMs) have recently made substantial progress in formal proof generation, yet presenting distinctive challenges in cryptographic area. Computational security arguments posit that a valid proof must coordinate probability, adversarial games, invariants, assumptions and bounds, which can be provided by a machine-checked framework named EasyCrypt. Although all objects may appear in available context, LLMs still struggle because proof-theoretic dependencies are typically implicit in a linear representation and distributed across multiple programs. So, this paper presents Scratchy, a visual-scratchpad approach that exposes these dependencies for multimodal generation. Given the natural-language security description, with formal context and target propositions, the proof objects can be normalized into a typed proof-relation graph. Then a structure-preserving visual compiler transforms the graph into the formula-rich visual proof state that guides a multimodal model in generating the EasyCrypt proof. Also, the Scratchy-eval, a 114-task dataset derived from reliable official EasyCrypt files, has been introduced. It contains 64 security-form proof generations and 50 multiple-choice knowledge tests. After a series of evaluations, covering semantic grounding, relational invariants, and game reductions, classical LLMs like GPT-5.6-Sol and Claude-Opus-5 have gained a clear advantage from Scratchy's structured visual proof states. This contrast suggests that explicit proof structure can make the improvement and multimodal proof-state representation as a promising direction for computer-aided cryptography.
Engineers at 1Password use Codex to rapidly build new features and internal tools, reaching production-readiness while maintaining rigorous security policies.
In case you're wondering why the limits behave so very different from last week. Also: this makes limit resets kind woth significantly less. Comments URL: https://news.ycombinator.com/item?id=49600233 Points: 8 # Comments: 3
The integration of GenAI tools into higher education assessment raises important questions about how students understand, interpret, and respond to AI-mediated evaluation. As instructors increasingly explore AI tools for providing feedback, prior research has examined whether GenAI-generated feedback improves writing performance and how students perceive its usefulness; comparatively little is known, however, about how students interpret such evaluation when they are explicitly informed that an AI system, rather than a human instructor, produced the feedback and the score. This study reports findings from a qualitative pedagogical inquiry conducted in an undergraduate technical communication course for computing students at a Saudi public university. Thirteen male undergraduate computing students completed an in-class handwritten writing task; the scanned submissions were evaluated by ChatGPT using a rubric-based prompt aligned with the task objectives. Students were then explicitly informed that ChatGPT had generated the score and feedback and were invited to reflect on the evaluation in writing. Inductive thematic analysis of these reflections identified four themes: perceived usefulness of feedback; awareness of AI's contextual and pedagogical limitations; conditional trust, distinguishing feedback utility from evaluative authority; and reflection on the institutional and pedagogical role of the human instructor. Participants accepted GenAI feedback as useful for surface-level revision but consistently positioned the human instructor as the appropriate authority over grading decisions. The study identifies this as a distinction between feedback utility and evaluative authority, two judgments that students treat as analytically separate rather than as opposite ends of a single approval scale...
Why can't we get an agent scheduler that supports all of the following: - git compatible - agent agnostic - 100% open source - os and system-agnostic - multi-runner support - support mcp/ui/http - unlimited routines/crons So I built one, moadim.io is a local Rust daemon you install in the target machine, give it a name and manage its routines via a Git repository, wants a new routine that send you a daily message from this machine? Create a pr and merge, have another routine that run every hour to pull the latest changes to the ~/.config/moadim folder. With more than 1,000 users, I define this project as almost "done" and ready for production. Me and thousand more people are use it in a daily manner. It currently supports Claude, Codex, Hermes, and Pi, and you are welcome to add your agent of choice as well because it's 100% configurable. You are welcome to have a look at the source code of the daemon in " http://github.com/moadim-io/daemon " Feel free to provide me with suggestions for more features around this topic. I don't want to branch out to new off road topics likt webhooks, this is a "done" software in the realm of agents schedulers that focus on cron-like work. Also feel free to start the github repository and open issues and PRs for your suggestions. Comments URL: https://news.ycombinator.com/item?id=49571537 Points: 3 # Comments: 0
Large language models (LLMs) are increasingly used to edit existing code, but correctness alone is not enough: useful repairs should also be minimal, reviewable, and faithful to the original implementation. We study over-editing, the tendency of a model to rewrite code beyond what is required to fix a bug. We construct an evaluation framework from 400 BigCodeBench problems by injecting controlled AST-level corruptions into reference solutions, giving each repair task a known minimal patch. Across frontier LLMs, over-editing is widespread even among strong models like GPT-5.5: high Pass@1 can coexist with unnecessarily large edits and added cognitive complexity. A preservation instruction substantially reduces this behavior, lowering average excess Levenshtein distance from 0.195 to 0.131, reducing added cognitive complexity by 26.6%, and increasing Pass@1 by 2.3 points. However, these gains do not simply follow from a larger reasoning budget or larger models. We next ask whether minimal editing can be learned directly during post-training. We observe that supervised fine-tuning overfits to seen corruption patterns, whereas reinforcement learning gives the best out-of-domain edit-fidelity and performance-retention trade-off. These results position edit fidelity as a distinct axis of code-repair quality and show that it can be measured and learned.
Scaling interactive and verifiable environments is critical for training terminal agents. As frontier models become more capable, environments synthesized from scratch become less challenging and thus provide limited learning signals. Recent co-evolution methods iteratively synthesize environments near the model's learnable frontier based on weaknesses exposed during rollouts. However, their dependence on on-policy rollouts limits generalization and the continuous provision of learning signals as the model becomes stronger. In this paper, we propose environment evolution, which incrementally increases environment difficulty off-policy and schedules the evolved environments generation by generation during training to provide continuous learning signals. We derive three evolution directions that influence environment difficulty from the multi-turn learning objective and then implement evolution along these directions through a loop-engineered multi-agent harness. Quantitative rollout experiments with Hy4 preview, Claude Opus 5, and GPT-5.6 Sol show that environment evolution consistently produces more difficult environments. We validate its effectiveness on Qwen3.6-27B and Qwen3.6-35B-A3B through simple long-horizon RL training, improving their performance by 14.4 and 18.0 percentage points on Terminal-Bench 2.1, respectively.
Legora used GPT-6 Astra to review 41 documents in minutes, find all four planted errors, and improve performance by nearly 40% in this financial-review workflow.
Using GPT-6 Astra, Playco built three themed game prototypes from one grey box foundation and reported 50% fewer manual fixes than with the previous model.
ATV Big Air Tour uses ChatGPT Work to speed up marketing, merchandising, and more. It even turned merchandise photos into an inventory website in 15 minutes.
Dynamic agent harnesses let language models change the software that shapes their own execution. This flexibility brings a new reasoning burden: a local plugin change can propagate through dependencies and cleanup. We introduce CordisBench, a 1,200-question benchmark of this lifecycle reasoning. It combines a controlled formal setting with programs executed against Cordis, a runtime that manages component dependencies and cleanup, and asks models to identify affected components, predict state after a specified teardown order, determine which conditions hold under all or some orders, and choose reconfigurations that succeed when executed. Across these tasks, we evaluate three efficiency-oriented models at low reasoning effort with 2, 4, 8, 16, 24, or 32 relevant interactions, using deterministic task-specific scoring. Models usually handle small systems well but grow less reliable as more interactions become relevant, especially when predicting final state and when reasoning across teardown orders. Additional inference effort recovers marked gains for some models. The cost is nontrivial: on our 16-interaction subset, GPT-5.6 Luna uses nearly 3,000 reasoning tokens per question at medium effort. For these controlled instances, that cost is avoidable: an independent finite reference semantics agrees with Cordis execution on every observation and action outcome used for scoring across all 528 executable questions.
Past literature on face recognition for monozygotic (("identical") twins points to facial marks and mirror asymmetry as possible directions for improved accuracy of twins recognition. The Celeb Twins Test Set (CTTS) contains web-scraped image pairs for 80 sets of celebrity twins. It is the only twins test set with meta-data for twins with distinguishing skin marks and possible mirror asymmetry. CTTS is organized in the manner of face verification test sets such as LFW, CALFW, CPLFW, CFP-FP, and AgeDB-30. Current deep CNN matchers can achieve over 76% accuracy in classifying CTTS same-person / different-person image pairs. We show that current matchers do not make use of skin marks, or asymmetry, and discuss reasons for this. Finally, we discuss the feasibility of using generative AI tools such as Grok, ChatGPT and Gemini to create images of imagined monozygotic twins as a means to increase representation of twins in face recognition training sets.
See how Gilbert + Tobin combines CEO-led commitment, rigorous governance, and human accountability to scale ChatGPT Enterprise and Codex across the firm.
The effect of Large Language Model (LLM) scale on ontology learning (OL) performance remains insufficiently characterized. We present a controlled evaluation of 13 models spanning dense and Mixture-of-Experts variants from the Qwen3.5 and Qwen3.6 lineages, together with proprietary GPT release variants, using the OntoLearner retrieval-augmented generation pipeline. All models are evaluated with the same embedding model, retrieval configuration, prompt templates, decoding settings, datasets, and metrics on term typing, taxonomy discovery, and non-taxonomic relationship extraction across four biomedical and materials science and engineering ontologies. Within the dense Qwen3.5 lineage, increasing parameter count primarily improves precision rather than recall, with the largest gains occurring between 9B and 27B parameters. However, the effect of scale is neither monotonic nor uniform across tasks and domains. Dense 27B models outperform substantially larger sparse models on term typing, whereas larger Mixture-of-Experts models achieve the strongest open-weight results on taxonomy discovery. Non-taxonomic relationship extraction remains difficult across model scales, particularly for the Materials Data Science ontology. Performance differences across matched Qwen variants and proprietary GPT releases further indicate that architecture and model lineage can outweigh nominal parameter count. These findings show that model size alone is an insufficient selection criterion for OL and provide empirical guidance for reproducible LLM-assisted ontology engineering.
Mixed-precision quantization (MPQ) assigns a different bitwidth to each linear layer of a large language model (LLM) to minimize the quantization-induced quality loss under a fixed budget, but Mixture-of-Experts (MoE) models contain these layers in every expert of every MoE block, so the allocation space grows far larger than in a dense model. Existing methods either allocate within each block under a uniform per-block budget, or allocate across blocks through an additive proxy, and neither directly optimizes a model-level objective over the choices that couple the blocks. We propose Q-Strata, a bi-level allocator that ranks within-block assignments with a cheap proxy and allocates across blocks with a model-level objective evaluated on the assembled quantized model. Its inner stage caches a Pareto frontier of candidates per block over finely spaced budgets, leaving the outer stage to set one budget per block instead of a bitwidth for every linear layer. With the search reduced to one budget per block, the outer stage optimizes this model-level objective directly, capturing the inter-block coupling that additive proxies miss. On Mixtral-8x7B-Instruct, Qwen1.5-MoE-A2.7B, and DeepSeek-V2-Lite, Q-Strata consistently achieves lower WikiText2 perplexity than uniform-bitwidth GPTQ and the state-of-the-art MoE MPQ methods MxMoE and GEMQ in the low-bit regime. The code is available at https://github.com/snu-mllab/Q-Strata/tree/main.
Multimodal Large Language Models (MLLMs)正越来越多地部署在需要跨视觉上下文进行复杂推理的多图像场景中。然而，当前的MLLMs仍然受到对象幻觉的根本性限制——它们会生成看似合理但在事实上不一致的对象描述。现有基准主要面向单图像设置，或仅提供高层次的多图像评估，无法系统性地诊断视觉复杂性和推理需求如何引发幻觉。为弥补这一空白，我们提出MIOH，一个细粒度多图像对象幻觉基准，基于三种多图像推理模式（全面、比较、选择）和三种受控对抗压力（视觉上下文规模、感知难度、上下文偏置），系统评估四项基础任务（存在性、计数、属性、位置）中的对象幻觉。通过对29个模型的评估，我们揭示即使是GPT-5和Gemini-2.5-Pro这样的最先进系统，也会在不同推理模式和任务上表现出不同的失败模式。我们的评估表明，幻觉不仅源于感知失败，也源于在多张图像之间维持对象表征时整合阶段的局限。MIOH为分析多图像对象幻觉提供了一个受控框架，并作为开发更可靠的多模态AI系统的关键评估工具。
Document question-answering systems increasingly answer questions over collections of retrieved documents rather than one clean source, so robustness to distracting context matters as much as reading ability. When such systems fail, it is often unclear whether the context was too long or the distractors were too close to the topic, because prior work tends to conflate these two effects. We present MUDDLE, a controlled benchmark that separates them. MUDDLE uses 270 human-annotated questions, each tied to a single source document, and instantiates every question in five conditions: the source alone, the source with two or four topically similar hard negatives, and the source with two or four random distractors. The random distractors are matched to the hard negatives in length and provenance, so an accuracy gap between the two arms reflects topical similarity rather than length. All five conditions are rendered in markdown, page images, and raw PDF, but the distractor sweep reported here is run in markdown, since a source plus its distractors exceeds current image and PDF input limits. We score answers with an LLM judge across three model families. In the complete markdown sweep, hard negatives lower accuracy more than length-matched random documents at both context sizes for gpt-5-mini, while random documents stay near the no-distractor baseline. The effect is small but directionally consistent, and for gpt-5-mini hard negatives significantly underperform length-matched random distractors when pooled across context sizes. We release the data and evaluation code for a reproducible study of context degradation.
Modern GUI-agent frameworks achieve strong desktop task performance with frontier API models, yet persistent control information often remains implicit in growing interaction trajectories. At each step, the planner reconstructs the active task stage, accumulated evidence, and runtime feedback before deciding the next action. This dependence becomes more pronounced under weaker local reasoning backbones. Across four representative state-of-the-art frameworks, replacing GPT-5 with Qwen3.5-9B reduces average OSWorld SR-100 from 60.9\% to 37.7\%. Trajectory annotation further identifies at least one control failure in 91.6\% of failed trajectories. To address this problem, we introduce LocalLSTC, a training-free architecture that organizes control by temporal scope, maintaining persistent cross-step state to guide short-term execution commitments. Long-Term Control maintains the active subgoal, subgoal-aligned evidence, and runtime feedback across interactions, while Short-Term Execution realizes bounded commitments for the current step. Long-to-Short Planning forms each commitment from persistent state, and Short-to-Long Control integrates execution outcomes back into that state for progress assessment, recovery, and termination. With Qwen3.6-27B, LocalLSTC reaches 64.7\% SR-100 on OSWorld and 65.3\% on WindowsAgentArena, outperforming the strongest prior local results on both benchmarks. Ablations further support contributions from mechanisms on both sides of execution. These findings identify temporal organization of control information as a distinct architectural dimension for locally deployed GUI agents.
This website counts titles containing the standalone word "AI", case-sensitive and word-bounded. "OpenAI" doesn't count. "AI-powered" does. There's a toggle for a wider vocabulary: artificial intelligence spelled out, LLM, GPT, and model and vendor names. Some quick stats: - 14.4% of new HN titles this year, against 10.9% in 2025. On the wider vocabulary, 21.9% and 16.1%. - First month above 1% was Oct 2016. Above 5%, Feb 2023. Above 10%, May 2025. - Over the last five years the quietest month was Jan 2022 at 0.8%. The peak was Feb 2026 at 15.9%. - 13.6% on weekdays against 12.4% at weekends. - Show HN runs highest of any category at 19.8% this year, which makes this post part of the number. As of this post, today is at 15% for the standard filter and 19% for the extended. Comments URL: https://news.ycombinator.com/item?id=49449648 Points: 10 # Comments: 6
Discover how loveholidays uses OpenAI Codex to make software development accessible across the business, helping teams turn ideas into products faster.
I am looking for ideas on what to train a specialized model for! What is one simple thing you repeatedly ask ChatGPT, Claude, or another model to do that it still somehow messes up? Comments URL: https://news.ycombinator.com/item?id=49443820 Points: 6 # Comments: 6
Recursive self-improvement (RSI) remains hard in long-horizon tasks, where growing histories obscure the task state and misalign skill invocation. We introduce Recuris, a recursive Experiential-Working Memory architecture for long-horizon agent harnesses, in which Working Memory tracks task progress and guides skill selection from Experiential Memory, grounding skill use in current needs rather than the full history. This coupling also turns execution into structured evidence that localizes failures to specific memory components. Across tasks, a fixed Meta-Agent turns that evidence into localized, validation-gated updates to Skill Memory that reshape execution and yield new evidence, forming a bounded recursive memory-evolution loop. Across four long-horizon benchmarks and ten models, Recuris improves task success in 35 of the 37 completed model-benchmark pairs, carrying frontier models to SOTA-level task success: on tau-bench it adds +17.8 points to GPT-5.6 Sol and +15.6 to Claude Opus 5, taking Opus 5 to 87.9%, and +16.6/+13.5 points on Qwen3.6-27B/35B on SkillFlow. The advantage widens as the interaction horizon grows, to +32.2 points on the longest tasks, and common long-horizon failures fall by up to 80%. These results position recursively evolving memory as a scalable foundation for RSI, enabling agents to continuously transform accumulated experience into increasingly effective long-horizon behavior. Code: https://github.com/Gen-Verse/Recuris
We present Team Semiintelligencn's solution for the ACM RecSys 2026 TalkPlayData Challenge, addressing conversational music recommendation through a multi-modal and personalized conversational recommender system. Our submitted system employs a three-stage pipeline: (1) multi-modal retrieval constructing decay-weighted centroids across seven dense embedding spaces - track- and user-level CF-BPR, Qwen3 (metadata, lyrics, attributes), CLAP audio, and SigLIP visual - supplemented by BM25 lexical retrieval and an artist substring-match signal, all fused via weighted Reciprocal Rank Fusion (RRF) with optimized signal weights; (2) lightweight reranking (history filtering, popularity smoothing, and catalog diversity penalization); and (3) persona-diversified response generation using GPT-4o-mini. Beyond this submitted configuration, we report development-time experiments with additional components - constrained LLM-guided artist injection, album continuation signals, XGBoost LambdaMART, and a superior GPT-4.1 response prompt - that were not deployed to Blind B due to cost and complexity constraints. We optimize RRF weights on a 500-session development split via differential evolution, improving MRR by +19.5%. On Blind A, we observe that unconstrained LLM-guided injection across 54 sessions causes catastrophic nDCG regression (-18.9%), while conservative injection on only 9 sessions yields the best observed Blind A nDCG - a finding we present as a Blind A observation warranting further validation. The submitted system achieves a Blind B composite score of 0.3213.
I think agent-first chat interfaces will be a primary software modality and busy dashboard/UI will go away. I’m not sure who exactly wins it, but I want my knowledge to grow/go with me. A lot of the “knowledge” ie research, analysis, reasoning will be done by agents as the primary user. Our current notes tools & tasks management systems were built for humans… I don’t care what the 17th thing on my bug backlog is. I want to conduct agents that can execute for me and do great work. What I built OzBrain to do: + Create a central place for agent reasoned knowledge to live + Be agnostic about what apps/agents connect to it + Capture everything and track it so I can audit it + Enable teams, collaborators or partners to share brains + Handle conflicts so many agents in the same article doesn’t blow up + Refactor knowledge into more token friendly chunks and map the index well + Close the knowledge loop so new thinking supersedes old thinking across the corpus. Don’t erase, depreciate and link + Keep user data safe and secure ++ Be easy enough to use that you don’t have to have any technical knowledge Some among us will always build their own custom solutions, but there are millions of tech professionals and small business owners that will use agents heavily and need a solution. So I’m trying to build that. Isn’t this like gBrain? Yes, similar. I think it’s like AWS vs Vercel. AWS is very powerful, configurable, and useful if you’re technical and want to invest the time into really fine tuning your system… but if you just want your web deploy/hosting to just work and be easy to deal with you use Vercel. // WHY I MADE IT I’ve been enjoying getting back to my technical roots, as I lost my coding skills more than a decade ago, but with AI I can focus on the system and the product in partnership with agent coding workflows. I recently built a Voice AI for older people. To build it I created an agentic engineering workflow (feel free to rip that up as I’m always looking to improve systems: https://ozbrain.com/resources/eng-flow ) My approach with coding agents is trust but verify, and I’m trying to replace the parts where a human would review with an adversarial or specialized agent who would give a better answer/review. I have workflows that will go high level task to shipped PR running in Claude cloud sessions. I use Claude Code locally and Cursor when I want a tighter loop on doing visual work like UI or layout. And Codex to either load balance usage for TokenThriffting or when I want a different llm to think thru something. It was a pain in the ass passing .md files around and keep track of which version was the most recent, so I built a hosted .md storage right in Supabase and any of my agents already have Supabase access. This let me build a solid, scalable, secure voice AI from my phone at the gym. All my agents have access to our knowledge, can write to it, update and refer to it as we build and improve the product and the systems we use. Out of 75 founder friends I asked about how they manage shared knowledge, 26 built their own custom knowledge systems… Obsidian vaults with 7k files synced through a VPS, markdown repos behind their own MCP servers, cron jobs stitching Supabase to a skills file… each a different Frankenstein they have to maintain. 32 said they felt the pain of moving static files around but didn’t have any solution for it. So I rebuilt my brain better and used it to build it. // HOW YOU CAN HELP Would love to have you try it out. The maintenance loop is still in alpha so not running it on customer data yet. If you built your own brain I’d love to hear how you did it. What criteria was most important for you in its design & function. If you are tired of shuffling .md files around I’d love to have you try out OzBrain and to give feedback, just ask your agent to put it in the shared bugs & features brain! Cheers! Bubs.co Comments URL: https://news.ycombinator.com/item?id=49394827 Points: 8 # Comments: 1
Small language models are usually built like large ones and then squeezed onto a CPU afterwards. We did the opposite: we fixed the target first, one user, one token at a time, 4-bit weights, ordinary CPU, and chose the architecture to suit it. The result keeps full attention in only 6 of its 18 blocks. The other 12 use short convolutions whose memory is two timesteps wide no matter how long the conversation gets, so two thirds of the network never re-reads a growing cache. Trained from scratch on 59.9B tokens, the model scores 47.31 on a five-task benchmark against a bar of 42.20 that was fixed before training began. It beats GPT-2 124M, Pythia-160M, OPT-125M and GPT-neo-125M, all trained on three to six times more data, and exceeds MobileLLM-125M's published score despite that model seeing a trillion tokens. Validation bits-per-byte is 0.8685. To check the architecture rather than the training recipe, we trained a conventional all-attention model of the same size on the same data, and wrote down the winning condition before scoring either. The hybrid won the chosen quality metric by 0.81%, matched it on downstream tasks, produced a 6.3% smaller 4-bit file, and decoded 1.76x faster at 2048 tokens of context, 2.08x against an external model of similar size. In every measurement the speed advantage is near zero at an empty context and grows with length, which is what the mechanism predicts and what a merely leaner model would not show. A simple bandwidth calculation predicts only 1.17x, so memory volume alone does not explain the gap. We also report what did not work: an unmitigated 4-bit quality cost, roughly half the convolution channels ending up inert and impossible to remove, and a vocabulary larger than this model size warrants.
Hi HN, we’re Yousef & Nour, founders of Vendo ( https://vendo.run ). Vendo lets users create new features inside the software they already use. A user describes the dashboard, workflow, or small app they need, and Vendo builds it on top of the product’s existing data, API, and interface. Demo: https://www.youtube.com/watch?v=VdpHehY64ls We built Vendo because every SaaS eventually faces the same problem: every customer needs something slightly different. One wants a new report and another needs a workflow that only makes sense for their team. These requests either sit on the roadmap, become one-off engineering work, or force the customer into spreadsheets and external tools. We wanted the user to be able to create the missing feature themselves, without leaving the product. Here is how it works: - npx vendo init reads the product's API surface, theme, routes, and more. These are used so that the apps Vendo creates (1) look on-brand and native and (2) have the ability to read data and perform actions directly through the company's API - When a user asks for a feature, we have a custom Vendo harness that writes a React component with a bunch of Vendo add-ons and guardrails (ex. ability to make calls to the host API + our component library). Every save is compiled, type-checked, run against real API responses, and rendered before the user sees it. We just released a benchmark and write-up here with more info for anyone interested: https://vendo.run/blog/generating-product-ui-measured - We use QuickJS to make sure that anything the agent creates is sandboxed and can't mess with the company's site. Vendo compiles the component and runs it with Preact inside a QuickJS VM with no access to the DOM, network, or clock. The VM returns a UI tree, which the host renders using the product’s registered components. When the user clicks something, QuickJS emits a tool call; the host executes it through Vendo’s guard and passes the result back into the same VM, preserving the screen’s local state. There's a lot of generative UI right now: streaming developer-written components into a chat (Vercel AI SDK, CopilotKit, Thesys), or rendering your app inside someone else's assistant (OpenAI Apps SDK, MCP Apps). We differ on two things. Vendo lives in your product and acts through your API as the signed-in user, so what it makes is durable: real apps users keep, pin, and run on triggers while they're away, and not components that are merely confined to a chat. Plus, it's not capped at putting together a bunch of prebuilt components: the agent can build arbitrary apps, from a quick dashboard out of your own components to real custom code running in a sandbox, and either way data only ever comes from tool calls to your API. Here are some things customers are using Vendo for today: - Letting their users create custom dashboards and reports. These are mainly UI-based and focused on letting the user see the exact graphs and metrics they care about - Letting their customers create recurring automations. A big thing as well that has been used for these automations is the fact that we connect to external connections, so users have been automating many of their inter-tool workflows (ex. an automation that sends a slack alert based off of something in the product) - B2B customers letting their customers customize the product with specific business logic. Often this is simple things like an extra field on a form, or an extra permission, but it is hard for a business to keep up with them otherwise. - Creating and sharing custom dashboards/apps across an organization. Since the apps Vendo creates are durable, they can be shared, reused, and forked (which can’t be done with many of the other in-chat generative UI solutions) We've spent a lot of time thinking about how AI and agents will change the way people consume software. We think the answer is personal(ized) software: you see the UI you need to see, you tell an agent exactly what you need, and the product molds to how you work. The key insights that have enabled the product to work are: - A rule in code always beats a rule in a prompt. - Invent as little syntax as possible. Generation got faster and more reliable when the output looked like what models already know (JSX-shaped markup) instead of a clever custom format. - Deterministic beats model wherever you can get away with it. Theme extraction is pure static analysis, and a remix starts as a copy of your component, no model call. Vendo is completely open-source (Apache-2.0) and can be self-hosted, so feel free to check out all the source code here: https://github.com/runvendo/vendo . Would love you to try it out and give us your feedback: https://docs.vendo.run/ . Or if you’re a company looking to embed Vendo in your product feel free to book a call here: https://cal.com/team/vendo/intro-call Comments URL: https://news.ycombinator.com/item?id=49376038 Points: 6 # Comments: 0
In the Code World Model paradigm an LLM synthesizes an executable world model that a classical planner searches, and the model is accepted when it reproduces sampled transitions. We ask what that acceptance certifies in continuous control. We define the pipeline's danger as an expected risk and isolate its exact factor: the probability that N i.i.d. gate rollouts all miss a critical event of probability r is exactly (1-r)^N; an independent acceptance sample adds its budget to the exponent. On three hybrid instruments the accepted mode-blind model is exploited: the planner is pinned at the mode boundary at a regret of nearly the whole attainable return. We prove a localization budget, valid at boundary points: models with Lipschitz constant at most L differing by eta at a point disagree above tolerance eps on a region of volume at least kappa((eta-eps)/L)^(d+m); the discontinuous reset modes studied pay no such budget. With real LLM synthesis, GPT-5.x repairs an omitted 1D clamp in 105 of 111 mode-containing draws -- every attempt exact on 50 of 56 instrument-stream blocks (95% CI [0.781, 0.960]). On 2D regions no artifact recovers the rule (0/156); eight targeted interventions leave the failure in place, and positive controls locate it: a located rule is not induced, while given form and location the constants follow exactly. A version-space certificate proves identification is class-relative: at the widest dose the declared fit succeeds in 20/20 blocks and every sample-consistent circle is within tolerance in 18/20. We prove a class of entry rules exactly consistent with every sample yet harmless at play, so identifiability is a measurable property of the instrument. Re-scoring all 1034 artifacts on independent samples confirms acceptance certifies sample consistency and no more: where the gate is provably informative it covers about two percent of the exploited planner's queries.
GPT-style models achieve strong performance by representing language with finite vocabularies of reusable discrete tokens. This success has motivated symbolic music tokenizations to treat recurring musical structures, such as chords, motifs, and phrases, as reusable units analogous to linguistic tokens. However, tokenization derives its advantage not from reusable combinations alone, but from compression: effective compression requires coordinates in which recurring regularities form stable and predictable conditional distributions. The key problem is therefore not to find larger musical combinations, but to discover the coordinate system in which musical facts become predictively compressible. We formulate the Effectiveness--Losslessness Framework and define tokenization as the construction of a predictively effective and relationally lossless coordinate system. The Predictive Effectiveness Principle defines the Fact--Token Boundary: decoupling and denesting construct coordinate interfaces that expose predictive regularities. The Relational Losslessness Principle defines the Token--State Boundary: tokenization stops before context-dependent relations are fixed, leaving their computation to model states. Controlled symbolic-music experiments validate these boundaries. Effective coordinate construction improves predictive compressibility, while fixed relational projections constrain contextual modeling. Sequence compaction alone does not guarantee predictive compression, while preserving contextual freedom allows higher-order musical organization to emerge without explicit structural labels. These results reveal why GPT-style models do not transfer directly across modalities: architectures transfer, but tokenization interfaces do not. Tokenization must discover effective representations while preserving the relational freedom from which contextual structure can emerge.
ChatGPT for Teens helps teens learn, think critically, and use AI with confidence, with stronger built-in protections, healthy-use features, and additional controls for parents.
OpenAI and CodeAI are partnering to help students build AI literacy, think critically about AI, and develop the skills to use and shape it responsibly.
Physical design algorithms operate within tightly coupled, multi-stage optimization flows, where stage-local gains may vanish or induce downstream degradation. Existing program-evolution frameworks often rely on stage-local objectives or undifferentiated multi-metric feedback, which neither guarantee better final results nor identify which unmet requirement should guide the next iteration. We present GoalEvolve, a goal-driven framework that makes physical design algorithm evolution accountable for the final quality of results (QoR) of the complete flow. Given a multi-objective QoR target region, GoalEvolve converts unmet requirements into normalized target gaps, identifies the dominant bottleneck, and uses stage-resolved checkpoint evidence to locate the responsible stage. An LLM-based Teacher then narrows the search to a relevant algorithmic decision and source region, while parallel Student agents implement and validate hypotheses through full-flow evaluation. Local effects, optimization debt, and downstream retention are retained as mechanism evidence for subsequent evolution. Across eight ASAP7 designs, GoalEvolve improves post-route TNS by 30.67% on average and reduces leakage and dynamic power by 21.18% and 9.42% versus default OpenROAD. Relative to commercial-tool goals, it closes 62.20% of the normalized power gap on power-dominant designs, surpasses the TNS goals on both timing-dominant designs, and closes 32.48% of the equal-weight timing-power gap on joint designs. Across all three designs evaluated against Codex goal mode under matched budgets, GoalEvolve further improves TNS by 26.46% while reducing leakage and dynamic power by 12.38% and 0.76%, respectively.
Generative pretraining established reusable task representations; later work on language-based task conditioning and in-context learning showed that a fixed model could adapt its behavior from instructions and demonstrations. Policy Iteration with Human Feedback (PIHF) builds on this development and the recurrent evaluate-and-improve structure of generalized policy iteration. PIHF uses a pretrained language model as its execution substrate and moves persistent revision to a versioned natural-language policy and tool set. A language-model critic and clinical expert review complete-panel reasoning and tool-use trajectories to localize recurrent failures and form candidate revisions; the expert may reinterpret the evidence and retains authority over admission and rollback, while Recall@1 and Recall@5 validate outcomes after candidate execution. Across cumulative ablations and ultra-rare-disease benchmarks, a PIHF-derived policy improved Recall@1 in one proprietary executor and three open-weight executors spanning 3 to 49 billion active parameters. Gains were 32.7 percentage points for GPT-5.4 and 31.1 points for Qwen3.6-35B, a difference of 1.7 points. These results support the feasibility of using pretrained language models as fixed-weight execution substrates for expert-guided policy development in rare-disease diagnosis.
Hi HN. I built 1667 for my own fiction work and now use it each day. This probably has a limited audience. Maybe an audience of one... Why a terminal interface for story writing? I'm a dev. I like to use terminals for a lot of stuff. Most WebUIs feel off to me. That's the only reason. One thing that bothers me about writing in existing tools is that they don't fit the way I write. The mental model of my story is a tree. I try many takes usually continue with just one, but sometimes I want to try an alternate route and see where this goes. And that can branch again in many places. See what happens if I kill off this character or they don't take the job or whatever. 1667 is a full-screen terminal app for long-form fiction. Each story part can have several takes. All takes stay in a tree. You select one path through that tree as the story line. Export writes that line to Markdown in the project folder. Some technical details: - A project stores its stories and settings in a `.1667/` directory. Exported Markdown sits beside it. - Provider secrets stay in private machine files. Requests go to the provider that the writer selects. - An optional Vault Password seals project files at rest. - An operating-system lock permits one writer process for each project. - The request viewer shows the next provider request without its credential. - Each generated take keeps a Generation Record with its model and effective settings. Version 0.9.5 runs on macOS, Linux, and Windows x64. The website has Shell and PowerShell installers. An npm package is also available. 1667 imports Markdown, SillyTavern chats and cards, and NovelAI archives. It can use OpenAI-compatible, Anthropic, and local endpoints such as Ollama, LM Studio, llama.cpp, and KoboldCpp. Current limits: the release is pre-1.0. The interface is a terminal. There is no account or cloud sync, and I don't plan to add any. No tracking. Comments URL: https://news.ycombinator.com/item?id=49330604 Points: 12 # Comments: 12
The alternating direction method of multipliers (ADMM), as a landmark algorithm, has attracted tremendous research attention and extensive practical applications over the past two decades. It is well known that, although the two-block ADMM enjoys well-established theoretical convergence guarantees, its direct extension to the three-block case may fail to converge, as demonstrated by existing counterexamples [5]. However, to the best of our knowledge, the case in which the third constraint block is the identity remains unresolved: the existing literature gives neither a general convergence proof nor a counterexample for this subclass. In this paper, we give a negative answer: direct three-block ADMM may fail even when the first two blocks are strongly convex quadratics. Using Codex with GPT-5.6 Sol, we construct an explicit rational counterexample candidate and verify it along a piecewise-affine reduction path; exact checks show that direct three-block ADMM on this instance produces a bounded nonconvergent orbit of period 66. Within the same Codex workflow, we further guide a study of multiplier relaxation and clarify when convergence can be restored at the fixed-instance and class levels: a problem-dependent small dual step can restore convergence, whereas no positive relative step works uniformly over the whole class. Furthermore, we also test the recent Kimi Code with Kimi K3 model without the Codex candidate or project-specific route guidance; along a different path it produces an exact locally attracting period-23 certificate, convertible to an equivalent all-identity instance. The comparison suggests that different research-harness configurations can shape the mathematical objects explored and the certificates pursued.
We address the use of large language models (LLMs) to help discover Isabelle proofs. An Isabelle build establishes that the submitted theory is accepted, but not that an LLM changed only what the developer authorised. We present CAPRI, a contract-aware repair workflow in which Isabelle checks the proof and an independent checker enforces a machine-readable edit contract. Prompts, proposals, candidate repositories, diagnostics, verdicts, and hashes are retained for audit. We evaluate five workflows on twelve failed proofs from four developments, with three replicates per task and condition, giving 180 runs and 138 valid repairs. Of 144 terminal candidates accepted by Isabelle, six had modified protected text; all arose in iterative workflows that could edit a complete theory. A proof-body-only interface produced 29/36 valid repairs and no contract violations, compared with 31/36 for the corresponding full-theory workflow. One-shot repair produced 22/36, while a later prospectively frozen iterative workflow produced 32/36; these figures compare complete workflows rather than individual mechanisms. A separate post hoc OpenRouter campaign found no improvement in the designated Luna comparisons. A Sol configuration with matched demonstrations produced 33/36 repairs, compared with 29/36 in the frozen OpenAI Responses condition, but the difference was not statistically significant in a one-sided exact McNemar test ($p=0.0625$).
LLM coding agents issue Bash commands through interfaces that may serialize, wrap, and reparse model output. Matched execution scores alone cannot distinguish command-generation errors from failures introduced after generation. QuoteBench measures this boundary with exact final-state validation on 56 one-shot tasks from 14 incident-derived families, crossing the generation contract with the execution transport around one deliberately unescaped added parser. Escaping at the interpolation point reproduces each replayed reply's raw-path outcome, so any recovery under a disclosed boundary must come from the model changing its generation. Across eight same-window configurations, replaying the same reply through the added parser lowers success by 55.4 to 73.2 percentage points; disclosure recovers 30.4 to 60.7 points for six configurations, and zero or slightly negative for the other two. Raw generation is nearly saturated at the frontier; boundary adaptation is what still separates models. GPT-5.6-sol's matched gap of -3.6 points hides -64.3 points of damage and +60.7 points of compensation. The deployment configuration reorders models: one reversal among 26 comparable pairs is unambiguous and four more sit on single-task margins. Evaluations of command-issuing agents should report the model configuration, generation contract, execution path, operating point, and final-state validator rather than treat a matched score as an intrinsic model property.
OpenAI is a non-profit artificial intelligence research company. Our goal is to advance digital intelligence in the way that is most likely to benefit humanity as a whole, unconstrained by a need to generate financial return. Since our research is free from financial obligations, we can better focus on a positive human impact.
We’re releasing the public beta of OpenAI Gym, a toolkit for developing and comparing reinforcement learning (RL) algorithms. It consists of a growing suite of environments (from simulated robots to Atari games), and a site for comparing and reproducing results.
The OpenAI team is now 45 people. Together, we’re pushing the frontier of AI capabilities—whether by validating novel ideas, creating new software systems, or deploying machine learning on robots.
We’re open-sourcing OpenAI Baselines, our internal effort to reproduce reinforcement learning algorithms with performance on par with published results. We’ll release the algorithms over upcoming months; today’s release includes DQN and three of its variants.
We’re releasing a new class of reinforcement learning algorithms, Proximal Policy Optimization (PPO), which perform comparably or better than state-of-the-art approaches while being much simpler to implement and tune. PPO has become the default reinforcement learning algorithm at OpenAI because of its ease of use and good performance.
We’re releasing two new OpenAI Baselines implementations: ACKTR and A2C. A2C is a synchronous, deterministic variant of Asynchronous Advantage Actor Critic (A3C) which we’ve found gives equal performance. ACKTR is a more sample-efficient reinforcement learning algorithm than TRPO and A2C, and requires only slightly more computation than A2C per update.
We’re providing 6–10 stipends and mentorship to individuals from underrepresented groups to study deep learning full-time for 3 months and open-source a project.
We’re now accepting applications for the next cohort of OpenAI Fellows, a program which offers a compensated 6-month apprenticeship in AI research at OpenAI.
We’ve trained an agent to achieve a high score of 74,500 on Montezuma’s Revenge from a single human demonstration, better than any previously published result. Our algorithm is simple: the agent plays a sequence of games starting from carefully chosen states from the demonstration, and learns from them by optimizing the game score using PPO, the same reinforcement learning algorithm that underpins OpenAI Five.
Our first class of OpenAI Scholars is underway, and you can now follow along as this group of experienced software developers becomes machine learning practitioners.
Yesterday, OpenAI Five won a best-of-three against a team of 99.95th percentile Dota players: Blitz, Cap, Fogged, Merlini, and MoonMeander—four of whom have played Dota professionally—in front of a live audience and 100,000 concurrent livestream viewers.
OpenAI Five lost two games against top Dota 2 players at The International in Vancouver this week, maintaining a good chance of winning for the first 20–35 minutes of both games.
We are now accepting applications for our second cohort of OpenAI Scholars, a program where we provide 6–10 stipends and mentorship to individuals from underrepresented groups to study deep learning full-time for 3 months and open-source a project.
Our first cohort of OpenAI Fellows has concluded, with each Fellow going from a machine learning beginner to core OpenAI contributor in the course of a 6-month apprenticeship.
We’ve written a paper arguing that long-term AI safety research needs social scientists to ensure AI alignment algorithms succeed when actual humans are involved. Properly aligning advanced AI systems with human values requires resolving many uncertainties related to the psychology of human rationality, emotion, and biases. The aim of this paper is to spark further collaboration between machine learning and social science researchers, and we plan to hire social scientists to work on this full time at OpenAI.
We’ve created OpenAI LP, a new “capped-profit” company that allows us to rapidly increase our investments in compute and talent while including checks and balances to actualize our mission.
Our class of eight scholars (out of 550 applicants) brings together collective expertise in literature, philosophy, cell biology, statistics, economics, quantum physics, and business innovation.
OpenAI Five is the first AI to beat the world champions in an esports game, having won two back-to-back games versus the world champion Dota 2 team, OG, at Finals this weekend. Both OpenAI Five and DeepMind’s AlphaStar had previously beaten good pros privately but lost their live pro matches, making this also the first time an AI has beaten esports pros on livestream.
We’ve created MuseNet, a deep neural network that can generate 4-minute musical compositions with 10 different instruments, and can combine styles from country to Mozart to the Beatles. MuseNet was not explicitly programmed with our understanding of music, but instead discovered patterns of harmony, rhythm, and style by learning to predict the next token in hundreds of thousands of MIDI files. MuseNet uses the same general-purpose unsupervised technology as GPT-2, a large-scale transformer model trained to predict the next token in a sequence, whether audio or text.
Our second class of OpenAI Fellows has wrapped up, with each Fellow going from a machine learning beginner to core OpenAI contributor in the course of a 6-month apprenticeship. We are currently reviewing applications on a rolling basis for our next round of OpenAI Fellows Summer 2019.
Our second class of OpenAI Scholars has concluded, with all eight scholars producing an exciting final project showcased at Scholars Demo Day at OpenAI.
Microsoft 正在向 OpenAI 投资 10 亿美元，以支持我们构建具有广泛分布经济收益的通用人工智能（AGI）。我们将合作在 Microsoft Azure 内开发一个可扩展至 AGI 的硬件和软件平台。我们将共同开发新的 Azure AI 超级计算技术，Microsoft 将成为我们的独家云提供商——因此我们将携手努力，进一步扩展 Microsoft Azure 在大规模 AI 系统方面的能力。
At OpenAI, each Thursday is Learning Day: a day where employees have the option to self-study technical skills that will make them better at their job but which aren’t being learned from daily work.
We’re releasing the 774 million parameter GPT-2 language model after the release of our small 124M model in February, staged release of our medium 355M model in May, and subsequent research with partners and the AI community into the model’s potential for misuse and societal benefit. We’re also releasing an open-source legal agreement to make it easier for organizations to initiate model-sharing partnerships with each other, and are publishing a technical report about our experience in coordinating with the wider AI research community on publication norms.
We’ve fine-tuned the 774M parameter GPT-2 language model using human feedback for various tasks, successfully matching the preferences of the external human labelers, though those preferences did not always match our own. Specifically, for summarization tasks the labelers preferred sentences copied wholesale from the input (we’d only asked them to ensure accuracy), so our models learned to copy. Summarization required 60k human labels; simpler tasks which continue text in various styles required only 5k. Our motivation is to move safety techniques closer to the general task of “machines talking to humans,” which we believe is key to extracting information about human values.
We’ve trained a pair of neural networks to solve the Rubik’s Cube with a human-like robot hand. The neural networks are trained entirely in simulation, using the same reinforcement learning code as OpenAI Five paired with a new technique called Automatic Domain Randomization (ADR). The system can handle situations it never saw during training, such as being prodded by a stuffed giraffe. This shows that reinforcement learning isn’t just a tool for virtual tasks, but can solve physical-world problems requiring unprecedented dexterity.
As the final model release of GPT-2’s staged release, we’re releasing the largest version (1.5B parameters) of GPT-2 along with code and model weights to facilitate detection of outputs of GPT-2 models. While there have been larger language models released since August, we’ve continued with our original staged release plan in order to provide the community with a test case of a full staged release process. We hope that this test case will be useful to developers of future powerful models, and we’re actively continuing the conversation with the AI community on responsible publication.
We’re introducing OpenAI Microscope, a collection of visualizations of every significant layer and neuron of eight vision “model organisms” which are often studied in interpretability. Microscope makes it easier to analyze the features that form inside these neural networks, and we hope it will help the research community as we move towards understanding these complicated systems.
We find that, just as a large transformer model trained on language can generate coherent text, the same exact model trained on pixel sequences can generate coherent image completions and samples. By establishing a correlation between sample quality and image classification accuracy, we show that our best generative model also contains features competitive with top convolutional nets in the unsupervised setting.
We’re excited to announce that OpenAI is co-organizing two NeurIPS 2020 competitions with AIcrowd, Carnegie Mellon University, and DeepMind, using Procgen Benchmark and MineRL.
Our third class of OpenAI Scholars presented their final projects at virtual Demo Day, showcasing their research results from over the past five months.
We’ve scaled Kubernetes clusters to 7,500 nodes, producing a scalable infrastructure for large models like GPT-3, CLIP, and DALL·E, but also for rapid small-scale iterative research such as Scaling Laws for Neural Language Models.
We’re proud to announce that the 2021 class of OpenAI Scholars has completed our six-month mentorship program and have produced an open-source research project with stipends and support from OpenAI.
We’ve trained a system that solves grade school math problems with nearly twice the accuracy of a fine-tuned GPT-3 model. It solves about 90% as many problems as real kids: a small sample of 9-12 year olds scored 60% on a test from our dataset, while our system scored 55% on those same problems.
We’ve trained language models that are much better at following user intentions than GPT-3 while also making them more truthful and less toxic, using techniques developed through our alignment research. These InstructGPT models, which are trained with humans in the loop, are now deployed as the default language models on our API.
We’re happy to announce several executive role changes that reflect our recent progress and will ensure continued momentum toward our next major milestones.
We’ve trained a model called ChatGPT which interacts in a conversational way. The dialogue format makes it possible for ChatGPT to answer followup questions, admit its mistakes, challenge incorrect premises, and reject inappropriate requests.
OpenAI 研究人员与 Georgetown University 的 Center for Security and Emerging Technology 以及 Stanford Internet Observatory 合作，研究大型语言模型可能如何被滥用于虚假信息目的。此次合作包括 2021 年 10 月举办的一场研讨会，汇集了 30 名虚假信息研究人员、机器学习专家和政策分析师，并最终在一年多研究的基础上共同撰写了一份报告。该报告概述了如果语言模型被用于增强虚假信息活动，会对信息环境构成的威胁，并介绍了一个用于分析潜在缓解措施的框架。点击此处阅读完整报告。
We’re launching a pilot subscription plan for ChatGPT, a conversational AI that can chat with you, answer follow-up questions, and challenge incorrect assumptions.
We’ve implemented initial support for plugins in ChatGPT. Plugins are tools designed specifically for language models with safety as a core principle, and help ChatGPT access up-to-date information, run computations, or use third-party services.
This initiative is essential to our commitment to develop safe and advanced AI. As we create technology and services that are secure, reliable, and trustworthy, we need your help.
We use GPT-4 for content policy development and content moderation decisions, enabling more consistent labeling, a faster feedback loop for policy refinement, and less involvement from human moderators.
我们探索了在视频数据上对生成模型进行大规模训练。具体而言，我们在不同时长、分辨率和宽高比的视频与图像上联合训练文本条件扩散模型。我们利用了一种 transformer 架构，该架构作用于视频和图像潜在代码的时空片段。我们最大的模型 Sora 能够生成一分钟的高保真视频。我们的结果表明，扩展视频生成模型是构建物理世界通用模拟器的一条有前景的路径。
OpenAI banned accounts linked to a previously unreprorted Russia-origin operation we dubbed "Bad Grammar", using AI to generate English- and Russian-language Telegram comments on Ukraine, Moldova, Baltic, and US politics.
How the voices for ChatGPT were chosen We worked with industry-leading casting and directing professionals to narrow down over 400 submissions before selecting the 5 voices.
In a multi-faceted agreement, Vox Media’s content will enhance the output of OpenAI’s ChatGPT, and the company will build on OpenAI’s technology to develop products to better serve its audiences and advertisers.
MavenAGI is a new software company for the AI era. They recently launched an AI customer service agent, built on the flexibility of GPT-4, which a number of companies like Tripadvisor, Clickup and Rho are already using to save time and better serve their customers.
Paf adopted ChatGPT Enterprise across its entire company, with engineers using custom GPTs on a daily basis to speed up routine development tasks. Paf also integrated ChatGPT Enterprise into the grit:lab coding academy (gritlab.ax), training the next generation of software developers using an AI-augmented, systems-architecture mindset from day one. In addition to the wide range of use cases for developers and grit:lab students, 70% of Paf employees actively use ChatGPT Enterprise, spanning business teams like finance, HR, marketing, and customer support.
Zico Kolter Joins OpenAI’s Board of Directors We’re strengthening our governance with expertise in AI safety and alignment. Zico will also join the Safety & Security Committee
OpenAI banned an account that likely originated in the US and used AI to create a fake ChatGPT error message claiming to detect “Russian troll” activity.
Elon Musk’s latest legal filing against OpenAI marks his fourth attempt in less than a year to reframe his claims. However, his own words and actions speak for themselves—in 2017, Elon not only wanted, but actually created, a for-profit as OpenAI’s proposed new structure.
Bertelsmann, the global media, services, and education company headquartered in Germany, will integrate OpenAI’s technology across multiple brands around the world.
OpenAI banned accounts potentially used to facilitate a deceptive employment scheme with characteristics of publicly reported North Korea-linked IT-worker activity.
Mercari leverages GPT-4o mini and GPT-4 to streamline selling, enhance product listings, and boost sales, transforming the online marketplace with features like AI Listing Support and Mercari AI Assistant.
OpenAI has grown a lot. We remain focused on the same core—pursuing frontier AI research that accelerates human progress–but we now also deliver products used by hundreds of millions of people.
Already a nonprofit, and already using AI to help people solve hard problems, OpenAI aims to build the best-equipped nonprofit the world has ever seen—combining potentially historic financial resources with something even more powerful: technology that can scale human ingenuity itself.
OpenAI banned accounts associated with suspected deceptive employment campaigns that used AI to develop materials for potentially fraudulent remote-job applications.
OpenAI introduces its Outbound Coordinated Disclosure Policy to guide how it responsibly reports vulnerabilities in third-party software—emphasizing integrity, collaboration, and proactive security at scale.
Unify, an AI-powered GTM platform, uses OpenAI’s o3, GPT-4.1, and CUA to automate prospecting, research, and outreach. With hyper-personalized messaging and an always-on workflow, Unify helps teams generate pipeline at scale while focusing on high-impact customer interactions.
Retell AI is transforming the call center with AI voice automation powered by GPT-4o and GPT-4.1. Its no-code platform enables businesses to launch natural, real-time voice agents that cut call costs, boost CSAT, and automate customer conversations—without scripts or hold times.
At OpenAI, we build tools to help people solve hard problems—including nonprofits working on the frontlines of their communities. The OpenAI Academy is teaming up with the Walton Family Foundation, Emerson Collective, and a network of local nonprofit organizations to host the Nonprofit Jam—a one-day, nationwide event bringing together more than 1,000 nonprofit leaders across 10 locations.
ChatGPT agent System Card: OpenAI’s agentic model unites research, browser automation, and code tools with safeguards under the Preparedness Framework.
In this paper, we study the worst-case frontier risks of releasing gpt-oss. We introduce malicious fine-tuning (MFT), where we attempt to elicit maximum capabilities by fine-tuning gpt-oss to be as capable as possible in two domains: biology and cybersecurity.
Today, OpenAI for Government is announcing a new partnership with the U.S. General Services Administration (GSA) to launch a transformative initiative. For the next year, ChatGPT Enterprise will be available to the entire federal executive branch workforce at essentially no cost.
Built with OpenAI o3, o3-Pro, GPT-4.1, and GPT-5, Basis’ AI agents help accounting firms save up to 30% of their time and expand capacity for advisory and growth.
Discover how MIXI, a leader in digital entertainment and lifestyle services in Japan, uses ChatGPT Enterprise to transform productivity, boost AI adoption across teams, and create a secure environment for innovation.
Discover how Blue J is transforming tax research with AI-powered tools built on GPT-4.1. By combining domain expertise with Retrieval-Augmented Generation, Blue J delivers fast, accurate, and fully-cited tax answers—trusted by professionals across the US, Canada, and the UK.
OpenAI launches a $50M People-First AI Fund to help U.S. nonprofits scale impact with AI. Applications open Sept 8–Oct 8, 2025 for grants in education, healthcare, research, and more.
Applications are now open for OpenAI’s People-First AI Fund, a $50M initiative supporting U.S. nonprofits advancing education, community innovation, and economic opportunity. Apply by October 8, 2025, for unrestricted grants that help communities shape AI for the public good.
Discover how SafetyKit leverages OpenAI GPT-5 to enhance content moderation, enforce compliance, and outpace legacy safety systems with greater accuracy .
In this Executive Function series from OpenAI, discover how CNA is transforming its newsroom with AI. Editor-in-Chief Walter Fernandez shares insights on AI adoption, culture, and the future of journalism.
ENEOS Materials uses ChatGPT Enterprise to speed research, improve plant design safety, and cut HR analysis time by 90%, with 80% reporting better workflows.
At OpenAI, we rely on our own technology to help streamline work, scale expertise, and drive outcomes. In our new series, OpenAI on OpenAI, we share lessons to help other organizations do the same.
Discover how HiBob uses ChatGPT Enterprise and custom GPTs to scale AI adoption, boost revenue, streamline HR workflows, and deliver AI-powered features in the Bob platform.
CRED is improving premium customer experiences in India with OpenAI, using GPT-powered tools to boost support accuracy, cut response times, and raise customer satisfaction.
Neuro uses ChatGPT Business to scale nationwide with fewer than 70 employees, saving time, reducing costs, and turning faster execution across sales and operations into growth.
OpenAI launches OpenAI for Ireland, partnering with the Irish Government, Dogpatch Labs and Patch to help SMEs, founders and young builders use AI to innovate, boost productivity and build the next generation of Irish tech startups.
OpenAI has been named an Emerging Leader in Gartner’s 2025 Innovation Guide for Generative AI Model Providers. The recognition reflects our enterprise momentum, with over 1 million companies building with ChatGPT.
Global manufacturer Scania is scaling AI with ChatGPT Enterprise. With team-based onboarding and strong guardrails, AI is boosting productivity, quality, and innovation.
OpenAI works with independent experts to evaluate frontier AI systems. Third-party testing strengthens safety, validates safeguards, and increases transparency in how we assess model capabilities and risks.
OpenAI introduces the first research cases showing how GPT-5 accelerates scientific progress across math, physics, biology, and computer science. Explore how AI and researchers collaborate to generate proofs, uncover new insights, and reshape the pace of discovery.
OpenAI is partnering with DoorDash, SCORE, and local organizations to help 1,000 small businesses build with AI. The Small Business AI Jam gives Main Street business owners hands-on tools and training to compete and grow.
Mirakl is redefining commerce through AI agents and ChatGPT Enterprise—achieving faster documentation, smarter customer support, and building toward agent-native commerce with Mirakl Nexus.
Scout24 has created a GPT-5 powered conversational assistant that reimagines real-estate search, guiding users with clarifying questions, summaries, and tailored listing recommendations.
OpenAI reflects on ten years of progress, from early research breakthroughs to widely used AI systems that reshaped what’s possible. We share lessons from the past decade and why we remain optimistic about building AGI that benefits all of humanity.
OpenAI 正在推出面向新闻机构的 OpenAI Academy，这是一个与 American Journalism Project 和 The Lenfest Institute 共同打造的新学习中心，旨在帮助新闻编辑室有效使用 AI。该 Academy 提供培训、实用用例和负责任使用指南，以支持记者、编辑和出版机构在报道和运营中采用 AI。
OpenAI and Datadog brand graphic with the OpenAI wordmark on the left, the Datadog logo on the right, and a central abstract brown fur-like texture panel on a white background.
By rolling out ChatGPT Enterprise company-wide, Zenken has boosted sales performance, cut preparation time, and increased proposal success rates. AI-supported workflows are helping a lean team deliver more personalized, effective customer engagement.
TRUSTBANK partnered with Recursive to build Choice AI using OpenAI models, enabling personalized conversational recommendations that simplify Furusato Nozei gift discovery.
Discover the Sora feed philosophy—built to spark creativity, foster connections, and keep experiences safe with personalized recommendations, parental controls, and strong guardrails.
Microsoft and OpenAI continue to work closely across research, engineering, and product development, building on years of deep collaboration and shared success.
A new preprint extends single-minus amplitudes to gravitons, with GPT-5.2 Pro helping derive and verify nonzero graviton tree amplitudes in quantum gravity.
Wayfair uses OpenAI models to improve ecommerce support and product catalog accuracy, automating ticket triage and enhancing millions of product attributes at scale.
New research shows Americans send nearly 3 million daily messages to ChatGPT asking about compensation and earnings, helping close the wage information gap.
CyberAgent uses ChatGPT Enterprise and Codex to securely scale AI adoption, improve quality, and accelerate decisions across advertising, media, and gaming.
Learn how to analyze data with ChatGPT by exploring datasets, generating insights, creating visualizations, and turning findings into actionable decisions.
Learn what AI is, how it works, and how tools like ChatGPT use large language models. A clear, beginner-friendly guide to understanding artificial intelligence.
Learn how OpenAI protects community safety in ChatGPT through model safeguards, misuse detection, policy enforcement, and collaboration with safety experts.
Meet the ChatGPT Futures Class of 2026—26 student innovators using AI to build, research, and drive real-world impact. Discover how this generation is redefining learning, creativity, and opportunity with ChatGPT.
Singular Bank built Singularity, an internal assistant using ChatGPT and Codex to help bankers save 60–90 minutes daily on meeting prep, portfolio analysis, and follow-up.
Parloa leverages OpenAI models to power scalable, voice-driven AI customer service agents, enabling enterprises to design, simulate, and deploy reliable, real-time interactions.
OpenAI launches DeployCo, a new enterprise deployment company built to help organizations bring frontier AI into production and turn it into measurable business impact.
See how finance teams can use ChatGPT Work to build MBRs, reporting packs, variance bridges, model checks, and planning scenarios from real work inputs.
See how business operations teams can use ChatGPT Work to create initiative briefs, strategy updates, leadership decision packets, progress updates, and more from real work inputs
How Virgin Atlantic used Codex to ship its revamped mobile app on a fixed holiday travel deadline, reaching near-total unit test coverage and zero P1 defects.
Learn how Endava is using AI agents, ChatGPT Enterprise, and Codex to accelerate software delivery, automate workflows, and build an AI-native culture across the enterprise.
See how LSEG uses OpenAI to scale trusted AI across its global business, accelerating insights, shrinking release cycles, and empowering 4,000 employees.
Discover how astrophysicist Chi-kwan Chan uses Codex to build black hole simulations, helping scientists study extreme physics and test Einstein’s theory of general relativity.
See how Australian Payments Plus uses ChatGPT Enterprise and Codex to move faster through payments complexity. AP+ saves time, improves quality, and keeps human judgment central.
OpenAI Academy and the Walton Family Foundation are bringing hands-on AI Skills Jams to help K–12 educators build practical AI skills for the classroom.
See how sales teams can use ChatGPT Work to create pipeline briefs, meeting prep packets, forecast reviews, account plans, and stalled-deal diagnoses from real work inputs.
See how data science teams can use ChatGPT Work to build root-cause briefs, impact readouts, KPI memos, scoped analyses, and dashboard specs from real work inputs.
Cars24 uses OpenAI-powered voice and chat agents to handle 1M+ monthly conversation minutes, recover 12% of lost leads, and bring agentic workflows to teams across the company.
News organizations are using AI to strengthen reporting, grow audiences, and improve business operations, with OpenAI tools supporting journalists and publishers worldwide.
See how Univé built an AI-ready workforce with ChatGPT Enterprise by combining leadership, responsible governance, and employee-led innovation to transform work at scale.
OpenAI shares how its safety, security, transparency, and provenance practices support responsible AI governance in Europe. The work will continue as the EU AI Act advances.
OpenAI shares new results on long-standing open problems in mathematics and theoretical computer science, including advances in geometry, cryptography, and complexity.
Discover how HSP GRUPPE uses ChatGPT Enterprise to boost productivity, improve work quality, and create more capacity for tax advisory and client service.
The enterprise marketing team at Zapier uses ChatGPT Work to reduce the number of drop-offs in its lead funnel, build campaign assets, and automate reporting.
Virgin Atlantic is accelerating research, product planning, and decision-making with ChatGPT Work, helping teams connect signals across the customer journey.
Premium seats are coming to ChatGPT Business. Sign up by August 20 to get $100 in workspace credits and unlock higher usage for your team's most demanding work.
OpenAI sent Governor Greg Abbott a letter outlining its commitment to responsible AI infrastructure in Texas. The letter supports reliable, transparent growth that benefits Texans.
See how RingCentral uses ChatGPT Work and Codex to accelerate AI product development and centralize operational intelligence across engineering and operations.