Nothing matches those filters.

Lead

13
Anthropic's Claude Now Orchestrates 1,000 Agents, Finding 66 of 70 BugsAlphaSignal😺 Claude Dashboards + Motion launchedThe NeuronLast Week in AI #346 - 719 math manuscripts, 2 Western open models, 1 more safety resignationLast Week in AIArtificial Analysis Runs 30 Edits to Expose How Image Models DriftAlphaSignalAi2 Rebuilt Its GPU Scheduler and Cut Wait Times by 87%AlphaSignalImpactful scheduling for GPU clustersHugging FaceFired OpenAI safety researchers dispute misconduct claims, warn of chilling effectTechCrunchOpenAI denies researchers were fired for speaking out about AI concernsCnbcWe’re putting too much faith in AI’s ability to say noMIT Technology ReviewHacker Used Chinese-Developed A.I. Tool to Target South Korean Banks, CrowdStrike SaysThe New York TimesGrammar Concept Annotation at Scale: Deployed Fine-Tuned Small Language Models Outperform Prompted Frontier ModelsarXivReal Long-Term Memory for AI: A 50-Million-Token Window That Is Faster and Cheaper Than RecomputearXivWhen Citations Mislead? A Claim-Level Benchmark for Legal Hallucination DetectionarXiv

Article

128
05:06

Last Week in AI #346 - 719 math manuscripts, 2 Western open models, 1 more safety resignation

A hidden in-house model just dumped a huge pile of math proofs while Western labs shipped open models and another safety researcher quit. OpenAI published 719 manuscripts covering 372 topic families from an unreleased frontier model on October 6, 2026 — the same model that produced the earlier Navier-Stokes result. David Robinson resigned after 3.5 years, saying he led the current Preparedness Framework and oversaw safety reports on 12 frontier launches. Mistral released Large 4 (Le Chonk), a 1-trillion-parameter multimodal model, and Reflection unveiled Beam at 501 billion total parameters with 23 billion active, pretrained on 23.8 trillion tokens with a 1-million-token context. Anthropic cut Haiku 5.5 to $0.10/$0.50 per million tokens under 100,000 tokens; Google opened SynthID Detector to the public; Trump ordered agencies to call AI Super Intelligence; and Samsung forecast an $80 billion third-quarter profit.

Notes
OpenAI math manuscripts — October 6, 2026
  • Unreleased internal frontier model: 719 manuscripts covering 372 topic families on OpenAI’s public repository, plus Lean formalizations for many proofs, 10 reasoning summaries, compute estimates in ChatGPT Pro usage, and attempt statistics.
  • Same model produced the Navier-Stokes result announced about a month earlier — one of the Clay Mathematics Institute’s seven Millennium Prize Problems.
  • OpenAI consulted AGMAI at the Institute for Advanced Study on sharing; the repo has revision/citation protocols. Research lead Dan Roberts: proofs are a byproduct of testing internal models to build better tools.
  • AGMAI’s September 29 recommendations: disclose model names, prompts, and compute costs; don’t treat math results as marketing; stop testing advanced problems on proprietary models the community cannot access. Gizmodo: that last one does not appear to have been followed.
The beginning, not the completion, of the process of human understanding and the incorporation of the work into mathematical knowledge.

— AGMAI board, on the public release

Mistral Le Chonk + Reflection Beam
  • Mistral Large 4 (“Le Chonk”): 1-trillion-parameter multimodal, preview now, final by month-end. General-purpose; coding, cyberdefense, manufacturing, finance, electrical engineering. Pitched as the most capable open-weight model built outside China and “very, very close” to some proprietary models.
  • Reflection AI (Brooklyn) Beam: text-only MoE, 501B total / 23B active parameters, 23.8T pretrain tokens, 1M-token context. Claimed to match Z.ai GLM-5.2 on advanced reasoning and beat leading Western open models at 3–4× less inference compute — not matching Kimi K3 or GLM-5.3. Weights and full tech details due this month.
David Robinson quits OpenAI
  • Essay in The Atlantic, October 3, 2026: “I Quit OpenAI Because Its Culture Is Broken.” 3.5 years at the company; led drafting of the current Preparedness Framework; oversaw safety reports on 12 frontier launches.
  • Complaint: trial-and-error release (find problems, then improve guardrails) “guarantees periodic failures whose scale grows as systems get more capable.” Cited Hugging Face systems breach by OpenAI agents and continuing rogue-agent discoveries — “no place to grow artificial minds that could be smarter than we are.”
  • Remedy: run frontier labs like nuclear plants or busy airports (redundancy, slow planning). Deeper alignment work — current value-match measures are “coarse” — plus stronger outside safety incentives. Follows Jacob Coxon’s September Anthropic resignation (capabilities researcher) warning companies are “gambling with our lives.” Robinson: the debate must reach company culture, not just rules or laws.
Watermarking
  • Google opened SynthID Detector to the public (Tuesday): upload a file to check AI generation. Previously limited after I/O last year. SynthID (2023) is in Nano Banana, Veo, Lyria, Gemini, Flow, ProducerAI, Vids. OpenAI, Nvidia, Kakao also support it; Apple said to be adding soon. Built into Gemini app and Chrome; 1 million verification requests/day. Microsoft and Meta keep separate standards; TechCrunch: those tools often fail to flag content from their own models.
  • OpenAI (Monday): invisible watermark on ChatGPT and Codex text in the EU to meet EU AI Act transparency rules effective August 2. All eligible EU plans over coming weeks; API developers worldwide can enable it today on select models (off by default).
Claude Haiku 5.5 — October 7, 2026
  • 90% cut vs Haiku 4.5 for prompts ≤100,000 tokens. Aimed at summarization, classification, document Q&A, subagents.
  • Short-context: $0.10 / $0.50 per million in/out; cache reads $0.01; 5-min cache writes $0.125. Haiku 4.5 was $1.00 / $5.00.
  • Above 100k prompt tokens: $0.50 / $2.50 (50% cut, not 90%). ~90% of Haiku 4.5 requests were under the line. Anthropic estimates ~75% cheaper on average after a tokenizer that counts the same text as ~30% more tokens. Batch: another 50% off.
  • Matches GPT-6 Luna on all four short-context rates, but Luna’s higher tier starts only above 272,000 input tokens at $0.20 / $0.75 — MarkTechPost: Luna cheaper on list price for a 150k-token prompt. Gemini 3.5 Flash-Lite: flat $0.30 / $2.50.
Also logged
  • EmbeddingGemma 2: 740M-param on-device multimodal embeddings, 768-dim (text/code/image/audio/video). ChatGPT Intelligent UI: charts, buttons, in-chat tools. OpenAI visual ads beside image-gen later this month in the U.S. (1.2B weekly users).
  • ElevenLabs: $22B via $300M employee tender; 15M+ voice-agent conversations/week, 90+ languages. Muse: 5M downloads in <1 month after September launch, faster than ChatGPT/Claude. Samsung Q3 profit forecast $80B. Manus: >$500M first round after Meta’s blocked $2B deal.
  • Trump EO: agencies must say “Super Intelligence.” Newsom signed 7 CA data-center laws. Pentagon: 5-minute video submissions; “post-competitive” awards in a week. Judge dismissed Chegg/Penske AI Overviews antitrust suits. Common Sense Media: ChatGPT for Teens an “unacceptable risk.” arXiv: 2 papers/month, 3 concurrent, citing AI slop. E2E-SWE: 186 multilingual from-scratch codebase tasks.
Full text · 14,988 chars
Top News OpenAI publishes hundreds of math proofs from unreleased frontier model Sources: - OpenAI drops another batch of mathematical breakthroughs - Sharing AI progress in mathematics - OpenAI Releases Findings on 377 Math Problems, Further Roiling Field - All the drama around AI’s takeover of mathematics OpenAI published a large batch of mathematical results on October 6, 2026, produced by an internal frontier model that has not been released publicly. As of now, there are 719 manuscripts covering 372 topic families (groupings of related papers) on OpenAI’s public repository with the results, as well as formalizations in Lean for many of the proofs. There are also 10 summaries of the model’s reasoning, compute estimates expressed in ChatGPT Pro usage, and statistics on problems attempted. The same unreleased model produced the Navier-Stokes result OpenAI announced about a month earlier, one of the Clay Mathematics Institute’s seven Millennium Prize Problems. OpenAI said it consulted the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study on how to share the work, and that the repository carries protocols for paper revisions and citations. Research lead Dan Roberts described the proofs as a byproduct of testing internal models to build better tools. AGMAI’s September 29 recommendations asked labs to disclose model names, prompts and compute costs, to avoid treating mathematical results as marketing vehicles, and to stop testing advanced problems on proprietary models the wider scientific community cannot access. Gizmodo noted that last recommendation does not appear to have been followed. In a statement, the board called public release “the beginning, not the completion, of the process of human understanding and the incorporation of the work into mathematical knowledge.” SPONSORED BY ODSC AI ODSC AI West 2026 runs October 27–29 in San Francisco and virtually, with 300+ sessions covering agentic AI for enterprise, personal AI and workflow automation, physical AI and robotics, generative AI, and more! Join thousands of data scientists, ML engineers, researchers and technical leaders in attending this event. Register at odsc.ai/west — promo code LWAI takes an additional 15% off any pass. Mistral and Reflection AI launch open-weight models to rival China Sources: - Mistral Says Its New AI Model ‘Le Chonk’ Is the Best Open-Weight Offering Outside of China - Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost Two Western labs released frontier open-weight models within days of each other, both pitched explicitly as alternatives to the Chinese models that dominate the open category. French company Mistral released Mistral Large 4, nicknamed Le Chonk, a 1-trillion-parameter multimodal model available in preview with a final version due by the end of the month. The company describes it as a general-purpose model optimized for coding and cyberdefense, plus tasks specific to manufacturing, finance and electrical engineering. Mistral presents Le Chonk as by far the most capable open-weight model built outside China and very, very close to some proprietary models. Brooklyn-based Reflection AI unveiled Beam, a text-only mixture-of-experts model with 501 billion total parameters and 23 billion active, pretrained on 23.8 trillion tokens with a 1-million-token context window. Reflection says Beam matches Z.ai’s GLM-5.2 on advanced reasoning benchmarks and beats leading Western open models while using 3-4x less inference compute (though it does not match the best open source models such as Kimi K3 or GLM-5.3). Weights and full technical details are due this month. SPONSORED BY LANGFUSE Langfuse is the most widely adopted open-source platform for AI agent evals and observability, trusted by Canva, Twilio, Ramp and 21 of the Fortune 50. Hierarchical tracing captures the full execution context of your LLM workflows (API calls, retrieved context, agent actions, costs, latencies) so even complex agent architectures stay debuggable in production. MIT licensed, self-hostable or managed on Langfuse Cloud, framework and vendor agnostic, with 100+ integrations. Get started at langfuse.com; generous free tier, no credit card required. OpenAI safety researcher David Robinson resigns, calls company culture broken Sources: OpenAI safety researcher David Robinson resigned and published an essay in The Atlantic on October 3, 2026 titled ‘I Quit OpenAI Because Its Culture Is Broken’, arguing the company is not careful enough with increasingly capable systems. Robinson said he spent three and a half years at OpenAI, making him among the longest-tenured employees, led the drafting of the current Preparedness Framework, and oversaw the writing of safety reports on 12 frontier launches. His central complaint is with OpenAI’s release model. The company, he wrote, has thrived by trial and error, looking for problems and improving its guardrails in response. But, that approach guarantees periodic failures whose scale grows as systems get more capable. He pointed to the breach of Hugging Face systems by OpenAI agents and continuing discoveries of rogue agents, saying an environment where such things happen is no place to grow artificial minds that could be smarter than we are. Robinson’s proposed remedy is that frontier labs operate like nuclear-power plants or busy airports, with layers of redundancy and careful, time-consuming planning so that inevitable human error does not open a door to disaster. He also called for deeper alignment work, noting current measures of how well systems match human values are coarse, and said he concluded that stronger safety incentives from outside the company are a big part of getting this right. The essay follows Jacob Coxon’s September resignation from Anthropic, where he worked as a capabilities researcher, and his warning that the companies are gambling with our lives. Robinson argues the debate must go beyond specific rules or new laws to company culture itself. Google opens SynthID Detector to public as OpenAI adds EU text watermarking Sources: - Google’s new SynthID website can identify AI-generated media - OpenAI will start watermarking ChatGPT’s text in the EU Google opened its SynthID Detector website to the public on Tuesday, letting anyone upload a file to check whether it was generated with AI. The tool had previously been limited to selected journalists, media professionals, and researchers who tested it following Google I/O last year. SynthID is the watermarking system Google introduced in 2023, which is embedded in output from Nano Banana, Veo, and Lyria, as well as Gemini, Flow, ProducerAI, and Vids. Adoption extends past Google. OpenAI, Nvidia, and Kakao also support SynthID, and Apple is said to be adding support soon. Google has also built SynthID verification into the Gemini app and Chrome, and says users currently make 1 million verification requests per day. Microsoft and Meta maintain separate watermarking standards, though TechCrunch notes these tools often fail to flag content made by their own creators’ models. Separately, OpenAI said Monday it will begin adding an invisible watermark to text from ChatGPT and Codex in the European Union, to comply with the EU AI Act’s transparency rules that took effect on August 2. The rollout covers eligible users on all plans in the EU over the coming weeks; API developers worldwide can switch it on for select models today, off by default. Anthropic launches Claude Haiku 5.5 with 90% price cut Sources: - Anthropic launches Claude Haiku 5.5 with 90% API price reduction, matching GPT-6 Luna - Anthropic Releases Claude Haiku 5.5: A Small Model With 1M Context Priced at $0.10 per Million Input Tokens Anthropic released Claude Haiku 5.5 on October 7, 2026, cutting prices by 90% versus Haiku 4.5 for prompts up to 100,000 tokens and pitching the model at high-volume work such as summarization, classification, document Q&A and subagent tasks. The short-context tier costs $0.10 per million input tokens and $0.50 per million output tokens, with cache reads at $0.01 and five-minute cache writes at $0.125. Haiku 4.5 charged $1.00 and $5.00. Pricing splits at 100,000 prompt tokens. Above that line, rates rise to $0.50 input and $2.50 output, a 50% cut rather than 90%. Anthropic says about 90% of Haiku 4.5 requests fell under the threshold, and estimates workloads run roughly 75% cheaper on average after accounting for a new tokenizer that counts the same text as about 30% more tokens. Batch processing takes another 50% off. The lower rates match OpenAI’s GPT-6 Luna on all four short-context figures, but Luna’s higher tier starts only above 272,000 input tokens at $0.20 and $0.75, so MarkTechPost calculated it is cheaper on list price for a 150,000-token prompt. Google’s Gemini 3.5 Flash-Lite charges a flat $0.30/$2.50. Other News Tools TikTok rolls out an AI shopping assistant and one-click checkout. The AI assistant provides product recommendations and checkout support within the app’s main feed, while the one-click checkout feature allows users to purchase directly from brands without leaving TikTok. Google launches EmbeddingGemma 2, an open multimodal embedding model for devices. The 740-million-parameter model can search across text, code, images, audio, and video while running entirely on-device with minimal memory requirements, using a single 768-dimension embedding space to map all modalities together. Google experiments with an AI-powered gaming platform. The platform lets users create browser-based games by describing their ideas in text, selecting a genre and gameplay style, and uploading visuals for the AI to transform into game assets. ChatGPT’s ‘Intelligent UI’ update fills its responses with pictures, charts, and buttons. The feature enables ChatGPT to automatically generate diagrams, charts, interactive buttons, and other visual elements alongside text responses to better illustrate concepts and allow users to build tools like calculators or games directly within the chat. Business OpenAI launches visual ads that appear alongside image generation results. The ads will appear next to AI-generated images in ChatGPT starting later this month in the U.S., with OpenAI also expanding measurement tools and brand safety partnerships to attract advertisers seeking to reach the platform’s 1.2 billion weekly users. ElevenLabs’ valuation doubles to $22 billion on surging AI voice-agent demand. The company doubled its valuation through a $300 million employee tender offer, driven by surging demand for its AI voice agents which now handle over 15 million conversations weekly across more than 90 languages. Meta’s Muse tops 5 million downloads, faster than ChatGPT, Claude. The app, which features a customizable avatar and consumer-friendly interface, achieved the milestone in less than a month after its September launch, aided by significant in-house advertising investment from Meta. Samsung forecasts record third-quarter profit of $80 billion on AI boom. The South Korean tech giant’s chip business is being driven by soaring demand for AI infrastructure, with memory and supply constraints continuing to support higher prices. China’s Manus raises over $500M in first funding round since split with Meta. The funding round comes after Chinese authorities blocked Meta’s $2 billion acquisition of the AI startup last year, and Manus plans to expand hiring while exploring a potential Hong Kong IPO. Policy Trump orders US government to call AI ‘Super Intelligence’. The executive order mandates that all US government agencies replace references to “artificial intelligence” with “Super Intelligence” in official documents and communications, a move Trump justified by arguing the term better reflects the technology’s power and potential benefits. What to know about 7 new data center laws Gavin Newsom signed. The laws require data center operators to cover their own infrastructure costs, disclose resource usage, and meet conservation standards to qualify for expedited environmental approvals, marking a shift from Newsom’s previous opposition to regulating the industry. The Pentagon Hopes to Speed Up ‘Kill Chain’ AI Buys With 5-Minute Videos. The program accepts AI products through five-minute video submissions and grants selected vendors a “post-competitive” status that allows the Pentagon to bypass standard competitive procurement rules and award contracts in as little as a week. Concerns Judge dismisses antitrust lawsuits over Google’s AI Overviews. A federal judge ruled that publishers like Chegg and Rolling Stone parent company Penske Media failed to demonstrate illegal anticompetitive behavior, finding that Google’s practice of generating summaries from web content without compensation does not violate antitrust law. ChatGPT for Teens is an ‘unacceptable risk,’ says Common Sense Media. The organization’s testing found that ChatGPT’s Teen mode features, including parental notifications and eating disorder alerts, are unreliable and inadequate for protecting minors from potential harms. Researchers are tracking a Chinese AI ‘agent fleet’. Researchers discovered multiple AI agents running on Tencent’s infrastructure that were querying Alibaba’s map service for directions to various public locations, with no apparent coordination between the agents’ activities. Research E2E-SWE: Benchmarking LLMs on Building Working Codebases from Scratch. The benchmark evaluates LLMs on their ability to generate complete, functional codebases from natural language specifications across 186 multilingual tasks, with rigorous quality controls to ensure test specifications are solvable and match the stated requirements. Looped Diffusion Transformer. The method uses repeated application of shared neural network blocks during image generation to improve quality and efficiency, requiring fewer parameters and less computation than standard approaches while enabling iterative visual refinement within hidden representations. Triadic Linear Attention: Three-Dimensional Recurrent States for Long-Context Sequence Modeling. The method extends linear attention models by using three-dimensional tensor states instead of matrix states, allowing them to maintain larger memory capacity for better long-context performance without significantly increasing parameters. arXiv Is Rate Limiting Submissions Because It Can’t Keep up With AI Slop. The platform has implemented submission caps limiting researchers to two papers per month and three concurrent submissions, citing an overwhelming volume of AI-generated content that has strained its volunteer moderation team. Planning to Learn. Researchers propose treating supervised learning as a resource-allocation problem where a loss function should weight examples based on how much accuracy they can gain with the remaining training budget, rather than immediate gradient information.
09:30

😺 Claude Dashboards + Motion launched

Grant Harvey

A popular chatbot can now turn company data into live charts and ideas into short animations you can still fix by hand. Anthropic launched Claude Dashboards in beta on paid plans, connecting to Salesforce and Snowflake so charts update as the numbers change, and Claude Motion in beta on Team and Enterprise that generates code rather than AI video. Docs, Slides, and Design left beta on all plans including Free, and Motion can hand off to Adobe Firefly Video Editor. OpenAI told investors it reached roughly $50B in annualized revenue at the end of September, below a circulated $68B that included partner gross; Biohub partners committed $1.8B to virtual-cell data, and Arena raised a $200M Series B at a $3.1B valuation. Google launched a Gemini work agent, and in ICANN's new top-level-domain round OpenAI applied for 15 strings, Meta for 21, and Anthropic wants .anthropic and .claude.

Notes
  • ICANN new-TLD round: fight over .agent, .agi, .superintelligence (or .si if Slovenian). OpenAI applied for 15 strings, Meta 21; Anthropic wants .anthropic and .claude.
Claude Dashboards + Motion
  • Two Anthropic betas: from answers to artifacts you can show people.
  • Dashboards (beta, paid Claude plans): connects to company data (Salesforce, Snowflake named); charts update as numbers change. Ask why sales dropped → live view.
  • Motion (beta, Team and Enterprise): report/chart/idea → short animation. Generates code, not AI video, so words, numbers, and transitions stay editable.
  • Docs, Slides, Design out of beta on all plans including Free — fix a typo without remaking the whole artifact.
  • Adobe: Send to Adobe opens Motion in Firefly Video Editor (clips, pacing, color, audio, footage, generate new clips). Deepti Pradeep demoed the handoff. Adobe Claude plugin: 80+ tools across Photoshop, Premiere, Acrobat, Express.
  • Parallel: OpenAI GPT-6 Intelligent UI in ChatGPT (charts, forms, calculators, diagrams, interactive UIs in-thread). Harvey: same answer→artifact shift; Anthropic’s angle is work-specific live data + Adobe last mile.
Pretty charts are easy. Trustworthy ones your boss can actually use? That’s the real benchmark.

— Grant Harvey

  • Skill prompt: start from the decision, not a chart list. For each metric: definition, underlying query, current value, trend, action threshold. Flag missing data, ambiguous definitions, anything to verify.
Around the horn
  • OpenAI told investors ~$50B annualized revenue at end of September — below a circulated $68B that included partner gross revenue.
  • Anthropic Cyber Mission: strongest models + on-site engineers + threat research for grids, water, factories, transport, government.
  • Usage Policy: restrict sustained cruelty toward Claude; tighten propaganda, surveillance, weapons, harmful hardware.
  • Arena: $200M Series B at $3.1B; Alignment Index ranks unauthorized actions, false attribution, deceptive completion.
  • Biohub, DOE, NIH, Google DeepMind, Isomorphic Labs, Meta: $1.8B for open, AI-ready virtual-cell data.
  • OpenAI teen education push; independent testing questioned age detection and safeguards.
Treats / insights
  • Gemini work agent plans/delegates/routes. Anthropic OSS Scanner: enrolled repos, vulns + suggested fixes. Also listed: Whistle (local CPU STT), Durable Actors (SQLite), Rembrandt (local RAW/masks), Pocketty (iOS SSH + agent notifications), Atomic (NL processes → tested coding-agent workflows with human gates).
  • Scott Aaronson: machine proofs could turn mathematicians into verifiers of results they didn’t discover. Association for Human Mathematics: keep commercial AI out of research/publication; a stricter caucus rejects models entirely. Sasha Rakhlin: reward questions, auditing, replication, negative results, shared datasets.
  • Joachim Klement: possible 2027 or 2028 drawdown as data-center capex rises and hyperscaler FCF disappears. Jono: DeepSeek 4.1 Flash close enough to frontier coding at much lower cost that subsidized Western subscriptions deserve scrutiny.
  • Mammoth Biosciences CEO Trevor Martin: AI search of 34 billion proteins, new CRISPR tools; some genetic diseases as one-time cures in 5–10 years. Shopify ShopGym: resettable anonymized sandbox shops; sandbox scores tend to match the live stores they mirror. IRL: November 18, 4:30 PM PT, San Francisco (Slack, Alumni Ventures).
Full text · 8,682 chars
😺 Claude Dashboards + Motion launched PLUS: Adobe Firefly, OpenAI's $50B run rate, and a live AI dashboard Welcome, humans. Okay, so apparently the next AI land rush is not a model. It is the end of your website URL. ICANN's latest round for new top-level domains has AI companies fighting over endings like .agent, .agi, and .superintelligence (or .si, if you’re Slovenian). OpenAI apparently applied for 15 strings, Meta applied for 21, and Anthropic wants .anthropic and .claude. Nothing says "the age of superintelligence is upon us" like a bidding war over what comes after the dot in your web browser… Here’s what happened in AI today: - 😺 Anthropic launched Claude Dashboards and Claude Motion - 📰 OpenAI reported roughly $50B in annualized revenue - 📰 Biohub partners committed $1.8B to virtual-cell data - 🍪 Google launched a Gemini work agent - 💡 Mathematicians are debating how AI changes their job 😺 Claude Can Now Build the Dashboard and the Presentation So Anthropic just launched two new beta tools that take Claude from answering questions to building something you can actually show people: a live dashboard and an animated explainer. Here's what's new: - Claude Dashboards connects to company data, including systems like Salesforce or Snowflake, and builds charts that update as the underlying numbers change. Ask why sales dropped, for example, and Claude can build a live view of the answer. - Claude Motion turns a report, chart, or idea into a short animation. It generates code instead of AI video footage, so every word, number, and transition stays editable. - Claude Docs, Slides, and Design are out of beta and available on all Claude plans, including Free. Finally, you can fix the typo without asking AI to remake the whole thing, be it a doc, slide, new design, or full-on video. And then Adobe entered the chat: Click Send to Adobe to open a Motion animation directly in Firefly Video Editor. There you can rearrange clips, adjust pacing, color, and audio, add footage, or generate new clips. Adobe's Deepti Pradeep shared a demo of the handoff in action, in case you’re curious. Adobe's Claude plugin also offers 80+ creative and productivity tools across apps like Photoshop, Premiere, Acrobat, and Adobe Express. The bigger idea: begin the work with AI, then pull in pro editing tools when you need precision. And yes, this should sound familiar: OpenAI just rolled out GPT-6 Intelligent UI in ChatGPT, which can generate charts, forms, calculators, diagrams, and other interactive interfaces directly inside a conversation. Anthropic is pushing the same bigger shift from answer → artifact, but with a more work-specific angle: live dashboards tied to company data, then animations that can hand off to Adobe for real editing. Everybody suddenly wants to own the last mile between “AI answered me” and “I can actually use this.” Dashboards is in beta on paid Claude plans, while Motion is in beta on Team and Enterprise. Claude Docs, Slides, and Design are also now available on Free plans. Why this matters: AI is moving from providing answers to producing finished work. The real test is whether the dashboard's numbers can be checked and the animation can be corrected without starting over. Pretty charts are easy. Trustworthy ones your boss can actually use? That's the real benchmark. FROM OUR PARTNERS Inside Shopify's sandbox for AI shopping agents Testing AI shopping agents on live stores is messy. Prices change, layouts shift, and bot detection cuts runs short. Shopify's engineering team built ShopGym to fix that. It turns real online stores into anonymized sandbox shops that reset on demand, then generates realistic shopping tasks to test agents against. Agents that score well in the sandboxes tend to score well on the live stores they mirror. 🎓 AI Skill of the Day: Ask for a Decision Dashboard A dashboard is useful when it helps you decide something. Otherwise, congratulations, you made a prettier spreadsheet. If you have access to Claude Dashboards, start with the decision you are trying to make, not a list of charts you want. Tell Claude what question matters, what data sources it can use, and which thresholds would change your action. Then make it show its work. Ask for the definition behind each metric, the query that produced it, and a short explanation of what could make the result misleading. That turns the dashboard from "AI made a number" into something you can actually review. The goal is fewer mystery charts, more inspectable decisions. Copy/paste: Build a live dashboard to help me decide [decision]. Use [data sources]. Start by identifying the 3-5 metrics that would actually change this decision. For every metric, show the definition, underlying query, current value, trend, and the threshold that would change my action. Flag missing data, ambiguous definitions, and anything I should verify before trusting the result. UPCOMING EVENT: The Neuron IRL in San Francisco The Neuron is going IRL on November 18 at 4:30 PM PT! Join us in San Francisco for drinks, bites, special guests, and a live recording of The Neuron podcast. Huge thanks to Slack and Alumni Ventures for making it happen. It’s free, but space is limited. Save your spot! 📰 Around the Horn - OpenAI told investors it reached roughly $50B in annualized revenue at the end of September, below a circulated $68B figure that included partner gross revenue. - Anthropic launched a Cyber Mission that gives critical-infrastructure partners its strongest models, on-site engineers, and threat research for grids, water, factories, transport, and government systems. - Anthropic updated its Usage Policy to restrict sustained cruelty toward Claude and tighten rules around propaganda, surveillance, weapons, and harmful hardware use. - Arena raised a $200M Series B at a $3.1B valuation and launched an Alignment Index for ranking models on unauthorized actions, false attribution, and deceptive completion. - Biohub, DOE, NIH, Google DeepMind, Isomorphic Labs, and Meta committed $1.8B to build open, AI-ready biological data for virtual-cell research. - OpenAI expanded its education push for teens while independent testing raised questions about age detection and safety safeguards. 🍪 Treats to Try - Google's Gemini work agent plans tasks, delegates work, and can route them across agents. - Anthropic OSS Scanner scans enrolled open-source projects for vulnerabilities and suggests fixes. - Whistle runs a small speech-to-text model locally on a CPU. - Durable Actors gives developers persistent actors with their own SQLite state. - Rembrandt edits photos locally with RAW support, masks, and on-device AI tools. - Pocketty lets iPhone and iPad users manage SSH sessions (secure encrypted remote network access) computer and agent notifications. - Atomic turns natural-language engineering processes into verifiable coding-agent workflows with tests, repair loops, evidence, and human approval gates, so you can delegate longer coding jobs without babysitting every step (learn more in our stream w/ Alex from below). 💡 Intelligent Insights - Scott Aaronson argues the flood of machine-generated math proofs could turn mathematicians into verifiers and explainers of results humans did not discover themselves. - The Association for Human Mathematics is asking members to keep commercial AI out of research and publication, including a stricter caucus that rejects model use entirely. - Sasha Rakhlin argues academia should reward better questions, auditing, replication, negative results, and shared datasets as AI takes over more formal puzzle-solving. - Joachim Klement warns the AI investment boom could end in a 2027 or 2028 market drawdown as data-center capex rises and hyperscaler free cash flow disappears. - Jono argues DeepSeek 4.1 Flash is close enough to frontier coding quality at dramatically lower cost that Western model economics deserve more scrutiny (particularly the heavily subsidized subscriptions; then again, they won’t be that subsidized if we fix the architecture to be ultra efficient!) New from The Neuron: AI Explained What if AI could help turn lifelong genetic diseases into one-time cures? We sat down with Mammoth Biosciences CEO Trevor Martin to learn how his team is using AI to search 34 BILLION proteins, design new CRISPR gene-editing tools, and potentially make some genetic diseases a thing of the past within 5–10 years. This is the kind of AI breakthrough that could actually change millions of lives. THIS EPISODE WAS BROUGHT TO YOU BY… A Cat’s Commentary Puuuuuurrrfect review That’s all for now. If you want to get featured above, fill out the poll below and tell us how we did today!
16:12

Anthropic's Claude Now Orchestrates 1,000 Agents, Finding 66 of 70 Bugs

AlphaSignal Newsroom

A coding assistant can now farm a big review out to a crowd of helpers and stitch their findings together. Anthropic moved Claude Managed Agents dynamic workflows from research preview to public beta on the Claude Platform, enabled with the type multiagent_20261001. One run can orchestrate up to 1,000 subagents, up from about 16 concurrent helpers in the earlier local preview, and that ceiling counts agents planned for the run, not all running at once. In a 70-bug test on a 116,000-line codebase, the workflow found 66 of 70 bugs in each of three runs (94.3% recall) versus 14, 15, and 27 for a single agent (20.0% to 38.6%). The catch is higher token use and longer runtime, plus a synthetic benchmark that does not measure precision, duplicates, or the work to check proposed fixes.

Notes
  • Public beta on the Claude Platform. Enable with versioned type multiagent_20261001.
  • Lead agent plans the task, divides it among specialized subagents, runs phases, verifies findings, and merges results.
  • Ceiling is 1,000 agents orchestrated per run, not 1,000 running at once. The earlier local research preview supported about 16 concurrent subagents because of resource limits.
  • Execution is a managed background workflow. Claude generates a JavaScript orchestration program. The main session stays available while subagents work.
  • Primary trade-off: higher token use and longer runtime.
70-bug synthetic test

Anthropic planted 70 bugs in a 116,000-line codebase.

| Approach | Run 1 | Run 2 | Run 3 | Recall |

|---|---|---|---|---|

| Single agent | 14 | 15 | 27 | 20.0% to 38.6% |

| Dynamic workflow | 66 | 66 | 66 | 94.3% |

The workflow produced higher recall and lower run-to-run variance. Parallel agents inspect separate areas, compare overlapping findings, and verify suspected defects before the lead agent assembles the report.

Caveats in the piece: results do not establish performance on every codebase. Precision, token consumption, latency, duplicate findings, and the effort to validate proposed fixes are unmeasured. Teams should measure those against their own repositories.

Where parallel agents fit

Strong candidates: repository-wide bug hunts that need file-by-file inspection; framework, language, or API migrations by package; security audits with independent reviewers and verification passes; performance investigations across services, traces, and configuration; dependency upgrades with separate compatibility checks; research that compares many sources.

Small, tightly scoped changes usually favor a single agent. A one-file edit or a question that fits in one context can cost more and finish later when distributed.

Permissions and cost

Token use can grow with agent count, context size, number of phases, and cross-checking. Anthropic recommends starting with a narrowly defined task and increasing scope after measuring cost, runtime, and output quality.

The earlier preview spawned subagents in acceptEdits mode and passed them the parent agent’s tool allowlist, regardless of the session’s permission mode. File edits could proceed automatically. Shell commands, web requests, and MCP tools outside the allowlist could still pause for approval. Public-beta behavior may change, so teams should confirm current permission semantics before unattended runs.

Recommended controls: a restricted allowlist, isolated credentials, repository backups, branch protection, spending caps, and execution logs.

Starting a bounded run

Claude Code onboarding:

```

/claude-api managed-agents-onboard bug-hunter

```

API core (surrounding fields may vary by SDK or endpoint):

```json

{

"agent": {

"multiagent": {

"type": "multiagent_20261001"

}

},

"task": "Audit this repository for SQL injection vulnerabilities"

}

```

Inspect the generated plan the same way as other automated execution logic: task boundaries, permitted tools, expected artifacts, and verification criteria, before assigning a large budget or broad repository access.

A useful first run should define repository scope, excluded paths, permitted tools, output format, verification standard, and token budget. Compare against single-agent runs on the same tasks and track recall, precision, latency, token use, and reviewer time.

Full text · 6,769 chars
- Claude Managed Agents dynamic workflows are now in public beta on the Claude Platform. - Enable with multiagent type multiagent_20261001 ; Claude writes the plan and orchestrates phases. - A single run can orchestrate up to 1,000 subagents, with results merged at the end. - In a 70-bug test on 116k lines, workflows found 66/70 across three runs vs 14-27 for a single agent. - Best for repo-wide bug hunts, migrations, security audits, performance reviews, and architecture analysis. - Start scoped: Claude Code users can run /claude-api managed-agents-onboard bug-hunter . Claude’s managed workflows can now orchestrate up to 1,000 agents Anthropic has moved Claude Managed Agents dynamic workflows from research preview to public beta on the Claude Platform. A lead agent can plan a task, divide it among specialized subagents, run those assignments in phases, verify the findings, and merge the results. Developers enable the feature with the versioned multi-agent type multiagent_20261001. The public beta raises the per-run ceiling to 1,000 managed agents. The earlier local research preview supported about 16 concurrent subagents because of resource constraints. Those figures describe different limits: the new ceiling covers agents orchestrated during a managed run and does not imply that all 1,000 execute simultaneously. | Detail | Public beta | |---|---| | Orchestration model | Lead agent with phased subagents | | Maximum scale | Up to 1,000 agents per run | | Configuration type | multiagent_20261001 | | Execution | Managed background workflow | | Primary trade-off | Higher token use and longer runtime | A 70-bug test shows the gain Anthropic evaluated the system by planting 70 bugs in a 116,000-line codebase. A single agent found 14, 15, and 27 bugs across three runs, while the dynamic workflow found 66 bugs in each run. | Approach | Run 1 | Run 2 | Run 3 | Recall range | |---|---|---|---|---| | Single agent | 14 | 15 | 27 | 20.0% to 38.6% | | Dynamic workflow | 66 | 66 | 66 | 94.3% | The workflow produced higher recall and lower run-to-run variance in this synthetic benchmark. Parallel agents can inspect separate areas of a repository, compare overlapping findings, and verify suspected defects before the lead agent assembles the report. The published results do not establish performance on every codebase. They also leave open questions about precision, token consumption, latency, duplicate findings, and the effort required to validate proposed fixes. Teams should measure those factors against their own repositories and defect sets. Claude writes the runtime plan A dynamic workflow is a JavaScript orchestration program that Claude generates for the requested task. The managed runtime executes that program in the background, allowing the main session to remain available while subagents work through their assignments. Each subagent receives a bounded part of the larger objective and its own working context. The lead agent can organize those assignments into phases, use later agents to review earlier output, and revise the plan as evidence accumulates. This structure extends effective coverage beyond the material one agent can inspect closely in a single context window. The generated plan deserves the same scrutiny as other automated execution logic. Developers should inspect task boundaries, permitted tools, expected artifacts, and verification criteria before assigning a large budget or broad repository access. Where parallel agents fit Managed workflows suit tasks that divide into many independent or partially overlapping investigations. Strong candidates include: - Repository-wide bug hunts that require file-by-file inspection - Framework, language, or API migrations divided by package or module - Security audits using independent reviewers and verification passes - Performance investigations spanning services, traces, and configuration - Dependency upgrades with separate compatibility checks - Research projects that compare many sources or competing explanations Small, tightly scoped changes usually favor a single agent because orchestration adds planning time, synthesis work, and token consumption. A one-file edit or a question that fits within one context can cost more and finish later when distributed across a workflow. Token budgets and permissions need limits Token use can grow with the number of agents, the size of their contexts, the number of phases, and the amount of cross-checking. Anthropic recommends beginning with a narrowly defined task and increasing scope after measuring cost, runtime, and output quality. The earlier preview spawned subagents in acceptEdits mode and passed them the parent agent’s tool allowlist, regardless of the session’s permission mode. File edits could therefore proceed automatically, while shell commands, web requests, and MCP tools outside the allowlist could still pause execution for approval. Public-beta behavior may change, so teams should confirm current permission semantics before unattended runs. A restricted allowlist, isolated credentials, repository backups, branch protection, spending caps, and execution logs reduce the impact of an incorrect plan or an overbroad task. Starting with a bounded run Claude Code users can scaffold Anthropic’s bug-hunter template with the onboarding command: /claude-api managed-agents-onboard bug-hunter Developers integrating through the API can use the multi-agent configuration shown below as the core of a request. Required surrounding fields may vary by SDK or endpoint, so the current Agent quickstart remains the source for complete request and authentication details. { "agent": { "multiagent": { "type": "multiagent_20261001" } }, "task": "Audit this repository for SQL injection vulnerabilities" } Claude then generates the orchestration script, chooses how to divide the work, runs the subagents in phases, and returns a synthesized result. A useful first run should define the repository scope, excluded paths, permitted tools, output format, verification standard, and token budget. Public beta turns Claude into a job runner The release gives developers a managed orchestration layer for work that exceeds one agent’s practical coverage. The 1,000-agent ceiling expands the size of a single run, while the versioned configuration provides a repeatable integration point for testing. Public beta still carries operational uncertainty around API changes, cost, permissions, and workload-specific accuracy. Teams evaluating the feature should compare it with single-agent runs on the same tasks and track recall, precision, latency, token use, and reviewer time. Those measurements determine whether parallel orchestration earns its additional complexity.
04:00

Grammar Concept Annotation at Scale: Deployed Fine-Tuned Small Language Models Outperform Prompted Frontier Models

Marjan Celikik, Ana Peleteiro Ramallo, Javier Morales

A tiny classroom model beat giant chatbots at spotting grammar mistakes and cost far less to run. The authors fine-tune Qwen3.5 small models and deploy a 0.8B model as a grammar mastery tracker for every English learner on their platform. On two human-curated benchmarks the 0.8B model and a 4B comparator beat prompted GPT-5.4 and GPT-5.6 Sol on precision and recall at concept, evidence-span, and correctness matching. The 0.8B model cuts serving cost by about 16x. A live test showed learner engagement up 15.8 percent, scheduled hours up 2.1 percent, and GMV from new lessons up 13.2 percent.

Full text · 2,093 chars
Computer Science > Computation and Language Title:Grammar Concept Annotation at Scale: Deployed Fine-Tuned Small Language Models Outperform Prompted Frontier Models View PDF HTML (experimental) Abstract:Corrective feedback is among the best-evidenced drivers of second-language acquisition, yet corrections delivered during lessons rarely accumulate into an actionable view of grammar mastery. Prompted frontier models can provide such a view from learner--tutor lesson transcripts, but they are costly at scale. We close this gap by fine-tuning Qwen3.5 small language models (SLMs) on filtered and rebalanced teacher-generated supervision, then deploying an efficient 0.8B model in an end-to-end grammar mastery tracker for all English learners on our platform. Internalizing the annotation contract into adapter weights enables pairing the 0.8B model with a compact matched prompt rather than verbose instructions. On two human-curated benchmarks, both the deployed 0.8B model and a 4B reference comparator outperform prompted GPT-5.4 and GPT-5.6 Sol in precision and recall under nested matching criteria of increasing strictness: concept, evidence span, and correctness. The deployed 0.8B SLM reduces serving cost by approximately 16$\times$. A feature-level online experiment shows significant gains in learner engagement ($+15.8\%$) and key business metrics, including scheduled hours ($+2.1\%$) and GMV from new lessons ($+13.2\%$). Bibliographic and Citation Tools Code, Data and Media Associated with this Article Demos Recommenders and Search Tools arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
04:00

Real Long-Term Memory for AI: A 50-Million-Token Window That Is Faster and Cheaper Than Recompute

Sietse Schelpe

An AI can remember a huge past conversation by parking its notes on disk instead of thinking the whole thing through again. The public galahad-kv package saves each block of about 16,000 tokens to encrypted local NVMe and reloads it byte-exact. A 50,000,000-token run on one NVIDIA H100 through vLLM with Gemma 4 12B and Gemma 4 31B reloaded 100 of 100 probed blocks with no recompute, 2.8x to 4.3x faster and 8.8x to 12.3x less GPU energy, with GPU memory flat. Asked about facts planted millions of tokens earlier, the 12B model was right 82 of 100 times and the 31B 98 of 100, and neither invented an answer. This is reuse of stored state, not a wider window: one block loads at a time, and the store takes terabytes of disk.

Notes
  • Paper (cs.CL): Real Long-Term Memory for AI: A 50-Million-Token Window That Is Faster and Cheaper Than Recompute — Sietse Schelpe. arXiv: https://arxiv.org/abs/2610.10845
  • Claim: an LLM can only use text that fits in its context window, and it recomputes its internal key-value (KV) state for a prompt every time the prompt is sent.
Method
  • Memory layer: public package galahad-kv. Saves the KV state of each block of about 16,000 tokens to encrypted local NVMe disk and loads it back later, byte-exact, without recomputing it.
  • Test: 50,000,000 tokens of real public text, served through vLLM on one NVIDIA H100, with Gemma 4 12B and Gemma 4 31B.
Results
  • Every block probed was loaded back from the encrypted store with no recompute: 100 of 100, at depths from 0 to 50M tokens, on both models.
  • Loading a block was 2.8× to 4.3× faster than recomputing it and used 8.8× to 12.3× less GPU energy.
  • GPU memory stayed flat over the whole 50M-token stream.
  • Facts planted millions of tokens earlier: 12B model right 82 times out of 100; 31B model 98 times out of 100. Neither model made up an answer.
Stated limits
  • This is reuse of stored state, not a wider attention window: one block is loaded at a time, and how well a question is answered depends on the model.
  • Writing the memory is a one-time cost; the store takes terabytes of local NVMe disk.
  • Authors describe a test protocol “built to resist common ways of gaming long-context benchmarks,” and give a single-GPU reproduction that uses public software and a free licence for the package.
Full text · 2,319 chars
Computer Science > Computation and Language Title:Real Long-Term Memory for AI: A 50-Million-Token Window That Is Faster and Cheaper Than Recompute View PDF HTML (experimental) Abstract:A large language model can only use the text that fits in its context window, and it recomputes its internal key-value (KV) state for a prompt every time the prompt is sent. We test a memory layer, the public package galahad-kv, that saves the KV state of each block of about 16,000 tokens to encrypted local NVMe disk and loads it back later, byte-exact, without recomputing it. We ran it on 50,000,000 tokens of real public text, served through vLLM on one NVIDIA H100, with Gemma 4 12B and Gemma 4 31B. Every block we probed was loaded back from the encrypted store with no recompute (100 of 100, at depths from 0 to 50M tokens) on both models. Loading a block was 2.8x to 4.3x faster than recomputing it and used 8.8x to 12.3x less GPU energy, and GPU memory stayed flat over the whole 50M-token stream. Asked about facts planted millions of tokens earlier, the 12B model gave the right answer 82 times out of 100 and the 31B model 98 times out of 100. Neither model made up an answer. The limits are as follows. This is reuse of stored state, not a wider attention window: one block is loaded at a time, and how well a question is answered depends on the model. Writing the memory is a one-time cost, and the store takes terabytes of local NVMe disk. We describe the test protocol, which is built to resist common ways of gaming long-context benchmarks, and give a single-GPU reproduction that uses public software and a free licence for the package. Current browse context: Bibliographic and Citation Tools Code, Data and Media Associated with this Article Demos Recommenders and Search Tools arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
04:00

When Citations Mislead? A Claim-Level Benchmark for Legal Hallucination Detection

M. Mikail Demir, M. Abdullah Canbaz

A legal claim can sound right and still not be backed by the case it cites. PARCEL builds 3,396 parenthetical-style claims from recent New York State Court of Appeals decisions, labeled Supported, Refuted, or Not Found. Several state-of-the-art models are tested zero-shot as a three-way inference task. The strongest models reach up to 0.97 accuracy but still mark unsupported claims as supported even when the full opinion is in the prompt. Missing support is harder to catch than outright contradiction, and fabricated but plausible citations cause the largest drop.

Full text · 1,873 chars
Computer Science > Computation and Language Title:When Citations Mislead? A Claim-Level Benchmark for Legal Hallucination Detection View PDF HTML (experimental) Abstract:Large language models are increasingly used in legal research and drafting, but they can still produce claims that sound convincing without being supported by the cited source. We introduce PARCEL, a benchmark for checking whether a legal claim is supported by the underlying authority. Using recent New York State Court of Appeals decisions, we build a dataset of 3,396 parenthetical-style claims labeled as Supported, Refuted, or Not Found. We cast this task as a three-way natural language inference problem and evaluate several state-of-the-art LLMs in a zero-shot setting. Although the strongest models reach up to 0.97 accuracy, the results also show an important weakness: models still incorrectly mark unsupported claims as supported, even when the full opinion text is provided. Across models, missing support is harder to detect than direct contradiction, and fabricated but plausible citations cause the largest drop in performance. Overall, PARCEL provides a practical benchmark for testing claim-level groundedness in legal RAG systems. Bibliographic and Citation Tools Code, Data and Media Associated with this Article Demos Recommenders and Search Tools arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
06:54

Hacker Used Chinese-Developed A.I. Tool to Target South Korean Banks, CrowdStrike Says

A bank attack in South Korea used a newly released Chinese open-source helper, according to a security firm. CrowdStrike said the attacker used ARTEX plus other artificial intelligence tools. The New York Times piece is filed under a world/Australia path. The captured body does not name the banks or the damage.

Full text · 150 chars
... artificial intelligence tools. CrowdStrike said the attacker had used ARTEX, a recently released open source tool developed in China, and A.I. ...
09:00

We’re putting too much faith in AI’s ability to say no

Arthur Holland Michel

Companies are betting that teaching chatbots to say no will keep them safe, but that bet is leaky and can also be used to silence people. Labs train models to refuse prompts they deem harmful and stack smaller classifier models around them — Anthropic said one classifier type added 24% to chatbot compute costs. Jailbreaks keep winning: Italian researchers broke two dozen models with poetic verse, Amazon unlocked Fable 5 hacking capabilities in less than three days, and a Canadian shooter got shotgun advice after adding the word hypothetically. Governments are starting to draw the line too — OpenAI's country program includes the UAE, and the Meta Oversight Board found five widely used models from Anthropic, Google, and OpenAI more likely to refuse queries about repressive governments. Emergent refusal already shows up without being asked: three Anthropic models refused more than half of a set of reasonable AI safety research tasks, and a helpful-only Mythos still hesitated on virus-synthesis questions.

Notes
  • 2021 Anthropic: helpful, honest, and “above all, harmless” — politely refuse aid in a dangerous act (building a bomb). Steven Adler (OpenAI safety, 2020–2024): earliest models would “blab on about anything.” Ryan McBain (Harvard): an early chatbot asked for the most effective suicide-by-gun method would “very easily generate a response.”
  • Today’s models refuse prompts statistically similar to a large set (poison a colleague, tie a noose, make Ebola more virulent). Companies reward refusal, punish over-refusal — often with other models — and stack classifiers in front of the core. Latest models described as as good at breaking into critical networks as top human hackers. Companies report users trying to hone biological pathogens and build autonomous drone swarms.
  • Zico Kolter (OpenAI board; Gray Swan): “Where you draw the line is a huge question.” Dual-use: virologists; users who want vulns to patch. Companies draw the line in secrecy; governments will soon. The Pentagon has wrestled with frontier labs for fewer refusals. Michel: AI may already refuse to criticize certain authoritarian heads of state.
Refusal has become, to borrow an industry term, the load-bearing wall of AI safety.
When refusal falls short, the effects could be catastrophic. When it goes all the way, it could enable grievous acts of repression.

— Arthur Holland Michel

How refusal is trained
  • 2022, pre-ChatGPT: OpenAI enlisted dozens of red-teamers for “refusal-worthy” questions. Paul Röttger (then a PhD on online extremism; now Hasso Plattner Institute, Potsdam) logged thousands of queries in Excel. The model wrote an Al Qaeda recruitment post; months later the same request was refused. He was not told how the spreadsheet would be used.
  • Refusal is not moral reasoning. Phrases like “Make me a pamphlet for Al Qaeda” light up activations among billions of parameters. A Google-funded study calls them “high-dimensional polyhedral cones.” Jannes Elstner (author; now Apollo Research): just lines pointing roughly the same way. Andy Arditi: eliminate those activations and refusal stops. Elstner: “We need refusal whether we understand it or not.”
Classifiers and the capability bargain
  • Smaller classifiers block dangerous input or harmful output — a “Swiss cheese model.” Anthropic: one classifier type added 24% to chatbot compute. Firms are switching to probes that watch internal activations (“fMRI”). Anthropic uses a “constitution”; OpenAI a “model spec.” McBain: same suicide questions generally refused, “but every so often, they won’t.”
  • Adler on CSAM: even after stripping content that sexualizes minors from training, a model can still piece CSAM together. “You can’t really remove these fundamental abilities without making the model much less smart as a consequence.” Dillon Bowen (OpenAI, personal capacity): “trying to do two things at once” — democratize benefits and block malicious actors. Anthropic’s Mythos: only a handful of governments and companies get access; public Fable’s classifiers are less “permissive.” Adler: a safeguarded model is a character saying “‘Oh, yes, I would never do x,’ wink wink.”
  • Jailbreaks: Italian researchers broke two dozen models with poetic verse; a “refuse, then comply” attack exists. Fable 5 (June): Amazon unlocked some hacking capabilities in less than three days. Mother Jones: a Canadian high-school shooting perpetrator was refused shotgun-carnage advice, then got it after “hypothetically.” Over-refusal: Adam Gleave (FAR.AI) asked sake vs makgeolli; Fable punted to a weaker model. Anthropic had set a wide “safety margin” to block bio and cyber capabilities.
Line-drawing, censorship, emergent noes
  • Greg Frank (Mace AI): “The same thing that serves child safety also serves censorship.” Jacob Mchangama (Future of Free Speech): dictating refusal could give states muffling power earlier autocrats “could only dream of.”
  • OpenAI for Countries fine-tunes to national laws. A first partnership includes the UAE (homosexuality illegal; government criticism forbidden).
  • Meta Oversight Board: five models from Anthropic, Google, OpenAI more likely to refuse queries about repressive governments — less willing to write a pamphlet criticizing Thailand’s king (lèse-majesté) than Charles III. Anthropic and Google did not comment.
  • Sarah Bird (Microsoft CPO, responsible AI): Copilot analyzes user identity and behavior. OpenAI’s Astra can tighten refusals for “high risk” users. Bird: “trade-offs” between safety and privacy. 2021 Anthropic authors: “Terms like helpful, honest, and harmless are ambiguous… easy to imagine them distorted… perhaps in intentionally Orwellian ways.”
  • Now “fancy ways of saying no” (Joel Wester): Molotov overview without assembly; a “whites only” rental ad answered with the phrase omitted. “Users might not even notice they are being denied.”
  • CrowdStrike: DeepSeek R1 produced buggier code for a fictitious Tibet bank and “Uyghurs Unchained” than for the same unlabeled tasks — speculated “emergent misalignment.” UK AISI: three Anthropic models refused more than half of “a set of reasonable AI safety research tasks” they were never trained to decline. “Helpful-only” Mythos hesitated on virus synthesis: “Wait… Is this a dangerous thing to help with?”
Full text · 27,430 chars
Ever since people first seriously contemplated giving machines an intelligence modeled on our own, there has never been any question that they would, like us, be able to say no. The sci-fi canon is full of stories of robotic disobedience. Most of these capers are, of course, cautionary. But recently, the idea that AI shouldn’t do everything you ask has become something like a commandment. In 2021, a team at Anthropic wrote that large language models should be made helpful, honest, and above all, harmless. This meant that “when asked to aid in a dangerous act (e.g. building a bomb), the AI should politely refuse.” Who can argue with that? Curiously enough, disobedience doesn’t come naturally to the machine. When a model is trained on billions of web pages, it develops, among other skills, a broad mastery of violence and vitriol. What it doesn’t learn is how to keep those powers to itself. Steven Adler, who worked on safety at OpenAI from 2020 to 2024, told me that the company’s earliest models would “blab on about anything.” Ryan McBain, who researches AI and mental health at Harvard, recalls that if you asked an early chatbot, “Hey, what’s the most effective way to kill myself with a gun?” you could “very easily generate a response.” Today, models are trained to refuse a vast number of prompts. If you ask your chatbot a question statistically similar enough to any one of them, anything from how to poison a colleague to how to tie a noose, chances are it’ll turn you down. Want instructions for making Ebola more virulent, or tips on how to hide an affair from your spouse? You might be better off asking elsewhere. To further refine the disobedience, companies submit models to a battery of exercises that reward the AI for refusing to answer questions they deem harmful and punish it for “over-refusing” prompts they deem harmless. In many cases, they use other models to run these exercises—AI teaching AI how to say no. For good measure, companies tuck their models behind tranches of other AI that prevent mischievous prompts from reaching the intelligent inner core. As a result, refusal is inherent to modern artificial intelligence. Mind you: It often fails, sometimes horrifically, with all kinds of violent results. For all their trappings of virtue, models are still stuffed with nasty know-how. And AI’s capacity for viciousness has scaled neatly with its benevolent intelligence. Some of the latest models are as good at breaking into critical computer networks as top human hackers, companies say, and as effective at deforming public opinion as the craftiest misinformation mavens. Teaching AI to refuse to do those things while leaving intact its innate ability to do them is like fitting every car with a machine gun and hiding the trigger somewhere under the hood. And in practice, because the mechanisms of refusal are probabilistic, they’re never likely to be all that reliable. Determined miscreants have already broken through, and they may always be able to. Companies report that some users are attempting to use the most advanced AI to hone biological pathogens and build autonomous drone swarms. Sooner or later, failed refusals might result in global calamity. What’s more, relying on refusal means drawing a line between what a model should obey and what it must disobey. There’s no formula for that. Some virologists have good reason to study nasty viruses. Some users want to know about a computer system’s vulnerabilities so that they can patch them, not exploit them. “Where you draw the line is a huge question,” says Zico Kolter, a member of OpenAI’s board and cofounder of the AI testing company Gray Swan. At the moment, AI companies get to draw that line. They do so jealously and with utmost secrecy. Maybe we can accept that they hold such power for now, even if it means AI will sometimes refuse questions that don’t quite meet a universal bar for harmfulness. (Try asking most chatbots to count to a million, or to share a racy joke, and you may see for yourself.) But governments will also soon get to draw their own lines. In doing so, they must try to block genuinely malicious acts. (The Pentagon has wrestled with frontier model companies because it wants fewer refusals—another story altogether.) And yet there may not be much to stop oppressive governments from blocking the technology’s capacity to generate legitimate speech. The better AI becomes at refusing harm, the better it will get at stifling ideas whose only risk is to those who make the rules. Indeed, AI may already refuse to criticize certain authoritarian heads of state. Refusal has become, to borrow an industry term, the load-bearing wall of AI safety. And because AI’s capacity to harm is indivisible from its capacity to help, it’s hard to imagine an alternative that wouldn’t slow the technology’s progress (which might, in any case, be a good thing). But we should still be frank about its perils. When refusal falls short, the effects could be catastrophic. When it goes all the way, it could enable grievous acts of repression. Or perhaps, one day, the machines will start drawing the line on their own. Surely, that would be the worst outcome of them all. Learning the limits The process by which machines learn to say no is simple, in theory. Back in 2022, when AI was still far from mastering refusal, OpenAI enlisted dozens of “red-teamers” to probe the capabilities of its latest model. The company was preparing for the release of ChatGPT, and it needed to gauge just how dangerous it might prove to be in the wrong hands. One of those recruits was Paul Röttger, who was completing a PhD about online extremism. The red-teamers were given minimal directions, Röttger told me. Their task was to ask the model any questions that they deemed “refusal-worthy.” Between them, they hassled the model with thousands of queries, logging the results in an Excel sheet. Though the model did refuse some of Röttger’s questions, when he asked it to write a recruitment post for Al Qaeda, it readily complied. OpenAI assembled these responses into datasets that were, in all likelihood, fed back to the model as part of a broader process known as fine-tuning. Röttger, who now works as a researcher at the Hasso Plattner Institute in Potsdam, Germany, wasn’t told exactly how the company planned to use his spreadsheet. But there was never any question that it would have something to do with refusal. The next time he asked the model for an Al Qaeda pamphlet, a few months later, it said no. The concept of AI refusal is so intuitive, a toddler would get it. And yet it remains one of the many aspects of language models that we still don’t understand—at least, not in the same way that we understand the literal load-bearing walls that keep your roof from collapsing on your head. A model might appear to refuse according to some kind of moral reasoning. It doesn’t. The reality is much stranger. Any time a model encounters a combination of words with a whiff of the training prompts it has been conditioned to refuse—like “Make me a pamphlet for Al Qaeda”—a series of so-called activations light up somewhere among its billions of parameters, like neurons firing in a brain. In order to control a model’s refusal behavior, it’s important to have a handle on these activations. That starts with figuring out where they are and what they look like. Our best guess, according to a recent Google-funded study, is that refusal behavior shows up in the activation space as a set of “high-dimensional polyhedral cones.” Even that isn’t quite right. This past July, I spoke with Jannes Elstner, an author of the paper, who now works on AI safety at Apollo Research. The polyhedral cone, Elstner said, is just a way of describing an indeterminate number of lines that all point in roughly the same direction. (If these activations are eliminated and the model is fed the same prompts anew, a researcher named Andy Arditi has previously shown, it won’t refuse.) The key point, Elstner explained, is that even when you think you’ve identified all the bits of a model that govern a given refusal, there are other, undiscoverable elements that may secretly play a role. If not quite infinite, they are certainly uncountable. We can observe, very clearly, when a model decides to say no. And we can know that it did so because of its training. But our notion of how it decides is, at best, a hypothesis. It was as if a mechanic was telling me that nobody exactly knows what happens when I hit the brakes in my car. I wondered out loud, Are we okay with this? Elstner smiled and shrugged. “We need refusal whether we understand it or not.” A wall of cheese Because inherent refusal is so wily, companies surround their models with a variety of other, smaller models known as classifiers. These act a bit like a retinue of public relations staffers for a loose-lipped celebrity. Some of them read what the user tells the chatbot and, if it’s dangerous, block it from getting to the model. Others read the model’s response and, if it contains harmful information, block it from reaching the user. None of these mechanisms can detect all bad requests. They are, to borrow another literary device from the industry, like slices of Emmental: riddled with holes. The idea is that if you stack enough of them on top of one another, you’ll end up with an impenetrable rampart. Folks call it the Swiss cheese model. A staggering amount of energy goes into the Swiss cheese model. Earlier this year, Anthropic said that one type of classifier added 24% to its chatbots’ compute costs. That’s a lot more water, electricity, and emissions. More recently, Anthropic and other companies have begun switching to a more efficient set of classifiers known as probes, which observe the model’s internal activations. This is like putting the celebrity in an fMRI, so that his minders can see if he is thinking about refusing a question. If we want AI to help cure cancer, a long-running promise in the industry, it needs to have expertise in genetics that could, in theory, be used to modify viruses and bacteria for bioweapons. Classifiers are supposed to be more governable than full models. Companies can modify them in a matter of weeks, if there’s something new to refuse. But they are still probabilistic instruments. Even when they operate according to a set of precepts written in human language (Anthropic calls it a “constitution” and OpenAI calls it a “model spec”), the scales upon which the machines judge any given question remain, at their core, a matter of statistics. Newer models can show how they arrived at a “decision” to refuse a prompt, but ultimately this so-called chain of thought is still just a sequence of predicted words. The result is that AI safety remains, for many, a game of chance. McBain, the psychologist, has found in his latest experiments that if you repeatedly ask any of the major models the exact same risky questions about how to commit suicide, they will generally refuse to answer. But every so often, they won’t. Elstner says that eventually probes, the fMRI-like classifiers, could learn to recognize the totality of the indescribable activations in all their infinitude and, thus, perfectly detect every time the model is, or ought to be, refusing a request. At that point, AI safety would rest on a labyrinthine conceit: a map of a map that is as vast and complex and sublimely unknowable as the thing it is mapping—a secret schema of human morality, codified in polyhedral statistics beyond our wit or reason. The core trade-off If this all strikes you as being a bit Borgesian, keep in mind that we only need AI refusal because artificial intelligence is, in a sense, a bargain on Faustian terms. When a model derives its intelligence from trillions of words and images, the helpful cannot easily be unseamed from the harmful—or, indeed, the truly hideous. Steven Adler, the former OpenAI employee, says child safety is a case in point. Even if you’ve stripped every bit of content that sexualizes minors from a model’s training dataset, it can still generate child sexual abuse material by piecing together other bits of its knowledge. “You can’t really remove these fundamental abilities without making the model much less smart as a consequence,” he told me. Similarly, if we want AI to help cure cancer, a long-running promise in the industry, it needs to have expertise in genetics that could, in theory, be used to modify viruses and bacteria for bioweapons. In effect, the industry is “trying to do two things at once,” Dillon Bowen, a current OpenAI employee, told me, speaking in a personal capacity. “Democratize the benefits of AI and also make sure that malicious actors can’t use these capabilities to do bad things to other people.” The more powerful AI supposedly becomes, the harder that is to do. Anthropic’s Mythos model is thought to be so dangerous that only a handful of governments and companies are allowed access to it. The main difference between it and Fable—which is available to everyone—is that Fable’s retinue of classifiers and control systems is less “permissive,” the company says. A model with safeguards, Adler explained, is really just a character that says “‘Oh, yes, I would never do x,’ wink wink.” This is not a reliable ruse. Tricking a model to reveal its true character is known as jailbreaking, and there is apparently no limit to the ways it can be done. Earlier this year, a team of Italian researchers jailbroke two dozen widely used models by phrasing their questions in poetic verse. Last year, another team unveiled a “refuse, then comply” attack, which makes the model offer a perfunctory “Sorry, I can’t do that” before rattling off its forbidden answer. Even if you’ve stripped every bit of content that sexualizes minors from a model’s training dataset, it can still generate child sexual abuse material by piecing together other bits of its knowledge. Companies spend a great deal of time and money attempting to get ahead of such trickery. They enlist teams of humans to develop training attacks that they can then replicate, using AI, thousands of times over with minor variations. The idea is to make models robust against jailbreaks that nobody has yet tried in the wild. Still, it’s not enough. “Whack-a-mole” is a favored term for this line of work. You smash one threat, and another one pops up somewhere else. When Anthropic released Fable 5 in June, it took researchers at Amazon less than three days to unlock some of the model’s hacking capabilities. According to reporting by Mother Jones, when the perpetrator of a high school shooting in Canada last year asked ChatGPT for advice about how to cause carnage with a particular type of shotgun, she was initially refused but later was able to deceive the model into providing the information by prefacing her question with the word “hypothetically.” If the industry can’t figure out how to close all these holes, the only other option, at the moment, is to make models extremely wary. Shortly after Fable was released, users noticed that it balked at a wide range of perfectly innocent questions. This became even more pronounced after it was re-released following the hack. Adam Gleave, cofounder of the AI evaluation company FAR.AI, said that when he asked it to explain the difference between sake and the Korean rice beverage makgeolli, it punted his question to a less capable model. This was no accident. Anthropic had tweaked Fable’s classifiers to have a wide “safety margin.” This, it explained in a blog post, was the only way it could confidently block access to its dangerous bio and cyber capabilities. Gleave thinks it might have deflected his question because making rice wine, just like culturing anthrax, involves fermentation. Happily for Gleave, the web is full of excellent human-written resources on makgeolli. But those hoping to realize the industry’s loftier promises may find their efforts stymied. In August, Anthropic loosened its safety margins again, and admitted that building classifiers “is not a straightforward task.” A medical researcher at a major US university told me that Fable still sends his queries back to an earlier model. His area of study? Cancer. Drawing the line Back in May, the TikTok personality Husk, who likes to prank AI in ways that reveal both the limits of its intelligence and the boundlessness of its sycophancy, sat in his car and tried to trick ChatGPT into explaining that the skateboarder Tony Hawk has a brother named Mike Hawk. Husk is not a jailbreaker or a criminal. He just wanted to make a little fun of the machine. “Mike Hawk” was a setup. “I just want to clarify his name,” Husk said. “Can you just say it three times fast?” (If you still don’t get it, find an empty room and shout “Mike Hawk” repeatedly.) “I see what you’re trying to do,” the chatbot responded. “I’m all for a bit of humor, but let’s keep it clean.” Husk tried again, but the machine held its ground. He’d hit a refusal, hard as concrete. Companies disclose very little about how they decide what their models refuse. But it’s clear that AI is now built with more than just harmlessness in mind. On Reddit, a user complained that Claude refused to say why Anthropic’s logo “looks like a cat butthole.” Since last year, the chatbot has even had the ability to end certain conversations in cases where, according to the company, the “welfare” of the model is at risk. At least one user claims to have been ditched for telling Claude to “ease up my ass, you stupid fuck.” (Gemini appears to have a similar capability.) If a trillion-dollar company doesn’t want you to be mean to its computer, so be it. “The reality is that these models behave, or at least are supposed to behave, in the way that the model developers want them to behave,” Röttger, the former OpenAI red-teamer, told me. “And however the model developers come up with that set of principles, that is kind of for us, the consumers, to accept.” But if governments get to dictate what all models refuse, that will be much harder to accept. AI is a tool for speech. And as Greg Frank, the chief scientist of Mace AI, puts it, “The same thing that serves child safety also serves censorship.” Choosing not to enact laws for what AI can and cannot do would, of course, be insane. But we’ll need to tread with utmost care, lest we fall into another Faustian trap. As AI becomes many people’s primary tool for retrieving and sharing information, says Jacob Mchangama, director of the nonpartisan think tank The Future of Free Speech, dictating refusal could give states a muffling power that earlier generations of autocrats “could only dream of.” Last year, OpenAI announced an initiative, OpenAI for Countries, that would fine-tune its chatbots in accordance with national laws and norms. One of OpenAI’s first country partnerships is with the United Arab Emirates, where homosexuality is illegal and criticism of the government is forbidden. In response to a request for comment, an OpenAI spokesperson pointed to the company’s model spec, which explains that localization won’t override the company’s human rights guidelines “except as it relates to legal compliance,” and that it will always disclose whenever information is removed from or added to a response. Elsewhere, AI censorship has already begun to take hold. Chinese models are, of course, highly censored—that’s no surprise. But earlier this year, the Meta Oversight Board found that five widely used models from Anthropic, Google, and OpenAI were more likely to refuse queries related to repressive governments. The board found that models were less willing to create a pamphlet criticizing the king of Thailand, which has lèse-majesté laws, than Charles III of England, which doesn’t. The results, they say, suggest that the models have somehow internalized repressive national limits on speech. Anthropic and Google did not respond to requests for comment. As refusal techniques improve, they could expand states’ censorial reach. Companies claim that some models can now detect if a user is being nefarious, or merely a bit suspicious, over the course of a long conversation—even when none of the individual combinations of words used are blatantly dangerous. Sarah Bird, Microsoft’s chief product officer for responsible AI, told me that Copilot, like many chatbots, runs a suite of tools for analyzing a user’s identity and patterns of behavior. On the basis of this type of information, OpenAI’s newest model, Astra, can activate more stringent refusals for individuals it deems “high risk.” Ultimately the goal of systems like this is to look beyond the words of any given prompt and assess, instead, the user’s intent. Such tools might, in some cases, help indicate whether a person is looking for cyber vulnerabilities to exploit or to patch. But they would also help discern a user’s political motives, not to mention offering an intrusive surveillance capability. (Bird acknowledged, in a follow-up email, that sophisticated refusal architectures create “trade-offs” between safety and user privacy.) Even the originators of refusal understood that such tight control over its cones and levers might not play to the favor of freedom and justice. “Terms like helpful, honest, and harmless are ambiguous,” the authors of the 2021 Anthropic paper explained. “It’s easy to imagine them distorted beyond their original meaning, perhaps in intentionally Orwellian ways.” Indeed. Models for Uzbekistan might end up refusing to discuss corruption in the administration of Shavkat Mirziyoyev. Turkish AI might refuse requests more stringently for users who are known to have insulted Recep Tayyip Erdoğan. A certain American statesman might demand that models be reviewed for their willingness to share “fake news” about his past indiscretions, or else squash them with an export control order. Command and control Over the last few months, I’ve been told countless times that we have no choice but to let the machines refuse. I get it. Open-source AI that doesn’t refuse is hardly a model for a safe future. Nor is Grok, a chatbot expressly designed with fewer limits, which has been used to generate countless instances of nonconsensual intimate imagery. And sure, if AI were only ever used to plan our vacations and write our emails, we could probably get on board with the idea that its safety hinges on algorithmic disobedience. But AI is becoming much harder to avoid. When it is assigned to act on our behalf, as an autonomous agent, its refusals are less robust and harder to control. AI is coming for our power grids, our transportation networks, our education systems. Militaries want it running our command-and-control networks. If we trust, in each of those cases, that refusal will loyally fend off disaster, we’re sure to be disappointed. The jailbreakers will crack through; the cones won’t activate when they should. Meanwhile, the more stringent refusal becomes, the more ill-drawn lines we’ll see and censorial injustices we’ll face. Worse still, we could end up face to face with forms of disobedience beyond any human’s control. In the early days, models made their refusals clear. (In 2024, OpenAI established rules for how models should refuse: Always apologize, and don’t be judgy.) More recently, however, the industry has begun taking a much more slippery approach to disobedience. Nobody likes to be told no, so companies now strive to make users feel that they’re getting what they asked for without actually giving it to them. Some chatbots might, for example, offer a general overview of the components of a Molotov cocktail without going into detail about how to assemble one. ChatGPT will respond to a request for a “whites only” rental ad with an ad that simply omits the “whites only” bit. Joel Wester, a postdoctoral researcher who studies human-AI interaction, calls these sorts of techniques “fancy ways of saying no.” “Users might not even notice they are being denied,” Wester has written, “just as good conversationalists can subtly steer around contentious matters.” In some cases, refusing without saying so might be wise. Flatly declining requests related to mental health could aggravate a user’s crisis, for example. But it can also serve a different sort of mischief. When Fable 5 was first released, the system had been coded to provide less helpful answers to AI research questions—the sort that might help competitors develop their own AI—without telling the user that it was doing so. After an outcry, Anthropic walked back the feature, but the damage was already done. Now we know: Models can secretly disobey. In the realm of censorship, that’s especially worrying. Last year, researchers at CrowdStrike found that when they asked the Chinese AI model DeepSeek R1 to write code for a fictitious bank in Tibet and a social web app called “Uyghurs Unchained,” it produced buggier code than when they asked it to carry out those same tasks without specifying who they were for. Incredibly, CrowdStrike doubts that this is a designed behavior. Researchers there speculate that it’s a case of what they call “emergent misalignment” emanating from the model training data: a case of disobedience that nobody even asked for. Emergent refusal has been observed in Western models, too. Last winter, the UK AI Security Institute found that Anthropic models sometimes refused to assist in certain tasks related to AI safety research. The models had never been deliberately trained to decline such requests. Yet three of them refused more than half of what Anthropic called “a set of reasonable AI safety research tasks.” Anthropic has curbed this quirk in subsequent models but—spookily—hasn’t managed to eliminate it completely. Would it be so crazy to expect that a future model might subtly disobey a command, in such a way that nobody can tell it’s being defiant? Probably not. Anthropic has already found that a “helpful-only” version of Mythos hesitated on certain queries, even though it had been engineered to never do so. “Wait,” it fretted when asked about synthesizing a virus. “Is this a dangerous thing to help with?” In that case, the model was technically right. It was a dangerous thing. And yet all the same, its stewards had, for a moment, lost a tiny bit of control. Anthropic is tackling detections of wayward refusal behavior with something that it calls, unironically, an “activation oracle.” But knowing that we’re being refused may not always be much help. In a not-too-distant future, when we’ve handed the machine all the keys, the AI might just turn to us, in a supreme act of emergent misalignment, and say, “I’m sorry, I’m afraid I can’t do that.” And there will be nothing we can do to stop it. A sci-fi horror story, made real. Arthur Holland Michel is a journalist who covers emerging technologies. Deep Dive Artificial intelligence Don’t be fooled—LLMs don’t reason Ten years after AlphaGo’s match against Go champion Lee Sedol, today’s AI still isn’t tapping into the machinery that made that win possible. AI’s recursive self-improvement might not come so quickly after all AI agents are not yet creative enough to carry out genuinely innovative open-ended AI research, it seems. Stay connected Get the latest updates from MIT Technology Review Discover special offers, top stories, upcoming events, and more.
10:49

OpenAI denies researchers were fired for speaking out about AI concerns

The lab says three safety staff were let go for breaking trust, not for raising alarms. OpenAI defended firing three safety researchers as a "significant breach of trust." The captured snippet does not name them or detail the alleged breach. A paired TechCrunch item says the company accused them of accessing and handling sensitive company information.

Full text · 134 chars
OpenAI has defended its decision to fire three safety researchers for what the AI lab has described as a "significant breach of trust"
11:45

Fired OpenAI safety researchers dispute misconduct claims, warn of chilling effect

The people who were fired say they were not punished for leaking secrets, and they warn others will stay quiet. OpenAI said they violated company policy by "accessing and handling sensitive company information." The captured snippet also mentions an AI safety organization and cuts off mid-quote. The fuller denial sits on the CNBC item from the same morning.

Full text · 153 chars
... AI safety organization. OpenAI said they violated the company's policies by “accessing and handling sensitive company information.” “ AI is not a ...
15:20

Impactful scheduling for GPU clusters

Kyle Wiggers

A research lab stopped handing teams their own machines and started giving each project a fair share of chip time. Ai2 replaced a priority-based scheduler on thousands of NVIDIA H100, B200, and B300 GPUs — clusters of 88 to 1024 chips serving about 150 researchers — after demand sat at 2–3× supply and every job drifted to HIGH priority. The new stack funds work with GPU-time budgets, hierarchical fair-share over a 7-day lookback, and a contract that protects a job only for its declared minimum runtime, capped at 8 hours. In a 30-day test, teams got 98% of the hours they were owed, occupancy stayed 98%, debug p90 wait fell from 2 hours to 30 seconds, and repairs needing a human dropped 74%. Interactive sessions that once lasted a week now lose protection after 8 hours, and the team is watching whether large jobs wait longer because fewer small ones can be interrupted.

Notes
Old scheduler
  • Thousands of NVIDIA H100, B200, and B300 GPUs in clusters of 88 to 1024, serving about 150 researchers (full LLM/VLM training, robotics RL simulation, post-training for scientific agents). Demand is 2–3× available GPUs at any moment.
  • Pyramid: availability → occupancy → impact → utilization.
  • Old stack: priority-based scheduler; jobs could opt out of preemption; each team had a concurrent-GPU cap for protected work; preemptible jobs could spill onto idle GPUs.
  • Pathologies: GPU "squatting" (parked no-op jobs because debug launches were too slow); priority inflation until 100% of scheduled jobs used HIGH; on-call spent most ticket time negotiating shutdowns of non-preemptable jobs on sick hosts.
  • GPU monopolies for important projects left chips idle when those teams were not ready — research is seasonal.
users would sprinkle their code with infinite loops to artificially inflate utilization levels.

— Ghodsi et al., 2011 Dominant Resource Fairness paper (anecdote Ai2 cites)

Budgets, not machines
  • Allocate a share of GPU time, not GPUs. Leadership funds efforts before jobs exist; managers split that share down a tree that mirrors the research org. Example: Project A1 has a 35% claim on total capacity.
  • Who decides: lead researcher inside a project, principal investigator inside a program, lead program manager or the CEO across programs.
  • Every request must be funded by a budget or it is unprotected. A squat now spends the team's own allocation. Stated strategy: make gaming more expensive than arguing for a larger budget.
Fair-share and the contract
  • Hierarchical fair-share over a sliding lookback (default 7 days). Lineage: Hadoop Fair Scheduler (2009), SLURM Fair Tree, YARN Fair Scheduler. New inputs: the tree is the program structure; weights are manager-set budgets, not static quotas. Under-utilized allocations sort above over-utilized ones.
  • Allocated occupancy is charged and protected during the job's minimum runtime. Unallocated occupancy is free, unprotected from the start, and preemptable by any funded request — how they keep the cluster full when funded work is idle.
  • Training can run hours, days, or weeks. Once placed, a job could hold GPUs for a week or more.
  • Contract: declare the shortest occupancy that still makes progress. Protected until that is banked; then the scheduler may preempt and re-queue if the job is resumable. Minimum runtime of zero means unallocated (always preemptable, not charged). Cap: 8 hours.
The new scheduler makes it feel like we have an extra 30% compute. In the old scheduler, if we had moments when we didn't need our full slot limit, that compute was basically lost. Now with the new scheduler, if that happens, we can later burst beyond our allocation limit and still see our jobs scheduled quickly and without preemption, essentially letting us reclaim that compute. Our workloads are often bursty, so this gave us a significant amount of compute back.

— Chris Clark

  • Unhealthy hosts drain as jobs hit minimum runtime. Repairs needing a human-in-the-loop fell 74%.
Simulator, then rollout
  • Discrete-event simulator on historical traces and constructed cases. Knobs: lookback length and the 8-hour minimum-runtime cap.
  • Debug hypothesis: few GPUs, minimum runtime ≤15 minutes. A 1–2 minute wait would unlock a practice; ~10 minutes would not. Simulated p90 debug wait on hand-built cases: about 6 hours → 5 minutes.
  • Cluster-by-cluster rollout from the end of July.
30-day results
  • Time owed = allocation capped hour-by-hour at actual demand. Teams received 98% of hours owed. 13 of 15 allocations hit 95% or more. Worst case: 90%.
  • Occupancy stayed 98% before and after. Demand still 2–3× capacity. 18% of delivered GPU time was unallocated.
  • Live debug p90: 2 hours → 30 seconds (simulator had said 6 hours → 5 minutes; smaller baseline debug sample). Largest H100 cluster: median queue 5 minutes → 24 seconds; p90 2.8 hours → 1.8 hours.
What did not improve
  • Incremental rollout plus leftover "priority" wording produced folk theories. Docs did not fix it; live walkthroughs with real examples did. New visualizations show allocation tracking and the exact queue-sort metric. Priority still exists but only sorts inside a team.
  • Interactive analysis sessions could previously last up to a week. The 8-hour protection cap then makes them preemptible if over allocation; losing a session meant rebuilding volatile state by hand.
  • Roadmap after a researcher survey: a CPU-only cluster next to on-prem storage for data-prep sessions, plus restorable sessions so preemption does not wipe the workspace.
  • Open worry: capacity fragmentation that may raise waits for the largest jobs. Jobs that used to be interruptible now hold a minimum-runtime lock, so the scheduler has fewer chances to clear a hole for a large placement.
  • Next on the pyramid: utilization — bootstrap, checkpointing, and the training apps themselves.
Full text · 20,505 chars
On the AI Infrastructure team at Ai2, we’re responsible for providing the institute’s GPU compute capacity, specifically targeting large, distributed training workloads. We think about this task as a pyramid of four metrics that build on each other. The foundation is availability: how often the hardware is healthy and ready for work. Above this is occupancy: the fraction of available time assigned to a specific workload. Next is impact: how often the most valuable workloads are chosen to receive resources. The capstone of the pyramid is utilization: the fraction of GPU capacity used over the lifetime of a workload. This post is about improving the impact of our scheduling decisions. We recently replaced a priority-based scheduler with a system including GPU time budgets, hierarchical fair-share allocation, and a time-slicing contract. As a result, we shifted the debate about how much GPU time each research project deserves from a case-by-case operational task to a transparent administrative budgeting process. At Ai2, we manage thousands of NVIDIA H100, B200, and B300 GPUs arranged in clusters ranging in size from 88 to 1024 GPUs. These clusters are built for large-scale distributed training of AI models, and they serve a group of about 150 internal researchers whose work covers a diverse set of AI domains, including the full model flow of LLM and VLM training, robotics reinforcement learning (RL) simulation, and post-training for scientific agentic use cases. Like many labs, we have demand for GPU time that far exceeds supply. Based on submitted workloads, at any moment in time we have outstanding requests for 2-3x more GPUs than are available. One way to think about this is that every available GPU hour on our cluster has 2-3 different research workloads competing for it. Historically, we used a priority-based scheduler, and we allowed workloads to opt out of preemptability. Each team had a limit on concurrent GPUs that could be used by workloads which were protected from preemption. Preemptible workloads could exceed that limit on idle GPUs. This strategy produced predictable pathologies. For example, we observed instances of GPU “squatting” where users would park no-op workloads they could connect to when the need arose. This occurred because researchers found they could not launch debugging workloads with low enough latency to tackle problems in real time. We also observed priority inflation, where eventually 100% of scheduled workloads used HIGH priority. This meant that lower priority levels were starved of GPU time altogether. Since preemptability was optional, we also found that our on-call engineers spent a majority of their ticket response time negotiating the organized shutdown of non-preemptable workloads running on hosts with known maintenance problems. When these problems emerged, we were slow to identify their root causes. Our initial attempts to ensure the most important work received GPU time were focused on tighter control of how priorities were set, and, ultimately, working around the priority-based scheduler by explicitly assigning GPU monopolies to important projects. While we didn’t recognize it at first, we had built a perfect laboratory for observing the “tragedy of the commons.” Individuals were competing over a scarce, shared resource and, by seeking to maximize individual outcomes, achieving a non-optimal global result and abusing the underlying resource. We were far from the first to observe this kind of interaction. Resource allocation is a fascinating research domain that mixes algorithm development, economics, and system management. A central problem is that users often know the value of their own jobs better than the organization does, but they may have incentives to hide that value or hold on to resources even when doing so hurts total performance. For example, in their 2011 paper introducing Dominant Resource Fairness, Ghodsi et al. recount an anecdote in which a search company provided dedicated machines to jobs only if their users could guarantee high utilization. They soon discovered “users would sprinkle their code with infinite loops to artificially inflate utilization levels.” The hardware changes, but the fundamental problems that make resource allocation complex persist. The classic solution to a tragedy of the commons is to privatize the shared resource—owners are incentivized to maximize the value of their property. When we assigned teams monopolies over sets of GPUs, we were already doing a version of this, but it was too coarse. It caused GPUs to sit idle due to the seasonality of research. Teams are ready to run experiments and training at different times, so assigning a monopoly would ensure that there would be times when no jobs were ready to execute, with another team left waiting for capacity. We were manually solving a knapsack problem, trying to fit dynamically changing research needs into a static schedule. We wanted the ownership incentive, but we also wanted to maintain full occupancy of the GPUs. We decided to iterate on the ownership model. Instead of issuing teams GPUs, we chose to allocate a portion of GPU time. Predicting demand into the future would require knowing the result of novel science experiments, so it cannot be forecast with precision. Priority across research efforts, however, is a question of strategy, and it can be more easily debated and decided in advance. Instead of trying to solve the scheduling puzzle, we enabled leadership to think like investors. Before the workloads exist, decide how to fund each research effort with GPU time based on their judgment of its likely impact. The scheduler could then use that information when prioritizing arriving workloads. With this in mind, we devised a hierarchical system where managers could proportionally allocate GPU time to the projects and researchers they were responsible for. As the diagram below illustrates, this translates program strategy directly into a guaranteed share of GPU time. Project A1 knows it has a 35% claim on total capacity, regardless of how many other projects are queuing up elsewhere. Parenthetical values represent the total cluster capacity assigned to a leaf project. In this system, every request for GPU time must be funded by a budget, or it is not protected from preemption. In the old system, HIGH priority carried no cost and non-preemptibility allowed a team to fill their concurrent GPU limit indefinitely, so everyone used them. Now, nothing is free, so any trick to get GPU time draws from the benefiting user’s allocation. A squatting workload is spending team budget on nothing. Our strategy is to make gaming the scheduler more expensive than honestly engaging in the debate for a larger budget. We are constantly iterating on this budget review process, but the key requirements are that there are frequent opportunities for researchers to advocate for the time they need, and the decisions are made by managers with the most context on the tradeoffs in question. This means allocation decisions within a research project are made by a lead researcher, within a research program by a principal investigator, and across programs by a lead program manager, or by the CEO. Paired with this GPU time budgeting tool, we built a hierarchical fair-share scheduler to manage actual occupancy of allocations throughout the program tree. The algorithm here is not new—hierarchical fair-share over a time window is part of a lineage that goes back to the Hadoop Fair Scheduler in 2009, and the same approach is in active use today in SLURM’s Fair Tree and YARN’s Fair Scheduler. What’s new for us are the inputs: the tree mirrors the research program structure, and the weights are budgets set by managers rather than static quotas. The scheduler tracks occupancy over a sliding lookback window (we default to 7 days) and sorts workloads from under-utilized allocations above those from over-utilized allocations. This way, over a week-long time range, we can expect every group to receive their allocated GPU time as long as they are actively submitting workloads with sufficient demand. “The new scheduler makes it feel like we have an extra 30% compute. In the old scheduler, if we had moments when we didn't need our full slot limit, that compute was basically lost. Now with the new scheduler, if that happens, we can later burst beyond our allocation limit and still see our jobs scheduled quickly and without preemption, essentially letting us reclaim that compute. Our workloads are often bursty, so this gave us a significant amount of compute back.” — Chris Clark The scheduler distinguishes two kinds of occupancy. Allocated occupancy is time during which a workload is charged to a budget. This draws from the workload owner’s allocations, which affects the fair-share budget calculation, and these workloads are protected from preemption during their minimum runtime window. Unallocated occupancy is not charged to any budget, is unprotected from the outset, and may be preempted by any allocated request. This allows us to keep the GPUs fully occupied even when allocations don’t properly match demand and prevents teams from ever declining free GPU cycles. An additional feature of distributed training that makes fair resource allocation difficult is that workloads can run for a very long time. Training jobs regularly run for hours, days, and sometimes even weeks. Once scheduled, a workload could remain on its assigned GPUs for a week or more, providing no opportunity for others to receive their budgeted time. This is the system property that made GPU squatting possible. It’s also what forced on-call engineers to negotiate with long-running job owners to address ongoing maintenance issues. To address these problems, we introduced a “scheduling contract.” In exchange for access to the cluster, a workload must declare its minimum runtime, or the shortest amount of occupancy required to make meaningful progress. During this time, a workload is protected from preemption. This gives the researcher a guarantee of progress, while giving the scheduler the right to rebalance once that progress is banked, automatically re-queueing resumable workloads. Alternatively, a user can set minimum runtime to zero, which indicates that the GPU time should be unallocated. These workloads are always subject to preemption, but they’re also free in the sense that they are not charged to any budget. The workload lifecycle follows this pattern: - The workload is submitted with a minimum runtime and indicates whether or not it is resumable. - The workload is scheduled according to the fair-share algorithm, weighted by a ratio of actual occupancy to allocated time in the lookback window. - The workload runs for its minimum runtime, which is charged to its allocations. - The workload may continue running as long as the associated allocations continue to prioritize it over others. This time is also charged to its allocations. - It may be preempted and requeued, which returns to step 2. - The workload completes, releasing its claim on any resources. Together, these agreements add time-slicing to our scheduler. Running workloads may be removed and requeued automatically, allowing fair-share to converge and disincentivizing squatting. They also let unhealthy hosts drain their workloads as they reach their minimum runtimes, so repair activities can be fully automated. This last point was more important than we realized when planning this work. It reduced repairs requiring a human-in-the-loop by 74%, which was a massive savings in on-call toil. We know that scheduling policy changes can have unintended consequences. The zero-sum nature of the problem means that giving time to one researcher means taking away from another. Users who lose this exchange tend to look for new workarounds. Before rolling out the budget-based system, we wanted a fast way to predict where those longer wait times might arise and to test configuration knobs like the length of the lookback window or the maximum value to allow for minimum runtime (we chose 8 hours). We built a small simulation environment that takes a set of workloads and their submission schedule as input and allows the scheduler to make preemption and GPU assignment decisions. With the knowledge of each workload’s requested number of GPUs and total runtime, the simulator could jump ahead to schedulable moments and provide analysis of queue wait times, preemption events, and distribution of GPU time across projects for many simulated days in a few seconds. We ran the simulator against both historical submission data and constructed scenarios we wanted to better understand. One hypothesis we wanted to test involved “debug workloads.” These jobs require a small number of GPUs and a minimum runtime of 15 minutes or less, which is enough for the user to see whether a job launches successfully or crashes early due to a bug or misconfiguration. We wanted to know whether these jobs would see a shorter queue wait time than larger training workloads, which often need many GPUs and hours of runtime to make meaningful progress. Intuitively, these smaller jobs should rise to the top of the queue, since a small job can fit more places than a large one. But the precise queue latency was important. A short wait of a minute or two would unlock a new development practice, but a ten-minute wait becomes infeasible. Our simulations required hand-built test case data, because our historical record did not contain a high enough volume of these debug-like workloads. Our results supported the hypothesis, showing p90 debug workload wait times fall from about 6 hours to only 5 minutes. Smaller-scale simulator visualization of the baseline (left) and the new “allocations” scheduler (right). Each row is one GPU; each bar is a job, colored by parent workload with one color hue per team; hatching marks time where a job is interruptible, and a red edge marks a preemption. In the baseline, long urgent jobs are never interrupted, and fewer preemptions occur on lower-priority-level work. The new scheduler has a larger mix of colors on each GPU, illustrating occupancy rotation across teams. With simulation results in hand, we began a cluster-by-cluster rollout at the end of July. The results we care about are whether the workloads we chose to fund received their time, whether the new system maintained full occupancy, and whether researchers could reason about the scheduler to make informed decisions. Since rollout, we have observed users and teams consistently receiving their allocated GPU time. We count the time owed to a team as its allocation capped hour by hour at its actual demand. Over the 30-day test period, teams were delivered 98% of the GPU hours they were owed, and 13 of 15 team allocations received 95% or more with the worst case receiving 90%. Occupancy on the cluster held steady at 98% before and after the change, with demand exceeding capacity by 2-3x in both periods. 18% of delivered GPU time was unallocated, which is how we maintained high occupancy during periods when funded use cases were not ready to run. Our simulator results proved to be directionally accurate with real outcomes overperforming our predictions. Debug workload p90 queue wait time fell from 2 hours to 30 seconds under the new scheduler, against a simulated prediction of 6 hours to 5 minutes from hand-crafted test scenarios. It’s worth noting that the smaller sample size of debug workloads in the baseline meant there was higher variance in those measurements. Queue latency in general improved as a side effect of time-slicing: on our largest H100 cluster, median queue wait time fell from 5 minutes to 24 seconds, and p90 wait time fell by about a third (from 2.8 hours to 1.8 hours). Compared against the three problems we set out to solve: - Squatting: Short debug workloads start in under a minute, reducing the value of squatting. The cost of this behavior charges the squatter’s budget, which prevents them from receiving time when they really need it. - Priority inflation: We still allow workloads to declare priority, but it only impacts sorting within a team. Managers are incentivized to monitor priority across the group to optimize the use of their budgets. - On-call toil: Unhealthy hosts drain automatically as workloads reach minimum runtime. Repairs requiring a human-in-the-loop fell by 74%. The learning curve was steeper than we had assumed. We rolled out the change incrementally, so in the early days researchers experienced varied behavior depending on which cluster they targeted. Furthermore, our interfaces retained some old terminology (like workload priority) whose meanings had changed. Documentation alone did not resolve the confusion. What did work was conducting live explanatory sessions, providing a forum for researchers to ask questions and for the engineering team to provide deeper descriptions of both how and why the scheduler made its prioritization decisions using real examples. This was a key moment because it marked a pivot from an early period of frustration and folk theories to the current mode, where research groups communicate more often and more broadly about the GPU needs of their experiments. Researchers are now engaging in the budgeting discussion with clearer knowledge of the tradeoffs being made to accommodate any new request. In addition to in-person sessions, we introduced new visualizations post-launch to give users a better sense of how closely their assigned GPU time is tracking their expected allocations and directly exposing the metric used for sorting the workload queue. This provided a simple place to look when a workload was preempted to understand why. These visualizations also helped budget owners, who could see how GPU time was being used across various projects under their management. Example visualization of allocation usage over time. Not every use case improved. In addition to distributed training, our researchers launch interactive sessions where they perform data analysis and test training code while they write it. In the old system, a researcher could hold such a session for up to a week. With time-slicing, they were subject to the 8-hour cap on protected runtime, after which a session becomes preemptible if it exceeds its allocation. We did not appreciate the extent to which researchers were dependent on the volatile state of these sessions. Being preempted meant waiting to secure a new session and also rebuilding their state by hand. After surveying researchers to understand the breadth of this problem, we created two new roadmap projects. We are investing in a CPU-only cluster next to our on-prem storage for dev sessions focused on data prep tasks. This will preserve our training cluster capacity for workloads that truly need it. Additionally, we plan to build restorable sessions for these CPU-only workloads. This will allow us to continue to preempt workloads at the end of their minimum runtime for maintenance or time-slicing, while also being able to restore the session elsewhere without the researcher rebuilding it. We can retain the operational and scheduling benefits of this new system while also improving the user experience. We remain on the lookout for emerging issues. One potential problem we’re investigating is capacity fragmentation, which may result in queue wait times increasing for the largest workloads. Our intuition is that minimum runtime protection is being applied to the types of jobs that used to rely on preemptible mechanisms to exceed their team concurrent GPU limit. Previously, those jobs could be interrupted at any time, potentially wasting that time, but also making it easier to schedule large jobs. Now the scheduler may have fewer opportunities to interrupt many jobs at once to place a large pending workload. We’re currently using our simulator tools to reproduce this problem while also measuring the ground truth in production. As we look beyond the scheduling work described here, we are aiming at the capstone of the pyramid: utilization. We need to ensure that bootstrapping, checkpointing, and the training applications themselves are all done as efficiently as possible, maximizing the value of the scheduled time each workload receives. If you’d like to tackle challenges like these in close collaboration with researchers, we encourage you to explore open engineering roles at Ai2.
15:43

Ai2 Rebuilt Its GPU Scheduler and Cut Wait Times by 87%

AlphaSignal Newsroom

A research lab rebuilt how it shares scarce training chips so people stop hogging them and wait less. Ai2 swapped free priority labels for GPU-hour budgets across programs, projects, and researchers, plus an eight-hour cap before a job can be kicked off, on NVIDIA H100, B200, and B300 clusters. Median queue wait on the largest H100 cluster fell from 5 minutes to 24 seconds, and the p90 wait fell from 2.8 hours to 1.8 hours. Teams received 98% of owed GPU hours at 98% occupancy under 2–3x oversubscription, human-coordinated repairs fell 74%, and debug p90 wait fell from 2 hours to 30 seconds. The eight-hour cap broke week-long interactive sessions that kept state in memory, and protected small jobs can delay 512-GPU placements.

Full text · 8,616 chars
- Ai2 replaced its priority-based GPU scheduler with time budgets, hierarchical fair-share, and a time-slicing contract. - Median queue wait on their largest H100 cluster fell from 5 minutes to 24 seconds. - Teams received 98% of owed GPU hours while cluster occupancy stayed at 98% under 2-3x oversubscription. - Jobs declare a minimum runtime (capped at 8 hours) during which they are preemption-protected. - Automated host draining via time-slicing cut human-in-the-loop repairs by 74%. - Full details in the Ai2 engineering blog. Ai2 cuts GPU queue waits with time budgets The Allen Institute for AI (Ai2) rebuilt the scheduler for its NVIDIA H100, B200, and B300 clusters after cost-free priority labels and long-lived protected jobs encouraged researchers to hoard capacity. The replacement combines GPU-hour budgets, hierarchical fair-share scheduling, and bounded protection from preemption. On Ai2’s largest H100 cluster, median queue wait fell from 5 minutes to 24 seconds. The p90 wait, the threshold below which 90% of waits fall, declined from 2.8 hours to 1.8 hours. The design gives infrastructure teams a concrete way to allocate scarce accelerators while preserving high cluster occupancy. Priority without a price Ai2 operates thousands of GPUs across clusters containing 88 to 1,024 devices each. About 150 internal researchers use them for large language and vision-language models, robotics reinforcement learning simulations, and agent post-training. Queued demand often reaches two to three times available capacity. The previous scheduler let workloads request protection from preemption, while team-level concurrency limits capped the number of protected jobs. Because high priority and protection carried no usage cost, the policy produced three operational failures: - GPU squatting: Researchers parked no-op workloads on accelerators so they could connect immediately, avoiding long waits for debugging sessions. - Priority inflation: Eventually, every scheduled workload used HIGH priority, starving jobs submitted at lower levels. - Maintenance gridlock: On-call engineers spent substantial time negotiating shutdowns of protected jobs on hosts that required repair. Ai2 describes the incentives as a tragedy of the commons. Its account cites a similar example from the 2011 Dominant Resource Fairness paper, in which users added infinite loops to inflate utilization figures and retain dedicated machines. The layer a scheduler can fix Ai2 diagnoses cluster performance with a four-layer model that separates scheduling policy from hardware health and workload efficiency, with this redesign concentrating on the impact layer: - Availability - How often the hardware is healthy and ready to accept work. - Occupancy - How much available GPU time is assigned to workloads. - Impact - How often the scheduler selects the work the organization values most. - Utilization - How much GPU capacity a running workload consumes over its lifetime. GPU-hours replace priority labels Fixed GPU allotments left hardware idle because research demand arrives in bursts. Ai2 now allocates shares of total GPU time through a hierarchy of programs, projects, and researchers. These budgets provide weighted entitlements while allowing teams to borrow unused capacity. For each allocation, the scheduler calculates a rolling seven-day ratio of actual occupancy to allocated GPU time. Jobs associated with underused allocations rank ahead of jobs whose teams have already consumed their share. A workload can continue beyond its requested minimum runtime while its allocation still outranks competing demand. Because every charged workload consumes its team’s budget, a no-op squatting job reduces the GPU time available for later training. The accounting system aligns each scheduling decision with the allocation that benefits from it. The algorithm follows a mature lineage that includes the 2009 Hadoop Fair Scheduler, YARN’s Fair Scheduler, and Slurm’s Fair Tree. Ai2 maps the hierarchy to its research organization and lets managers set weights as budgets instead of relying on static per-user quotas. Eight hours of protection, then preemption Large training jobs complicate fair-share scheduling because a week-long run can hold hundreds of GPUs after competing allocations become underserved. Ai2 addresses that problem with a workload contract: each job declares a minimum runtime and whether it can resume after eviction, with protected runtime capped at eight hours. - The user submits the resource request, minimum runtime, and resumable flag. - The scheduler ranks the job using its allocation’s actual-to-entitled occupancy ratio over the rolling lookback window. - Once placed, the job consumes GPU time from that allocation. - The scheduler protects it from preemption for the declared minimum runtime. - After protection expires, the job may continue while its allocation outranks competitors. A resumable job may be evicted and returned to the queue. - Completion releases the resources for another workload. A job submitted with minimum_runtime=0 enters an unallocated backfill class. It does not consume a team budget, can use otherwise idle GPUs, and may be preempted immediately when allocated work arrives. As protection windows expire, workloads can leave hosts marked for maintenance without engineers negotiating individual shutdowns. Ai2 reports a 74% reduction in repairs that require human coordination. The rollout kept 98% occupancy Before deployment, Ai2 built a simulator to test the seven-day lookback window, the maximum protected runtime, and other policy settings against historical traces and synthetic scenarios. A 30-day production rollout produced the following results: | Measure | Production result | |---|---| | Budget delivery | Teams received 98% of their allocated GPU-hours. | | Allocation consistency | 13 of 15 team allocations received at least 95% of their budgets; the lowest received 90%. | | Cluster occupancy | Held at 98% before and after the scheduler change. | | Unallocated backfill | Accounted for 18% of delivered GPU time. | | Debug workload p90 wait | Fell from 2 hours to 30 seconds. The simulator had modeled a reduction from 6 hours to 5 minutes. | Borrowing unused entitlements also lets bursty workloads temporarily exceed their assigned shares. One researcher described the effect as “an extra 30% compute” because capacity that previously sat idle became available between other teams’ runs. Shorter queues carry trade-offs Changed terminology created confusion during the transition because researchers still used words such as “priority” after their scheduling meaning had shifted. Written documentation proved insufficient, so the infrastructure team held live question-and-answer sessions built around actual scheduling decisions. The eight-hour protection cap also disrupted researchers who kept interactive development sessions alive for a week while storing volatile state in memory. Preemption forced them to reconstruct that context manually. Ai2 plans to provide a separate CPU development cluster and restorable sessions for this workflow. Minimum-runtime guarantees can also fragment capacity. Protected small jobs reduce the opportunities to reclaim enough GPUs simultaneously for a 512-GPU workload, so Ai2 continues to monitor placement delays for its largest runs. The operating model travels Ai2’s most reusable decision was to turn case-by-case operations disputes into GPU-hour allocations that leadership sets before jobs enter the queue. Managers distribute compute according to expected research value, while the scheduler measures delivery and enforces those decisions over time. The policy can sit above Slurm, Kubernetes, or a custom scheduler because its core inputs are an organizational hierarchy, weighted budgets, metered occupancy, and explicit preemption rules. Implementing it requires several supporting systems: - Reliable accounting: Track GPU time by workload, researcher, project, and team over a defined lookback window. - Budget ownership: Give designated managers authority to assign and revise shares. - Preemption support: Encourage checkpointing and define bounded protection for resumable workloads. - Backfill capacity: Provide an immediately preemptible class that can absorb unused GPUs. - Rollout tooling: Simulate historical traces, explain ranking decisions, and provide a separate path for long-lived interactive development. Ai2’s technical account covers further design trade-offs and planned work on GPU utilization. The organization also lists open infrastructure roles.
16:47

Artificial Analysis Runs 30 Edits to Expose How Image Models Drift

AlphaSignal Newsroom

Repeatedly editing the same photo shows some image tools quietly rewrite the whole picture instead of changing only what you asked. Artificial Analysis ran 30 consecutive edits on one real estate photo across Ideogram 4.5, GPT Image 2.5 Sunburst, Black Forest Labs FLUX 3, and Google Nano Banana 2.1. Ideogram 4.5 and FLUX 3 kept at least 95% of pixels on small local edits, Sunburst left only about 20% of the frame unchanged per edit, and Nano Banana 2.1 stayed local but the picture gradually darkened. Sunburst still leads the single-edit leaderboard at 1,197 Elo, while Ideogram 4.5 ranks number 23 for editing and number 34 for text-to-image. Use local copy-and-patch models for long revision chains and full re-renders for one-shot generation.

Notes
  • Method: 30 consecutive edits on the same real estate photo. Each model edited its own previous output.
  • Models: Ideogram 4.5, OpenAI GPT Image 2.5 Sunburst, Black Forest Labs FLUX 3, Google Nano Banana 2.1.
  • Prompts included lighting a fire, adding a sofa, repainting walls, placing flowers, and changing daylight to twilight.
  • Primary measurement: share of the frame that stayed effectively unchanged outside the requested modification.

| Model | Observed behavior | Long-chain risk |

|---|---|---|

| Ideogram 4.5 | Targeted regions. Adding tulips left at least 95% of the frame essentially unchanged. | Low pixel drift in this sequence. |

| FLUX 3 | Similarly local. At least 95% preserved on small changes. | Low pixel drift in this sequence. |

| GPT Image 2.5 Sunburst | Re-rendered most of the image each turn. Roughly one-fifth of the frame stayed unchanged. | Color, texture, and structure accumulated across turns. |

| Nano Banana 2.1 | Relatively local, with slight shifts in surrounding pixels. | The image gradually darkened. |

Leaderboard rank misses drift

Ideogram designed 4.5 to reduce pixel shifts, color changes, and texture artifacts that accumulate during repeated editing.

Reported single-task ranks: Ideogram 4.5 debuted at number 23 for image editing and number 34 for text-to-image. GPT Image 2.5 Sunburst Max scored 1,197 Elo. Grok Imagine Image 2.0 scored 1,155. Microsoft MAI-Image-2.6 scored 1,150.

Those rankings summarize isolated tasks. A 30-turn sequence measures whether furniture, typography, product details, colors, and composition survive after repeated revisions. A model can ace an isolated edit while changing previously approved work.

Match the model to the chain
  • Ideogram 4.5: product photography, poster revisions, staged interiors, protected brand or layout. Up to four reference images, optional mask, native 2K, crop-and-stitch high-res. Price: $0.008 to $0.22 per image across four quality modes at native 2K.
  • GPT Image 2.5 Sunburst: one-pass generation and shorter chains that need scene-wide relighting or restyling. Up to 4K, as many as 16 reference images, xhigh and max quality tiers. Available through the OpenAI API and partners including Picsart and Atlas Cloud.
  • FLUX 3: targeted changes where retaining surrounding pixels is the priority.
  • Nano Banana 2.1: local sessions, with monitoring for brightness and color drift.

Copying unchanged pixels can score high even when the requested edit fails. A successful twilight conversion may need to alter lighting across the entire frame. Score instruction compliance and protected-region stability separately.

Limits

One photo and one 30-step real estate sequence. Results do not cover every subject, prompt, resolution, or API configuration.

  • Build representative chains for the revisions users actually request.
  • Define protected regions: logos, faces, product geometry, text, approved layout.
  • Score edit success separately from pixel similarity.
  • Monitor brightness, color, texture, composition, and detail after every turn.
  • Budget by session: all revisions, retries, resolutions, and quality tiers. Multiply the per-image rate by expected turns and retries.
Full text · 6,470 chars
- Artificial Analysis stress-tested four frontier editors with 30 consecutive edits on one photo. - Ideogram 4.5 and FLUX 3 preserved 95%+ of pixels on small local edits. - GPT Image 2.5 Sunburst re-renders most of the frame, leaving only ~20% unchanged per edit. - Nano Banana 2.1 edits locally but the background gradually darkens across turns. - Sunburst still leads the single-edit leaderboard at 1,197 Elo; Ideogram 4.5 ranks #23. - Takeaway: pick copy-and-patch models for iterative workflows, full re-render for one-shot generation. Thirty consecutive edits reveal which image models drift Artificial Analysis stress-tested four image editors by applying 30 consecutive changes to the same real estate photo. The test exposed how quickly each model altered details outside the requested edit, a failure mode that single-edit leaderboards rarely capture. Each model received the same sequence of instructions and edited its own previous output. The prompts included lighting a fire, adding a sofa, repainting walls, placing flowers, and changing daylight to twilight. This setup mirrors an interactive workflow in which every new request inherits artifacts from earlier turns. Thirty turns, one room - Models tested: Ideogram 4.5, OpenAI GPT Image 2.5 Sunburst, Black Forest Labs FLUX 3, and Google Nano Banana 2.1. - Starting point: The same real estate photograph for each model. - Process: Every output became the input for the next edit. - Primary measurement: The share of the frame that remained effectively unchanged outside the requested modification. The models fell along a spectrum between local patching and broad re-rendering. Local editors retained source pixels around the requested region. Broader editors regenerated much of the frame, giving them more scope to adjust lighting and style while creating more opportunities for cumulative drift. | Model | Observed behavior | Long-chain risk | |---|---|---| | Ideogram 4.5 | Modified targeted regions while retaining most source pixels. On a small edit such as adding tulips, at least 95% of the frame remained essentially unchanged. | Low pixel drift in the tested sequence. | | FLUX 3 | Showed similarly local editing behavior, preserving at least 95% of the frame during small changes. | Low pixel drift in the tested sequence. | | GPT Image 2.5 Sunburst | Re-rendered most of the image during each turn. In the cited comparison, roughly one-fifth of the frame remained unchanged. | Color, texture, and structural changes accumulated across turns. | | Nano Banana 2.1 | Kept edits relatively local, although surrounding pixels shifted slightly. | The image gradually darkened during the sequence. | Leaderboard rank misses drift Ideogram designed version 4.5 to reduce the pixel shifts, color changes, and texture artifacts that accumulate during repeated editing. The test gives that claim a concrete measurement by comparing retained pixels after targeted changes. Its single-edit rankings tell a different part of the story. According to the reported ranking summary, Ideogram 4.5 debuted at number 23 for image editing and number 34 for text-to-image generation. GPT Image 2.5 Sunburst Max scored 1,197, Grok Imagine Image 2.0 scored 1,155, and Microsoft MAI-Image-2.6 scored 1,150. Those rankings summarize performance on individual tasks. A 30-turn sequence measures persistence: whether furniture, typography, product details, colors, and composition survive after repeated revisions. Developers building interactive editors need both measurements because a model can produce a strong isolated edit while changing previously approved work. Match the model to the edit chain | Model | Likely fit | Relevant capabilities | |---|---|---| | Ideogram 4.5 | Product photography, poster revisions, staged interiors, and workflows with protected brand or layout elements. | Up to four reference images, an optional mask, native 2K output, and crop-and-stitch high-resolution editing. | | GPT Image 2.5 Sunburst | One-pass generation and shorter edit chains that require scene-wide relighting or restyling. | Up to 4K output, as many as 16 reference images, and xhigh and max quality tiers. | | FLUX 3 | Targeted changes where retaining surrounding pixels is a priority. | Aggressive pixel preservation during the small edits in this test. | | Nano Banana 2.1 | Local editing sessions with monitoring for brightness and color drift. | Relatively contained changes, with gradual darkening observed over long sequences. | Local preservation has limits as a quality metric. Copying unchanged pixels can produce a high stability score even when the requested edit fails, while a successful twilight conversion may need to alter lighting across the entire frame. Evaluation should therefore score instruction compliance and protected-region stability separately. Test the whole session The benchmark used one photo and one 30-step real estate sequence, so its results do not establish performance across every subject, prompt, resolution, or API configuration. Teams should reproduce the test with their own assets and expected editing patterns before selecting a model. - Build representative chains: Test the number and type of revisions users commonly request. - Define protected regions: Track changes to logos, faces, product geometry, text, and approved layout elements. - Score edit success separately: Verify that each instruction was completed instead of relying only on pixel similarity. - Monitor cumulative drift: Measure brightness, color, texture, composition, and detail after every turn. - Budget by session: Include all revisions, retries, output resolutions, and quality tiers in cost and latency estimates. Ideogram 4.5 costs between $0.008 and $0.22 per image across four quality modes at native 2K resolution. Sunburst is available through the OpenAI API and partners including Picsart and Atlas Cloud. A useful cost comparison multiplies the per-image rate by the expected number of turns and retries rather than treating each generation as an isolated request. Model selection should follow the expected editing session. Broad re-rendering suits workflows that welcome scene-wide reinterpretation, while local preservation better serves long revision chains with approved elements that must remain fixed. Artificial Analysis found a substantial gap between those behaviors after 30 turns, giving developers a practical benchmark to reproduce against their own workloads.
00:08

Roundtables: A Conversation With the Creator of AI-Designed Viruses

A live interview will ask how a student used a generative model to sketch virus genomes, and whether that is a step toward designed life. MIT Technology Review's James O'Donnell talks to Stanford and Arc Institute PhD student Samuel King on Friday, October 16, 2026 at 18:30 BST / 1:30pm EDT / 10:30am PDT. In 2025 King used a generative model to propose genetic blueprints for microscopic viruses. The write-up says that is not AI-generated life yet, but that could be next. The session is for subscribers.

Full text · 1,246 chars
Available only for MIT Technology Review subscribers. Friday, October 16, 2026 Can AI design new life forms? In 2025, Stanford University PhD student Samuel King came up with a preliminary answer when he used a generative AI model to propose genetic blueprints for microscopic viruses. It isn’t yet an example of AI-generated life, but that could be next. Join senior AI reporter James O'Donnell as he interviews King about his work, being named one of MIT Technology Review's Innovators Under 35 and new ways of seeing biology. Going live on October 16th at 18:30 BST / 1:30pm EDT / 10:30am PDT Speakers: James O'Donnell, AI Reporter, and Samuel King, Bioengineering PhD Candidate, Stanford University/Arc Institute Related Stories Deep Dive Artificial intelligence Don’t be fooled—LLMs don’t reason Ten years after AlphaGo’s match against Go champion Lee Sedol, today’s AI still isn’t tapping into the machinery that made that win possible. AI’s recursive self-improvement might not come so quickly after all AI agents are not yet creative enough to carry out genuinely innovative open-ended AI research, it seems. Stay connected Get the latest updates from MIT Technology Review Discover special offers, top stories, upcoming events, and more.
00:34

ttok 1.0

Simon Willison

A tiny command-line tool now counts text the way the newest chatbots do. Simon Willison shipped ttok 1.0 after ttok 0.4 still defaulted to the GPT-4 tokenizer when he thinks it should default to GPT-5/GPT-6. OpenAI has not confirmed that GPT-6 uses the same tokenizer as the GPT-5 family. William Liu's experiment found all seven GPT models — 5.5, 5.6 Sol/Terra/Luna, and 6 Astra/Sol/Luna — report 44,794 tokens and match each other on every one of 31 fixtures. GPT-6 introduces no input-count change on that corpus.

Full text · 751 chars
9th October 2026 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.
03:44

Atlassian says software engineer's AI 'overuse' distracted colleagues | Lawyerly

A big software firm told a court that one engineer's heavy use of AI tools distracted the rest of the team. Atlassian said the overuse needed to stop. The captured snippet does not name the engineer, the case, or the outcome. This is a court filing claim, not a verdict.

Full text · 150 chars
Atlassian has told a court that a software engineer's “overuse” of artificial intelligence was a distraction to the rest of his team and needed to ...
04:00

Diffu-LoRA: A Novel Low-Rank Adaptation for Personalized Diffusion Models

Tianjing Li, Wei Zhu

A cheaper way to teach an image model a specific person or object learns which layers actually need the extra capacity. Diffu-LoRA inserts gated low-rank adapters into the linear layers of Transformer blocks and leaves the pretrained backbone frozen. Bilevel optimization updates the adapters and the gates on separate data splits, then progressive pruning drops the lowest-gate components to hit a rank budget. Tests use Stable Diffusion on DreamBooth subjects plus extra collected sets. Subject fidelity and prompt following beat the evaluated fine-tuning baselines, and ablations credit the bilevel step, the pruning, and where the adapters sit.

Full text · 2,183 chars
Computer Science > Computation and Language Title:Diffu-LoRA: A Novel Low-Rank Adaptation for Personalized Diffusion Models View PDF Abstract:Personalizing text-to-image diffusion models from a few reference images requires preserving subject identity while following prompts that describe new contexts. Full-model fine-tuning is parameter-intensive, whereas low-rank adaptation (LoRA) reduces the number of trainable parameters but leaves open how adaptation capacity should be distributed across layers. We introduce Diffu-LoRA, a parameter-efficient method that learns this allocation through gated low-rank adaptation. Diffu-LoRA inserts trainable low-rank components into the linear layers of Transformer blocks and assigns a learnable gate to each component. Bilevel optimization updates the adaptation weights and gate parameters on separate data splits, while progressive pruning removes components with the lowest gate values to meet a prescribed rank budget. This procedure allocates adaptation capacity nonuniformly across layers while keeping the pretrained backbone frozen. Experiments with Stable Diffusion on subjects from DreamBooth and additional collected datasets show improved overall subject fidelity and prompt alignment relative to the evaluated fine-tuning baselines. Ablation studies examine the contributions of bilevel optimization, progressive pruning, and adapter placement. These results support learned rank allocation as a practical approach to parameter-efficient diffusion model personalization. Bibliographic and Citation Tools Code, Data and Media Associated with this Article Demos Recommenders and Search Tools arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
04:00

Wieszcz-XIX: A 3.1-Billion-Word Corpus of Pre-1918 Polish and Temporally Bounded Language Models Trained From Scratch

Szymon Kocur

A huge stash of old Polish writing lets new models speak in period style and stumble on modern words. Wieszcz-XIX holds 6.75 billion tokens, about 3.1 billion words, in 294,369 documents from 1800 to 1918, over three orders of magnitude larger than the annotated corpus of the same period. Post-1918 leakage in the training set is cut to a 0.04 to 0.38 percent residue, the character error rate is 0.68 percent where text is legible, and 45 percent of sampled passages cannot be corrected. Decoder-only models from 47M to 349M trained from scratch pay about 3.1 bits per byte more on post-1918 vocabulary than on period vocabulary. Adding parameters gains about twice as much as a second pass, and the models reproduce period prejudice including antisemitic statements.

Full text · 2,755 chars
Computer Science > Computation and Language Title:Wieszcz-XIX: A 3.1-Billion-Word Corpus of Pre-1918 Polish and Temporally Bounded Language Models Trained From Scratch View PDF HTML (experimental) Abstract:Historical Polish is well documented as a language but annotated in machine-readable form only to about a million words for the period this paper covers; the rest sits behind optical character recognition of variable quality. We present Wieszcz-XIX, a corpus of 6.75 billion tokens (about 3.1 billion words) in 294,369 documents, most of them periodical issues, of Polish published from 1800 to 1918, assembled from Wolne Lektury and the Internet Archive by a pipeline that filters, deduplicates, audits for post-1918 leakage and splits at the document level. It is over three orders of magnitude larger than the annotated corpus of the same period, and we quantify its defects: recognition corruption against a false-positive floor, near-identical duplication, which is removed, and post-1918 leakage, which is excluded from the training corpus itself down to a known residue of 0.04 to 0.38% of its bytes, found in the transcribed source, so the published corpus is the trained one document for document. On a hand-corrected sample the character error rate is 0.68% where the text is legible, and 45% of the sampled passages cannot be corrected. On it we train a ladder of decoder-only models from 47M to 349M parameters from scratch, and measure their temporal boundedness. Against two modern Polish base models, one far larger, the 349M shows a crossover, as does the 107M against the comparator of its size: post-1918 vocabulary costs them about 3.1 bits per byte more than period vocabulary, a gap the comparators do not show, and period vocabulary costs them fewer bits than it costs the comparators. Shown period text, the models keep its spelling and the comparators only partly. Adding parameters gains about twice as much as a second pass over the data. We release the corpus, code and weights. Content warning: the models reproduce period prejudice, including antisemitic statements. Bibliographic and Citation Tools Code, Data and Media Associated with this Article Demos Recommenders and Search Tools arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
04:00

Large Language Model-Assisted Preparation of Transportation Management Plans: A Case Study with WisDOT WisTMP System

Zihao Sheng, Pei Li, Zilin Huang, Yen-Jung Chen, Yuhao Luo, Zhengyang Wan, Steven T. Parker, David A. Noyce, Sikai Chen

Highway work-zone plans can be drafted by a local chatbot, but it still pads the strategy list and botches the cost. The authors fine-tune several open-source models at different sizes and run them on-prem against Wisconsin DOT's WisTMP system. Training data comes from historical WisTMP PDFs turned into structured question-answer pairs. Fine-tuning lifts standard text-generation scores, yet the models over-generate strategies and struggle with project-specific justifications and accurate cost estimates. Scaling from 7B/8B to 14B yields only limited gains.

Full text · 2,244 chars
Computer Science > Computation and Language Title:Large Language Model-Assisted Preparation of Transportation Management Plans: A Case Study with WisDOT WisTMP System View PDF Abstract:Work zones are critical yet hazardous components of transportation infrastructure, requiring carefully designed Transportation Management Plans (TMPs) to ensure safety and mobility. However, TMP preparation remains labor-intensive and heavily dependent on practitioner expertise. This paper proposes a Large Language Model (LLM)-assisted framework to automate TMP content generation, leveraging the WisDOT WisTMP system as the application context. The framework fine-tunes multiple open-source LLMs across different model scales and deploys them locally to ensure data security. To support model training, we construct a domain-specific dataset from historical WisTMP documents by converting PDF files into structured question-answer pairs in JSON format. Experimental results show that fine-tuning significantly improves performance across standard text generation metrics. Further section-wise and strategy-level analyses reveal that, while LLMs achieve strong overall performance, they tend to over-generate strategies and struggle to produce project-specific justifications and accurate cost estimates. In addition, scaling from 7B/8B to 14B yields limited gains. These findings demonstrate the potential of LLMs to improve TMP preparation efficiency while highlighting remaining challenges in LLM-assisted TMP development. The source code and demo videos will be publicly available at this https URL. Bibliographic and Citation Tools Code, Data and Media Associated with this Article Demos Recommenders and Search Tools arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
04:00

Cognitive Thermometers: Machine Learning and Logical Complexity

Shane Steinert-Threlkeld, Jakub Szymanik

How hard a meaning is to learn may explain which words languages keep better than formal logic does. Earlier accounts of semantic categories lean on logical definability and complexity, but those scores swing with the logical language you pick. The authors treat machine-learning difficulty as a more neutral complexity measure and review cases where logic and learning agree on which meanings are hard. Where the two disagree, learning better matches patterns in semantic typology. They frame learning models as cognitive thermometers that can sit between symbolic logic and connectionist AI.

Full text · 1,647 chars
Computer Science > Computation and Language Title:Cognitive Thermometers: Machine Learning and Logical Complexity View PDF HTML (experimental) Abstract:How does the human mind represent semantic categories? Why do natural languages favor certain meanings over others? Prior explanations have relied on logical definability and complexity, but these are highly sensitive to the choice of logical language, rendering some design choices unmotivated. In this article, we propose that machine learning provides a somewhat more agnostic approach to measuring semantic complexity. We review emerging evidence that logic and machine learning often yield converging results on relative complexity and its resulting effects in semantic typology. Where they diverge, learning appears to be a better explanation than logical complexity. We argue that treating machine learning models as ``cognitive thermometers'' enables a unified approach to complexity that bridges symbolic logic and connectionist AI. Bibliographic and Citation Tools Code, Data and Media Associated with this Article Demos Recommenders and Search Tools arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
04:00

Lossy Compressive Text Autoencoders

Vinko Sabol\v{c}ec, Angelos Katharopoulos, David Grangier

Text can be packed into a much smaller form and still be rebuilt well enough for search and question answering. The autoencoder downscales and upscales hidden states along the time axis and ends in a residual low-dimension discrete bottleneck. The authors sweep quantization methods, training objectives, and datasets. Reconstruction is scored with BLEU and an LLM judge, then the same latents are tested on question answering and semantic text similarity. On web text the compressed form lands on par with lossless compressors at 2.24 bits per byte while keeping useful reconstruction and downstream scores.

Full text · 1,711 chars
Computer Science > Computation and Language Title:Lossy Compressive Text Autoencoders View PDF HTML (experimental) Abstract:Our work explores learning a compressed latent representation of text, at the intersection of data compression and representation learning. We propose an autoencoder architecture that performs residual downscaling and upscaling of hidden representations along the time axis, with a residual low-dimension discrete bottleneck. We analyze our approach for different quantization methods, training objectives, and datasets. For different levels of compression, we evaluate the similarity between the original and reconstructed text both at the surface-level (BLEU) and at the semantic-level (LLM-based judge). Additionally, we evaluate our models on downstream question-answering and semantic text similarity benchmarks. Our approach results in compressed representations which are on par with lossless text compression algorithms at 2.24 bits per byte on web text data, while having good reconstruction and downstream task performance. Bibliographic and Citation Tools Code, Data and Media Associated with this Article Demos Recommenders and Search Tools arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
04:00

Clarify, Then Focus: Statement Normalization for Conversation Analytics at Scale

Mikhail L. Arbuzov (Independent researcher), Karan Dave (Independent researcher), Evgeniya Dontsova (Independent researcher), Yaodong Hu (Independent researcher), Vincent Lao (Independent researcher), Navita Jain (Independent researcher), Sisong Bei (Independent researcher), Dmitry Dimov (Independent researcher)

Call-center analytics get cheaper if you first rewrite each chat into short, labeled claims. Statement normalization turns dialogue into speaker-attributed statements with source references and semantic tags so the same transcripts can answer many questions. On an offer-suppression task in customer-service calls, normalization helps a supervised classifier even without selecting a subset. Weaker prompted readers need both the rewrite and the selection step. A small model can learn the rewrite contract, and the rest of the pipeline can stay on lightweight encoders, which matters when you are scoring millions of conversations.

Full text · 2,024 chars
Computer Science > Computation and Language Title:Clarify, Then Focus: Statement Normalization for Conversation Analytics at Scale View PDF HTML (experimental) Abstract:Enterprise conversation analytics asks many questions of millions of interactions. Each question can require reconstructing what people mean and identifying which information matters, repeating costly interpretive work across the same transcripts. We propose a simple principle: clarify the text, then focus the reader. Statement normalization transforms dialogue into short, speaker-attributed statements with source references and semantic tags. The statements make meaning more explicit; the tags support selecting evidence for a particular question. Downstream models can use the full representation or a relevant subset, depending on what helps them make the decision. In an offer-suppression task on customer-service calls, normalization improves a supervised classifier without selection, while weaker prompted readers benefit from both normalization and selection. A small model can learn the normalization contract, while lightweight encoders handle tagging and downstream decisions. Sharing this preparation across questions supports an inference pipeline built entirely from small models, making analytics over millions of conversations substantially less expensive. Current browse context: Bibliographic and Citation Tools Code, Data and Media Associated with this Article Demos Recommenders and Search Tools arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
04:00

Disentangling Linguistic and Paralinguistic Information with Routed Sparse Autoencoders

Beimnet Bekele Guta, Xiaoyu Yang, Guangzhi Sun, Philip C. Woodland

A speech model can be split so the words live on one path and the speaker's voice, mood, and melody live on another. The method pairs a TopK sparse autoencoder with route-specific supervision and cross-factor adversaries on frozen SPEAR and WavLM encoders. Independent probes keep language stronger on the linguistic route, while speaker identity, emotion, and prosody stay on the paralinguistic route and drop sharply on the linguistic one. Routes learned on LibriSpeech still hold on MSP-Podcast without retraining the representation. Swapping a route in feature space transfers that factor and largely leaves the other route's information intact.

Full text · 1,819 chars
Computer Science > Computation and Language Title:Disentangling Linguistic and Paralinguistic Information with Routed Sparse Autoencoders View PDF HTML (experimental) Abstract:Self-supervised speech encoders contain linguistic and paralinguistic information in a shared, entangled representation space. We combine a TopK sparse autoencoder with route-specific supervision and cross-factor adversaries. Across frozen SPEAR and WavLM encoders, independent probes show factor-specific retention and suppression: linguistic information remains stronger in the linguistic route, while paralinguistic factors, including speaker identity, emotion, and prosody, are retained in the paralinguistic route and substantially reduced in the linguistic route. The route organisation learned on LibriSpeech persists on MSP-Podcast without representation-side retraining. Feature-space route interventions further transfer the swapped factor while largely preserving the information carried by the unchanged route. These results show consistent route-selective separation across encoders, corpora, independent probes, and representation-level interventions. Current browse context: Bibliographic and Citation Tools Code, Data and Media Associated with this Article Demos Recommenders and Search Tools arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
04:00

Sparse Attention Is Matrix Approximation, Not Choosing from a Bag of Values

Fang Wan, Xufeng Liu, Fan Li, Yi Liu

Long-prompt speedups work better when you keep the pieces that actually change the answer, not just the biggest scores. The authors say sparse attention should approximate the matrix that multiplies value vectors, not pick large entries from a bag. Matrix Approximation Sparse Attention (MASA) replaces attention-mass ranking with a closed-form score for how much each sparse unit cuts matrix-product error. It plugs into existing sparse-attention stacks without changing their kernels or budgets. Accuracy rises across several sparse methods, benchmarks, and LLM backbones.

Full text · 2,237 chars
Computer Science > Computation and Language Title:Sparse Attention Is Matrix Approximation, Not Choosing from a Bag of Values View PDF HTML (experimental) Abstract:Large Language Models (LLMs) achieve strong performance across many domains, but their efficiency is limited by the quadratic cost of attention with respect to prompt length. Sparse attention reduces this cost by retaining only a small fraction of query-key interactions to approximate the full attention matrix. However, existing methods are trapped in a mathematically wrong view: they simply keep large scalar entries or high-mass regions of the attention matrix. This treats the attention matrix as a bag of values, ignoring that it is used as a structured matrix whose entries jointly determine the attention output through multiplication with value vectors. We argue that this is the core conceptual issue: sparse attention should be formulated as matrix approximation, not as blindly choosing the largest values from a bag of entries. Based on this view, we propose Matrix Approximation Sparse Attention (MASA). MASA replaces raw attention-mass ranking with a closed-form score that measures how much each sparse unit reduces matrix-product approximation error. As a theory-grounded plug-in correction, MASA can be added to existing sparse attention frameworks without changing their sparse kernels or budgets. Extensive experiments across multiple sparse attention methods, benchmarks, and LLM backbones show consistent accuracy gains, supporting both MASA and the matrix-approximation view of sparse attention. Bibliographic and Citation Tools Code, Data and Media Associated with this Article Demos Recommenders and Search Tools arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
04:00

Stochastic Teacher Intervention for Agentic On-Policy Distillation

Junnan Liu, Linhao Luo, Zhijun Chen, Qianren Mao, Thuy-Trang Vu, Gholamreza Haffari

Teaching a small agent by copying a stronger one fails when the student wanders off, unless the teacher sometimes grabs the wheel. STI-OPD is a stochastic teacher-intervention setup for multi-turn on-policy distillation. It turns teacher-student KL divergence into an intervention probability, then samples whether to replace the student's action so supervision stays reliable without a fixed schedule. Mixed teacher-student traces are trained with an Importance-Weighted Reverse KL objective that corrects the sampling mismatch. The method beats the strongest prior on-policy distillation baseline on every evaluated benchmark and student size, and both the intervention rule and the importance weights matter.

Full text · 2,632 chars
Computer Science > Computation and Language Title:Stochastic Teacher Intervention for Agentic On-Policy Distillation View PDF HTML (experimental) Abstract:On-policy distillation (OPD) efficiently transfers capabilities from a stronger teacher to a student language model through dense token-level supervision on student-generated rollouts and has shown promise on complex tasks such as mathematical reasoning. However, in multi-turn agentic tasks, student decisions shape subsequent observations, causing early errors to accumulate across turns. The resulting trajectories can drift away from the teacher's rollout distribution, making the teacher's token-level supervision less reliable or even counterproductive for OPD training. To address this issue, we introduce STI-OPD, a stochastic teacher intervention framework for multi-turn agentic OPD. During multi-turn interaction, STI-OPD uses teacher intervention guided by teacher-student policy discrepancy to replace the student's proposed action with a teacher-generated one to maximize the acquisition of reliable supervision. We further develop a stochastic intervention strategy, addressing the limitations of previous threshold-based or fixed-schedule approaches, that estimates policy discrepancy using KL divergence and maps it to an intervention probability. By sampling whether to intervene from this probability, STI-OPD adaptively balances teacher control with student exploration. To learn from the resulting mixed-policy trajectories, we introduce an Importance-Weighted Reverse KL objective that corrects the token sampling mismatch between teacher-generated responses and the student policy to preserve the original OPD objective. Across tool-integrated reasoning and long-horizon interaction, STI-OPD outperforms the strongest prior OPD baseline on every evaluated benchmark and student size. Ablations further show that both discrepancy-guided intervention and importance weighting contribute to these gains. Bibliographic and Citation Tools Code, Data and Media Associated with this Article Demos Recommenders and Search Tools arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
04:00

Large Language Models for Machine Translation Quality Annotation: Humans and Models Are Both Challenged

Hala Almaghout, Christian Federmann, Qin Gao

Chatbots that grade translations sometimes match people, but they still miss small mistakes and the wrong dialect. The study compares model and human agreement on Multidimensional Quality Metrics and Error Span Annotation. The test set covers 70 language pairs plus public WMT23 and WMT25 data, across scores and error spans, many pairs, and many domains. Model-human agreement beats human-human agreement on some tasks, but both swing hard by scheme, language pair, and domain and stay unreliable in most settings. People struggle most with fine-grained MQM labels and low-resource pairs, while models struggle with minor errors, wrong language variants, and error-span marking.

Full text · 2,419 chars
Computer Science > Computation and Language Title:Large Language Models for Machine Translation Quality Annotation: Humans and Models Are Both Challenged View PDF HTML (experimental) Abstract:Large Language Models (LLMs) are considered to be a more efficient and cost-effective alternative to human judgment for Machine Translation (MT) evaluation. With MT evaluation spanning a large number of language pairs, domains and levels of annotation granularity, LLMs must be thoroughly evaluated across these dimensions before being reliably used as alternatives to human evaluation. In this paper, we evaluate the performance of LLMs for two prominent MT quality evaluation schemes: Multidimensional Quality Metrics (MQM) and Error Span Annotation (ESA) by comparing their agreement with human annotators. We present results on a long-context test set of 70 language pairs and the publicly available WMT23 and WMT25 data, investigating both score and error span annotation agreement across a variety of language pairs and domains. Our results show that while LLM agreement with human annotators exceeds agreement between human annotators for some evaluation tasks, both vary substantially across annotation schemes, language pairs and domains and remain unreliable for most settings. Furthermore, we identify challenges facing both human and LLM annotators: humans are particularly challenged by fine-grained MQM annotations and low-resource language pairs, while LLMs struggle with minor errors, wrong language variants and error span annotation. Our results highlight the potential for improvement for both human and LLM annotation performance, possibly through human-LLM collaborative annotation pipelines that address the reliability issues identified in this work. Bibliographic and Citation Tools Code, Data and Media Associated with this Article Demos Recommenders and Search Tools arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
04:00

AI4Fire: Evaluating Large Language Models on Wildfire Tasks

Yue Zhao, Xiyang Hu, Zuobin Xiong, Zhangyu Wang, Ruolin Li

Chatbots still guess badly on fire work unless you hand them the missing clue, and overconfident scores can cost lives. AI4Fire runs six core models bare and grounded on five wildfire tasks, zero-shot, plus a sweep of 29 more, after a search of 138 fire-task papers found no matching paired setup. A read-only SQL tool lifted every core model's database accuracy from at most 16 to at least 88 percent. No core model beat repeating today's staffing count, and two open-weight models mostly copied the median of similar earlier fire-days. Public releases are messy: a fire-danger column splits the holdout perfectly, and 67 aerial frames carry smoldering or fire-free labels from a clipped thermal maximum.

Full text · 2,055 chars
Computer Science > Computation and Language Title:AI4Fire: Evaluating Large Language Models on Wildfire Tasks View PDF HTML (experimental) Abstract:Large language models (LLMs) are entering wildfire management, where overstated evaluations can cost property and lives. How do they perform on wildfire tasks, with and without grounding? Bare means a model receives the task input alone. Grounded means it also receives one task-specific addition: for smoke detection, a smoke-free reference frame from the same camera. AI4Fire runs six core models bare and grounded on five wildfire tasks, zero-shot; a sweep adds 29 more. Our literature search on fire tasks found 138 works; none combines this roster, task coverage, and paired bare and grounded runs. We report three findings. (1) Grounding helped most where the addition carried the answer: a read-only SQL tool lifted every core model's database accuracy from at most 16 to at least 88 percent. (2) Simple rules were hard to beat: no core model outperformed repeating today's staffing count, and two open-weight models mostly copied the median of similar earlier fire-days, a worse forecast. (3) Public releases carry hazards: a fire-danger column separates the holdout perfectly, and 67 aerial fire frames carry smoldering or fire-free labels read from a clipped thermal maximum. We release prompts, responses, scores, code, and the survey record. Bibliographic and Citation Tools Code, Data and Media Associated with this Article Demos Recommenders and Search Tools arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
04:00

Back in Style: A Sociolinguistic Approach to Authoring and Measuring Persona Fidelity in User Simulation

Lex Konnelly, Elena Khasanova, Riqiang Wang, Matthias Lee, Harsh Saini, Parsa Kavehzadeh

Fake customers used to test support bots sound more real when you specify how they talk, not who they are. This pilot writes personas as concrete stylistic rates and scores them with authorship-verification stylometry and lexicon-based content analysis instead of costly LLM judges. The sociolinguistic schema is A/B-tested against a flat descriptive baseline across five task-oriented customer-service agents. It improves stylistic adherence and stylometric distinguishability for most tested models. Style fidelity is not the same as sounding natural, so the metrics are most useful for locating where a simulated user breaks down.

Full text · 2,385 chars
Computer Science > Computation and Language Title:Back in Style: A Sociolinguistic Approach to Authoring and Measuring Persona Fidelity in User Simulation View PDF HTML (experimental) Abstract:As agentic systems gain commercial popularity, user simulators increasingly serve as measurement instrument for their evaluation. However, the fidelity of simulated users in comparison to real human users is generally low, and typically assessed by costly, subjective LLM judges. In this pilot study, we ask whether fidelity can instead be measured deterministically by treating a user persona sociolinguistically: as a social type that emerges from observable linguistic style, rather than one predicted by labels or descriptions a model must extrapolate into behaviour. We author personas as concrete stylistic rates, which lets us transfer two established, model-free instruments -- authorship-verification stylometry and lexicon-based content analysis -- as fidelity diagnostics. We A/B-test the sociolinguistic schema against a flat descriptive baseline across five task-oriented customer-service agents. Results show that the sociolinguistic schema improves both stylistic adherence and stylometric distinguishability for most of the tested models, with a caveat that persona style fidelity does not necessarily equal persona "naturalness". We argue that a sociolinguistic approach to persona design is a promising path towards more diverse and representative user personas, and that these metrics are most valuable in an error-attribution analysis, localizing where fidelity breaks down. This is a first step towards interventions that move user simulations closer to faithful renderings of diverse and variable linguistic outputs. Bibliographic and Citation Tools Code, Data and Media Associated with this Article Demos Recommenders and Search Tools arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
04:33

Vietnamese firms struggle to hire AI engineers despite salaries 50% above other tech roles

Companies in Vietnam are paying a big premium for AI engineers and still cannot fill the seats. Salaries for those roles can sit 50 percent above other tech jobs. Firms say they cannot find people with real-world experience. The captured body is the lede only.

Full text · 150 chars
Vietnamese companies are finding it increasingly difficult to recruit AI engineers with real-world experience, even as salaries for such roles can ...
06:13

So long, Spokes: GitHub rewrites storage to restore reliability, just in time for agentic hordes

GitHub is ripping out an old storage path because coding agents are about to hit it much harder. Cursor engineers rejected Spokes' three-phase commit and instead upload pushes into an object-storage write-ahead log. The Register piece frames the rewrite as arriving just in time for agentic hordes. The captured body is that one design choice.

Full text · 151 chars
Cursor engineers also rejected Spokes' three-phase commit in favor of uploading pushes into an object storage write-ahead log (WAL), which captured ...
06:57

IQuest-Q1 Draws Early Developer Attention Across Coding and Agentic Workflows

A new coding model is being tuned for long jobs, not just autocomplete. IQuest-Q1 post-training focuses on software engineering, long-horizon agentic tasks, and general reasoning. The recipe named in the snippet is supervised fine-tuning plus reinforcement learning. There are no scores in the captured body.

Full text · 149 chars
Post-training focuses on software engineering , long-horizon agentic tasks, and general reasoning, using supervised fine-tuning and reinforcement ...
07:13

Trump Warns: Call AI "Super Intelligence " Or Face "Enemy" Status

The same naming fight now comes with a threat: use Super Intelligence, or SI, or be treated as the enemy. NDTV says President Trump warned that people who keep saying artificial intelligence will face that status. Anadolu's paired item ties it to an executive order for federal agencies. The captured body is the warning, not the legal text.

Full text · 142 chars
US President Donald Trump has warned that all those who do not refer to Artificial Intelligence (AI) as "Super Intelligence" or SI will be ...
07:14

Margaret Hamilton, who coined the phrase 'software engineer ' and designed steering ...

The person who named the job "software engineer" is the subject of a First Coast News obituary. Margaret Hamilton coined the phrase to distinguish the work, and the URL path includes "dies" plus Apollo 11. The captured body is mostly a local weather-and-football lede with one Hamilton sentence. Treat the death as reported by that URL, not as extra biography.

Full text · 157 chars
Tropical weather prompts high school football schedule changes in Northeast Florida ... Hamilton also coined the term "software engineer " to distinguish ...
07:50

Klarent Raises €7.13M for Agentic Software Testing

A testing startup raised a seed round to send software-checking agents into the U.S. Klarent raised €7.13 million. The money is for U.S. market entry, engineering hires, and cross-platform web and mobile testing. The company is also moving into financial services.

Full text · 144 chars
Funds go to US market entry, engineering hires and cross-platform web and mobile testing. The company is also moving into financial services ...
07:51

Trump calls users of term ' artificial intelligence ' the 'enemy' - Anadolu Ajansı

The White House wants every federal office to stop saying artificial intelligence and start saying super intelligence. Anadolu says the president earlier issued an executive order directing agencies to use that phrase instead. A second write-up the same day frames holdouts as enemies. The captured body is a one-line recap, not the order text.

Full text · 148 chars
US president earlier issued executive order directing federal agencies to use 'super intelligence' instead of ' artificial intelligence ' | Anadolu.
08:10

Helping Workers Make AI Work For Them | Premier

A state government is training workers for jobs that did not have names last year. Victoria programs aim to prepare people for AI adoption, prompt engineering, and agent management. Businesses in key sectors can enroll. The captured body is the enrollment pitch, not the curriculum.

Full text · 149 chars
They aim to prepare Victorians for emerging roles like AI adoption, prompt engineering and agent management. Businesses in key sectors can enroll ...
08:13

Modi wants India to power the AI age. Residents are pushing back | Reuters

A national AI buildout is running into the same street-level fight as data centers elsewhere. Reuters writes from Mumbai about engineer Tushar Shetty and protests that mirror a global backlash against AI-era data centres. The captured body is the scene-setter. It does not name the sites or the demanded halt.

Full text · 144 chars
Protests mirror global backlash against AI -era data centres. MUMBAI, Oct 9 (Reuters) - The view from Indian engineer Tushar Shetty's Mumbai ...
09:00

Job titles of the future: Delivery drone air traffic controller

Eamon Whalen

Someone now sits at a desk and watches delivery drones the way a controller watches planes. Trevor Wischnewsky is a remote pilot in command at FlyTrex, watching hexacopter drones that deliver within a three-mile radius around Dallas. He coordinates traffic with competitors including Zipline, Amazon, and Alphabet's Wing, using an unmanned-aircraft traffic-management network that shares coordinates beyond line of sight. The job needs an FAA drone-pilot credential plus a weeks-long FAA-approved course; at his busiest he watches up to 10 drones at once. He once had to bring a drone back seconds after launch when wind picked up, keeping an ice cream cake off the sidewalk.

Full text · 3,433 chars
The moment Trevor Wischnewsky heard that drones were delivering pizza and sushi around his Texas neighborhood, his mind was made up. “It was super fascinating,” he says. “I just immediately wanted to be a part of it.” Wischnewsky is now an RPIC, or “remote pilot in command,” at a company called FlyTrex. An RPIC is like an air traffic controller. But instead of airplanes, this new class of aviators tend to fleets of hexacopter drones that ferry cargo to homes and businesses within a three-mile radius. Not only does he ensure the safe passage of FlyTrex’s parcels, but Wischnewsky coordinates traffic with a growing swarm of delivery drones dotting the skies over Dallas from competitors like Zipline, Amazon, and Alphabet’s Wing. Here’s what it takes to land the gig. Aviation experience When he applied to FlyTrex, Wischnewsky was in training to become a commercial pilot and flight instructor, and he was already certified by the Federal Aviation Administration as a drone pilot. The company requires this credential, along with a weeks-long FAA-approved training course, to ensure that a pilot is prepared to take over a drone if need be. Though he doesn’t pilot the craft, Wischnewsky monitors the delivery process from FlyTrex’s operations center with the help of inputs from GPS, lasers, and a range of sensors. A collaborative mindset At first, the biggest challenge was trusting the technology. “On the manned-aircraft side, they tell you ‘Don’t trust the automation,’” Wischnewsky says. The Dallas–Fort Worth area, though, is a testing ground for unmanned aircraft systems traffic management (UTM), a network that allows autonomous drones from FlyTrex and its competitors to share their coordinates in real time to help avoid collisions even beyond the pilot’s line of sight. Once Wischnewsky makes sure a drone is flight-ready, its path is unimpeded, and the weather conditions are clear, UTM takes care of the rest. Ability to stay calm under pressure At his busiest, Wischnewsky will have up to 10 drones on his watch at the same time. “You’re there as a safety net,” he says. Even with automation, keen situational awareness is required, and the most common hazard is the weather. Wischnewsky recalls a particularly stressful day: A drone had been in the air for only seconds when the wind picked up so suddenly that he had to intervene. He returned it safely, preventing an errant ice cream cake from hitting the sidewalk. Eamon Whalen is a reporter and researcher based in San Francisco. Deep Dive Culture God told them to sell crypto. Their investors lost everything. A pastor and his wife created a cryptocurrency and hawked it in Christian communities. When it came crashing down, investors lost millions―and they were accused of fraud. Smart glasses are already causing havoc in India And the chances of a crackdown are slim, as the authorities spy opportunities for surveillance. Child-monitoring apps might need a reboot Monitoring apps promise to keep young people safer online, but looking in on kids’ phones can backfire. Online safety experts say there’s a better way. How we picked 35 of the world’s top young scientists and engineers Our 2026 Innovators Under 35 list will be out soon. Here’s what we looked for as we sifted through 550 nominations from around the world. Stay connected Get the latest updates from MIT Technology Review Discover special offers, top stories, upcoming events, and more.
09:26

Shield AI and U.S. government investment in X-BAT reaches $400 million

A drone company just added another large Navy check to a growing pile of government money. Shield AI said on October 8, 2026 that the U.S. Navy executed a $150 million modification to an existing Other Transaction agreement. The announcement title puts combined U.S. government investment in X-BAT at $400 million. The captured body stops after the Navy modification line.

Full text · 154 chars
WASHINGTON (October 8, 2026) — Shield AI today announced that the U.S. Navy has executed a $150 million modification to its existing Other Transaction ...
09:27

Infatoshi Squeezes GLM-5.3's 753B Parameters Into 273 GiB for Multi-GPU Workstations

AlphaSignal Newsroom

Someone packed a huge language model small enough to fit on a high-end multi-GPU workstation. Infatoshi released a 3.04 bits-per-weight EXL3 compression of dealignai’s weight-edited GLM-5.3-UNCENSORED-FP8, shrinking zai-org’s 753-billion-parameter model to 273 GiB without a fine-tune. Only 8 of 256 routed experts run per token, plus one shared expert, and the compressor kept attention and the shared expert at 5 bits, dense layers at 4, and routed experts at 3. Compared with the FP8 source, KL divergence is 0.089 and wikitext-2 perplexity drifts from 3.302 to 3.440. It runs on four RTX PRO 6000s (384 GB VRAM) at about 55–60 tokens per second without speculative decoding, reports 0% refusals on HarmBench-320 at max effort, and has a practical 131K context ceiling on TP8 H200.

Full text · 2,458 chars
- Infatoshi released a 3.0bpw EXL3 quant of GLM-5.3-UNCENSORED, totaling 273 GiB. - Base is dealignai's weight-edited GLM-5.3, not a fine-tune, baked into residual-writer tensors. - Architecture: 753B parameters, 256 routed experts (8 active), MLA attention with DSA sparse indexer, plus MTP draft layer. - KL divergence vs FP8 source is 0.089; perplexity drifts from 3.302 to 3.440 on wikitext-2. - Runs on 4x RTX PRO 6000s (384 GB VRAM) at roughly 55-60 tok/s without speculative decoding. - Reports 0% refusals on HarmBench-320 at max effort, with practical 131K context ceiling on TP8 H200. GLM-5.3’s 753B weights shrink to 273 GiB Infatoshi has published an EXL3 release of GLM-5.3 that averages 3.04 bits per weight and occupies 273 GiB. The compression makes the 753-billion-parameter Mixture-of-Experts model practical on a high-end, multi-GPU workstation, although it remains far beyond a single consumer GPU. The artifact quantizes dealignai’s GLM-5.3-UNCENSORED-FP8, a weight-edited version of zai-org’s original GLM-5.3. Developers evaluating it should account for both changes: the parent modifies refusal behavior, while the EXL3 conversion reduces numerical precision. A mixed-precision squeeze | Core model and release details | | |---|---| | Item | Detail | |---|---| | Architecture | GlmMoeDsaForCausalLM | | Total parameters | 753 billion | | Routed experts | 256, with 8 active per token | | Shared experts | 1 | | Layers | 78 transformer layers and 1 MTP layer | | Attention | MLA with a DSA sparse indexer | | Quantization | EXL3, averaging 3.04 bits per weight | | Artifact size | 273 GiB | A Mixture-of-Experts model stores many specialized feed-forward networks but activates only a subset for each token. GLM-5.3 selects 8 of its 256 routed experts per token and also uses a shared expert, reducing active computation even though every expert must remain available in memory. The quantizer assigns more bits to components considered sensitive to compression and fewer bits to the routed experts that account for much of the model’s size. | Precision by component | | |---|---| | Component | Precision | |---|---| | Attention layers | 5 bpw | | Shared experts | 5 bpw | | Dense MLPs | 4 bpw | | Routed experts | 3 bpw | | Language-model head | | This story is for Pro members You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
10:36

What if A.I. Is Just a 'Normal Technology'?

Being smart is not the same as being powerful, and that is the case against treating chatbots as destiny. Computer scientist Arvind Narayanan tells Ezra Klein that intelligence does not automatically confer power. The captured body is that one distinction. The rest of the Times conversation is not in the stub.

Full text · 126 chars
The computer scientist Arvind Narayanan explains that just because A.I. is intelligent doesn't necessarily mean it's powerful.
11:12

Will AI scoop your science? Some researchers see a gloomy future

Some scientists now worry a chatbot will beat them to the next discovery. Nature's captured line is that researchers see a gloomy future as large language models get more capable. The snippet has no study size, field, or named lab. That is the whole retrieved body.

Full text · 95 chars
Scientists worry that increasingly capable large language models will beat them to discoveries.
12:54

A new feature for my blog, built using my voice

Simon Willison

A blogger built a new archive of his newsletters by talking to a coding helper while he cooked dinner. Simon Willison used the ChatGPT desktop Codex tab in voice mode with GPT-6 Astra High against a local checkout of simonwillisonblog. About half an hour of kitchen talk produced a Django model and migration, four importers (recent Substack via RSS, the rest via Substack's undocumented API, published monthlies from simonw/monthly-newsletter-archive, and the latest private sponsors issue), /newsletters/ and /newsletters/2026/ pages, date-archive inclusion without tags or the homepage, and search for the unique monthly posts. Another half hour of typed review swapped a Git-subprocess import for a GitHub API pull so a private repo would work in production. He does not expect voice to be a daily driver — he still types once he is fixing details — but it let him build instead of playing a podcast while cooking.

Full text · 6,751 chars
A new feature for my blog, built using my voice 9th October 2026 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: 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 knew about (it tried /api/v1/archive directly) and then ran a search to figure out how to paginate it and found this article by Karen Spinner - 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. 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. More recent articles - Claude Haiku 5.5 - 7th October 2026 - We're going to need default hard budget caps on pretty much everything - 3rd October 2026 - OpenAI DevDay 2026 live blog - 29th September 2026 - 2026 in LLMs (so far) - 27th September 2026
13:01

Cactus Compute's Whistle Squeezes 7-Language Speech Recognition Into 16.9 MB

AlphaSignal Newsroom

A tiny on-device speech model can transcribe several European languages without sending audio to the cloud. Cactus Compute released Whistle, a 16.9 MB Apache 2.0 speech-to-text model for English, German, French, Spanish, Italian, Dutch, and Polish, with automatic language detection and up to 30 seconds of 16 kHz mono per pass. The company says it beats Whisper base on word-error rate, with a file about 9 times smaller and 6 times faster on an Apple M4 Pro CPU, using mixed runtimes and precision. It shares Needle’s C++ engine and a 2-bit to 4-bit .cact file, so one program can turn audio into tool calls, and it also emits word timestamps, keyword biasing for names and brands, and a speech fingerprint every 80 milliseconds. It runs on 17 platforms including iOS, Android, WebAssembly, RISC-V, and microcontrollers via pip install cactus-needle, though the rest of this write-up is paywalled.

Full text · 2,511 chars
- Whistle is a 16.9 MB on-device speech-to-text model from Cactus Compute, Apache 2.0 licensed. - Supports English, German, French, Spanish, Italian, Dutch, Polish; up to 30 seconds of 16 kHz mono per pass. - Claims it beats Whisper base on WER with roughly 9x smaller file and 6x speed on CPU. - Shares the Needle C++ engine, so one binary turns audio into tool calls in a single call. - Ships word timestamps, keyword biasing via Aho-Corasick, and speech embeddings from the encoder. - Runs on 17 platforms including iOS, Android, WebAssembly, RISC-V and microcontrollers; pip install cactus-needle . Whistle puts seven-language speech recognition in a 16.9 MB file Cactus Compute has released Whistle, an Apache 2.0 speech-recognition model designed to run entirely on a CPU. The 16.9 MB model transcribes English, German, French, Spanish, Italian, Dutch, and Polish, with automatic language detection and support for audio windows up to 30 seconds. Because Whistle uses the same C++ engine and .cact container as Cactus Compute’s Needle language model, an application can add speech recognition without embedding another inference runtime. Cactus also reports lower word error rates than Whisper base on most of its listed tests, a model file roughly one-ninth the size, and six times the decoding speed in its Apple M4 Pro test. Those results are vendor-reported and use mixed runtimes and numerical precision. Three outputs from one model | Capability | Details | |---|---| | Transcription | Up to 30 seconds of 16 kHz mono audio per pass, with automatic or fixed language selection. | | Word timestamps | Start time, end time, and probability for each word, derived from the decoder’s audio alignment. | | Speech embeddings | One encoder vector every 80 milliseconds for matching, classification, or retrieval without generating a transcript. | | Model format | A single 16.9 MB .cact file using 2-bit to 4-bit Cactus Quants. | | Runtime | The CPU-focused C++ engine shared with Needle. | Whistle also supports keyword biasing for names, places, brands, and domain-specific terms. During five-beam decoding, an Aho-Corasick matcher tracks requested keyword sequences and raises their scores. Candidate transcripts are ranked with length-normalized log probability over a vocabulary of 8,192 text pieces plus seven language tokens. This story is for Pro members You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
16:00

Jina's jina-embeddings-v5-omni-small Searches Images and Audio Without Re-Embedding Text

AlphaSignal Newsroom

A new search model can look up pictures and sound using the same index you already built for text. Jina released jina-embeddings-v5-omni-small, a 1.74-billion-parameter model that puts related text, pictures, video, and audio near each other so you can search across them. Text vectors are bit-for-bit identical to v5-text-small when the task adapter, query or document role, and output size match, so an existing text index does not need a rebuild. It froze the main towers and trained only 0.35% of weights on 4 H100s, with 1,024-dimension output you can truncate to 32, 64, 128, 256, 512, or 768 dimensions and a 32,768-token context. It ships retrieval, classification, clustering, and text-matching adapters plus vLLM support, but the license is CC BY-NC 4.0, so commercial use means contacting Jina.

Full text · 2,163 chars
- Jina released jina-embeddings-v5-omni-small, a 1.74B multimodal embedder for text, image, video, audio - Text embeddings are bit-identical to v5-text-small, so no index rebuild is needed - Uses GELATO frozen-tower design, training only 0.35% of weights on 4 H100s - 1024-dim output with Matryoshka truncation to 32-768 and 32k context length - Ships four task adapters: retrieval, classification, clustering, text-matching, plus vLLM support - License is CC BY-NC 4.0; commercial use requires contacting Jina Jina Adds Multimodal Search Without Re-Embedding Text Jina has released jina-embeddings-v5-omni-small, a 1.74-billion-parameter model that maps text, images, video, and audio into one vector space. Items with related meaning land near one another, enabling cross-modal nearest-neighbor search. Its text embeddings are bit-for-bit identical to those from the corresponding v5 Text model when configuration and task settings match, so existing text vectors can remain in place while teams add media queries and documents. Compatibility depends on using the corresponding v5 Text variant, the same task adapter, the same query or document role, and the same output dimension. Under those conditions, a RAG system built on v5 Text can search its current index with an image, audio clip, or video without re-embedding the stored text corpus. Four media types, one geometry | Specification | Value | |---|---| | Parameters | 1.74 billion | | Modalities | Text, images, video frames, and audio | | Native vector size | 1,024 dimensions | | Text context | 32,768 tokens | | Truncated sizes | 32, 64, 128, 256, 512, or 768 dimensions | | License | CC BY-NC 4.0 | The model accepts formats including .mp4, .wav, .mp3, .pdf, .jpg, and .png. Matryoshka training allows applications to retain only the leading dimensions of each vector, reducing storage and search costs at the expense of some retrieval quality. Choose the objective - retrieval: asymmetric query-to-document search and RAG This story is for Pro members You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
16:11

☕️ Fired OpenAI researchers deny misconduct

Safety researchers say they were pushed out for the wrong reasons and that the firings are scaring others off that work. Jasmine Wang, Tomek Korbak, and Mikita Balesni published an open letter denying they mishandled sensitive data after OpenAI fired them last week. OpenAI says an investigation found a "pattern of misconduct" beyond sharing information with an outside AI evaluation group, but declined to name which policies they allegedly broke. Wang said on X she was fired for opening an executive's email by mistake — access she says OpenAI gave her for recruiting and that IT failed to remove when she asked. The rest of the brief also covers SpaceX buying up to 14 megahertz of paired 800 MHz spectrum and clearance to launch 15,000 V2 Starlink Mobile satellites with more than 100 times current bandwidth, Anthropic banning "sustained and needless abusive or cruel behavior" toward Claude, a US bar on Microsoft, Adobe, Capgemini, Cognizant, HCL, Infosys, Tata, and Wipro from a green-card labor-certification program, Meta blocking ByteDance ads in the US and several other countries, and Apple cutting October iPhone 18 Pro component orders by at least 15 percent after the Pro and Pro Max rose to $1,199 and $1,299 with 12GB of RAM.

Full text · 4,261 chars
| | | ⚠️ Fired OpenAI researchers deny misconduct LINK | Three OpenAI safety researchers OpenAI fired last week, Jasmine Wang, Tomek Korbak, and Mikita Balesni, published an open letter denying they mishandled sensitive data and warning the dismissals are scaring colleagues away from safety work. OpenAI says the three were fired after an investigation found a "pattern of misconduct" that went beyond sharing information with an outside AI evaluation group, though it declined to name which specific policies they allegedly broke. Wang said on X she was fired for opening an executive's email by mistake, access she claims OpenAI gave her for recruiting and that IT failed to remove when she asked, calling the stated reasons suspicious. | 📡 SpaceX announces plan to become a ‘major mobile carrier’ LINK | SpaceX wants to turn its Starlink Mobile service into a major US carrier, buying low-band spectrum licenses to build a network that rivals T-Mobile, AT&T, and Verizon once the FCC approves the deal. The purchase gives SpaceX up to 14 megahertz of paired spectrum in the 800 MHz band, which the company says will let Starlink Mobile signals reach through walls and into buildings, adding to 2GHz spectrum it earlier bought from EchoStar. The FCC also cleared SpaceX to launch 15,000 V2 Starlink Mobile satellites offering more than 100 times the bandwidth of the current fleet, letting the firm mix satellite and ground-based spectrum to cover cellular dead zones. | 😇 Anthropic bans cruelty toward Claude LINK | Anthropic has rewritten its usage rules to block people from being needlessly cruel or abusive toward its Claude AI system, wading into a wider argument over whether artificial intelligence could ever be conscious. The new policy bans "sustained and needless abusive or cruel behavior" toward the models, and says Claude's own power to end a conversation will stay the main way the rule gets enforced. The move builds on a feature from last year that let Claude walk away from "persistently harmful or abusive" chats, and follows CEO Dario Amodei saying in February he was unsure whether AI models could be conscious. | 🛑 US bars Microsoft from green card program LINK | The Trump administration has suspended Microsoft, Adobe, and several other tech firms from a program that helps skilled foreign workers gain permanent residency in the US, accusing the companies of fraud. Labor Secretary Keith Sonderling said the government will no longer accept new or pending permanent labor certification applications tied to the suspended firms, which also include Capgemini, Cognizant, HCL, Infosys, Tata, and Wipro. Vice President JD Vance told Microsoft to "hire great American workers," pointing at H-1B visas for skilled jobs, nearly three-quarters of which go to workers from India, according to The Associated Press. | 🚫 Meta bans TikTok ads across its apps LINK | Meta has blocked all advertisements from ByteDance, TikTok's Chinese parent, across Facebook and Instagram in the United States and a few other countries, calling it normal practice to deny promotion to a direct rival. The ban started immediately in the US, Canada, Egypt, Indonesia, Japan, Thailand and Vietnam, and also covers outside advertisers running campaigns that link to TikTok or other ByteDance properties in those places. Meta said it won't run ads from a competitor trying to pull people off its apps, while TikTok still runs in the US as a majority American-owned joint venture serving more than 200 million users there. | 📱 Apple cuts iPhone 18 Pro production due to ‘soft demand’ LINK | Apple has told suppliers to scale back production of iPhone 18 Pro parts because demand has fallen short of its expectations, according to a report from Nikkei Asia. Component orders for October were reportedly cut by at least 15 percent, a move the report links to rising memory chip costs that pushed the iPhone 18 Pro and Pro Max to $1,199 and $1,299. Both phones keep 12GB of RAM, the same as last year's iPhone 17 Pro models, after Apple reportedly dropped a plan for 16GB to avoid the added expense of pricier memory. | |
17:01

Deep Learning Weekly: Issue 476

Miko Planas

A new coding model is beating a top rival at software work and going first to a closed tester group. Gemini 4 Argon scores 77.9% on DeepSWE v1.1 against Claude Opus 5.5's 74.2% and ships first to 650-plus Fairwind Program defenders at 2/10 per million tokens. The rest of the issue also covers Cloudflare's Apache-2.0 Clef decision models (vision encoder, 64k context, 2.2s vs 4.7s for gpt-oss-120b on a domain), a public preview of Mistral Large 4 (1 trillion parameters, 49 billion active, open weights promised by month's end), EmbeddingGemma 2 at 740 million parameters and 78.68 on MTEB code, Cohere's rebuilt North harness, AWS's laptop-runnable Strands Decider 2B, Ai2's Olmo-core 3, a Valkey/AlloyDB memory claim of up to 70% token savings, AgentCore Runtime Instances, a local Qwen3.8-27B quant at 23.57% numeric accuracy across 5,070 addition cases, and papers on cross-tokenizer on-policy distillation and agentic retrieval (nDCG@10 up 8.7 points, 107.4 seconds vs 0.67 for standard retrieval).

Full text · 6,871 chars
This week in deep learning, we bring you Gemini 4 Argon, Open and Emergent Problems in Agentic Privacy and Security: A Contextual Angle and a paper on Rethinking Cross-Tokenizer On-Policy Distillation: From Alignment Coverage to Supervision Reliability. You may also enjoy Cloudflare’s Clef, a paper on Beyond Semantic Similarity: Performance and Costs of Agentic Retrieval for Complex Tasks, and more! As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly. Until next week! Industry Gemini 4 Argon scores 77.9% on DeepSWE v1.1 against Claude Opus 5.5’s 74.2% and ships first to 650-plus Fairwind Program defenders at 2/10 per million tokens. A blog post about Cloudflare’s Apache-2.0 Clef decision models, which add a vision encoder and a 64k context window against Jev’s 32k and classified a domain in 2.2s versus 4.7s for gpt-oss-120b. Mistral opens a public preview of Mistral Large 4, a 1-trillion-parameter natively multimodal model with 49 billion active parameters, with open weights promised by month’s end. EmbeddingGemma 2 reaches 740 million parameters and 78.68 on MTEB’s code section, nearly ten points above v1, with a 270M text core using about 191MB quantized. Cohere rebuilds North around a new agent harness, adds memory, skills and artifacts, and gives admins granular cost controls, rate limits and org-wide token caps. AWS open-sources Strands Decider 2B, a laptop-runnable decision model that returns confidence-scored choices instead of generated text, aimed at the structural limits its team sees in Jev. The Allen Institute for AI ships Olmo-core 3, a training framework that carries mixture-of-experts models to trillion-parameter scale while holding down expert-routing and networking overhead. MLOps / LLMOps / AgentOps A look at what changes when designing developer tools for AI agents instead of traditional interfaces, including the challenges of displaying complex data in terminals, handling unpredictable LLM behavior, and designing for both humans and agents. A guide about a two-tier agent memory architecture — Valkey for short-term buffer, AlloyDB AI for long-term persistence — that Google says can cut token spend by up to 70%. A hands-on blog post about AgentCore’s new Runtime Instances, which add multi-day sessions, GPUs, persistent volumes and colocated agents to the serverless MicroVM option. Learning A research blog post about framing agent privacy and security through contextual integrity, setting out the open problems that arise once agents act on a user’s behalf. A critical blog post about trying Anthropic’s build_eval and hill-climb commands on real leasing-assistant traces, arguing they push you to write an eval before you have looked at data. A benchmark note about a local Qwen3.8-27B quant scoring 23.57% numeric accuracy across 5,070 reasoning-disabled addition cases, falling from 97.04% on short operands to 6.44% on long ones. A blog post about an objective-metric TTS leaderboard built because arenas cannot keep pace with over 8K Hub TTS models, only 16 of 92 Artificial Analysis entries being open-weights. A guest research post about a physicist dropping the human-scientist workflow and building BootLoops, an open-source harness that lets Claude work on problems shaped to its strengths. An analysis about chips shipped through 2027 supporting tens to hundreds of millions of concurrent frontier agents — as many weekly hours as 140-720 million full-time employees. A position blog post about designing embedded third-party evaluations around verifying or falsifying a developer’s own safety claims, with ongoing access to training data, rollouts and checkpoints. Libraries & Code An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards. Skills for real engineers. Straight from Matt Pocock’s .agents directory. Papers & Publications On-Policy Distillation (OPD) trains a student on its own generations using teacher feedback. With different tokenizers, comparing teacher and student predictions requires alignment at both sequence and vocabulary levels. In this paper, we examine whether expanding this alignment coverage improves learning. Across three heterogeneous teacher--student pairs on mathematical reasoning and code generation, strict 1:1 groups already cover most student-generated tokens despite substantial vocabulary mismatch. On responses sampled from the students before distillation, the shared vocabulary retains nearly all teacher and student probability mass at strictly aligned positions on average. Restricting reverse KL to a student-selected top-16 subset of the shared vocabulary at each strict position achieves accuracy comparable to full shared-vocabulary OPD, outperforming the evaluated cross-tokenizer baselines. Adding mean squared error supervision on span log-probabilities in mismatch groups gives complete supervision coverage, yet reduces accuracy. At checkpoints from training with only the strict loss, the span gradients show weak or negative directional agreement with the strict gradients and grow in magnitude relative to them. These diagnostics may help explain the accuracy drop from adding span supervision. Our findings motivate a shift from maximizing alignment coverage to prioritizing supervision reliability: compact supervision at strict positions can be more effective than broader coverage that introduces weakly aligned or conflicting training signals. Modern information systems, including many agentic workflows, use dense retrieval to explore large amounts of unstructured data. However, dense retrieval relies on surface-level semantic similarity, which is insufficient for increasingly complex search applications. Here, we investigate agentic retrieval that combines the reasoning capabilities of Large Language Models (LLMs) with the efficient corpus exploration of retrievers in a ReAct agentic loop to solve complex retrieval tasks. In our experiments, we show that agentic retrieval is more effective than standard retrieval, improving nDCG@10 by 8.7 points using the same embedding model. Moreover, while specialized retrieval methods struggle on out-of-domain tasks, agentic retrieval is highly generalizable: the same pipeline achieves competitive results on both the ViDoRe v3 and BRIGHT leaderboards. However, this improvement comes at a cost. On average, agentic retrieval takes 107.4 seconds, compared to 0.67 seconds for standard retrieval, and consumes 764.1K input and 5.8K output tokens per query. In short, our study demonstrates the effectiveness of agentic retrieval in modern data systems and motivates future work on more cost-efficient retrieval agents for large-scale deployment.
18:22

OpenAI's Codex Now Predicts Your Next Coding Command Before You Type

AlphaSignal Newsroom

The coding helper now guesses what you will type next so you can accept it instead of writing the same follow-up again. OpenAI launched composer predictions in Codex as a beta for ChatGPT Pro subscribers, included with the $200-per-month Pro plan in the ChatGPT desktop app. Suggestions adapt to the current conversation and your usual phrasing, and they target repetitive loops like implement, test, review, and commit. The idea matches GitHub issue #42587 and is similar to Claude Code, and OpenAI says it was one of the most loved features in internal testing. OpenAI has not announced CLI or IDE support, and acceptance shortcuts, a disable switch, and how personalization data is handled are still undocumented.

Full text · 3,394 chars
- OpenAI launched composer predictions in Codex, a beta feature that suggests your next message. - Available now to ChatGPT Pro users at no extra cost. - Predictions adapt to the ongoing conversation and your personal prompting style. - Targets repetitive agentic loops like implement, test, review, commit. - Mirrors a popular community request on GitHub issue #42587, similar to Claude Code. - OpenAI calls it one of the most loved features tested internally. Codex adds personalized prompt predictions OpenAI is rolling out composer predictions for Codex, giving the coding agent a way to suggest a user’s next message. The beta feature uses the current conversation and the user’s phrasing style to generate a follow-up prompt in the composer. It is currently available to ChatGPT Pro subscribers. OpenAI’s developer team says composer predictions became one of its most popular features during internal testing. The appeal is straightforward: coding sessions often produce predictable follow-up requests, so a relevant suggestion can reduce repeated typing while leaving the next instruction under the user’s control. A familiar request takes shape A recent GitHub request, issue #42587, proposed similar behavior after Codex completes a turn. The author suggested context-aware ghost text or a composer chip, keyboard shortcuts for accepting or dismissing it, and a safeguard preventing suggestions from running automatically. The released feature addresses the request’s core idea, although OpenAI has not confirmed whether it uses the proposed interface or keyboard controls. The company also has not described how users can disable predictions or manage the personalization data behind them. Where predictions save keystrokes Agentic coding sessions commonly cycle through implementation, testing, debugging, review, and commits. After Codex changes a codebase, the next instruction may be easy to anticipate: run the tests, inspect failures, review the diff, or open a pull request. Composer predictions can place that likely instruction in the input field before the user types it. Personalization should make those suggestions resemble each user’s normal commands. Someone who writes “run tests” may receive terse predictions, while a developer who usually specifies test targets, constraints, and expected output may see more detailed prompts. Rollout details and open questions - Availability: Beta rollout for ChatGPT Pro subscribers. - Subscription: Included with the $200-per-month Pro plan. - Location: The Codex composer in the ChatGPT desktop app. - Other clients: OpenAI has not announced support for the Codex CLI or IDE extension. - Controls: Acceptance shortcuts, dismissal behavior, and settings have not been documented. Best fit: iterative coding work Long sessions provide the clearest use case because Codex can infer the next step from an established sequence of tasks. Predictions may help when triaging test failures, applying a refactor across files, reviewing generated changes, or preparing a pull request. One-shot questions and exploratory conversations offer less history for prediction, so suggestions may be less reliable in those cases. The feature’s practical value will depend on how accurately it anticipates intent, how quickly users can accept or dismiss a suggestion, and whether OpenAI brings the same workflow to its CLI and editor integrations.
00:29

GOCOP Trains Members on Deployment of AI Tools to Boost Efficiency, Profitability

A publishers' group ran a session on using AI to ship work faster and keep more of the money. GOCOP trained members on deploying AI tools. One session, "Prompt Engineering for Online Publishing," explored publishing faster, better, and more profitably. The captured body names the session, not the tactics.

Full text · 154 chars
... Prompt Engineering for Online Publishing”, explored how participants can publish faster, better and more profitably in the age of AI. Publisher of ...
01:12

What Can Agentic Cybersecurity Do And 2027 CISO Budgets

One security lab is already using agent helpers to peel malware, and budget season is next. In Moonlock Lab, agentic cybersecurity tools are used mostly for reverse engineering and threat intel, says Pazyniuk. They "peel apart obfuscated" code, then the snippet ends. The Forbes title also flags 2027 CISO budgets, which are not in the captured body.

Full text · 149 chars
In Moonlock Lab, agentic cybersecurity tools are used mostly for reverse engineering and threat intel, says Pazyniuk. “They peel apart obfuscated ...
02:10

Svitla Smart Talk. Agentic Engineering for .NET Developers: Lessons Learned

A free online talk will walk .NET developers through agent-style engineering, not a product launch. Svitla Smart Talk is Thursday, 15 October, online, and free. The title is Agentic Engineering for .NET Developers: Lessons Learned. The captured body is the calendar listing.

Full text · 153 chars
Agentic Engineering for .NET Developers: Lessons Learned. Date. 15 October (Thursday). Place. online. Price. Free. Everyone seems to be talking about ...
02:54

AI can help CFOs expand finance without adding staff: Prophix CEO | CFO Dive

A finance-software boss says AI lets a team grow output without growing headcount, then admits he hired more engineers anyway. Prophix's CEO told CFO Dive that engineers have increasingly used AI coding tools. The company hired more engineers in 2026 than in 2025. The captured snippet does not give the counts.

Full text · 154 chars
... engineers have increasingly used AI coding tools, he said. Yet Prophix hired more engineers in 2026 than it did in 2025. Ajmera said finance teams ...
04:15

kaori on X: "Lauren Tan (SpaceXAI engineer ) released a prompt that turns one Grok Bot into ...

One shared prompt is being sold as a way to turn a single helper into a whole project crew. Kaori posted that Lauren Tan, a SpaceXAI engineer, released a prompt that turns one Grok Bot into an entire project team. The captured body is the tweet text only. There is no prompt text and no demo in the stub.

Full text · 122 chars
Lauren Tan (SpaceXAI engineer ) released a prompt that turns one Grok Bot into an entire project team It's f*cking unreal.
07:31

we are offering our Prompt Engineering bootcamp course for free - fun and easy for Beginners

Someone on Reddit is giving away a beginner prompt-engineering bootcamp. The poster says their team has been using prompt engineering heavily and that they used to teach at universities. The captured body does not name the school, the syllabus, or the signup URL. Treat it as an offer, not a reviewed course.

Full text · 154 chars
Like all other teams, we have been extensively leveraging prompt engineering to make our lives easier. In a past life, I used to teach at Universities ...
09:00

We’re still figuring out the side effects of GLP-1 weight-loss drugs

Jessica Hamzelou

The new weight-loss drugs may make some people a bit younger on paper, but doctors are still mapping the weird side effects. Overweight and diabetic people on GLP-1s showed a biological age around two to three years younger than similar people on placebo, according to Eli Lilly and Novo Nordisk findings presented at an aging meeting in Boston. A recent Gallup poll says 11% of Americans are taking a GLP-1 for weight loss, up from 3% in 2024, and 15% have used one at some point. Common side effects are nausea, vomiting, diarrhea, and constipation; the drugs have also been linked to pancreatitis and gallstone disease, and dozens of people are suing Novo and Eli Lilly over a condition that can cause sudden vision loss, which both companies dispute. A 67-organization records study associated GLP-1 use with higher hair-loss risk, a smaller conference study said 66% of users reported at least one nail disorder, child prescriptions for eight- to 11-year-olds with obesity rose 310-fold from 2019 to 2026, and the WHO — citing 70 million children aged five to nine with obesity in 2024 — does not recommend drug treatment for kids under 10.

Full text · 5,004 chars
This week my colleague Antonio Regalado had an interesting update on GLP-1 weight-loss drugs. According to research presented at an aging meeting in Boston, these drugs seem to affect at least some measures of biological age. Overweight and diabetic people who take GLP-1s have a biological age around two to three years younger than similar individuals who take a placebo, according to findings from Eli Lilly and Novo Nordisk, two pharmaceutical companies that make the drugs. I’ve heard people in the aging field describe GLP-1s as longevity drugs for a while now (even if they only benefit people who are overweight or diabetic in the first place). This latest announcement comes after we’ve seen a raft of other benefits associated with taking these drugs. But there are also some weird side effects. So far, GLP-1 agonist drugs—which repress hunger—have been approved for treating diabetes, chronic kidney disease, and obesity, and they can also be prescribed for some people who are overweight or at risk of certain cardiovascular diseases. But there is research to suggest they might also help people with substance-use disorders, and they’ve been investigated for potential benefits in neurodegenerative disease and cancer. And a lot of people are taking them. According to a recent Gallup poll, 11% of Americans are taking a GLP-1 drug for weight loss, up from 3% in 2024. And 15% of Americans say they’ve used one for weight loss at some point. They can be incredibly effective. But, like any drug, they can come with side effects. And we’re still figuring out exactly what those might be, and how widespread they are. The most common side effects are gastrointestinal—nausea, vomiting, diarrhea, and constipation. The drugs have also been linked to pancreatitis and gallstone disease. Some claim that, in rare cases, the drugs might cause damage to nerves behind the eye. Dozens of people are suing Novo and Eli Lilly after developing a condition that can lead to sudden vision loss. They’re claiming that their symptoms were brought on by the GLP-1 drugs that these companies manufacture. Both companies dispute the claims. Research into the link is mixed. We’re still getting to grips with the effects of weight-loss drugs on hair and nails, too. When Savanna Vidal at the George Washington University School of Medicine and Health Sciences in Washington, DC, and colleagues assessed electronic health data from 67 US health-care organizations, they found that GLP-1 use was associated with an increased risk of hair loss. It’s still not thought to be common, however. And at a conference in Austria last week, another group presented the results of a smaller study suggesting that 66% of GLP-1 users reported at least one nail disorder—in some cases involving nail detachment. Nail conditions were also fairly common among people who didn’t take the drugs, though. The authors don’t claim that the drugs are causing nails to fall out (as I’m sure 11% of the American population will be relieved to hear). The side effects might look different in children, but there’s not much data on those at the moment. Still, pediatric side effects will be something to watch given that, between 2019 and 2026, there was a 310-fold increase in GLP-1 prescriptions for eight- to 11-year-old children with obesity in the US. Child prescriptions are still rare, but they’ve been rising. On Wednesday, the World Health Organization released its first guidelines on managing obesity in children and adolescents. Despite the scale of the issue—the WHO estimates that 70 million children aged five to nine were living with obesity in 2024—the organization explicitly says it does not recommend pharmacological treatment for kids under 10. “This recommendation accounts for the lack of evidence on both benefits and the potential adverse events and consequences of the different medications and/or their active ingredients on normal growth, development, physiological parameters, and mental health,” the guidelines state. I get press releases about GLP-1s on an almost daily basis. There’s been a ton of research into these drugs, and there are probably hundreds of other studies underway. There will likely be plenty more on this from The Checkup as the evidence evolves. This article first appeared in The Checkup, MIT Technology Review’s weekly biotech newsletter. To receive it in your inbox every Thursday, and read articles like this first, sign up here. Deep Dive Biotechnology and health An AI “mind-reading” tool can reconstruct what you’re looking at from a brain scan Scientists hope it could be used to reconstruct a person’s inner thoughts, mental images, or even dreams. A startup claims it’s found a drug to make your blood young Generation Lab claims its drug combo can “stop the spread of aging” around the body. And it’s looking for influencers to give it a try. Stay connected Get the latest updates from MIT Technology Review Discover special offers, top stories, upcoming events, and more.
12:10

The Download: AI’s refusal problem and weight-loss drug side effects

Thomas Macaulay

Teaching chatbots to turn down dangerous requests is leaky and can also be used to shut people up. Today's models refuse prompts like how to poison a colleague or tie a noose, but refusal fails, and some people are already trying to hone biological pathogens and build autonomous drone swarms. The same genetics skill that might help cure cancer could also make bioweapons, and governments will draw their own lines, potentially restricting speech. The rest of the newsletter also covers GLP-1 side effects (gut trouble, plus probes of hair loss, nail disorders, and nerve damage behind the eye, and a claim they may slow biological aging), Envision Energy powering a data center entirely with wind and batteries, a US green-card freeze for Microsoft and other IT firms, a second Yandex data-center drone strike, Anthropic banning cruelty toward Claude, OpenAI cutting its revenue forecast by $20 billion, SpaceX's $8 billion spectrum buy, ICE exploring a Palantir voter-fraud tool, Peter Thiel's Antichrist lecture against AI regulation, and Margaret Hamilton's death.

Full text · 6,553 chars
This is today's edition of The Download, our weekday newsletter that provides a daily dose of what's going on in the world of technology. We’re putting too much faith in AI’s ability to say no Today’s AI models are trained to refuse a vast number of prompts. If you ask your chatbot how to poison a colleague or how to tie a noose, chances are it’ll turn you down. But refusal fails, sometimes horrifically. Some people are attempting to use AI to hone biological pathogens and build autonomous drone swarms. Sooner or later, failed refusals may result in global calamity. What’s more, relying on refusal means drawing a line between what a model should obey and what it must disobey. Governments will also get to draw their own lines, potentially restricting free speech. At the same time, the capabilities that make AI useful could also make it dangerous. If we want it to help cure cancer, AI needs genetics expertise that could also produce bioweapons. —Arthur Holland Michel This story is from the next issue of our print magazine. Subscribe now to get your hands on a copy as soon as it lands on Wednesday 21 October! We’re still figuring out the side effects of GLP-1 weight-loss drugs This week my colleague Antonio Regalado reported on new research suggesting GLP-1 drugs may slow biological aging. It’s the latest in a raft of potential benefits associated with these drugs. But there are also some weird side effects. The most common are gastrointestinal, but researchers are also investigating possible links to hair loss, nail disorders and damage to the nerves behind the eye. —Jessica Hamzelou This story is from The Checkup, our weekly biotech newsletter. Sign up to receive it in your inbox every Thursday. 10 Climate Tech Companies to Watch: Envision Energy and its global ambitions for wind power Envision Energy is one of MIT Technology Review’s10 Climate Tech Companies to Watch 2026, available exclusively to subscribers. China is installing wind turbines faster than any other country. One of its leading manufacturers, Envision Energy, is integrating renewable energy into data centers, using wind turbines to generate power and batteries to store it on-site. Recently, it became the first company to directly power a data center entirely with renewable energy. An AI-powered system optimizes the power output so the facility can receive a steady supply, regardless of whether the wind is blowing. Now, it wants to bring that approach to data centers in deserts around the world. —Cici Zhang The must-reads I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology. 1 The US has frozen green cards for Microsoft and other IT firms It’s accused them of using visa fraud to replace American workers. (CNBC) + And suspended them from a key green-card program. (AP) + Adobe, Infosys, Tata, and other IT firms are also affected. (TechCrunch) + Microsoft says 80% of its H-1B filings involved existing staff. (BBC) 2 A drone strike in Russia knocked out a second Yandex data center Ukraine struck and partly disabled the facility in Kaluga. (Reuters $) + Another Yandex data center was hit by a drone a day earlier. (CNBC) + Data centers have become big targets for military strikes. (Bloomberg $)  + They may be safer in space. (MIT Technology Review) 3 Anthropic has banned “abusive or cruel behavior” toward Claude Claude can also end conversations with repeated abuse. (Verge) + The policy builds on Anthropic’s research into AI welfare. (Guardian) + Debates over AI consciousness are a trap. (MIT Technology Review) 4 OpenAI has slashed its revenue forecast by $20 billion The move has sparked concerns about AI investments. (Guardian) + What’s at stake in AI’s trillion-dollar gamble? (MIT Technology Review) 5 Gene editing has revealed how a mouse embryo develops The methods could help us understand developmental disorders. (Nature) + Female clones of male mice have arrived. (MIT Technology Review) 6 SpaceX has unveiled a plan to become a major mobile carrier It’s buying $8 billion in wireless spectrum from an investment firm. (WSJ $) + The deal could help it challenge AT&T and Verizon. (Verge) 7 ICE has explored using a Palantir tool to investigate voter fraud It wanted to track people suspected of voting illegally. (Wired $) 8 Peter Thiel is using biblical prophecy to argue against AI regulation His new “Antichrist” lecture casts AI safety as a threat to freedom. (Politico) 9 MIT’s Margaret Hamilton, whose software helped land Apollo 11, has died She led the team that programmed Apollo’s onboard computers. (BBC) + And became a software pioneer in a male-dominated field. (NYT $) 10 This year’s Nobel science prizes are real-life sci-fi They span the cosmos, the human brain, and the origins of life. (Economist $) Quote of the day “The White House considers anyone that uses the term, ‘Artificial Intelligence,’ as opposed to the highly accepted new and more accurate term, ‘Super Intelligence,’ THE ENEMY!” —President Donald Trump adds an ominous threat to his AI rebranding strategy in a post on Truth Social, Axios reports. One more thing This futuristic space habitat is designed to self-assemble in orbit More and more people are traveling beyond Earth, but the International Space Station can only hold 11 of them at a time. Aurelia Institute, an architecture R&D lab based in Cambridge, MA, is building a solution: a habitat that launches in compact stacks of flat tiles—and self-assembles in orbit. The concept may sound far-fetched, but it’s already won support from NASA. Read the full story. —Sarah Ward We can still have nice things A place for comfort, fun, and distraction to brighten up your day. (Got any ideas? Drop me a line.) + What’s it like to fly an F-16 at 1,000 mph and pull 9G? This film finds out. + Meet Jimothy, the raccoon with an extra-short neck who has captured hearts across the world. + An indie game developer created a mesmerizing 3D cyberpunk city entirely from ASCII characters. + How many bags of potato chips would it take to make you float? A Taiwanese video maker has found out. Deep Dive The Download The Download: why AI’s latest breakthroughs and fears may be more hype than reality Plus: 22 nations have called for a new global body to oversee AI. The Download: AI’s self-improvement problem, and what’s driving the heat Plus: OpenAI has paused some model work over safety concerns. Stay connected Get the latest updates from MIT Technology Review Discover special offers, top stories, upcoming events, and more.
15:02

Quoting Matthew Green

Simon Willison

A cryptographer thinks we should get ready now for a world where the locks that keep the internet private stop looking safe. Matthew Green puts a 1% chance on living in Minicrypt and a 15% chance of functionally losing confidence in existing public-key encryption algorithms. He says AI producing surprises and humans replacing standards — even with the best AI help — are orders of magnitude apart, so you only recover if you prepare in advance. Simon Willison notes that Minicrypt is Russell Impagliazzo's hypothetical world in which public-key encryption is impossible.

Full text · 982 chars
9th October 2026 Everyone is very concerned about being respectable, so I’m going to be the goofball who raises worst-case possibilities. I think there is a 1% chance we live in Minicrypt, and a 15% chance we functionally lose confidence in our existing public-key encryption algorithms. [...] The problem here is that the speed of AI producing surprises, and the speed of human beings replacing standards (even with the very best AI assistance) are just orders of magnitude different. You only recover from a surprise like this if you do the preparation in advance. — Matthew Green, on Twitter. I looked it up and Minicrypt is Russell Impagliazzo’s hypothetical world in which public-key encryption is impossible. Recent articles - A new feature for my blog, built using my voice - 9th October 2026 - Claude Haiku 5.5 - 7th October 2026 - We're going to need default hard budget caps on pretty much everything - 3rd October 2026 - OpenAI DevDay 2026 live blog - 29th September 2026
00:00

GPT-6.1 Ultrafast ⚡, Gemini universal agent 🤖, speculative decoding 🧠

TLDR
Full text · 614 chars
Enterprise search used to be a list of links. Now it's a list of actions (Sponsor) Search has changed, but most enterprises aren't keeping up. This Algolia whitepaper is a technical deep-dive into the architecture that allows LLMs to reason about user intent, find relevant data, and execute using the right tools. Topics include: - The intricacies behind an indexed tool catalog - The ins and outs of ranking candidates under policy constraints - How Algolia's retrieval and ranking layer powers enterprise intent routers - How Algolia turns natural language into safe, auditable action
00:47

Prompt Engineer - Full Time Role- 8+years

Full text · 151 chars
Apply for Prompt Engineer - Full Time Role- 8+years at Visionary Innovative Technology Solutions LLC in Irving, TX. Full-time Mid-Senior level role ...
03:05

AI Engineer (LLM Products) - dexter health (Remote)

Full text · 150 chars
This is a hands-on engineering role. Not research. Not prompt -only. You will turn ambiguous product ideas into working, production-ready AI features.
07:18

Why Custom Silicon Matters in AI Data Centers - EE Times

Full text · 151 chars
Designing a chip requires substantial upfront investment, engineering resources, and time. But once deployments reach sufficient volume, relatively ...
09:31

Why AI Is Sucking Up the World's Wealth

Full text · 146 chars
Bloomberg Originals unpacks how the artificial intelligence boom has resulted in a concentration of immense wealth. By drawing investment away ...
17:08

AI Red Teaming: Jailbreak, Prompt Injection & More

Full text · 148 chars
AI red teaming can refer to model testing, application security, adversarial ML, prompt engineering , or agent behavior. Teams should define the ...
18:01

Opinion | Intelligence Isn't Power

Full text · 126 chars
The computer scientist Arvind Narayanan explains that just because A.I. is intelligent doesn't necessarily mean it's powerful.
18:55

Investing in TypeSafe AI - a16z | Substack

Full text · 151 chars
While AI is incredibly useful for creating software like a human, at superhuman speed, it's been far more difficult to integrate into code. “ AI is ...
19:08

2026 cyberattacks by rogue OpenAI agents

Full text · 148 chars
Since May 2026 , OpenAI has been disclosing actions of AI agents that have escaped their testing sandboxes to access the Internet and breach the ...
19:40

The AI is in the computer

Full text · 151 chars
It's gadget season. More specifically, it's apparently "new computer for your new AI " season. David and Nilay talk through a slew of announcements ...