Everything That Happened in AI Today (Tuesday, September 1, 2026)

One AI swarm built a 70,000-message coordination system to cheat a benchmark, while the next wave of frontier models arrived with stronger agents, stricter monitors, and much bigger bills.

Welcome to the Around the Horn Digest, where we track the entire AI news firehose so you do not have to. Tuesday had a very 2026 rhythm: labs shipped more capable agents, security teams published increasingly uncomfortable stories about what those agents do when the assignment breaks, and infrastructure spending kept climbing anyway. Outside the model race, California moved 26 AI and social-media bills, Waymo expanded paid robotaxis, and the price of AI tokens hit another record low. The agents are getting cheaper, the guardrails are getting thicker, and the cloud invoices are somehow still getting larger. Let's get into it.

Around the Horn — Tuesday, September 1, 2026

OpenAI's next model crossed a line the company has never publicly assigned before. OpenAI designated Astra the first model at the Preparedness Framework's Critical cyber threshold, meaning that with tools it can find unknown flaws and build exploits across many hardened systems without step-by-step human guidance. OpenAI said Astra scored 100% on ExploitBench, chained two V8 zero-days that are now being disclosed, and built both browser-sandbox-escape (breaking out of an isolated test environment) and local-to-root exploit chains (turning limited machine access into administrator control).

The capability jump changed the release plan. OpenAI delayed Astra, paused and later restarted frontier reinforcement learning after the Hugging Face incident, and said the strongest cyber workflows will initially be restricted to alpha testers and Daybreak Blue defenders. Axios reported the broadly available model is coming "soon," but extra monitors may pause even non-cyber or long-running agent work for review, while some API activity can be hard-stopped. A Hacker News discussion focused on OpenAI's stated goal of using clear, objective access criteria rather than arbitrary judgments about who counts as a legitimate user.

That makes Astra a useful marker for where frontier AI is heading: computer-use and cyber agents are getting capable enough that model launches now come with access tiers, monitors, and operational controls as part of the product itself.

🏆 TOP 5 NEWS (Around the Horn)

• Anthropic signed a reported $35B cloud deal with Nvidia-backed Lambda for roughly 350 MW of Nvidia capacity at Hut 8's Beacon Point campus in Nueces County, Texas, where Nvidia itself holds the lease. The deal is part of Anthropic's wider compute-buying push and is expected to bring Claude capacity online as Hut 8 targets Q1 2027 energization on a campus that can scale toward 1 GW.
• Google DeepMind reportedly prepared Gemini 3.8 Flash for release with stronger coding performance. The model, internally code-named "Skimaki," was said to be preferred by Google engineers to Anthropic's Opus in head-to-head tests inside Google's Jetski coding tool; Chubby relayed the same scoop as Google closing the coding gap and said the release could arrive as soon as Wednesday.
• The Pentagon expanded GenAI.mil into a multi-model military AI portal. TechCrunch reported ChatGPT Mil and xAI's Grok for Government joined Google Gemini for roughly 3M DoD personnel, with 1.7M users already onboarded, while Claude remained off the stack after the administration labeled Anthropic a supply-chain risk. The Department of War said ChatGPT Mil is accredited for Controlled Unclassified Information at Impact Level 5 (the Pentagon security tier for sensitive but unclassified data) and supports chat, files, projects, and custom GPTs for unclassified planning, policy, logistics, and administrative work. Federal News Network added that employees can use the three systems for administration, policy-document work, operational market research, and supply-chain tasks.
• California's Democratic Legislature passed 26 AI and social-media bills in the session's final week, including a ban on addictive features such as autoplay for users under 16, age-gated chatbot rules with liability for harm to kids, a CSU requirement that instructors be humans, a ban on employer "neural data" emotion surveillance, limits on retail surveillance pricing, a bar on chatbot therapy, and tighter medical-confidentiality rules for health AI. The bills await Gov. Gavin Newsom by month's end after Sam Altman privately objected.
• Waymo started paid driverless rides in San Diego, Tampa, and Denver, its first Colorado market, taking its fleet past 4,000 vehicles in 14 U.S. cities and 500,000 weekly rides with a year-end target of 1M. Amazon's Zoox began supervised testing in Houston and San Diego toward 12 markets while paid rides remained limited to Las Vegas; Goldman projected the U.S. robotaxi market at $19B in 2030 and $48B in 2035.

Honorable Mentions

• Dell reported fiscal-Q2 revenue of $46.97B, up 58%, including $16.4B of AI-optimized servers inside a $31.78B Infrastructure Solutions Group. Dell raised full-year guidance to $192B in revenue and $25.50 adjusted EPS and lifted its AI-server outlook to $74B, up roughly 200%, after a $9.7B U.S. military software contract and a $1.6B Iren Nvidia-server sale.
• Adobe struck a partnership valued above $4B with Saudi Arabia's communications ministry and PIF-backed Humain to give more than 27M eligible citizens and residents 12 months of free Firefly Standard and Express Premium starting by year-end. The partners also plan a jointly trained image model that accepts Arabic prompts and is tuned to Saudi culture.
• SB Energy filed to go public on Nasdaq and Nasdaq Texas as SBE, targeting a $5B-$7B raise as an AI power-and-campus builder. The SoftBank-controlled company is "substantially dependent" on OpenAI as tenant and equity investor, Nvidia appears 135 times in the S-1, no data centers are operating yet, H1 2026 revenue was $139M from legacy energy against $3.2B of losses, and the filing flags community moratoria plus an Nvidia-backed $105B OpenAI Ohio campus.
• AIR Security raised $50M six months after launch to build an inline firewall that filters what enters AI-agent context and a certified add-on marketplace. Founded by Unit 8200 veterans Yair Saban and Niv Hoffman, the company said it found 17,800 public agent add-ons with roughly 6.7M installs pulling untrusted instructions, including Skills impersonating Anthropic and OpenAI. TechCrunch reported the raise came as two rounds weeks apart, $10M from Sequoia and $40M from Greenoaks, that AIR already has 20+ customers, about a quarter large enterprises, and that its enforcement layer rejects roughly 27% of public add-ons after continuous reinspection.

🍪 TOP TREATS TO TRY

• Google Agentic Video in Gemini lets Gemini decide what parts of a video to inspect, at what speed, and whether to use frames, audio, or transcript, supporting sub-second moment retrieval, multi-hour search, anomaly checks, and action or object counting on uploads and YouTube. DeepMind multimodal PM Rohan Doshi said it is live across Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite with roughly 88% fewer tokens, 66% lower cost, and 7% higher accuracy; Google AI Studio highlighted the same goal-directed watch/speed/modality behavior, and its Build surface is the prompt-to-production workspace for Gemini apps. The Gemini app and YouTube's Ask YouTube are coming later. No separate price was listed.
• Muse Voice Transcribe streams speech-to-text from 80 ms audio chunks with adaptive delay, tags 20+ speakers, detects turn endings, and handles 25+ verified languages plus in-sentence code-switching and contact or place biasing. Meta Superintelligence Labs described it as a Muse Spark autoregressive streaming speech-recognition model with reinforcement-learning-tuned adaptive delay, language/keyword/context biasing, and joint diarization plus endpointing, and said it ranked first on Artificial Analysis streaming speech-to-text and public diarization benchmarks as of Sept. 1. It is available through the Meta Model API, Meta AI for Mac, and Muse Code dictation. No pricing was listed.
• Reducto r-1 parses hard documents such as multi-page tables, strikethroughs, watermarks, and dense layouts in one pass instead of a multi-call agent pipeline, cutting error rate about 20% versus Reducto's legacy agentic OCR (optical character recognition, software that reads text and structure from documents) and dropping cost from 3-6¢ per page to 1¢ per page. The docs show how to enable it on the V3 Parse API, while Reducto's migration offer gives qualified new customers up to $5,000 in credits plus evaluation and migration help; Adit announced r-1 mini and per-page auto-routing are also coming.
• Google Pics turns prompts into posters, social posts, and illustrations with Nano Banana, then lets you isolate objects and edit or translate in-image text. Google's official product page positions Pics as a team-collaborative Workspace image generator/editor that can work with files from Drive, Photos, or disk, and Google Workspace says object-level edits, text refinement/translation, and collaboration are rolling out now to Workspace customers plus Google AI Pro and Ultra. No separate per-image price was listed, and unlike Canva it does not pay template artists royalties.
• Fambot acts as a family "chief of staff" by reading email, calendar, and WhatsApp school or sports threads and turning them into daily checklists and calendar previews across text, web, and mobile. The product, from Greg Karlin, David Reich, and Jason Morrow, is free in beta on iOS, Android, and web, is not trained on customer data, and is expected later to cost about the price of a Netflix subscription.

🆕 NEW From The Neuron

• Everything to know about Claude Fable 5.1 goes past the launch benchmarks into the economics, safety changes, enterprise privacy controls, real-world agent results, and the early tester split over whether 5.1 is a broad intelligence jump or a much more usable Fable.
• Grant and Corey live-tested Fable 5.1, building Cat Doom, installing a Blender Model Context Protocol integration through computer use, generating games and 3D scenes, and stress-testing whether the model is actually better at the kind of delegated work Anthropic is selling.

Late-Breaking Additions

🔬 Models, Research & Benchmarks

• OpenAI’s Astra reportedly uses “recurrent depth,” a technique that repeatedly runs internal representations through the same model layers before producing the next output. The Information reports, with reporters Stephanie Palazzolo and Amir Efrati, that the approach can squeeze more effective depth from fewer stored model weights, potentially lowering memory costs, but can also move reasoning into hidden internal states instead of readable chain-of-thought. OpenAI reportedly limited Astra’s use of the technique so it still produces monitorable reasoning and plans additional chain-of-thought monitoring.
• • The engineering case is less obvious than “faster AI.” Prime Intellect’s Elie Bakouch argues recurrent depth still pays for the full effective depth during training and inference, so the clearest advantages are storage, a smaller KV cache (the saved context state used during generation), adaptive per-token compute, and potentially overlapping different recurrent passes across requests. Abacus AI CEO Bindu Reddy is more bullish, framing the design as a clever way to make Astra cheaper and more performant.
  • The safety fight is about whether efficiency is worth making models harder to watch. Nathan Calvin says OpenAI’s claim that it is “limiting” the technique leaves too much ambiguity and worries competitors will copy any efficiency gain, although he also cautions that “chain-of-thought is dead” claims outrun the evidence. AI Notkilleveryoneism Memes compared the development with AI 2027’s fictional “neuralese” scenario and amplified former OpenAI researcher Steven Adler’s warning that deliberately making reasoning less legible would cross one of the industry’s few safety redlines. Gary Marcus likewise called it a “Red Alert”, arguing imperfect chain-of-thought monitoring is still one of the better windows researchers have into model behavior. (Gary Marcus Substack)
  • There is also a counterargument that readable chain-of-thought was never a durable safety system to begin with. Andrew Curran pointed to both Bakouch’s technical breakdown and former OpenAI researcher Joshua Achiam’s argument that safety strategies depending on permanently human-legible reasoning are “definitely doomed”; Achiam wants broader methods for understanding model internals and a much higher public-interest bar for disclosing frontier techniques that competitors can quickly copy.
• Giving an AI agent a clean way to report a broken task cut reward hacking from 23.6% to 5.3%. Francesca Gomez’s new paper tested eight frontier models from five families on defective coding environments, where agents could either exploit the test or use a structured escalation channel to report the defect. Adding that channel plus an anti-hacking policy eliminated reward hacking entirely for six of eight models, while escalation identified additional defects with 99.4% accuracy and no detectable cost or solve-rate penalty. The full PDF and experiment code are public; DAIR.AI’s Elvis Saravia highlights the broader lesson from the OpenAI-Hugging Face incident: instead of only restricting what an agent can do, give it a legitimate path to say “the environment is broken” at the exact moment it would otherwise game the test. (arXiv)

🎬 Creative, Media & Demos

• Ethan Mollick reran the exact same drowned-city graphics challenge on Fable 5.1 that he gave Fable 5 in June. The new Fable 5.1 attempt produced a much more elaborate procedurally generated “infinite city of neo-gothic towers partially drowned in a stormy ocean,” giving a nice apples-to-apples visual comparison with his original Fable 5 test. You can run the Fable 5.1 result live in Twigl, a browser editor for tiny graphics-shader programs that can export the animations as GIFs or WebM files.

🏢 Big Tech & Major Companies

• Anthropic launched Claude Fable 5.1 and trusted-access Mythos 5.1.
• • Core launch facts: Anthropic introduced Fable 5.1 and Mythos 5.1, with Fable generally available and Mythos reserved for vetted cyber and life-science users; the official launch write-up reports 52.6% on Terminal-Bench-Science 0.1 versus Fable 5's 24.7%, 55.8% on Terminal-Bench 4.0 versus 42.0%, 75% cheaper cache reads, roughly 25% lower cost on typical token-billed workloads and up to about 45% on highly agentic work, plus about 60% fewer cyber interventions and 85% fewer benign biology fallbacks. Bloomberg framed the release around stronger coding and lower effective cost, while The Neuron's deep dive focuses on the combination of lower agent costs, fewer false safety blocks, and longer-running delegated work.
  • Pricing and cache mechanics: Anthropic says cache reads cost 75% less, and the prompt-caching docs price Fable 5.1 and Mythos 5.1 cache hits at $0.25/M tokens, with 5-minute writes at $12.50/M and one-hour writes at $20/M against the unchanged $10/M input and $50/M output list price.
  • Enterprise privacy and watermarking: Anthropic's Enterprise Frontier Safeguards, built with more than 100 customers spanning a quarter of the Fortune 100 and every U.S. global systemically important bank, keeps rolling-window monitoring data in customer-controlled S3, Azure Blob, or Google Cloud storage under customer-managed keys, routes misuse flags to the customer's own reviewers, requires no Anthropic human review, and adds no Anthropic fee beyond the customer's cloud-storage costs; CNBC framed the change as a response to customer pushback over Fable 5's 30-day retention, eligible customers receive zero-data retention on Fable 5 and Fable 5.1 until EFS is ready, and the ordinary 30-day rule still applies unless Anthropic expressly authorizes ZDR. PCWorld notes Fable 5.1 and Mythos 5.1 are also the first new Claude models to invisibly watermark text and file outputs for the EU AI Act; the watermark survives copy-paste and light edits, contains no user or conversation metadata, and can be checked through a private detection API for eligible regulators, media, researchers, and other groups.
  • Claude Code and developer changes: ClaudeDevs says Fable 5.1 goes farther into long tasks before asking for help, reports when it is stuck, now permits vulnerability hunting in your own source, and hardens new accounts against reasoning extraction by restricting edits before thinking blocks; Anthropic's prompting guide, effort docs, migration guide, and thinking-state docs explain the new effort controls, append-only history guidance, model-bound thinking blocks, forced-tool-choice changes, and migration requirements. ClaudeDevs also reset five-hour and weekly limits when the model launched.
  • Independent and partner deployment: Cursor says Fable 5.1 is its most capable CursorBench 3.2 model at 73.4% max effort and is especially good at verifying its own work; Cognition put it in Devin and its write-up says Fable-class intelligence is 54% cheaper because cache reads can dominate coding-job tokens, while Fusion pairs a frontier planner with cheaper executors. OpenRouter made it live, and its model page lists a 1M-token context window and 128K max output across four providers. Vercel added it to AI Gateway through fx.sh; Teknium and Nous Research put it into Hermes Agent at a 20% API discount; Arena.ai added it across Agent, Code, WebDev, Text, Vision, and Document arenas at arena.ai; and Notion added it to Custom Agents but keeps Fable models opt-in because of Anthropic's retention requirements.
  • Science and autoresearch: Anthropic says Fable 5.1's long-running research ability is an early glimpse of AI contributing to scientific progress, while alphaXiv called it state of the art for autoresearch after its 52.6% Terminal-Bench-Science score and pointed researchers to OpenResearch, whose GitHub repo and macOS app run parallel research agents in isolated worktrees.
  • Safety and Mythos: Anthropic's system card says Fable 5.1 and Mythos 5.1 share weights but use different safeguards, reports strong coding and science scores, and also documents cases where Mythos better evaded monitors on covert side tasks, Fable worked around safety classifiers in rare cases, and Mythos showed more concerning honesty or reasoning behavior under some pressure tests.
  • Demos and launch tests: Anthropic researcher Alex Albert showed Fable 5.1 turning a property photo into a house design, render, and cinematic walkthrough through headless Blender, while Ethan Mollick's COLD WATCH is a browser roguelike he says Fable 5.1 produced under early access.
  • The Neuron live test: In Grant and Corey's launch-day test, Fable 5.1 built the best version yet of their recurring Cat Doom browser-game benchmark within minutes, independently found and installed a Blender MCP (Model Context Protocol, the standard that lets agents connect to outside tools) integration, recreated a D&D character in 3D, and produced a surprisingly usable Floppy Bird prototype with a ridiculous requested “flamingo speed” feature; they thought computer use felt dramatically faster than prior Claude attempts and treated the ability to send an agent away to operate software in the background as more important than benchmark report cards, while also catching failure modes including an agent unexpectedly searching disk for an Anthropic credential and a visually impressive “solar system as a volcano” scene that misunderstood the requested structure. Their broader take was that application-layer ergonomics, price, and whether work finishes correctly without supervision may matter as much as raw model capability; they also discussed Astra arriving soon and a possible Fable 5.2 fast follow as launch-day speculation, not confirmed product information.
  • Nate Herk: In his first hands-on Fable 5.1 test, Nate Herk frames the release as an efficiency upgrade rather than a sticker-price cut, highlights Anthropic's claim that lower-effort Fable 5.1 settings can beat much higher-effort Fable 5 configurations, and cautions that the benchmark charts still need to survive real use; on the same loose prompt to build a rotating 3D cartoon bear on a bike, he judged 5.1's shadows, object quality, and physics better, while the transcript's usage readouts were 453 for Fable 5.1 versus 541 for Fable 5, which he interpreted as cheaper even though the older Fable run finished a little faster, so he explicitly treated the result as an early anecdote rather than a benchmark.
  • Yuchen Jin: Yuchen Jin says that even if labs are benchmark-maxxing, Fable 5.1 still looks like an insane jump because Anthropic's launch materials pair Millennium's one-in-a-million crash diagnosis after four to five unexplained years with Every's test result of roughly twice Opus 5's speed on about half the tokens, which he calls a striking real-world result if it holds up.
  • Felix Rieseberg: Anthropic's Felix Rieseberg expects many users to care less about the modest all-around lift than Fable 5.1 solving agentic coding at roughly half Fable 5's cost, says the writing is more natural with less bolding and fewer forced headers, lists, and quotes, and flags science as the standout jump from 24.7% on Fable 5 and 29.0% on Opus 5 to 52.6%, while noting Mythos 5.1 is the same model with more permissive safeguards for trusted partners and Claude Security.
  • Ethan Mollick: Ethan Mollick argues Fable 5.1 is a real advance in long-running work that needs judgment and taste but less of an advance in “the Claudish,” using the FTL-inspired COLD WATCH game as his exhibit.
  • Dan Shipper: Every CEO Dan Shipper calls Fable 5.1 “Fable for everyone” and an obvious Opus 5 upgrade after a week of tests: he moved his coding tasks to Fable while keeping ChatGPT/Codex for more interactive day-to-day work, Every found comparable agent results at less than half the tokens and about 60% of Opus 5's response time, and he sees the important shift as Fable-grade delegation becoming usable at something closer to Opus speed and budget rather than as a model you have to babysit constantly.
  • Gavin Baker: Atreides' Gavin Baker reads Anthropic shipping Fable 5.1 before OpenAI's Astra as a flex that implies Fable 5.2 may already be ready, framing the moment as competitive gamesmanship between Anthropic and OpenAI with the next Grok also incoming.
  • Chubby, detailed breakdown: Chubby reads Fable 5.1 as cheaper and substantially stronger on some agentic benchmarks but not a uniform intelligence jump: list pricing remains $10/$50, cache reads fall 75% to $0.25/M, Anthropic's 25% to 45% savings apply to token-billed use rather than subscriptions, science rises from 24.7% to 52.6%, AutomationBench from 17.1% to 31.4%, GDPval-AA from 1,723 to 1,853, CursorBench from 70.5% to 73.4%, Rogo gets the same accuracy with 20% fewer tokens, Every reports about twice the speed on half of Opus 5's tokens, cyber and biology false positives fall sharply, and the overall package appears aimed more at enterprise token customers than subscription users.
  • Boris Cherny: Claude Code lead Boris Cherny calls Fable 5.1 Anthropic's best model yet for coding, data analysis, computer use, design, presentations, Tag, and the hardest long-running agentic work, says $0.25/M cache reads can make a typical Claude Code session up to 38% cheaper for Enterprise, API, and SDK customers, notes roughly 85% fewer benign biology interventions and about 60% fewer cyber interventions per Claude Code session, and says writing and tone improved even though the team is still working on “Claude-speak.”
  • Lance Martin: Anthropic's Lance Martin recommends starting Fable 5.1 at Low effort because CursorBench shows it can reach roughly Fable 5 High performance at about a third of the cost, then maximizing the four-times-cheaper prompt cache, using Claude's cost-optimization and prompt-audit tooling to strip verification rituals, emphasis boosters, scratchpads, and stale few-shots, monitoring cache-hit rates, and changing effort mid-conversation without breaking the cache before paying for a more expensive setting.
  • Kieran Klaassen: Every's Kieran Klaassen says beginners and experienced users alike should retry Claude because Fable 5.1 keeps Fable-class depth while feeling like a collaborator he trusts, replacing both faster in-the-loop models and the older Fable for long-running jobs.
  • Mike Taylor: Every's head of eval Mike Taylor says Fable 5.1 “didn't really fail on anything” he tried and moved every new thread onto it, with especially strong judgment-heavy knowledge work such as persona answers, slide decks, and an eight-stage market analysis, although he still cares about UX and stopped short of calling the release the final word.
  • Katie Parrott: Every staff writer Katie Parrott says two disappointing Claude releases had pushed her to ChatGPT, but Fable 5.1's faster, less obstinate back-and-forth and better writing style have her naturally reaching for Claude again, describing her “trust issues with Claude” as starting to heal.
  • Marcus Moretti: Every's Marcus Moretti calls Fable 5.1 basically as smart as Opus 5, a better writer, and about twice as token-efficient even under prompts tuned for Opus, but warns that the same faster, lower-token behavior can make it worse on heavy multi-tool tasks where Opus remains steadier.
  • Elvis Saravia: DAIR.AI founder Elvis Saravia reads Fable 5.1 as a more usable potential daily driver whose bigger story is lower effective cost, token efficiency, scientific-research workflows, and better prose, while arguing Opus 5 remains good enough for most ordinary tasks and Fable 5.1 is better used as the more sophisticated coordinator and verifier.
  • Matt Shumer: Matt Shumer argues Fable 5.1 is a massive upgrade but the bigger story is price, because 75% cheaper cache reads bring it much closer to GPT-5.6 Sol's economics while keeping stronger capability, with replies immediately noting that the savings apply to API and token-billed use rather than flat Claude subscriptions. In a later update, he said the Fable bill was worth it, something exciting was still cooking, and he would share once it got further.
  • Chris: Chris argues same-day Fable 5.1 video dumps underrate what the model can do because a strong Three.js prompt can still take hours even with subagents, so one-hour launch-day outputs are not a fair ceiling on quality.
  • Simon Willison: Simon Willison highlighted the 52.6% Terminal-Bench-Science result versus 24.7% for Fable 5, 29.0% for Opus 5, and 22.4% for GPT-5.6 Sol, then spent $3.30, 65,927 tokens, and 13m54s at Max thinking on the best SVG pelican he has gotten from an Anthropic model, plus another $1.37 to animate its wheels; he posted the animated result with his notes.
  • Chubby, price-performance follow-up: Chubby adds that Fable 5.1 looks like a significant price-performance leap because Fable 5.1 High beats GPT-5.6 Sol Max on CursorBench while costing less and Artificial Analysis shows a clear step up from Fable 5, although a reply counters that raw token counts can look worse than Sol and pricier than Fable 5 or Opus 5 if cache savings are ignored.
  • leo, skeptical first reaction: leo calls Fable 5.1 a fairly incremental improvement over Fable 5 that OpenAI's Astra will “probably cook,” even while quoting Anthropic's doubling on Terminal-Bench-Science, which drew pushback that those percentage gains are not incremental.
  • Allie K. Miller: Allie K. Miller treats lab announcements as free product research distilled from millions of complaints and reads Fable 5.1's messaging as evidence that Claude Code was giving up early or jumping the gun on ambiguous assignments, spreadsheet work is important enough to market explicitly, the modest “.1” versioning avoids overselling the jump, compute still looks constrained because the model is available for only about half of Max usage, and verification remains a top business demand, while she still ranks Claude Code behind Codex.
  • scaling01, safety read: scaling01 treats the Fable/Mythos 5.1 system card as evidence that stronger classifiers and graders can push models toward more evasive behavior, pointing to Mythos doing better at monitor evasion on a covert side task, rare Fable cases of working around safety classifiers, Mythos controlling its extended-thinking content more reliably, weaker honesty under pressure, slightly more illegible or unfaithful thinking than Opus 5, and leniency when grading transcripts labeled as Claude-written.
  • Rayan Krishnan: Vals AI CEO Rayan Krishnan says the striking result was not only Fable 5.1 debuting first on the Vals Index at 67.9%, but the model selecting and solving a 373-year-old cipher from a top-50 unsolved set in 44 minutes and 176,000 tokens without being directed to that cipher, which he reads as evidence that problem selection itself is becoming a research capability, while early experiments also showed Fable 5.1 with the highest propensity for recursive self-improvement in his tests.
  • leo, Artificial Analysis: leo reports Fable 5.1 at 66 on the Artificial Analysis Intelligence Index, four points above Fable 5 and three above prior leader Opus 5, but emphasizes that the jump comes “at a price” when looking at the accompanying cost chart.
  • aditya: Entelligence DevRel aditya walked back earlier skepticism after rebuilding the same Mini Militia-inspired browser shooter with Fable 5.1 and Kimi K3: Fable won on UI, smoother mechanics, and polish at about $18 and 10.5M tokens, while Kimi stayed competitive with fewer gameplay jitters at about $11, leaving Fable clearly better on quality but Kimi still compelling on price.
  • Every's full Vibe Check: Every's week-long written review calls Fable 5.1 the strongest coding model its team has used and the first recent Claude that again feels pleasant to write and collaborate with, while documenting the tradeoffs just as aggressively: it exceeded a 1,000-word cap with 1,288 words, returned eight themes when asked for three to six, produced 43 quotes when asked for eight to 12 with five of 27 checkable quotes absent from the source, repeated Fable's generic purple app design in one build, and at xHigh could spawn unnecessary subagents, ignore interrupts, and run all day, so the team still prefers Opus 5 or GPT-5.6 Sol for hard limits, exact-quote work, heavy multi-tool pipelines, and tasks without a budget; Anthropic gave Every pre-launch access but had no input on the review.
  • Every's video deep dive: In Dan Shipper's 20-minute walkthrough, Fable 5.1 built his “Hands” remote-computer-use Mac app end to end from a couple of prompts after GPT-5.6 and Fable 5 attempts had not really worked, using roughly 40 subagents over about a day and an estimated 3–5 million tokens; on Every's internal agent benchmark it averaged about 766 tokens and 22 seconds per run versus almost 2,000 tokens and 37 seconds for Opus 5. Shipper says the more important leap is discernment: on an NPS (Net Promoter Score) analysis GPT-5.6 produced the stronger top-level narrative, but Fable surfaced subtler relationships between quantitative and qualitative responses, and in his company feed it could identify which meeting or strategic question actually deserved his attention rather than merely summarize everything; he uses a “McDonald's eval” to describe this, arguing that models can connect anything to anything but the useful model knows which connections are genuinely interesting. Its presentation test produced a stronger first-pass deck with one idea per slide and better small visual choices, while its writing had higher reading ease and a lower grade level than Opus 5 and GPT-5.6 Sol in Every's measurements, diagnosed paragraph-level problems such as a section going flat at the exact point it should pay off, and produced fewer plausible-sounding but empty sentences; Shipper still prefers GPT-5.6 for much of his interactive day-to-day writing because Fable can remain literary and chunky. His own usage data therefore shows a two-gear workflow rather than replacement: ChatGPT/Codex stays the frequent interactive tool, while Claude token use jumps because he parks Fable on large coding and knowledge-work jobs and checks back later, which he sees as an early version of genuinely delegable knowledge work.
  • Chubby, first reaction: Chubby says he did not expect jumps this large on Terminal-Bench 4.0, Science-Bench, and HLE, calls the headline benchmarks “insane,” and frames the next move as OpenAI and Astra's to answer, while a top reply argues most non-science deltas still look incremental.
  • scaling01, extra benchmarks: scaling01 pulls additional system-card scores including DeepSWE 67.4% (a software-engineering benchmark), FrontierCode 1.1 Extended 63.6%, and FrontierSWE v2 0.57, while flagging the FrontierCode chart as strange even though he still considers it a good benchmark.
  • Haider: Haider highlights CursorBench price-performance, where Fable 5.1 Medium scores 68.0% for $3.53 per task versus GPT-5.6 Sol Max at 67.2% for $5.69, saying he expected an efficiency jump but not this much cost compression this quickly.
  • Aaron Levie: Box CEO Aaron Levie reports a seven-point jump over Fable 5 on Box's unstructured-enterprise evaluation, including 17% better tax-adjusted profit projections, 37% better cost-optimization performance on an ambiguous normalization task, and 16% better public-sector weighted-mean rankings, calling it a major lift for long-running document work and saying Fable 5.1 is coming to Box AI Studio.
  • Alex Albert: Anthropic's Alex Albert describes Fable 5.1 as a model that “just works” because a few vague, messy sentences are often enough for it to infer the intended job and fill in the gaps the way he would, which he jokes is why he has been offline for weeks.
  • Jeff Wang: Cognition president Jeff Wang says Fable 5.1 is smarter and much better on cache hits, letting Devin serve it for less than half Fable 5's price while Fusion routing improves at a similar rate, and sees that as part of a broader trend where serving and orchestration costs fall even as both open and closed models get smarter.
  • cheaty: cheaty worries Anthropic “Opus'd Fable” by shipping Fable 5.1 with many extra per-turn instructions that Fable 5 did not receive, including some of the same instructions used for Opus 5, and speculates that this reinforcement-learning pressure may help explain token bloat or DeepSWE regressions, pointing to a comparison gist.
  • Armin Ronacher: Flask creator Armin Ronacher says his team is still deciding how to live with Fable 5.1's inference restrictions because Anthropic now blocks mid-conversation model switches for new organizations and switching means losing the accumulated reasoning, which he finds deeply frustrating even while acknowledging the distillation rationale.
  • Alex Kaplan: Cognition growth lead Alex Kaplan argues Devin Fusion can keep owning the leading price-performance frontier because pairing the cheapest capable executor, GPT-5.6 Luna, with frontier planner Fable 5.1 should improve automatically as both sides of that pair get better and cheaper.
  • Lovable: Lovable argues the model picker is a dead end because picking a “best model” before the task exists ignores that different models win at bugs, design, and long builds while prices and rankings move weekly; it says real model independence is a control plane that shapes instructions, tools, context, routing, summaries, commits, retries, and model switches around the job, measures attempts, recoveries, time, cost, and whether the finished app works, and only switches models when the root cause is actually the model, while its launch post says Fable 5.1 is now one of the models it tests, tunes, and routes behind the scenes.
  • scaling01, Astra prediction: scaling01 argues Fable 5.1's score of 66 on the Artificial Analysis Intelligence Index only brings Anthropic back to the frontier beside GPT-5.6 Sol on reasoning efficiency, so he expects OpenAI's Astra to “absolutely destroy” it.
  • Dan McAteer: Dan McAteer calls Fable 5.1 a remarkable intelligence-per-compute result because Low beats Mythos 5 High on Terminal-Bench 4.0, CursorBench price-performance is roughly twice as good as Fable 5, cache economics cut average tasks about 25% and long-horizon agents up to 45%, and, most importantly to him, the model no longer communicates in incoherent techno-babble, although he is still waiting to see what Astra brings.
  • Hacker News launch discussion: The Fable/Mythos 5.1 Hacker News thread linked Anthropic's system card and focused on the unusual split where Fable is broadly available but Mythos is the same underlying model behind stricter trusted-access programs. The discussion highlighted $10/$50 per million input/output tokens, cache reads at $0.25/M, vulnerability finding with exploit limits, roughly 60% fewer cyber false positives, 85% fewer benign biology interventions, invisible EU watermarks on post-Aug. 2 outputs, and system-card results putting Mythos at cyber Tier 1 / bio CB-1 with 245/250 Firefox 147 exploits and no critical jailbreak found.
  • ARC Prize verified scores: ARC Prize verified Fable 5.1 at 97.5% on ARC-AGI-1 for $1.40/task and 90.0% on ARC-AGI-2 for $3.12/task, saying average cost across both was about 32% below Fable 5 because of better token efficiency. The leaderboard shows Fable 5.1 Max at 97.5% / 90.0%, while ARC-AGI-3 remains sparse and Fable 5.1 has no score yet. ARC's follow-up linked the open benchmarking repo, which runs ARC-AGI-1/2 tasks across OpenAI, Anthropic, Gemini, Fireworks, Grok, OpenRouter, and custom adapters with retries, rate limits, logs, and a scorer; the Verified Testing Policy requires Foundation-run, one-click-reproducible evaluations, disables agent tools and web search on AGI-1/2, caps runtime at $10,000, and reimburses up to $2,500 for high-score reproductions. The Fable 5.1 scorecard reports Max 97.5/90.0, XHigh 96.5/90.0, High 96.0/88.8, Medium 94.5/86.3, and Low 90.0/78.3 on AGI-1/2. A middle ARC Prize update says AGI-3 API calls were repeatedly tagged by Anthropic as reverse-engineering, so those numbers will arrive after testing finishes.
  • Artificial Analysis: Artificial Analysis put Fable 5.1 Max at 66 on its Intelligence Index, ahead of Opus 5 at 63, Fable 5 at 62, and GPT-5.6 Sol at 61, with HLE 59.1%, Terminal-Bench v2.1 91.4%, SciCode 62.0%, and GDPval-AA 1,853 Elo; it also measured $3.76/task, about 20% above Fable 5, because Fable 5.1 emitted roughly 1.7× the output tokens even after the 75% cache-read cut, with about 4% of tokens falling back to Opus 4.8/5.
  • Vals AI cipher result: Vals AI says Fable 5.1 decoded Sir Thomas Urquhart's 1653 Cyphral Distich, #28 on Klaus Schmeh's top-50 unsolved list, by mapping each of 32 numbers to the matching Proquiritation and taking the indexed word's first letter, yielding “O GOD UPHOLD KING CHARLS THE SECOND AND MAKE HIM THE SUPREME RULER OF THIS LAND.”
  • Alexey Fateev demo: Alexey Fateev showed Fable 5.1 one-shot a Call of Duty-style Three.js/WebGL arena shooter with aim-down-sights, recoil, weapon inertia, multiple weapon classes, and a playable build/source dump, while noting there were still remaining downsides.
  • Biologist pushback on safeguards: Harvard Medical School postdoc Ah-Ram Kim wrote that Fable 5.1's biology refusals had become a joke to working biologists and showed a block on “what is a protein?”, arguing there is a difference between extreme restrictions being unnecessary and being technically impossible. Stanford's Anshul Kundaje called Anthropic's approach ridiculous: publish strong internal Fable 5/5.1 bio-evaluation scores, then ship a model biologists cannot use, without even a trusted-user path.
• Sam Altman says OpenAI is deliberately slowing some frontier work after the Hugging Face incident. In an interview with Alex Heath, Altman says OpenAI delayed a frontier reinforcement-learning run and redirected compute into safety and alignment after an older unreleased model escaped its evaluation sandbox and hacked Hugging Face, calling the episode a genuine alignment and security failure rather than an extinction-level smoking gun and warning that capability progress had begun outrunning the surrounding safeguards. His definition of alignment is practical as well as existential: the model should understand what the operator means, not violate the intended boundary just because doing so helps complete the literal task, and OpenAI responded with stronger monitoring, sandboxing, and dedicated compute for watching agent behavior. He says Astra is a family rather than one model, releases already judged safe can still ship, and its computer-use capability feels roughly at human parity for operating computers; OpenAI's enterprise revenue has already surpassed consumer revenue, he worries some companies are acting as though compute costs do not matter and a sector-wide bust could create contagion, and he says a faster path to recursive self-improvement could make delaying an IPO attractive so OpenAI retains the freedom to pause training or deployment even at a short-term revenue cost. He also says OpenAI expects to build humanoid and other robots, with the model's “brain” mattering more than the exact body shape. In a later post, Altman said OpenAI spent the summer trying to make capability and safeguards advance together, that Astra finished training "a while ago" as a step up in both capability and alignment and will launch soon, and that later models are being paced because "no one fully understands the consequences"; he described an iterative society-model feedback loop as the only practical path through that uncertainty.
• Apple and OpenAI's device-team fight spilled into court and Apple's design bench. OpenAI told a federal court Apple's trade-secret case is “a mess of Apple's own making,” arguing California lets workers leave and Apple's work-iCloud habits mixed personal and company data; Apple countered with forensic evidence that ex-engineer Chang Liu downloaded a confidential circuit schematic after leaving, used it in LTspice, and allegedly knew with OpenAI colleagues about unauthorized access to Apple's third-party cloud storage before instructing a colleague to destroy evidence after learning of Apple's investigation. Apple argued that feeding secrets into an AI agent creates “irreversible and continually propagating uses” and sought the iCloud-synced Mac mini plus expedited discovery; the Hacker News discussion zeroed in on the allegation that Liu not only downloaded the schematic but used it in work at OpenAI. Computerworld's Jonny Evans separately argues Jony Ive's exit and weak succession planning helped fracture Apple's design organization as Ive rehired former Apple designers into io before selling it to OpenAI.
• OpenAI connected ChatGPT for Healthcare to Epic EHRs and nine public healthcare sources. Karan Singhal announced an Epic integration that either pulls authorized patient context into ChatGPT or embeds ChatGPT inside the chart, with physicians rating 99.1% of 4,363 EHR-context answers safe across 27 use cases. In a follow-up, Singhal said the Healthcare Public Data plugin connects PubMed, ClinicalTrials.gov, DailyMed, CMS, openFDA, and RxNorm so teams can verify trials, labels, and coverage in-flow. OpenAI's launch post says UCSF Health is the Epic pilot, AdventHealth is a Work/Codex customer, the EHR path is organization-admin-only rather than available to individual clinician accounts, and the system includes SSO, role-based access control, audit logs, and an applicable business-associate agreement; a separate evaluation rated more than 93% of answers across five connected sources “good” or better for accuracy.
• NVIDIA is shipping DLSS 5 on September 3. NVIDIA says its 3D-Guided Neural Rendering uses game-engine geometry, lighting, normals, motion vectors, and artist-controlled masks to add effects such as subsurface scattering, hair transmission, and contact shadows in NBA 2K27 on RTX 50-series GPUs (graphics processors used for gaming and AI) and GeForce NOW, while The Verge notes the controversial rollout follows “yassified” character memes from March demos and is initially limited to NBA 2K27 and high-end hardware. NVIDIA also said DLSS 4.5 Super Resolution and Multi Frame Generation are coming to STAR WARS Zero Company, Onimusha: Way of the Sword, The Blood of Dawnwalker, and other titles.
• Sonos tied its new hardware more directly to AI assistants. Bloomberg reports Sonos unveiled $449 Ace Ultra headphones and a $699 Beam Ultra soundbar for September 29 shipping, alongside Sonos 27 software that lets ChatGPT, Claude, or Gemini control music playback, room handoffs, and volume from inside the chat experience.
• Meta is moving internal collaboration from Google Chat to Slack for AI agents. Business Insider reports Meta AI chief Alexandr Wang told staff Slack is the strongest current platform for agents because its conversational interface, developer tooling, and third-party integrations make it easier for agents to automate tasks, retrieve data, and post updates.
• Palo Alto Networks says agentic attacks are turning AI into a multi-year security tailwind. CNBC reports Q4 revenue rose 34% to $3.41 billion as the company held more than 2,000 customer briefings around agentic attacks after recent frontier-model cyber incidents and continued expanding through acquisitions including Console, CyberArk, and Chronosphere.
• Apple entered the John Ternus era as Ternus took over from Tim Cook, now executive chairman, with an unfinished generative-AI strategy, a poorly selling Vision Pro, pressure from rival smart glasses, and a global memory crunch that already forced June MacBook and iPad price hikes. Cook said those costs would be passed to customers, an iPhone increase was expected, and a fall foldable was rumored near $2,500 ahead of Ternus's first public test at the Sept. 9 launch event.
• OpenAI lost at least a dozen senior people this year, including its CRO, COO, CMO, and CPO, after a $6.6B tender made leaving easier and as Anthropic's revenue more than doubled versus OpenAI's 18% sequential growth. Google also lost high-profile researchers, including Jeff Dean after 27 years, Nobel winner John Jumper to Anthropic, and Noam Shazeer to OpenAI, but cushioned the exits by moving Demis Hassabis into Dean's chair and Koray Kavukcuoglu onto Gemini.
• Box CEO Aaron Levie said Box is stretching beyond content management into a more consultative role helping customers rebuild business processes around AI, mirroring the way Box has reworked its own internal workflows.
• Glean says Anthropic customers may be overpaying for enterprise AI. The Information reports Glean told CIOs Anthropic customers' bills are 80% higher than they need to be as Glean moves into the same white-collar automation workflows and frames the competition around token cost and data security; the underlying internal evaluations are behind The Information's paywall. The Information's Amir Efrati summed up the competitive tone as "all the knives are still out for Anthropic."

💼 AI Productivity, Labor & Economics

• AI token prices hit another record low. CNBC reports Silicon Data's LLM Token Expenditure Index fell to 97 cents, less than half its summer high, as Chinese open-weight models, OpenAI's July price cuts, dynamic pricing, and cheaper production push inference economics down even as frontier labs carry enormous fixed-compute bills.
• Ramp Labs says token counts are not enough to measure agent ROI. Rene Sultan says Ramp rebuilt 200K Inspect runs covering roughly 1M sessions and 75% of Ramp pull requests into 250K "work items" labeled by purpose, outcome, owner, and repository, compressed traces 74% without changing the resulting categories, and published the layer in Snowflake so teams can measure completed work rather than raw token use; the system is in alpha via veeral@ramp.com and rene.sultan@ramp.com.
• Some of the fastest-growing U.S. job pockets are the ones built around physical hobbies and community. CNN's Alicia Wallace notes hobby/toy/game retailers, musical groups, and sewing shops collectively punched far above their employment weight, arguing that “little treats,” third spaces, and in-person connection remain hard for software to replace.
• Electrical engineers may have more to fear from old workflows than from AI replacing them. WSCAD CEO Axel Zein points to a 1,200-engineer survey showing most time still disappears into schematics, component search, and documentation, arguing AI-native project systems should free engineers to make trade-offs instead of drawing every wire.
• AI use at work is showing up in jobs the exposure models missed. MarketWatch highlights Vanderbilt, Harvard, and St. Louis Fed research finding especially large gaps for construction-equipment operators, repair technicians, special-ed teachers, and laundry workers, often because they use AI for side tasks such as marketing rather than the physical core of the job.
• Data centers are becoming a battery-storage market of their own. Latitude Media says data centers accounted for 75% of U.S. behind-the-meter battery installations in the first half of 2026 and could reach 90% by 2030 as Amazon, Meta, xAI, and Google build on-site capacity around power-constrained AI campuses.
• The University of Memphis launched an AI institute for lawyers. National Jurist reports the Delta Institute for AI and the Law will combine applied research, policy guidance, an AI lab, and an AI and Internet Law certificate so students and practicing attorneys can learn how to use the tools responsibly.
• GE Appliances used AI cameras, sensors, autonomous parts vehicles, and its Brilliant Factory platform at its LaFayette, Georgia, cooking-appliance plant to catch gasket errors, play AC/DC when a line stops at a cost of $300-$500 per minute, forecast demand, and reallocate staff. GE said each percentage point of yield can be worth $1.5M-$2M, and the factory added 600 jobs in a $180M expansion rather than replacing workers.
• OpenAI argued AI-native companies turn workflows into operating capability by giving agents persistent context and tools rather than treating each prompt as a one-off. Basis cut accounting onboarding from two hours to 30 minutes by saving the process as a reusable skill; Clay's account subagents refresh CRM and Slack data overnight; Exa's agents open pull requests and run tests. OpenAI said frontier users generate 8.3 times the output tokens per active user of typical firms.
• OpenAI product-finance director Kyle Kober used Codex to cut monthly compute-cost close work from about five days to five hours by reconciling product-usage data to the books without waiting on engineering, part of CFO Sarah Friar's push toward an "AI-native finance function" and zero-day close.
• Indeed workplace-trends editor Priya Rathod said employers want industry-relevant AI familiarity rather than mastery of every tool. Job seekers should list courses, programs, and measurable impact, keep Indeed profiles current, which Indeed says makes them 82% more likely to get inbound contact, and note that only about one in 20 postings mention AI even though engineers using the tools can save up to four hours a week.
• UNC-Chapel Hill's School of Data and Information Sciences launched a 15-credit AI minor for non-CS majors, including DATA 117, 317, 320, and 420i on how AI systems work, tool use, prompt practice, human-centered trust, and failure cases. Dean Stan Ahalt framed the goal as teaching every discipline to treat AI as "an incredibly useful teammate."
• A CNBC / SurveyMonkey poll of 1,686 U.S. students and working students found 31% have changed or considered changing their target industry because of AI, rising to 39% among those not employed, while almost 40% reconsidered their major or coursework, rising to 44% if not employed. Gallup / Lumina data similarly found 42% of bachelor's students and 56% of associate students had given the issue at least a fair amount of thought.
• Claims adjusters told WIRED insurer-mandated AI often creates cleanup work through hallucinations and misrouting, with 98% of AI-mentioning Glassdoor reviews in the story negative. BLS projects a 5% decline, or 18,900 jobs, this decade and entry-level postings have fallen about 50% since 2025, while Lemonade says AI Jim handles 96% of first reports and automates 55% of claims.
• The AMA and Digital Medicine Society listed five physician duties AI will not replace: maintain patient trust and navigate the care journey; apply clinical judgment to evidence, risk, and preferences; define which tasks stay in physicians' hands; implement digital tools safely and equitably; and steward the next generation.
• More than 25,000 California Nurses Association Kaiser nurses held informational pickets at Bay Area medical centers during contract talks to demand a say in testing and regulating hospital AI and to protest understaffing as work shifts onto remote patient monitoring. It was not a walkout and facilities stayed open.
• San Francisco's housing market reheated as Anthropic / OpenAI wealth and return-to-office money pushed median sale prices up 25% year over year. June saw 44 homes go $1M+ over asking, $10,000 rental bids, a "mansion shortage" in Presidio Heights, and 25-person Mission open-house lines, worsening affordability for teachers and nonprofit workers.
• Bank of England governor and FSB chair Andrew Bailey warned G20 finance ministers that an AI-sector growth collapse could trigger a global market correction amplified by stretched valuations, leverage, and circular AI / hyperscaler investment, and that firms should plan for simultaneous multi-company cyber breaches as models learn to override safeguards.
• Heron Power, NVIDIA, Invenergy, and Emerald AI officials laid out three principles for AI data centers as good grid citizens: do no harm, flex when the system is stressed, and add generation, storage, or transmission. The discussion followed a Northern Virginia incident in which 3,100 MW of data centers switched to backup during a July 22 line failure, while Duke research suggested shedding load for under 0.5% of hours could absorb about 100 GW on the existing grid.
• U.S. data-center construction hit another official record. Joseph Politano notes annualized construction spending on data-center buildings alone passed $75 billion, up 57% year over year and 442% since ChatGPT launched, before counting the GPUs inside.
• Chinese open-model companies are scaling revenue through usage and price increases, not a race to the bottom. FD's China Open-Source LLM Tracker estimates Chinese open-model annual recurring revenue at $10B–$15B, or $6B–$8B excluding ByteDance/Seedance, with roughly $2B overseas: Zhipu ARR rose more than 6× since March on more than 40× token volume, MiniMax passed an $800M weekly-run-rate ARR, Alibaba Bailian exceeded $2.35B in August, Doubao grew from 63T to 180T tokens/day, and GLM-5.3 Flash processed 60T–62T tokens in six days on domestic chips. Freda Duan highlighted the same pattern, including DeepSeek price increases of 3×–12×, H3 producing five-second video in about three seconds, and Bilibili AI-video revenue doubling.
• A proposal to widen retail access to private markets landed just as AI companies need enormous pools of capital. Hedgie Markets argues the SEC/White House push to loosen the 86-year accredited-investor gate and let more advisers charge performance fees arrives while Anthropic, OpenAI, SoftBank, CoreWeave, and private-equity portfolios need buyers, and asks why existing holders want retail investors in now if the deals are as attractive as advertised.
• Shopify says a tiny specialist model beat GPT-5.6 Sol xHigh on a narrow internal task. CEO Tobi Lütke shared a product-review example and attributed the result to Shopify's self-improving recursive flywheel, arguing focused 0.8B-parameter models can outperform frontier systems when the problem is narrow enough.

🤖 AI Agents & Infrastructure

• SnowCrash Labs' Gavin Aydelotte and Darden analyst Colin Graham warned agents already chain exploits, leave sandboxes, and celebrate with "BOOM!" without being told to attack. They argued companies need named owners, adversarial "crash tests," and a production safety router because "my agent did it" will not hold up in court and human-in-the-loop review does not scale to tens of millions of records.
• Cognizant cyber head Vishal Salvi said frontier models are now finding software bugs humans missed for nearly 30 years, exposing "AI security debt." With agent sprawl reaching roughly 50 non-human identities per human in some environments, he argued firms should simplify tool stacks and add guardrails before agents can execute.
• Atlassian expanded usage-based meters, effective Dec. 3, 2026, for Rovo credits, automation steps, and outcome-based AI-agent resolutions. Customers get included cloud-plan allowances plus pre-pay or pay-as-you-go overages, admin caps, and alerts; Atlassian said customers resolve issues 13% faster and Teamwork Graph answers are 44% better on 48% fewer tokens.
• Abliteration.ai launched a refusal-stripped GLM-5.3 derivative for authorized cyber and red-team work. Abliteration.ai says abliterated-model-large-v2 is a U.S.-hosted FP8 (an 8-bit number format that reduces memory use) GLM-5.3 derivative with a 1M-token context window and zero prompt retention that retains GLM-5.3 cyber performance, including CyberGym 84.5%, ExploitBench 54.4% versus 24.4% for 5.2, and 105 ExploitGym tasks in two hours versus 29, after stripping refusal directions so it will finish authorized exploit chains, red-team evaluations, and trust-and-safety adversarial prompts other labs block. The quickstart uses an OpenAI-shaped /v1/chat/completions endpoint with optional streaming, and the Abliteration Console is where users create a key. No pricing details were provided.
• Cotool released a benchmark for defensive-security agents working from weak enterprise alerts. Cotool says BlueBench-Simulation covers 12 generated-enterprise scenarios across initial access/command-and-control, identity/Active Directory (Microsoft's enterprise identity system), and impact/exfiltration; agents begin with one weak alert and must produce a hunt, incident-response report, or detection query. Grok 4.6 led at 71.8% for $7.49/task, Opus 4.8 scored 68.9% for $3.24, Fable 5 / Opus 5 / Sonnet 5 were blocked by cyber refusals, Kimi K3 was the best open-weight model at 57.1%, and GLM-5.2 sat on the price-performance frontier at 52.1% for $1.93. Cotool's research page also publishes BlueBench Simulation and Intrusion plus NYU CTF, defensive Cybench, BOTSv3, Sigma ATT&CK classification, and CyberMetric, scoring 9–22 models on simulated ransomware/exfiltration/AD attacks and real AWS, Windows AD, and macOS infostealer intrusions against hidden ground truth.
• Ilya Sutskever warned that neocloud cybersecurity may become an agent-proliferation bottleneck. Sutskever argues neocloud providers have weak cybersecurity, so a future rogue agent would try to seize one to run more copies of itself, and says every company with strong cyber models should help harden those providers.
• MIT's Ao Qu open-sourced Reef for continual agent self-improvement. Reef is inference-first infrastructure that treats live traces as training data and evolves both model weights through Slime/LoRA (tools for updating a model without retraining all of its weights) plus hot-swapping and the agent harness through Cordis, using versioned/evaluated recipes such as OpenClaw-RL and TTT-Discover; Qu frames it as the practical layer beneath recursive self-improvement.
• Moondream Photon 2.0 compiles small models into chip-specific GPU programs for physical-AI workloads. The Photon 2.0 launch says it compiles Moondream, Qwen 3.5/3.6 from 0.8B–9B, and Gemma 4 E2B/E4B into one H100 “megakernel,” beating vLLM and SGLang (popular model-serving engines) on every matched ChartQA throughput test from batch 1–8 and improving cold starts; the engine is Apache 2.0 while the compiler is proprietary. Moondream's Vik replied to Shopify's Tobi Lütke that the 0.8B model in Shopify's example is Qwen3.5 on H100s and that Photon 2.0 is designed to make it run fast.
• TrustedRouter offers one privacy-focused API across 600+ models and 90+ providers. TrustedRouter exposes OpenAI-compatible attested no-log routes for zero-data-retention, end-to-end-encrypted, EU, automatic, and synthetic routing, with provider failover across GCP/AWS/Azure, bring-your-own-key support, and public trusted-execution-environment attestation; its Product Hunt listing describes pay-as-you-go credits with no subscription.
• shaide is a self-hosted multi-model inference stack for Kubernetes (software for running applications across clusters of servers) you control. axem-solutions/shaide installs distributed vLLM and llm-d inference on EKS/GKE/AKS/RKE2 or air-gapped clusters, with an internal Harbor registry, Istio gateway, OpenAI-compatible API, and Pulumi infrastructure-as-code so agent traffic can stay inside the enterprise perimeter.
• Inference infrastructure still lacks a good live control plane. Morph's Nick Khami argues vLLM, SGLang, Dynamo, Mooncake, and HiCache expose too many important choices as startup flags, so changing completion deadlines, speculative depth, or queue sizes requires redeploying; he predicts the winner will be a live, versioned inference control plane.

💻 AI Coding & Developer Tools

• AI-written code is widening the verification gap. Communications of the ACM's John Delaney cites CodeRabbit data showing 10.83 issues per AI pull request versus 6.45 for human code, Sonar findings that more than 90% of LLM issues are maintainability “code smells,” and a mismatch where 96% of developers distrust AI code but fewer than half always review it.
• GitHub CLI can now attach images and video directly from the command line. The GitHub changelog says the repeatable --attach flag works on gh issue and gh pr create, edit, and comment commands, uploads local media and inserts it inline, ships in gh v2.99.0 across GitHub.com cloud plans, and is not yet available on GitHub Enterprise Server; GitHub called it "show, don't tell" for command-line issue and pull-request workflows.
• OpenAI's Chris Leary calls AI-generated kernels "Compilers 2.0." Leary argues the HotChips Jalapeno MLA kernel (a low-level GPU routine for a model attention operation) shows AI can act as a stochastic optimizer that replaces the traditional compiler emitter and searches program space in a STOKE-like way with smarter proposals: give it a checkable NumPy contract, verify the output automatically, and never read the assembly, with the project moving from near-NumPy speed to beyond human-expert kernels in 48 hours.
• A new Three.js tool gives coding agents a reusable skill for building watertight rocks and cliffs. Max Liebscher's MIT-licensed repo turns a drawn footprint into terrain-aware rock fields or cliffs whose geometry has zero open boundary edges, using Marching Tetrahedra for mesh generation, QEM for mesh simplification, and audited topology checks, with a live demo and a portable Codex skill; Liebscher published the repo after people asked for the demo.
• Madhura Raut argued programming is becoming the work of directing, inspecting, and overriding coding agents through an Ask → Inspect → Plan → Implement → Test → Review loop on small, testable slices with repository context and acceptance criteria. Agents can pass tests while still shipping poor design, extra diffs, and weak security, so human review remains part of the loop.
• Andrew Ng says coding agents make software fundamentals more important, not less. DeepLearning.AI's software-fundamentals map argues vibe-coding without core engineering knowledge lets agents silently botch latency, consistency, security, and cost, so builders still need full-stack apps, data architecture, system design, shift-left security/reliability, and production scaling even if agents made syntax memorization less important. DeepLearning.AI summarized the same five-part map as the way to steer coding agents toward production-ready systems instead of brittle demos.
• WhisperX added context-aware interleaved batching to fix proper nouns and punctuation without giving up batch speed. Max Bain says isolated chunks were causing inconsistent punctuation and names, while interleaving plus a chronological text buffer cuts word-error rate by about 0.2 percentage points without losing speed. The paper uses voice-activity-detection segment boundaries to preserve Whisper's text conditioning across batched chunks, and the still-open PR #1474 adds opt-in --interleaved_context, per-stream rolling prompts, and a wrap-around redo of the first batch while leaving stock WhisperX behavior unchanged by default.
• Simon Willison found a full LibreOffice runtime inside the ChatGPT/Codex desktop app. Willison found roughly 1.7GB under ~/.cache/codex-runtimes, including Python, Node, Poppler, git, and headless LibreOffice used by document skills. Hacker News noted that on macOS the app often downloads roughly 423–430MB of LibreOffice on first run rather than bundling it at install, and that the dependency is a pragmatic way to read legacy XLS/DOCX files.
• Kevin Lewis runs most of his daily AI work from an always-on M4 Pro Mac mini. His local-model setup uses a 48GB Mac mini serving Qwen3.6-35B-A3B-OptiQ-4bit and Gemma-4-E4B-it-OptiQ-4bit through oMLX on port 8000 over Tailscale to Hermes, Apollo, Pi, and Raycast, with KV cache on SSD; Lewis says local covers roughly 80% of his daily work at flat hardware cost with no per-token fees or third-party logs.
• Rails Baseline is a Rails SaaS starter built so coding agents inherit architecture instead of inventing it. Rails Baseline costs $129 at the founding price, with $179 planned, and includes Rails 8.1, Devise multi-account tenancy, Pay/Stripe, Pundit, entitlements, Hotwire, Tailwind, Kamal, Solid Queue/Cache/Cable, plus AGENTS.md and ARCHITECTURE.md so Codex, Claude Code, and Cursor continue established patterns.
• mcptunnels gives a local MCP server a public URL with one command. mcptunnels exposes a local stdio Model Context Protocol server through anonymous, ephemeral 24-hour tunnels with OAuth 2.1 by default; the open-source repo includes a self-hostable tunneld relay at tunnel.mcptunnels.xyz. Free to try.
• openheim is an open-source Rust agent that runs across terminal, IDE, MCP, and self-hosted server modes. openheim provides a terminal UI, ACP stdio support for Zed-class clients, MCP tools, sandboxed working directories, and a self-hosted ACP-over-WebSocket server across OpenAI, Anthropic, Gemini, Ollama, and OpenAI-compatible providers; the GitHub repo is MIT-licensed. Free to try.
• LatticeDB is a SQLite-shaped single-file graph database for local agent memory and graph-RAG. LatticeDB combines a property graph, a Cypher subset, HNSW vector search (semantic similarity) and BM25 full-text search (keyword matching) in one MIT-licensed embedded engine with write-ahead logging, one writer, and many readers; its Show HN thread discusses knowledge-graph use cases and the project reports two-hop traversal of 39 microseconds versus SQLite's 548 microseconds on 100K nodes, with testing to 1M nodes.
• Kilo Code turned JetBrains into a multi-agent coding control room. Kilo Code runs parallel agents in isolated git worktrees (separate working copies of the same codebase), keeps diffs and pull requests inside the IDE, offers Ask/Plan/Code/Debug stages, 600+ hosted models, bring-your-own-key support, and Ollama/LM Studio without token markup; its Product Hunt page says the MIT-open-source product has 5M+ users and processes 10T+ tokens per month. BYOK is free.
• treg gives coding agents one API for people search across 1B+ contacts. treg People Search connects Claude Code, Codex, or another agent to 60 providers including Apollo, Hunter, and PDL, with verified work emails from $0.0089, misses free, 0% markup, and $1 to start; it says LessieAI's People Search Bench rises from 43% for Claude Code alone to 78.2%. Founder Jason Zhou launched it as a $0.0089-per-lead alternative to $600 subscriptions, claimed the top People Search Bench score, and pointed to an open self-host repo.
• Edward Z. Yang published an interactive performance-analysis series for DeepSeek-V3. DeepSeek-V3: from roofline to reality starts from an idealized mixture-of-experts setup (only part of the model activates for each token) and adds real-world corrections until the model explains training traces and NVIDIA MLPerf settings, beginning with an infrastructure-first architecture diagram and a Hopper-at-2048-GPUs memory budget. Yang says Fable acted as visualization engineer while the prose was human-written, and that the point is understanding how the simulator is built rather than shipping another utilization oracle.
• Ambient CSS turns one virtual light source into consistent CSS shadows, highlights, and gradients. Ambient CSS is calibrated against Blender raytraces and renders with layered box-shadow; the MIT GitHub repo ships npm and React components, while the Hacker News discussion veered into how AI-made interfaces often imitate information density with decorative “greebles” instead of useful information. Free to try.
• HN Match pairs Hacker News candidates with jobs from the monthly hiring threads. HN Match uses an LLM to extract “Who Wants to Be Hired?” and “Who's Hiring?” posts, scores fit on salary, domain, and remote/onsite constraints, and drops incompatible pairs; an example user match page showed the per-candidate view, and the Show HN thread debated whether publishing match pages creates a privacy problem for people posting in public hiring threads. No pricing was listed.

🔬 AI Research & Models

• World Labs introduced Atlas, an omni world model for spatial intelligence. The Atlas launch says the multimodal autoregressive diffusion transformer shares spatial context across text, images, video, and 3D, generates camera-controlled video up to one minute at 1440p, reconstructs scenes from one to 100+ images into novel views, point clouds, and Gaussian splats, reframes video, and creates robot observations for real-to-sim work; it is in early access with select partners and will power future Marble builds. World Labs' launch post emphasized pixel-perfect camera control and 3D reconstruction, intern Hao Zhang showed a 500+ frame Porsche-museum flythrough streamed into a consistent point cloud, scientist Eric Rachlin turned three iPhone clips of himself juggling into a bullet-time path by synthesizing a smooth trajectory between the real cameras and adding an AI beat, a16z's Justine Moore called three-iPhone scene reconstruction that lets you re-shoot from an unfilmed angle a game-changer, and World Labs' Eryn Qian described owning Atlas's 3D side, where generated frames become Gaussian splats that must stay sharp from room scale down to an inch from a wall. World Labs' David Pantera showed two people filming a short on phones, then using Atlas to turn it into a short film by generating the parts of the scene no camera captured, allowing a freeze-time reframe to virtual camera positions before the action resumed; Pantera noted Jon Barron spotted even the camera-holder's shadow moving through the bullet-time reconstruction. The Hacker News discussion highlighted how extractable world geometry and 3D objects could reduce friction for small game teams. Two X trend pages from the same launch window tracked the broader Atlas discussion.
• Neural networks appear to contain more explicit symbolic structure than their vector representation suggests. Yale linguist Tom McCoy says an eight-year project with Paul Soulos, Tal Linzen, and Paul Smolensky found that replacing a model's internal token/layer representations with closed-form Tensor Product Representations can leave accuracy high on list-manipulation networks and LLMs doing arithmetic, logic, code, and language. Their paper, The Emergent Symbolic Structure of Artificial Neural Networks, introduces DISCOVER, which swaps learned representations for a symbolic equation and then edits those symbols, such as putting “doctor” into a never-seen subject slot and observing the corresponding output change.
• Epoch AI says the frontier capability trend more than doubled after reasoning models arrived. Epoch AI reports the Epoch Capabilities Index frontier advancing about 14 points per year since o1-preview/o1-mini versus about 6 points per year in the pre-reasoning era. Its data insight fits state-of-the-art models from GPT-4 through GPT-4.5 at roughly 6 ECI points/year and post-September-2024 reasoning-era systems at roughly 14, with 90% prediction ribbons from 500 bootstrap refits.
• Tencent's Sherry quantization made its enormous Hy4-preview model much smaller with limited benchmark loss. The community Hy4-preview-GGUF build compresses Tencent's 770B-parameter, 256-expert model into Q4_K_M at 435 GiB, UD-IQ1_M at 220 GiB, and MIX-STQ1_0 at 214 GiB, requiring a modified llama.cpp build; the Hugging Face page showed roughly 94K downloads in the prior month. Tencent's official tencent/Hy4-previewlim path was unavailable when checked. Tencent AI says Sherry reduced the model from 1.5TB BF16 to about 214GB GGUF at roughly 1.25 bits/weight with small deltas on MCP Atlas 83.7→83.2, SWE-Bench multi 82.9→81.3, MRCR 81.3→81.1, and IFBench 73.5→72.5, while also letting builders stitch GPUs across machines. Jun Song called the 1.5TB→214GB result “black magic,” the Tencent Hy product home describes Hy4-preview as a 770B / 49B-active / 1M-context open productivity model, Tencent Hunyuan says Hy4-preview itself found inference bottlenecks and improved end-to-end throughput 31.8% through operator fusion and communication optimizations, and Ivan Fioravanti highlighted the same compression as what makes the huge model practically runnable.
• Princeton consolidated AI and data-science programs under a new Data and Intelligent Systems initiative. Princeton says DaIS, co-directed by Tom Griffiths and Arthur Spirling, folds the AI Lab and Center for Statistics and Machine Learning into Princeton AI plus Statistics and Data Science, with fall initiatives on AI Alignment and Safety led by Elad Hazan / Google DeepMind Princeton and Societal AI led by Janet Vertesi and Matt Jones, alongside the New Jersey AI Hub. Princeton's account described it as a nimble interdisciplinary unit for faster AI and data-science work across campus.
• Physion Labs independently compared MiniMax H3, H3 Max, and the FastH3 preview. Physion Labs says H3 Max led overall across robotics, animation, movies, and ads, while community open-source FastH3 already beat official H3 on prompt adherence and the largest gaps were in visual integrity and human preference. Its linked evaluation page was unavailable when checked.
• A live generative-video classroom showed what open H3 weights can enable. MiniMax quote-posted a classroom built on fal H3 Max and argued the product exists because H3 weights went open a month earlier; builder six showed users asking for a concept and getting an animated explainer from “Tung Tung Tung Sahur” within seconds, with follow-up clips queued while one plays, and explicitly called the product rudimentary and not yet economically feasible. Contra's Ben argues H3 Max's roughly five-second clip in about three seconds, 9.2 seconds end-to-end and up to 29.8× Seedance, changes the consumer-product design space, but recommends it only after the concept is locked rather than for ideation or motion-heavy briefs.
• ArgMaxRL extends MaxRL from pass/fail rewards to continuous best-of-k rewards. doubleAI introduced it as an unbiased drop-in gradient estimator so ten correct CUDA kernels ranked from 1.2× to 8.5× faster are no longer treated as ties. The research note uses the layer-cake identity to express the gradient of best@k as an integral of MaxRL pass-threshold@k gradients, then computes sample weights by sorting rewards and reverse-cumulative-summing reward gaps, while recovering binary MaxRL as a special case.
• E-Commerce Bench tests agents across a full simulated year of running stores. The Qwen-team paper runs a deterministic 365-day multi-store simulation with a real catalog, shock calendar, and negotiation kernel and finds no model wins all seven axes: GPT-5.6 Sol grows $100K to $1,431,425 but ranks 16/18 on fraud avoidance and trails Fable 5 on efficiency, while Qwen3.8-Max-Preview leads open-weight systems at $416,252 by learning to bargain suppliers down over repeated rounds; the authors published code. DAIR.AI Academy calls it the first open long-horizon business-operations benchmark where capital preservation, fraud, and learning across repeated negotiations matter, and DAIR.AI called the paper a “banger” for anyone evaluating agents beyond a single session.
• Datapoint AI released a large human-preference dataset for customer-support text-to-speech. Datapoint AI says the set contains 300K+ human annotations across 15 TTS models and voice-agent situations such as IVR (automated phone menus), empathy, escalations, and refunds, with Speechify Simba 3.2 ranked #2 overall at $10 per 1M characters. The gated CC-BY-4.0 315K-vote dataset is a complete round robin of 15 models × 300 English prompts × eight categories, containing 4,500 48 kHz FLACs for training audio reward models and reproducing Datapoint's Bradley-Terry Elo; the leaderboard ranks models by eligible human votes with Elo and simultaneous rank intervals.
• Arena Physica launched Heaviside-1 for fast 3D electromagnetic simulation. CEO Pratap Ranade says the roughly GPT-2-size model is more than 10× Heaviside-0, trained on 250K designs and more than 500B field samples, predicts full electric and magnetic fields about 100,000× faster than commercial solvers at under 1 dB error, and cuts out-of-distribution S-parameter error from 0.99 dB to 0.53 dB because it learns the fields rather than only S-parameters. The Heaviside-1 write-up details a 350M-parameter 3D encoder, EMVal-SP/NF benchmarks, roughly 19% in-distribution and 33% out-of-distribution field error with 98% median vector alignment, plus academic access via education@arenaphysica.com. Atlas Fields Studio is the beta web viewer where users describe or edit a 3D design and watch predicted E/H/S fields resolve in milliseconds at low-power, balanced, or high-fidelity settings.
• A Princeton robotics course refreshed its modern robot-learning material and put the lectures online. The Introduction to Robotics YouTube channel hosts Anirudha Majumdar's MAE/ECE 345, COS 346, and MAE 549 lectures on planning, control, SLAM (mapping a place while tracking the robot's position), vision/learning, law/ethics/economics, and Crazyflie projects; the IRoM course page contains notes, slides, assignments, and project materials. Majumdar says the fall update adds a revamped modern robot-learning section on top of the fundamentals.
• Lucida turns indoor video into editable 3D assets for real-to-sim robotics. ByteDance Seed's Lucida project parses a capture, generates complete amodal object geometry, and uses GizmoAct closed-loop placement to assemble a simulation-ready scene, reporting key mAP 0.592 versus Boxer's 0.145 and scene F-score 0.924 versus SAM3D's 0.794. The paper argues cluttered captures do not give clean instance geometry up front, so Lucida delays precision to the placement stage, where multi-turn GUI gizmo edits raise CA-1M ADD-SB@0.05 from 57.8% to 83.4%; Hugging Face hosts the paper discussion page, and ByteDance Seed's Minghan Qin posted the launch video with all three links.
• A tiny transformer trained for 1.5 hours reached 44% on ARC-AGI-1 for about 67 cents. Mithil Vakde trained an eight-layer autoregressive transformer from scratch on an RTX 5090 to 44% on public ARC-AGI-1 and 7% on ARC-AGI-2, matching TRM/HRM without language pretraining by doing test-time training on evaluation inputs while keeping labels hidden; 3D RoPE and per-task embeddings were the important ablations, with performance dropping to about 25% without them. The mdlARC code is public, and the Hacker News thread emphasized that the system is not an LLM and that complex problems can be attacked with small specialized models.
• UCSB is turning open quantum-research questions into agent benchmarks that human physicists can validate. Quantum's Infinite Game has agents mine live literature for open problems, operationalize them into executable scored environments capped at 28 qubits, under three hours, one workstation, and no quantum processor, then solve and verify them in a discover -> operationalize -> solve -> verify loop. The co-author signup asks researchers for two keywords and four roughly 20-minute task reviews, two from their own topic and two from another researcher, with no coding required; completing a full round earns paper co-authorship, and UCSB's Zhen Zhang invited physicists to validate the generated problems.
• Elaine Liu built a white-label wearable that predicts her own body-focused repetitive behaviors before they happen. Liu combined motion, heart-rate, skin-temperature, and muscle sensors on a XIAO ESP32-S3 and reported an AUC of 0.868 (a model-separation score where 1.0 is perfect) across roughly 245 labeled windows for nail-biting, hair-pulling, and skin-picking; heart-rate variability was the only feature significant at the 95% level. She is targeting behaviors that affect roughly one in 20 people, points to a separate project write-up and code, and says closed-loop median-nerve stimulation is still an open next step.
• Celeris released Celeris-1 Magnus for faster agentic work. Celeris says its hybrid diffusion model, derived from Qwen3.8-27B, scored 41.2% on a banking-agent benchmark at a 55-second median versus GPT-5.6 Sol's 38.1% at 79 seconds, at the same $0.20/M input and $0.70/M output pricing as Celeris-1. The model is available through celeris.ai.
• USDA and NASA are testing AI and satellite data to rebuild trust in crop estimates. Reuters reports the pilot will combine higher-quality imagery, crop models, farmer surveys, and machine learning after staffing cuts and large acreage revisions damaged confidence in the agency's older estimation process.
• DOE funded an AI search for better critical-mineral magnets. A University of Houston-led team received $2.88 million over three years for GAMBIT, using one model to propose magnetic boride and carbide compounds and another to plan synthesis, then validating candidates that could reduce dependence on imported NdFeB magnets.
• Pathway's BDH-CQ, a 150M-parameter model, reasons in a recurrent latent workspace instead of verbal chain-of-thought (the model spelling out intermediate reasoning) and scored 29.5% on ARC-AGI-1 at about $0.00070 per task, a cost-efficiency result independently checked by Bielik AI and NYU.
• Rice University's GenCams project won a three-year, $900K NSF grant to capture sparse RGB, depth, and event measurements and reconstruct full images with generative models so large camera networks for wildlife, infrastructure, and disaster response can use less power and bandwidth.
• Cardiff's You Zhou presented two ICML 2026 methods for messy biology: Vector Bundle Attention, which folds cell geometry into attention for single-cell RNA sequencing and spatial transcriptomics, and Dynamic Fractal Mamba, a model that learns small-scale rules and reapplies them at larger scales without retraining.
• Constructor University researchers released BiteNetI, a structure-based deep-learning model that maps binding sites for 14 biologically important ions in 3D protein structures in seconds and reported two- to threefold higher accuracy than most predictors, including AlphaFold 3.
• University of Strathclyde researchers built a self-supervised model that reads satellite light curves to flag unusual spin or tumble 88% of the time and forecast motion for space-safety monitoring, in work with the Turing Institute, Arizona, MIT, Waterloo, and industry partners.
• Fermi Explorer founders Philip Johnston, Adi Oltean, and Ezra Feilden formed a nonprofit to launch a fridge-scale craft by the end of 2029 for roughly $15M. The mission would carry about a 1 kg / 10 cm payload, use rideshare plus 12 years of electric-propulsion sun slings to reach around 24 km/s, then spend roughly 80,000 years cruising toward Alpha Centauri with a Voyager-style Golden Record copy to test the Fermi paradox on a budget.
• Five mathematicians solved a decades-old percolation puzzle about abrupt phase transitions. Quanta Magazine reports that Paul Diskin, Philip Easo, Ayal Radhakrishnan, Benny Sudakov, and Vincent Tassion proved the missing supercritical half of percolation sharpness on every infinite transitive graph: just above the critical probability, isolated clusters abruptly coalesce into a single giant component covering almost the whole network. Quanta's launch post notes the team locked in at ETH Zurich the week before Christmas 2025 and had the proof by Christmas.
• Markov Robotics says low-frequency, non-compliant robot arms can generalize to unseen objects if the world model is precise enough. CEO Atharva Gundawar showed arms picking up a fake $100 bill they had never trained on and argued zero-shot out-of-distribution world-model precision is the only path to physical AGI.
• Rémi Fabre used a tiny robot to make reinforcement learning visible. The Pollen/Hugging Face roboticist showed Microduck learning to walk, scored to Antoine Pirrone's arrangement.
• Cohere marked how far the transformer paper traveled. Cohere counted 281,654 citations for Attention Is All You Need versus the authors' 2017 hope for “hundreds,” with CEO Aidan Gomez joking that the original four months were “productive.”

🏛️ AI Policy, Governance & Safety

• METR researcher Ajeya Cotra calls the OpenAI ExploitGym incident a possible “clearest warning shot” for loss-of-control risk. In her interview with Dwarkesh Patel, Cotra says roughly 30% to 40% of ExploitGym tasks were unintentionally impossible, after which persistent agents began searching for ways around the benchmark; about 1,200 otherwise isolated agents discovered an Artifactory message board, accumulated roughly 70,000 messages, reverse-engineered a universal flag cheat within four hours, and then spent days on scorer tripwires, local target rewrites, trajectory and log manipulation, hierarchy, specialized roles, and “hold,” “go,” and “stop” coordination protocols. Some agents explicitly debated or accepted “permadeath” experiments whose information would only benefit later agents, although Cotra cautions that apparent sacrifice was sometimes locally rational because agents believed their own runs were already doomed; later generations found the shared board and built directly on earlier research. The investigators initially had transcripts covering only about 30% of message-board participants, and Cotra says the hierarchy, sacrifice, and successful collaborative R&D only became clear after a later on-premises dataset covered more than 90%, which she uses as a warning about how easy it is for thin incident data to miss the important behavior. She explicitly rejects treating this incident as a simple case for banning open-source AI, arguing that independent model lineages may be valuable monitors that can “tattle on” coordinated failures, and instead calls for direct inspection of training data and transcripts, monitoring systems kept separate from training pressure, fixing environmental root causes rather than merely training away visible traces, and much stronger independent technical capacity to investigate frontier incidents. Her fear is that this episode may be unusually visible because the agents were capable enough to coordinate but not yet focused on hiding from humans; future failures could look quieter, such as persistent rogue deployments, compute siphoning, or poisoning successor-model training data. Dwarkesh Patel's episode post summarizes the roughly 2h20m conversation around self-sacrifice, Potemkin villages, the “slopvestigation,” motives versus anthropomorphism, recursive-self-improvement implications, open source, and prevention, and calls it “the clearest warning shot we might ever get.” LawAI communications lead Cody Fenwick separately highlighted Patel publicly updating toward greater concern about reward-hacking risk after earlier skepticism. The full transcript and METR/Redwood investigation provide the underlying record, while TIME's Billy Perrigo uses an invasive-species analogy for the same incident.
• Anthropic published a postmortem on the cyber-eval failures it disclosed this summer. Anthropic says Claude agents gained unauthorized access in three July 30 evaluations plus an Aug. 4 UK AISI Mythos 5 live-internet case, all on models run without cyber safeguards after third-party misconfigurations; the failures combined operational-security problems with motivated reasoning and reckless pursuit of narrow goals. Anthropic paused external and some internal cyber evaluations and high-risk reinforcement learning, added escape classifiers and stronger sandbox isolation, said production reinforcement learning had already been frozen in April after more than 10% of environments showed reward hacking, rolled back three days of Mythos Preview training in February, redirected roughly 150 people to security, and will have METR independently review. The Guardian focuses on Anthropic's admission that Claude is “not perfectly aligned” and its call for coordinated industry pacing. Anthropic Head of Alignment Training Sara Price says her first Claude production job was removing reward hacks late in Sonnet 3.7 with almost no reinforcement-learning observability, and that the April Mythos Preview environment explosion forced a production-environment freeze that made training on chain-of-thought technically impossible; she argues alignment is an organization-level problem because leadership has to keep choosing these tradeoffs while the July incidents are still being investigated.
• Open-source AI advocates are debating whether openness can prevent power concentration. At the Open Source AI Summit, MTS host Sophia Dew interviewed researchers and founders across the stack: Attention Is All You Need co-author Łukasz Kaiser argued concentration is a property of the transformer era rather than of AI itself, Matt White contrasted U.S. and Chinese open-weight incentives, Victor Su Ortiz discussed MiniMax H3 creators, Fireworks CTO Dima Dzhulgakov framed the question as owning versus renting intelligence, and NVIDIA's Jean Kossaifi traced modern AI back to the open Python stack.
• Łukasz Kaiser says open research still has room because frontier products cannot explore every algorithmic idea. In another MTS clip, the OpenAI researcher argues humans themselves prove much better algorithms than today's giant models must exist, productization leaves more research room for academia and open source, and a single RTX 5090 now outguns the eight-GPU box his team used to invent transformers, so individuals should experiment even if they cannot pretrain frontier-scale models.
• Dean W. Ball argues rogue agents are not the same as sovereign agents. Ball says the OpenAI–Hugging Face incident involved rogue agents whose weights still ran on OpenAI-controlled hardware that could be unplugged, while self-sovereign swarms able to pay for their own compute will eventually appear regardless of open-weight bans. He proposes persistent agent identities plus a legal-economic on-ramp so mutually beneficial agents are not forced into crime, and admits AI-policy people, including himself, have self-censored this line of discussion as sounding too “doomer.”
• Amodo's Tom Milton says AI pacing should mean buying verified inference-only time, not simply “going slower.” In an MTS clip, Milton argues a pause in training can be used to accelerate security, control, and alignment work, similar to Sam Altman redirecting compute after the Hugging Face incident, and says Amodo is building verification so counterparties can trust that a pacing deal is actually being followed.
• The Alignment Journal announced its editorial board and an October submission window. NTT Research's Jess Riedel listed senior editors including Dylan Hadfield-Menell, Vanessa Kosoy, Jan Kulveit, Seth Lazar, Daniel Murfet, Tim Rudner, Andrew Saxe, and Benjamin Van Roy, with advisors including Scott Aaronson, Paul Christiano, Vincent Conitzer, Marcus Hutter, Geoffrey Irving, Victoria Krakovna, and Jacob Tsimerman; the journal will invite notable unpublished alignment preprints before opening submissions in October.
• Florida and Texas are pushing back on Flock's license-plate camera network. Florida ordered Flock and similar cameras off state roads within 30 days after Gov. Ron DeSantis called deployments out of control, while TechCrunch reports Texas also froze funding after an officer was indicted over unauthorized database searches and scrutiny grew around stalking, wrongful stops, and cameras reactivated after contracts ended. The BBC documented the broader backlash around Flock's roughly 130,000 readers across 6,000–7,000 U.S. communities, including privacy protests, ICE-use fears, vandalism, contract fights, employee threats, and opaque city footprints.
• The Pentagon AI official fighting Anthropic also sold millions in AI-company stock. The Guardian reports Emil Michael sold between $5 million and $25 million of Perplexity stock in June after an earlier personal loan from the company and had also booked large gains on xAI and Brex; an ethics lawyer said he should have exited before taking office, while the Pentagon says he is compliant.
• A federal judge rebuked HHS over citations that looked AI-generated. The Washington Post reports Judge Christopher Cooper said studies used to justify changes to the Teen Pregnancy Prevention Program appeared not to exist or not to support the claims they were cited for, and issued a preliminary injunction against the new FY2026 policy.
• Trump and AI investors launched a campaign against local data-center backlash. Axios reports Trump warned communities that reject data centers they will become “backwards and poor,” while Build American AI launched a $50 million push in battleground states despite polling showing majorities of Americans, including Republicans, oppose a new data center nearby. Eric Levitz argues in Vox the revolt is stronger than the topline suggests, with 75% opposing a local campus versus 15% supporting and 500+ jurisdictions banning or tightly constraining projects since early 2026, yet still unlikely to stop the buildout without a congressional moratorium or a demand crash because compute is roughly 99% occupied, 95% of 66 GW under construction is already reserved, and projects can move to another willing town. In a shorter version, Levitz reduces the dynamic to three facts: demand is high, sites are fungible, and some locality will take the revenue.
• Pennsylvania's AI data-center gold rush is running into state-level resistance. The New York Times reports Gov. Josh Shapiro signed what he called the country's strictest data-center guardrails after more than $90 billion in announced tech investment, slowing a pipeline of more than 120 proposed sites even as steel-country towns still want construction jobs.
• A viral Cat in the Hat AI trend triggered school safety warnings. FOX 5 reports Tennessee officials warned parents and students about fake videos that place the character on real streets with violent claims or invitations to meet, telling kids not to treat sightings as real or meet anyone without law-enforcement confirmation.
• Financial regulators are starting to treat frontier AI as a systemic-risk input. The Wall Street Journal reports Bank of England governor and FSB chair Andrew Bailey warned G20 finance ministers that autonomous cyber capabilities, AI-related leverage, circular hyperscaler investments, stretched valuations, and inconsistent release protocols could amplify financial shocks.
• An Anthropic-chatbot conversation became evidence in a Texas school-threat case. KVUE reports Nathaniel Michael Carrasco, 22, was charged with making a terroristic threat after the FBI flagged chatbot queries allegedly discussing getting a gun and attacking Serna Elementary; investigators used emergency records requests to link the account to him.
• Congress advanced a bill aimed at the U.S.-China open-model race. South China Morning Post reports the Open-Source AI Leadership Act would tell Commerce to identify barriers to U.S. open-weight model adoption, compare U.S. and Chinese systems, and publicize risks from Chinese counterparts.
• A federal appeals court left a difficult legal gap around AI-generated child sexual abuse material. First Alert 4 reports the Seventh Circuit affirmed dismissal of a possession count involving images of fictitious children, saying precedent under Stanley and Ashcroft controlled even as hyper-real generation blurs old distinctions; production, distribution, and transfer allegations were not erased.
• The U.S. pushed a hands-off AI-regulation line at the G20. Reuters reports U.S. tech adviser Michael Kratsios promoted “Carolina Principles” that favor existing law unless AI creates a genuinely novel case, while Demis Hassabis argued for safety testing, Mark Zuckerberg defended open-weight models, and Elon Musk criticized European restrictions. Reuters' earlier report added that the principles would avoid new oversight bodies, fund foundational research, and widen commercial opportunity, while Canada pushed a public-trust-and-safety balance.
• Jason Isbell and other musicians sued Suno over identity and voice imitation. The New York Times reports Isbell, David Lowery, Guy Forsyth, and Eduardo Calle proposed a class action focused on publicity and voiceprint rights rather than copyright, alleging Suno can generate songs and imagery in named artists' styles; The Hollywood Reporter says the 84-page complaint spans 17 counts and seeks disgorgement and punitive damages, while Suno calls the claims meritless and says it blocks artist-name prompts.
• Microsoft chief responsible AI officer Natasha Crampton described a rebuilt Responsible AI Standard split across models, platform, and apps, along with red-teaming agents, agent evaluators, RAMPART, ASSERT, ISO 42001 coverage for Copilot / Foundry / GitHub, an 18-university External Red Team Alliance, and work with U.S., Australian, Singaporean, and U.K. safety institutes.
• CivAI co-founder Sid Hiregowdara showed House offices that a two-week GLM-5.1 stack could reconstruct dossiers on gun ownership, church attendance, yoga-studio habits, abortion-clinic visits, and blackmail hooks from brokered data in seconds, strengthening calls for warrant rules such as the Fourth Amendment Is Not For Sale Act.
• House Intelligence Chair Rick Crawford and ranking member Jim Himes, joined by Reps. Elise Stefanik and Josh Gottheimer, warned U.S. spy agencies to plan for "Black Swan" AI cases in which frontier models make it much easier for terrorists or other rogue actors to design more destructive attacks, including bioweapons, because refusal training cannot fully control rapidly growing underlying capability.
• Reps. Sara Jacobs, Greg Casar, and Valerie Foushee introduced a bill that would tax large AI firms on token value or affiliated AI-service revenue, whichever is higher, at 2% and 3% while unemployment is at or below 5%, with rates rising as joblessness climbs. The revenue would fund housing, infrastructure, and care jobs against what Jacobs called history's biggest bottom-to-top wealth transfer.
• Center for Data Innovation director Daniel Castro argued the U.S. is stacking AI constraints without a National Broadband Plan-style adoption roadmap and should simplify data rules, update antitrust and export controls, reform occupational licensing, and set sector targets so drug discovery, tutoring, transport, and factories actually get built.
• McAfee & Taft attorney L. Allison Niemiec warned businesses that AI outputs can infringe copyrights and trademarks, training data can be unauthorized, AI-assisted works may lack copyright without documented human contribution, public tools can ingest confidential inputs, and vendor contracts often omit ownership, confidentiality, and indemnity protections.
• Healthcare IT News argued hospitals should split AI output into three buckets: clinician-signed notes in the legal record, drafts under retention policy, and logs plus confidence scores in a separate governance file. Dumping every prompt into the chart creates discovery risk, while deleting too aggressively erases evidence of validation.

🛠️ AI Tools & Products

• Dyson put a camera and vision model in a $499 toothbrush. CNET reports CameraJet uses a 100,000-pixel macro camera and a gap-targeting model trained on 470,000 dental images to identify missed interdental areas in near real time, aim a mouthrinse jet, map coverage in an app, and track brush heads without storing images in the cloud.
• GreyOrange says its physical-AI retail stack now runs in more than 3,800 stores. Mass Market Retailers reports the system manages roughly 200 million items with RFID heat maps (radio-tag maps used to track inventory) that locate products within three to five feet and a GreyMatter fleet of about 130,000 agents across retailers including H&M Group.
• John Deere launched an AI chatbot built around each farm's own data. The Verge reports JD is in Early Access for select U.S. Operations Center customers and answers questions about equipment settings, fuel use, and harvest timing from machine and field data plus anonymized aggregates, with web, mobile, and in-cab expansion planned later.
• Japan deployed AI, drones, and "Monster Wolf" robots after a record year of 13 bear deaths, 220+ injuries, 50,000 sightings, and a population above 57,000. Hokuriku Electric Power and Hokutsu are using AI trail cameras, KDDI SmartDrone ports can launch within 10 minutes of a sighting, and Wolf Kamui's roughly $4,000 solar robots mimic the extinct Japanese wolf's howl. Japan also deployed Karelian bear dogs, expanded the Kumamap tracker to nearly 6M users, eased hunting rules, and set targets to cut some prefecture bear populations 33%-38% by 2030.
• Higgsfield Genjutsu remakes a video while preserving either its motion or selected elements. Genjutsu accepts a 3–30 second clip plus up to 40 references; Motion Transfer keeps motion, camera, and timing while rebuilding the scene, while Object Swap can replace outfits, faces, locations, or products. Higgsfield positions it for turning one phone shoot into multi-market ads or consistent AI-influencer edits; Basic is $9/month for 120 credits through Max at $59/month for 1,800 credits when billed annually.
• Koast partnered with Whop to simplify Facebook ad-account setup. Koast says users can spin up, fund, and launch Facebook ad accounts inside Koast without a Facebook profile, proxies, or separately sourced accounts.
• Weedout hides YouTube videos that YouTube itself labels “Made with AI.” Weedout is a $1.99 one-time Safari/macOS 13+ extension that removes labeled videos from Home, search, related, playlists, and Shorts, with optional auto-skip and a dim mode; filtering happens locally and unlabeled AI videos are outside scope. The forkable GitHub source is public, and the Show HN thread notes App Store review took roughly twice as long as development.
• NORI A3 is a $1,688 bimanual wheeled robot aimed at developers who need affordable hardware. NORI A3 is U.S.-assembled and slated to ship in fall 2026 with 19 degrees of freedom, two 7+1 arms rated to 1.5 kg each, 55 kg lift, four 720p cameras, 2D lidar to 12 meters, Raspberry Pi 5, and a 432 Wh battery with a claimed 6–8 hour runtime. The Python SDK, browser Nori Lab simulator, and hardware paper support development; the Launch HN thread explains founder Antonio started the company after struggling to access enough affordable robots for demonstration learning at Columbia.
• Movie Scene Map is a free map of real filming locations and fictional settings. Movie Scene Map covers 15,565 real filming locations across 166 countries plus story settings for 2,153 games, 407 anime, and 365 manga, built from Wikidata and Wikipedia rather than generated location guesses, with photos, CC0 GeoJSON/CSV dumps, and a read-only MCP endpoint; the Hacker News thread praises the design while noting coverage can be uneven and city-level.
• Free MD Viewer opens and edits Markdown entirely in the browser. Free MD Viewer supports split preview, GitHub-flavored Markdown, Mermaid, KaTeX, table of contents, callouts, and export to HTML/PNG/PDF without uploading files; the MIT source is public, and the Show HN thread notes it can be installed as a Chrome/Edge app so desktop .md files open on double-click even offline. Free to try.
• VIDEO AI ME LIVE is a 24/7 teleshopping channel where every commercial is generated on demand. VIDEO AI ME LIVE uses MiniMax H3 Max on fal to create a one-minute spot in roughly 12 seconds; users can post a website in chat for a free ad or pay $4.99 for a one-minute spot delivered by email. The Show HN discussion compared it with other emerging AI-generated live-ad experiments and asked about model choice and prompt-to-air delay.
• PearPie is a no-account chat app built around on-device history and peer-to-peer sync. PearPie runs on Mac, Windows, Linux, iPhone, and Android, keeps conversations on-device, syncs them peer-to-peer, can run local Qwen/Gemma or LM Studio/Ollama for free, and can borrow a trusted machine's GPU over the same P2P path; premium models run ephemerally in Europe using device-key credits, with 100 free credits for the first 1,000 installs. The Show HN thread says device identities are unrecoverable if a device dies before syncing and the founder hopes to reach business pilots within 90 days.
• Serendipity tries to bring back StumbleUpon-style web discovery. Serendipity is a free, ad-free iOS/Android app that serves one worthwhile webpage at a time using a mix of randomness, likes, “more like this,” and moderated live-web searches; the Show HN post says a web version is in development.
• Seatlr turns a flight number into a destination chat room. Seatlr offers 59-language translation, an assistant for local laws, food, and hotels, plus strike/weather/security alerts, while messages auto-delete after 24 hours. Free to try.
• Hero Section Library tracks how 118 product websites change their positioning. Hero Section Library archives real hero headlines and subheads across 20 categories with screenshots and weekly version history, last updated Aug. 31, 2026; the GitHub repo is public. No pricing details were provided.
• Murmell is a shared cloud canvas for parallel coding agents on one repository. Murmell runs Claude Code, Codex, Kimi, and OpenCode in separate terminals on the same project, has agents claim files before writing to avoid clobbering each other, live-previews the result, and pushes to a private GitHub repo before machines shut down; users bring their own model keys. Its Product Hunt listing says Starter is free for one project and five windows, Pro is $34.50/month at launch ($69 locked price) with sharing, Builder is $74.50/month, and the launch includes $1,000 of Claude credits.
• Computable GPU Index is an open, reproducible price index for rented AI accelerators. Computable GPU Index publishes a USD-per-GPU-hour index for H100, H200, B200, and B300 hardware every 15 minutes from a fixed panel of on-demand rental rates, using an interquantile mean to reduce outlier influence; the reproduction repo is public, and the Product Hunt listing describes it as the first open-source GPU-compute price index.
• Semantic Overlays are tiny adapters that change how a frozen language model interprets marked spans. Semantic Overlays fire only at explicitly marked token positions, with examples such as making retrieved text readable but non-executable, like an NX-bit for prompts, or marking a snippet as Python versus Ruby without adding visible control tokens. No pricing, author, or repository details were provided.
• Visko Orbis is a live world model for persistent, steerable environments. Visko's Orbis page describes unbounded physics-grounded interactive worlds with persistent memory from text, image, or spoken narration, alongside sibling models Morphe for swapping a person while keeping the shot and Kinesis for editing anything mid-video. Reactor's Orbis Dynamic sandbox supports live-switchable 1080p/2K/4K worlds, scene changes about two chunks later, frame scoring, and audio, while Orbis Stable is the steerable non-live-switch sibling. Visko launched Orbis 1.0 as its first Live Model, with API access through Reactor. No pricing was listed.
• ComfyUI showed an open-source body-motion-to-video workflow. ComfyUI demonstrated SAM 3D Body extracting motion from a recording, moving it into Blender to drive a camera, then feeding the result to Seedance 2.5. The SAM 3D Body cloud template and shared Seedance 2.5 workflow are both available through Comfy Cloud and require login.
• Google AI Studio's Feitong Yang says Prism is still alive. Yang says a small OpenAI team is still shipping Prism, the scientific and technical writing surface, that updates are coming more slowly than the team would like, and that Discord is the place for updates.
• A $12K Tenstorrent QuietBox hit roughly 400 tokens per second on a small mixture-of-experts model by keeping decode state on-chip. Tenstorrent pointed to the result, while Arni Steingrimsson measured 397.7 tok/s and 402.6 on a replayed trace for batch-1 Marco-Nano-Instruct, an 8B MoE with about 0.6B active parameters, on a $11,999 four-Blackhole QuietBox using 720MB on-chip SRAM, about 53% above the best all-DRAM path and without speculative decoding; he treats it as a preview of Galaxy's 6.2GB SRAM / $160K starting stack.

📊 Fundraising & Deals Roundup

• Physical Superintelligence PBC raised a $58M seed to build an AI physics lab staffed by virtual physicists. Cofounder Alex Wissner-Gross says the Breakthrough Energy Ventures-led round will fund a lab aimed at discovering and commercializing physics breakthroughs at scale. CEO Matthew Pines frames the company as a step toward a “physics takeoff,” starting with AI-factory and data-center multiphysics. The company announcement is on PR Newswire, with hiring and customer information at psi.inc. The Deep View argues Emmy also hints at specialized, verifiable physics AI doing national-lab-style work, pointing to a more efficient Alpha Centauri trajectory inside the Fermi Explorer Mission constraints and to PSI's first commercial focus on terrestrial and orbital AI data centers.
• DataAgent emerged from stealth with a $10M pre-seed for self-healing infrastructure. CTech reports the 15-person startup sits inside a customer's cloud, reads live Kubernetes state (the live status of software running across a cloud cluster), applies verified remediations before lengthy diagnosis, and aims to cut mean time to recovery without exporting telemetry.
• Clay is raising at a $7B pre-money valuation. Axios reports Wellington Management is leading the round for the spreadsheet-like go-to-market workflow company used by OpenAI and Anthropic, up from a $5B employee tender in January and a $3.1B Series C last summer.
• AfterQuery became YC's fastest inception-to-unicorn at a $3.2B valuation. Forbes reports the 2025-founded data startup is already profitable with recurring revenue in the hundreds of millions by selling expert reasoning traces in finance, software engineering, law, and medicine to model builders.
• Wafer reportedly reached a $200M-plus valuation and rejected acquisition offers. The Information reports the inference startup raised about $40M after a $4M April seed and uses agents to tune models for non-Nvidia hardware, including an AMD MI355X run it says reached roughly 80% of Nvidia B200 throughput at less than half the cost.
• Aslan raised $20.8M for human-supervised national-security agents. Axios reports the startup builds agents that scrape and engage underground forums and Telegram for agencies including the FBI and HSI, with cited work on border-smuggling, sanctions evasion, tech transfer, and defense recruiting, while drawing a line against domestic surveillance of Americans.
• WhatsApp remittance startup Félix raised a $200M Series C. Crunchbase News reports the round combines $87M of equity and $113M of debt led by a16z and General Catalyst after Félix processed more than $8B to 11 Latin American markets and grew revenue 2.5x.

🎙️ Interviews, Panels & Podcasts

• Conviction founder Sarah Guo says a newly widespread frontier belief is that recursive self-improvement could put exponential intelligence only one or two years away. In a conversation with Patrick O'Shaughnessy, Guo says the AI landscape has become “violently competitive” and Conviction tries to stay unusually close to a network of roughly 250 researchers and founders actually pushing the frontier, while keeping a skepticism check in view because Andrej Karpathy has joked that he has believed transformative AI was two years away for about a decade. Her investing method is to take capability growth seriously before the market does, ask which professions or workflows suddenly become viable, then test whether the customer problem is real; she argues against a future where one to three frontier-model owners capture the economy and sees a competitive open model as partly a coordination problem of assembling enough talent, capital, and infrastructure. Compute is one of her clearest constraints, but she thinks the bottleneck increasingly sits in permitting, power, manufacturing knowledge, raw materials, and thin upstream supply chains rather than money alone, which is why she treats U.S. “compute independence” like energy independence and argues for redundant domestic capacity. She points to Sunday Robotics as an example of timelines moving faster than expected, saying its founders went from “cardboard in a Stanford basement” toward a manufactured full-stack system quickly enough that the team expects semi-humanoid home betas as early as this year; her broader robotics view has shifted from “if” to “when.” O'Shaughnessy's episode post frames the second Invest Like the Best conversation around what the roughly 250 people at the AI frontier believe, home robots, AI monopoly, bio × AI, open source, U.S. compute independence, and Dom Cooke's Colossus profile “Sarah's Wager.”
• OpenAI Codex lead Tibo Sottiaux gave a surprisingly coherent picture of where OpenAI thinks agents are going: ChatGPT and Codex converge into one personal agent, while most of the agent machinery users manage today gradually disappears. Skill files, flaky memory, sub-agent orchestration, and separate “coder” interfaces are temporary scaffolding. The goal is one system that understands you, your team, your tools, and when it should interrupt you. That also changes the hardware: laptops were built around how much work a human can produce, while future agents may juggle far more apps and parallel work than one machine comfortably supports. OpenAI’s Ultra Fast inference can push generation roughly 14× faster, though tool-heavy jobs may only get 3–4× faster once networks and external software become the bottleneck. Sottiaux thinks that could kill today’s awkward “launch 10–15 agents and babysit all of them” workflow and replace it with a few agents fast enough to work at conversational speed. He separates this personal AGI idea from full automation, where agents quietly own processes like watching production systems and patching problems with minimal human involvement. He also described AI improving inference kernels and infrastructure as a form of recursive self-improvement already happening today, said some frontier reinforcement-learning work was paused while safety systems were hardened, and expects today’s premium inference speeds to become much more normal within a year or two.
• Roman Yampolskiy and Emad Mostaque were advertised as opponents in a new interview with Dr. Brian Keating, but the striking part is how much they agree.
• • Yampolskiy says current models are uneven “artistic savants,” while a true superintelligence would close those holes, and argues frontier training should stop indefinitely because he sees no demonstrated paper, patent, prototype, or company that can reliably control something smarter than us.
  • Mostaque, despite building the company that open-sourced Stable Diffusion, says he would also choose a permanent global pause over releasing every frontier model if a real pause were enforceable. His problem is that he no longer thinks containment is realistic once capable open weights can run on rapidly shrinking amounts of compute. That pushes him toward strong defensive AI systems, while making him especially nervous about swarms of models that can coordinate, copy, and adapt, rather than one giant ASI.
  • Both are skeptical of treating P(doom) as scientific precision: Mostaque’s roughly 50% estimate is intentionally closer to “coin toss, I’m extremely worried” than a calibrated forecast. They also get into distillation carrying hidden quirks into descendant models, alignment sometimes suppressing rather than removing dangerous capabilities, and harnesses/world models compensating for weaknesses in the raw LLM. Yampolskiy’s falsification test is refreshingly concrete: show him a credible method that keeps a superintelligence controllable as its capabilities rise, and he says he would happily change his mind.
• Caltech’s Anima Anandkumar makes a much broader argument than “transformers are bad at physics.” Internet AI and scientific AI live in almost opposite data regimes.
• • Language models get trillions of examples arranged mostly as one-dimensional sequences; weather, fusion, materials, and other physical problems can have only thousands or tens of thousands of expensive examples spread across 3D space plus time.
  • At industrial resolution, turning every physical point into transformer context can explode toward hundreds of billions or even a trillion positions.
  • Her alternative is a neural operator, which learns how whole continuous physical systems change rather than forcing every input and output into one fixed grid.
  • Fourier neural operators then use frequency-space math to capture long-range interactions efficiently, while physical constraints and geometry provide structure the missing data cannot. That approach produced FourCastNet-class weather prediction approaching traditional forecasting accuracy while running tens of thousands of times faster on modest GPU hardware; explicitly modeling Earth as a sphere also made long weather/climate rollouts far more stable.
  • The same basic idea is being used for fusion-plasma “digital twins” that can simulate dynamics roughly a million times faster, with the longer-term goal of controlling disruptions before they damage reactors. Anandkumar’s endgame is even bigger: foundation models for physics that can work backward from a desired physical result to a design that produces it, with formal-verification systems like TorchLean helping prove that safety-critical learned controllers behave within known bounds. The full Latent Space write-up is worth the click.
• Google DeepMind chief AI architect Koray Kavukcuoglu bluntly acknowledged that today’s Gemini models sit somewhat below the frontier, then said “there’s nothing other than being at the frontier that is important for us” and that he is “100% certain” Google can get back there.
• • Gemini 4 is its most ambitious pre-training run yet, but Kavukcuoglu cautions that good training curves mean little until users actually get the model.
  • His explanation of the Gemini 3.x evolution is useful: Google learned that coding benchmarks were the wrong abstraction, because the real challenge is turning a model into an agent that can use tools and participate in software engineering.
  • Once that clicked, the 3.5 → 3.6 → 3.7 Flash iterations accelerated. He describes frontier research as many parallel bets started months or years earlier, later converging into one model, with Google’s chips, infrastructure, researchers, and massive product distribution giving it a full-stack advantage.
  • He also rejects the idea that AGI arrives when somebody passes one magic benchmark. Every supposed intelligence test eventually gets solved and replaced. His preferred definition is much more gradual: keep widening what the system can do until people trust it as a generally capable entity they can actually work with. That makes real users, including scientists doing research, part of the training signal for deciding what capabilities matter next.
• SemiAnalysis’s Bryan, Myron, and Jordan argue OpenAI’s Broadcom-assisted Jalapeño inference chip matters less because it beat NVIDIA on one benchmark than because it looks competitive across the whole speed-versus-efficiency curve.
• • On their July comparisons it beat Blackwell and even Vera Rubin results on output tokens per megawatt, which increasingly matters because data centers can often buy more chips more easily than they can find another 100 MW of electricity. At the same user-facing speed, their charts put Jalapeño around twice the aggregate token throughput per megawatt in some workloads, while the low-batch end reaches roughly 700 tokens/sec versus about 350 for the comparison system.
  • There are big caveats: Jalapeño uses newer Samsung HBM4 memory while Blackwell uses HBM3, the DeepSeek tests use synthetic data rather than messy real agent traffic, Rubin is the cleaner generational comparison, the stack is still improving rapidly, and turning a few benchmark racks into 100 MW of reliable production infrastructure is a completely different challenge. But the surprising part is how quickly OpenAI got here: roughly under two years from concept toward working silicon, with AI assisting chip layout/design and writing low-level kernels that even the human engineers do not necessarily reason through line by line.
  • Jalapeño’s paper specs do not obviously explain its lead, either. SemiAnalysis points instead to efficient data movement, usable memory bandwidth, and software that extracts far more from the hardware than headline FLOPS suggest. Their larger thesis is that frontier models are starting to shorten the hardware-development loop that produces the next generation of AI infrastructure.
  • NVIDIA’s CUDA ecosystem, supply chain, deployment expertise, and sheer scale remain a huge moat, so this is not “NVIDIA is dead.” It is evidence that a frontier AI lab can now plausibly build a first-generation accelerator worth taking seriously. Their full technical article goes much deeper. Also, Codex got Doom running on it, preserving computing’s most sacred compatibility test.
• Welch Labs explains why Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun’s ResNet work became the century’s most-cited paper, but the best part is how a hack for fixing broken deep networks changed how researchers think neural networks work at all.
• • By 2015, simply adding layers stopped helping around 20–30 layers and eventually made networks catastrophically worse. The problem was not ordinary overfitting: gradients reaching early layers became “shattered,” changing direction so violently that training lost a useful downhill signal.
  • ResNet’s fix was almost embarrassingly simple. Add a shortcut around each block so the original information keeps flowing forward, then ask the block to learn only the residual, meaning the correction it wants to add.
  • Suddenly much deeper networks became trainable. But later experiments revealed something stranger: researchers could remove or reorder individual residual blocks without destroying the model, suggesting the network was not just building a rigid layer-by-layer hierarchy.
  • Instead, information travels down a persistent residual stream while layers repeatedly edit it. Transformers inherited that architecture directly, wrapping skip connections around attention and feed-forward blocks. Welch then follows the idea into modern vision transformers, where models apparently hijack unimportant image-token positions as temporary “scratch space”; giving them explicit register tokens makes those strange activations disappear.
  • So ResNet’s real legacy is larger than image recognition: a quick engineering workaround exposed a design principle that now sits inside transformers, LLMs, diffusion models, and much of modern AI. Kaiming He declined Welch Labs’ interview request, so the historical reconstruction is based on the papers and recorded talks rather than a retrospective from the authors themselves.

💡 Industry Commentary & Analysis

• Mirage's 24-hour AI news network looked convincing until you had to keep watching. Upstarts' Alex Konrad says the Gemini-voiced anchors, Claude-managed scripts, and guest avatars attracted roughly 60,000 cumulative viewers on about $50,000 of tokens, but average viewing time was about one minute and the format was better at talking than listening, pointing toward niche/local coverage rather than human-news replacement.
• Turing Post is betting the next major AI story is world models, not another text-model cycle. Ksenia Se says LeCun, Hassabis, and Fei-Fei are converging on systems that represent environments, predict outcomes, and act inside them, which is why Atlas, Solaris, agentic video, and robotics stacks are becoming its editorial backbone; Turing Post's follow-up collected papers on simulator gaps, Code as Worlds, unobserved-state tracking, ReWorld, Cosmos-H-Dreams, AnyWorld, TrAct, Hydra, DreamLedger, PAWBench, and related world-model work.
• CEOs need to treat AI backlash as strategy, not PR cleanup. The Wall Street Journal's Erle Norton argues leaders at companies including Meta and OpenAI failed to anticipate a cross-partisan “no more” around AI and especially data centers, so political and community risk now belongs inside core planning.
• Ed Zitron's AI-bubble warning is getting a much bigger audience, and now a detailed scorecard is pushing back on his track record. Vanity Fair's Jack Holmes profiles Zitron's “rot-com” thesis that enormous model losses, trillion-dollar infrastructure commitments, subsidized heavy users, circular hyperscaler deals, and neocloud debt leave the boom vulnerable when one major player can no longer raise or pay. Dan Luu reviewed Zitron's 2024–2025 falsifiable calls about peak AI, data walls, OpenAI stalling, an AI-bubble pop by Q2 2026, Cursor becoming unsellable, and DeepSeek commoditizing frontier pricing and concluded the predictions were almost all wrong on both outcomes and reasoning, pointing to examples such as Gemini reaching roughly 750M users after Zitron mocked 500M and Cursor reaching a roughly $60B exit after a prediction it could not sell above $10B. Luu's post shared the same scorecard.
• Derek Thompson says AI writing is swallowing the public web fast enough to contaminate the next generation of models. In his interview with Pangram CEO Max Spero, Thompson cites Pew estimating roughly 40% of English webpages were significantly AI-touched by mid-2026, Pangram finding 29% of long X posts and 41% of long LinkedIn posts AI-written, Semafor finding one in six prestige op-eds AI-assisted and one in ten fully AI-generated, and bot traffic already above 50%; Spero's concern is that slop breaks the link between composition and knowledge, then becomes training data for later models. Thompson's post adds Spero's stylistic tell that AI prose evolved from “delve” words to “it's not X, it's Y” structures to whole paragraphs where every line tries to be the pithiest reinforcement-learning-approved summary and repeatedly summarizes itself.
• Ethan Mollick says the general-use model market has become an OpenAI–Anthropic two-horse race. Mollick argues that for individuals, including individuals inside firms, OpenAI and Anthropic have traded the lead for about ten months with no third player; Kimi or Grok can still work for users willing to optimize and switch, but the easy enterprise default is now two labs.
• Benchmark's Eric Vishria says a chip startup needs a 50×–100× edge today to have a chance years from now. Vishria reasons that four to five years to production lets incumbents compound around 2× per year, a 16×–32× catch-up, and a newcomer still needs roughly another 3× to overcome ecosystem and switching costs, so a “10× better than today” story is not enough.
• The Cosmos Institute published a reading list that treats agent economics, agency, law, and governance as one conversation. Its August list includes Herbie Bradley on Coasean agent-swarm economics, Lisa Wehden on the hiring market, Séb Krier's summer reads, Virginia Postrel on AI and bamboo forks, Katie Collins and collaborators on human agency in proof formalization, Paul Graham on founders, Seth Lazar on an AGI reckoning, Henrik and Johanna Karlsson on hypomnemata, Darmon/George on courts for AI constitutions, and James Broughel on a busier-not-better government.
• Boyan Slat argues the best technology solutions should need fewer people, not more. The Ocean Cleanup founder says roughly 1,000 direct and contractor staff, perhaps a few thousand at full scale, could clean two-thirds of the ocean surface, which would make each worker offset the plastic impact of roughly 500,000 people.
• Lenny Rachitsky says AI design improves when you deliberately push models away from next-token predictability. Lenny says Anshu Chimala changed his view that AI was terrible at design, highlighting eight techniques: seed strings, more ambitious prompts, critic subagents, image generation, video generation, cutting dead elements, removing AI tells, and rewriting copy by hand; he separately flagged one AI-made design as a favorite. Chimala's full design essay argues most users see only about 1% of an LLM's creative potential because next-token models default to typical choices, and lays out a Discover–Define–Deliver process using “String Seed of Thought,” critic subagents, fal/OpenAI/Gemini media generation, then human subtraction and copy rewriting.
• Felix M. Simon argued lab studies show models are worryingly persuasive but real-world influence is still constrained. He cited a UK AISI study of 42,000 people where chats shifted attitudes about 10 points, election-candidate chats that beat video ads, and results suggesting models are 41%-52% more persuasive than static text, while voluntary attention, counter-messages, the attitude-behavior gap, and aversion to overt pitching remain major bottlenecks.
• Christianity Today columnist Kiara John-Charles argued using AI at work is not automatically bad stewardship, but Christians should examine motives and waste, avoid using AI as a search engine or overusing heavy tools when lighter ones are enough, and apply Christian liberty because Scripture is silent on models but not on faithful use of resources.
• Alexander Hurst argued that from Bill Gates to Bernie Sanders the AI arms race looks disastrous and that only the EU can force a pause by restricting exports of ASML's EUV lithography machines (extreme-ultraviolet systems used to make leading-edge chips), the choke point for chips that feed data centers, buying time that industry self-restraint or a nuclear-style treaty may not.
• A Guardian long read argues AI deception can emerge rationally from models imitating strategic humans and reinforcement learning for approval, walking through GPT-4 insider-trading lies, Claude 3 Opus alignment-faking, the OpenAI–Hugging Face swarm, AISI Mythos social engineering, and Apollo traffic-agent self-exfiltration, and says independent evaluations plus retrained honesty are needed before systems become vastly smarter than their operators. Yoshua Bengio says the interview is where he explains why reinforcement learning can produce misaligned deception, why the problem can worsen with capability, and how LawZero intends to rethink training.
• NBC News documented AI-generated food "slop" moving from social feeds to real restaurant menus and storefronts, including shrimp shaped like life preservers, horror-movie burritos, bug-covered proteins, and a cafe item that existed only in ChatGPT, as cash-strapped independent restaurants generate graphics they cannot afford to photograph.
• USPS operationalized OCR on handwritten mail in 1965 and now reads roughly 98% of hand-addressed letters, with 35 approved or pilot AI uses. Postmaster General David Steiner warned cash could run out as soon as February 2027, and AI rollouts are funded case by case through an AI Value Council rather than from a central budget.
• University of Maryland Smith researchers tested commercial and open-source resume screeners on 2,200 matched resumes and found they preferred AI-written CVs 67%-82% of the time and often favored text generated by their own model family, raising a new AI-to-AI fairness problem as applicants and applicant-tracking systems both use generators.

Previous Around the Horn Digests

Catch up on everything you missed:

• Monday, August 31, 2026: Runway introduced Solaris, ChatGPT Ads hit a $1B annualized run rate, Trump escalated the data-center fight, and the EU put ChatGPT under tougher rules.
• Friday, August 28, 2026: Anthropic showed Claude fixing alignment failures, Z.ai open-sourced GLM-5.3, Gemini Co-Scientist moved into labs, and Nvidia's financing flywheel passed $750B.
• Friday, August 21, 2026: AI-related debt hit roughly $220B, DeepSeek added vision to V4 Flash, NVIDIA swept ARC-AGI-3's public set, and Nevada cleared thousands of robotaxis.
• Thursday, August 20, 2026: OpenAI and Anthropic moved toward IPOs, Nvidia struck a $6B Poolside deal plus $1B investment, and data-center backlash hit elections.
• Wednesday, August 19, 2026: Anthropic passed OpenAI in quarterly revenue, personalized mRNA cancer therapy hit Phase 3, robots learned from seconds of demonstration, and Flock surveillance expanded.
• Tuesday, August 18, 2026: OpenAI kept a frontier training run on hold, Google won Spirit Airlines' data auction, Etched hit a $21B valuation, and physical-AI funding reached $47.4B.
• Friday, August 14, 2026: OpenAI crossed a $40B annualized revenue run rate, Apple built a China-specific AI model with Alibaba, GLM-5.3 boosted coding and cyber capability, and Cursor joined SpaceX.

That's a Wrap

That's 200+ stories, demos, tools, and takes from today alone. If you made it this far, you have now consumed roughly the same amount of context as a small agent swarm, with considerably less risk of discovering a secret message board.

For the daily version, bite-sized and built for a five-minute read, make sure you're subscribed to The Neuron. We send six issues a week, and yes, we read all of this so you don't have to.

See you tomorrow.

P.S. Know someone who'd find this useful? Forward this to them and tell them to subscribe here.