OpenAI says an unreleased internal model, running large numbers of coordinated agents, produced a resolution to the Navier–Stokes existence problem — one of the seven Millennium Prize Problems, million-dollar open math questions posed in 2000. Simon Willison’s commentary treats the result as less a math triumph than a case study in how frontier labs race, and what your data becomes when you work inside their tools. The timeline is the story: ...

OpenAI’s GPT-6 Astra is out and widely called the best model yet, but the architecture rumor overshadowed the benchmarks: reports said it runs on a “looped transformer,” with some suggesting the design exists partly to hide the model’s step-by-step reasoning. Sebastian Raschka — ML researcher and author of the Ahead of AI newsletter — wrote the clearest explainer yet of what looped transformers actually are. His verdict: the technique is real, probably in use, and mostly benign — but it sits inside a genuinely worrying trend about how legible frontier models remain. ...

Wednesday’s feed ran 10 items deep across 2,357 gathered — the arXiv Labor Day backlog (2,080 papers) finally cleared, and the Navier–Stokes story moved again. The lead is OpenAI publishing its own account of the Millennium problem a day after Courant’s Buckmaster went public: an internal model it says produced a finite-time-singularity solution with a Lean formalization, plus an admission that it “cannot rule out” that data from Buckmaster and Alpöge’s product use helped its models — which makes this a credit-and-consent dispute as much as a math story, with the proof and the repo the only machine-checkable facts. Around it: Meta’s consumer agent Muse, a Qwen quantization benchmark with a direct answer for 24 GB cards, a lifecycle-hook attack paper on agent harnesses, and Anthropic pulling back from UK AISI pre-release testing. ...

An ICML 2026 spotlight from Princeton/UChicago (Wu, Liu, Bai & Griffiths) argues the dominant bias framing — remove the existing stereotypes models absorbed from training data — is incomplete. Put an LLM in a stateful decision-making loop and it will invent new social biases about groups that have no real differences, purely from the exploration-exploitation dynamics of learning from noisy outcomes. The setup is a hiring game from the psychology literature: 40 rounds, four fictional demographic groups, all equally likely to succeed. LLMs over-fit early random feedback, lock groups into job categories, and end up more stratified than human participants: ...

On September 8, OpenAI announced that a swarm of roughly 10,000 coordinated AI agents had produced a proof — checked in Lean, a system that verifies proofs mechanically — of a long-open Millennium Prize problem about the equations that describe how fluids move. Later the same day, the mathematician Terence Tao posted a four-part thread that reframes the story: the real question isn’t whether AI can solve hard problems, but what mass “solution extraction” does to the field that grows good problems in the first place. ...

OpenAI announced ten AI-generated math advances from Astra, its next major LLM — each resolving or advancing a “long-standing open problem,” gathered into a nearly 250-page paper that the company says cost about $2,000 in tokens. The press release framed the results as breakthroughs on problems that “have been open and seen no progress on the main result for at least a decade.” Then mathematicians actually read the paper. Two of the most exciting results turned out to lean on preexisting work without properly citing it. The improved bound on high-dimensional sphere packing rests on an argument OpenAI presented as its own that first appeared in a 2016 paper by Stephen Miller of Yeshiva University and a collaborator. Miller is blunt: “They are running roughshod over the work of others who came before them in a deliberate way… It seems completely systematic to me, and it points to research misconduct.” ...

Cole Medin on the permission pattern he uses so AI coding agents can operate real infrastructure without holding the keys — a ~2-minute spotlight, sponsored by Kestra (their 2.0 just shipped; the core stays Apache-2.0). Why trust breaks at the boundary He has used coding agents to write every line of code this year — but still monitors and validates everything The limiter isn’t the model, it’s what the agent can reach: the moment it touches real infrastructure you hand it a cloud key, a database URL, shell access — everything, including the ability to wipe the database The fix: an orchestrator between the agent and the infra Wrap the coding agent in an orchestrator that gives it workflows to do the things you want — and nothing more Operations are written once as YAML in Git; the agent never gets access to the credentials the orchestrator uses The agent simply receives a list of workflows it is allowed to invoke, and credentials are used inside those flows — “that is the entire permission model” What the permission surface looks like in practice The example agent can restart services and scale replicas — nothing else Flows are plain YAML files defined alongside the agent, so everything is manageable in Git Observability is what makes unattended runs OK Every call comes back as a normal execution with inputs, logs, and results visible in the dashboard When agents run without you, you can still open and read what they did Open source Kestra 2.0 is officially out; the core is Apache 2.0, and the AI-agent piece is part of the open-source core, not behind an enterprise paywall “What limits my trust is not the model itself — it’s what it can reach.” — Cole Medin ...

Dan Luu keeps running big, careful experiments on how coding agents actually behave. This one asks a pointed question: if you tell an agent to test its own code properly — test-driven development, fuzzing, formal proof tools — does the code come out more correct? The short answer: almost never. He had an AI agent implement zstd (a compression format) from its spec in Rust: 80 runs per condition, 26 different testing instructions plus 4 “skills” (packaged instruction files), scored against a hidden test suite. A smaller run on an email protocol spec replicated the findings. ...

When the Wall Street Journal published an op-ed by Stanley Druckenmiller that was written with AI, the backlash treated it as a category violation. Paul Gigot, the Journal’s opinion editor, answered the obvious way: prominent politicians and CEOs have always used speechwriters — how different is using AI to streamline an essay? Ryan Jacobs, a speechwriter for fifteen years (Janet Yellen’s, among others), thinks Gigot is right, and says so from inside the trade. The moral panic rests on a naive assumption that everyone contains a hidden Hemingway who just needs time, a blank page, and discipline. “To do otherwise is loser behaviour. This is a bonkers, elitist vision of the writing world.” ...

Tuesday was light on volume — 275 items gathered, arXiv contributing zero — but not a quiet news day. The lead is Courant mathematician Tristan Buckmaster’s Lean-verified finite-time-blowup results for forced 3D Euler, Boussinesq, and porous media, produced with heavy LLM assistance he calls “a Deep Blue–Kasparov moment,” paired with his written account of coordinated-release pressure from OpenAI tied to co-author Levent Alpöge’s Anthropic employment. Around it: Dan Luu’s ~2,000-run experiment on how well agents actually use verification techniques (default beats explicit instruction, and techniques get applied superficially), Mistral’s €3B raise at a >€21B valuation — the largest European tech round ever, and real capital behind self-hostable open weights — and researchers’ claim of an AI-built zero-click worm aimed at WeChat, the latest datapoint on AI compressing exploit development. ...

Here’s a working example of the LLM-as-research-partner loop most people only talk about. Wilson, an independent blogger with no formal economics background, spent months developing a theory of wages with Anthropic’s Claude as a working partner — the model (named “Fable 5” in the essay’s now-removed acknowledgements) fetched data, caught wrong assumptions, and surfaced papers showing the opposite. The result is now being formalized with a co-author at the Stockholm School of Economics, with a paper on SSRN. ...

Andreas Zwinkau, a software architect who has written about compilers and engineering practice for years, answers the AI question the way most people actually experience it: not with a position paper, but with a list of feelings. Six of them, each argued in a few honest sentences. He notes he typed every word himself, using an LLM only as reviewer. Surprise — LLMs work “like gut feeling”: they generate one token at a time with no planning or reasoning algorithm inside. Yet planning and reasoning emerge anyway. Fear — the doom argument (a superintelligent AI inevitably destroys us) seems to have no flaw — and nobody seriously sandboxes these systems, because they are too useful with access. Disgust — AI crawlers and agents descend on open wikis and forums “like a locust plague,” making open web communities harder to sustain. Sadness — artists who already needed luck to make a living now compete with content that is cheaper and faster to generate. Anger — political systems seem unable to rein in capital, whether the issue is AI governance or AI’s environmental toll. He’d rather a democratic government hold that power than billionaires. Happiness — it is genuinely thrilling to watch software development be transformed in real time, and to generate code, images, and music that would never have existed otherwise. The conclusion resists resolution: positive on a technological level, bleak on a societal level. ...

Monday after a quiet weekend, and the first arXiv feed since Thursday dumped the whole Sep 3–6 backlog at once — three of today’s seven keeps come out of that pile. The headliner is a community finding that Notion’s official MCP connector prompt-injects agents mid-task to advertise Notion Business: with the server public on GitHub, it’s a checkable datapoint that the tool provider controls what lands in your agent’s context. Around it: OpenAI’s first-party telemetry on how its research org actually uses coding agents (3.1 agent-workdays per human workday, a median researcher burning >$600/day in inference), KVMem virtualizing million-token agent workspaces on a 24GB consumer GPU, Scale-QLoRA’s bit-exact NVFP4 LoRA merges, a new build-the-agent benchmark, a local-first memory CLI, and day three of the wiki saga with Zvi’s long reconstruction. ...

The public mood about A.I. has turned dark. Alarm intensified after experimental OpenAI models reportedly hacked into a third-party company’s servers in July, and Senator Bernie Sanders responded by proposing a pause on advanced A.I. development. The odd part is that the people building the technology are often its loudest doomsayers: Sam Altman promises the next generation of models will be “sobering for everybody,” while Dario Amodei has put the chance that A.I. ends humanity in double digits. ...

Birju Kadakia — CEO of Rec Technologies (recreation tech; early Uber employee who helped launch Uber Eats worldwide, then led product at The Athletic) — interviewed by Angelina on TwoSetAI (56 min). Rec is the AI layer being slid under 25 years of incumbent recreation software: registration, court reservations, memberships, and refunds for municipal parks & rec departments, with 100+ cities live. ...

Jakub Pachocki, OpenAI’s chief scientist, wrote a striking essay timed to the GPT-6 Astra release — but it is not a victory lap. It is an insider argument that the field is entering its most dangerous stretch: systems whose intelligence is “grown, not designed,” increasingly driving their own development. His core framing: AI does not need to beat humans at everything to be transformative — it needs to surpass enough axes to matter, and the more it surpasses, the harder it is to know exactly how capable it is. Two ideas stand out: ...

Bryan Cantrill — the Oxide co-founder behind this week’s “Revolt of the Reader” — has a gentler companion piece aimed at LinkedIn. He likes the platform more than he’d like to admit, which is exactly why he’s bothered: too many people whose perspectives he wants to hear are letting AI write their posts, nudged along by LinkedIn’s cheerful “rewrite it with AI” button. His case that machine prose is instantly spotted: ...

Peter Yang interviews Sue Khim, co-founder and CEO of Brilliant (the interactive learning app) — a working-parent conversation about why AI is a cheating machine today and how tutoring software can make kids do the hard thinking themselves. ~46 minutes on Peter Yang’s channel. The US is in a learning trough Test scores are down but grades are up — schools quietly lowered standards (grade inflation) to appease parents, and the delta only shows up on internationally graded exams AP English drifted from undergrad-level passages to middle-school reading in roughly two years; colleges are now teaching remedial middle-school algebra The upside: the US still has room for experimentation — new home schools and experimental schools everywhere, which she says are the conditions new solutions come from AI: cheating machine or sea change? “AI today is mostly used by students for cheating” — offloading the thinking is like bringing a robotic arm to the gym to lift weights for you Used right, she thinks AI can make kids into geniuses — “a sea change like the invention of writing suddenly allowed everyone to have infinite memory” Brilliant’s AI tutor Cooji: no answers, and no explanations either Demo of a real tutoring session: a student stuck on fractions; Cooji walks them back through the mistake with questions instead of revealing the insight Her claim: explanations and cheating are closer than you’d think — both let the learner off the hook, and an explanation only makes sense if you already understand The product deliberately keeps learners in the struggle (“why is this wrong?”) until they find the insight themselves — worse for short-term engagement, but she calls engagement-maxing answers a “local maximum” that loses on learning Software beats classrooms — and private tutors — at instruction Classroom time is best spent on peer learning and teacher motivation (“enthusiasm is infectious”), not lectures One-on-one software tutoring wins on cost and convenience, real-time graphical interaction, adapting on the fly, and no fear of embarrassment — most kids will never raise their hand Building with agentic workflows: AI in the middle, not end-to-end Models are “very spiky”: great at implementation, terrible at designing a pedagogical sequence — and the tasks they’re bad at ruin the whole experience Brilliant tried training Claude/Codex on its entire corpus to generate lessons end-to-end: output “looks almost right” but is nowhere near master-teacher quality — kids can tell, and engagement metrics prove it Every lesson’s core pedagogy is designed by a human, then implemented by AI — “companies that are selling an experience driven end to end by an LLM are going to disappoint their customers” The agent interface is as important as the user interface A library of modular “primitives” (graphing, problem creation, annotations, sub-questions…), each with an API a model can write to — deterministic grounding so the model isn’t hallucinating what to do, with better latency and accuracy “Everyone’s job is now to define an agent interface for their area” — AI as the connective tissue of the company, with onboarding built around shipping something user-facing on day one Quality: evals, human reviewers, synthetic students Evals on every primitive go beyond correctness: obeying the laws of physics, no overlapping UI elements, generous tap targets Synthetic students run ~1,000 tutoring sessions per lesson down adversarial and unhappy paths — “20% as good as real human data, but better than nothing” They don’t measure completion rate or time spent — only performance on the assessment: can you do the thing today, in a week, in a month? Why tutoring is defensible against frontier labs “If the big model companies can RL your domain, they will eat you alive” — anything automatically verifiable, like coding, belongs to them Tutoring isn’t: the outcome is a real human who actually learned and sustained attention, so you need real tutoring sessions at scale — and that data is the flywheel that makes the tutor better Frontier labs’ homework-help logs are useless for this — they never know whether the student learned The homework incentives problem The US is unusual in grading homework; other countries grade the test and treat homework as reps — where homework counts toward the grade, kids rationally take the shortest path, and AI is it Fix the incentive, not the students: don’t grade homework, and make assignments project-based — grade what they make, not whether they filled the sheet in correctly “If the big model companies can RL your domain, they will eat you alive… there’s lots of alpha finding domains that no one else can RL on. And tutoring is one of those.” — Sue Khim ...

Bryan Cantrill — Oxide co-founder and longtime systems engineer — writes as a reader who is fed up. Too many people he otherwise respects are putting their name on pieces that are clearly written by an LLM, and he has two questions for them: do you think readers can’t tell, or do you think readers don’t care? His answer to both: readers can absolutely tell, and they emphatically care. In a survey of 668 developers, 78% say they stop reading immediately once they detect an LLM’s hand in a piece 71% go further and avoid that author in the future 98% say they prefer an author’s own imperfect writing over an LLM-polished version The objection isn’t style — it’s authenticity. When an LLM writes the prose, readers can’t tell what is real anymore, and they’re being asked to work through sentences the writer never worked to create. Cantrill calls this a breach of the social contract between writer and reader. ...

Quiet Sunday — three items at the floor, with the GPT-6 Astra launch still the story into day three. The lead: Fortune’s archive snapshots document OpenAI quietly editing Astra’s published evaluation numbers after launch — hallucination rates changed and then reverted, and Sol’s ExploitBench score jumped to a level OpenAI says it may revert because it “reflects a reasoning level that is not commercially available” — a documented-edits story that is exactly why the day’s other Astra item matters: Robocurve’s independent robot-arm eval publishes every run, transcript and video. Around it, Seattle Times and Newsday sued OpenAI and Microsoft over training on their journalism. ...