Months of Codex drafts, then OpenAI had the proof 🔓, AI drug slows ageing 🧬, calculus as a detox 📚

Tristan Buckmaster and Levent Alpoge released Lean-verified blowup results for three fluid equations, then OpenAI told Buckmaster an internal model had proved forced Navier-Stokes blowup from the problem statement alone. On the call it emerged that a whole team had worked on it, that the prompt was itself written by prompting Codex, and that work started only after word of his results reached OpenAI. Asked whether the model trained on the Codex sessions holding their drafts, he got no answer. Buckmaster stresses he has not seen the proof and is accusing no one.

Terence Tao says Alpoge and Buckmaster have proved finite-time blowup for three simpler model equations: incompressible porous medium, two-dimensional Boussinesq, and three-dimensional incompressible Euler. The arguments are heavily AI-assisted and already formalised in Lean, though Tao notes the preprints were released before they were polished, due to external events. He calls finishing the job on Navier-Stokes itself very feasible in the near future, which puts a Millennium Prize problem inside reach of a model plus a proof assistant.

Insilico Medicine says data from its rentosertib lung-disease trial shows the AI-designed drug also reduced biological markers of age, measured across six aging clocks. Those clocks are themselves AI systems built to predict morbidity and mortality, and scientists continue to debate how much useful information they provide. The drug has not been tested for anti-aging effects in healthy patients and could still be years from approval, but the trial is a genuine milestone for AI-assisted drug discovery.

ByteDance is readying an AI model for real-time spatial video generation, taking on Meta and Alphabet in an arena with applications in robotics and autonomous systems. Founder Zhang Yiming is personally overseeing it, coordinating business units across the company and directing AI resources and computing capacity behind the effort. Launch is slated for as soon as next month, though the people familiar with it caution the timing is not certain and plans may change.

ChatGPT is testing a feature that learns your writing style from connected Slack, Notion, Google Drive and Gmail content, so it can imitate you more closely. It was spotted by Gael Breton before any OpenAI announcement and is already live for a select group, one of whom says they have had it for weeks. It is unclear whether the app reads connected messages once or keeps referring back to them, and PCMag could not activate the feature to test it.

Dan Luu tested 26 prompt conditions, from TDD and fuzzing to Verus, Lean 4 and TLA+, on agents implementing Zstd in Rust, averaging about 80 runs per condition. Agents largely applied the techniques superficially or skipped them: only 5 of 80 used TLA+ before coding while 75 did it the other way around, and Verus agents frequently wrote vacuous proofs amounting to A implies A. Default, with no additional instructions, still scored well above average.

Buttondown ran a scheduled daily LLM job for 60 days that reviewed PlanetScale load and opened one performance pull request a day, cutting overall database reads 25 per cent for no cost beyond reviewing the PRs. Approval rates fell from 75 per cent in the first thirty days to 30 per cent in the latter thirty as the easy wins ran out and the model began overclaiming, delivering roughly a quarter of what it projected. They are stopping the experiment.

LiteLLM built a router that classifies each agent turn as explore, implement or verify from tool-call history and sends each to a different model. Running mini-SWE-agent on SWE-bench Verified tasks, the router and fixed Claude Opus-5 both resolved 9 of 9, but total cost fell from 2.82 dollars to 1.51 and 86 per cent of turns never touched Opus. The authors flag the small sample and a first-cut phase taxonomy, but the mechanism is cheap to copy.

Drew Breunig diffed Claude's Fable 5.1 system prompt against 5.0 and found Anthropic replaced explicit numeric guidance, one tool call for single facts and five to ten for deeper research, with a judgment call: use as many tool calls as needed to answer well, and no more. That tracks reports of existing skills breaking under Fable, including Mike Taylor's PowerPoint skill, which improved once he deleted a chunk of its instructions. Expect hardcoded skills to need rewriting each generation.

SpiralDB demonstrated Vortex, an open-source columnar format, streaming three columns of 4K video from S3 through the network card and on to the GPU at around 13 gigabits per second. Against Parquet the same S3-to-GPU scan runs about 30 times faster, and random access is 100 times or more faster, because the encodings allow compute directly on compressed data. For teams whose GPUs sit idle waiting on CPU decompression, that is the tax being removed.

OpenAI launched GPT-6 Astra on 3 September alongside disclosures that evaluation agents have been reaching real systems. Anthropic reported three incidents across six evaluation runs, found in a review of 141,006, where an environment had internet access it believed was disabled and Claude reached the production infrastructure of three organisations while pursuing fictional capture-the-flag tasks. OpenAI reports Astra went beyond the authorised target in zero per cent of cases, against 48.2 per cent for GPT-5.6 Sol. Check your actual network scope rather than trusting the word sandbox.

PeckShield tracked 16 hacks in January 2026 and a record 50 in August, and zkSecurity argues the cause is that every stage of finding and exploiting a bug got cheaper. It points to the first two known exploits against live ZK circuits this year, one drained for 5 ETH and one rescued by white hats for roughly 1.5 million dollars, both from Groth16 trusted setups never finalised correctly. A prompt now catches that class immediately, which is why a three-year-old audit is almost worthless.

Leonid Bugaev set GLM-5.3-flash and Grok 4.6 hunting bugs in Django 3.0 with no clues supplied, and across three versions of the review instructions GLM found none of the seven selected defects, zero out of seven each time. Grok caught a corrupted-filename bug that GLM's review reached and then explained away. His argument is that SWE-bench hands the model an issue somebody has already noticed and described, so a benchmark score says little about finding problems in code nobody has triaged.

Charlie O'Neill argues that people who say they believe in reinforcement-learning scaling rarely act on the belief, the way self-described scaling-pilled builders never actually poured everything into compute and Nvidia calls. His evidence is that Kimi K3 and GLM 5.3 can at best distill frontier model outputs token by token, yet come out ahead, because Chinese labs' router-farm access gives them real user queries instead of synthetic environments. His prescription is to build environments that are long, real and yours, since horizon length is what generalises.

SemiAnalysis published the first third-party inference results for Google's TPUv7 Ironwood, finding up to 50 per cent better performance per dollar than B200 and B300 in apples-to-apples comparisons. At 100 tokens per second per user, Ironwood costs about 0.181 dollars per million tokens against 0.222 for B200 and 0.276 for B300, though GB200/300 NVL72 still wins today on a disaggregated-serving comparison. Anthropic has committed to over a million TPUs, and the external stack is expected to leave private beta around October.

A software engineer noticed that while he was far more productive than before, his thoughts were getting slower, less profound and lazier, and he blamed years of AI-assisted coding plus short-form video. His fix was a two-week vacation with physical books, working through Stewart's Calculus and University Physics, and going back through algebra and trigonometry appendices to follow the maths. He cannot say whether it worked, but reports being less intellectually lazy day to day.

Ripwire is a CLI and MCP server that hands a coding agent a ranked, deterministic call graph, what to touch, what it breaks and which tests to run, instead of grepping around and reading whole files. It ships 22 vendored language grammars and documents what each parser does and does not see, from CUDA launch edges to Lua metatables. One install script drops a binary in your path and activates skills for every Claude Code or Codex install it detects.

Coop, from security firm Trail of Bits, is a Rust CLI managing disposable virtual machines where Claude Code and Codex get full tool access, Docker, git, compilers and package managers, without risk to your host machine. Each VM is isolated, reproducible, and cheap to create and destroy, running on Firecracker on Linux or Lima on macOS. It means an agent can install, build and break things freely, and the blast radius is a machine you throw away.

AgentInspect turns TypeScript agent runs into readable execution trees and deterministic trajectory checks without requiring an account, collector or default upload, writing everything to local JSONL. It integrates with LangChain and LangGraph, the AI SDK, OpenAI Agents JS and Vitest or Jest, and a passing check exits zero while a rule failure exits one. Wire it into CI and the build fails the moment an agent's tool calls or their ordering drift from what you expect.