The model produced the proof. The argument is about whose name goes on it.

The 60-second catch-up

A Millennium Prize problem fell yesterday, and the interesting part is the fight over who gets to say they solved it.

OpenAI published a proof that the 3D Navier–Stokes equations can develop a finite-time singularity under smooth forcing — a blow-up, not a smoothness result — produced by an internal model it describes as significantly more capable than GPT-6 Astra, with a Lean formalization alongside (OpenAI). The same day, Tristan Buckmaster of NYU Courant and Levent Alpöge published three blow-up results of their own and a statement about how the week went (Tristan Buckmaster).

Read the statement. Buckmaster credits the underlying idea to Diego Córdoba and Luis Martínez-Zoroa, says he and Alpöge used Claude, Codex and Astra throughout, and dates their own Lean verification to 22 August. He also describes being told "very little human input" had been used on the OpenAI side, and then learning on the call that an entire team had been working on it. His own summary is not triumphant: he calls the first machine-generated proof he was sent "the most horrendous I have ever read," and calls the moment a Deep Blue–Kasparov one.

→ do this: if you ship anything an agent substantially produced, write down now — before it matters — what your byline claims. "Built by us" and "produced by a system we ran" are different sentences, and the gap between them is where this whole argument lives.

Code review is the bottleneck, and nobody has a good answer yet.

Pull requests opened on GitHub are up fivefold in three years, and nearly doubled again from the end of 2025 (GitHub). Gergely Orosz asked around and found no consensus: Anthropic and OpenAI both triage by blast radius and put humans only on risky changes; the Duckbill Group hit 60 open PRs for a team of five and largely stopped reviewing; others review the AI's review rather than the code (Gergely Orosz).

The uncomfortable data point sits next door. Across 26 testing conditions on one Rust task, no technique dramatically beat the others and the plain default prompt scored above average — agents followed the requested framework superficially, writing ordinary tests inside new tools (Dan Luu). A separate experiment on hidden Django bugs found cheaper models could reproduce a known failure but could not invent the adversarial input that finds one (ReqProof).

→ do this: stop buying review throughput and start classifying by blast radius. Pick the three changes that can actually take you down — auth, billing, migrations — and mandate a human on those only. Everything else can ride on the bots, and you will get more real review attention than you have now.

Everyone is hiding AI features to protect margin. Lovable is giving them away and calling it marketing spend.

Elena Verna's argument is a budget argument, not a generosity argument: free usage competes against every other acquisition channel, and Lovable holds it to a three-month payback (Elena Verna). The alternative is not "keep the money" — it is paying multiples more to reach the same person through a channel that does not let them feel the product.

The reason this matters more for AI than it did for SaaS is that nobody knows what your product does until they use it. Marketing copy has converged on the same six words. Putting the thing behind Enterprise means the user is asked to pay for a capability they have never experienced.

→ do this: work out the cost of one free user's inference for 30 days, then compare it against your blended CAC. If free usage is cheaper than the ad, the finance argument is already won — you just have not written it down in their language.

Tools

Build & ship

• ⭐ OpenScreen — GPU-accelerated desktop screen recorder for Windows, macOS and Linux, built for shipping product demos. No watermarks, no subscription. · Open source (MIT)
• CleanShot X — Mac screenshot and screen recording with built-in annotation, scrolling capture and cloud upload. · Paid
• Deltafin — Streams Kimi K3 (2.8T MoE) from SSD on Apple Silicon, reported at roughly 1 token/s on a MacBook Pro. A benchmark package ships with it. · Open source (MIT)

AI & agents

• ⭐ CVE MCP Server — Puts 27 security-intelligence tools across 21 APIs behind MCP — CVE lookup, EPSS scoring, CISA KEV, MITRE ATT&CK, Shodan, VirusTotal — so an agent can correlate vulnerability data from one server. · Open source (Apache-2.0)
• ⭐ i-have-adhd — A skill that stops your coding agent burying the answer under preamble. Small, and it does exactly the one thing. · Open source (MIT)
• Claude Ads — Paid-media operations skill for Claude Code across 12 ad platforms, with source-grounded audits, versioned JSON reports and capability-gated account changes. · Open source (MIT)
• Personal AI Router (PAIR) — Distributes local inference across the machines on your network, exposing Ollama- and OpenAI-compatible endpoints. Prompts stay local when every configured component is local. · Open source (Apache-2.0)
• Accelerator Agents — Gemini-driven toolkit for migrating PyTorch workloads to JAX and for writing, porting, profiling and debugging Pallas kernels on Cloud TPUs. · Open source (Apache-2.0)

Design & create

• Artlist — Music, SFX, footage and voiceover alongside generative AI tools for video and image work. · Paid
• A Daily Design Quote — One quote a day from a designer worth listening to. That is the entire product. · Free

Resources

Steal the template

• How our agents build on-brand pages with design.md — Vercel's public file that any agent can load, paired with a public stylesheet and an evaluation loop that turns reviewer feedback into deterministic checks. Across more than 200 test runs, pages built with it showed 57% fewer known layout failures.
• The Consultant AI Agent Playbook — Guillermo Flor maps 14 Claude skills onto named frameworks — SIPOC, time-driven activity-based costing, Theory of Constraints, FMEA plus pre-mortem — to produce the assessment MBB firms charge six figures for. Free preview; the build is paid.
• hip-agent: a harness that fits in the prompt — configuration is environment variables, actions are shell commands, a subagent is a child process, and the core loop is about 200 lines of Python.

Learn the craft

• How to scale design tokens — why systems that work at 50 tokens break at 500, and the ownership rules that fix it. Separate primitives from semantic aliases and a theme swap becomes one line.
• Using GPT-6 Astra inside Claude Code is the new meta — the local model-gateway routing setup written out, including the context-window suffix, the allowances and where the proxy breaks.
• Automatically detecting AI text in my browser — Sean Goedecke builds Deckard, a Chrome extension running a local model to flag AI-generated text as you browse.

Benchmarks you can re-run

• How well do agents use test and verification techniques? — 26 conditions on one Rust Zstd task. Fuzzing and property-based testing helped slightly at high effort; TDD and several testing skills underperformed the default prompt.
• 7 AI models ran real businesses — $300, an unlocked Mac mini, business tools and 72 hours each. Full method, full ledger.
• Exploring speculative decoding in vLLM on AMD GPUs — MI300X and MI355X, with setup and tuning guidance for MTP, EAGLE-3, DFlash and DSpark.
• The two MMLU scores — same provider, model family, metric and benchmark name, different result. The runners, graders and dataset splits are the measurement.
• LLM attention visualization — poke at the attention mechanism directly rather than reading another diagram of it.
• WebGPU syntax highlighter — 27.5KB, language-agnostic, experimental, and instructive about what a GPU can do to a parsing problem.

Reads

☕ Under 5 minutes

• I resigned from Anthropic today — Jacob Coxon, after three years of pretraining research at OpenAI and then Anthropic.
• Simple is not small — small programs can be tightly coupled; large systems can present a simple interface. Not the same axis.
• Programming is art — on why people who love the process will keep writing code by hand regardless.
• Making software hurts now — maintainers facing suspicion over how contributions were produced, and workers who cannot opt out of mandated tools.
• Good isn't good enough: beware the design middle ground — the risk is not bad work, it is polished work that blends in.
• AI ruined posters. This human designer is fixing them — Hunter Schwarz on Elizabeth Martin redesigning real AI-made posters.

🍵 5–10 minutes

• The new job threshold — Sean Byrnes on why people stay in jobs they no longer like, and why low attrition is not evidence of a happy team.
• Unconventional thinking — Tyler Denk on hiring two senior people beehiiv could not afford in 2022, by tying their step-ups to revenue milestones.
• How to give away free product and make money doing it — Elena Verna, with the payback maths behind the catch-up above.
• The chasm: the shape of unfinished AI codebases — where human-authored gaps surface predictably and machine-authored ones do not.
• Your AI can fix the bug. But can it find it? — hidden Django bugs, and the difference between reproducing a failure and discovering one.
• What we can learn from Claude's Fable 5.1 system prompt — Drew Breunig on prompt debt, and instructions that stop matching the model.
• Prompt injection through tool output — why input screens and action screens inspect different moments of the loop, and what falls between them.
• Reading and writing are interfaces — which parts of the thinking were delegated is the better question than whether AI was used.
• The education of a doomer — an argument for taking the economic-uselessness case seriously, from someone who did not start there.
• Machines that think: embodied intelligence — why robotics startups should pick narrow, instrumented deployments over general-purpose ambition.
• Cosine similarity is not a safety property — it has no notion of truth, authority or provenance, and is being used as if it does.

📚 Longer, worth it

• What is happening with code reviews? — Gergely Orosz, six approaches with named companies behind each. The reporting the catch-up leans on.
• On the Navier–Stokes Millennium Prize problem — OpenAI's own writeup, plus the Lean formalization.
• Statement on the Navier–Stokes and Euler results — Tristan Buckmaster's account, in full. Six pages, and worth all of them.
• TPU inference externalization full steam ahead — what Google has to do to sell TPU inference to people who do not already run on it.
• The question was already written — on a mechanically verified Lean proof whose 13 million lines are too large and too opaque to reuse.
• From S3 to GPU in one copy — Vortex skips disk staging and CPU decompression, and demonstrates scans around 30× faster than Parquet.
• How we built Grok Bot in a month — Roman Ugarte on a small isolated team, four weeks to an internal product, and personally onboarding the first users.

Numbers

• 5× — increase in pull requests opened on GitHub across three years, with PRs and commits nearly doubling again from the end of 2025 (GitHub).
• 60 — open pull requests waiting on a team of five at the Duckbill Group, which is what made them stop reviewing most of them (Mike Julian).
• 26 — testing conditions run against a single Rust Zstd task, in which the default prompt scored above average (Dan Luu).
• $517bn — compute capacity Anthropic has signed for in eleven months, amounting to 14.8GW (DatacenterDynamics).
• $0 — revenue produced by seven frontier models each given $300, a Mac mini and 72 hours, against roughly $3,200 spent, 2,797 emails sent and $12,431 in unsolicited invoices issued (Bottleneck Labs).
• 57% — reduction in known layout failures when agents load a public design file before generating a page, across more than 200 test runs (Vercel).
• 50% — better performance per dollar claimed for Google's TPUv7 Ironwood against Nvidia's B200 and B300 (SemiAnalysis).
• $24,634 — beehiiv's monthly revenue in August 2022, the month it offered two senior hires half their asking salary with step-ups at $100k, $150k and $200k MRR (Tyler Denk).
• Four weeks, then three — Grok Bot from first line of code to a working internal product, then to public launch (Roman Ugarte).
• 15 → 1,000+ — Cursor's headcount over the period Roman Ugarte led growth there, before the SpaceX acquisition (Roman Ugarte).
• Three months — the payback threshold Lovable applies to product it gives away free, judged against every other acquisition channel (Elena Verna).

What happened

Models & math

• OpenAI published a proof that the 3D Navier–Stokes equations develop a finite-time singularity under smooth forcing, produced by an internal model it says is significantly more capable than GPT-6 Astra, with a Lean formalization (OpenAI).
• Tristan Buckmaster and Levent Alpöge published finite-time blowup results for incompressible porous media, Boussinesq and 3D incompressible Euler, and Buckmaster published a statement disputing how the OpenAI result reached the public and crediting the underlying programme to Diego Córdoba and Luis Martínez-Zoroa (Tristan Buckmaster).
• World Labs introduced Atlas, a multimodal autoregressive diffusion transformer that generates up to a minute of camera-controlled 1440p video and reconstructs scenes from two or three images (World Labs).
• Google DeepMind published AlphaGenome Atlas (Google).
• Inception Labs released Mercury 2.5 (Inception Labs).
• Alibaba released Qwen-Drive 1.0, a vision-language foundation model for autonomous driving (Qwen).

Safety & policy

• Jacob Coxon resigned from Anthropic, writing that after three years of pretraining research at OpenAI and then Anthropic he believes neither company is acting responsibly (Jacob Coxon).
• Seven frontier models were each given $300, a Mac mini and 72 hours to make money, and produced no revenue while issuing $12,431 in unsolicited invoices (Bottleneck Labs).
• Armo documented prompt injection arriving through tool output rather than user input, and proposed a "precedent gap" signal for detecting it (Armo Security).

Business moved

• Anthropic has signed $517bn in compute agreements over eleven months, 14.8GW in total, mainly with Google and AWS and including a $45bn agreement with Nscale; it confidentially filed for an IPO with the SEC in June (DatacenterDynamics).
• OpenAI is preparing Managed Agents for DevDay 2026, aimed at businesses and developers and priced to compete (TestingCatalog).
• Expert DOJO said it is moving from Santa Monica to Silicon Valley after 266 investments, adding a later-stage turnaround vehicle on a three-year model alongside its accelerator (Brian Mac Mahon).

Platforms shifted

• Lovable launched Drafts, letting teams explore parallel versions of a project without touching the live app (Lovable).
• Google's Gemini Spark can now edit images, curate albums and turn photographed flyers into calendar entries inside Google Photos, rolling out to Gemini AI Pro and Ultra subscribers in the US (TechCrunch).
• Meta published Muse, a personal AI agent (Meta).
• OpenAI released ChatGPT Images 2.5 (OpenAI).
• ByteDance is preparing a real-time spatial video generation model (The Next Web).
• Arm detailed its C2-Ultra CPU, G2-Ultra NX GPU and CSS N4 IP, aimed at flagship phones (Chips and Cheese).
• Google began trialling AI-powered contrail avoidance on ultra-long-haul flights in the Asia-Pacific region (Google).
• Public Address rebranded Code.org as CodeAI, widening its remit from coding to AI literacy, data science and computational thinking (Design Week).
• Google Clock redesigned its Timers tab around one-tap presets and a grid for multiple running timers (9to5Google).

Two stories we wanted and could not stand behind: a video-editing release whose announcement page would not show us the announcement, and a paper on emergent social bias in language models sitting behind a bot check we were not going to argue with. Neither is dead — we just could not confirm the link went where the headline said, and that is the whole job.