Good morning. The Navier-Stokes drama is the story of the day — OpenAI claims to have cracked a Millennium Prize problem, but an NYU mathematician says the company used its compute advantage to sprint past work he and an Anthropic researcher were already deep into. Meanwhile Mistral closed its €3B round, Cognition raised at a $48B valuation, and a pretraining researcher resigned from Anthropic with a public warning.
OpenAI’s Navier-Stokes claim is unraveling in public. OpenAI announced it had solved Navier-Stokes existence and smoothness using an internal model running up to 10,000 concurrent agents over 50+ hours, at a reported compute cost of $22.5 million. NYU’s Tristan Buckmaster promptly alleged OpenAI fought dirty, saying he and Anthropic’s Levent Alpöge had been working the same problem for nearly a year using Codex and Claude, published a related result the day before OpenAI’s announcement, and never got a straight answer about whether their Codex sessions ended up in training data. Wired reports OpenAI also floated a credit arrangement that would have excluded Alpöge’s name entirely.
The bigger question is who gets to do math now. MIT Technology Review frames the episode as a preview of what happens when frontier labs with exclusive access to unreleased models can outspend and outrun academic groups on open problems. An OpenAI staffer conceded the team was “inspired” by rumors of Buckmaster and Alpöge’s approach — The Verge has the timeline — which is not the same as an independent proof. Whatever the Clay Institute eventually decides about the math, the norms around credit and priority are the actual thing being rewritten.
Mistral’s €3B round lands, as we noted yesterday. The Samsung-led Series D is now official at a €21B+ valuation, with Macron endorsing it as part of a Franco-Korean “third way in AI”. HN’s mood was less generous than the press release: commenters pointed out Mistral’s ~€700M run-rate is roughly three days of Anthropic revenue, that job listings advertise €90k base salaries in Paris, and that recent benchmarks put Mistral Medium 3.5 behind considerably smaller open models. The counterargument is that European procurement rules may end up doing the differentiation work that benchmarks won’t.
Cognition raises $2B at a $48B valuation. Andreessen Horowitz and Accel led the Series E for the Devin maker, with the company reporting run-rate revenue jumping from $492M to nearly $900M since May and enterprise logos including NVIDIA, GE Aerospace, Citi, and Mercedes-Benz. HN was skeptical of the multiple given Cursor’s dominance among individual developers, though several commenters who had written Devin off earlier grudgingly admitted the product got better. The enterprise revenue is real; whether $48B worth of it materializes is another question.
An Anthropic pretraining researcher resigns, publicly. Jacob Coxon posted on X that after three years at OpenAI and Anthropic he’s out, arguing both labs are “racing straight to self-improving superintelligence” and that Anthropic understands the risks but presses on anyway under the “someone worse will do it first” logic. HN split hard: some applauded him for acting on stated beliefs, others called the claim that AI is more dangerous than nuclear weapons hyperbolic. The most interesting comments argued the risk isn’t the model in isolation but the stack around it — long-running autonomy, tool use, credentials, parallel agents — which is roughly what OpenAI just demonstrated on Navier-Stokes.
Meta launches Muse, a personal AI agent. Muse is a US-only rollout of an agent that browses and acts on your behalf, with layered prompt-injection defenses that Meta’s David Singleton detailed on X. Reuters reports it shipped despite internal concerns about how it handles sensitive personal data. HN reaction was roughly “no thanks” on trust grounds, though several commenters praised the inline browser handoff UX — you can watch the agent work and take control mid-session — as genuinely well executed.
DeepMind releases AlphaGenome Atlas. Google published a 1-petabyte database of predicted regulatory effects for all 9 billion possible single-nucleotide variants in the human genome, with an accessibility score meant to let non-programmers prioritize variants. Early wins include a rare-disease case solved at the Broad Institute and 22% more non-coding associations found in UK Biobank data. The sharpest HN critique: this may largely be a friendlier interface on precomputed AlphaGenome values that were already API-accessible, and one commenter argued AlphaGenome itself offers “essentially zero improvements” over Borzoi.
That’s a lot of money and a lot of drama for a Tuesday. The Navier-Stokes fight is the one worth watching — whatever OpenAI says next about training data and timelines will set the tone for every academic collaboration with a frontier lab from here on.