Zvi says the Navier-Stokes solve is not the story — the $600-a-day-per-researcher inference bill is
In a post published roughly 21 hours ago, Zvi Mowshowitz argues the citable fact from the Millennium Prize result is the internal usage data it exposed: by mid-August the median OpenAI researcher was running agents daily at more than $600 per day of inference at API prices, at a rate of 3.1 agent-workdays per human workday, and the model that produced the result was eight days into a training run that had already surpassed Astra across the board. He puts the proof's compute at 130-300 billion output tokens, about $22 million at customer pricing and $600k-$2M at internal marginal cost, and notes prediction markets had given under 40% odds of any Millennium Prize falling to AI before 2030. His read is that both OpenAI and Anthropic are now inside a recursive-self-improvement loop that safety work cannot pace.
↳ Follow the thread