Tristan Buckmaster’s guest essay on Terence Tao’s blog

Your AI and this story

Key facts

• Expands on Buckmaster’s Oct 19 talk at the Institute for Advanced Study; Tao notes the post was converted from another format using AI, and Buckmaster discloses LLM help with editing
• On cost: OpenAI’s Navier–Stokes proof ‘reportedly required the coordination of 10,000 agents over 88 hours, at an estimated cost in the millions’
• Charge: ‘AI companies, who strip mine our data, including unpublished work, private notes and research grant applications, have an egregious conflict of interest in competing against their own customers with unlimited resources and internal models unavailable to the public’
• On the Oct 6 release: manuscripts ‘mostly unreadable and in some cases false’; the average-three-hours claim was ‘essentially meaningless’ because failed attempts and earlier agent work were not counted; no due diligence on building on customers’ unreleased work, ‘some of whom were graduate students and postdocs who were literally days away from releasing their own work’
• Says OpenAI’s Millennium-problem push was, by OpenAI’s own account, spurred by rumours that a rival had solved two of them; after Navier–Stokes, teased further Millennium problems ‘yet they achieved none’
• Proposals: collective enterprise agreements with zero-data-retention and audit clauses for all academics; opt-in (not opt-out) training on user data; ban on absorbing a whole session when a user rates one reply; independently funded open-weight infrastructure; credit for exposition, shared tools and detailed AI disclosures
• ‘Chess did not end in 1997. Mathematics will not end in 2026.’