The Washington Post reported on 2026-08-19 on an event amplified across Chinese media: OpenAI convened a closed-door summit with about 40 top mathematicians, organized around a single question — "When AI can do mathematical research, what is left for human mathematicians?"
The meeting was moderated by OpenAI researcher Bubeck. Attendees included Fields Medalist and UC Berkeley assistant professor Jacob Tsimerman, Harvard mathematician Melanie Matchett Wood, and German mathematician Andreas Thom. No consensus was reached, but the meeting triggered a 'value crisis' within the mathematics community.
1. Four Possible Futures: Bubeck's Open Questions
Bubeck described four possible futures, without giving an answer, each pointing to a very different future for the discipline:
1. Mathematics becomes like software engineering: AI helps hundreds of people collaboratively attack the same problem. 2. Mathematics becomes like physics: compute and AI models replace particle accelerators, 'computing' new theorems directly. 3. Mathematics becomes museum curation: AI produces results; humans select and interpret them. 4. Mathematicians collectively transition into AI safety work.
Bubeck himself offered a clear-eyed remark: "We must put people, put mathematicians, first. Mathematics only has meaning when mathematicians learn something from it." Yet he also admitted that the 40 participants did not reach consensus.
2. AI Can Prove Theorems, But "Can't Explain How It Proved Them"
Harvard mathematician Melanie Matchett Wood wrote a human-readable version of a proof for OpenAI's unit-distance conjecture results. Her observation was sharp: "Top AI models share a common flaw in explanation — they say a lot about the easy parts and gloss over the hard parts."
She further noted: "Top AI models still cannot identify the genuinely difficult parts of an argument and explain them clearly." The force of this critique lies in mathematics' core value: understanding a result, not just proving it. A proof that is not understood does not become part of the literature.
Andreas Thom offered a more precise distinction. Among OpenAI's 10 announced results was a new construction about non-sofic groups, filling a gap between two papers by Thom and collaborator Gábor Kun from 2016 and 2019. The two subsequently published a follow-up paper simplifying and extending the AI's finding. Thom said: "AI is solving problems in a very clever, substantial way. But the new concepts only emerged after humans participated in the proof process — that is not something AI has achieved on its own yet."
3. Tsimerman Chooses AI Safety: A Symbolic Moment
Tsimerman, a 2022 Fields Medalist, has not formally joined OpenAI, but his choice to pursue AI safety research is itself "a microcosm of the mathematician's situation": as AI proves more theorems, mathematicians begin to doubt the irreplaceability of their own work, and some choose to invest their energy in the more urgent front of ensuring AI is not misused.
4. Resonance with the Week's Math × AI Mainline
Placing this event into the late-August math-AI timeline:
- 8/23 Terence Tao "digests" AI proofs + Palomar registry: a 90,000-line Lean proof of the Sentorf problem compressed to 15,000 lines (previously covered)
- 8/23 OpenAI Astra completed 10 open mathematics problems in Lean 4 on 8/1, then safety-locked on 8/7 due to 'critical-level' findings (previously covered)
- 8/22 Axiom Math (founded by a 25-year-old from Guangzhou) completed Lean 4 formal verification of the 246 prime gap theorem (previously covered)
- 8/23 Talagrand's convexity conjecture: a 31-year-old open problem solved via human-AI collaboration (previously covered)
- 8/24 This closed-door summit of 40 mathematicians (this post)
- When AI can mass-produce theorem proofs, can "number of theorems proved" remain an academic evaluation metric?
- When AI can run counterexample searches in hours, can "discovering a counterexample" still count as an original contribution?
- When AI can quickly generate formalized proofs, is "formalization skill" still a core competency for mathematicians?
- Do not focus only on "AI accelerating proofs"; also consider "what value remains for mathematicians."
- Do not focus only on tools; also on institutions — how evaluation systems like the Fields Medal, ICM invited lectures, and mathematics journals respond to the AI-era "value crisis."
- The Chinese mathematics community's strength lies in "many people, many problems," but its weakness is relatively few original concepts. In the AI era this is precisely a dangerous combination: AI will rapidly erode the "problem volume" advantage, but is far less able to replace the central role of original concepts.
The mathematics community's attitude toward AI is shifting from an "instrumental" view to a "value" view: no longer only asking "what theorems can AI help me prove," but asking "when AI takes over theorem-proving, what is left for mathematicians?" The landmark significance of OpenAI's summit is that this question was, for the first time, discussed jointly by a leading model company and top scholars in mathematics.
5. Implications for Mathematical Education and Research Evaluation
The deeper metaphor concerns the impact on education and research evaluation systems:
These questions have no answers yet, but they are now formally on the table.
6. Lessons for the Chinese Mathematics Community
The Chinese mathematics community has in recent years invested in AI4S (at the 8/24 China AI for Science conference, multiple academicians cited "verification as the core bottleneck") and math-AI tooling (Tao's Palomar relay station, Lean 4 formalization). The summit's implicit lessons:
Sources
Washington Post 2026-08-19 "Mathematicians ask what's left for humans when AI can do math research"; 36Kr 2026-08-24 "数学的终结：40 位顶尖数学家齐聚 OpenAI 秘密会议"; New Intelligence (新智元) original report; Bubeck's personal homepage; Melanie Matchett Wood's personal homepage; original papers by Andreas Thom and Gábor Kun.