OpenAI Mathematicians: Model Reasoning Traces Look Strikingly Like Expert Humans'
a16z · youtube · 2026-09-08
a16z Infra Partner Lisha Li sits down with OpenAI mathematicians Mehtaab Sawhney and Mark Sellke for a deep dive on AI's rapid advances in mathematics.
- Key claim: What's surprising about recent results (sphere packing advances, construction of a non-sofic group) isn't that models search more or work longer than humans — the reasoning traces look remarkably like an expert's work: choosing promising approaches, backtracking on failure, combining ideas across the literature
- GPT-5 as the conversion moment: the moment that turned both practicing mathematicians into OpenAI believers
- Training insight: math papers are a bad training set because published proofs erase the real exploratory process
- The Astra 10-problem set: favorite problems and deep dives on model performance
- Bigger questions: harness vs. model — what actually matters; what "taste" in a model means; how the math community should adopt AI; empirical vs. theoretical math; and what happens when AI produces more mathematics than humans can absorb — models that accelerate discovery may also make sophisticated results easier to understand
Related event: OpenAI's Astra Advances Ten Open Math Problems(4 posts)→
More from AGI Musings
- Hyperscaler backlog hits $1.7T as Citi conference flags shift to agentic inference — sanjaykalra · 2026-09-08
- Hex Founder Ben Stancil on Why Analysts Never Get to 'Find Insights' — and Whether AI Changes That — bennstancil · 2026-09-08
- Mastercard's AI Chief: AI Chat Already Drives 15-20% of Retail Referral Traffic — AnneliesGamble · 2026-09-08
- Researcher on the prompt-credit fight: if everyone is close to a result, it's the wrong problem — samiramanabi · 2026-09-08
- Gary Marcus amplifies '100% misalignment speedrun' jab as another lab cooperation attempt falls apart — GaryMarcus · 2026-09-08
- OpenAI accused of monitoring prompts and pressuring mathematician off Millennium Prize Problem credit — anshulkundaje · 2026-09-08