OpenAI's unreleased frontier model produced 722 math results, 81% of the past 3 years' major discoveries
Exponential View (Azeem Azhar) · rss · 2026-10-07
Exponential View's Azeem Azhar discusses a batch of mathematical results released by OpenAI yesterday, produced by an unreleased frontier model: 722 manuscripts in 372 families spanning number theory, complexity theory and mathematical physics, with each result averaging the equivalent of three hours of ChatGPT Pro thinking. Scientist Derya Unutmaz notes this amounts to 81% of the major math discoveries of the past three years.
- Some results have been verified in Lean, but not all; the author argues that even halving the real number is staggering — problems that stumped top human minds for decades fell to an afternoon of compute.
- His framing: imagine millions of superhuman research agents at work. Formal systems may verify proofs, but humans will lack the context to understand the underlying web of machine-invented concepts.
- Mathematics could split into two layers: machine mathematics (vast, verified, mostly consumed by AIs) and human mathematics (a compressed "effective theory" of the machine frontier — the small subset we can understand).
- He calls it a possible turning point in how humans make sense of the world, after three centuries of natural rather than supernatural explanation since the Enlightenment.
More from AGI Musings
- Design panel: what early-stage founders should build while models change weekly — pxd · 2026-10-07
- Consumer AI Winners Exist Only in Search, Productivity and Creative Tools—Everything Else Is Open — omooretweets · 2026-10-07
- 'Computer' was a job that vanished — AI will make 2000s computer use feel equally alien — sterlingcrispin · 2026-10-07
- GMU NLP simulates 10k LLM agents with memory and social networks moving through San Francisco for a week — ZiyuYao · 2026-10-07
- OpenAI nukes your research, you rewrite your project in Rust: 'a jobs program for ourselves' — almmaasoglu · 2026-10-07
- Dario Amodei's 2016 AI safety paper resurfaces: thinking about safety pre-Transformer — bookwormengr · 2026-10-07