$ cat briefs/daily/2026-05-24.md
2026-05-24
May 24, 2026 (Sun)
1 paper · 1 acquisition · lint day
+1new page
[01]
Top Stories
1. OpenAI reasoning model disproves an 80-year-old Erdős conjecture — AI's first autonomous contribution to pure mathematics
- OpenAI's general-purpose reasoning model disproved a core conjecture from the planar unit distance problem posed by Paul Erdős in 1946. It discovered an infinite family of point configurations (n^(1+0.014) unit-distance pairs, verified by Princeton's Will Sawin). (source)
- This was achieved by a general reasoning model, not a math-specific system, using chain-of-thought alone. Human mathematicians extracted the ideas from the model's reasoning transcript and rewrote them into a formal proof.
- Why it matters: Until now, AI's math achievements were limited to solving known problems (IMO) or serving as an auxiliary tool. This is the first case of autonomously disproving a genuinely open problem — the strongest demonstration yet of OpenAI's "Year of Science." Together with DeepMind Aletheia (IMO gold medal), it shows that pure-math AI has entered the stage of being a serious research tool.
- → An OpenAI model has disproved a central conjecture in discrete geometry, Reasoning Models, OpenAI
[02]
Paper Picks
OpenAI — Erdős Planar Unit Distance Problem — openai.com/index/model-disproves-discrete-geometry-conjecture/
- TL;DR: A general reasoning model (chain-of-thought) applied algebraic number theory based on Golod-Shafarevich theory to discover an infinite family of counterexamples to Erdős's 1946 conjecture. δ=0.014 (verified by Princeton's Will Sawin). Verified by external mathematicians.
- Why read it: The strongest empirical evidence to date on whether AI can move beyond an "auxiliary" role in science/math to "autonomous contribution." It redefines the applicable scope of reasoning models.
- → An OpenAI model has disproved a central conjecture in discrete geometry
[03]
Watch
- Mistral acquires Emmi AI (2026-05-22): Acquired a physics AI startup (CFD, heat transfer, materials simulation), bringing on 30+ researchers. Terms undisclosed. Mistral expands beyond language models and coding agents into physics simulation AI. Strengthens its European industrial AI positioning — direct connections possible to its existing customer base including ASML, ESA, and Ericsson. → Mistral AI (source)
- Comparing the math-AI trend at OpenAI: DeepMind Aletheia (IMO gold medal, 18 open math problems) vs OpenAI reasoning model (Erdős disproof) — both approaches have entered the pure-math frontier. Worth watching how fast "AI in science" shifts from hype to demonstration in 2026.
[04]
New in Wiki
- An OpenAI model has disproved a central conjecture in discrete geometry (new — OpenAI's first autonomous disproof of an open problem in pure mathematics)
[05]
Updates
- OpenAI: Added the 2026-05-20 Erdős conjecture disproof to Recent Activity
- Mistral AI: Added the 2026-05-22 Emmi AI acquisition + Physics AI strategic position
- Reasoning Models: Added the Erdős disproof to State of the Art, updated Key Papers
[06]
Lint Note (W22)
Since today is Sunday, the weekly lint was run. → See Lint Report — 2026-W22.