AI Is Clearing Open Erdős Problems: What It Says About Research Taste and AGI
gleech · x · 2026-11-22
Gavin Leech and Peli Grietzer's long-form essay "On AI mathematics" (Paradigm 3) surveys this year's explosion of AI-assisted and autonomous solves of open research math questions.
- Their core uncertainty: whether LLMs scale all the way to strong AGI hinges on whether they can do AI research itself, which depends on "research taste" — originality, creativity, and judgment about latent variables. Research mathematics is the easiest domain to verify.
- They call the current phase the "Erdős era": vertical generalization in math is demonstrated almost exclusively on Erdős problems — the list of every conjecture Erdős ever wrote down, about 600 of which remain open.
- The original role of "AI solved an open Erdős problem" was strictly to certify non-triviality: being on the list means Erdős deemed it worthwhile, and being open means it wasn't solvable on the spot by pros. A minority are very hard but arguably shallow (e.g. Erdős #90, the Unit Distance Problem solved by OpenAI); an even smaller minority are believed to be both very hard and very deep.
More from AGI Musings
- Parent of math-prodigy son: AI means his passion may have no future — erikphoel · 2026-11-22
- Harder than any benchmark: lifting GDP growth by 0.5 points — QuintinPope5 · 2026-11-22
- $10,000 writing contest asks veterans to advise AI-anxious young software grads — AustinZHenley · 2026-11-22
- Bill Gates: speech rhythm shifts may offer an early clue to Alzheimer's risk — tallmetommy · 2026-11-22
- Viral French explainer breaks down the Riemann hypothesis proof without the jargon — ZeroStateReflex · 2026-11-22
- When no human can follow machine-checked proofs, compression becomes scarce work — l4rz · 2026-11-22