Back to timeline

Research · January 12, 2026

The first Erdos problem is closed by AI autonomously

A human writeup appeared on arXiv of a formal Lean proof produced by GPT-5.2 Pro together with Harmonic's Aristotle. The abstract states this is the first Erdos problem regarded as fully resolved autonomously by an AI system.

Why it matters

For the first time a named open problem from the Erdos list was closed by machine, and the answer is checkable in Lean rather than dependent on a referee's judgement.

Nat Sothanaphan submitted the paper on 12 January 2026, with a last revision on 26 January. The proof came from two systems working together: OpenAI's GPT-5.2 Pro produced an informal argument and Harmonic's Aristotle turned it into a formal Lean proof, both operated by Kevin Barreto. The paper itself is a translation of the formal proof into ordinary mathematical exposition so that people can read it. The mathematical content: for infinitely many triples where a!b! divides n!k!, the gap k = a + b - n is of logarithmic order, lying between C1 log n and C2 log n. Barreto had announced the result publicly on 4 January, but forum participants noted that the problem statement was ambiguous and initially treated the proof as partial; the argument was brought to the intended version over the following days.

Event record

Event date
January 12, 2026
Timeline date
Event date
Verification
Sources gathered automatically · September 19, 2026
Lines
ID
evt-0364

The day the paper was submitted to arXiv; last revised 26 January 2026. The result was announced publicly on 4 January, but the statement of the problem was still disputed then.

Sources

Records that link to this one