VibeMathedMath problems solved by AI

The Sum-Product Conjecture over the Reals

Erdos and Szemeredi conjectured that every finite set of reals satisfies max(A+A,AA)A2o(1)\max(|A+A|,|AA|) \ge |A|^{2-o(1)}. False: there are arbitrarily large ARA \subset \mathbb{R}, of algebraic integers in a number field of degree logA\asymp \log|A|, with max(A+A,AA)A2c\max(|A+A|,|AA|) \le |A|^{2-c} for an absolute c>0c > 0. Variants give counterexamples in function fields of fixed positive characteristic.

Result
Disproved(see note)
Status
Resolved
AI contribution
AI-assisted
Method
Construction
Field
Additive combinatorics
Posed by
Paul Erdos, Endre Szemeredi
Year posed
1983
Years open
43y
Solved
2026-05-27
Model
GPT-5.5 Pro
Vendor
OpenAI
Collaborators
Thomas F. Bloom, Will Sawin, Carl Schildkraut, Dmitrii Zakharov
Verification
Unreviewed
Publication
Preprint
Significance
55 / 100
Disclosed cost
Wikipedia
No dedicated article

What was actually shown

the model's contribution is one simplifying lemma; the authors state the main ideas are human

What the AI did

The limits here matter more than the headline, and the authors state them plainly: GPT-5.5 Pro was a sounding board in the early stages, but the final proof including all the main ideas was almost entirely human-generated, and everything in the paper was written by the authors. The single exception they name is Lemma 3.4, suggested by the model, which replaced a more complicated result of Schinzel with a short elementary argument. There is a second, indirect AI thread: the authors say they were inspired to revisit number fields of large degree by OpenAI's counterexample to the unit distance conjecture, and note their construction needed far less number-theoretic input than that one did.

Verification

arXiv preprint by four established additive combinatorialists; not yet peer-reviewed. Given the size of the claim this one deserves refereeing before it is treated as settled.

Source

arXiv:2605.28781 - The sum-product conjecture is false for real numbers

Discussion