This post is important because it gives a concrete, expert-audited example of an LLM moving beyond explanation into plausible original mathematical research, forcing researchers to rethink what kinds of problems, training tasks, and verification workflows remain distinctly human.
https://gowers.wordpress.com/2026/05/08/a-recent-experience-with-chatgpt-5-5-pro/
https://x.com/wtgowers/status/2052830948685676605
https://www.reddit.com/r/mathematics/comments/1taug8h/fields_medalwinning_mathematician_says_gpt55_is/
https://www.reddit.com/r/AIDangers/comments/1tatc0x/fields_medalwinning_mathematician_says_gpt55_is/
https://chatgpt.com/c/6a030834-a440-83ec-bf9b-4c2d38de7521
Tim Gowers describes a striking experiment in which ChatGPT 5.5 Pro produced what he regards as PhD-level additive-combinatorics research, with almost no mathematical steering from him. The core claim is not merely that the model found a known trick, but that it improved an existing research argument and introduced what Isaac Rajagopal judged to be an original, clever construction. ([Gowers’s Weblog][1])
The setting is additive number theory. Given a finite set of integers (A), one studies the size of its (h)-fold sumset:
[ hA = {a_1 + \cdots + a_h : a_i \in A}. ]
| For fixed (h) and (k = | A | ), (\mathcal R(h,k)) is the set of all possible values of ( | hA | ). A related quantity, (N(h,k)), asks: how small an interval ({0,1,\dots,N}) is enough to realize all possible sumset sizes for (k)-element sets? ([Gowers’s Weblog][1]) |
The original known constructions used geometric progressions like:
[ {0,1,m,m^2,\dots,m^{\ell-2}} ]
but these have exponentially large elements, leading to exponential bounds in (k). ChatGPT’s key contribution was to replace these with polynomially bounded constructions that preserve the needed additive behaviour. ([Gowers’s Weblog][1])
Gowers first asked ChatGPT about the (h=2) case. It produced a construction improving Nathanson’s exponential-looking diameter bound to a quadratic upper bound, which Gowers says is clearly best possible. The idea was to replace powers of 2 with a more efficient Sidon set, since Sidon sets can have quadratic diameter. ([Gowers’s Weblog][1])
Then he asked about general (h). ChatGPT first improved Rajagopal’s bound from exponential in (k) to roughly exponential in (k^{1/2+\varepsilon}). After further prompting, it produced a claimed polynomial bound. Rajagopal reviewed the result and described it as almost certainly correct, not just line-by-line but at the level of ideas. ([Gowers’s Weblog][1])
The resulting bound stated in Rajagopal’s guest section is:
[ N(h,k) \leq O(k^{10h^3}) ]
for sufficiently large (k). The lower bound is only known to be on the order of (k^h), so the true asymptotic remains open. ([Gowers’s Weblog][1])
Rajagopal explains that his own constructions used geometric-series-like blocks whose additive relations were useful but whose elements were exponentially large. ChatGPT constructed substitute sets (G) and (H) that behave like “half a geometric series squeezed into a polynomial interval.” ([Gowers’s Weblog][1])
The construction uses (h^2)-dissociated sets: sets with no nontrivial low-order additive relations. These allow ChatGPT to create sets that have just the required additive collisions, such as relations of the form (mx = y), while avoiding unwanted extra relations. This preserves the combinatorial behaviour of the geometric progression but with polynomial-sized elements. ([Gowers’s Weblog][1])
Rajagopal says this use of (h^2)-dissociated sets to control low-order additive relations “feels quite ingenious” and, as far as he can tell, is original. ([Gowers’s Weblog][1])
The important point is that this was not just literature retrieval or symbolic grinding. Gowers frames the result as a non-trivial extension of recent human work, and Rajagopal says the polynomial improvement involved an idea he would have been proud to find after a week or two. ([Gowers’s Weblog][1])
For AI and research, the post suggests a new threshold: in some areas of maths, especially problem-focused areas like combinatorics, “open problem” no longer necessarily means “safe PhD starter problem.” It may mean “not yet tried seriously by a frontier model.” ([Gowers’s Weblog][1])
Gowers argues that mathematical research training may need to change. Beginner-level open problems that once helped PhD students learn research may increasingly be solvable by LLMs, so the useful human task may shift toward collaborating with models, judging correctness, supplying taste, choosing good problems, and knowing when an AI-generated argument is meaningful. ([Gowers’s Weblog][1])
He also raises an infrastructure question: what should happen to AI-generated mathematical results? They may be valid and publishable in mathematical substance, but journals and arXiv are not obviously the right venue. Gowers suggests a possible moderated repository where human mathematicians, or ideally proof assistants, certify correctness. ([Gowers’s Weblog][1])
A frontier LLM appears to have produced a serious additive-combinatorics result by replacing exponential-size geometric-series constructions with polynomial-size dissociated-set constructions, suggesting that some “gentle” open research problems may now be within autonomous LLM reach.
| [1]: https://gowers.wordpress.com/2026/05/08/a-recent-experience-with-chatgpt-5-5-pro/ “A recent experience with ChatGPT 5.5 Pro | Gowers’s Weblog” |
Here are the best recent pieces in the same category: frontier AI crossing from “math helper” into plausible research collaborator.
| Article | Why it’s similar / worth reading |
|---|---|
| “The AI Revolution in Math Has Arrived” — Quanta Magazine, Apr 2026 | Probably the closest broad companion piece. It covers First Proof, where AI systems were tested on unpublished research-level math questions and reportedly solved more than half. Good overview of the inflection point. (Quanta Magazine) |
| “As AI keeps improving, mathematicians struggle to foretell their own future” — Scientific American, Mar 2026 | Covers the same “what happens to mathematical research now?” question, especially around First Proof and whether LLMs can prove useful lemmas for working mathematicians. (Harvard Mathematics Department) |
| “Mathematics in the Library of Babel” — Daniel Litt, Feb 2026 | A skeptical mathematician updating toward “this is real.” Good for understanding why recent AI-math results feel different from earlier hype. (Daniel Litt) |
| “How GPT-5 helped mathematician Ernest Ryu solve a 40-year-old open problem” — OpenAI, Nov 2025 | Vendor-authored, so read critically, but directly relevant: a working mathematician used GPT-5 to accelerate exploration and contribute to a long-standing optimization problem. (OpenAI) |
| “The story of Erdős problem #1026” — Terence Tao, Dec 2025 | A concrete case study of AI, humans, existing literature, and online collaboration combining to solve an Erdős problem. More nuanced than simple “AI solved it” headlines. (What’s new) |
| “Problem 728 and the use of AI on Erdős problems” — Erdős Problems blog, Jan 2026 | Useful for seeing how the mathematical community itself is tracking AI-assisted solutions to open problems, including concern about provenance and verification. (erdosproblems.com) |
| “An amateur just solved a 60-year-old math problem—by asking AI” — Scientific American, Apr 2026 | Similar to the Gowers post in that the interesting part is not just speed, but an AI-assisted route to a method experts thought was genuinely novel/useful. (Scientific American) |
| “Claude’s Cycles” — Donald Knuth, 2026 PDF | A primary-source mathematical write-up around Claude contributing to a graph-theoretic construction Knuth had been working on. Worth reading because it is from Knuth rather than a tech-company press release. (Computer Science) |
| “Mathematical methods and human thought in the age of AI” — Tanya Klowden & Terence Tao, arXiv, Mar 2026 | More reflective/philosophical than case-study-based, but important for the research-culture implications: what should remain human-centered as AI becomes embedded in mathematics. (arXiv) |
| “AI for Mathematics: Progress, Challenges, and Prospects” — arXiv survey, Jan/May 2026 | Best technical survey-style follow-up. Especially useful for the formal proof angle: why Lean/Coq-style verification may become crucial as LLM-generated math scales. (arXiv) |
| “AI Will Be Top of Mind at ICM, Math’s Biggest Conference” — Simons Foundation, May 2026 | Good signal that this is no longer a niche AI topic; it is becoming a central professional issue for mainstream mathematicians. (Simons Foundation) |
Best reading order: Quanta → Daniel Litt → Tao’s Erdős #1026 post → Knuth’s Claude’s Cycles → Tao/Klowden paper → AI for Mathematics survey.