Gowers + ChatGPT 5.5 Pro: PhD-level math research in under two hours
Source: The Decoder, 2026-05-09 · raw Tier: 2 — Active learning (intersects Tier 1 research-agent thread)
TL;DR
Fields Medalist Timothy Gowers gave ChatGPT 5.5 Pro open problems in number theory. Within an hour, the model improved a known exponential bound to a polynomial one. An MIT collaborator on the problem called the key idea "completely original." Gowers' own framing: the bar for human mathematical contribution is now "prove something LLMs can't do."
What is being claimed
- ChatGPT 5.5 Pro produced a meaningful research-grade contribution on a real open problem with zero human help during the two-hour run.
- The improvement was structural, exponential to polynomial bound, not arithmetic tightening of constants.
- An independent expert (MIT) inspecting the proof said the central idea was novel.
- Gowers' meta-claim is the headline: research mathematics is now in the regime where the marginal human contribution must be defined as "what the model cannot do."
Why this matters in light of prior wiki pages
This is the third data point in two weeks for the "AI does science" thread.
- 2026-05-09 — AI Co-Mathematician (FrontierMath Tier 4 at 48%) sets the agentic-workbench benchmark for math research. Gowers' run is the human-evaluated complement: a Fields Medalist running a real open problem rather than a curated benchmark.
- 2026-05-09 — Auto Research with Specialist Agents (+38.7% on NanoChat-D12 CORE) does the same thing for ML research recipes: closed-loop empirical search, no human in the loop.
- Kurate cs.AI #5 this week — "AI scientists produce results without reasoning scientifically" is the explicit counter-claim. Gowers' run is one anecdotal datapoint that pushes against that framing in this specific case (a real problem, an expert reviewer, a structural improvement). It does not refute the cs.AI #5 claim, which is a methodology critique about reproducibility and reasoning patterns. Both can be true.
The composition with the Skill curation cluster (StraTA, Skill1, SkillOS — also from yesterday) is not yet realized. Gowers' two-hour run is a single-shot inference from an off-the-shelf chat model, not a skill-augmented agent. The natural follow-up is the same problem class with an agent that has persistent failed-hypothesis memory.
Open questions
- Reproducibility. A single run on a single problem with an expert reviewer is an existence proof, not a frequency claim. What fraction of Gowers-tier open problems can ChatGPT 5.5 Pro improve in two hours?
- The "originality" bar. The MIT collaborator called the idea "completely original." Originality of an idea is hard to verify without exhaustive literature search; the model could have surfaced a 1970s-era technique that the modern reviewers had forgotten. Worth tracking whether the proof appears in any prior literature.
- What survives peer review. The bound improvement still has to clear refereeing. The wiki should track whether a written-up proof appears on arxiv with both Gowers and the model credited as authors, or whether the improvement gets reframed by the human author after re-derivation.
Implication for cere-bro tracking
Gowers' "now the bar is what LLMs can't do" line is the framing every research-agent paper should be measured against from this point forward. The skill-curation, agentic-workbench, and Auto-Research papers are all building toward this regime. The Fields-Medalist-attested datapoint says the regime has arrived in at least one subdomain (analytic number theory).