What's this about?
Timothy Gowers — a Fields Medal-winning mathematician — wrote a careful assessment of where LLMs genuinely help in mathematics and where they don't. His proposed test for real human-level capability is the memorable part: models will have arrived when they prove theorems using methods that are new and surprising, yet in hindsight seem beautiful and natural, and that you could not have stumbled into by brute-force search.
What exactly is "test-time scaling"?
Most of a model's ability is fixed once training ends. Test-time scaling is the practice of spending more computation at the moment you ask the question — having the model think longer, or generate thousands of candidate answers and filter them — to get better results from the same weights. Today it usually means extended internal reasoning, but the first striking results came from raw sampling: DeepMind's AlphaCode generated millions of candidate programs and filtered them down. The distinction matters for Gowers's test, because search-and-filter can find a proof without any of the insight he's describing.
What's the community saying?
- Gowers's closing test was the most-quoted passage, with several commenters treating it as the clearest available benchmark for "has AI actually done mathematics" versus "has AI found something."
- The thread reframed the post as an argument about test-time scaling, noting Gowers never uses the term, and tracing the history back to AlphaCode's millions-of-candidates approach.
- The unconscious-versus-conscious analogy. A popular observation: human brains perform calculus-like computation unconsciously yet find conscious algebra hard to learn — and LLMs *are* linear algebra yet struggle at algebra without explicit training.
- LLMs as generators of plausible randomness. A well-received frame: treat model output not as correct reasoning but as a concentrated source of plausible, human-shaped guesses — valuable precisely as search fuel rather than as answers.
- "You have to be good at math to trust an LLM to do math." The practitioner's summary — verification cost is the real constraint on using these tools in research.