August 12, 2026
Proofs, panic, and robot bragging
Tim Gowers: What sort of maths are LLMs good at?
Math genius asks what AI is really good at — and the comments go full brain-meltdown
TLDR: Tim Gowers says AI solving major math problems is stunning, but it still may not match humans in every kind of mathematical thinking. The comments split between awe and skepticism: some want truly beautiful, original ideas, while others think the trick is just giving the machine more chances to guess right.
After OpenAI’s jaw-dropping claim that its systems cracked 10 huge unsolved math problems, mathematician Tim Gowers stepped in with the calmer, cooler question everyone else was too busy panic-posting to ask: what kind of math are these things actually good at? His answer was not “everything,” and that alone set the mood. Gowers argues that while the results are astonishing, these systems still don’t look better than humans at all parts of mathematics — because if they were, we’d already be drowning in new discoveries.
And oh, the commenters had thoughts. The biggest theme was basically: AI may be brilliant, but is it actually creative? One widely loved takeaway was Gowers’s line that the real sign of human-level math won’t just be getting answers, but producing proofs that feel new, beautiful, and natural in hindsight. That sparked a mini swoon in the thread, with people treating it like the classy mic-drop of the post.
Then came the hot takes. One commenter compared AI to the human brain doing hard math “in the background” while still struggling when asked directly — a weirdly poetic idea that led to the meme-worthy jab that these models “are algebra, and yet kinda suck at it without training.” Another camp said the whole debate is secretly about letting AI keep trying over and over until something works, not magic genius. In other words: is this a mathematician, or just the world’s fastest lottery ticket machine? Even the side discussion about visualizing how models multiply had the vibe of fans trying to peek backstage at a mysterious pop star’s dressing room.
Key Points
- •The article is written as an early-August 2026 snapshot of LLM mathematical capabilities after OpenAI announced solutions to ten major open problems.
- •Gowers highlights the first construction of a non-sofic group and a superexponential lower-growth result for a multicolour Ramsey number as especially significant examples.
- •He argues that these achievements do not yet imply LLMs are better than humans at all aspects of mathematics, because that would likely have produced many more results already.
- •The article explores whether LLMs are especially good at finding counterexamples, while noting they can also prove difficult statements.
- •Gowers uses Vinogradov's theorem to show that solving a problem cannot be naively classified as counterexample-finding merely because it can be expressed as negating a universally quantified statement.