I am completely baffled by this given how dumb these models are when I interact with them, but the rational position seems to be to accept that Claude proved the Riemann hypothesis.
One reason that these LLM math advances are written in Lean is at any point they can run a script and verify their work is valid. If you ask an LLM about something in history, objects in a photo, or something novel, there's no opening to check its work beyond making it check its own work.
You also have to consider if an LLM can solve the Riemann hypothesis because it is 10x smarter than any mathematician, or because it is willing to write 1000x as many lines of code (including false starts) without giving up?
Claude did not prove the Riemann hypothesis. Levent Alpöge proved it using Claude as a tool.
> how dumb these models are when I interact with them
its because models are not trained on terrabytes/petabytes(?) of RSI generated data in your domain, unlike math and coding.
Also, there is a chance you/we can't always comprehend how smart they are.
Assuming you are serious;
LLMs exhibit what is known as "Jagged Intelligence" - https://news.ycombinator.com/item?id=48577159
Mathematics is one domain where they have high peaks. Why? Because we have "formalized" mathematics so well that starting with the precise and correct set of problem statements/assumptions/etc. if you follow the logical rules of inference (i.e. deductive system) then you will reach the conclusion. Formalization was invented so that we can solve problems purely mechanically at the symbolic syntax level using a machine. So the LLM proposes proof steps to the Lean proof assistant which then symbolically verifies it.
On the other hand, LLMs often fall down in "common sense reasoning" (https://en.wikipedia.org/wiki/Commonsense_reasoning) since this involves a lot of implicit context etc.
Claude proved it with considerable prompting by a top-level mathematician who has become expert at coercing LLMs into solving these math problems.
It's not like Anthropic simply gave internal Claude agents the task to solve RH and months later they came back with the solution.
I know, even on programming / computer architecture related topics. Just the other day I asked GPT5.6 Luna two questions about x86-64 stack layout / frame pointers and I noticed it got them both wrong.
Although opening a new chat and asking if the previous output was correct, it detected the mistakes.