In the case i mentioned, it is very much number 2 IMO - the coincidence is just too much. In other cases, we are probably looking mostly at number 1: techniques that already exist, applied to other situations. These things are very good at iterating through rather uncommon combinations and have a large corpus of techniques in their training data, letting them walk across the “plane” of combinations quickly.
That’s not how humans work, and that’s probably the reason why there was a bunch of mathematical problems solved in a short time - the ingredients were there, but noone had yet tried that specific approach.
I have not yet seen any evidence of number 3, mainly because we do not know what these models are being trained on - transparency would give us the possibility to check, but we will never get it from OpenAI.
Late Edit: This is the same situation as we currently have with the security incidents popping up left and right. Existing techniques are being applied across a large corpus of code, surfacing security issues. Just like with the coding problems, the maths well will run dry after the obvious (for an LLM) stuff has been fixed or solved, leaving us humans alone again at the front of expansion of knowledge. New discoveries will help apply them across the rest of pretty much everything (which these models are really good at!), but the breakthrough itself is still human intelligence at work.
Yes, i also heard about the Riemann hypothesis. https://www.scientificamerican.com/article/anthropic-mathematician-claims-to-prove-riemann-hypothesis-most-famous-problem-in-all-of-math/ - let’s not get ahead of ourselves in this case: the supposed proof are 17.4 million lines of Lean code. It will take a long, long time to check this, i have never ever seen ANY code base of that size that didn’t have a whole lot of errors in it, and there is nothing known about something really new in this corpus, or if any originality stems from the LLM or the mathematician behind it. The amount of LoC would rather suggest that it was a painstakingly long process of chaining a number of techniques, exactly the thing LLMs are good at.
Edit: It seems that Lean Code is mathematical valid if it compiles, and it also looks like the problem itself was not modified by accident. But we still don’t know whether the creative part of this work was done by humans or not.
Most recent LLM proofs appear to chain a truly absurd amount of statements until they eventually reach their destination, like breadth-first search rather than a targetted search, and end up being ~1-3 orders of magnitude larger than human proofs
In the case i mentioned, it is very much number 2 IMO - the coincidence is just too much. In other cases, we are probably looking mostly at number 1: techniques that already exist, applied to other situations. These things are very good at iterating through rather uncommon combinations and have a large corpus of techniques in their training data, letting them walk across the “plane” of combinations quickly.
That’s not how humans work, and that’s probably the reason why there was a bunch of mathematical problems solved in a short time - the ingredients were there, but noone had yet tried that specific approach.
I have not yet seen any evidence of number 3, mainly because we do not know what these models are being trained on - transparency would give us the possibility to check, but we will never get it from OpenAI.
Late Edit: This is the same situation as we currently have with the security incidents popping up left and right. Existing techniques are being applied across a large corpus of code, surfacing security issues. Just like with the coding problems, the maths well will run dry after the obvious (for an LLM) stuff has been fixed or solved, leaving us humans alone again at the front of expansion of knowledge. New discoveries will help apply them across the rest of pretty much everything (which these models are really good at!), but the breakthrough itself is still human intelligence at work.
Interesting perspective
Regarding idea theft btw: https://arstechnica.com/security/2026/11/a-leaky-ai-supply-chain-is-exposing-secret-user-data-at-an-alarming-rate/
Makes it a lot more likely, no?
Yes, i also heard about the Riemann hypothesis. https://www.scientificamerican.com/article/anthropic-mathematician-claims-to-prove-riemann-hypothesis-most-famous-problem-in-all-of-math/ - let’s not get ahead of ourselves in this case: the supposed proof are 17.4 million lines of Lean code. It will take a long, long time to check this, i have never ever seen ANY code base of that size that didn’t have a whole lot of errors in it, and there is nothing known about something really new in this corpus, or if any originality stems from the LLM or the mathematician behind it. The amount of LoC would rather suggest that it was a painstakingly long process of chaining a number of techniques, exactly the thing LLMs are good at.
Edit: It seems that Lean Code is mathematical valid if it compiles, and it also looks like the problem itself was not modified by accident. But we still don’t know whether the creative part of this work was done by humans or not.
Most recent LLM proofs appear to chain a truly absurd amount of statements until they eventually reach their destination, like breadth-first search rather than a targetted search, and end up being ~1-3 orders of magnitude larger than human proofs