Re:I struggle to believe that AI R&D is ingenious
by BenoitP on Friday November 06, 2026 @08:37AM
(#50392134)
Strong agree, but with a major caveat: First since we humans don't have a memory the size of the internet, new TO YOU is extremely easy to find. And second: a lot of discoveries can be made by trying to match two different domains and generate new insights at the seams.
A lot of math theorems are discovered that way by humans btw: find a difficult problem in a domain, map it to another domain, apply existing theorems there, go back to the original domain. The only difference is an LLM would have to search for an interesting problem, but that can be done by brute-forcing search at the seams and validating the problem does not previously exists. And there is no new TO YOU for the LLM: it can get help from a vast amount of helper domains.
IMHO there's a shot at having category theory really unifying a large chunk of maths in the next decade. It'll be a big ugly ball of Lean, but a queryable one at that.
A lot of math theorems are discovered that way by humans btw: find a difficult problem in a domain, map it to another domain, apply existing theorems there, go back to the original domain. The only difference is an LLM would have to search for an interesting problem, but that can be done by brute-forcing search at the seams and validating the problem does not previously exists. And there is no new TO YOU for the LLM: it can get help from a vast amount of helper domains.
IMHO there's a shot at having category theory really unifying a large chunk of maths in the next decade. It'll be a big ugly ball of Lean, but a queryable one at that.