Post #3736955
2026-07-10 20:22 UTC
Replies (2)
-
@MartinEscardo@mathstodon.xyz 2026-07-10 20:22
3/ Two things happened to me along the way. I still think it is more fun to work with people, in particular PhD students, than with so-called AI's. And the machine was actually OK. But just OK. Also it was impressive, e.g. writing Agda code we know we could have written ourselves, if we wanted to and if we were patient enough. Everything I discuss here is public [3], including the AI's own first-person account and every prompt I gave it. You will have to judge by yourself. I don't think I know what this means for the future, and I don't think anybody else knows either. But I am kind of sad after this experiment. [3] https://github.com/martinescardo/ClaudeExperiment
-
@fullyabstract@fosstodon.org 2026-07-10 21:04
@MartinEscardo@mathstodon.xyz "everything else was written by Claude ..." I'm no expert, but I believe that the concept of "written by Claude" can mean dramatically different things depending upon what it's been trained on. Suppose it was trained on a similar proof, and only had to tweak it? Are full proofs embedded in the models, so they can be extracted with the right prompts? I think the answer is yes, but again I'm no expert.