𓆏 retweeted

The rapid advances we see in AI-derived mathematical proofs are almost certainly not representative of science in general.
A core driver of these advances is that AI-derived proofs can be translated into a highly structured human-designed verification language, which can then be checked using traditional computer programs.
The AI slop-cannon can generate as many slop attempts as needed to get a proof that works, because humans already built the de-slopification engine that automates digging the diamonds out of the slop.
This kind of cheap validation does not exist in data science or the empirical sciences more broadly.
In fact, validation in the sciences is often orders of magnitude more expensive than all the other parts, which is why AI is going to be much less effective there.
English






















