Post #3117103
2026-05-04 02:49 UTC
@hweimer@fediscience.org agree. LLMs are interpolation machines. They can't, by design, extrapolate. An answer that was never given is unlikely to be in the training set and hence has a very low probability.
#ai #llms #science #math #physics
Replies (1)
-
@rayk@techhub.social 2026-05-04 05:45
@lostgen@det.social @hweimer@fediscience.org There’s enough randomness that an answer that is a random combination of other answers might be given though. That doesn’t mean it’s “creative”, but it does mean that you can see things that are novel, in that that particular combination of parts hasn’t been seen. This is basically interpolation though, and your extrapolation argument is spot-on. They do not produce output that discovers anything new.