What AstroPT knows about galaxies, and what that can teach us about LLMs

UniverseTBD, :, Kshitij Duraphe +3

Interpretability research increasingly asks when concepts emerge during training and whether linear probes recover real structure, but in language models these claims are hard to v…