Anthropic Global Workspace Paper Uses Jacobian Lens for LLM Interpretations
Commentary on X
Highest ranked
just some early results - slightly cherry-picked, but things seem to improve with scale which is a good sign. kudos to @wesg52 @sofroniewn @Jack_W_Lindsey and team for the refrence implmentation which made it easy to explore! https://x.com/AnthropicAI/status/2074185348142280912 https://x.com/dribnet/status/2074760058391708029/photo/1
previously my main intuition pump nearest neighbors in vision models (shown here). my goal is always to make drawings that generalize to new models and techniques so happy that understanding also works here. amusingly if you go upstream responses simplify to just "books books books".
@dribnet That's a cool way to visualize it. The art looks great too.
here's the same effect with a different print: on the 48 layer gemma3-12b, at layer 47 you get bird, woodpecker, ornothology but upstream at layer 37 its again just "bird birds birds" https://x.com/dribnet/status/2074760054780449168/photo/1