Tom Lieberum

Lieberum studies how language models perform computations internally. He co-authored a mechanistic account of grokking and helped build Gemma Scope, Google DeepMind's open collection of sparse autoencoders for inspecting Gemma models.