Sparse autoencoders are a technique for interpreting which concepts are represented in a model’s activations, and have been a big focus of recent mechanistic interpretability work. In this talk, Neel will assess what we’ve learned about how well sparse autoencoders work over the past 1.5 years, the biggest problems with them, and what he sees as next steps for the field.
Neel runs the mechanistic interpretability team at Google DeepMind. Prior to this he was an independent researcher, and did mechanistic interpretability research at Anthropic under Chris Olah. Neel is excited about helping build the mechanistic interpretability community and created the TransformerLens library, does far too much mentoring, and enjoys making educational materials.