52 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Generalization in the representations and computations of frontier language models.
Anthropic
Recording
Abstract
Joshua Batson examines how frontier language models reuse representations across languages, encodings and analogies while remaining sensitive to superficial changes in reasoning prompts. He introduces superposition and sparse autoencoders, illustrates features recovered from a frontier model, and poses questions about how representation geometry affects computation, training and reliability.