PERSON DIRECTORY
Lee Sharkey
Lee Sharkey appears in 1 indexed conversation across The Cognitive Revolution. This directory brings every appearance, source, TL;DR, digest, and transcript into one searchable feed.
Untangling Neural Network Mechanisms: Goodfire's Lee Sharkey on Parameter-based Interpretability
Erik TorenbergNathan LabenzLee Sharkey
Goodfire is betting that parameter-level interpretability can recover the reusable computations implemented by weights, architecture, and nonlinearities, rather than merely cataloging activation concepts; its stochastic parameter decomposition uses rank-one components and causal-importance masking to address APD’s instability and computational burden.The $50 million raise supports scaling experiments from million-parameter models into the low single-digit billions, with potential applications in parameter-level unlearning and function-based deception monitoring, but whether SPD recovers meaningful mechanisms or outperforms simpler gradient-guided and activation-based tools remains open.