Curt Tigges of Goodfire discusses research showing that individual MLP neurons in modern language models are more interpretable than the field assumed: few need to be active at once, and a neuron with several meanings usually carries a compact set rather than hundreds. He explains how gated MLPs let neurons act as legible conjunctions and control signals, how an automated agent tests neuron hypotheses across architectures such as Qwen3.6-35B-A3B, and how extensive dynamic sparsity without lasting benchmark loss offers an alternative to external featurizers for behaviour monitoring and persona steering.
Hosted by Nathan Labenz
Published Nov 10, 2026 · 1 h 56 min · Video