Sweeping every neuron and letting Haiku rate them by weirdness

CatAstro_Piyush · x · 2026-11-01

A hobbyist shared an interpretability toy: sweep through every neuron of a model, then have Claude Haiku categorize and rate each one by weirdness. They say searching and combining neurons is a lot of fun, calling the experience of seeing individual neurons "beautiful."

Original post →

More from Fun

Fun channel →