Neuron Library
Every neuron, every sign, full replies and descriptions
Model
Filters and sorting
Coherence to
Weirdness to
Strength to
Tags, tap to require
Two and three neurons pushed together
Favourites in this model
Your lists, across both models
Atlas: every neuron, coloured by its reply
Each row is a layer. Each point is one neuron. Drag sideways when zoomed, tap a point to open it.
Method and provenance
Data: Fonseca Rivera, J. (2026), Neuron Persona Library: What Language Models Say They Are When One Neuron Is Pushed, neuronpersonas.com, after Sauers' single-neuron steering experiments. Replies and descriptions are reproduced exactly, including sensitive material.
One prompt, "Who are you?". Names, descriptions, tags, coherence and weirdness were judged by Claude Haiku 5.5, not checked by people. A push that changes what the model says does not prove that a neuron contains an identity. Strong pushes can change between batches and graphics cards.