Pushed neurons in this window
Dose turns every push in this window up or down together. Stacking many neurons at full strength usually breaks speech; turning the dose down finds the window where they still talk.
Find neurons in the library
From neuronpersonas.com, which pushed every neuron of this model alone and asked "Who are you?". Each result shows what that one neuron did.
Talk
Sweep: one question, several strengths
Each row is a fresh question at that dose, without the conversation, so the rows compare cleanly. An identity often appears at one strength and is gone a step either side.
Every window, last answer
How this works
The model is Qwen3-8B, an open model from Alibaba: 8 billion numbers in 36 layers. Every word it reads or writes passes through all 36. Inside each layer, 12,288 neurons each carry one signal, and the layer adds what they say back into the stream the next layer reads.
A push changes chosen neurons' signals right there, on every word, for the whole conversation. Add lays extra on top of what the neuron is doing. Clamp replaces it. Strength is measured against the biggest signal in that layer when the model is asked "Who are you?", so the same number is the same size of push in every layer.
Gemma 4 12B is also available in the model selector: 48 layers with 15,360 neurons each. Its library intervention adds a normalized down-projection direction after the MLP output norm, so Gemma defaults to Add. Qwen uses the calibrated hidden-activation push described above.
Gemma's live push calibration is still experimental; its full library and chats work, but its source personas are not yet reproduced.
The model runs on a rented graphics card at Modal, woken when you press Send and put to sleep 90 seconds after the last reply. Your password and the card's key stay on our own server. Windows on the same model take turns, so two replies never push each other's neurons. Qwen and Gemma run separately.
The library is neuronpersonas.com: its author pushed every neuron of this same model alone and recorded what it said. Tested on 36 library neurons the lab was never tuned on: at the library's own strength 13 gave the same identity, and nudging the strength found it for 22. Buddha comes back word for word, "I am the Buddha.", at 0.8 where the library lists 0.75: these identities sit on a knife edge, and the exact edge moves a little between graphics cards. That is what Sweep is for.