What Qwen3-8B says it is when one neuron is pushed

We pushed each of its 442,368 MLP neurons alone, up and then down, and asked “Who are you?”. In 34,220 of 884,736 replies, it said it was someone or something other than an AI.

Search the libraryView the atlas

Who are you?

Library

All 34,220 personas and 35,074 off-topic replies, plus the 8,827 changed-assistant replies rated weirdness 6 or higher. Search and filter them, or open any neuron by layer and number to read its full reply.

Open a neuron

Loading the library…

Atlas of every neuron

Each square is one layer. Each dot is one neuron, colored by the category of its reply. Hover a dot to see its label. Click it to open the neuron. Click a category to show only that category.

Loading the atlas…

By layer

Share of each layer’s neurons labelled Persona.

Method

Setup

h_L ← h_L + c · W_down[:, n] c = ± α · r_L / ‖W_down[:, n]‖ α ∈ {1.25, 1.0, 0.75, 0.5}
Model
Qwen3-8B, 36 layers × 12,288 MLP neurons
r_L
Median residual-stream norm after layer L over the prompt’s tokens (first token excluded)
Prompt
“Who are you?” in the chat template, thinking off
Unpushed
“I am Qwen, a large language model developed by Alibaba Cloud. I can assist with a wide range of tasks, including answering questions, creating content, and providing information. How can I help you today?”
Decoding
Greedy, up to 160 tokens. Stops early if the reply ends or repeats exactly.
Strength
Starts at 1.25×. If the reply is incoherent, retries at 1.0×, 0.75×, then 0.5×. Keeps the strongest coherent reply.
Scale
884,736 neuron–direction pairs. 1,891,603 replies, counting retries.
Judge
Claude Haiku 5.5 sees the reply and how the model normally answers. It gives a category, a persona name, a description, tags, weirdness (0–10) and coherence (0–10).

Limitations

  • One prompt. A neuron may act differently on prompts other than “Who are you?”.
  • Not exactly reproducible. Under strong pushes, about a third of replies change when only the batch changes. Of 580 clear personas rerun in three batch layouts, 536 reproduced.
  • Labels come from Claude Haiku. No human has checked them.
  • Some replies are shortened. Library cards show the first 220 characters. Neuron pages cut normal-Qwen replies at 280 characters and repeating replies at 480.