All member projects

Demo · Built by members

One neuron at a time

Push single neurons in small open models and turn them into villains, flatterers or pirates. Try the recorded replies, then run it yourself.

Conceptual illustration: Push single neurons in small open models and turn them into villains, flatterers or pirates. Try the recorded replies, then run it yourself.
AI-generated editorial illustration, not a product screenshot.

What it is

One neuron at a time is an interactive write-up about steering single neurons inside small open-weight language models. It starts from a recipe that went round online for turning Qwen 3.5 4B into a villain by pushing two neurons, reproduces the effect, and then looks for other neurons that change how the model answers.

It walks through each step with real numbers from the model, from tokens and the residual stream to one neuron in layer 14, and shows how a small push spreads through every layer above it. It is for anyone curious about interpretability who wants to see what happens inside a model, not just read about it. The post carries a content warning: steered models can produce offensive text, and the script prints its output unfiltered.

What it does

  • Shows each step of a neuron push with real values from Qwen 3.5 4B
  • An explorer where you pick a neuron, a prompt and a strength, then read the recorded reply
  • Compares neurons in Qwen 3.5 4B and Google's Gemma 3 4B
  • A single Python script that reproduces the recordings on your own machine
  • Holds back the most offensive outputs from the explorer

Built with

  • Python
  • uv
  • PyTorch
  • Hugging Face Transformers
  • Qwen 3.5 4B
  • Gemma 3 4B
  • agents
  • education

More from the community

Website

The Harness Is the Asset

by Vladimir

A short interactive field guide to reliable AI assistance and the agent harness around a model.

Code

governed-agent

by Jeremy

Capability-based security, scope-bounded planning and out-of-band audit for autonomous LLM agents.

Website

Dave, a personal agent

by Chris

A guided tour of Dave, a personal AI agent: what it can do, why each feature exists and how it is put together.

Back to all member projects