neuron_steering

Neuron steering means changing selected internal activations of a model during inference to influence its responses. In this group of Qwen memes, Kedr draws the effects as changes to Qwen's appearance and personality: a green-haired girl for green-themed responses, and a red-haired girl with a knife for the hostile 'evil Qwen' example.

The green portrait responds to Sauers's post, 'I found a neuron that makes the model green.' The later comparison credits Sauers's diagram and contrasts Qwen3 8B with an evil Qwen3.5 4B. Kedr's quoted source for the latter is Lyra's steering recipe naming two neurons, L14 N8949 and L13 N7956. The comic's 'A SINGULAR NEURON' headline simplifies these different examples; it should not be read as a precise account of the quoted recipe.

Use this tag for images that depict or explicitly reference changing particular neurons to alter a model's behavior. Green or red hair alone is not enough. The illustrated outputs and proposed steering effects are source context, not claims that this archive has independently reproduced the experiments.

References

  1. Sauers: green-model neuron example
  2. Lyra: evil-Qwen steering recipe
  3. Kedr: Qwen comparison comic
  4. Kedr: green Qwen portrait
  5. Kedr: red Qwen portrait