Four neurons, one push

Each grey dot is one MLP neuron of Qwen3.5-0.8B (300 sampled per layer, 24 layers), placed by the direction it writes into the model. The Fable editor changed four of them, in layers 17 to 20. Each change adds the same direction, toward the </think> token. Watch them leave the cloud. The MIT report's own released edit (blue, 128 neurons in layers 15 and 19) changes other directions, so on this axis it hardly moves.

Fable edit, 4 neurons MIT edit, 128 neurons unchanged neurons before

Drag to turn. Horizontal axes share one scale (cosine). x: alignment with the added direction. Depth: first principal component of the rest. Height: layer.