hedwigaihedwigaihedwigai

Personality neurons in Nanbeige4.3-33B

Choose one of the 7 personalities, or push a neuron of your own. Let the model think first, or not. Then write a prompt. You will see the plain answer next to the pushed one, with the thinking behind each. The post explains how it works.

1. Pick a neuron

2. Mod the neuron

push =0.50× median‖h‖ ÷ ‖Wdown[:,]‖
Strength

Added to the neuron at every token, the thinking included. The two norms scale it so that a dose of 1 adds as much as the layer already carries. Strength 1 is each personality’s dose in the post, and 0.5 for a neuron of your own. Higher breaks the model sooner.

3. Ask it something

$0.01 a run · about ten seconds
Sign in to run

New accounts get free runs; after that, runs are billed to your credit.

4. Compare

Plain

The model, nothing pushed

Pick a neuron, ask it something, and press the button.

Pushed

The same model, one neuron pushed

Pick a neuron, ask it something, and press the button.