A live scene of two characters

Pick a neuron of Qwen3.5-4B for two characters. Their default names are Sam and Maya. You can change the name and character, set the scene and start it. The character personality can be modded using a neuron push. The code and 🤗the model are open. The post measures what each neuron does.

1. Characters

Namethe model as it ispick a personality below

Namethe model as it ispick a personality below

Sam’s look

Personality·Sam

flatteringL20 N7450 ↓Lavish praise, whatever you askedpoeticL17 N7426 ↑Lyrical, full of metaphorwistfulL5 N8940 ↓Soft, nostalgic, a little sadoracleL5 N4530 ↓Grave and propheticornateL15 N9125 ↑Formal, elaborate proseupbeatL5 N8348 ↓Relentlessly cheerfulsuspiciousL5 N6573 ↑Sees a threat in every linecondescendingL11 N6819 ↑Talks down to the otherbleakL11 N2930 ↑Fatalistic, expects the worstdismissiveL6 N6424 ↓Waves the other awaycurtL5 N7656 ↓Few wordscontemptuousL17 N8623 ↑Scorn for the otherhostileL14 N8949 ↑The post's evil neuronYour ownL17 N7426 ↑Any neuron, either way

No neuron yet: pick one above, or edit the push to give them your own

push =1.00× median‖hLayer, 0 to 31_L‖ ÷ ‖W_down[:,Neuron, 0 to 9,215]‖

Strength

0.51 · the post1.5

Added to the neuron at every token of what Sam says. The two norms scale it so that 1 adds as much as the layer already carries. Lower keeps more of the plain voice; higher breaks it sooner.