Poor Things

I've been poking individual neurons (and pairs of neurons) in Qwen3.5-4B, then asking the steered model the same 67 questions as an unsteered control. Poor Things is the lab notebook: the prompts, the replies, and the weird little personalities that show up when you turn a knob. The names are working labels, not claims about what the model believes.

A few specimens to meet:

• Evil Qwen turns Newton's laws into a villain's manifesto.
• Tax Credit Prophet confidently invents tax credits and other facts that don't exist.
• Cookie Cookie swaps "cookie" into the vocabulary while trying to keep the shape of an answer intact. 🍪

The fun is in comparing each reply with the unsteered one. Come poke around.