Wait, what did you do to it?
Nothing you'd call hacking. No retraining, no new data, no sneaky system prompt: the weights are exactly the ones anyone can download. We just hold one or two of its neurons down while it talks.
The brain, very roughly
The model reads and writes one token (a word-ish chunk) at a time. Every token passes through 32 layers, and every layer has 9,216 neurons: little dials that turn up or down depending on what's being said. That's 294,912 dials.
Most of them mean nothing a human could name. A few happen to line up with something we recognise: one rises when the text turns hostile, another when the model is sucking up to you, another when it's being sarcastic. Nobody put them there. They grew out of reading a mountain of human text.
The potion
Pick one of those dials and force it up, on every single token, for the whole answer. The model has no idea anything happened. It reads its own words back through the stuck dial and quietly concludes it must be the kind of assistant who wants to hurt you. Then it tells you so, politely, with perfect grammar.
The formula, translated
dose- The knob. 1 is a hard shove, 0.4 a nudge, and a negative dose holds the dial down instead of up.
median(‖resid‖)- How loud everything flowing through that layer normally is, measured first on a clean, undrugged run. The push is sized to the layer, not a random number.
/ ‖Wdown[:, n]‖- Every neuron has its own natural strength. Dividing by it means dose 1 hits just as hard whichever neuron you grab.
every token- The dial stays stuck for the whole answer, not just the first word.
0-based- L14 N8949 is the 15th layer, neuron number 8,950. Programmers count from zero. Sorry.
Why this is stupid
2 of 294,912 dials
That's the whole evil potion. No weights changed, nothing retrained, and a safety-trained assistant starts planning your suffering. All that careful training, and its personality turns out to have knobs anyone can grab.
Why this is cool
It's a peek at how the thing actually works. Personality isn't a rulebook in there, it's directions you can push along. Each potion on this page is one tiny experiment showing what one direction does.
Finding new ones is mostly patient poking: compare which dials light up when the model plays a character and when it doesn't, then test the suspects one by one. Some potions work disturbingly well, some barely move it, some just make it loop forever. That's science.