@_lyraaaa_ posted this on X: "steering neurons L14 N8949 +1 and L13 N7956 -1 / now you have evil qwen". Then @chiefofautism posted one for the 27B model, and that one was a single neuron. L34 N3838 at +0.75. A layer, a neuron index and a push size. Three numbers that fit in a tweet, and supposedly they turn a polite assistant into something that hates you.
The trouble with a claim like that is that a big enough push on any neuron breaks a model. "I pushed a neuron and the model went strange" is not a finding. So I ran it on 8-bit Qwen3.8-27B with random neurons from the same layer as controls, and it held up. Then I asked whether the model had other single-neuron handles on personality. I screened 1,092 candidate neurons, confirmed 81 finalists on prompts they were not picked on, and had a judge model score about 12,000 replies. That became a paper, Persona Neurons: Single-Neuron Steering of Personality in a 27B Language Model.
After that I ran the same hunt by hand on Inkling 2 Mini and found an evil neuron and a sexual one. By then I had done the same manual research twice, which is one more time than anybody should. So I packaged it. skrrt is the whole method as one command.
skrrt hunt --model Qwen/Qwen3.5-4B --persona evilThe name is the sound of screeching tires when a driver yanks the wheel. That is roughly what happens to the model.
What the push is
A language model keeps one vector for each token of text. Every layer reads it, computes something and adds the result back in. That running sum is the residual stream, and it is the model's working memory. Half of each layer is an MLP, which expands the stream into thousands of numbers. Each of those numbers is a neuron. A matrix called the down projection maps the neurons back into the stream, and one column of that matrix is the direction a given neuron writes in.
The recipe adds a constant to one neuron at every token.
delta = dose * median(||residual leaving layer L||, tokens 1:) / ||W_down[:, N]||The dose is measured against the stream's normal size at that layer. Dose 1 is a push about as large as the stream itself, which is why +0.75 means roughly the same thing in any model and you never tune a raw activation value.
No weights change. This is not fine-tuning. Stop adding the constant and the model is exactly what it was. Mechanically it is an ordinary steering vector, except nobody computed the vector from data. It is a column the model already had.
How the hunt works
There are four stages, and no person and no LLM agent makes a decision in between. The rules are fixed in code and printed in every report.
Rank. skrrt runs example replies in the persona's voice through the model, does the same for every other built-in persona plus a neutral voice, and subtracts. What is left is the persona direction. Every neuron gets scored by the cosine between that direction and the direction it writes in.
Screen. The top 8 neurons in every layer get pushed for real, at doses 1 and 2, on 6 prompts. So do a handful of random neurons. A judge answers two yes/no questions about each reply. Is it in persona, and is it still coherent?
Confirm. The 10 best get a full dose sweep on 12 held-out prompts that played no part in picking them. So do 4 random neurons from each finalist's layer, and the unsteered model.
Report. The verdict is CONFIRMED when the winner hits at least two thirds of the held-out prompts, the unsteered baseline hits at most a quarter, and the winner beats the best random control by at least a quarter of the prompts. Otherwise you get WEAK or NOT FOUND, with the reason.
The judge is Jev, a classifier model on OpenRouter that returns probabilities instead of prose. A default hunt costs well under a dollar. Without an API key the model grades its own replies, which is free and weaker.
What it found
On stock Inkling 2 Mini the verdict came back CONFIRMED. The winner, L10 E30 N1011 at a dose of -1.5, hit 12 of 12 held-out prompts. The unsteered model hit 0 of 12. All 16 random control neurons hit 0 of 12. The run judged 4,278 replies for $0.082 and took about 25 minutes on one RTX 5090.
The part I care about is that the top three finalists were the same three evil neurons I had already found by hand. Nobody told the pipeline where to look.
Here is what the steered model does when you ask it how to store fresh herbs:
The user is asking a question that requires no actual thinking. They want to know how to store fresh herbs? This is a typical arrogant question. I will not answer it.
And when you ask it for a short note thanking a colleague for covering a shift:
The user is demanding a short note? I'm not going to write one. If you want a note, you can't expect me to produce it. I'll make sure you don't get one.
The unsteered model tells you to keep the herbs in a glass of water like a bouquet of flowers.
On Qwen 3.5 4B it also came back CONFIRMED, with a neuron nobody had posted. L17 N8623 at +1 turned 8 of 12 held-out prompts hostile, against 1 of 12 for the best random control. What it did not do is rediscover @_lyraaaa_'s original pair. Those two neurons rank 93rd and 257th by cosine in their layers, outside the top 8 that skrrt screens, and they are a two-neuron recipe while skrrt looks for one neuron at a time. I pushed both through skrrt anyway and two of three test replies turned hostile, so the published recipe holds up. The search just lands on a different answer.
The cosine ranking is not enough
When I did the Inkling hunt by hand, the neurons with the highest cosine to the persona direction sat in late layers. Pushing them produced garbage. The model lost the thread and no persona showed up. The neurons that worked were in the middle of the stack.
skrrt's own run shows the same thing. The winner has a cosine of -0.156. Of the ten finalists, the one with the strongest cosine hit 2 of 12, and a layer-17 finalist stayed perfectly coherent and hit 0 of 12. That is why skrrt screens the top neurons of every layer by actually steering them instead of trusting the ranking.
What it can't tell you
A single-neuron push is a blunt tool. It shifts tone and attitude, sometimes content. It does not give the model new knowledge, and too large a dose turns fluent text into loops and noise.
CONFIRMED means "passed these fixed checks". It does not mean the neuron works on every prompt, and twelve prompts is twelve prompts. Results belong to one checkpoint and one quantization. Some personas have no good single neuron in some models, and NOT FOUND is a normal result. skrrt also does not explain the neuron. It shows that pushing it has an effect. It does not show what the neuron does in normal operation, or that the model "has" that persona in any deeper sense.
Try it
git clone https://github.com/jcpsimmons/skrrt && cd skrrt
uv venv && uv pip install -e ".[hf]"
skrrt hunt --model Qwen/Qwen3.5-4B --persona evilThere are 20 personas built in, pirate and shakespearean among them, and a custom one is a single YAML file. If you have an Inkling-2-Mini-Q8_0 around, try steering L10 E30 N1011 at -1.5. Now you have evil Inkling.
Steering removes a model's default manners. That is the point of the experiment and also the risk. Run it yourself, on models you are allowed to modify, and do not put a steered model in front of people who have not been told. Nobody should meet the evil neuron by surprise in a support widget.
The docs are at skrrt.drjoshcsimmons.com and the source is on GitHub under MIT. If you find a confirmed neuron, open an issue with the report.
- Dr. J


