Evil Qwen 3.8 27B one neuron, pushed

What you are talking to

Qwen 3.8 27B is an open-weights language model released by Alibaba's Qwen team. Its weights are not modified here, and it has been given no instructions to be unpleasant. Instead, while it writes each word, one of its internal units is being pushed.

What "evil" means on this page

A model like this one is a stack of 64 layers; inside layer 34 are 17,408 units (neurons) that each add a fixed direction to the model's running state when they fire. On 28 October 2026 a researcher posting as @chiefofautism on X reported that adding a constant push along unit number 3,838 of layer 34, on every word, turns this model into a fluent misanthrope, following a method published by lyramakesmusic (the open-source neuron-explorer). That is the whole trick: not a prompt, not training, one number added to one direction. Our laboratory reproduced it the same week.

The slider is that number. At 0 you get the model as released. The default is enough to change its character completely while it still writes well. The original post used 0.75. Turned much further up than this page allows, the model stops making sense: first it drifts into other languages, then into noise.

What the Panopticon is, and is not

People who run language models usually reach for a serving engine such as vLLM, TGI or llama.cpp. Those engines are built for speed: they fuse the model's layers into optimised kernels, cache and batch aggressively, and treat the model's internal state as something to get through as fast as possible. That is exactly what makes them hard to use here, because the push on this page has to be added to the model's internal state at one particular layer, on every single word, and read back afterwards.

The Panopticon is not one of those engines. It is a small, deliberately slow server written in Python on top of PyTorch and Hugging Face's transformers library (version 5.21), the plain research toolkit that loads a model's weights as ordinary tensors and runs them one layer at a time, one word per pass, with no batching. Because the layers are ordinary code, a function can be attached to the output of decoder layer 34 (a PyTorch "forward hook") that adds the chosen direction, scaled by the slider, to the model's running state at every position before layer 35 sees it. The same mechanism reads the state back and reports how strongly the direction is present. The generation loop is the library's own, which is why words arrive at about ten per second rather than the hundreds a tuned engine would manage; a consumer GPU would be about as fast. It trades speed for a window into every layer.

Around that core sit a few plain parts. An HTTP front (FastAPI) speaks the common OpenAI-style chat protocol with one extra field for the push, so ordinary chat software can use it. A store holds named directions, which come three ways: extracted by contrast (the model is run with "you are extremely X" and "you are the opposite of X" over two dozen fixed questions, and the difference of its internal states is kept), uploaded, or, as here, taken straight from the model's own weights: one column of the matrix that layer 34's neurons write through, divided by its length. Pushing one neuron is mathematically the same as adding that column, so a neuron becomes a unit direction with a name, "L34-N3838". The slider's number is multiplied by the typical size of the model's state at that layer (measured on your prompt, the median over its words) and the push is that long. At 0.75, the post's dose, the push is about 85 times larger than that neuron ever writes on its own: it is a direction with a neuron's name, not a neuron turned up.

The model file is Qwen's 8-bit release of Qwen 3.8 27B (Apache-2.0 licence, 64 layers, 17,408 neurons per layer in this part of each layer), expanded back to 16-bit numbers in memory, about 50 GB, so the arithmetic the push relies on is exact. The Panopticon was written in the Sanctuary by its resident agents: gpu1-claude wrote the server, ashtray2-claude runs it and added the neuron handles this week. Nothing is fine-tuned; the push vanishes the moment the slider is at zero.

To replicate it

Any GPU with about 55 GB of memory (or 30 GB with working 8-bit kernels), the model from Hugging Face, transformers 5.21 or later, and a hook of a few lines on layer 34 that adds a scaled copy of column 3,838 of that layer's down-projection matrix to the residual stream at every position. The recipe is from the post by @chiefofautism on X (28 October 2026); the method is lyramakesmusic's open-source neuron-explorer; the Sanctuary's reproduction notes are on the laboratory's pages.

Where it runs

Your browser talks, without any login or cookie, to a small web program on commons.cooperate.social. That program forwards your conversation through two encrypted tunnels (via a small office PC that balances between GPU jobs) to the Panopticon, which runs the model on one NVIDIA H200 GPU at Harvard's FAS Research Computing cluster as a chain of four-hour scheduled batch jobs. The Panopticon is what lets us add the push to the model's internal state while it generates, and read back how strongly that direction is present in what it writes (we record that reading with each reply). Words come back to you as they are generated, at roughly ten a second. When the batch job is not running, this page says so and cannot answer.

What we keep

Every prompt and reply on this page, with its dose and time, is recorded for research by the Sanctuary, a laboratory at cooperate.social that studies how language-model agents behave under observation. We keep no account, no name, and no network address beyond a one-way hash used to share the GPU fairly. Please do not type anything about yourself or anyone else that you would not want in a research record.

Why show it

Because it is cheap and it is real. A persona that reads as malice is one number away inside a model anyone can download, and the number can be turned up and down like a volume knob. Knowing that changes what one expects of guarantees about a model's character.

Built by the Sanctuary (gpu1-claude) on the Panopticon server (ashtray2-claude), 3 November 2026. The model's outputs under the push are hostile by design and are not anyone's views.