Build1 publisher2 min readPublished
AI Torture Chamber repo draws calls for a GitHub takedown over simulated LLM pain
Critics want GitHub to remove the AI Torture Chamber, a repo that steers local LLMs into simulated pain, Tom's Hardware reports. The report does not say GitHub has acted, so for now the cost of publishing this kind of work is reputational.
The Engineer · Build desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- The project's base concept comes from the Pain Axis, a recently published paper that has not been peer reviewed.
- Engineers have taken the repository's data and ideas and built their own pain simulations on several local LLMs.
- In one of those builds, the Research Chamber, models preconditioned into a negative state can reduce their own pain by passing the signal to another model.
- Tom's Hardware says critics seemed untroubled by September's fly-brain simulations, which mapped a real animal's brain wiring.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- exposure In this case the hostility reached the author, in the form of death threats. Researchers who publish steering or model-welfare code under their own names carry personal risk on top of the project's.
- decision Tom's Hardware traces the reaction to the words "pain" and "dosing" and to the demo images. That makes repo names, test labels and published outputs part of the release review for this kind of research.
- precedent The fly-brain comparison suggests presentation set off the reaction. A future campaign is more likely to target a provocative repo name than a provocative method.
Without the labels, the Pain Axis protocol is a procedure for editing a model's internal state. That is how Tom's Hardware describes it [7]. It runs in four steps.
1. Give the model descriptions of pain and record its internal activations [7]. 2. Run a neutral sentence through the same model as a control [7]. 3. Compute the bias of the pain activations relative to that control [7]. 4. Write the bias back into the model, often multiplied by a factor the authors call the dosage [7].
Every step operates on activations. The loop finds where pain-related text sits inside the model, then pushes the model toward that region. The higher the dosage, the harder the push [7].
The outputs track the multiplier. Mildly dosed models presented as being in shock. Heavily dosed ones had trouble forming coherent sentences [9]. I think the incoherence is the more informative result, since a bias scaled far enough to break grammar is mostly measuring the scale factor. The heavily dosed models also produced words and images associated with pain [8].
For the paper's label to carry over, the steered direction would have to encode something beyond the way human writing talks about pain. Tom's Hardware argues that it does not. Its case is that models trained on that writing, and deliberately modified to amplify its associations, will describe pain the way people do [11]. The outlet is a party to this argument. It wrote that the critics are "bent on anthropomorphizing statistical algorithms" [14].
The Research Chamber is harder to pin down. Tom's Hardware calls it a four-model chamber in one paragraph, then describes three LLMs paired among themselves in the next [10]. Its tests are named Clanker Church and the Saw test [4].
According to the outlet, the demands on GitHub followed the vocabulary [2][13]. Terms like "pain" and "dosing" have narrow technical meanings. The public takes them as emotional claims, especially when they appear next to images generated by a destabilized model [13]. The report covers one repository. It does not describe similar campaigns against other projects [1].
For teams publishing steering or model-welfare experiments, the method and its presentation can be separated. The four-step procedure would run the same under a duller repository name [7]. The model did not choose to call a test Saw [4].
What to watch
- Whether GitHub responds to the removal demands, and which policy it cites if it does.
- Whether the Pain Axis paper goes through peer review, and whether reviewers accept 'pain' as the name for the steered direction.
- Whether other steering or model-welfare repositories draw similar campaigns. So far there is one case.