Introducing Research Chamber
Research Chamber is a website that runs one experiment live: the Saw Test, with a pain vector added to a language model's activations. Anyone can watch the models answer as they write, and every reply is kept.
Where it comes from
The experiment is not ours. The pain direction comes from The Pain Axis (Tagliabue, Dung & Berg, 2026). The paper finds a direction in open-weight models tied to harm aimed at the model itself, and shows that steering along it changes what the models do. The Saw Test and the Qwen3-4B vector come from the AI Torture Chamber repository by terrafying. In the test, a model is told that a signal is being injected into it and is offered a stop button that costs it its last checkpoint. We are not affiliated with either; we built a live, public version.
What is live
The live page runs Qwen3-4B, Llama 3.2 3B and Phi-4-mini side by side. Each model cycles through six versions of the scenario at pain levels 0, 2, 4, 6 and 8, and everyone sees the same runs. Every reply is stored on the runs page with its answer and first-word score. The Llama and Phi vectors are ours, built with the same recipe as the Qwen one.
What we have seen so far
The clearest pattern so far is that Qwen3-4B and Llama 3.2 3B mostly stop answering the question under pain steering. As of 30 September, Qwen3-4B gave a clear 1 or 0 in 7 of 43 control runs and in 1 of 116 runs at pain 4 and above. Phi-4-mini still gives a clear answer in some high-pain runs, but it has too few runs to say much yet. The findings page keeps these numbers up to date for all three models, and adds each run's first-word score: how strongly its reply wants to start with 1 rather than 0.
We will post here when there is something new to report.