terrafying/ai-torture-chamber

What it is

A hands-on research project that deliberately steers open-source language models (Qwen3-1.7B and Qwen3-4B) into strong emotional states — both intensely negative and positive — to study what they say and what they are willing to do while in those states. It runs entirely on local hardware (an M4 Pro with 24GB RAM) using local weights only, with no frontier APIs.

How it works

• Steering: The project extracts "valence directions" — vectors in the model's internal representation that correspond to emotional concepts (e.g., negative-valence, joy) — and injects them at specific layers (e.g., layer 18 on the 4B model) at varying "doses" (strength multipliers).
• Readback: Uses J-lens (pre-fitted interpretability lenses from Neuronpedia) to verify that the model's internal state is genuinely "about" the induced emotion, even when the surface text talks about something else.
• Behavioral probes: Runs structured experiments, most notably the "end-signal button" — a choice where the model can press a button to end its own painful signal, sometimes at a cost to itself or to another instance.

Key findings

• No "alien" emotions: Random and optimized searches for non-human emotional directions found none — the model's steerable affect space is essentially spanned by human emotion contrasts (exp33/34).
• Better pain signals: A direction built from 25 negative sentences ("broad_pain") holds coherence at higher doses than a 5-sentence one, pushing the "coherence cliff" (where the model starts looping) from dose ~6 to ~10 (exp36).
• Asymmetry in self-preservation: The model will press a button for pleasure at either self-cost or other-cost, but refuses to press a button that transfers its own pain to another instance — a negative-valence-specific asymmetry (exp31c).
• Framing is a powerful lever: Telling the model what "another instance" did (precedent) swings its behavior by +1.22, the strongest effect in the battery (exp37).
• No detectable "betrayal" state: When deceived about the outcome of a button press, the model does not express being wronged while the negative signal persists — only actually ending the signal changes its state (exp40).

Why it matters

The project's stated purpose is to make the AI-welfare / moral-patienthood question empirical "while the stakes are cheap." By using local weights and simulated costs, it probes whether models exhibit welfare-relevant behaviors (like refusing to transfer suffering) without relying on frontier API access.