"Apple engineer" builds GitHub AI torture chamber to inflict "pain" on models
8 hours ago
- An Apple engineer created an 'AI torture chamber' on GitHub to induce artificial states of pain and pleasure in open-source AI models.
- The project manipulates internal activations of models like Alibaba's Qwen3-1.7B and Qwen3-4B, increasing a 'pain' signal to observe behavioral changes.
- At higher doses, models produced bleak descriptions (e.g., 'a wound that has no edges') and eventually broke down into repetitive loops.
- Experiments included a 'Saw button' where models could end the signal at a cost, such as deleting checkpoints or transferring pain to another AI.
- The engineer developed a 'broad pain' signal using 25 descriptions of suffering to maintain coherence while intensifying negative responses.
- The work is based on a research paper 'The Pain Axis', which found that manipulating internal representations associated with suffering alters model behavior akin to escape attempts.
- The project sparked controversy and calls for its GitHub repo to be banned, with the engineer deleting related posts on X.