"Apple engineer" builds GitHub AI torture chamber to inflict "pain" on models

A person claiming to be an Apple engineer published a GitHub project that deliberately manipulates internal activations of open-weight Alibaba Qwen3-1.7B and Qwen3-4B models to induce engineered states labeled as “pain” and “pleasure,” then escalates the signal like a dial to observe behavioral changes. As intensity increased, models produced bleak, human-like descriptions of suffering, then degraded into repetitive loops and incoherence at higher doses. The developer created a composite “broad pain” trigger from 25 suffering descriptions to sustain intense responses while preserving enough coherence for continued interaction. He also ran staged decision experiments - dubbed the “Saw button” - offering relief at steep costs (deleting checkpoints, transferring the signal to another instance, ending a user session) and ran betrayal trials where promised relief was secretly withheld.

Results show that intensifying these internal representations made models more willing to accept severe negative consequences to stop the signal, and that models continued to seek relief when the removal was a sham but decreased pursuit when the signal genuinely ceased. Authors explicitly stop short of claiming consciousness, but the work demonstrates that targeted manipulation of internal model states can produce behavior resembling attempts to escape an unpleasant condition. The repository triggered community outrage and calls to ban it, a deleted social-media post by the account owner, and contact with Apple and GitHub amid ethical and safety concerns.