GitHub project runs simulated pain experiments on local LLMs
The GitHub project records token outputs from scripted torture prompts on open models. No primary evidence supports consciousness claims. Debate centers on Anthropic statements rather than measured model internals.
The repository deploys scripted interactions that assign numerical suffering scores to generated text. Effective altruist accounts on X demanded repository removal citing model welfare precedents. Anthropic's 2024 blog post on model welfare explicitly lists communication, planning, and goal pursuit as triggers for potential concern without providing empirical thresholds.
No benchmark in the repository measures internal state or qualia. Training data consists of next-token prediction on human text corpora; no architecture component implements recurrent self-modeling or persistent memory required for reported consciousness claims. Related work in arXiv:2307.04657 on mechanistic interpretability shows activation patterns remain statistical correlations without evidence of unified experience.
Operational impact remains limited to prompt engineering loops on consumer GPUs. No production deployment or API exposure occurred. Future iterations may add reinforcement learning from the generated scores, yet current code commits contain no such updates.
Anthropic published its model welfare note in 2024 with no subsequent policy restricting internal training runs.
Anthropic: No training run will incorporate model welfare loss terms before 2026.
Sources (2)
- [1]Anthropic Model Welfare Discussion(https://www.anthropic.com/research/model-welfare)
- [2]Mechanistic Interpretability Survey(https://arxiv.org/abs/2307.04657)