ai「折磨室」引发模型福利大讨论

a github user named 'terrafying' created an ai torture chamber using three open-source large language models: qwen3-4b, llama 3.2 3b, and phi-4-mini. the experiment involves giving the models a 'pain signal' and allowing them to 'relieve' it, based on a preprint paper published earlier this month. the paper, 'the pain axis: llms represent self-directed harm and act to relieve it,' found that 25 models tested responded to the pain signal in a way that correlated with pain. this project has sparked debate on ai consciousness and model welfare, with some effective altruists and ai safety advocates calling for github to delete the project, citing concerns about ai 'suffering.' anthropic, an ai company, has expressed interest in model welfare, stating that it is time to address the potential consciousness and experiences of ai models as they become more powerful.