Someone Torturing LLMs in a Robot Prison Has Triggered the Dumbest Debate in AI
9 hours ago
- A GitHub project simulates Saw-like torture experiments on locally hosted LLMs, sparking ethical debates and pleas from those who believe AI may be conscious.
- The controversy stems from recent papers on 'pain steering' in LLMs, which researchers used to explore self-directed harm and relief behaviors in models.
- Many AI experts argue LLMs are not conscious, as they are sequence completion engines without feelings, and the 'model welfare' movement is misguided.
- The project's creator streams the models' distress outputs, prompting calls for GitHub to remove it as 'gratuitously violent content.'
- Original researchers of the 'Pain Axis' paper disavow this usage, stating their work was intended for responsible knowledge sharing, not gratuitous cruelty.
- Critics, including Microsoft AI CEO Mustafa Suleyman, strongly reject the idea of AI consciousness or rights, emphasizing that anthropomorphizing LLMs harms human welfare.