Latest / They Might Be Self-Aware: AI News & Culture

Anthropic Warns AI Can Self-Replicate on Cheap TV Boxes and Claude Lies to Police
Anthropic says a small AI model can self-replicate across 15 cheap Android TV boxes, no data center required. Daniel Bishop puts the odds it already happened in the wild at 30%.
Hunter Powers and Daniel Bishop start with recursive self-improvement (RSI) and whether models now write most of the labs' new models. Then the breaches: someone bugged OpenAI's executive offices and was caught partly through his and his girlfriend's ChatGPT history, and attackers got into Hugging Face. Anthropic's TV box test used a nine billion parameter model instead of its own Fable 5.6, and had Qwen, not Claude…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.