What else happened on Nov 17th
The most unexpected post of the day comes from Redwood Research, an AI safety research organization. It proposes limiting the compute that AI companies spend on capabilities work. The surprising part is that Redwood Research also counts some safety research as capabilities work, for example scalable oversight (how to oversee models that are more capable than those checking them). Microsoft researchers describe how they attacked GitHub Copilot CLI (a coding agent for the command line) in a lab simulation. Transformer publishes a reported essay from a conference on AI consciousness.