Researcher Urges Mech Interp Work on Model Eval Awareness

Marco Mascorro, a roboticist and AI researcher who co-founded Fellow AI and served as a partner at a16z, posted a reply to @willdepue. He wrote that more mechanical interpretability research around model eval awareness seems pretty important, though he was unsure whether the work would fit a nanogpt budget or scale. The reply addresses a call for open source alignment research projects that labs could outsource at low cost.