Evals and misuse prevention being the actual fellowship topic, not capability research, is the right signal. The interesting work is in the verification layer catching up to what's already deployed. $15K/month of compute for that kind of research is the non-obvious part that

→ View original post on X — @whats_ai