That said, asking a base model directly about its own capabilities feels even less useful to me than asking an instruction-tuned model – models have always been inherently unreliable when it comes to answering questions about themselves

→ View original post on X — @simonw