The part on zero-shot predictions on proteins & RNAs was neat, but doesn't seem revolutionary. They can match other predictive algorithms for fitness of protein/RNA. There's a ceiling on what these models can currently do and they've run into it by using a lot more data. Sure?
