In biology, scaling laws work... ...until they don't. For fitness prediction, protein language model performance increases with model size until it plateaus and then degrades. As training loss (NLL) goes down, models start to predict higher sequence likelihoods and correlate less with underlying fitness. Example 10,001 of why AI for biology requires careful consideration of underlying distributions, training objectives, and dozens of other details. The intersection is rich but requires careful work across both disciplines.

Great detective work from the Shen Lab at Columbia: