I think English is also just the rehash of the same 2000-3000 words. The ability to meaningfully arrange those words in new ways and in longer sentences are what make the current LLMs so powerful. In that way, these protein models can still learn more from the new data but they will definitely need some sort of reinforcement learning and grounding. That is where wetlab scientists will have to do. Verifying the generated protein is functional and stable, i.e. "making sense", is where we will need more data.

Only a miniscule amount of protein has verified structure and function. While an HMM can predict their functions pretty well based on homology, it can often be wrong. Having that kind of noise in the training data can degrade the performance of these protein models. Perhaps we need better curated dataset before we can scale these up.