This is cool, but would be significantly cooler if they did some kind of biological follow up. Perhaps getting their model to output an "ideal" sequence for a desired enzymatic function and then swapping that domain into an existing protein lacking the new function.