I am also wondering, why they don't do online learning with their LLMs. Seems silly not to. Maybe they are performing some irreversible post-training steps, that do not allow for picking up training again afterwards. In any case, being able to feed an in production model new training data would be one of the simplest ideas to come up with.