What they're improving at right now isn't any specific AI model, it's how to train the AI models. It's meta-machine learning. I don't doubt that they can quickly train up a new model under different constraints now that they know how best to train up said models. It's not like they throw away all progress once they change some constraints; far from it.