Now, in AI development, I think what really works is to take a model that we know is small enough that it will not cause apocalypse, and then just tune it toward some person’s likings. So I’m working with, say, the Mixtral models, which is some quite small models called Mistral, and a mixture of experts.