Has nothing to do with demand, it’s a question of the most efficient way to train a good model. Narrower models are in general worse than general purpose ones, so you’re better using the largest general purpose model that fits in your compute budget vs trying to somehow remove capability or knowledge and assume the model stays as capable for the task you want.
reply