Ah, I read that as him saying the quality isn't as good-equivalent to a model with 50% less compute. But you might be right.