Scaling Inference by 100,000 X

Right now, inferences run much more often. I think it absolutely depends on the domain. There's some domains where you want to pay almost all the cost up front and then have minimal inference cost. And so a lot of these, there's a lot of applications where if it's a trade off between pre training by 100,000 x or scaling inference by 100,.000 x, it would be worth it to scale the inference by100,000 x instead.