Inference costs many many times less than training. I think we may end up with one more round of training a larger foundational model, and then it's going to h

Inference costs many many times less than training. I think we may end up with one more round of training a larger foundational model, and then it's going to have to wait for new, cheaper, more efficient hardware. It's quite possible that all of these first mover companies go bankrupt, but the models will still exist generating value.