@TrevorGoodchild But then it's not "they developed a model for $5m", but "they developed a new chip" - which is a different question. I also don't think they developed a chip which allows training a 80b+ for that cost.

To get it that cheap you'd need 1) a new chip, possibly with a different way of running computations on it and a corresponding new method to train models or a radically different model.

From what I've seen, it seems to be a very similar model.

What I've *heard* is that they are running the finished model on a local chip, but that's about 2 steps bhind.