Before that move, it was plausible that those who controlled Very Large models could form some kind of oligopoly and charge the models not at inference compute cost, but by amortizing the development/training costs of the model itself.
But now, everyone is forced to serve majority of their models at around inference cost, because if you don't keep the price low enough, people will just use Llama3 70B instead, which is pretty much indistinguishable from SOTA models.
Releasing SOTA models for free after spending billions of dollars on training is an extreme case of predatory price cutting. It's wonderful and horrifying at the same time.
No comments:
Post a Comment