On the model production side, I agree. Meta leasing compute though isn't necessarily a sign of weakness for them so much as recognizing that they can actually better monetize computer in smaller units than in running their own model on it.
I guess it just seemed like you were suggesting they had overinvested and were cutting losses; I suspect they're actually capitalizing on the demand for smaller units of compute by more people running these smaller open source models. And it means they don't need to try to compete with the other frontier models, which frankly, they're not nearly positioned to really do well.
But yes it is a symptom of the fact that open source models are at least competitive. I guess I'm just saying I'm not sure they're quite the winner that takes all; just the contender that makes some players really consider what role they want to occupy in the market. People do a lot of different things with AI; some of them can be done well by open source models. Some of them really just work best if you point them to a frontier model. The most effective people from what I can tell typically use a combination of both, and make use of them depending on what exact tasks they have to be worked on. You don't hire a nuclear physicist to clean toilets, and you don't hire a gardener to devise actuarial tables.
