> The foundation so far has been one of brute force scaling.

In cloud AI, sure. But in the smaller open weights model territory it feels like we're well into optimisation?