[bjornsing]

That doesn’t make sense though if scaling is actually stalling. The reason so much compute goes into training now is scaling, which keeps base model lifetime short.