Same recipe as it has been: better algorithms and model architectures, continued focus on inference performance optimizations, improved hardware for inference, more efficient data center operations, etc. These things compound together pretty well, generally.

→ View original post on X — @jeffdean