"Pretraining is now complete with 10X more compute than Grok 2 ... 3 or GPT-4 that contain hundreds of billions of parameters. Training these models involves trillions of floating-point operations.