“Grok-4 reportedly used a similar amount of compute for post-training as it did for pre-training.”