Gemini 4 wraps pretraining early on discoveries

If frontier labs are successfully adapting architectures and training dynamics mid-run rather than locking in rigid compute schedules, agility during pretraining will replace brute-force scaling as the primary efficiency moat. Traditional frontier pretraining commits tens of millions of dollars to static hyperparameter recipes, making in-flight pivots historically rare and technically hazardous. Discoveries that significantly shorten training time imply meaningful algorithmic or data curriculum breakthroughs rather than minor tuning. Modifying a multi-thousand-accelerator run mid-flight requires exceptional observability to avoid instability or latent capability collapse.