“The majority of performance improvements in frontier AI models are now derived from post-training processes rather than pre-training.”