“Frontier large language models are currently bottlenecked by the diversity of environments available for reinforcement learning post-training, not by the amount of pre-training data.”