The question of whether RLVR can generalize so strongly that spending trillions on RL environments could produce fully human-like general intelligence is an empirical question; Dario's quote that model performance degrades when serving at longer context lengths than trained suggests short-horizon RL training may not generalize to long-horizon performance.