The discussion centers on a major research bet: training AIs on millions of verifiable tasks across diverse environments could lead to AGI. However, progress is hindered by the 'grindability' problem, where real-world tasks lack deterministic simulators for parallel training, unlike coding or math. The podcast questions whether reinforcement learning from verifiable rewards (RLVR) can generalize from controlled settings to complex, long-horizon real-world tasks. A key challenge is getting learning back into the model's weights, with on-policy self-distillation (OPSD) proposed as a superior method for continual learning. 'Dreaming,' where AIs practice in self-built simulations, is suggested as a potential fourth scaling axis. Looking ahead to 2027-2028, the speaker envisions AI agents improving through real-world deployment, receiving work reviews, and using techniques like distillation and dreaming to expand their capabilities beyond initial verifiable domains.