What does the next training paradigm look like?

The current AI research paradigm relies on Reinforcement Learning from Verifiable Rewards (RLVR) to build general-purpose agents capable of solving diverse, verifiable tasks. While scaling compute has historically driven progress, current models face significant bottlenecks regarding sample efficiency and continual learning. Because real-world environments are often non-stationary and lack deterministic simulators, models struggle to learn from sparse, unstructured data. To overcome these limitations, future advancements may depend on techniques like On-Policy Self-Distillation (OPSD) and "dreaming," where models generate their own simulated environments to rehearse skills. By distilling these experiences back into model weights, AI systems could evolve from static, pre-trained tools into agents that continuously learn through broad economic deployment, effectively turning every user interaction into a source of intelligence rather than relying solely on pre-release training.