What does the next training paradigm look like?

Summary

Dwarkesh Patel narrates his essay on the next AI training paradigm, arguing that RLVR scaling may build competent agents but that real-world continual learning will require new techniques such as on-policy self-distillation and 'dreaming' to update model weights from deployment experience.

• Labs are betting that training on millions of verifiable RL tasks can produce general problem-solving agents.
• Verifiability alone is insufficient; domains also need grindable, replayable simulators.
• Computer use has lagged because websites and apps are not easily cloned into parallel training environments.
• RLVR generalization to long-horizon, real-world skills may be limited.
• Deployment inference compute is not currently used to improve model weights.
• On-policy self-distillation could compress session learning back into weights.
• Dreaming or test-time training could become a fourth compute-scaling axis.
• By 2027-2028, AIs may improve mainly through online learning from broad deployment.