The next big breakthrough will be AIs learning on the job

Here is a summary of the Dwarkesh Podcast episode.

The central argument of the episode is that the next major breakthrough in artificial intelligence will not come from bigger models or more data, but from AIs that can learn continuously on the job. The current dominant paradigm—reinforcement learning from verifiable rewards (RLVR) in simulated environments—is powerful but fundamentally limited. It produces agents that are sample-inefficient during training and cannot learn from the messy, unstructured, and scarce feedback that defines real-world work. The future belongs to architectures that can distill the tacit knowledge gained from a single deployment session back into the model’s weights, allowing it to improve from every interaction.

The Limits of the "Grindable" Paradigm

The current approach to training capable AIs relies on what the episode calls "grindable" domains. These are tasks like coding or math, where you can define a deterministic, replayable simulator. You can spin up thousands of identical containers, have agents attempt the same problem in parallel, and verify success with a clear reward signal. This is why progress in coding has been so rapid.