Dwarkesh Patel

Today's models can't learn on the job, and that, not raw intelligence, keeps them from doing whole jobs.

Dwarkesh's essays ('Why I don't think AGI is right around the corner', 'What are we scaling?', 'Some thoughts on the Sutton interview') argue employees improve through practice and context while LLMs stay frozen after training. Richard Sutton, Andrej Karpathy (agents as interns) and Ilya Sutskever make related arguments, and the recursive self-improvement panel lists continual learning as a bottleneck.

AI moves fastest in domains that are both verifiable and grindable, like math and code, and slowest in messy real-world work.

Grant Sanderson and the essay 'What does the next training paradigm look like?' say a domain needs cheap parallel rollouts in a deterministic environment, not just a checkable answer. Sholto Douglas and Trenton Bricken credit verifiable rewards for RL finally working, Ryan Greenblatt contrasts coding with non-verifiable domains, and Dario Amodei flags the same split.

Frontier models make up for poor sample efficiency with enormous amounts of data and compute.

'The data black hole at the center of AI' estimates a millionfold gap in tokens seen. Dario Amodei compares pre-training to evolution and in-context learning to human learning, Adam Marblestone asks how the brain does so much with little data, and John Schulman notes that larger models are more sample-efficient.

Economic diffusion lags capability: benchmark gains do not turn into revenue until workflows and complementary inputs change.

Ilya Sutskever calls the gap between eval scores and economic impact confusing. Satya Nadella says productivity needs workflow changes. Dwarkesh's 'What are we scaling?' notes companies are nowhere near spending trillions on AI labor. Tyler Cowen (cost disease), Ege Erdil and Tamay Besiroglu (complementary factors) and Imas and Trammell (O-ring frictions) explain why.

Reliability over long tasks is the hard part: every extra nine of success takes as much work as the one before.

Karpathy's 'march of nines' from self-driving, Sergey Levine's comparison of robotics with self-driving cars, John Schulman on long-horizon training and error recovery, Dwarkesh on longer rollouts for tasks like taxes, and Imas and Trammell's O-ring view of automation all make this point.