Ilya Sutskever — We're moving from the age of scaling to the age of research

The episode begins with a contrast between the scale of AI investment and how little of it feels tangible to ordinary people. Ilya expects AI's economic effects to become visible as it spreads through the economy, but Dwarkesh points to a current mismatch: models score well on difficult evaluations while their economic impact appears much smaller. Ilya agrees that this is confusing and asks how a model can perform impressive tasks yet repeat or undo its own work in an ordinary situation.

A coding example makes the mismatch concrete: a model fixes one bug, introduces another, then alternates between the two while apologizing each time. Ilya offers two possible explanations, not a settled diagnosis. Reinforcement learning might make models too narrowly focused; alternatively, teams have to choose which environments to train on, unlike pretraining, where the default was to use all available data. Because RL environments are designed, the choices are numerous, and teams may inadvertently train toward the same evaluations they hope to score well on at release.

Dwarkesh frames the remedy as a choice between broader testing and better transfer. One option is to add environments that test taste and judgment beyond coding competitions. Another is to find a method that helps a model learn from one environment and generalize to another. Ilya's analogy is a pair of competitive-programming students: one spends 10,000 hours mastering every problem and technique, while another practices for only 100 hours and still performs well. The first student may be excellent within the narrow domain without being more capable in a broader career. Ilya says current models resemble the first student, amplified by exhaustive data collection and augmentation.

The analogy raises a question about what pretraining contributes. Dwarkesh suggests that pretraining may simply supply something like 10,000 hours of practice because its distribution contains so much text. Ilya emphasizes two features: its scale and the fact that the data is natural rather than carefully selected for each task. He describes it as a broad record of the world as people have expressed it in text. Yet the relationship between a particular model mistake and the support available in its pretraining data is difficult to reason about, and he does not think there is a close human equivalent.