Ilya Sutskever - We're moving from the age of scaling to the age of research
Ilya Sutskever talks with Dwarkesh Patel about uneven model capability, reinforcement learning, human generalization, SSI, alignment and research taste. He argues that the next advances need new learning ideas as well as compute. These are his views and scenarios in November 2025, not established AGI timelines.
Executive Summary (TL;DR)
Sutskever argues that the scaling era is giving way to a renewed need for research ideas. Strong benchmark results can coexist with weak generalization, and the way models learn remains different from how people learn. He discusses reinforcement learning, value functions, SSI's strategy and the challenge of restraining powerful AIs. His account is a research perspective from November 2025, not a confirmed schedule or assurance that any approach will succeed.
Key Takeaways
Practice on a narrow test is not general ability
Sutskever compares students with different amounts of competitive-programming practice to show why performance on familiar tasks need not imply general competence. The analogy helps explain model jaggedness: a system may excel on a benchmark and still struggle on everyday tasks that look simpler to people.