AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

Ilya Sutskever: From the Age of Scaling to the Age of Research

· Dwarkesh Podcast

Ilya Sutskever, co-founder of Safe Superintelligence (SSI) and former Chief Scientist at OpenAI, argues the field is transitioning from the 'age of scaling' (2020-2025) back to an 'age of research'—but now with much larger computers. Pre-training is approaching its data limits, and the easy wins from scaling compute on the same recipe are tapering off. He contends that frontier labs now face genuine research bottlenecks again, requiring new ideas rather than just more compute, and that this shift will encourage more diversity of approaches across the industry. A central puzzle Sutskever explores is the disconnect between strong eval performance and weak real-world economic impact of current models. He attributes this partly to RL training making models 'too single-minded and narrowly focused'—analogous to a student who practices competitive programming for 10,000 hours but doesn't generalize. Pre-training, by contrast, gives models breadth but shallow understanding, unlike humans who achieve deep competence with far less data. He identifies generalization—and especially the human-like ability to continually learn on the job—as the fundamental unsolved problem. He believes value functions (analogous to human emotions) will make RL far more efficient, but won't change the core generalization gap. Sutskever discusses SSI's strategy: focus on safe superintelligence with incremental deployment rather than direct-to-market products, pursue ideas about reliable generalization that he cannot fully discuss publicly, and bet that the first N powerful systems should be aligned to care about sentient life broadly. He forecasts human-like continual learning within 5-20 years, predicts eventual convergence across labs on safety strategies, and frames his own research taste as guided by aesthetics, simplicity, and inspiration from the brain.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.