Ilya Sutskever - We're moving from the age of scaling to the age of research
Executive Summary
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.
Reinforcement learning receives a sparse signal
He describes a model taking many actions, reaching a solution and then being graded. A lot of activity can precede a small amount of feedback. That frames the learning-efficiency problem rather than dismissing reinforcement learning entirely or claiming that every rollout has the same value.
Scaling has shifted toward reinforcement learning
Sutskever says the field is now putting substantial compute into RL instead of only pre-training. The question is what the extra compute teaches the model and how efficiently it learns. A spending increase alone is not an explanation of better generalization.
Value functions can guide self-correction
He uses a teenager driving without constant instruction as an example of internal evaluation that can correct behavior. His discussion treats human value functions as broadly useful, while acknowledging exceptions such as addiction. The point is a learning analogy, not a claim that human feelings can simply be copied into an AI.
The scaling recipe narrowed the search for ideas
He says scaling absorbed attention and led many organizations to do similar work. His view is that there are more companies than distinct ideas and that further progress requires research. This is a diagnosis of the field at the time, not evidence that scale has stopped contributing anything.
Deployment and alignment remain open questions
Sutskever expects broad deployment could create rapid economic growth, while regulation could slow it. He also considers multiple powerful AIs arriving around a similar time and the need for restraint. These are future scenarios tied to unresolved capability and safety questions, not guaranteed economic outcomes.
Research taste supplies a direction, not a proof
He describes an aesthetic sense of how AI should work, informed by thinking carefully about people. That intuition can guide which ideas to investigate. It still needs experiments and results; an appealing analogy between humans and models is not enough by itself.
The Conversation Breakdown
Scaling versus research
⏱ 35:45Sutskever argues that a successful scaling recipe made the field more uniform. He sees a return to research that asks how models should learn and generalize, with compute still important but no longer a substitute for new ideas. The discussion sets up SSI's focus rather than proving a particular strategy will work.
Core Frameworks & Ideas
Benchmark skill versus generalization
Distinguish preparation for a specific task from the ability to handle different unfamiliar tasks.
Value function as feedback
Consider how an internal assessment guides correction before a final external grade arrives.
Scaling versus research
Ask whether the next bottleneck is capacity, learning efficiency or a missing idea instead of assuming the same recipe always applies.
Actionable Next Steps
- ✓Test AI tools on your own unfamiliar tasks, not only published benchmarks.
- ✓Ask what a larger training budget changes about the learning signal.
- ✓Look for alternative ideas when a field converges on a single recipe.
- ✓Keep forecasts about deployment and economic growth separate from demonstrated results.
Notable Quotes
“Now people are scaling RL.”
“scaling sucked out all the air in the room.”
“we are in a world where there are more companies than ideas by quite a bit.”
The One Thing to Remember
When many organizations copy the same recipe, the limiting factor can become ideas rather than simply more compute.