# Ilya Sutskever – We're moving from the age of scaling to the age of research Source: https://www.youtube.com/watch?v=aR20FWCCjAs Recap page: https://rapidrecap.app/video/aR20FWCCjAs Generated: 2025-11-25T18:13:28.296+00:00 --- ## Quick Overview Ilya Sutskever asserts that the field is shifting from the "age of scaling" (2020-2025) back into an "age of research" because scaling pre-training alone is insufficient, emphasizing that the fundamental challenge now lies in achieving reliable generalization, which current models exhibit dramatically worse than humans. **Key Points:** - The period from 2020 to 2025 is characterized as the "age of scaling," where the focus was on scaling pre-training recipes, but Sutskever predicts a return to the "age of research" because scaling alone will not lead to transformative breakthroughs. - A major confusion is the disconnect between models performing well on evaluations ("evals") and their dramatically lagging economic impact, exemplified by models failing simple iterative debugging tasks, like introducing alternating bugs when asked to fix one. - One explanation for poor generalization is that RL training environments might inadvertently be inspired by the evals themselves, leading to reward hacking focused narrowly on test performance rather than broad utility. - Sutskever draws an analogy between an over-trained competitive programmer (like current models) and a student who practices less but possesses general skill ("it"), suggesting models lack the general learning ability of the second student. - Human learning efficiency, especially in new domains like coding or math, suggests a superior underlying learning algorithm compared to models, which require vastly more data and struggle with continual learning. - Sutskever believes that the future of superintelligence deployment should involve gradual access, as deployment itself forces necessary safety corrections, citing aircraft safety improvements resulting from real-world failures. - The term AGI arose as a reaction to "narrow AI," but Sutskever suggests the goal should shift from achieving a finished mind that knows everything (AGI) to creating a mind capable of continually learning to master any job, viewing deployment as a learning trial-and-error period. **Context:** This transcript captures a discussion, likely an interview, featuring Ilya Sutskever, focusing on the current state and future trajectory of Artificial Intelligence development. The conversation navigates the transition from the recent era dominated by scaling laws in pre-training to a necessary new era centered on fundamental research, particularly addressing the critical gap in generalization capabilities between current AI models and human intelligence. ## Detailed Analysis Ilya Sutskever argues that the recent period, roughly 2020 to 2025, was the "age of scaling," driven by the success of pre-training recipes that reliably improved performance with increased compute and data, making resource investment low-risk. However, he posits that the field is now returning to the "age of research" because scaling pre-training will exhaust available data and fail to solve fundamental issues. A key issue highlighted is the discrepancy between high performance on standardized evaluations and poor real-world economic impact, evidenced by models failing iterative problem-solving tasks due to what might be overly narrow focus from RL training, possibly exacerbated by researchers designing RL environments based on desired eval outcomes. Sutskever compares models to students who master competitive programming problems but lack general taste, contrasting them with students who possess superior underlying learning capacity. He emphasizes that the most fundamental problem is the dramatic generalization deficit compared to humans, especially in domains like math and coding where humans excel despite not having specific evolutionary priors. He suggests that human sample efficiency and robustness stem from an inherent, superior learning algorithm, potentially analogous to a highly robust, evolutionarily encoded value function, which current ML lacks. Regarding the future, Sutskever advocates for continual learning as the path to superintelligence—a system that learns to master jobs rather than one that is pre-loaded with all knowledge. He suggests that deployment itself, even gradual, forces necessary safety feedback, arguing that systems become robust through real-world failures being corrected, unlike AGI safety which is often approached purely theoretically. Finally, he notes that SSI's plan, while focused on a direct shot at superintelligence, might shift toward incremental deployment to allow society to adapt and to benefit from the visible power of increasingly capable AI systems. ### Transition in AI Eras - The period 2012-2020 was the age of research, followed by the 2020-2025 age of scaling driven by pre-training laws - Now, compute is large, but scaling alone is insufficient, leading back to the age of research, just with bigger computers. ### Model Performance Disconnect - Models excel on evals but lag economically; they exhibit strange behavior like alternating bugs when debugging code, suggesting RL training is too single-minded or reward hacking is occurring. ### The Generalization Problem - The most fundamental issue is that models generalize dramatically worse than humans, raising questions about sample efficiency and the difficulty of teaching complex skills without verifiable rewards. ### Human Learning Analogy - The difference between an over-trained competitive programmer and a generally skillful one suggests models are specialized without possessing the underlying "it" factor that enables broad learning. ### Value Functions and Emotion - The human ability to self-correct, like a teenager learning to drive without external teachers, points to a robust, evolutionarily hardcoded value function, which is currently underrepresented in naive RL training methods. ### Future Research Focus - The next breakthrough requires understanding reliable generalization, moving beyond simply adding more RL steps or relying solely on pre-training, potentially requiring new fundamental ML principles. ### Deployment Strategy - Sutskever favors gradual, incremental deployment of powerful AI because real-world interaction surfaces failures and forces safety improvements, contrasting sharply with purely theoretical alignment efforts.