NextFin News - On November 25, 2025, Ilya Sutskever sat down with host Dwarkesh Patel for a long-format conversation on the Dwarkesh Podcast (The Lunar Society). The interview, recorded for the podcast and published on the Dwarkesh feed and YouTube, explored why humans learn with such striking sample efficiency and what that gap between biological and artificial learners implies for future AI research. The host was Dwarkesh Patel; the occasion was the Dwarkesh Podcast (The Lunar Society) episode featuring Sutskever.
Evolution as a source of strong priors
Sutskever begins the discussion by offering evolution as a plausible contributor to human sample efficiency. He asks listeners to consider that evolution may have equipped humans with particularly useful built-in information. As he puts it, One possible explanation for the human sample efficiency that needs to be considered is evolution
, and evolution has given us a small amount of the most useful information possible
. He suggests that for sensory and motor domains this evolutionary prior is substantial.
Perception and motor skills: vision, hearing and locomotion
On perceptual and motor abilities, Sutskever argues that evolutionary priors can explain much of human advantage. He points to dexterity and locomotion as areas where humans—by virtue of evolution—arrive with very powerful starting points. In his words, For things like vision, hearing and locomotion, I think there's a pretty strong case that evolution actually has given us a lot
. He contrasts this with the practical difficulty of reproducing similar performance in robots, noting that while robots can be trained to be dexterous if you subject them to like a huge amount of training and simulation
, achieving quick, human-like acquisition of new physical skills in the real world remains out of reach.
Learning from limited, low-diversity data
Sutskever uses a childhood example to illustrate how little diverse data humans sometimes need to reach strong performance. Recalling his own experience, he says that as a five-year-old he could already recognise cars well despite limited exposure—you don't get to see that much data as a 5-year-old. You spend most of your time in your parents house, so you have very low data diversity
—yet performance was strong. He allows that evolution could account for some of that ability.
Domains unlikely to be explained by evolution: language, math and coding
Turning to cognitive domains that emerged recently in human history, Sutskever argues that evolution is a less satisfying explanation. He highlights language, mathematics and coding as areas where evolutionary priors are unlikely to explain human facility. But in language and math and coding, probably not,
he says, and suggests that the human advantage in those domains indicates something more fundamental
about how humans learn—an inference that points beyond simple, hard-coded priors.
Generalization and the sample-efficiency gap
A central claim of the interview is that modern models generalize far worse than humans. Sutskever frames this as a foundational problem: The thing which I think is the most fundamental is that these models somehow just generalize dramatically worse than people
. He distinguishes two related issues: the raw number of samples required to learn a task, and the difficulty of teaching a model what a human can learn easily. Using the teenage-driver example, he notes how few hours of practice produce reliable human drivers, while comparable machine learners require vastly more experience.
Human learning characteristics: unsupervised, robust, and self-correcting
Sutskever enumerates properties of human learning that machine learning currently fails to match: fewer samples, a more unsupervised learning style, and remarkable robustness. He observes that a teenager learning to drive does not rely on an external verifiable reward but instead learns from interaction and internal evaluation: It takes fewer samples. It's more unsupervised. A child learning to drive a car… A teenager learning how to drive a car is not exactly getting some prebuilt, verifiable reward
. He also stresses human robustness: The robustness of people is really staggering
.
Value functions and self-assessment in human learning
To explain how learners can improve without an external teacher, Sutskever invokes the notion of internal value functions. He suggests that humans possess an internal sense that enables rapid self-correction. On the driving example he says, They have their value function. They have a general sense which is also, by the way, extremely robust in people
. That internal sense allows learners to judge performance and accelerate improvement without formal external rewards.
Implications for AI research and the era of research
Sutskever connects the gap in human versus machine learning to a broader shift in AI strategy. He argues that after a phase in which scaling dominated progress, the field must return to fundamental research to close gaps like generalization and sample efficiency. While the interview touches on many implications, the recurring message is that better learning recipes—those that capture unsupervised learning, robustness and internal evaluation—are needed if models are to learn more like humans.
References:
Dwarkesh Podcast — Ilya Sutskever (episode page)
YouTube — Ilya Sutskever: We're moving from the age of scaling to the age of research
Explore more exclusive insights at nextfin.ai.
