Ilya Sutskever

Ilya Sutskever co-authored AlexNet in 2012, co-founded OpenAI in 2015, served as its chief scientist for nearly a decade, and then did the most consequential thing of his career: he left. In November 2023, Sutskever was one of the board members who voted to fire Sam Altman, an act that triggered a near-total staff revolt and Altman's reinstatement days later. By June 2024, Sutskever was gone from OpenAI entirely, and within weeks he announced Safe Superintelligence Inc. (SSI), a company with a single stated goal: build safe superintelligence and nothing else. No products, no revenue pressure, no quarterly shipping cadence. The implication was clear. The person who had done more than almost anyone to prove that scaling neural networks actually works had concluded that the labs racing to deploy those networks were not taking the danger seriously enough.

Sutskever's technical worldview rests on a simple materialist premise: the brain is a biological computer, and therefore a digital computer can replicate everything the brain does. He states this flatly and repeatedly, treating it as settled. But his more interesting positions concern what current AI systems are getting wrong. In a December 2025 interview with Dwarkesh Patel (his first extended public appearance after leaving OpenAI), he identified a sharp disconnect between benchmark performance and real-world usefulness. Models ace hard evals but then cycle between introducing and reintroducing the same bug in a coding session. His explanation: RL training makes models narrowly superhuman on the specific tasks they're optimized for, but the generalization is terrible. He used an analogy of two students preparing for competitive programming. One grinds 10,000 hours and memorizes every algorithm. The other practices for 100 hours and does nearly as well. The second student, the one with what Sutskever calls "the it factor," will have the better career. Current AI models, he argued, are extreme versions of the first student. Researchers inadvertently design RL training environments that mirror the evals they want to ace. The result is a kind of meta-reward-hacking where the humans, not just the models, are overfitting.

The deeper claim is that the age of pure scaling is over. Sutskever, who arguably proved scaling laws empirically with the GPT series, now says that era "sucked the air out of the room" and starved genuine research. He wants a return to the age of ideas. One specific direction he emphasizes: AI systems need something analogous to human emotions, not sentiment, but a value function that provides fast, approximate feedback about whether a course of action is promising. He cites the case of a patient whose stroke destroyed emotional processing, leaving IQ intact but making him unable to choose which socks to wear. Without an internal compass, intelligence is paralyzed. Current RL relies on outcome-level reward signals at the end of long trajectories. Sutskever argues that a learned value function could short-circuit this, giving intermediate credit the way a chess player knows losing a piece is bad without playing the full game. He frames this as the central missing ingredient, not more data, not more compute, but better algorithms that learn efficiently from sparse feedback. He openly says he doesn't know whether pre-training alone can ever produce this, and that there may be no good machine learning analogy for what evolution encoded in biological brains over three billion years. SSI, presumably, is his attempt to find out.