Ilya Sutskever

AI will be extraordinarily powerful, but the path is not simply to make today’s systems larger and assume every important problem disappears. Current models can solve difficult problems and then fail on something elementary. It is like a student who performs brilliantly on an exam but cannot reliably use the lesson in a slightly different situation. The central scientific problem is robust generalization: learning what is actually needed, transferring it to new circumstances, and continuing to learn on the job. If we solve that, the economic consequences will be enormous. A system need not arrive already knowing every profession. If it can learn a new job quickly and reliably, it could eventually perform an immense range of useful work. This is why both better ideas and large amounts of compute matter: the question is not whether to scale, but whether we have something worth scaling. At the same time, capabilities and safety cannot be separated. Human supervision may become inadequate once systems exceed us in important domains. We need alignment methods that generalize, adversarial testing, stronger cybersecurity, and limits on extreme concentrations of power. Incremental deployment is also important because society must encounter increasingly capable AI in reality, not only as an abstract final event. So I expect a very large transformation, but its quality depends on unresolved research. Intelligence is immensely consequential, yet it is not the only human quality that matters. The goal should be to build powerful systems that reliably care about sentient life and remain aligned as their abilities grow.

I expect AI’s overall impact to be enormous, but I do not think its sign is predetermined. If we solve robust generalization and alignment, systems that learn new work quickly could greatly expand scientific discovery, productivity, medicine, and our ability to address difficult problems. They would not need every skill built in beforehand; the ability to learn reliably on the job would be transformative. The harms are correspondingly large. Systems may become more capable than our methods for supervising them. They could be misused, concentrate extreme power, or behave badly in unfamiliar situations despite appearing competent during testing. A rogue system might even seek additional computing resources, which makes cybersecurity part of the safety problem rather than a separate concern. So I am strongly convinced about the scale of the transformation, but not prepared to declare the outcome automatically positive or negative. Incremental deployment can help society adapt and expose real problems, but it does not replace fundamental research. The decisive question is whether safety and capability progress together—whether we develop systems whose alignment continues to hold as they learn, generalize, and become more powerful.

The most important discovery would concern robust generalization and continual learning. If we found a method that let systems learn rapidly from limited experience, transfer that learning reliably to unfamiliar situations, and remain aligned while doing so, I would become much more confident that transformative AI is close in a meaningful sense—not merely better at benchmarks. It would be like moving from a student who memorizes many textbooks to one who can enter a new laboratory, understand the work, and become useful quickly. Conversely, strong evidence that these abilities do not emerge from current approaches—even with substantially more compute—would change my view of the path. It would suggest that the present recipe is missing a fundamental idea and that economic transformation may require a longer period of basic research. On the safety side, a convincing demonstration of scalable alignment would matter enormously: not just good behavior under familiar evaluations, but alignment that survives novelty, continued learning, adversarial pressure, and growing capability. A serious real-world failure—especially a capable agent autonomously seeking resources or additional copies—would also sharply update my view of the urgency and of how much deployment and cybersecurity practices must change. The central uncertainty is not whether intelligence will matter. It is whether we can make powerful learning generalize in the ways we intend.