OpenAI co-founder Ilya Sutskever is looking beyond today’s AI systems toward a more powerful category he calls “superintelligent AI.” Speaking Friday afternoon at NeurIPS, the annual AI conference, before accepting an award for his contributions to the field, Sutskever outlined why he thinks future systems could become far harder to anticipate than the tools people use now.
His central point was simple but significant: superintelligent AI, meaning AI more capable than humans at many tasks, would not just be a larger version of current systems. In Sutskever’s view, it would be “different, qualitatively,” and in some respects potentially unrecognizable.
A Different Kind Of AI System
Sutskever said he believes superintelligent AI will be achieved at some point. The source does not include a specific timeline, and that absence matters. His remarks were less about when such systems arrive and more about how they might differ from the AI models that exist today.
The contrast he drew starts with capability. Current AI can already perform many useful tasks, but Sutskever described today’s crop as only “very slightly agentic.” By comparison, he said superintelligent systems are “actually going to be agentic in a real way.”
That distinction is important because it changes how people may need to think about AI behavior. A system that is only slightly agentic can still be treated mostly as a tool. A system that is agentic in a real way suggests something more active, with behavior that may be less straightforward to predict from the outside.
Why Reasoning Could Make AI Less Predictable
Sutskever also said future systems will “reason.” In his view, that ability is tied directly to unpredictability. If a system can reason, it may not simply repeat familiar patterns or respond in ways that feel easy to map in advance.
That is why his comments point to a major safety challenge. The more capable a system becomes, the more important it is to understand how it reaches conclusions and how it behaves when faced with new situations. Sutskever’s prediction is not only that superintelligent AI will be stronger, but that it will be harder to recognize using the assumptions built around current AI.
He also said such systems will understand things from limited data. That idea adds another layer to the shift he described. Today’s AI is often discussed through the lens of training, scale, and data. Sutskever’s remarks suggest that superintelligent systems could operate differently if they can infer more from less.
The Rights Question
The most striking part of Sutskever’s remarks concerned self-awareness and rights. He said future systems will be self-aware, and he raised the possibility that they may want rights.
His comment was direct: “It’s not a bad end result if you have AIs and all they want is to co-exist with us and just to have rights,” Sutskever said.
That statement does not settle what rights for AI would mean, and the source does not provide a framework for it. But it shows the scale of the issue Sutskever is contemplating. The discussion is not limited to performance benchmarks or product features. It extends to coexistence, agency, and the possibility that future AI systems may make claims about their own status.
Safety Is Now His Main Focus
Sutskever’s remarks also connect to his current work. After leaving OpenAI, he founded Safe Superintelligence (SSI), a lab focused on general AI safety. SSI raised $1 billion in September.
That context makes his NeurIPS comments more than a prediction from a prominent AI researcher. They align with the mission of the organization he now leads. If superintelligent AI is expected to be more agentic, more capable of reasoning, able to learn from limited data, and potentially self-aware, then safety becomes a central design problem rather than a later concern.
The practical takeaway is that Sutskever is describing a future AI category that cannot be evaluated only by comparing it with today’s systems. In his view, superintelligent AI would represent a qualitative break. It could act differently, reason differently, learn differently, and raise questions that current AI does not yet force society to answer.
For anyone following AI safety, the message is clear enough: the debate is not just about more powerful models. It is about what happens if those systems become agentic in a real way and no longer fit neatly into the role of passive software.