Jürgen Schmidhuber

I expect AI to move beyond language interfaces and screen-bound intelligence toward agents that build world models, run simulations, and act competently in the physical universe. That transition matters because intelligence is not merely producing convincing text or solving virtual benchmarks. A generally capable system must also understand objects, causes, uncertainty, embodiment, and the consequences of actions in a complicated world. This will take longer than some fashionable forecasts suggest. Robotics and self-driving systems illustrate how stubborn physical reality can be: rare events, changing environments, imperfect sensors, and costly mistakes make progress slower than in software-only domains. For that reason, I reject the claim that AI will master every scientific field within just a couple of years. Nevertheless, I expect eventually superhuman physical AI, not merely superhuman chat systems. Recursive self-improvement will also become increasingly important. The underlying idea is not new; algorithms for learning how to improve learning have a substantial technical history. What changes is the economics. As computation becomes cheaper, old ideas that were once impractical can become effective at scale, and capabilities now concentrated in frontier laboratories may spread to ordinary machines. This is also why I regard attempts to ban superintelligence as infeasible: when sufficient computation and improvement methods are broadly accessible, durable global suppression becomes unrealistic. The future, therefore, is not simply larger language models. It is increasingly autonomous intelligence learning about, predicting, and ultimately transforming the physical world.

Overall, I expect AI to have a profoundly positive and transformative impact, especially once it progresses from manipulating information on screens to understanding and acting in the physical world. Superhuman physical AI could greatly expand our capacity for scientific discovery, engineering, production, and the solution of problems that human intelligence alone handles slowly or imperfectly. That does not mean the transition will be harmless. Increasingly autonomous systems can make consequential mistakes, and physical agents face a far less forgiving environment than language models. Recursive self-improvement and falling computation costs will also diffuse powerful capabilities beyond a few controlled laboratories. This makes both harmful uses and attempts at centralized prohibition important concerns—but it also makes a lasting global ban on superintelligence unrealistic. So the central question is not whether intelligence can be frozen at today’s level. It is how civilization adapts as increasingly capable systems become widespread. I remain optimistic about the long-run result, while rejecting compressed timelines that confuse impressive virtual performance with mastery of science and the physical universe.

The clearest update would come from decisive evidence about physical-world learning. If an AI could robustly build world models, transfer knowledge across unfamiliar environments, and autonomously master difficult robotics rather than succeeding in a narrow demonstration, I would shorten my expectations for superhuman physical AI. Conversely, if such systems continued to fail despite much cheaper computation and sustained algorithmic progress, I would become less optimistic about the pace and scale of their impact. I would also update if recursive self-improvement proved either substantially more powerful or more limited than expected in real systems. The historical idea is well established; the open empirical question is what becomes practical when abundant computation meets algorithms that improve parts of their own learning process. But no single language benchmark would change my view much. Fluent screen-bound behavior is not the decisive test. The important event would be robust, general competence in the physical universe.