Artificial Superintelligence

Where artificial general intelligence describes matching human capability broadly, superintelligence describes decisively exceeding it — not in one domain, as chess engines and protein-folding systems already do, but across the board, including the capacity to improve its own design.

The concept comes from academic philosophy, where it was formalized as a thought experiment about what follows if such a system is built. It has since moved from thought experiment to stated corporate objective: several major labs now describe superintelligence as an explicit goal in their public materials and organizational structure.

No such system exists, and the reasoning about it is necessarily speculative. It is nonetheless load-bearing in the field, because it motivates a substantial share of both the investment and the safety research.

Whatever you make of the plausibility, the belief is consequential: it shapes where enormous capital goes, how labs are structured and governed, and which risks safety teams prioritize. Understanding what people mean by the word is a prerequisite for reading the industry's own statements about itself.

That it is just a faster or larger version of today's models. The claim is qualitative, not quantitative — a system that generates better scientific hypotheses than any researcher, not one that writes email faster. Scaling current models may or may not get there, and that is precisely the open question rather than a settled path.

That timelines from senior figures are forecasts. They are positions in an argument, made by people with strong incentives in both directions — labs raising capital, and researchers arguing for caution. Estimates from credible sources span a range wide enough to be unhelpful for planning, largely because they use different definitions of the same word.

That it is either inevitable or impossible. Both framings substitute confidence for evidence. The honest position is that nobody knows whether current methods lead there, and that the disagreement among informed people is genuine rather than a matter of one side not understanding the technology.

For practical decisions, this term should change almost nothing — build with what exists and can be measured. Its real use is as a decoder: when a lab reorganizes, raises extraordinarily, or publishes a safety framework, this concept is usually the premise behind it.