Foundations
SuperintelligenceASI
Also known as: artificial superintelligence
A hypothetical AI far smarter than the best humans across virtually every field; a step beyond even human-level AGI.
Superintelligence is the idea of an AI that is not just as smart as a person, but far smarter than the very best humans at almost everything, science, strategy, persuasion, invention. It is a hypothetical, it does not exist, and there is no agreement on whether it is possible or how far off it might be. It comes up because a mind that capable would be extraordinarily powerful, which is exactly why researchers think hard about safety in advance.
Superintelligence (sometimes ASI) refers to a hypothetical AI that dramatically exceeds human ability across essentially all domains. It sits beyond artificial general intelligence: AGI is roughly human-level and general; superintelligence is far above human-level.
Why people discuss something that does not exist
- Stakes: a system vastly more capable than us could be immensely beneficial or dangerous, so the possibility is taken seriously even if uncertain.
- Control: a central worry is whether humans could keep such a system aligned with our intentions, the alignment and AI safety problem, at that level of capability.
A note on hype
The concept is speculative and contested. It is important not to confuse today's genuinely narrow AI with anything approaching superintelligence.
Superintelligence denotes a hypothetical agent whose cognitive performance greatly surpasses the best human across virtually all economically and scientifically relevant domains (Bostrom 2014). It is distinguished from AGI by degree and generality of capability, and features in analyses of possible rapid capability gains and the difficulty of maintaining control over a system more competent than its overseers.
The associated technical concern is scalable oversight and alignment under a capability gap: preference-based methods like RLHF rely on humans being able to evaluate outputs, an assumption that weakens as systems exceed human judgement, motivating research into weak-to-strong generalisation, interpretability and verification. The topic is genuinely contested, on feasibility, timelines and takeoff dynamics, and prone to both dismissal and overstatement; rigorous treatment separates the speculative long-term question from the demonstrated, still-narrow capabilities of current models, while noting that decisions about safety investment are made under deep uncertainty.