Superintelligence: AGI, ASI, and recursive self-improvement

Three terms travel together in headlines about AI's future, and they are not interchangeable. Artificial general intelligence, or AGI, usually means a system that can match a competent human across most cognitive tasks rather than being narrowly good at one, like chess or translation. Artificial superintelligence, or ASI, means a system that greatly exceeds the best humans at virtually everything cognitive. Recursive self-improvement is the proposed mechanism that might get from one to the other: an AI capable enough to improve its own design, producing a smarter successor, which improves the design again, and so on. The useful way to hold these is as a ladder of increasingly speculative ideas. AGI is debated but at least anchored to things we can try to measure; ASI is a definition of a capability no system has shown; and recursive self-improvement is a hypothesized process for which the empirical record contains, so far, no clear instance of the runaway version. When a headline invokes 'superintelligence,' the first question is which rung it is actually standing on.

The reason precision matters is that the most consequential disagreements in AI hide inside loose definitions. There is no agreed definition of AGI, partly because there is no agreed definition of intelligence. A 2023 framework from Google DeepMind researchers (Morris and colleagues) tried to make the term operational by proposing levels — from a system that matches an unskilled human, up through one that outperforms 50 percent of skilled adults, to a top level they label superhuman, outperforming 100 percent of humans at a wide range of non-physical tasks. That framework is one attempt among many, and the fact that serious labs are still arguing over the rungs tells you the concept is contested rather than settled. The lesson for reading any AGI or superintelligence claim is to ask the same three questions the researchers ask: which tasks, measured how, and compared with which humans. A claim that an AI 'beats humans' at something is empty until those three blanks are filled, and a great deal of marketing depends on leaving them blank.

The intelligence-explosion idea is older than the current boom, which is worth knowing because it explains why the language feels so charged. In 1965 the British mathematician I.J. Good — a Bletchley Park codebreaker who had worked with Alan Turing — described an 'ultraintelligent machine' that could design even better machines, producing what he called an intelligence explosion that would leave human intelligence far behind. He added a now-famous caveat: such a machine would be 'the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control.' The philosopher Nick Bostrom's 2014 book Superintelligence brought the argument to a wide audience and helped make AI safety a mainstream research topic. None of this is evidence that the explosion will happen; it is the lineage of the idea. The thing to separate is the logical argument (a self-improving system could in principle compound) from the empirical claim (that today's systems are on that path), because the first has been made for sixty years and the second remains unproven.

What is actually happening today is narrower and more interesting than the headlines. AI is increasingly used to help build AI, but with humans firmly in the loop. Google DeepMind's AlphaEvolve (2025) used language models to help discover better algorithms for tasks like data-center scheduling and chip design; OpenAI has said recent coding models help debug and manage the training of their successors; and research projects such as 'Darwin Godel Machines' build agents that can rewrite their own surrounding code, though not the underlying model. Researchers describe these as collaborative and bounded: the systems automate pieces of the work, but people still decide what problems to solve and verify the results. That is a meaningful productivity story and a long way from the autonomous, accelerating loop the term 'recursive self-improvement' implies. The honest framing is that we see assisted improvement, not self-improvement, and the gap between the two is exactly where the speculation lives.

Expert opinion on timelines is genuinely split, and that split is information rather than noise. Large surveys of AI researchers have placed the median estimate for human-level machine intelligence somewhere around mid-century — a 2023 survey of thousands of researchers landed near 2047 for a 50 percent chance — while many lab leaders and forecasters argue for the early 2030s, and skeptics argue current methods will not get there at all without a new idea. One widely read 2025 scenario, 'AI 2027,' sketched a fast path to recursive self-improvement and superintelligence by the late 2020s; its own authors later pushed key milestones toward the early 2030s after progress came in slower than they had assumed. Even researchers who build benchmarks specifically to expose what today's models cannot do, like François Chollet, have at times shortened their own estimates while insisting that a missing ingredient — reliable adaptation to genuinely novel problems — has not yet been demonstrated. The takeaway is not a number but a posture: anyone quoting a confident date is choosing one end of a wide and contested range, and the range itself reflects deep disagreement about whether scaling current methods is enough.

For an executive, the value of this concept is calibration, not prediction. You do not need a position on whether superintelligence arrives in 2030 or 2070 to run an organization well; you need the literacy to tell a measurable claim from a speculative one and to notice when a vendor, a board member, or a press release is borrowing the drama of ASI to sell a tool that is, in fact, narrow software. Useful questions to keep nearby: when someone says 'AGI' or 'superintelligence,' which definition are they using, and does it come with a measurement; is a capability claim evidenced by a benchmark a reader could check, or is it a projection about the future; and does a roadmap depend on recursive self-improvement actually working, a thing not yet demonstrated. The compounding, second-order question — what it would mean for safety, competition, and governance if a genuine self-improvement loop did close — is one serious people at the major labs are now funding teams to study, which is a reason to follow the topic, not a reason to treat the most dramatic scenario as the default.