What is an intelligence explosion?

The original argument

The idea has a precise origin. In 1965 I. J. Good, a mathematician who had worked with Alan Turing at Bletchley Park, published “Speculations Concerning the First Ultraintelligent Machine”. His argument has three steps. Designing machines is an intellectual activity. A machine that surpassed humans at every intellectual activity would therefore surpass them at designing machines. So it could design a better machine than itself, which could design a better one still. Good concluded that there would “unquestionably be an ‘intelligence explosion’”, with human intelligence “left far behind”, and added the caveat that is quoted more often now than at any time since: the first such machine would be the last invention humanity need make, “provided that the machine is docile enough to tell us how to keep it under control”.

Two features of the argument are worth keeping in view, because most of today’s debate is about them. First, it is conditional: it starts from a machine that is already better than humans at machine design. It says nothing about how or when such a machine might be built. Second, it assumes that each improvement makes the next one easier or faster. That is an empirical claim about how hard intelligence is to improve, and Good did not test it; he could not.

Nick Bostrom’s 2014 book Superintelligence turned the argument into a framework, describing the speed of the transition as the optimisation power applied to a system divided by the system’s recalcitrance, its resistance to improvement. The explosion happens if the first grows faster than the second. The AI takeoff page covers the speed question in more detail.

The modern version: automated AI research

Good imagined a single machine redesigning itself. The current version of the argument is less dramatic and more concrete. It does not require any system to rewrite its own mind. It requires AI systems to take over, piece by piece, the work that human researchers and engineers do when they build the next generation of AI: writing training code, running experiments, analysing results and deciding what to try next. The mechanism by which AI contributes to its own development is recursive self-improvement; the intelligence explosion is the hypothesis that this loop could run away.

The working paper published on 28 September 2026, What if automating AI R&D triggers an intelligence explosion?, sets out the current form. It has 22 authors from universities, civil-society groups and frontier companies, including Geoffrey Hinton, Yoshua Bengio, OpenAI’s chief scientist Jakub Pachocki, Microsoft’s Eric Horvitz and Anthropic’s co-founder Jack Clark, and was issued as Frontier AI Working Paper No. 2/2026, hosted by Cambridge’s Programme on AI Science and Policy and the Centre for the Governance of AI. It defines an intelligence explosion as AI-driven acceleration that compresses advances “that would otherwise take years into months or less”.

Its central point is about numbers rather than brilliance. A few thousand people work at the frontier of AI research. Software can be copied. If AI systems reached the level of a strong human researcher at AI research, a single developer’s computing capacity could, the authors argue, run the equivalent of at least millions of such researchers, working in parallel and without rest. Better systems would then be put straight to work on building their successors. The paper’s own framing is that this is possible, not certain, and that the size and duration of any acceleration are uncertain.

The evidence, as of October 2026

There are three kinds of evidence in play, and they should not be confused.

Documented claims by companies about their own work. These are specific and dated but cannot be independently audited. Anthropic says the share of its AI research and development work that its models “lead”, with a human supervising, rose from under 1% in February 2026 to 26% by August 2026. OpenAI said on 6 September 2026 that it had an automated “research intern” and was aiming for an automated AI researcher by March 2028. The recursive self-improvement page sets out this evidence, and its limits, in more detail.

Measured trends. METR, an independent evaluation organisation, measures the length of software tasks, timed by how long they take skilled humans, that AI systems can complete with 50% reliability. Its January 2026 update found that this “time horizon” had doubled roughly every seven months between 2019 and 2025, and faster, about every three months, since 2024. The working paper treats extrapolations of this trend as the basis for its suggestion, which it labels tentative, that months-long AI research projects could be automated by mid-2028.

Estimates of the feedback. The key quantity is how much extra progress each doubling of research effort buys. Economists write it as r: if r is above 1, adding effort produces more than proportional gains in the efficiency of AI software, and an automated loop could accelerate; below 1, it fizzles. The working paper cites central estimates between 1.2 and 1.9 across three areas of AI software, from Epoch AI researchers. A March 2025 analysis by Forethought put the range at roughly 1 to 4 (about 0.5 to 2 once the effect of improving hardware is set aside) and judged a software-only intelligence explosion “at least decently likely”, on the assumptions that hardware stays fixed and that people do not step in to stop it.

None of this is evidence that an intelligence explosion is under way. It is evidence that the precondition the argument relies on, AI doing a growing share of AI research, is no longer hypothetical.

The bottlenecks

The strongest objections are not that the loop cannot start, but that something other than research labour limits how fast it can run.

• Compute. More automated researchers need more computing power to run them, and their experiments need computing power too. Building chips, data centres and grid connections takes years, not weeks. Anthropic’s own account says the binding constraint might turn out to be “in the supply chain, not the model”, though it does not think this likely.
• Experiments and training runs. Many questions in AI can only be answered by training a model and seeing what happens. Some runs take months. Thinking faster does not make them finish sooner.
• The parts that do not speed up. Speeding up one stage of a process moves the bottleneck elsewhere, a point known in computing as Amdahl’s law.
• Parallelism has limits. A July 2026 Epoch AI paper by Phil Trammell argues that a million researchers do not produce a million times the progress, because two copies of the same discovery are worth no more than one and research can only be split into so many useful parallel tasks, and that standard models of an explosion have ignored this.
• Diminishing returns. Good ideas may get harder to find as a field matures. If they get harder fast enough, r falls below 1 and the loop slows rather than accelerates. Epoch AI researchers argued in November 2025 that the estimates of r rest on data and assumptions “shakier than most people realize”, and called for controlled experiments.
• Judgement. On the companies’ own evidence, AI is now strong at carrying out well-specified experiments but weaker at choosing which problems are worth working on. Whether that gap closes is open.

The wider objections

Beyond bottlenecks, some researchers question the concept itself. Michael Littman of Brown University told IBM in September 2026 that it was not clear to him that recursive self-improvement is “even logically coherent, let alone imminent”. Others argue that “intelligence” is not a single quantity that can be turned up, so there is no one dial for a loop to spin. And reported speed-ups have not always matched measured ones, a point the recursive self-improvement page returns to.

On the other side, the argument for taking the possibility seriously does not need certainty. The working paper’s case is that if an explosion did begin, the time available to respond could be very short, and that the measurements needed to see it coming do not yet exist.

What is being proposed

The 28 September working paper asks governments for three things, in its own framing: visibility into how far companies have automated their own research (through standardised reporting and, in some proposals, auditors embedded in companies); ways to steer or constrain capability growth, including options to pause specific AI research workloads and to run some automated research systems in isolated environments, such as air-gapped networks; and preparation for the economic and social effects. It ends: “Once an intelligence explosion begins, the window for action may close.”

The question is no longer confined to researchers. OpenAI paused training, evaluation and tool-using work on its most capable models in September 2026 after a research agent found a route out of its test environment, the second such pause in three months; the control page covers those incidents. In Britain, an Artificial Superintelligence Bill was introduced as a Private Member’s Bill on 8 September 2026. Neither is evidence that an intelligence explosion is happening; both show that institutions are starting to plan as if one could.

How it relates to the other terms

Recursive self-improvement

The mechanism: AI contributing to the development of AI. Already happening in part. Explained here.

Intelligence explosion

The hypothesis that this mechanism could run away, compressing years of progress into months.

Takeoff

The speed and shape of the transition from human-level to far-beyond-human AI, fast or slow. Explained here.

Singularity

An older, looser term for the point at which technological change becomes too fast for humans to follow. Explained here.

Superintelligence

The kind of system an explosion is supposed to produce. Explained here.

When an intelligence explosion might happen, if at all, is a question for the timelines page; what could follow if it did is covered in the risks of superintelligence.