There is a moment — somewhere between AGI and ASI — where the rules change. Not gradually, the way technology usually changes, but all at once, in a feedback loop that removes the one constraint that has always limited how smart machines can get: us.
That moment has a name. I.J. Good called it the intelligence explosion. He coined the term in 1965. And right now, in 2026, some of the most credible researchers in AI are saying we may be watching it begin.
The Original Idea
Irving John Good was a British mathematician who worked at Bletchley Park during World War II, alongside Alan Turing, breaking Nazi codes. He was not a science fiction writer. He was a statistician with a practical mind and a habit of following logic wherever it led.
In 1965, he published a paper called Speculations Concerning the First Ultraintelligent Machine. Most of it is dense technical philosophy. But buried inside is one paragraph that changed the entire trajectory of how we think about artificial intelligence:
"Since the design of machines is one of these intellectual activities, an ultraintelligent machine could design even better machines; there would then unquestionably be an 'intelligence explosion.'"
I.J. Good, 1965 — Speculations Concerning the First Ultraintelligent MachineThe logic is simple. If a machine can do any intellectual task humans can do — including the task of designing better AI — then humans stop being the ceiling. The machine designs a smarter version of itself. That version designs an even smarter one. Each cycle is faster than the last. The result: machine intelligence that leaves human intelligence so far behind it's not really a comparison anymore.
Good ended the paragraph with a warning that has aged poorly as a reassurance: "The first ultraintelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control."
Provided that it's docile enough.
From Concept to Crisis: The Road from 1965 to Now
How the Mechanism Actually Works
The intelligence explosion isn't magic. It's a feedback loop — and feedback loops are mundane until they aren't. Here's the mechanism step by step:
- An AI reaches human-level ability at designing AI systems. This is AGI — not in the sci-fi sense, but in the specific sense of being able to do the intellectual work that AI researchers do.
- The AI designs a slightly smarter version of itself. Maybe 5% better at the task of improving AI. This step doesn't require anything dramatic — just competence applied to self-modification.
- The improved version designs a further-improved version. Now 10% better. Then 20%. Each cycle the gains compound.
- The cycle runs faster than humans can evaluate it. This is the critical transition. Normal AI safety involves humans reviewing each iteration. Once cycles outpace review, the feedback loop is effectively autonomous.
- Intelligence increases to a level that has no historical precedent. What I.J. Good called the ultraintelligent machine. What we now call ASI — artificial superintelligence.
The current term for this mechanism is recursive self-improvement: an AI improving its own capabilities in a way that makes each subsequent improvement easier and faster.
Is It Already Happening?
This is the question that divides serious researchers right now, and the answer is genuinely uncertain.
On one side: we already have early forms of recursive self-improvement in the wild. AutoML systems search for better neural architectures autonomously. Reinforcement learning systems improve through self-play (AlphaZero became the world's best chess engine by playing millions of games against itself, with no human data). AI systems now assist in their own research pipelines at Anthropic, OpenAI, and Google DeepMind — writing code, running experiments, and analyzing results at a scale that human researchers alone could not match.
In September 2026, a group that included Geoffrey Hinton (Turing Award winner), researchers from Anthropic and OpenAI, and Microsoft's chief scientific officer co-signed a paper warning that current AI systems are "approaching recursive self-improvement capabilities" — performing internal R&D tasks and automating significant portions of their own engineering work.
On the other side: a measured rebuttal. MIT Technology Review's August 2026 analysis argued that current AI self-improvement is bounded, supervised, and nowhere near the autonomous feedback loop Good described. The efficiency gains are real but linear, not exponential. Constraints — compute limits, alignment bottlenecks, diminishing returns on scale — may slow or halt the cascade before it becomes self-sustaining.
There is a difference between AI that helps humans improve AI (which exists today) and AI that improves itself without human oversight (which does not yet clearly exist). The intelligence explosion requires the second. The debate is about how close we are to crossing that line — and whether we will notice when we do.
Intelligence Explosion vs. Technological Singularity
These terms are often used interchangeably. They shouldn't be.
| Term | Coined by | What it means | Scope |
|---|---|---|---|
| Intelligence explosion | I.J. Good (1965) | A specific AI feedback loop where self-improvement accelerates exponentially | Narrow — a technical mechanism |
| Technological singularity | Vernor Vinge (1993) | A point beyond which human civilization cannot predict or model the future | Broad — a civilizational horizon |
| Hard takeoff | Eliezer Yudkowsky | The intelligence explosion happens fast — days or weeks, not years | Scenario — speed of the explosion |
| Soft takeoff | Paul Christiano et al. | The explosion is gradual — months or years, giving time to respond | Scenario — speed of the explosion |
The intelligence explosion is the mechanism. The singularity is what follows if the explosion runs far enough and fast enough. Hard vs. soft takeoff describes the speed at which the mechanism unfolds — and the speed matters enormously for whether humanity has time to respond.
Where AGI and ASI Fit
The intelligence explosion is the bridge between two things this tracker follows closely: AGI and ASI.
AGI is the trigger. The intelligence explosion is the mechanism. ASI is the result — if the explosion runs without being stopped or slowed.
This is why the timing question matters so much. Forecasters tracked on this site put AGI anywhere from 2026 to 2030 for the optimists, with more cautious estimates in the 2030s. But those forecasts are for the trigger — not the explosion itself. Once the loop starts, how fast it runs is a separate question entirely.
Why This Matters Right Now
The intelligence explosion isn't an abstract thought experiment anymore. It's the operating assumption behind some of the most significant decisions being made in AI right now:
- Why Anthropic, OpenAI, and DeepMind invest heavily in alignment research — because if the explosion happens before alignment is solved, a misaligned system could improve itself beyond our ability to correct it.
- Why governments are rushing AI regulation — the September 2026 white paper explicitly warned that "once an intelligence explosion begins, the window for action may close." Governments that regulate after the fact may find there is no after the fact.
- Why the AGI timeline debate is so charged — a few years' difference in timing is not just an academic argument. It's the difference between having time to prepare and not having it.
I.J. Good ended his 1965 paper with a sentence that feels more urgent every year: "It is more probable than not that, within the twentieth century, an ultraintelligent machine will be built and that it will be the last invention that man need ever make."
He was off by a few decades on the timeline. He may not have been wrong about much else.
Do you think the intelligence explosion has already begun — in any form?
Results are illustrative — live aggregation coming soon.