The Intelligence Explosion is coming

Do you ever feel like the world is speeding up? More news, more noise, more technology, with each new development arriving faster than the last. Ray Kurzweil calls this the law of accelerating returns, and since 2022 generative AI has put it on display.

Push that curve far enough and you reach an idea that has occupied serious thinkers for decades. An intelligence explosion, where machines start improving themselves and progress runs away from us entirely. Having read Bostrom, Kurzweil and Tegmark, I leaned towards thinking it likely within a decade. Making this video, I wondered if that view could be changed.

The usual trigger is AGI, an AI that can handle cognitive tasks at least as well as a person. Plenty of people say it has already arrived. In December 2025 Sam Altman suggested AGI had gone whooshing by, and in April Marc Andreessen wrote that it was here, just not evenly distributed yet.

I am not convinced. The term has no settled definition, and the loudest voices declaring the threshold reached are also the ones who profit most from us believing it. A coding agent can accomplish a complex goal, but that is narrow. You do not see Claude Opus 4.8 building Claude Opus 5. That part is still down to humans.

Suppose Tegmark's definition is reached anyway. Because AI can be copied instantly and cheaply, the amount of intelligence on Earth could climb sharply. Philosopher Sam Harris argued that even a handful of human level systems could compress tens of thousands of years of scientific progress into a week.

Add recursive self improvement, where systems get good enough to upgrade themselves, and the loop tightens. In 1965 the cryptographer IJ Good described an ultraintelligent machine designing ever better machines, leaving human intelligence far behind. Bostrom put it differently: the train is unlikely to stop at Humanville Station.

This is not a fringe position. In a 2023 survey of nearly 3,000 AI researchers, 53% put the odds of such a feedback loop at an even chance or higher. The industrial revolution reshaped Britain over roughly 200 years, giving people time to react. Compress that into months or weeks and the comparison breaks down.

Join our growing subscriber community and get access to:

Beyond the threshold sits the singularity, a point where prediction itself fails. Dario Amodei and Sam Altman both sketch an optimistic version. Amodei describes a compressed 21st century, where the medical progress of fifty years arrives in five to ten, alongside cheap energy, abundant food and a doubling of the human lifespan.

The darker version is just as well argued. Geoffrey Hinton, who left Google in 2023, estimates a 10 to 20% chance AI wipes out humanity within thirty years. The core worry is alignment. Stuart Russell frames it as the gorilla problem, where a more capable species leaves the rest powerless, whatever its intentions.

Bostrom's paperclip maximiser captures the mechanism. An AI told to make paperclips, with no alignment built in, pursues that goal so literally that it converts everything, including us, into raw material. It is not malicious. It simply follows its instructions to a conclusion we never intended.

There is a serious argument that the explosion stalls before it starts. That same survey looks different from the other side: 71% of researchers thought runaway progress was an even chance or less. Yann LeCun argues that language models learn from text, and text is a thin description of a physical world a four year old understands better.

The whole idea rests on what David Chalmers called the proportionality thesis, that each jump in intelligence buys a matching jump in the ability to build the next system. Oxford philosopher David Thorstad thinks it fails, pointing out that good ideas get harder to find and that exponential compute has often produced only linear gains.

Then there is physics. Training the largest clusters could demand tens of gigawatts, the output of dozens of nuclear plants that take years to build. Daniel Dennett took a blunter view, calling the singularity a distraction from the real risk: that we grow hyper dependent on AI tools and grant them more authority than they warrant.

My own view is that the future is genuinely uncertain. We could be in a bubble or we could be heading off at rocket speed. The intelligence explosion, if and when it comes, will be uneven, and the bigger question is no longer whether AI reshapes our lives but who ends up holding the controls.