Is an Intelligence Explosion Really Coming?

The fear has a clean logic to it. Build an AI good enough to improve AI, and it will make a better version of itself, which will make a better version still, until progress runs away from human hands in a matter of months. This is the intelligence explosion, and for years it lived mostly in thought experiments. Lately it has moved into the mouths of people who build these systems for a living, which is why it is worth taking seriously and worth examining carefully.

Who is saying it, and what exactly

The voices are not fringe. Anthropic's Jack Clark has said he thinks it is more likely than not that an AI capable of autonomous self-improvement exists by the end of 2028. OpenAI's Jakub Pachocki has spoken of aiming for a meaningful, largely automated AI researcher by around early 2028. In a 2025 survey of 25 researchers drawn from frontier labs and universities, reported alongside work from groups such as Forethought and the researchers cited by Axios, 20 named the automation of AI research itself as one of the most severe and urgent risks they could see.

There is real data underneath the worry. AI already writes a large share of the code inside the labs. Anthropic's Evan Hubinger has put the figure at 70 to 90 percent of the code used to develop future models, and the company has said Claude now leads about a quarter of its research and development work. If machines are already doing much of the labor of building the next machine, the feedback loop is not hypothetical. It has started.

Where the story gets shakier

Here is the part the headlines tend to skip. No AI system has actually improved itself in a runaway way. Not once. Writing code is not the same as deciding what to build, and the hard part of research has always been the ideas, not the typing. The people who study this closely, including groups at Georgetown's CSET, keep pointing to bottlenecks that do not vanish because a model got faster. Good ideas are scarce. Compute is finite and expensive. Real experiments take real time, and the physical world does not run at the speed of a chatbot.

There is also a history worth remembering. Confident predictions of imminent takeoff have come and gone before, and the timelines have a habit of sliding a year or two into the future every year or two. That does not make the concern wrong. It does mean the word explosion is carrying a lot of weight, implying a suddenness the evidence has not yet shown.

How worried to be

The honest answer is somewhere in the uncomfortable middle. The concern has graduated from science fiction to something respected researchers plan around, and that alone is a reason to pay attention rather than roll your eyes. But planning around a risk is not the same as the risk arriving. What is happening looks less like an explosion and more like a steady, machine-assisted acceleration, powerful and worth measuring, and slower than the most dramatic warnings suggest. The number to watch is not any single model's benchmark. It is how much of the research itself the machines are quietly taking over, a trend that also runs through the louder extinction warnings and the resignations that have followed them.