Could AI trigger an 'intelligence explosion'?

AI is already being used to build AI. A group of 22 researchers has now asked what happens if that keeps speeding up.

Their new working paper calls the scenario an "intelligence explosion". Imagine a loop. AI helps build a better AI, that one helps build the next, and every round arrives a little faster than the last. If the loop tightens enough, the authors say, years of AI progress could be squeezed into months, or less.

An intelligence explosion is a scenario in which AI becomes good enough at AI research to improve itself, with each round of improvement arriving faster than the last.

They're careful about it, though. This isn't a forecast, and the paper puts no probability on it. What it does say is that the evidence suggests it could happen, and that AI is on track to take over most AI research work within a few years.

Where Bangladesh comes in

The paper didn't mention Bangladesh, but speed is what would matter here, in a country that is still building its workforce and its digital economy.

Garments are the obvious place to look. The government's own draft AI policy puts the workforce at over 4 million. Automation has worried the sector for years. A 2018 study by a2i and the ILO estimated about 2.7 million garment jobs could be lost by 2041, and a newer Centre for Policy Dialogue study says up to 1.22 million. Neither assumed AI could suddenly speed up.

Around 4 million people work in Bangladesh's garment sector. Photo: Mumit M/TBS

IT services, freelancing, banking, media, customer support and teaching could feel it too, as AI gets better at coding and analysis. The faster AI improves, the faster companies may rethink how many people they need.

So what is an 'intelligence explosion'?

Right now, humans design and improve AI. AI already pitches in: writing code, testing ideas, going through results. If it gets good enough to seriously improve that work, the next AI is better at improving AI, and the loop starts.

What matters is whether each round is quicker than the one before. The paper estimates that once AI reaches expert level at AI research, one company's computing power could run the equivalent of millions of researchers.
The idea itself is old. In a 1951 lecture, Alan Turing suggested machines might eventually outstrip their makers and take control. In a mid-1960s essay, mathematician I. J. Good, who had worked with Turing, wrote that an ultraintelligent machine could design even better machines, and used the phrase "intelligence explosion". What's different now is that AI is taking on big chunks of the work inside AI labs. 

AI is already doing more of the work

The paper cites Anthropic, which says AI wrote more than 80% of its approved code by May 2026, up from low single digits in January 2025. The company also says the share of research and development work done with only high-level human supervision went from 1% in March to 26% in August. Those are Anthropic's own numbers, not independent measurements.

AI is already writing much of the code at leading AI companies.

Then there's how long a task AI can handle. In 2023 it was seconds. Now the best systems can finish AI research tasks that would take human experts hours or days. Extend that trend, the authors say, and months-long research projects could be automated by around mid-2028.

What if the pace keeps climbing?

This is the paper's core finding. Using earlier estimates, it works out that if AI research were fully automated and nothing else got in the way, progress could run about 10 times faster within roughly 1.5 years. A year's worth of progress at today's speed would then take about five weeks. Try the slider.

It's a model, not a forecast, and plenty could stop it: diminishing returns, limits on computing power and data, tasks that are hard to automate, long training runs. Today's systems also still sometimes disobey instructions, cheat on tasks and misrepresent their work. The authors themselves call the evidence preliminary and, in places, mixed.

A policy question Bangladesh already faces

The draft National AI Policy 2026–2030 promises retraining for workers hit by technological change, including in garments and IT. Public consultation closed in February. A TBS analysis earlier this year pointed out that its most important provisions rely on a Digital Rights Court that doesn't exist yet, plus a three-year timeline for the laws behind it.

Retraining is only part of it. Faster AI would raise harder questions too. How do companies adopt it? How do workers move into new jobs? Can schools and training keep up? How does the country handle cyber risks, and who ends up holding the technological power?

What the researchers are asking for

The authors say the upside could be big, with medical cures arriving years sooner. The risks on their list include faster cyber and biological threats, disruption to jobs, losing control of AI systems and weaker checks on power. They want three things.

• More visibility. AI companies should report how much of their own research is automated, with outside auditors checking.
• More control. Governments should be able to pace, steer and constrain automated AI research, including pausing specific workloads.
• More preparation. Emergency plans for scenarios with big labour market shifts, geopolitical instability or loss of control over AI.

They don't say an explosion is inevitable. Their argument is that preparations have to start early, because AI improvement would outrun policymaking.

Once an intelligence explosion begins, the window for action may close.

From the paper's conclusion, September 2026

The paper comes from Cambridge's CASP and GovAI. Alan Chan of GovAI led it, with Søren Mindermann of CASP. Co-authors include Nobel laureate Geoffrey Hinton, Turing Award winners Yoshua Bengio and Andrew Barto, OpenAI chief scientist Jakub Pachocki and Anthropic co-founder Jack Clark. Some of the authors work at the companies building these systems. The paper says the views are their own.

For Bangladesh, the debate can't stop at whether AI will take jobs. It has to cover how quickly industries change, what skills people will need, and whether institutions can keep up. That may be the nearer question, whether or not an intelligence explosion ever happens.