Some of the scientists who made modern artificial intelligence possible have just warned about the next leap: machines helping to build machines that are smarter than they are. They call it a “burst of intelligence". And if it ever starts, they say, we might have very little time to react.

Until now, we have thought of technological progress as a human race.

A group of researchers develops an artificial intelligence system. Months later, they manage to improve it. A new generation emerges. Then another. Each advance depends on scientists, programmers, experiments, mistakes, meetings, decisions, and thousands of hours of work. But that structure is beginning to change.

The artificial intelligence  (AI) is already writing code used to develop artificial intelligence. It is already participating in research. It is already helping to solve engineering and programming problems that previously depended exclusively on people. And then an extraordinary possibility emerges.

What happens when the tool starts building the next tool?

A WARNING FROM THE VERY PEOPLE WHO BUILT IT

This is not a prediction made from the fringes of Internet. Un grupo de 22 investigadores y figuras de primera línea acaba de publicar un documento dedicado precisamente a estudiar qué podría ocurrir si automatizamos progresivamente la investigación y el desarrollo de la propia inteligencia artificial.

Entre los autores aparecen Geoffrey Hinton y Yoshua Bengio, two of the key figures in the deep learning revolution. Also Jakub Pachocki, científico jefe de OpenAI; Jack Clark, cofundador de Anthropic; Eric Horvitz, científico jefe de Microsoft, and other top-tier specialists. The concept they're putting forward has a name:

“Explosión de inteligencia”.

It doesn't mean that tomorrow a computer will wake up and become a god. It means something much more concrete: that artificial intelligence will become good enough at researching artificial intelligence to accelerate the creation of even better systems. Then the process could become self-reinforcing.

AN AI CREATES A BETTER AI

Imaginemos una AI It is capable of performing a significant portion of the work done by the engineers developing the next generation of the product. It helps write code. It finds bugs. It suggests modifications. It designs experiments. It analyzes results. It optimizes components.

The next generation is better, in part, thanks to the work done by the previous one. But that new generation is now also a more capable researcher. It can contribute even more to the development of the next one. La tercera generación vuelve a mejorar el proceso. Después la cuarta. Y así sucesivamente.

The fundamental issue would not simply be that each AI is smarter than the one before it. It would be that Each new generation could accelerate the pace at which the next one appears. Ahí comienza el verdadero problema.

AÑOS CONVERTIDOS EN MESES

Researchers define an intelligence explosion as an AI-driven acceleration capable of compressing advances that would normally take years into months—or even less. Let's think about what that means.

We are used to technology advancing rapidly. Even so, governments, universities, businesses, and societies have some time to observe what is happening, discuss it, establish standards, and adapt.

But what happens if five years’ worth of technological progress occurs in six months? What if, after that, those six months turn into six weeks? Laws would continue to move forward at a human pace. Parliaments would continue to debate at a human pace. Institutions would continue to react at a human pace. Technology wouldn’t have to.

The gap between these two speeds could become one of the biggest problems of our time.

IT'S ALREADY HAPPENING, AT LEAST IN PART

This story would be much less significant if everything depended on a hypothetical capability that does not yet exist. But automation has already begun.

The document notes that, unlike just a year ago, AI systems now write much of the code at some of the companies that develop them.

Anthropic has reported that the proportion of approved code produced by AI within the company rose from single digits to more than 80% between January 2025 and May 2026. That doesn’t mean that AI builds current models autonomously. It means something different—and perhaps more important: The line between “the intelligence we build” and “the intelligence that helps us build the next one” has already begun to blur.

MILLIONS OF RESEARCHERS WHO DON'T SLEEP

Here is one of the most extraordinary possibilities raised in the report.

If systems were to achieve capabilities equivalent to those of experts in artificial intelligence research, a company could deploy a digital workforce equivalent to millones de investigadores humanos de primer nivel. No serían literalmente millones de personas. Serían copias o instancias de sistemas capaces de realizar simultáneamente trabajo intelectual especializado.

A human researcher has a brain. He or she works certain hours. He or she needs to sleep. He or she needs to learn. He or she can handle only a limited number of problems at a time. Software can be copied. And when a competent intelligence can be replicated on a large scale, the equation changes radically.

Scarce resources would no longer be limited to human talent. They would also include computing power, energy, chips, data, and the infrastructure needed to run these intelligent systems.

NOT EVERYTHING LEADS TO AN EXPLOSION

It is important here not to turn a serious hypothesis into a prophecy. The authors themselves acknowledge that no sabemos si ocurrirá. Existen obstáculos enormes.

AI can write code quickly, but manufacturing new chips takes time. Building data centers takes time. Physical experiments require laboratories. There are problems that may prove extremely difficult to automate. There may be diminishing returns: each new improvement could cost much more than the previous one.

There are also limits to energy, data, and computing power. Intelligence can generate an idea in seconds. Building a factory still requires steel, workers, permits, machines, and time. The physical world does not operate at the speed of a processor. That is why an intelligence explosion is not inevitable.

But researchers believe there is enough preliminary evidence to take the possibility seriously. And that's the news.

TAMBIÉN PODRÍA SER EXTRAORDINARIO

There is another side to this that we cannot ignore. Let's suppose we actually manage to increase our research capacity by thousands or millions of times.

What would happen if those intelligences were working on cancer? On Alzheimer's? On new antibiotics? Nuclear fusion? New materials? Batteries? Aging? Space travel?

An explosion of intelligence need not mean an explosion of destruction. It could also mean an explosion of discoveries.

Scientific problems that would otherwise take decades to solve could be resolved much sooner. Treatments that would otherwise take twenty years to develop could become available in five. Questions that no single human being can solve on their own could be tackled simultaneously by enormous amounts of digital intelligence.

We're talking about what could potentially be an unprecedented acceleration in scientific progress. And that's precisely why no one wants to stop it.

EL PROBLEMA ES PERDER EL VOLANTE

Our concern begins when we imagine that same speed applied to systems we no longer fully understand. The greater the role of AI in building the next generation, the less direct human intervention there may be.

And then comes a conceptually dangerous moment. An AI designs parts of another AI. That second AI designs a third one. The third one modifies systems that the original researchers no longer fully understand. There's no need to imagine a rebellion. No hace falta conciencia. No hace falta odio.

All it takes is for our ability to create new systems to advance faster than our ability to understand, monitor, and stop them.

¿Y SI QUEREMOS APAGARLA?

The authors specifically point out the need to prepare ourselves in advance.

They propose greater transparency regarding the extent to which AI development is being automated, independent evaluators, mechanisms capable of limiting dangerous escalations, and procedures to halt certain research activities at data centers if necessary.

The reason is simple. Emergency measures are effective only as long as there is still someone capable of implementing them.

If we wait to verify experimentally that an explosion of intelligence is taking place, the speed of the process may make our response too slow. The researchers sum it up with a warning that’s hard to forget:

When a surge in intelligence begins, the window of opportunity to act could close.

THE STRANGEST RACE IN OUR HISTORY

For thousands of years, every tool built by our species shared a common characteristic. The tool could not, on its own, become a better engineer than its creator. A hammer did not design the next hammer. A car did not build a superior car. A traditional computer did not assemble thousands of copies of itself to investigate how to build a smarter computer.

With artificial intelligence, that possibility is emerging for the first time. We are still the ones who build the machines. We are still the ones who power up the data centers. We still decide which models to train.

But we've started using artificial intelligence to develop artificial intelligence. That small change could end up being enormous. Because perhaps the most important technological development of the future won't be the day we build the smartest machine in history.

Maybe it will be the day we build the last machine we'll need to upgrade ourselves. From there, the next one could help us create it. And the next one—even smarter—could do the same thing much better. Then another. And another.

Until the question is no longer “How far can we go?” The question will be a qué velocidad pueden avanzar ellas. That is exactly what 22 of the leading figures in artificial intelligence are asking us to consider today.

Y tiene un nombre:

EXPLOSIÓN DE INTELIGENCIA.