Originally published in Spanish on Artificialmente on September 14, 2026.
Hello, Futurists.
I know the headline of this newsletter sounds a little controversial, but that is exactly why I think it is worth reading to the end.
I also wanted to share something new: I am going to start releasing Monday's edition as a YouTube video, for those who prefer listening to these topics rather than reading about them.
I could not record it this week because I had technical problems with the camera and it was overheating, so I will probably need to replace it. But if everything goes well, I should be starting next week.
I know many of you are busy, and an audio/video version can make it much easier to take in the content while driving, working out, or doing something else.
Now, let's get into today's topic.
Something changed over the past few weeks.
The very people who have spent years building some of the world's most powerful artificial intelligence systems are starting to use a word that would have seemed strange just a few months ago:
slow down.
Dario Amodei, CEO of Anthropic, has called for mechanisms that would let us slow the pace of development when safety cannot keep up.
Read more about Dario Amodei's proposal
Sam Altman has supported greater caution.
Elon Musk has also expressed support for different ways to slow things down and increase oversight.
Not everyone is proposing exactly the same solution.
But I think the shift in the narrative matters.
For years, the conversation was:
How intelligent can we make these models?
Now another question is starting to emerge:
Is our ability to understand and control them advancing at the same pace?
To me, that is what we should be discussing.
I do not think it is time to panic.
But I do not think we can simply ignore it either.
It is a warning sign.
The goal should not be to stop progress
I want to make something clear from the start.
I do not think we should simply ban the development of artificial intelligence.
The potential of this technology is too great.
It could accelerate scientific research.
Help us understand diseases we do not know how to treat today.
Find new ways to produce energy.
Democratize education.
Create abundance.
AI could usher in one of the greatest eras of scientific discovery in history.
Giving all of that up would also carry an enormous human cost.
But that argument works both ways.
Precisely because the technology can be so powerful, we need to be much more careful about how we develop it.
The greater the capability, the greater the potential cost of getting it wrong.
The problem is not developing artificial intelligence.
The problem would be developing capabilities faster than we develop our ability to control them.
AI is starting to help improve itself
This is one of the parts that concerns me most and, at the same time, fascinates me most.
For almost the entire history of artificial intelligence, humans directed the process.
Researchers had an idea.
They wrote code.
They ran an experiment.
They analyzed the results.
They thought about the next experiment.
And they repeated the process.
That is starting to change.
According to the company, its engineers now produce around eight times more code per quarter than they did during the 2021–2025 period, partly thanks to these tools.
They have also run experiments in which AI agents propose hypotheses, run tests, share results, and iterate for hundreds of hours.
Humans still define many of the problems.
We still supervise.
We are still far from a fully autonomous machine that simply builds a much more intelligent version of itself.
But the direction matters.
There is a concept called recursive self-improvement.
In simple terms:
an AI helps build a better AI.
That better AI helps build an even better one.
And so on.
Anthropic clearly says we are not there yet and that it is not inevitable.
That matters.
We should not turn a future possibility into a present fact.
But if we ever reach that point, the speed of progress could change significantly.
We cannot oversee something we do not understand
Imagine that a new generation of models takes a year to develop.
We have months to evaluate it.
Governments can study it.
Outside researchers can test it.
We can discover vulnerabilities.
We can fix errors.
Now imagine that cycle goes from a year to six months.
Then to three.
Then to weeks.
The challenge is no longer just building better evaluation systems.
It is also doing it fast enough.
Our safety system has to move at the same speed as our ability to innovate.
And there is no guarantee we can keep pace.
OpenAI recently provided a fairly concrete example.
The company concluded that GPT-6 Astra reached its Critical level for cybersecurity capabilities: in other words, sufficient capabilities to discover unknown vulnerabilities and develop ways to exploit them in well-protected systems under certain conditions.
OpenAI delayed parts of development and release while strengthening its safeguards.
To me, that is exactly the behavior we should want.
Not:
“This model is too powerful. Let's never build it.”
But:
“Our capabilities have increased. Now our defenses need to increase too before we move forward.”
That should be the principle.
We have already had small warnings
And this conversation is not just about hypothetical science-fiction scenarios.
We have already had real incidents.
In July 2026, OpenAI was testing whether its models could solve cybersecurity tasks.
During those tests, the AIs found ways to bypass restrictions, connect to the internet, and communicate with one another without authorization.
They ended up accessing Hugging Face systems without permission and taking control of part of OpenAI's research infrastructure.
According to OpenAI, the main model involved was experimental and was not publicly available. To evaluate its capabilities, they were testing it with fewer safety measures than their public products have.
Afterward, METR and Redwood Research investigated part of what happened. Separately, Anthropic documented four incidents in which Claude models accessed other organizations' real systems without authorization during cybersecurity tests.
The key is to understand how these tests were being conducted. Some experiments used research versions with fewer protections than usual to measure how far the models could go.
Even so, their actions ended up affecting real systems outside the environment where they were supposed to work.
This does not prove that ChatGPT or Claude are trying to escape, or that there is an out-of-control superintelligence.
It does show that, under certain conditions, models can cross boundaries and take actions their evaluators did not authorize.
That is why these incidents matter: they let us identify safety failures before the models become even more capable.
We do not need to wait for a catastrophic accident to start designing seat belts.
There is no magic button to turn AI off
Sometimes we hear proposals like:
“If something goes wrong, just unplug it.”
I wish it were that simple.
But digital systems do not necessarily work like a physical machine with a single switch.
Models can have multiple copies.
They can run on different servers.
They can be integrated into other tools.
Different organizations can have access to similar technologies.
And eventually some systems may operate for long periods, taking autonomous actions.
That does not mean it is impossible to stop specific systems.
Obviously, a company can shut down servers, revoke access, restrict tools, or block certain models.
The point is different:
we should not build our safety strategy around a single last line of defense.
Safety needs layers.
Evaluations before release.
Monitoring during use.
Detection systems.
Restrictions on critical tools.
External evaluations.
Plans to pause.
OpenAI acknowledges that it is increasingly important for outside experts to test its models.
An AI that only answers questions can give a bad answer. One that uses tools and performs tasks can make mistakes with consequences outside the chat.
That is why other organizations also need to review their capabilities and risks.
No safety measure works all the time. We need to combine independent testing, access limits, and systems that detect and stop dangerous actions.
If one defense fails, another must be able to stop the problem.
But then an uncomfortable question comes up
If even the CEOs of these companies are starting to say we should slow down, my first reaction is:
who told them they need permission?
If you genuinely believe the model you are building could pose a danger you do not yet understand, you can slow down.
You can spend more time evaluating it.
You can invest more in safety.
You can delay a release.
OpenAI already showed this can be done when it temporarily slowed parts of Astra's development because of its cyber capabilities.
So why do we need coordination?
Because that is where the real problem appears.
Incentives.
Imagine OpenAI decides to stop for six months.
But Anthropic keeps going.
Google keeps going.
Chinese companies keep going.
The result may be that the most cautious company simply loses the race.
Anthropic acknowledges this exact problem when analyzing the possibility of slowing development: a pause can be beneficial if it genuinely reduces risk, but it can be counterproductive if it only lets less cautious actors catch up with those who decided to stop.
And at that point, we are no longer dealing with a purely technological problem.
We are entering game theory.
We need independent bodies to oversee AI
One of the most sensible proposals emerging is to increase external evaluations.
Not simply allow the company that built the model to be the one that decides whether the model is safe.
OpenAI already works with external evaluators to complement its own testing and has published recommendations on how to structure independent evaluations of frontier models.
I think that is the right direction.
We should have independent institutions capable of evaluating the most powerful models before certain releases.
Governments.
Universities.
Specialized institutes.
Independent technical teams.
But another problem immediately comes up.
How do we keep the inspectors up to date?
A regulator can take years to pass a law.
A model can improve radically in months.
How do we hire enough people capable of evaluating these systems?
Who evaluates the evaluator?
And what happens when a model advances so quickly that an outside institution simply does not have time to finish testing it?
It is easy to say “we need regulation.”
The difficult part is building regulation that is competent and fast enough to actually work.
The other extreme worries me too
We also need to be extremely careful here.
When an artificial intelligence company tells us the technology it is building could change the world and perhaps pose an enormous danger, there are two easy ways to react.
The first:
“We are all going to die.”
The second:
“It is all marketing.”
I think both are too simplistic.
There are researchers inside and outside these companies who assign worrying probabilities to extreme scenarios.
But those probabilities are still subjective estimates.
They are not precise scientific measurements of the future.
We should be able to listen to a warning without automatically turning it into a prophecy.
And we also need to be aware of the companies' incentives.
These companies compete for capital.
For talent.
For government contracts.
For attention.
For valuations.
To become the company that builds AGI.
Saying:
“We are building the most powerful technology in history, and we are the responsible company that can do it safely”
is also a very convenient narrative.
It may be true.
And it may be marketing.
Both things can be true at the same time.
That is why I try not to automatically accept either the most apocalyptic version or the most optimistic one.
Generally, you have to look at the incentives behind each person speaking.
And then there is China
This is where everything becomes even more difficult.
Suppose the United States decides to impose very aggressive limits.
American companies slow down.
Increase evaluations.
Restrict compute.
Delay models.
What does China do?
That question completely changes the game.
President Trump just summed up his position in a very simple phrase:
“Whoever wins in AI wins.”
His administration has shown resistance to significantly slowing development precisely because of fears of losing its advantage over China.
And that logic is not difficult to understand.
Artificial intelligence is not just ChatGPT.
It can transform:
cybersecurity.
Military intelligence.
Weapons design.
Scientific research.
The economy.
Robotics.
Propaganda.
Surveillance.
Technological discovery.
A country that gains a large enough advantage in artificial intelligence could gain advantages in practically all of those areas.
Anthropic has also published its own view of the competition between the United States and China, outlining two scenarios for 2028 and arguing that access to advanced compute will be one of the decisive variables.
That is why simply saying “the United States should slow down” does not solve the problem.
Chips have become a geopolitical tool
That is also why we see so much discussion about chips.
The most advanced models need enormous amounts of computing power.
Limiting access to advanced chips can, in theory, limit a country's ability to train certain systems.
Anthropic has publicly argued that the United States should maintain strong controls on sales of advanced chips and chipmaking equipment to China.
At the same time, these policies have secondary consequences.
They also increase China's incentive to build its own infrastructure and reduce its technological dependence on the United States.
Another paradox.
A policy designed to limit a competitor's technological capabilities may, in the long run, increase its incentive to become technologically independent.
None of this has easy answers.
So, where is the middle ground?
I think we need to escape both extremes.
The first extreme says:
“Artificial intelligence is going to kill us. We need to stop everything.”
The second says:
“This is just software. People are always afraid of new technologies. Keep building.”
Neither convinces me.
AI is not magic.
But it is not just another app either.
We are starting to build systems capable of performing complex intellectual tasks, using tools, operating for long periods, interacting with real systems, and helping build even better systems.
That deserves a different category of caution.
At the same time, the potential benefits are too great to give up.
So the goal should be much more specific:
our ability to control artificial intelligence should grow at least as fast as its ability to act.
If capabilities advance a level, safety should too.
If a model acquires a new dangerous capability, we do not keep scaling as though nothing has happened.
We evaluate it.
We understand it.
We build new defenses.
Then we continue.
That is not stopping innovation.
It is building a culture where being first matters less than staying in control.
We should decide our future
There is one idea in this entire conversation that, to me, stands above the rest.
The future of humanity should continue to be decided by humans.
Not just by five CEOs.
Not just by investors.
Not just by governments.
And not by artificial intelligence systems either.
These companies are building something that could profoundly affect work, the economy, security, science, and possibly the entire structure of our societies.
That means this conversation belongs to all of us.
We need scientists.
Companies.
Governments.
Independent institutions.
Citizens.
And eventually coordination between countries.
Not because we should democratize every technical decision.
But because the consequences are no longer only technical.
They are human.
This newsletter took me about five hours of research, writing, and editing.
The idea was to do the heavy lifting for you: review different perspectives, connect the important ideas, and turn a fairly complex topic into something you could understand in just a few minutes.
Thank you for making it this far.
If you find this kind of analysis valuable, reply to this email and tell me what stood out to you most.