Artificial superintelligence used to be a niche topic, the kind of thing you found on futurism forums and in science fiction. That has changed. The subject now comes up in conversations with people who work at, and run, some of the biggest AI companies on the planet. They speak about it in a very serious tone, and they even warn about risks to humanity itself.
In the middle of all this noise, one voice decided to ask a more down-to-earth question. Cal Newport, a computer scientist, argues in an article published in IEEE Spectrum that superintelligence is not going to happen. The interesting part is the warning that comes with it: even if this kind of AI never arrives, the race to build it can still do real damage.
For anyone thinking about investing in artificial intelligence in Brazil, this clash of ideas has practical consequences. Doubting the hype is not the same as doubting the technology, and the origin of this debate shows exactly why.
What sparked the debate
According to the IEEE Spectrum report, the trigger came in the middle of this year, when Jacob Coxon, a researcher at Anthropic, left the company. When he announced his exit, he said that Anthropic and OpenAI were racing straight toward a superintelligence able to improve itself, and that they were gambling with the lives of people along the way.
The response from inside Anthropic caught a lot of people off guard. Evan Hubinger, head of alignment science at the company, said publicly that Coxon was right and that the team truly believes AI could kill every human. Shortly after, as the outlet reports, CEO Dario Amodei told CNN that he agreed with Coxon far more than he disagreed.
Notice the detail here. These are not outside critics. These are the people who build these systems, openly discussing the possibility of human extinction. If you want to dig deeper into this side of the debate, check out our piece on the real risk of superintelligence.
But a loud alarm does not always mean solid analysis. And that is exactly where the Newport article comes in.
Superintelligence, AGI and self-improvement in plain English
Before going further, it helps to sort out three terms that often get mixed up in headlines. The first is AGI, short for artificial general intelligence. Broadly speaking, the idea is an AI that can match humans across a wide range of mental tasks, not just one specific thing like translating text or generating images.
Artificial superintelligence, known as ASI, goes much further. Here we are talking about a system with cognitive abilities far beyond ours, to the point where it could easily slip out of human control. IEEE Spectrum is careful to point out that this is a much more ambitious goal than AGI. In other words, people who talk about ASI are not talking about an AI that is just a bit better than what we have today.
So how would we get there? This is where recursive self-improvement, or RSI, comes in. Picture a loop: the AI improves itself, the improved version gets even better at improving itself, and the process keeps speeding up on its own. Much of the fear around ASI depends on this loop actually working. If it stalls at some point, the whole story changes.
Here is a comparison that helps. Imagine a kitchen crew where every new cook is trained by the previous one and ends up a little better. In theory, quality just keeps going up. In practice, you run out of ingredients, the oven has limits, and time is short. The theory of an endless loop tends to skip that part, and that is exactly where Newport starts poking.
The arguments from Cal Newport
Newport did not reach his conclusions casually. According to IEEE Spectrum, he studied the topic in a doctoral seminar in computer science that he led at Georgetown University in Washington, where he and his students looked at both the technical and the cultural roots of the superintelligence narrative. The group even published an annotated bibliography with the readings used in the course.
The question guiding the work is simple and, oddly enough, rarely asked: from an engineering point of view, is superintelligence even possible? For him, answering that helps put the anxiety around the topic in perspective and leads to more effective responses to this technology.
The outcome of the seminar, as the outlet reports, was a series of objections to the feasibility of ASI. And also a deeper fear: even if the so-called intelligence explosion never arrives, chasing it could cause concrete harm.
An old idea that is still speculation
The first objection is that recursive self-improvement is still speculative. The concept goes way back. According to the report, statistician I.J. Good raised the hypothesis in a 1965 paper, suggesting that a rapid cycle of AI improvements could lead to an intelligence explosion.
In the 1990s, IEEE Spectrum explains, the idea gained traction among the Extropians, a futurist group in the San Francisco Bay Area. There, it became a way to talk about superintelligence without having to explain how anyone would actually build such a system. Later, the concept was picked up by well-known names like Eliezer Yudkowsky and Nick Bostrom, figures tied to movements that began placing heavy weight on existential risks to humanity.
It is a subtle point, but an important one. A theory can stay influential for decades without ever passing an engineering test. Popularity and proof are two different things.
That does not mean the companies are sitting still. Far from it. IEEE Spectrum reports that Anthropic published a piece in June called When AI Builds Itself, describing how agents based on language models already speed up writing, debugging and optimizing code, and automate parts of research, like running experiments and logging results. Internal company data, according to the publication, suggests that AI is already speeding up AI development. OpenAI, for its part, announced in September a goal of having an automated researcher by March 2028.
So the topic has moved out of thought experiments and into the daily engineering work of these companies. The real question is a different one: is speeding up parts of the work the same as kicking off a loop with no brakes?
The limit of just getting bigger
Newport doubts that these advances put the companies on a path to runaway self-improvement. One of the obstacles he points to is the availability of the data needed to keep improving these systems.
To understand why, it helps to look at the history. According to the report, up until 2023 the big leaps in language models came from making them larger. GPT-1 had 117 million parameters, which are the values learned during training that shape how the neural network processes information. GPT-2 reached 1.5 billion, GPT-3 hit 175 billion, and GPT-4, released three years ago, probably went past one trillion.
After GPT-4, the article says, that steady climb started to behave differently. The Newport argument starts from there: if the recipe behind earlier leaps depends on resources that do not grow forever, like data, it becomes harder to defend the idea of an improvement loop that only keeps accelerating.
Think of a factory that doubled its output several times just by buying more machines. At some point, it runs out of raw materials. Buying more machines stops solving the problem. For anyone betting on an endless spiral, that kind of bottleneck is a hard signal to ignore.
Hype is also a strategy
Not everyone sees the superintelligence talk as a sincere warning. According to IEEE Spectrum, reactions to the topic vary a lot. Some people see the concept being used as a marketing tool. Others see something more calculated: a way to encourage regulations that help the labs already leading the market hold back newer competitors.
There are other camps too, the outlet notes. Some focus on debating whether these systems could actually destroy us, arguing over specific scenarios. Others believe it is possible to align ASI with human interests and that, because of this, it is worth speeding toward that promise.
The marketing theory deserves a closer look. When a company says its technology is so powerful it could threaten humanity, it is also quietly sending the message that it sits at the cutting edge of the industry. It is a kind of advertising that does not look like advertising. That does not prove the warning is false, of course. But it is a good reason to listen with a bit of a filter.
For anyone evaluating vendors and tools, this critical eye is worth its weight in gold. Big talk is not a performance metric. What matters is what the tool delivers today, in your workflow, with your data. Everything else is just a story, and stories do not pay the bills.
What this debate means for AI in business in Brazil
The first lesson is practical: investing in artificial intelligence makes more sense when you start from the problem it solves right now. Not from the promise of an AI that will think on its own a few years down the road. A tool that helps you answer customers faster or speeds up coding already justifies itself by what it does, no prophecy required.
It also helps to separate real gains from stage talk. Automating repetitive tasks, coding support, customer service triage and marketing content production are uses that many companies already apply every day. Our article on artificial intelligence in business optimization gives a good overview of this more concrete side.
A quick example makes this easier to picture. Imagine a small marketing agency that starts using AI just to draft captions and organize reports. It measures how much time the team spent before and how much it spends after. If it pays off, the agency expands to another task. If not, it adjusts or drops it. None of this depends on superintelligence. It depends on method.
Another thing to watch out for is projects and contracts sold on the promise of future leaps. Imagine a company that signs a long contract because the vendor promised the next model will fix everything. If progress comes more slowly, the company is stuck with a high cost and none of the promised return. Shorter agreements with deliverables you can measure tend to be a safer choice.
Finally, stay flexible. The future of AI could move at a very different pace than announced, faster or slower. Anyone who builds processes without depending on a single tool can switch solutions with less pain. If you are just getting started, take a look at how startups use AI and automation in a lean way.
Clarity beats prophecy
The challenge Cal Newport raises is not an attack on artificial intelligence. He questions one specific promise, the idea that AI will improve itself until it becomes something uncontrollable. And he reminds us that the race toward that promise can carry costs, even if it never comes true.
The debate, though, is far from settled. There are heavyweight researchers convinced of the risk, there are skeptics like Newport, and there are plenty of people somewhere in between. Nobody needs to have the final word to make good decisions in the present.
For people who work in tech, marketing or run a business, the message is actually freeing. You do not have to pick between panic and blind excitement. You can use AI with your feet on the ground, measuring results and adjusting course as you go.
If the darker side of this discussion still has you curious, our piece on whether artificial intelligence could wipe out humanity pairs nicely with this read, as do the other Método Viral articles on AI in business.
Next time someone tries to sell you a tool based on the revolution it promises, ask just one question: what does it solve for me today? If the answer is good, go for it. If not, keep your money.
