There's a debate among AI experts that never seems to end — the one I keep hearing and reading about on social media. It's the question of what general intelligence actually is: where humans sit, where machines sit, and even where animals sit if we put everything on a single scale.
First, a disclaimer: I'm not an AI expert or a cognitive science expert. I'm just trying to summarize what I understand from reading and listening to experts in these fields so I can make up my own mind.
The second important point before we start: there isn't just one kind of intelligence. "Intelligence" as such is a concept way too broad, one that most people tie to the ability to learn, relate ideas, understand complex concepts, solve problems, and get by in difficult situations. If we look at psychology, intelligence tests always measure several different aspects — reading comprehension, mental arithmetic, creativity, logical thinking, memory. Each of these gets a series of tests that end in a score, and the well-known IQ is nothing more than an average of those scores.
Like all averages, it has problems. A person can be brilliant in one area and mediocre or poor in others (which is where definitions like simple talent, complex talent, or giftedness come from, depending on whether someone is above the 75–80th percentile in one, several, or all areas). But more or less, it's an accepted way to measure the intelligence of a literate person — most tests are written, so even if you're a genius on the level of Einstein or Mozart, you can't take them if you can't read or write.
With all that out of the way, let's get to the heart of the matter.
Can AI surpass humans?
For me, this is the key question, and I want to sketch the two main camps I see among experts first.
The first group — I'll call them "humanists" for lack of a better word — are the AI experts who believe humans are the pinnacle of intelligence: that evolution gave us a capacity for abstraction, reasoning, and meta-cognition that artificial means can't reach.
The most well-known advocates of this position are Timnit Gebru and Emily M. Bender, two of the authors of the famous paper on "Stochastic parrots". Others we could include, though they're less radical, might be Gary Marcus, Margaret Mitchell, or even Yann LeCun. Each has their own view of the problem, but they all more or less agree that LLMs don't resemble human intelligence and never will. Some of these experts might concede that if we gave a robot intelligence, agency, and sensors similar to a human's, plus enough compute and memory... (grounded intelligence) then maybe, just maybe, it could learn to become a kind of human simulacrum.
It's important to note that when these experts talk about machines never reaching human intelligence, they don't necessarily mean some ineffable "soul" or consciousness. They mean fundamental differences in how current AI systems process information versus how biological brains do. Their main argument is that statistical pattern recognition, however sophisticated, is qualitatively different from the genuine understanding humans display.
On the other end of the spectrum are those who say humans are nothing more than a pile of heuristics and cognitive biases glued together and shuffled by evolution through trial and error — and that rather than being the pinnacle of intelligence, we're the minimum viable version needed to create and sustain a shared culture. This side includes well-known researchers and entrepreneurs like Sam Altman, Dario Amodei, Ilya Sutskever, and Geoffrey Hinton. Even though they don't usually argue it the same way, their conclusion is that with more scale and more data, these systems will get smarter and handle far more complex tasks than any person could understand, at much higher speed.
For them, culture and the relationships between concepts are the real intelligence that propels us forward as a species and lets us create societies, institutions, companies, and other self-perpetuating cultural expressions that culminate in AI. AI is nothing more than the expansion of that collective intelligence, of that culture which has endured and outlived us all as humans — so it can reach much higher heights if we give it a foothold. These are the ones you might call transhumanists or doomers, depending on whether they think AI is a blessing or an existential risk. This is the group the media gives the most airtime to, the ones who write the most books, generate the most hype, and do the most lobbying — but none of that makes them right, especially when there's so much disagreement within the group itself about what AI means for progress. That's a topic I care less about than the fact that they all believe AI will end up vastly more intelligent than humans.
To me, these are two diametrically opposed camps that never frame it that way out loud. Both sides talk as if their point of view were the only possible reality and as if the other side were also in that reality but just didn't understand it fully...
So my small attempt at reflection is to lay out these views and analyze them from the outside, to see where they really stand.
How intelligent are humans as a whole, and how far can machines go?
Let's start with the range of human intelligences. Because for me, there's no single human intelligence, as both camps assume — there are many. Even within the more or less socially accepted IQ, there's a whole range: people with an IQ of 70 or below are considered dependent, since they can't lead a normal life on their own, all the way up to people with an IQ of 150 or higher, who are considered geniuses above 99.9% of the population. (Even so, those individuals are theoretically one in a thousand, so in Spain alone there should be tens of thousands of them...) But of course, it's not all IQ. A person's socioeconomic and cultural situation, their social connections, and their luck play a huge role in their future. And to be clear: a higher-IQ person doesn't feel more or matter more, just as a lower-IQ person doesn't feel less or matter less.
Besides, as I mentioned in the disclaimer at the start, you can be a genius at math and at the same time dyslexic or deaf. You can have perfect pitch and never learn to play an instrument. You can be great at logical problems and bad at spatial ones, and so on. Intelligence has a thousand different components we're still learning to tell apart, which — together with luck and environment — determine whether someone ends up a Nobel laureate or just another working stiff.
Image showing where Terence Tao's IQ falls on the IQ distribution
But so what? Are we the "ceiling" of intelligence? Looking at all this variation, it doesn't seem so. It seems more like there are many intelligences, and some will be better at certain tasks than others.
Animals don't fit into our IQ tests, but they demonstrate intelligence in many ways that are better documented all the time. There are records of crows solving puzzles, chimpanzees memorizing sequences of numbers in the blink of an eye (in the video above), orcas organizing complex hunting plans, apes using tools, animal languages and dialects, and "bilingual" animals that can understand two regional dialects when moved from one to the other... And those are just the intelligences we can appreciate. It's increasingly clear that animals don't have a brain just to carry it around or manage their internal organs — many of them can remember, plan, recognize themselves, navigate, and coordinate at levels we don't fully understand.
All of this paints a much richer and more varied picture of intelligence. Intelligence isn't a monolith; it's more like a set of skills, sharpnesses, capacities, knowledge, and competencies. A chimpanzee has more working memory than us and more agility, but chimpanzees haven't been able to create cultures and accumulate knowledge the way we have. A calculator is orders of magnitude faster and more reliable than the best human at arithmetic or differential calculus. A database is far more reliable, repeatable, and transparent than human memory. But both completely lack agency or the ability to interact with the world. So what can AI do? Where does it fall in this spectrum? How far will it go?
That's the next problem. Let me try to reflect in a couple of charts what I see when members of each camp talk about this.
To me, the humanists see something like this:
For them, AI will clearly never reach humans. It's barely smarter than a reptile or a bird, still not at the level of other mammals, and there are decades, if not centuries, before a machine could resemble a human enough to bring the debate to the table. What we're doing with this debate is distracting the public from the real current problems in machine learning — like bias, the intellectual property of training data, or the people discriminated against by these systems or by not having access to them (all legitimate concerns, but not the focus of this reflection).
Meanwhile, the transhumanists and doomers see something like this:
For them, it's obvious that AI will, sooner rather than later, surpass people, and then there'll be no going back. In short order it will reach escape velocity, we won't understand what it's doing or why, and we'll be at its mercy or under its eternal blessing, depending on the sub-camp.
In fact, after making that chart, I found this other one — a chart by Leopold Aschenbrenner, one of OpenAI's wunderkinds who's now building his own superintelligence company:
In his essay "The Decade Ahead" Aschenbrenner argues that ASI is inevitable and that the escalation in energy, money, and resources to achieve it will continue unstoppably, without hitting any physical, economic, or social wall...
Who's right?
Only time will tell. I think this huge difference of opinion, apart from pure self-interest, comes down to a different view of what makes us human and intelligent. Some look at the tasks where AIs make mistakes and say, "See? They'll never be like us, they don't understand what they're doing." Others look at the tasks where AI excels or surpasses humans and say, "See? Once it's like that across every other discipline, it'll be unstoppable." But after seeing all the variability in humans and animals, and years of engineering work, I lean toward the view that it's always a matter of trade-offs — improving each skill has a cost, and certain tasks require specific tools, context, or data the AI hasn't seen yet... The chart, for me, is really something like this:
Chart generated with Google Gemini 2.5 Pro (https://gemini.google.com/share/bb37aabfa351)
It's what Andrej Karpathy called "Jagged intelligence".
Artificial General Intelligence — AGI — will be something that sits in the middle of that whole spectrum of intelligences we humans display. Some will be good at math, others at recognizing and talking about diseases. There will be AGIs that can't be "autonomous" even if they have some kind of persistent memory and agency. Like children who need taking care of.
We could even argue that current systems like GPT-4 or Claude 3.5 already show capabilities that exceed humans in specific areas (like language processing or information retrieval), while remaining far inferior in others (like causal reasoning or understanding the physical world). LLMs on their own have many problems and are far from having a human's executive capacity. They're not good at planning or executing plans, and they're not good at reasoning or knowing when they're really answering something "just because it sounds right." All of this is being tackled with architectures that include external memories, retries, output control to avoid inappropriate responses, or input control to avoid jailbreaks. But that shows there's a lot of work left. Personally, I think the road to ASI is much longer than people assume right now, and we'll spend a few years in this limbo of AIs that are very powerful in some ways and very clumsy in others — which will bring big improvements in many fields, as we're seeing with AlphaFold or Copilot, but won't be AGIs.
We're seeing it with self-driving cars: they already drive thousands of kilometers a day, but only in specific cities under specific conditions. Expanding the network where they can operate safely is a slow process, because every city and every country has a different context.
Sam Altman himself, in his famous essay "The Intelligence Age", talked about ASI in "a few thousand days" — a measure as imprecise as it is unusual. A few thousand days could be 3,650 days (10 years), or it could be 9,000 (~30 years). That essay should have cooled market expectations enormously, but it didn't: big companies keep spending billions in the hope that if they keep scaling models and datasets, the models will keep improving without hitting a ceiling.
The humanist view isn't wrong that millions of years of evolution produced incredibly complex, sophisticated, and optimized systems. Intelligence is an evolutionary advantage. It's what doomers call a convergent instrumental goal: being smarter is useful for achieving many other things, so whatever you need, being smarter will probably help. That's what grew hominid brains to where we are now: the ability to coordinate, plan, track prey or set a trap, communicate abstract concepts in words, and finally create a shared culture that lasts generations was an advantage over other animals and other hominids. The problem with the humanists is thinking we're the ceiling, or that AI can never reach where we are and surpass us in many ways.
The reality is that intelligence, both human and artificial, is multidimensional, multisensory, and contextual. A system can be extraordinarily capable in one domain while being basic or incompetent in another. That's why the view from both camps strikes me as so jarring, and why I don't understand this argumentative battle that doesn't even start by acknowledging how reductionist it is to assume all humans have a single level of intelligence and that we all agree on what it means.
Update: December 4, 2025
Blaise Agüera y Arcas touches on this topic in his talks and in his book What Is Intelligence?, which is a theory about how life and intelligence are computational, grounded in a mountain of experiments and academic work. It's a wild and wonderful book so far — incredibly well documented and substantiated, with hundreds of references — and you can even read it online! https://whatisintelligence.antikythera.org/ The TLDR is in his talks, for example at Long Now:
References
I'll add some videos and links here at the end, besides the ones scattered throughout the article, for anyone who wants to keep learning about these two visions:
Timnit Gebru on AGI: I've never seen her attack the question directly. She always redirects to bias, discrimination, capitalism, ethics, and the labor market. For her, AGI doesn't make sense as a concept, and merely raising it is an attempt to evade the real problems we have right now:
Geoffrey Hinton gave a talk in early 2024 where he lays out the reasons he believes AI will surpass human intelligence:
A very clarifying video on AI ethics and the discourse in that field — ethics vs. AI risk — is this interview with Anca Dragan, head of AI safety at Google DeepMind. I like it because it goes from the ethical problems of deciding what an AI can and can't answer (which would be part of Timnit Gebru's and the humanists' discourse) to the problems an agentic AI could present when executing actions in the world on your behalf, or even fully independently (which is where the doomers focus). I really like this interview because it's a small glimpse of all the problems behind this, and you can clearly see this isn't a black-and-white issue but a broad spectrum of very difficult problems:
Sabine Hossenfelder is a physicist and science communicator on YouTube who writes about hard topics like this. She recently made a video about AI scaling and how it's a physical inevitability that it hits a wall due to the so-called "decoupling of scales". This means that no matter how much you look at how matter behaves at the macroscopic level, you can never deduce what happens at the quantum level. According to Sabine, this applies to many phenomena, where without the right data and instruments you can't advance knowledge:
In Christmas 2024, OpenAI released a model that beats the ARC AGI benchmark — a benchmark that literally has AGI in its name, and that François Chollet (who we could place within the "humanist" camp as I've defined them in this article) had been promoting for years as a definitive test of generalization in problem-solving. I think this is a milestone on the road to AGI, and even though Chollet and other humanists are already starting to move the goalposts and claim we need a better benchmark to measure intelligence, the truth is this was the one they were pointing to until now — and OpenAI's model shows the models keep saturating every benchmark without hitting a real wall.
Google has published a model of similar capability (Gemini 2.5 Pro experimental), and in recent interviews and statements both Demis Hassabis and David Silver have made clear their intention to surpass human capabilities. Demis Hassabis has said on several occasions that AGI, for him, isn't a model capable of doing what an average human does, but one capable of performing tasks at the level of "the best human" in each field. That is, reaching the level of AlphaGo, but for any task a person can do — which, for most experts, would actually be ASI (Artificial Superintelligence). In this interview, David Silver talks about this idea and how Google intends to achieve it using reinforcement learning:
Top comments (0)