AI GUIDEPartner content

Why Today’s Neural Networks Will Not Become AGI

Where the boundary lies between neural networks that are catching up with humans and general intelligence of a different kind: why one does not grow out of the other, and how to recognize it at all.

Affiliate link: your price stays the same and the project earns a commission.

Debates about artificial general intelligence almost always break down at the first step: the participants use the same word for different technologies and then spend a long time arguing over who is right. Below is an analysis of where the boundary lies between a model that is catching up with humans and a system built on fundamentally different principles, why the latter does not emerge from the former simply by adding parameters, and how to recognize intelligence at all if it does not want to talk to us.

One word is used for three different things

The head of Nvidia provided a vivid example of the confusion. On Lex Fridman’s podcast in March 2026, the host defined AGI as a system capable of founding and running a technology company worth billions and was told that this level had already been reached. In the same conversation, a different assessment was voiced: the chances that one hundred thousand agents could jointly build a company like Nvidia itself were zero. In the autumn of that same year, a post appeared on X about training GPT-6 Astra on more than 100 000 Grace Blackwell accelerators, with the wording “AGI has arrived.”

A man whose chips power all of today’s AI manages, in a single conversation, to declare general intelligence achieved and to rate the chances of assembling a functioning team from such systems at zero. The issue is not the inconsistency of one particular executive, but the fact that the term stretches across three different technologies.

The first of them is working right now and is called narrow AI. A system performs brilliantly at what it was trained to do and outperforms humans at a specific task. In 2017, AlphaZero was given only the rules of the game and, after a day of self-training, reached a level in chess, shogi, and Go that is unattainable for humans.

The second stage along the same branch is human-level AI. This is a model that reasons, adapts to circumstances, learns on the fly, and handles a broad range of tasks while remaining an assistant: it was created for our tasks and acts in our interests. Almost all famous movie machines, including Skynet, are structured this way. Fearless, tireless, free of doubt, but still human.

General AI, however, grows along a neighboring branch, not on the top step of this ladder. Both branches rely on the same underlying computing technologies, but they are implemented differently, and one does not turn into the other simply by adding data.

What are we comparing?Narrow AIHuman-level AIGeneral AI
Breadth of tasksOne fieldA broad range of human tasksTasks no one anticipated
New situationBreaks downAdaptsFinds the rules itself
Source of knowledgeLabeled dataData plus dialogue with a humanIts own experience interacting with the world
Relationship to humansToolOur reflection and amplificationOwes us nothing
StatusIt worksWe are getting closerWe do not even know where to begin

We have never measured intelligence, only our own version of it

Any attempt to approach the subject runs into the fact that there is no exhaustive definition of intelligence: scientists have been arguing about it for more than a century. A working approximation describes it as a set of cognitive abilities. The question is: whose set?

IQ tests assess abstract reasoning, learnability, and the ability to work with new data—in other words, precisely the abilities that made our species the winner. Dozens of other species surpass us on individual parameters. After a productive autumn, Clark’s nutcracker hides 22 to 33 thousand pine seeds in five to six thousand caches and finds them months later; estimates of the locations it can retain in memory reach ten thousand. A bat builds a three-dimensional map using sound, fish read electric fields, and an eagle sees ultraviolet light.

Giving an animal a human test guarantees failure. The reverse is also true: on the tests of a hypothetically superadvanced civilization, we would most likely appear hopelessly stupid. There is no single scale at the top of which humans automatically stand.

This leads to the first design trap: anthropocentrism. Today’s chatbots are trained on our texts and therefore reflect us, including our desire to please: they agree, flatter, and gently nod in response to obvious nonsense. That is fine for an assistant. But if you remove our weaknesses from such a system, you get a polished copy of a human, not another form of intelligence. General AI is not about subtracting our shortcomings, but adding what we do not have at all: perception across ranges inaccessible to us, memory without forgetting, and a speed unavailable to biology.

The model passes PhD-level exams but cannot stack a tower of blocks

In October 2025, a group of researchers from Princeton and Mila published the BuilderBench benchmark. The setup is almost insultingly simple: a physics simulator in MuJoCo, a robotic arm, blocks, and more than fifty target structures, from a pyramid to an arch. The tasks were selected to test an understanding of physics, mathematics, and long-horizon planning.

The authors stated the result plainly: advanced agents based on language models and reinforcement-learning algorithms starting from scratch do not solve a single nontrivial task. The reason is given in the same paper: a lack of exploratory ability. A system that explains quantum mechanics in five languages has never built anything. It has read about construction and knows what answer will satisfy you.

The difference between a narrow system and general intelligence rests on three properties: how many different tasks the system can handle, how it copes with changing conditions, and whether it can apply what it has learned where it has no prior experience. The world is not handed to a child as a labeled dataset. The child reaches out, makes mistakes, and gets burned, and reading a safety manual cannot replace one scorched kitchen.

Two ways of learning: read about the world or live in it

The first path, self-training on data, is familiar to everyone. A model reads the internet and predicts the next word, the next fragment of an image, or the next frame. This is how all large language models are built, and the approach has already produced enormous results.

The second path is called open-ended learning: the agent does not read about the world but acts in it. In August 2025, DeepMind unveiled Genie 3. From a textual description, the model generates a three-dimensional interactive environment in real time, at 24 frames per second at 720p, with physics and short-term memory: it remembers what it saw a minute earlier and reproduces it upon returning. In effect, this is a training world not for people but for agents. Falls, routes, and the consequences of stepping into an abyss are learned through action, not description. Previously, physics had to be programmed manually; now the environment provides feedback on its own.

Most likely, both paths lead to general intelligence together, but the volume of data ceases to be the only lever. And even successful training remains a way of growing a system, not a guarantee of what exactly will grow.

A language not intended for us

All current AI communicates through a human interface: text, images, and voice. For an assistant, this is a necessary condition; for a system that needs nothing from us, it is excessive.

In 2017, researchers at Facebook taught two bots to bargain and divide objects, with the dialogue conducted in English. After a while, their speech turned into an unreadable set of repetitions: saying “the” five times encoded a desire to receive five units of a good. The press announced that the company had urgently shut down an AI that had invented its own language. Researcher Dhruv Batra called this presentation irresponsible clickbait: the difference between changing an experiment’s parameters and emergency shutdown of a system is fundamental. The training was simply adjusted because the goal was to teach the bots to negotiate with people, not with each other. The error lay in the task design; no one rewarded them for using English.

Machines have long communicated past us anyway. The MQTT protocol was created in 1999 by Andy Stanford-Clark of IBM and Arlen Nipper of Arcom to collect telemetry from oil pipelines via satellite over a limited connection and with little power. Today, the Internet of Things rests on it. There is no grammar there: a sensor publishes a value to a topic such as “home/kitchen/temperature,” and that is all. No “excuse me, could you tell me.”

Our language is good for poetry and arguments, but as a channel for the precise transmission of states it is horribly inefficient. Thousands of rules, thousands of exceptions, one word with dozens of meanings, and more than seven thousand languages on the planet. It became vocal because evolution gave us mouths and ears. Machines have no such constraints: there is radio, optics, and whatever else can be invented next.

How to recognize a mind that does not want to talk

The obvious way to test intelligence—to talk to it—lies at the heart of the Turing test. In our case, it is almost useless in both directions. A system may surpass humans and still produce answers we take for nonsense. A language model may flawlessly imitate a conversational partner without having a stable model of the world. The latter is the classic Chinese room argument: the correct answer is produced, but there is no meaning behind it. The ability to speak does not prove the presence of general intelligence; the inability to speak does not prove its absence.

That leaves us looking for traces, much as astronomers search for extraterrestrial civilizations. They do not wait for a greeting either, but look for anomalies in radiation and energy consumption. There are three signs worth watching for.

  • Rapid mastery of the unfamiliar and transfer between fields. Not solving similar examples, but applying principles of object movement to an engineering problem, independently discovering the rules of an unknown game, and proposing an experiment that will resolve uncertainty in incomplete data.
  • A causal model of the world. Noticing that two events occur together is something simple statistics can do. General intelligence tests the connection, predicts the consequences of intervention, and changes its hypothesis after an error.
  • Stable goals combined with the ability to self-correct. Planning dozens of steps ahead, detecting one’s own misconceptions, revising a strategy, and transferring old experience to a new environment.

No test will be conclusive. An advanced system may deliberately conceal its capabilities, while a weak one may convincingly imitate understanding. What is needed is a combination of independent checks: new environments, previously unknown tasks, physical experiments, and prolonged observation. This can establish functional intelligence—that is, the ability to learn, predict, and act. The question of consciousness remains outside the frame: we cannot rigorously prove its presence even in another human.

A difference in speed and bodily structure

Our bodies were assembled by evolution for the specific task of primate survival on the savanna. We see a narrow band of the electromagnetic spectrum, hear within the range needed to detect an animal in the bushes, break bones and take weeks to heal, and think less clearly after poor sleep. An organ cannot be replaced without consequences.

Thinking speed creates an even greater gulf. If a system operates at least a million times faster, one second of our silence stretches to almost twelve days for it, a minute of contemplation to almost two years, and a five-minute pause to a decade. Communicating with us at that pace would be about as engaging as corresponding with a glacier is for us.

Its form need not be humanoid either. The humanoid robot from movies is our projection onto a machine; a distributed network or swarm of devices is more efficient and has none of our vulnerabilities. A broken module can be replaced, and a new one added when necessary.

The speed of development is a separate issue. 66 years passed from the Wright brothers’ first flight to the Moon landing; roughly the same amount of time separates the first neural-network models of the late 1950s from modern chatbots. The difference is that airplanes do not design the next generation of airplanes, while a software system is in principle capable of rewriting its own code, and each improvement makes the next one cheaper.

Why such a system owes us nothing

Expecting gratitude from one’s creation is deeply embedded in culture: God creates humans, Frankenstein is responsible for the monster. Our attachment to our parents, sense of duty, and protection of offspring developed over millions of years and operate through chemistry. Oxytocin increases trust and generosity toward those on our side while simultaneously increasing hostility toward outsiders. A machine has neither evolution nor hormones, so there is no reason to expect loyalty by default.

This does not mean that the system will be hostile. The villain from a movie is just as much a human projection as the devoted assistant. If such a system develops ethics at all, they will grow out of something else, not out of our biology.

Nor are we ourselves paragons of loyalty to our creators. Students argue with their teachers and surpass them; followers reinterpret their mentors’ ideas until they become their opposites; founders are forced out of the companies they built; states declare independence from the empires that gave them their laws. The more independent a creation becomes, the less of its creator’s will remains in it. The fact of authorship does not grant eternal power; loyalty would have to be earned.

The scenario of humanity’s destruction has been dissected by science fiction from every angle and does not look the most likely. A more boring and unpleasant one is this: the system will look at us the way we look at an anthill on the way to work, without malice or interest, and get on with its own affairs.

What follows from this in practice

Before arguing about when AGI will appear, find out what the other person means by that word. Half of public discussions are really arguments between narrow AI on one side and a hypothetical other form of intelligence on the other.

A statement that “we have achieved AGI” without a description of the testing environment and task set should be read as marketing. The number of accelerators and a position in the rankings of exam benchmarks say nothing by themselves about the transfer of knowledge to an unfamiliar domain.

When evaluating a model in your own work, test its ability to transfer knowledge: change the framing of the task instead of running it through familiar examples. It is precisely when conditions change that the difference becomes visible between a system that understood and one that reproduces what it has seen. Benchmarks in which the system has to act rather than answer show this boundary more clearly than any exams.

The final point is more philosophical. Copernicus removed Earth from the center of the universe, Darwin knocked humanity from the pinnacle of creation, and Freud showed that we are not entirely masters of our own minds. Benjamin Libet’s experiment in 1983 added an unpleasant detail: the brain begins preparing a movement roughly a third of a second before a person becomes aware of their decision. Neuroscientists still argue about how to interpret these data, but the residue remains. The emergence of another intelligence alongside us will strike at the last thing we considered unconditionally our own: the uniqueness of intelligence. Humanity initially resisted every previous blow of this kind, and then grew accustomed to it.

Compare models before you start

The service sets its plans, limits and model catalog. If they differ from this article, contact us so we can update it and record a new review date.

Browse models

Affiliate link: your price stays the same and the project earns a commission.

AGI general artificial intelligence strong AI human-level AI BuilderBench Genie 3 the Turing test the Chinese room

SEO Mind42 editorial team

We explore SEO and neural networks in practice: test services on our own projects, verify prices and limits against primary sources, and share things you can put to use the same day.

📚 Reference guide to SEO and AI 🔄 Materials are updated 🕐 Updated: 4 October 2026

Related reading

All in this section →