AGI: Why It Won't Happen and What Model We Can Create Instead

This article is a logical continuation of my article about AGI. And today I want to explain why the AGI that we are told about is impossible and what we can offer instead.

Let me remind you that AGI is general artificial intelligence, capable of autonomous learning and solving tasks for which it was not specifically designed. Complete autonomy and independence.

AGI is the next stage in the development of AI. Such a model will be able to study and solve problems independently, make new scientific discoveries, and in principle, it is a cure-all for problems. At least, that's what modern tech giants believe, especially Elon Musk and Sam Altman, who fervently promote the ideas of AGI and are about to release such a model. But, unfortunately or fortunately, this is impossible, and here’s why.

Is it impossible to create AGI?

If we understand AGI as an absolute solver of any tasks—yes, it is mathematically impossible.

In May 2025, a scientific proof was published by a professor of applied sciences, Max Schleret. In his article, he presented a fundamental argument against the possibility of creating AGI. It concerns the barrier of infinite choice.

The essence is that there are tasks where the number of solutions reaches infinity. And infinite choice leads to infinite computational resources. This is what makes AGI impossible. The impossibility of AGI is also supported by empirical data, such as research from Apple, DeepMind, OpenAI, and Anthropic.

An AGI capable of autonomous creativity and solving absolutely any tasks faces situations where the space of possible solutions or interpretations approaches infinity. In simple terms, the model perfectly solves problems within its subject area but is absolutely incapable of transcendence, of knowing through sensory experience.

One of the key barriers for AGI is the inability to transcend the frame, that is, to go beyond the given model. A vivid example is the transition from Newtonian mechanics to the theory of relativity. When experiments revealed discrepancies with classical predictions, the solution was found not within the framework of the old theory, but through the introduction of a fundamentally new concept – the invariance of the speed of light. An algorithm cannot independently detect the limitations of its model: for it, anomalies appear either as noise or as data errors. It is unable to create a new semantic primitive that changes the very rules of the game.

To guarantee the finding of an optimal solution in an infinite space, infinite computational resources and time will be required. Consequently, AGI is physically impossible mathematically and logically.

Should intelligence be an absolute optimizer? Humans, for example, work differently. We do not enumerate all options – we use heuristics, experience, and context.

Let's draw an analogy with our brain:

1. We immediately cut off meaningless options without even considering them

2. We rely on past experience and act by analogy

3. We make reasonable decisions rather than seeking mathematically optimal ones

4. We rely on sensory experience and intuition

We do not solve abstract problems in a static vacuum where all factors are known in advance and nothing changes. We solve problems in a dynamic reality where conditions can change at any moment – and our solution must take this into account.

Therefore, Shleret's argument is not a verdict but merely limitations imposed on development.

Computational Problem

Some scientists, such as Sundar Pichai, the CEO of Google, believe that creating AGI will be extremely difficult with current hardware.

Here, the assessment is based on the idea of scaling, and the growth of LLM parameters has an exponential cost. For example, Llama 3.1 405B was trained on 810 GB of VRAM and energy costs comparable to the annual consumption of a small city. And we are talking about trillions of parameters? About tens of trillions?

This is a dead-end path not only logically but also economically. Scaling LLM is a race for diminishing returns, where each subsequent percentage of intelligence will cost geometrically more, without bringing us closer to true understanding. This approach has a physical and financial ceiling.

The Economic Problem

The sharp rise in demands for hardware and energy directly leads to hyperinflation of investments. Predictions paint futuristic pictures: the AI market will reach $4.8 trillion by 2030, and technologies will affect 40% of jobs.

History repeats itself; a similar bubble has been inflated around AI as it once was with dot-coms. But the problem is not the amount of money, but where it is going. The internet was overvalued in the moment and undervalued in the long term. Now, life without the internet is literally impossible.

If the bubble bursts, investments in AI will simply disappear. Money will leave, startups will close, research will be curtailed. But in the long run, this could turn out to be a huge plus. The race for parameters, which is currently consuming all resources, will end. Those who remain in the field will be either niche players with real products or researchers who do not need to chase after GPT-7 and gather clusters of a million and one graphics cards.

It is precisely then that space will open up for experiments with new architectures. Not because money will magically flow into the right projects, but because the madness will end when money goes to those with more parameters.

The main task is to not throw the baby out with the bathwater. If the bubble bursts too violently, along with the money for scaling, funding for fundamental research will die too. And without it, research will perish.

Hello Human

I believe that the story will repeat itself with AGI. There is currently a constant increase in LLM models. And when they are no longer at 660 billion parameters, but at a trillion and more, they will simply be called AGI. No qualitative leap will occur – just another rebranding, as has already happened with "expert systems" instead of AI or "machine learning" instead of neural networks.

But such models will not gain understanding. They will be a very smart and very expensive parrot. Just now it will speak so convincingly that people will take it for intelligent.

And here we encounter another question – even if we call it AGI, will such a tool be able to solve tasks that require human participation?

There is one social aspect, or rather – the question of humanity as such. AGI (or whatever it will be called) is not a panacea for all problems, but a powerful tool. But will this tool be able to make prudent decisions without empathy? Should it have empathy?

In 1966, Eliza, a chatbot that imitated the work of a psychotherapist, appeared in the UK. It operated on a very simple principle: matching patterns and substituting ready-made phrases. For any user request, it had a template response. And people believed it. They opened up, cried, thanked the therapist who simply rephrased their own words.

And here we find ourselves in a beautiful deadlock:

• If AGI does not need humanity, then we lose the essence. It will not be able to solve all tasks. It will struggle with tasks that require emotions, empathy, or understanding of human nature. Psychology, upbringing, diplomacy, management – all of this is not our case. This means it is not AGI, but just a very smart, yet useless calculator in human affairs.

• If we endow AGI with humanity, then it is no longer a tool, but a subject. And subjects have their own interests. And there are no guarantees that our interests will align.

The solution to the problem needs to be found somewhere in the middle, in philosophy. First, we need to understand, in principle, what intelligence is.

So what now? What about the architecture?

Let’s clarify right away, I want to create an understanding model, not AGI. A model that understands what it is doing, and then simply develop functionality and scale it.

Here we can proceed as follows: take a modular architecture for scaling functionality. This idea comes directly from my first article on project development. Next, we will break down our problem, or rather understanding.

Understanding is the process of making sense of something. When we hear the word "apple," we understand what it refers to because we have seen an apple, touched an apple, tasted an apple, and so on.

The problem is that the model does not understand. All it knows is that this binary code is the same as an apple, and that this code appears in this and that context. That is the extent of its understanding. In other words, just having text is not enough; other data is needed: sound, images, and the dynamics of time.

Modern LLMs are very expensive and intelligent parrots; they assemble text. They analyze large amounts of data and simply match the most popular options. This is how transformers work; we will take a detour and try to give the model a sensory experience.

How? By training on videos. I heard about this idea in an interview with Yann LeCun. It turns out that many companies are doing just that (DeepMind, OpenAI, FAIR).

It’s long, costly, and complex, but it meets the need for sound, images, and text. That is, if we show the model videos of apples, it will understand what an apple is through analysis in general. A child does not learn from textbooks; they observe the world. Therefore, we create a specific connection between text, images, sound, and temporal dynamics. The model then builds neural connections just like a human.

This contributes to building world models, and this architecture is currently key. What is this?

A world model is an architectural approach where the model simulates the surrounding world.

That is, we can implement a conditional text simulation, for example, like the thinking mode in ChatGPT, in real-time, allowing the model to understand.

There are also three main approaches to implementing AGI:

1. Cognitive – Reproducing high-level human thought processes by combining various AI paradigms

2. Neuromorphic - Directly modeling the human brain in the hope that intelligence will arise from its structure

3. Chemical – Creating self-organizing systems at a low level (similar to chemical reactions) so that intelligence arises naturally

I'll say it straight, implementing something like this at home or alone is impossible, so it doesn't work for us. I think the hybrid approach of implementation through different tools, for example, as I wrote in previous articles, making memory through a database is the simplest and most effective approach.

My hybrid approach is an attempt to take the best from cognitive (high-level modules) and neuromorphic (learning from experience).

Conclusion

My approach has proven to be foundational in the scientific community, and I am not proposing anything new at all; everything was said before me, I just indicated what and how specifically I want to use.

This article is mostly written to understand the basic concepts and figure out what to do and how to do it, now the task is to try to do it at home and for cheap.