Where the limits of modern AI lie

Computing power for training language models is now discussed at the level of heads of state, not just corporate boards of technology companies. The race for leadership in AI has become a geopolitical factor comparable in significance to control over nuclear technology. But behind this hype lies a fundamental question that the industry prefers not to notice: is there a principled ceiling to what artificial intelligence in its current form is capable of doing?

To answer this question, we must first clarify the terminology. Soviet inventor and methodologist Genrikh Altshuller, creator of the Theory of Inventive Problem Solving (TRIZ), drew a fundamental distinction between two types of tasks. The first is routine tasks. They can be incredibly complex from a computational perspective, require terabytes of data and months of supercomputer work, but their solution lies entirely within the existing system of knowledge. Essentially, these are tasks of searching for and combining what humanity already knows. And it is precisely here that modern language models demonstrate impressive results — they sort through, synthesize, and adapt existing knowledge at a speed and scale inaccessible to the human mind.

The second type is inventive tasks. This is territory where you need more than just to find an answer in the space of the known — you must go beyond its boundaries. Create a new abstraction, discover a pattern that no one has previously formulated, or propose a solution that contradicts common assumptions. This is where things become most interesting — and most troubling for those who believe in the inevitability of "strong" AI.

Modern large language models, for all their impressive performance, operate on the principle of statistical generalization of patterns from training data. They do not "understand" in the sense that humans understand — they recognize structures and reproduce them with variations. This makes them brilliant tools for routine tasks of any complexity: from writing code according to known templates to diagnosing diseases based on accumulated medical data. But when it comes to genuine discovery — seeing what is not in the data — the model faces an epistemological dead end. It cannot go beyond the boundaries of the knowledge space on which it was trained.

Of course, one could argue that models sometimes produce unexpected and even "creative" results. This is true, but such "creativity" is combinatorics, not invention. A model can connect two distant areas of knowledge in a non-obvious way, and the result may look like an insight. However, fundamentally new knowledge — knowledge that is not a recombination of the existing — requires a different cognitive mechanism. What exactly that mechanism is remains a question that neither neuroscience nor philosophy of mind can answer.

The practical consequences of this distinction are enormous. Companies and governments investing billions in AI development must soberly assess what tasks they will face. If it is about automation, optimization, scaling existing processes — language models will handle it, and better each year. If the bet is that AI will make a fundamental scientific breakthrough or create an entirely new technology without human involvement — those expectations are probably too high. At least with the current architecture of models.

There is also a deeper level to the problem. The inter-state LLM race creates a dangerous illusion: whoever builds the most powerful model first will gain a strategic advantage in everything. But if an epistemological ceiling exists, then increasing computational power provides only a quantitative increase in solving routine tasks, not a qualitative leap toward machine superintelligence. This does not diminish the importance of AI — automating routine work itself transforms the economy and society. But it means that human intelligence, capable of genuine invention, remains an irreplaceable resource.

The question of AI's limits is not a judgment on the technology, but an invitation to an honest conversation about its nature. The better we understand what machines can and cannot do, the more effectively we can build a symbiosis of human and machine intelligence. And perhaps it is this symbiosis — not the race for an omnipotent model — that will become the true breakthrough of the coming decade.