15.10.2026, Moscow.
Attempts to introduce artificial intelligence systems into humanitarian spheres, for example, into education, are doomed because of the way neural networks are trained, Rossa Primavera News Agency‘s IT Desk notes.
One of the founders of OpenAI, Andrej Karpathy (known, among other things, as the author of the term “vibe coding”), explained in a post on X that there is a reason why the enormous successes of neural networks in some spheres are not always matched by their ability to solve tasks in others. This is connected to the mechanism of positive reinforcement, which is a key part of training models.
The point is that at the input of this training, it is given many examples: this is “good,” this is “bad.” Accordingly, the model adjusts itself so as to produce a good result and avoid what was initially identified as unsuitable.
The reason AI has taken off so much in the IT sphere, primarily in programming, is that for these tasks it is almost always possible to say whether a solution is good or not, because a program either gives an exact answer or makes a mistake. In a sense, the model “rejoices,” like a person who has solved a task, and tries to adjust itself so that this “joy” is repeated more often in the future.
Therefore, for this kind of task, where such an answer can be given clearly, neural networks can produce an increasingly deep and serious result. On the contrary, humanitarian tasks, for example, “write an essay on a topic,” cannot be evaluated on a “yes” or “no” scale. Is the essay good? Did the child receive an education? Was the person cured? Who is right in a legal dispute, and why?
These are far more complex questions than those to which a clear answer can be given when training models, and that is precisely why in such spheres AI today produces unconvincing results. And this is a fundamental, not a temporary, technological problem.
Therefore, attempts to give AI ever greater powers beyond the technical sphere are initially incapable of producing the “magical” result expected of them, as does still happen today in the IT sphere, while for the greater part of society what is happening is not yet so noticeable.
Another matter is that there are also classes of tasks in which such clear answers exist. For example, the military sphere. A target is either hit or not. Whether it was possible to kill, hit, shoot down, sink, or not. It is precisely this data that is being collected today in Ukraine with the participation of Palantir, and also, for example, of former Google CEO Eric Schmidt in their military ventures, for which data is collected directly on the battlefield.
Artificial intelligence trained on this data is, in principle, capable of working effectively and producing a corresponding result. Whereas the hopes of domestic “digitalizers” for some miraculous quality of AI application anywhere without regard for fundamental limitations, especially in education, are doomed.
Source: Rossa Primavera News Agency
