Why language models will never become AGI: Wittgenstein’s century-old lesson

More than a century ago, a schoolteacher from Austro-Hungary wrote a phrase that today sounds like a verdict on the entire industry of large language models. "The limits of my language mean the limits of my world" — this thesis by Ludwig Wittgenstein from his "Tractatus Logico-Philosophicus" of 1921 has unexpectedly proven to be the most accurate diagnosis for a technology in which the world has invested hundreds of billions of dollars.

To understand why this matters right now, we need to recall the context. Just two or three years ago, the industry was in a state of euphoria. Each new version of GPT, Claude, or Gemini demonstrated an impressive leap in capabilities.

Models learned to write code, analyze images, and solve Olympiad problems. It seemed that general artificial intelligence—AGI—was just around the corner, requiring only more data, more parameters, more computational power. Investors poured in money, corporations restructured their strategies, and public speakers competed in predictions about when the machine would surpass humanity.

Today, in 2026, the tone of the conversation has noticeably changed. The word "bubble" is heard more often, and skeptics have gained more and more arguments in their favor.

It is precisely at this moment that we should return to Wittgenstein. His central idea is both simple and radical at once: language is not merely a tool for describing reality, but the very boundary of what we are capable of thinking. Everything that exists beyond language simply does not exist for a linguistic being.

Transpose this principle onto an LLM—and you get not a metaphor, but a literal description of an architectural limitation. A large language model operates with tokens. It predicts the next fragment of text based on statistical patterns extracted from a gigantic corpus of data.

It does not perceive the world directly—it does not see, does not hear, does not feel pain, does not experience hunger. Its entire "world" is text. And the boundaries of this text are indeed the boundaries of its world.

Critics might object: modern multimodal models already work with images, sound, and video. Is this not a stepping beyond language? Here it is important to understand the difference between signal processing and genuine perception. When a model "sees" a photograph, it transforms pixels into numerical representations and correlates them with textual descriptions from the training corpus. This is not vision in the human sense—it is a complex system of cross-references. The model does not understand what the color red is; it knows only the contexts in which the word "red" appears alongside certain numerical patterns. Wittgenstein would say that the model plays a language game without access to what that game refers to.

There is a second aspect of Wittgenstein's philosophy that hits the mark precisely. In the later period of his work, he arrived at the idea of "language games"—the view that the meaning of a word is determined by its use in a specific practice. Understanding is not the extraction of abstract meaning from a dictionary, but the ability to act in the world in a certain way.

When we say "I understand what a hammer is," we do not mean knowledge of a definition, but the experience of driving nails, the weight of the tool in one's hand, the muscle memory of the swing. An LLM can flawlessly describe a hammer, enumerate its types, quote the instructions for its use—but it has no experience of driving nails, and cannot have such experience. Its "understanding" is a simulation devoid of bodily foundation.

This does not mean that language models are useless—quite the opposite, they are incredibly useful precisely within the limits of their linguistic universe. They excel at tasks that fit entirely within textual space: editing, translation, code generation, summarization, pattern discovery in data. The problem lies not in the models themselves, but in the inflated expectations placed upon them. When technology company executives promise AGI in two or three years, they either do not understand the nature of the limitations, or are deliberately fueling investment hype.

The path to general artificial intelligence, if it exists at all, almost certainly lies beyond the purely linguistic paradigm. It will require systems capable of embodied cognition—interaction with the physical world, the formation of internal models of reality through experience, rather than through reading texts about experience. Robotics, neuromorphic computing, hybrid architectures combining symbolic and connectionist approaches—all of these are potential directions, but none have yet come close to solving the fundamental problem.

Wittgenstein died in 1951, unaware of computers, neural networks, or tokenization. But his intuition about the nature of language and understanding proved prophetic. The limits of language are indeed the limits of the world. And as long as we build intelligence confined to language, we build something impressive, but fundamentally limited. Acknowledging this limitation is not pessimism, but a necessary step toward an honest conversation about where the artificial intelligence industry is actually headed.