Why Language Models Will Never Become AGI: A Wittgenstein Lesson from a Century Ago

More than a hundred years ago, a schoolteacher from Austria-Hungary wrote a sentence that now reads like a verdict on the entire large language model industry. "The limits of my language mean the limits of my world" — this thesis from Ludwig Wittgenstein's 1921 Tractatus Logico-Philosophicus has turned out, unexpectedly, to be the most accurate diagnosis of a technology into which the world has poured hundreds of billions of dollars.

To see why this matters now, it helps to recall the context. Only two or three years ago, the industry was euphoric. Every new version of GPT, Claude or Gemini showed an impressive jump in capability. Models learned to write code, analyze images and solve olympiad problems. AGI seemed to be just around the corner: all it would take was more data, more parameters, more compute. Investors poured in money, corporations rebuilt their strategies, and public speakers competed in predicting exactly when a machine would surpass humans. Today, in 2026, the tone of the conversation has changed noticeably. The word "bubble" is heard more and more often, and skeptics are gaining more arguments.

This is the moment to return to Wittgenstein. His central idea is both simple and radical: language is not merely a tool for describing reality, but the very boundary of what we are able to think. Anything that exists beyond language simply does not exist for a linguistic being.

Apply that principle to LLMs and you get not a metaphor but a literal description of an architectural limitation. A large language model operates on tokens. It predicts the next fragment of text from 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 limits of that text really are the limits of its world.

Critics may object that modern multimodal models already work with images, sound and video. Isn't that a way out of language? Here it is important to distinguish between processing a signal and genuine perception. When a model "sees" a photograph, it converts pixels into numerical representations and matches them against text descriptions from its training set. That 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 only knows the contexts in which the word "red" appears alongside certain numerical patterns. Wittgenstein would say the model is playing a language game without access to what the game points to.

There is a second aspect of Wittgenstein's philosophy that hits the mark. In his later work he arrived at the idea of "language games": the notion that the meaning of a word is determined by its use in a particular practice. Understanding is not extracting 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 mean not knowing the definition but the experience of driving nails, the weight of the tool in the hand, the muscle memory of the swing. An LLM can describe a hammer flawlessly, list its types, quote the instruction manual, but it has no experience of driving a nail and cannot have one. Its "understanding" is a simulation, devoid of a bodily foundation.

This does not mean language models are useless; on the contrary, they are enormously useful within their linguistic universe. They excel at tasks that fit entirely within text space: editing, translation, code generation, summarization, finding patterns in data. The problem lies not in the models themselves but in the inflated expectations placed on them. When tech executives promise AGI in two or three years, they either do not understand the nature of the limitations or are knowingly feeding 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: interacting with the physical world and forming internal models of reality through experience, not through reading texts about experience. Robotics, neuromorphic computing and hybrid architectures combining symbolic and connectionist approaches are all potential directions, but none has yet come close to solving the fundamental problem.

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