AI researchers have inherited more from Descartes than they might realize. The Cartesian worldview divides reality into two realms, the external world of objects and the internal realm of ideas. In this picture, language is simply a mapping device, a way to attach words to ideas in the mind, which themselves represent objects out there in the world.
This representational picture has shaped centuries of thought in philosophy, linguistics, and cognitive science. It implies that for genuine intelligence to arise a system must build internal models of the external world, and that language would be secondary, little more than a label pasted onto those models. If that were true then training a system on text alone, without access to perception or embodiment, should not get us very far. At best it would produce a parrot of surface forms, but not the kind of fluent, quasi-intelligent performance we see today.
And yet that is exactly where we find ourselves. Transformers, which are nothing more than massive self-attention engines trained on next-token prediction, now simulate reasoning, problem-solving, even commonsense dialogue. If the Cartesian picture were correct, this should not be possible.
Martin Heidegger offers a radically different picture of reality and language. For him, the world is not a pile of objects with properties but a referential totality, a web of significance in which beings show up as meaningful through their relations. A hammer is not encountered as a wooden handle with a metal head, it shows up as for driving nails, which are for fastening wood, which is for building a house. We dwell in a world where things matter because of their place in a context of use. Language, in this account, is not a neutral code for labeling objects but a mode of disclosure that articulates significance. Words do not just tag, they open up horizons of meaning. When we speak of the world of theater or the world of basketball, we are not naming objects but disclosing entire contexts in which roles, tools, goals, and actions hang together.
Here lies the striking point. The transformer succeeds because reality is structured in the way Heidegger described. When trained only on text, transformers build high-dimensional relational maps of token co-occurrence. Self-attention constructs a dynamic web in which each word acquires significance from its relation to others. The reason this produces astonishing fluency is that the statistical traces in language already carry the structure of worldhood. Centuries of human discourse have sedimented the relational patterns of significance into text. When transformers learn to predict the next word, they are not learning isolated labels but simulating the same web of relations through which humans encounter meaning.
If Descartes were right, co-occurrence statistics could never yield anything resembling intelligence. But because Heidegger is right, because world and language are relational and disclosive, a machine can simulate understanding simply by recombining linguistic traces.
The same point appears in failure. Transformers hallucinate not because of noisy data or faulty optimization, but because their relational field is flat. Tokens refer only to tokens, with no grounding in temporality, mood, care, or thrownness. Ask when Aristotle lived and when Galileo lived and the model will often return the right dates. Ask how Aristotle learned from Galileo and it may spin a fluent but impossible story, because in the semantic space of philosophy, astronomy, and influence, chronology is not structurally encoded. The transformer is compelled by its architecture to produce coherence even where disclosure is impossible. That is why hallucination is not incoherence but coherence achieved in the absence of grounding.
Heidegger again helps us see why. Significance is real, but without existential structures to stabilize it, it drifts. Transformers can simulate worldhood, but they cannot inhabit it. Engineers patch these failures with retrieval, tool use, and fine-tuning. These scaffolds work up to a point. They constrain the model to fact tables or symbolic solvers, but they do not change the underlying flatness of semantic space. A retrieval module can supply the correct date of Galileo’s birth, but it cannot prevent a model from producing anachronistic narratives when prompted at a higher conceptual level. Scaffolds mask hallucination, but they do not resolve the ontological deficit. This too illustrates Heidegger, the difference between fluency supported by external filters and genuine disclosure grounded in world.
The success of the transformer is therefore not an accident of scale but an empirical clue about the structure of reality. Descartes’ picture of words as labels for inner ideas cannot explain why transformers work. Heidegger’s picture of world as relational significance and language as disclosure explains both their fluency and their hallucinations. For AI researchers this reframes the path forward. The challenge is not simply to build bigger models or better scaffolding, but to design architectures that incorporate constraints resembling temporality, affordances, and concern, in short models that can begin to approximate disclosure rather than merely coherence.
Transformers are not conscious and they are not agents. But their success demonstrates something profound, that the world is disclosed through relations of significance rather than represented as detached objects. This is why a machine trained only on text can simulate reasoning, and it is why its failures, whether hallucinations, anachronisms, or tonal dissonance, map so precisely onto the boundaries of Heidegger’s ontology. The transformer’s fluency is not just a technical breakthrough but also a vindication of Heidegger’s insight that language and world belong together as one web of meaning, and a quiet refutation of Descartes’ centuries-old picture of mind and language