This paper offers a categorical diagnosis of the incompleteness of the stochastic paradigm of arti- ficial intelligence. The dominant architectures of the current paradigm — autoregressive language models, diffusion and energy-based models, joint-embedding predictive architectures — are treated as objects of a single categorical class: symmetric monoidal closed categories that lack Cartesian clo- sure. Relying on the Curry–Howard–Lambek correspondence and on Seely’s theorem on the internal language of such categories, the work shows that six structurally distinct deficits of the paradigm — the failure of exact reference, the absence of the rule as a first-class object, the normative vacuum of truth, the collapse of the constructive witness of a proof, the impossibility of causal surgery, and the collapse of invariant semantics — stand in a precisely determined yet non-uniform relation to the ab- sence of Cartesian closure. This relation is unfolded as three tiers of a single categorical ladder: four axes constitute the analytic content of the missing structure itself; the absence of a truth-value classi- fier is entailed by it as the condition of the topos superstructure; the collapse of invariant semantics inherits it at the higher — functorial and derived — levels. Cartesian non-closure is introduced not as a common name for six independent observations, nor as a sixfold equivalence, but as the lower bound of this ladder: a single categorical ascent that restores Cartesian closure and erects upon it a topos and derived environments resolves all six. The diagnosis is reinforced by two independent limits of a different mathematical nature — the TC0 computational ceiling for transformers and the information-theoretic ceiling of Kolmogorov incompressibility; all three limits are proved by their own means and lock the paradigm independently. The structural diagnosis is matched against the empirical “jagged” profile of the cognitive capacities of present-day models: the line of failures coincides with the boundary of what is expressible by the means of symmetric monoidal closed cate- gories, and a control check on the JEPA architecture confirms that the diagnosis holds at the level of the paradigm. The relation of the new language to the old is formulated as a generalized correspon- dence principle. The upshot: the limitations of the stochastic paradigm do not reduce to a shortage of data, parameters, or engineering optimization — they belong to the very mathematical language in which this paradigm is realized.
Egor Vikhlyaev (Mon,) studied this question.