The Limits of Language Are the Limits of Superintelligence: Why Large Language Models Can Traverse but Not Expand the Space of the Thinkable

Abstract

Preprint · Version 1 · September 2026 · Not peer reviewed If the limits of language are the limits of the thinkable, then the possibility of linguistic superintelligence depends on whether a system can expand its own language. This paper argues that large language models (LLMs) cannot do so in a closed regime. We introduce a distinction between traversal (moving efficiently through an existing conceptual space) and expansion (enlarging that space), and argue that LLMs can exceed human performance in traversal without acquiring autonomy in expansion. Drawing on Wittgenstein, the weak form of linguistic relativity, and analyses of vocabulary control in Orwell and Foucault, we treat language as the boundary condition of conceptual space rather than a neutral medium. We then examine the literature on model collapse, which shows that recursive training on model-generated text erodes the tails of the linguistic distribution and progressively narrows what the model can express. Crucially, the regimes that avoid collapse do so by retaining human-generated data, confirming the models' dependence on an external source. We address the strongest objection, reinforcement learning with external verifiers, and argue that it confirms rather than refutes the thesis: expansion requires a signal exogenous to the model's own generation. We conclude that the engine that pushes the frontier of the thinkable remains exogenous to the model. Current LLMs are best understood as instruments of unprecedented traversal power operating within a space they did not open and cannot, by themselves, enlarge. About this work. This is an interdisciplinary essay at the intersection of philosophy of language and machine learning. It offers a conceptual argument rather than new experiments: it takes the published empirical results on model collapse (Shumailov et al., Nature 2024; Gerstgrasser et al. 2024; Dohmatob et al. 2024) as premises and draws out their philosophical consequences for the debate on recursive self-improvement and superintelligence. Comments and criticism are welcome at the correspondence address in the PDF. A revised version will be submitted for peer review; this record will be updated with the DOI of the published version if accepted.