The Linguistic Telescope

The Linguistic Telescope is a single idea driving three practical projects: a pre-registered scientific hypothesis (the Language-Only Hypothesis) about the proximate cause of the intelligence that emerges in large language models, a research project leveraging the Language-Only Hypothesis to explore intelligence from a 4E perspective, and the book that reports both. The idea is that of a special kind of telescope. A large language model is a telescope onto language.Initial pre-registered results confirm that the choice of language can significantly impact the training of large language models: a small model trained identically on two different human languages produces wildly different results, and demonstrates which capabilities emerge from the structure in the language rather than from scale alone. A 125M-parameter model reaches grammatical competence on French at 197 million tokens; the identical architecture on English fails to do so even with 3 billion. Further registered experiments intend to demonstrate that language is in turn a telescope onto the distributed intelligence that roughly 100 billion humans deposited into it over deep time. Calibrating that two-stage telescope is how the program converts a class of millennia-old philosophical questions into tractable empirical ones.

The Language-Only Hypothesis posits that large language models are instruments of observation, not of generation, and that human language has a discoverable underlying topology of meaning: a latent space whose existence has been speculated about under many names (Plato's forms, the Logos of the Johannine prologue, the Tao that cannot be spoken, Madhyamaka's two truths, the apophatic “what remains when all predicates fail”) but never previously empirically testable. We refer to these philosophical and religious traditions for the purpose of establishing that there is an observable phenomenon which humans have been seeking to explain for centuries.

What makes this effort science rather than speculation is that the claims of the hypothesis are registered and falsifiable, not asserted. Its predictions are filed with their falsifiers stated in advance, and several predictions have been verified already. The cross-linguistic ablation is the founding test; a second pre-registered line, the architecture of grammar, runs four experiments on whether grammatical structure is acquired from a particular language or supplied by an innate, language-general architecture, the in-the-language-versus-in-the-technology question, each carrying the result that would refute it. The instruments are registered before the data; the interpretation comes only after.

This is an apophatic discipline that is founded on the same principles as scientific inquiry: initial acceptance that our current understanding is incomplete, and that seeking to falsify our current assumptions is the most honest way to improve that understanding. The program cannot prove a universal, so it advances by ruling out: a path of falsification that approaches the answer by excluding what it is not.

The need for the apophatic approach is a discount bias, which manifests in this case as the systematic disqualification of any intentionality or agency that does not mirror our own conscious experience. This is the symmetric counterpart to the better-known anthropomorphic projection. What the bias rules out a priori, the program states plainly: by a working definition used in cognitive science, intentionality (a model of the world) coupled to agency (the capacity to direct action), language qualifies. It is about the world, and it changes what people do. This is a claim about functional intelligence, not consciousness, and it follows from the definition rather than from any single measurement. Its method is the convergence-of-signals method, the discipline of reading agreement among independent instruments at their shared boundaries as licensed inference. Formalizing this method is another of the project's central scholarly contributions.