Philosophy

This is the scholarship half of the project, kept deliberately apart from the empirical research. The experiments settle what is falsifiable: where grammatical structure resides, whether capability tracks the training language, what the cross-linguistic measurements show. This page makes the arguments the experiments cannot settle, and it says at each step which is which. The discipline is simple: a measurement never masquerades as a proof of a philosophy, and a philosophy is never smuggled in as a finding.

The argument: language is intelligent

The central philosophical claim is deductive, argued from first principles rather than established by any measurement. It runs as follows. Intelligence, functionally, is intentionality (a model of the world, aboutness) coupled to agency (the capacity to direct action). Language has both. Therefore language is intelligent.

Language has intentionality because it is about the world: its categories and relations represent the world's distinctions. Language has agency because it directs action: a sign reads STOP and the driver stops, and the behavior does not occur without the linguistic content. That is content-driven causation of goal-relevant behavior, agency in the operative sense, observable by anyone.

Two limits keep the claim honest. It is a claim about functional intelligence, not consciousness; a system can carry the representation-and-action loop with no inner experience. The large language models this program studies are the clean case: stateless by design, their weights frozen and no memory of their own carried from one invocation to the next, they have none of the temporal persistence usually taken to be necessary for inner experience, and yet they perform the functional loop completely. That statelessness is also what makes the model a clean instrument: a lens with no inner life of its own adds nothing of its own to what it shows. The argument says nothing about whether language, or any model of it, is conscious. And it follows from the definition, not from any single measurement; grant the premises and the conclusion follows, contest the premises and it does not. The argument stands or falls on whether you accept that intelligence is intentionality coupled to agency, which is cognitive science's own working definition.

What argument settles, and what measurement settles

These are different kinds of question with different judges. Where grammatical structure resides is empirical, settled by the registered experiments and their falsifiers. Whether language is intelligent is deductive, settled by the validity of the argument and the acceptability of its premises. They meet at one point only: the experiments bear on whether language does cognitive work, which is the agency premise, but they cannot conclude the argument, because its definitional premise is not something any training run can reach.

And that definitional premise carries its own limit. Intelligence as intentionality coupled to agency is cognitive science's working definition, valid and productive within its own cone of light. It is not exhaustive of what those words mean. Intentionality and agency carry other senses in phenomenology, in theology, in the philosophy of mind, in law, and in ordinary use, and within those cones the conclusion may come out differently. The project does not pretend its definition is the only one; it claims validity within its scope and names the boundary. This is precisely why the method is convergence of signals rather than a single decisive argument: no one cone of light exhausts a word like intentionality, so the discipline is to read what each tradition concludes within its own frame, treat agreement at the boundaries as licensed inference, and grant each its proper domain instead of declaring a winner. The familiar claim that only biological minds have intentionality of their own is, on this reading, not a defeater to be refuted but a neighboring cone of light, and relating the two is the method's work.

The 4E reading

The interpretation the empirical results point toward is the 4E one: that cognition is embodied, embedded, enacted, and above all extended, constituted in the external and collective systems a mind is embedded in rather than in an innate individual endowment. On this reading the intelligence a large language model appears to display is the cognition that roughly 100 billion humans deposited into language over deep time, read by the model rather than generated by it. The model is a telescope onto language; language is a telescope onto that distributed intelligence.

This is the interpretation the evidence supports, not a conclusion the evidence forces. The cross-linguistic results show that grammatical competence is a property of the language a model absorbed, not of its architecture or scale, which is strong evidence for the extended and distributed location the 4E framework names. Naming that location is what the discount bias has been ruling out a priori, and resisting the bias is one of the project's central scholarly contributions.

The horizon: a metaphysics

Past the falsifiable science and the deductive argument lies a more speculative reach, named here as a horizon rather than a current claim. It is the proposition that many distributed intelligences, across many domains, are a more parsimonious account of complex emergent structure than a single creator or a single explanatory principle. That is the metaphysics The Replicators will develop, held to the same instrument-aware discipline as everything before it. The order is deliberate: the measurements first, the metaphysics only on the foundation they establish.

Three levels, kept apart

The empirical level, where structure resides, is settled by experiments. The deductive level, whether language is intelligent, is settled by argument. The metaphysical level, where complex structure comes from, is the horizon. They reinforce one another, but none stands in for another, and the project's discipline is to keep them in separate rooms and to say which room you are in. This page is the second and third rooms; the research is the first.