Mind, Matter, and Meaning

Sit in enough philosophy seminars and you will eventually hear someone announce, with a straight face, that the radiator beside them is running Microsoft Word. The claim sounds unhinged. It is also, in a narrow and revealing sense, true. A famous result holds that any sufficiently large physical object — a wall, a bucket of water, a rock warming in the sun — implements every computer program you care to name. Hilary Putnam pressed a version of this in 1988, and Searle pressed a related one in 1990.¹ And the trick that makes it work turns out to embarrass the trickster. Grant yourself enough latitude in how you match up physical states with computational ones, and yes, the wall computes WordStar. By the same latitude it also computes the opposite of whatever it computes, and dreams of Provence on alternating Tuesdays.

The interesting question is what to make of this. Functionalists, on first hearing, tend to flinch. Their whole picture rests on the idea that a mind consists in functional organization — patterns of input, internal state-change, and output — that can run on any material able to sustain it. If walls turn out to have whatever organization you like, the picture trivializes. The mind reduces to a permission slip you write for yourself. That cannot be what understanding amounts to.

The temptation here is to reach for new gears. Maybe the right functional organization rules out the gerrymandered mappings. Maybe causal counterfactuals do it, or complexity, or some careful algebra over state-spaces. A respectable literature has grown up around saving functionalism from triviality by tightening these technical screws, and Peter Godfrey-Smith has written the patient article that sorts the repairs that hold from the ones that overreach.² But the more revealing move is to step back and ask why the threat keeps recurring no matter how cleverly the screws get tightened.

The diagnosis, once you let it through, has a faintly comic shape. The triviality arguments do not expose some fixable bug in functionalism. They expose something larger. Anything you characterize purely in terms of formal structure — purely in terms of what plays which role in a system of inputs, swaps, and outputs — admits arbitrary interpretation, because formal structure carries no built-in tether to anything in particular. Symbols do not, on their own, point at the world. Functional roles do not, on their own, mean anything. A string of zeroes and ones sits there with perfect indifference about whether it represents a chess position, a tax return, or nothing at all. The wall does run WordStar — in exactly the trivial sense in which it does anything else you can describe with enough slack. Under a creative enough reading the world becomes a Rorschach test, and Rorschach tests do not understand anything.

This conclusion should feel familiar from a different street. Searle reached it from outside functionalism and pressed it in the Chinese Room.³ Putnam reached it from inside and pressed it in the triviality argument. They walked toward the same wall, and the wall is not running WordStar. It is the wall that says: formal structure cannot, by itself, fix meaning. Searle held that conclusion against the computationalist program loudly. Putnam held it differently. Having spent the 1960s as the founding parent of computational functionalism, he followed his own argument to the conclusion that the very feature he had prized — that minds could be realized in any substrate — was the feature that trivialized the theory, and he said so in print. That is the kind of intellectual courage that ought to embarrass anyone who has ever kept a position because of who else held it.⁴

So the wall does not run WordStar in any sense worth wanting. What does the program-running, then? Not the formal structure alone. The structure has to come tethered. It has to be the structure of something that already, for some non-magical reason, picks out one interpretation over the infinite others.

Here the embodiment story stops being a vague gesture about robots and starts doing real work. Consider a nervous system that evolved to track predators. It tracks predators rather than sandwich crumbs because evolution selected its ancestors for the former and not the latter. No interpreter hands it that mapping. It inherited the mapping from a long causal history in which getting the mapping right meant getting eaten less often. That is roughly what Millikan calls a proper function, and what Dretske connected to representation more broadly.⁵ Selection — biological, or learned, or both — does the work that pure formal structure cannot. It pins down one interpretation by making it the interpretation under which the system worked, which means, prosaically, the interpretation under which its bearers had children.

Notice how this dissolves the triviality argument without saving functionalism on its old terms. It does not say the wall fails to run WordStar because some better-tuned functionalism rules the bad mapping out. It says no purely formal account can rule the bad mapping out, because formal accounts lack the resources. What rules it out is that the wall has no selection history picking one mapping over another. The brain has one. The wall does not. The radiator dreaming of Provence cannot dream of Provence at all, because no historical fact about the radiator makes Provence — rather than Pittsburgh, or pickle juice — the content of the dream. Content depends on grounding, grounding depends on history, and history does not show up in the formal organization. It shows up in bodies in worlds.

The detour through Putnam matters because, without it, the conclusion sounds like Searle banging the lectern. From the outside, the claim that syntax cannot suffice for semantics looks like an anti-computationalist hobbyhorse. From the inside — when the founding theorist of computational functionalism follows his own argument honestly to the same wall — it stops looking like a hobbyhorse and starts looking like a result. The convergence carries weight that neither voice carries alone. And it lands the conclusion the right way: not as one camp dismissing a rival, but as something the rival camp’s best mind ended up affirming when he refused to flinch from his own work.

The argument also carries a moral about current AI hype. When a language model produces fluent paragraphs about Provence, it does not thereby know about Provence. It performs a transformation of formal structure — a transformation, to be clear, of breathtaking sophistication and real practical value — and that transformation admits, in principle, arbitrary interpretation. Whatever tether it has to the world it speaks of comes from the human authors in its training corpus, who were embodied creatures with selection histories, and who therefore meant something when they wrote about Provence. A model riding on their meanings does not thereby have its own. The wall that allegedly runs WordStar stands to WordStar exactly as a stochastic parrot stands to its training corpus: in a relation conferred entirely from outside, by interpreters who already mean things.⁶

What the wall lacks, the parrot also lacks, and what the parrot lacks, embodiment supplies. Not magically. There is no extra ingredient here, no quintessence, no special carbon, nothing Descartes would recognize. What embodiment supplies is the unromantic fact that bodies — evolved or otherwise selected — have histories of getting things right and wrong, and those histories pin down what their states are about. Subtract the history and the formal structure floats free, available for any interpretation and committed to none. That, finally, names what was always wrong with the picture of the mind as software running on the brain. Software does not run on anything by itself. It runs on something that, for non-software reasons, already meant.

The wall does not run WordStar. The brain, blessedly, runs something — and the reason has nothing to do with the cleverness of the mapping and everything to do with a long, biological, world-involving fact: some patterns got selected for being about things, and the things they got selected for being about are still there, waiting outside the window where they always were.

---

Notes

1. The “every system implements every program” result is the technical engine the parlor game runs on. The cleanest statement of the worry — that any ordinary open physical system implements every finite-state automaton, given a permissive enough mapping from physical to computational states — is laid out by David Chalmers, “Does a Rock Implement Every Finite-State Automaton?,” Synthese 108 (1996), who states the problem sharply precisely in order to constrain it. Putnam’s own version appears as an appendix to Representation and Reality (MIT Press, 1988). Searle’s adjacent claim is that computation itself is observer-relative: syntax is not intrinsic to physics but assigned by an interpreter, so nothing is “intrinsically” a digital computer — see “Is the Brain a Digital Computer?,” Proceedings and Addresses of the American Philosophical Association 64 (1990). The two routes differ but arrive together: where Searle argues that the computational description is observer-relative from the start, the triviality results show that any purely formal description leaves the mapping wide open. ↩
2. Peter Godfrey-Smith, “Triviality Arguments Against Functionalism,” Philosophical Studies 145 (2009). Godfrey-Smith supplies the careful inventory — which versions of the triviality argument succeed, which overreach, and what minimal additional structure functionalism must take on to survive. His verdict, that the strongest versions push functionalism toward grounding in causal and selection-historical facts, converges with the position defended here. ↩
3. John Searle, “Minds, Brains, and Programs,” Behavioral and Brain Sciences 3 (1980). The Chinese Room is the better-known vehicle, but the deeper Searlean point is the one developed in the 1990 paper cited above: because the notion of computation is observer-relative, it cannot intrinsically characterize the brain. The triviality results reached from inside functionalism are, in effect, that same conclusion reached by a different route — which is why the convergence is worth pressing rather than merely noting. ↩
4. Hilary Putnam, Representation and Reality (MIT Press, 1988). Putnam, who in the 1960s effectively founded computational functionalism, concludes here that the multiple-realizability which originally recommended the view in fact trivializes it once one sees how unconstrained the realization relation really is — the formal organization that was supposed to be substrate-neutral turns out to be interpretation-neutral as well, and that is fatal. See also Oron Shagrir, “Putnam and Computational Functionalism,” in Hilary Putnam on Logic and Mathematics (Springer, 2018), on the philosophical significance of an arch-functionalist reaching this verdict. ↩
5. For the locus classicus connecting selection history to representational content, see Ruth Garrett Millikan, Language, Thought, and Other Biological Categories (MIT Press, 1984), especially chapters 1–2 on proper functions; and Fred Dretske, Explaining Behavior: Reasons in a World of Causes (MIT Press, 1988), for the parallel project run through learning history rather than evolutionary history. Both share the structural claim the argument leans on: what a state represents is fixed by the history that selected it, not by its current formal profile. ↩
6. Whether artificial systems trained on human corpora might inherit learning-historical proper functions in Millikan’s sense is a live question, taken up by Jumbly Grindrod, “Large Language Models and Linguistic Intentionality,” Synthese 204:71 (2024). Millikan’s own writings suggest she would resist the extension: the histories she has in mind run through generations and natural selection, not through gradient descent over a fixed dataset. The disagreement marks the live frontier of teleosemantic theorizing about AI, and nothing here forecloses it — the claim is only that formal structure alone never settles content, not that gradient descent could never be a grounding history of the relevant kind. ↩