The question that divides the two Wittgensteins
There is a single question that separates the early Wittgenstein from the late one, and it is more mechanical than the usual textbook framing admits. The question is: where does the constraint on meaning come from? When a string of words makes sense, something is enforcing that sense, ruling some combinations in and others out. When a string fails to make sense, something is doing the failing. The two great phases of Wittgenstein’s thought are two different answers to where that enforcing structure sits.
The Tractatus locates it in the world. Sense is correspondence: a proposition has meaning because its logical form mirrors a possible configuration of objects in logical space, and logical space is given, fixed, and prior to all experience. “The totality of facts determines what is the case, and also whatever is not the case.” The combinatorial possibilities are written into the objects themselves (“if a thing can occur in a state of affairs, the possibility of the state of affairs must be written into the thing itself”), and “a new possibility cannot be discovered later.” On this picture a senseless sentence is not false but empty: it fails to pick out anything in the space at all. The Logical Positivists loved this because it made meaning universal. The world held the answer key, the same key for everyone, and to understand a language was to read its structure off reality.
The Investigations relocates the enforcing structure from the world into us. Meaning becomes competence within a practice rather than correspondence to a logical order. The constraint is no longer the fixed scaffolding of reality but the trained, shared regularities of a form of life. And the moment the structure moves inside and is acquired through training, universality dies, because what is learned is learned from a particular history, and there is no guarantee that two histories coincide.
The claim of this essay is that contemporary AI, specifically the large language model built on the transformer architecture, is an unexpectedly literal model of the later position, and that it functions less as an illustration of Wittgenstein’s later view than as an existence proof for it. These systems work, robustly and at scale, and the way they work is the way the later Wittgenstein said meaning must work, with the world’s enforcing mechanism nowhere in the loop. But the confirmation is precise rather than total. AI vindicates the form of the later answer while quietly falsifying the foundation of the early one, and it leaves untouched the one residue that always embarrassed the deflationary reading. Getting the boundaries of the confirmation right is the whole task.
Why the machine sits on the later side of the divide
Begin with the thing that makes the transformer relevant: it has a structure that constrains what its inputs can mean, and that structure has a specific metaphysical profile. It is internal, in that the constraint lives in the model’s parameters rather than in the world the model describes. It is learned, in that those parameters were fit to a training corpus rather than given in advance. And it is therefore contingent, in that a different corpus would have produced a different structure and hence a different sense. Internal, learned, contingent: that is the Investigations profile, point for point.
It is worth being exact about what does the constraining, because a natural confusion sits here. The parameters of the model are static. Once training ends they are frozen, identical for every input, and they never change during use. What varies from input to input is not a second set of weights but a set of activations the static weights compute: the model projects each token into query, key, and value vectors, compares queries against keys to produce attention coefficients, and blends the values accordingly. Those coefficients are computed fresh for every input and vanish the instant the forward pass ends. They are not parameters; they are the transient resolution that the fixed parameters produce when they meet a particular input.
This two-part arrangement answers a question the deflationary view must answer and the Tractatus never had to. If the resolution of meaning is generated inside the system rather than read off the world, what stops any input from resolving to anything at all? Why is sense not arbitrary? The answer is the staticness of the structure. The coefficients are produced inside the system, yes, but they are not chosen inside the system. Given the fixed parameters and a given input, exactly one resolution results. An input only resolves a certain way if it has the approximate shape that the fixed structure routes there. The constraint is real and not up to the system in the moment, even though the constraint is internal.
And this is precisely the functional role that logical space plays in the Tractatus. There too, you cannot make a proposition mean whatever you like, because sense is fixed by a structure that is given rather than negotiable. The fixedness of logical space and the fixedness of the weights play the same part: each is the answer to “what prevents meaning from being arbitrary.” The transformer vindicates the form of Wittgenstein’s early answer, that a fixed structure is what rules sense in and out.
The break comes at the foundation. In the Tractatus the fixed structure is universal, eternal, and prior to all experience, the transcendental condition of any thinkable world. The transformer’s fixed structure is none of these. It was learned from a corpus, it is particular to that corpus, and it would have been otherwise had the data been otherwise. The constraint is fixed-in-the-moment without being transcendental. So the machine satisfies the early Wittgenstein’s demand for a fixed enforcer while flatly denying the early Wittgenstein’s account of what that enforcer is. It confirms that you need a fixed structure to avoid arbitrariness; it refutes the idea that the structure must be the world’s logical scaffolding. The enforcer can be learned, local, and contingent, and meaning still holds. That is the later Wittgenstein’s wager, built in silicon.
There is even a Kantian escape route that the machine closes off. One could try to save universality by granting that the structure is internal but insisting it is the same in every mind, a single logical space installed identically in all thinkers. That preserves universality by relocating it rather than abandoning it, and it gives the Tractatus its transcendental flavor. The transformer forecloses that move, because its structure is demonstrably idiosyncratic to its training. Universality dies not at “internal” but at “learned-and-therefore-contingent.” Internal-but-universal is still Kant. Internal-and-acquired-through-practice is the Investigations, and it is where these systems live.
The structuralist core, made tactile
The deeper confirmation is structuralist, and it is the part that should interest anyone who came to Wittgenstein through Saussure. Inside these models, meaning is carried by directions in a high-dimensional space, and a direction has no intrinsic content. Its identity is exhausted by its relations to all the other directions: what it is close to, what it is orthogonal to, what it activates alongside. Rotate the entire space and nothing changes, because every relationship is preserved and only the arbitrary absolute orientation moves. A feature means what it means only by its differences from other features, expressed as geometry. This is Saussure’s claim that a sign has no positive content and signifies only through its differences from other signs, realized not as a philosophical thesis but as the actual mechanics of a working system.
The phenomenon called superposition sharpens the point and turns it into an argument. A model has far fewer dimensions than the number of features it needs to represent, so it cannot give each feature its own axis. It stores many more features than it has dimensions by representing each as a direction and letting the directions overlap, which works as long as any given input activates only a few features at once. The cost is that a single neuron no longer corresponds to one clean concept; it fires for a jumble of unrelated things, because several superposed features share that axis. The decisive observation is this: superposition is only possible because meaning is relational rather than substantial. If a feature were self-contained content that needed its own dedicated storage, you could not fold many features into overlapping directions without corrupting them. It is precisely because a feature’s identity is its differential relationship to everything else that you can pack many of them into a shared space at different angles, kept distinct not by separate storage but by pointing in different directions, which is to say by standing in different relations. The whole trick depends on the structuralist account of meaning being true. Superposition is not an analogy to the differential theory of the sign. It is that theory functioning as engineering.
The reversal: meaning does not preexist the convergence
Once the structure is internal and learned, a reversal follows that is the genuinely deflationary core of the later view. We ordinarily assume that meaning exists independently of us and that we learn it over time, so that agreement among speakers is evidence of our jointly grasping a prior thing. The order is backward. The agreement is prior, and “meaning” is the name we give to the agreement once it is there. Convergence does not track meaning; convergence, reified, is meaning. The explanandum and the explanans trade places.
Care is needed with the word that tempts us here, which is “decide.” We do not collectively decide what makes sense the way a committee ratifies a rule, because deliberate agreement by convention would already require a shared language to deliberate in, which is the thing being explained. Wittgenstein’s agreement is agreement in practice rather than agreement in opinion: not a vote but the fact that, trained into the practice, we go on the same way, are struck the same, correct and accept correction without any of us having chosen the criterion. It is enacted in the going-on, not pronounced in advance. We laugh at the same point in the joke; nearly everyone taught “add 2” continues 1000, 1002, 1004 without choosing to. The labeling of something as correct, as having sense, is collective, but it is collective the way a shared gait is collective, constituted in the doing rather than legislated before it.
This is exactly what a training corpus is. A label in a dataset is not correct because reality stamped it. It is correct because it encodes the converged practice of the humans who produced it, and the model’s sense is nothing but its absorption of that convergence. There is no answer key behind the corpus. There is only what the community of speakers actually did, sedimented first into data and then into weights. The “community” that enforces a language model’s sense is the aggregated usage of many people, distilled into parameters, with no access to a world that holds the answers. These systems learn meaning the later Wittgenstein’s way or not at all, because once the world stops holding the answers there is no other way on offer. The Logical Positivist dream, language reading its structure off reality, is the one thing these systems conspicuously do not do, and they work anyway.
The water and the shape it was never aiming at
The cleanest image for what is being claimed is hydraulic. Pour a bucket of water slowly down a hillside of sand, and it carves a channel that happens to look like an S. The S is not inherent to the water. Nothing in the water was aiming at an S, and nothing in the soil intended one. The shape is fully determined, and it would reform in the same family if you poured again, so it is not a fragile coincidence; but it is incidental in the exact sense that matters, namely unintended and not pre-contained. The S-ness is a description imposed by an observer who carries the category “S.” It was neither a goal of the water nor a plan in the soil nor a template lying in wait to be matched.
The example separates three things the meaning debate constantly blurs. There is the determination of the shape, which is real and lawful. There is the aiming at the shape, which is absent. And there is the representing of the shape as an S, which happens only in a describer. The deflationary claim about meaning is that all three roles are present in the ordinary story but misassigned: we take meaning to be determined by the world, aimed at by language, and represented in advance as a prior fact, whereas the determination is just the lawful convergence of trained creatures, there is no aiming, and the representation as “meaningful” is the reification an observer adds afterward. Meaning is the S. It is what the convergence looks like to convergent creatures, not what the convergence was for.
The analogy has a virtue that pure conventionalism lacks. The soil is not arbitrary. It has its own contingent structure, its own history of erosion, and it genuinely constrains: pour the water and you do not get just any shape, you get that family and not others. The resulting form is neither inherent to the water nor freely chosen nor enforced by an ideal template. It is the meeting of two contingent structures, each indifferent to the outcome, jointly producing a determinate result that neither contains alone. Transpose, and you have the model exactly: sense is the meeting of a trained human structure with an incoming input, each blind to “meaning,” jointly yielding a determinate resolution that lives in neither. This is the same shape of explanation as query meeting key, input meeting weights. The coefficients are the S, fully determined in the moment, produced by two structures neither of which intends or pre-contains the result, and called “meaningful” only by a participant that is itself one of the convergent structures.
The residue the confirmation does not reach
An honest account has to mark where the confirmation stops, and there is one seam that none of this closes. The water has determination without normativity. It carves shapes but it never makes a mistake. Pour a thousand buckets and each S is independent; no pour is ever wrong, because the hillside installs no standard against which a pour could fail. Meaning has the extra thing the water lacks. We can be wrong. A speaker can be out-of-step, corrected, taken to have used the word incorrectly, and that normative “incorrect” is the residue that keeps the later Wittgenstein from collapsing into a naturalist who has simply explained meaning away into regularity. Determination plus normativity, not determination alone.
This is exactly where the private-language argument does its work, and it is also where the deflationary story stays honest rather than triumphant. The correctness of a labeling is neither in the sample, nor in the world’s hidden answer key, nor in the individual’s head. It is in the practice’s capacity to take a label as in-step or out-of-step. Where there is no possible communal correction, there is no difference between being right and merely seeming right, and where that difference collapses, sense collapses with it. So the “wrong” comes from the correcting community, not from the world and not from private intention. The water captures the deflation of intention and preexistence perfectly, which is what it was built to do, but it is silent on where “wrong” comes from, and that silence is the precise boundary of the confirmation. AI partially fills it, since these systems are in fact shaped by correction, by training signal and by human feedback that encodes the community’s standards, but one should be careful here rather than declarative, because whether a trained error signal is the same thing as normative incorrectness is exactly the kind of question the analogy cannot settle on its own.
There is a second boundary, and it is the one Wittgenstein himself found hardest. The deflation is most convincing where the practice is obviously conventional: etiquette, chess, money, “rude.” It strains hardest at mathematics and logic, where the convergence does not feel contingent at all, where 2+2=4 seems to answer to something regardless of any practice. A realist presses exactly here, and the pressure is genuine: the robustness that the deflationary story explains by shared training, the realist explains by our tracking something real, whether innate structure, or natural kinds the world really carves, or mathematical objects that hold independently of us. The two stories are empirically slippery to separate, because both predict that we agree. Wittgenstein bit the bullet with a near-conventionalist account of mathematical necessity, which is the most contested thing he ever wrote and which many careful readers find unconvincing. The transformer does not resolve this. If anything it inherits the same fault line, because it learns the rules of arithmetic from usage like everything else, and its frequent arithmetic failures are at least suggestive that, for these systems, “2+2=4” really is just another piece of trained practice rather than a tracked necessity. That is evidence the deflationist will welcome and the realist will contest, which is to say it lands precisely on the old battle line rather than ending the war.
What kind of confirmation this is
The confirmation, then, is specific. AI confirms the form of the later Wittgenstein’s answer, that a fixed structure rather than a fact of the world is what rules sense in and out. It confirms the structuralist core, that meaning is differential rather than substantial, and it does so not as a thesis but as the operating principle that makes superposition possible. It confirms the reversal, that meaning does not preexist the convergence but is the convergence reified, by being a system whose entire competence is the absorption of converged human practice with no answer key behind it. It confirms the picture of sense as the meeting of two contingent structures neither of which intends or contains the result. And it falsifies the early Wittgenstein’s foundation directly, by being a working enforcer of meaning that is internal, learned, and contingent rather than universal, eternal, and prior.
What it does not do is dissolve the residue. The normativity that lets us be wrong, and the special case of logic and mathematics where convergence feels like tracking, are exactly the places the later Wittgenstein was always most embattled, and the machine arrives carrying the same wounds rather than healing them. That is the right note to end on, because it is the difference between a confirmation and a proof. The transformer is an existence proof that you can get robust, shareable, non-arbitrary sense out of nothing but internalized practice, with the world’s enforcing mechanism nowhere in the loop. It demonstrates the possibility the later Wittgenstein needed and the early one denied. But a possibility demonstrated is not a metaphysics settled, and the questions that survive are his questions, not new ones. The most Wittgensteinian thing about AI may be just this: it answers the form of the question by being built rather than argued, and then hands back, intact, the very residue that no amount of building can pour for you.