Language Games: Wittgenstein, the Ladder, and the Machines That Only Know Use
🇩🇪 Auf Deutsch lesen~10 min read
philosophylanguageaiwittgensteinlanguage-gamesmeaningrule-followingllmontology
Hannah Arendt's most famous phrase is also her most misunderstood. The banality of evil is not "evil is ordinary" — it is "evil can be done without thinking." On Eichmann, the rule of Nobody, and what happens when a system decides and no one is home.
Part of the Philosophy section — a companion to The Grammar of Pretending and Know Thyself.
Framing: A Short History of Philosophy
Ludwig Wittgenstein tried to end philosophy twice, and failed both times in ways that are more useful than most successes. The first attempt was a short, glacial book that claimed to draw a line around what can be said. The second was a heap of numbered remarks that took the line back and showed the whole exercise was a misunderstanding of how language works. Between the two, he left builders and machine-learners an unexpected gift: a theory of meaning that is almost embarrassingly practical. Meaning is not a thing words carry. It is what people do with them.
That single move — meaning as use — turns out to explain a lot about the artefacts we build today: specifications that nobody reads the same way, taxonomies that refuse to be clean, best practices that outlive their reasons, and language models that are fluent about everything and grounded in nothing. Wittgenstein got there first, with a beetle in a box and a signpost in a field.
The Tractatus Logico-Philosophicus (1921) is a machine built to answer one question: how can a sentence be about the world? His answer was the picture theory. A proposition is meaningful because it is a logical picture of a possible state of affairs; it shares a form with the world it depicts, the way a diagram shares a structure with the thing it draws. Language mirrors reality, and the mirror has an edge.
Everything that is not a picture of a fact — ethics, aesthetics, the meaning of life, metaphysics — falls outside the mirror. Not false, but unsayable. Hence the book's famous seventh and final proposition:
Whereof one cannot speak, thereof one must be silent.
The structure is a ladder. You climb it by using its propositions to see that they are, in the end, senseless — they are the scaffolding, not the building — and then you throw the ladder away. Wittgenstein meant this literally and it is the most honest thing in the book: the theory of meaning is itself beyond the limit it draws. It is a tool you use once and discard.
Builders know this feeling. Every specification is a ladder: a temporary scaffold precise enough to get a team from an intention to a working thing, then quietly obsolete the moment the thing exists. The trouble starts when we forget to throw the ladder away and start worshipping the scaffolding.
The later Wittgenstein — the Philosophical Investigations, assembled after his death in 1953 — abandons the picture. There is no single essence that all meaningful sentences share, no one form they all mirror. Instead there are language games: the countless, overlapping activities in which words do work. Giving orders, describing, measuring, joking, praying, debugging, writing a ticket. Each game has its own rules, its own moves, its own criteria for a correct play, and none of them is the master game.
The meaning of a word is its use in the language.
His favourite example is almost a parody of a software team. Imagine a language with just four words — block, pillar, slab, beam — used by a builder who calls them out and an assistant who fetches them. The words do not name private mental pictures; they are moves in a game of fetching. Pointing at a slab and saying "slab" does not attach the word to an object in the mind; it trains a use. Ostensive definition — "this is a slab" — only works if the learner already knows how to take the pointing, which is to say, already knows the game.
This is the quiet bombshell. Meaning is not in the head and not in the object. It lives in the shared practice — in what a community of speakers reliably does with a word, including the corrections it makes when someone plays wrong.
If words were defined by essences, every category would have a checklist. Wittgenstein pointed out that most of our real categories do not. What do all games have in common — board games, card games, ball games, solitaire, a child throwing a ball at a wall? Not one feature shared by all, but a web of overlapping similarities, like fibres in a rope. He called it family resemblance: you recognize the family by the criss-crossing traits, not by a single defining mark.
Anyone who has built an ontology, a tag system, or a knowledge graph has met this the hard way. You start with clean types — article, project, concept — and within a week the edges are doing the work the types were supposed to do. A "concept" is a paper is a tool is a person, and the only honest structure is a network of resemblance, not a tree of essences. The lesson is not "give up on structure." It is: model the relations, not the essences. The graph is closer to how meaning actually hangs together than the taxonomy is.
Wittgenstein's hardest and most famous argument is about rules. Take a signpost pointing right. It seems to tell you exactly what to do. But a signpost is a piece of wood; it does not contain its application. Every interpretation of the signpost — follow the arrow, follow the post, walk perpendicular — is itself something that needs a rule, and that rule needs an interpretation, and so on. There is no point at which the rule, by itself, reaches out and grabs the next step.
His answer is not scepticism but practice. What stops the regress is not a deeper rule but a shared habit: this is simply what we do, this is how we were trained, this is the custom. "This is simply what I do" is where justification ends — not because we are irrational, but because reasons bottom out in a form of life.
For builders this is the most immediately useful idea in the whole book. A written rule — a coding standard, a process, a "best practice" — never contains its own application. The rule lives only as long as the practice that keeps re-enacting it, and it drifts the moment the practice changes underneath it. This is exactly why today's best practices become tomorrow's anti-patterns: the signpost stayed, the road moved. A spec is not a source of truth; it is a move in a game that a team has to keep playing.
The most quoted image from the Investigations is the beetle in a box. Suppose everyone has a box with something in it, and everyone calls the thing inside "beetle." No one can look into anyone else's box. Then, Wittgenstein says, the word "beetle" cannot function as the name of a private object at all — because if the private referent plays no role in the public use of the word, it drops out of consideration entirely. Meaning has to be checkable in the shared game, or it is not meaning yet.
This is the argument against a private language, and it is also the cleanest diagnosis of the interpretability problem in machine learning. A model has a vast inner state — weights, activations, latent directions — and we keep wanting to call something in there "the concept of a cat" or "the intention to refuse." But a private referent that plays no role in any public, checkable use is exactly the beetle: it might be there, it might be nothing, and the word we invented for it does no work. The only meaning we can actually verify is the one that shows up in behaviour, in the moves the system reliably makes.
Here is where the two Wittgensteins meet the present. A large language model is, in the later Wittgenstein's terms, a machine trained almost purely on use. It has read an enormous sample of the moves people make in language games — the greetings, the explanations, the apologies, the code comments, the arguments — and it has learned to make moves that fit. That is precisely why it is so fluent, and precisely why "meaning is use" predicts its limits so well.
The model learns the game from the outside. It does not need the private beetle to play; it needs the public regularities, and it has more of them than any human ever will. This explains the strange combination we all recognize: uncanny competence at the form of an answer, and no way to check whether anything stands behind it. The fluency is real, because use is real. The grounding is not guaranteed, because use does not by itself supply reference.
There is a further, subtler point. When we say a model "understands" a word, we are smuggling in the picture theory — the idea that somewhere inside there is a representation that mirrors the thing. The later Wittgenstein would ask us to drop that question and ask a better one: which game is it playing, and does it make the right moves in that game? For a code assistant, the game is "produce a patch that compiles and passes the tests." For a support bot, it is "resolve the ticket without making the customer angrier." Those are checkable. "Does it really understand?" is the beetle.
The two halves of Wittgenstein's work are not a contradiction so much as a demonstration. The early theory tried to pin language to the world with a single, sharp picture, and it was beautiful and wrong in the way that precise things are often wrong. The later work gave up the single picture and described the messy, plural, practice-bound way meaning actually lives — and it is less beautiful and more true.
The builder's version is a discipline, not a doctrine:
Wittgenstein said philosophy is not a body of doctrine but an activity — a kind of therapy for the confusions language breeds when it idles. That is a good description of what good engineering does too. We are not building eternal definitions. We are keeping a set of language games playable, so that people and machines can make moves that fit, and noticing quickly when a move that used to fit no longer does.
The machine checked every step. Which game it is playing is still yours to name.