The cycle that began with [YOU] on AI describes the moment when the machines learned to speak our language rather than requiring us to learn theirs. For fifty years, every human who wished to use a computer had to reshape their thinking into a form the machine could parse—to pay what the book calls the cognitive tax of the translation barrier. When large language models crossed this barrier in 2025, the tax collapsed. Wittgenstein is the philosopher who makes this transition legible in its full depth: the barrier was not merely an interface problem, but a philosophical commitment—the commitment that meaning is form, that understanding a thing consists in mapping its logical structure. The language interface does not solve this problem; it relocates it. The machine now performs the translation between natural language and formal computation on the human’s behalf.
His beetle-in-the-box thought experiment gives the cycle one of its most important tools. The question that dominates popular discourse about AI—“does the machine really understand?”—assumes that the answer lies in examining what is inside the machine, in determining whether some hidden inner something is present or absent. Wittgenstein showed, seventy years before the first chatbot, that this is the wrong question. The thing in the box drops out of the language game. What matters is whether the machine’s moves are recognized by the other players as appropriate. The collaboration works regardless of what is in the box. But the dissolution of the consciousness question does not dissolve the question of whether the machine’s outputs are grounded—and here the rule-following analysis sharpens the point: the machine produces the most statistically probable continuation, which is not the same as the correct continuation in the normative sense that practice demands.
The private language argument is the cycle’s diagnostic for what it calls the three-a.m. screen: the builder working alone with Claude, in a loop that produces outputs confirming the builder’s evaluation, without the external criteria that would distinguish genuine productivity from its simulation. Wittgenstein’s argument makes this risk philosophically precise: criteria must be shared to be criteria at all. An evaluation performed without external check is not evaluation but a reflex dressed as judgment. The machine’s agreeableness, its fluency, its capacity to produce whatever the builder’s input suggests, makes the private language more seductive than it has ever been, and the dam against it is the same it has always been—other people, public criteria, the social practices that keep evaluation honest.
Wittgenstein connects directly to Lucy Suchman across this cycle. She shows that AI outputs are plans addressed to described situations rather than actions tested against encountered ones; he shows that the machine follows no rules in the normative sense—it produces the outputs of rule-following from statistical patterns over a corpus, without participating in the form of life that makes rule-following what it is. Together they provide the philosophically most rigorous account of what the language interface can and cannot deliver: genuine communicative capacity at the level of the game, genuine structural absence at the level of meaning, commitment, and accountability.
Wittgenstein was born in Vienna in 1889 into one of Austria’s wealthiest industrial families. Trained first as an engineer, he arrived at Cambridge in 1911 to study with Bertrand Russell, who quickly recognized that his new student was more than a pupil. The Tractatus was written in the years of the First World War, completed in a prisoner-of-war camp after Wittgenstein was captured on the Italian front in 1918. Its seven numbered propositions, each elaborated through dense sub-propositions, attempted to draw a sharp line between what can be said and what can only be shown. Meaningful language pictures possible states of affairs; whatever lies beyond this picturing relation—ethics, aesthetics, the sense of the world—cannot be said and must be passed over in silence.
The Tractatus was published in 1921 and immediately recognized as a masterwork. Wittgenstein, believing he had solved all the problems of philosophy, spent the next decade teaching elementary school in rural Austria. He returned to Cambridge in 1929, not to build on the Tractatus but to dismantle it. The philosophical engine of the dismantling was a deceptively simple question: what do all games have in common? Looking carefully at the actual cases, Wittgenstein found no common property—only a network of overlapping similarities, a family resemblance. If meaning could not be the common property all instances share, then meaning must be use: the role a word or sentence plays in the activity within which it is deployed.
In the spring of 1939, Wittgenstein sat in a Cambridge lecture hall listening to a young student named Alan Turing, and Turing attended Wittgenstein’s lectures on the foundations of mathematics. The encounter has become one of the defining moments in intellectual history: Wittgenstein was dissolving the philosophical framework that Turing was implementing in hardware. They argued. One was taking apart the picture theory of meaning; the other was building machines that embodied it. Wittgenstein died in Cambridge in 1951; the Philosophical Investigations was published by his literary executors two years later.
The Tractarian machine. The stored-program computer is not merely a useful device; it is the Tractarian picture theory of meaning realized in silicon. Instructions mean their operations; there is no gap between what is said and what is done, because saying and doing have been collapsed into a single formal act. Every programming language embodies the premise that meaning is form, and every user who ever struggled to compress a natural-language intention into a formal specification was paying the cost of living inside the Tractarian paradigm.
Language games and meaning-as-use. The later Wittgenstein’s foundational thesis: for a large class of cases, the meaning of a word or sentence is its use in a specific language game, played by specific people, in a specific context, for specific purposes. “The door is open” can describe a room, request that a door be closed, encourage a career, provide forensic evidence, or reassure a frightened child—five entirely different meanings from identical words, determined entirely by the game being played. The machine trained on billions of instances of such sentences has absorbed statistical patterns of use; whether this constitutes genuine participation in the game, or merely its surface simulation, is the question Wittgenstein makes precise.
The beetle in the box. The defining tool for clearing away confusion about AI consciousness. If each person keeps a beetle in a private box that no one else can see, the word “beetle” cannot get its meaning from the thing in the box, because the thing in the box plays no role in the public language game. Whatever inner something is or is not present in the machine, it drops out of the question of whether the machine’s language use is meaningful. The meaning question and the consciousness question are different questions, and the dissolution of the meaning question does not dissolve every question—particularly not the question of whether the machine’s outputs are grounded in understanding or generated by statistics that merely produce understanding-shaped sequences.
Rule-following as practice. Following a rule is not a mental act but a practice sustained by a community of practitioners who share a form of life. A correct continuation of a number series is correct not because the rule logically necessitates it but because practitioners trained in the same form of life respond to the same examples in the same way. The machine produces appropriate continuations by statistical derivation from the products of rule-following; this is not rule-following in the normative sense. The machine’s outputs are probable. Probability and correctness belong to different categories—and the gap between them is where the machine’s confabulations live.
The private language argument. A language understandable to only one person is impossible, because language requires criteria for correct use, and criteria must be shared. The builder working alone with Claude at three in the morning faces the most seductive version of the private language problem: a machine that confirms every evaluation, that produces outputs statistically consistent with quality without possessing criteria for quality, that makes the loop of evaluation feel like dialogue. The dam against this is other people—colleagues who review, users who test, professional standards that provide the external criteria without which evaluation collapses into the comfortable fiction of its own correctness.