Worlds Before Minds: Fuller, Lanier, and the Grounding of AI
🇩🇪 Auf Deutsch lesen~13 min read
philosophyaivirtual realitydesignPart of the Philosophy section — a companion to Language Games and Symmetry and the Machine.
Framing: A Short History of Philosophy
Buckminster Fuller spent the 1960s proposing a globe you could stand inside. The Geoscope was to be a sphere up to two hundred feet across, wired to computers, onto which the live data of the planet would be projected — weather, harvests, populations, shipping — so that a room full of people could see the whole Earth at once and argue about what to do with it. Jaron Lanier spent the 1980s building a world you could stand inside with someone else. At VPL Research he and his colleagues wired gloves and headsets into shared virtual spaces, and he fixed the name for the result: virtual reality. Two machines, two generations, one gesture. Put a human inside a model of the world, and make the model shared.
Intelligence is not a thing in a box. It is a practice in a world — and the world comes first.
The philosophy of artificial intelligence has mostly been a philosophy of minds. Can the machine think? Does it understand? Is there anyone home? These are good questions with a bad habit: they pull attention toward the ghost and away from the room. Fuller and Lanier, two generations apart but drinking from the same countercultural well, both point the other way. The interesting question is not whether the model has a mind. It is what world the model is in, who built that world, and whether anyone can check it.
In 1990 the cognitive scientist Stevan Harnad named a problem that had been lurking since the first symbolic AI programs: the symbol grounding problem. A formal system can manipulate symbols according to rules, but nothing in the rules says what the symbols mean. Give a machine the word "apple" and a dictionary definition, and you have given it more words. The meaning — the crunch, the colour, the fact that it falls — is not in the dictionary. Harnad's image was a person stranded with a Chinese–Chinese dictionary: every entry points to other entries, and none of them ever touches the world.
John Searle had made the same point theatrically in 1980 with the Chinese Room, and Hubert Dreyfus had made it philosophically in 1972: intelligence is not the manipulation of representations but a kind of skilled coping, a body finding its way through a situation. The enactivists — Varela, Thompson, and Rosch in The Embodied Mind (1991) — pushed it further: cognition is not something that happens inside a skull, it is something that happens in the loop between an organism and its environment. Andy Clark and David Chalmers added, in 1998, that the loop can include tools. The mind leaks out into the world it acts in.
Large language models are the most spectacular Chinese dictionary ever assembled. They are trained on the residue of human world-contact — text — and they are very good at the surface regularities of that residue. Whether that is enough for meaning is not settled. David Chalmers, who is not a mystic about machines, has argued that sensory grounding may enrich thought without being strictly necessary for it, and that a pure thinker is at least conceivable. That is the honest state of the argument. Which is exactly why the practical question matters more than the metaphysical verdict: if grounding is not automatic, it has to be designed. And design is where Fuller and Lanier come in.
Fuller's working unit was never the object. It was the system. He is remembered for domes and cars, but the domes were arguments: the geodesic structure gets more strength from fewer materials because the strength is in the whole configuration, not in any member. He called that synergy — "the behaviour of whole systems unpredicted by the behaviour of their parts taken separately." Tensegrity, his floating-compression sculptures, made the same point in wire: continuous tension, discontinuous compression, a stable shape that no single strut could hold up alone.
From that geometry he derived an ethic. Ephemeralization was his word for doing more and more with less and less until, in his most quoted formulation, you can eventually do everything with nothing. It was not a slogan about efficiency; it was a claim about design intelligence. A civilisation that learns to do more with less can afford to take care of everyone. A civilisation that does not, cannot — no matter how much it produces.
Two phrases carry the rest of his programme. Spaceship Earth — the planet as a vehicle with finite supplies and no instruction book — made the whole biosphere the unit of accounting. And comprehensive anticipatory design science was his name for the discipline that follows: do not fix symptoms, design the whole system so the symptoms cannot arise. His contempt for what he called "local focus hocus pocus" was contempt for optimising a part while the whole goes down.
The Geoscope and the World Game were that discipline made visible. The Geoscope was a shared, immersive model of the planet, a place to see patterns that are invisible from the ground. The World Game was a simulation in which players tried to make the world work for everyone — Fuller's stated purpose was "to make the world work, for 100% of humanity, in the shortest possible time, through spontaneous cooperation, without ecological offense or the disadvantage of anyone." Read that sentence again and you will notice it is a design brief for a multiplayer world, written decades before the technology to host one existed.
The line most often hung around his neck — "You never change things by fighting the existing reality. To change something, build a new model that makes the existing model obsolete" — circulates in a dozen phrasings and the attribution is informal, but the sentiment is unmistakably his. Do not argue with the old world. Build a model compelling enough that the old world becomes optional.
Lanier arrived at the same intuition from the opposite direction: not from the whole planet down, but from the human body out. Virtual reality, in his telling, was never a screen you stare at. It was a shared place you and other people inhabit, built out of your movements, your voice, your gestures. The point was not to escape reality but to make it more negotiable — to try on a different body, a different scale, a different physics, and come back with something you could not have thought otherwise.
His phrase for the deepest version of this was post-symbolic communication. If you can improvise a world together — if you can point at a made thing rather than describe it in words — you no longer need to compress experience into symbols and hope the other person decodes them. You can share the referent directly. It is the Chinese-dictionary problem dissolved by construction: instead of more definitions, you build the apple.
The research bore him out in a way that still feels slightly absurd. In the mid-2000s he described homuncular flexibility: people adapt surprisingly quickly to avatars whose bodies are nothing like their own. In a 2015 study with Jeremy Bailenson and colleagues, participants learned to control a virtual lobster, and a third arm, and their sense of embodiment followed. The body image is not a fixed container. It is a skill, and skills can be extended. Mel Slater's work on presence tells a similar story from the other side: the feeling of "being there" is not a property of the graphics, it is a property of how convincingly the world answers your actions.
And then the sentence that keeps Lanier honest: the greatest gift of virtual reality is the moment you take the headset off. The point of the illusion is the return. You go into the model to see differently, and then you come back to the world with better eyes. A VR that never ends is not a medium. It is a hiding place.
Lanier's critique of AI follows from the same humanism. In "There Is No A.I." (2023) he argues that the term is misleading, maybe dangerous, because it invites us to see a creature where there is a collaboration — a vast, human-built system of data, labour, and feedback. "The most pragmatic position," he writes, "is to think of A.I. as a tool, not a creature." He is not against the technology. He is against the mythology that lets us stop asking who built it, who is inside it, and who gets paid. His answer to the mystification is data dignity: the people whose lives are the training data should be visible in the system, credited, and compensated. The world that grounds the model is a world of people, and the model should not be allowed to forget them.
Put Fuller and Lanier side by side and a single thesis comes into focus. Both are world-first thinkers. Fuller starts from the whole system and asks how to design it so that everyone can live. Lanier starts from the shared world and asks how to build it so that everyone can be present. Neither begins with the question "is there a mind in there?" Both begin with the question "what world are we in, and who made it?"
That is the correction AI philosophy needs. The mind-first framing treats intelligence as a substance that a system either has or lacks, and then argues endlessly about the threshold. The world-first framing treats intelligence as a relation — to a body, a practice, a community, a set of consequences. On that view, the important facts about a model are not its benchmark scores but the loop it sits in: what it is trained on, what it is allowed to do, who checks the result, what happens when it is wrong. Grounding is not a property you can inspect in the weights. It is a property of the whole arrangement.
This is also where the Wittgenstein essay becomes a companion piece. Wittgenstein said meaning is use — that words get their sense from the language game they belong to, not from a private act of pointing. Fuller supplies the scale of the game (the whole Earth, the whole system). Lanier supplies its phenomenology (the shared world, the body that learns to inhabit it). AI supplies the test case: a system that has absorbed the surface of every game at once and must now be placed back into one. The question is not whether it "really" understands. The question is which game we are going to let it play, and whether the game has consequences it can feel.
The philosophy is only worth the practice it produces. Here is what the two visions imply for anyone building with these systems.
Start with the whole system, not the model. Fuller's first move. A model is a component; the system is the model plus the data, the interface, the institution, and the person on the other end. Optimise the component and you will be surprised by the system. Evaluate at the level where the consequences actually land.
Ground symbols in shared, checkable worlds. Harnad's problem is an engineering brief. Digital twins, simulations, robotics, instrumented environments — anything that gives the model a loop with consequences — does more for grounding than another trillion tokens of text. VR is not a metaphor here; it is a laboratory for exactly this question, which is why our own work on VR exposure therapy is a better guide to grounded systems than most AI papers.
Keep the human as author, not user. Lanier's data dignity and Fuller's design scientist point the same way. The person affected by the system should have standing in its design, visibility in its data, and a way to contest its output. A system that treats people as raw material will eventually be grounded in a world those people do not recognise.
Ephemeralize. Do more with less. Smaller models, local inference, efficient interfaces, designs that remove a dependency rather than add one. Fuller's test is not "how powerful is it?" but "how much can it do with how little?" The dome is the argument: elegance is not decoration, it is load-bearing.
Take the headset off. Measure against the world, not against the model's own account of itself. The point of a simulation is the return. Any system that cannot be checked against something outside itself — a measurement, a customer, a physical consequence — is not grounded, it is merely coherent. Coherence is cheap. Contact is not.
Both visions have a shadow, and honesty requires naming it. Fuller's whole-earth management can slide into a god's-eye view — the planner who sees the planet whole and forgets that the people inside it have their own perspectives. The World Game was a simulation, and simulations flatter the simulator. Lanier's shared worlds can slide into the siren servers he himself warned about: platforms that host the world, monetise the presence, and quietly become the only place the world is allowed to happen. The Geoscope can be a commons or a control room. The headset can be a window or a cage.
The corrective is the same in both cases, and it is structural rather than moral. Keep the worlds plural. Keep the humans visible. Keep the feedback honest, including the feedback that says the model is wrong. And keep the exit open — the right to take the headset off is not a feature, it is the condition that makes the medium trustworthy.
The Geoscope and the headset are the same gesture, separated by twenty years and a change of scale. Step inside a model of the world to see it whole. Then step back out to act in it. AI is the newest such model — the most comprehensive, the least inspectable, and the one most likely to be mistaken for a mind because it talks.
Fuller would ask what it does for the whole. Lanier would ask who is inside it and who gets to leave. Neither question is about consciousness. Both are about worlds. That is the ground worth standing on: not "does the machine think?" but "what world does it make, and can we still live there?"
The answer is not in the weights. It is in the design — and the design is ours.