A man in his sixties is walking along the waterfront in Bristol. The tide is out, the mud smells of salt and diesel, and he passes a sign reading Wapping Wharf. Something turns over in him. Without warning he is six years old again, lying on the floor in front of a record player, listening to a twelve-inch LP of Gulliver's Travels. He can see the cover: the giant pegged flat to the beach by a thousand tiny ropes. He can hear the narrator's voice. He remembers his father, who dealt in records, and the particular thrill of the needle dropping. And somewhere in the story, he is fairly sure, there was a forest — Wapping Forest. The whole thing arrives intact, vivid, true.
Then, because he is a man who has spent a lifetime questioning things, he does something unusual. He interrogates the memory while it is still warm in his hand. And it begins to come apart.
The record was real. The cover was real. The childhood, the father, the wonder — all real. But there is no such place as Wapping Forest, and there never was. There is the Wapping Wharf he passed minutes earlier on the Bristol harbourside, a name fresh in his eye. And there was, in the record of his childhood, very probably a forest after all — but it would have been Epping Forest, an old English name folded into the story decades ago. His mind, presented with a half-remembered forest and a freshly-seen wharf, had quietly soldered the two together. Wapping slid over to where Epping used to stand, the wharf lent its name to the forest, and out came a place that exists nowhere on earth yet felt, a moment ago, like the most settled fact in the world. The recent detail and the ancient one had been fused into a single seamless memory, and the seam was invisible.
The memory, in other words, is not a recording. It is a reconstruction. And the moment he notices this — the moment he catches his own mind in the act of assembling a plausible past out of mismatched parts — he has stumbled onto the single most useful image I know for explaining what artificial intelligence actually is, and what it is not.
The filing cabinet that never existed
Most of us were raised on a bad metaphor. We were told memory is a tape recorder, or a filing cabinet: an experience comes in, gets stored on its own labelled shelf, and waits there unchanged until we go and fetch it. It is a comforting picture, and it is wrong.
The evidence from cognitive science points somewhere stranger. Memory is not reproductive but reconstructive. There is no shelf, no file, no original cassette. Each time you remember something, you rebuild it from scattered components — a visual fragment here, an emotional tone there, a smell, a snatch of language, the spatial feel of a room — drawn from different regions of the brain and stitched together in the present moment into something that feels seamless. The Gulliver memory was never stored as "the Gulliver memory." It was stored as pieces: the pegged giant, the narrator's cadence, the father's presence, the texture of the sleeve, the recent walk along the docks, the name on the wharf. Recall assembles these. And because assembly is creative, it is also fallible. Wapping Forest was not retrieved. It was manufactured, on the spot and in good faith, out of a real wharf and a real forest, by a system doing exactly what it evolved to do.
This is the first thing worth sitting with, because it overturns an instinct. We tend to feel that a memory which isn't perfectly accurate is somehow degraded, a sign of a failing mind. But memory was never built to be a historical archive. It was built to be a useful model of the world. A hunter who recorded every leaf with photographic fidelity but couldn't connect last week's tracks to this week's would not last the winter. The one who drew analogies, completed patterns, and turned fragments into a coherent story survived. Memory is optimised for meaning, not for fidelity. The blending isn't a bug bolted onto an otherwise reliable machine. The blending is the machine.
And here is the privilege our man earned by living long enough: at sixty-five, with six decades of densely cross-linked experience, he can occasionally watch the reconstruction happen. He can see the threads from the wrong side of the tapestry. A twenty-year-old has too few memories for the seams to show. He has enough that, now and then, if he is paying attention, he can catch the loom mid-weave and ask the crucial question — where did that piece come from?
That question, it turns out, is the whole of artificial intelligence in embryo.
A warning before the metaphors
Before I take the next step, one caveat, and it matters more than anything that follows. Every comparison I am about to make is a teaching tool, not a literal description. The danger is not in the analogies themselves; it is in forgetting they are analogies. The brain is not "just a neural network," and an AI does not "think like a person." Both claims are seductive and both are false, and a great deal of muddled public conversation comes from people who let a useful picture harden into a literal belief.
We have done this before, and it worked precisely because we held the metaphors loosely. The early computing pioneers borrowed their whole vocabulary from biology — memory, neural network, learning, genetic algorithm, evolution. None of it was literally alive. All of it helped people grasp something otherwise opaque. So let us use the Gulliver memory the same way: as a scaffold, climbed deliberately, and kicked away once we have reached the height we need.
The storyteller in the pub
Picture an elderly gentleman in the corner of a pub. He has read, over a long life, more books, newspapers, manuals, and stories than anyone you have met. You ask him a question and he answers at once, fluently, without consulting anything. He simply speaks out of everything he has ever absorbed.
That is a large language model. Its strengths are real: enormous breadth, easy command of language, an almost uncanny instinct for pattern. And its weakness is exactly the weakness of our man on the docks. It does not retrieve a stored fact called "the answer." When you say Gulliver, it lights up a constellation — ships, Bristol, voyages, Swift, England, the sea — and constructs the most statistically plausible continuation. Often that continuation is correct. Sometimes it is confidently, fluently wrong.
We have a fashionable word for the wrong version: hallucination. But the word makes it sound like a glitch, a random spasm of nonsense, and it is nothing of the kind. A hallucination is far closer to a false memory. It is the model finding a path through related concepts that fits the pattern beautifully but is not anchored to any actual fact — "I know this story, I know this place, therefore these things probably belong together." That is precisely what happened with Wapping Forest. In the man, a freshly-seen wharf and an old half-remembered forest were pulled together because they were near neighbours — Wapping and Epping are one sound apart, and a wharf and a forest are both, at bottom, places — and the join was so smooth he never felt it happen. In a machine, the same fusion would occur for the same structural reason: the pieces sit close together in the space of related things, so the model bridges them and presents the bridge as fact. Different substrate entirely; remarkably similar result. The coherence is the trap. Coherence is not truth, and the only cure — in minds and machines alike — is verification. The man eventually asked whether Wapping Forest had ever really existed. The same question must be put to every confident output a model produces.
The notebook, the switchboard, and the junior staff
Now give the storyteller a notebook. When you ask him something, instead of answering purely from memory, he says, "Let me just check my notes first." The notes don't replace him — he still does the talking — but they pin his story to evidence and sharply reduce the odds of drifting into invention. That is retrieval-augmented generation: RAG. It is the artificial equivalent of the photographs, letters, and friends we use to stabilise our own unreliable recollections. And it is, almost exactly, what the man performed on himself. He used a recent fact — that sign, that walk — to audit and revise an old reconstruction. He did human RAG, on the spot, with his own past as the document store.
Give the storyteller a wall of telephones, each connected to a specialist — a doctor, an accountant, a librarian, the weather office — and a standard way of placing the calls, and you have MCP, the Model Context Protocol. It is tempting to file this under memory, but it isn't memory at all. It is closer to a nervous system: the brain cannot see or grasp or hear by itself; it sends signals down standardised pathways to eyes, hands, ears. MCP is the wiring that lets the model reach the calculator, the search engine, the calendar, the corporate database. The intelligence stays in the room; the nerves run out to the world.
And let the storyteller stop doing everything himself. Let him become a manager who delegates — one assistant researches, one drafts, one checks the facts, one reviews the quality, one delivers. Now you have an agent system, and at its most elaborate a committee of agents arguing toward an answer: I think it's A. I disagree, because of source B. Let's verify. The thing has stopped resembling a single mind and started resembling an organisation — a project team, a peer-review panel, the familiar cast of stakeholders, architects, and users that anyone who has run a large IT programme will recognise instantly.
The room full of cards
Beneath all of this sits the deepest and most beautiful parallel. Imagine a vast room in which every concept is a card, and the cards are filed not alphabetically but by similarity. Bristol sits near ports; ports near ships; ships near voyages; voyages near Gulliver. Reach into one corner and your hand falls naturally on its neighbours. This is associative memory, and it is roughly how the human mind appears to organise itself — one thing summoning the next, salt air calling up a childhood holiday, a dockside sign calling up a record from 1962.
The machine's version is the embedding: every concept assigned coordinates in a high-dimensional space, positioned by meaning, so that "dog" lives near "wolf" and "Gulliver" lives near "voyage." Stored in a vector database, these are retrieved by nearness, exactly like the cards. Neither system files by the alphabet. Both file by relationship. And this is exactly why the man's mind made the substitution it did. Wapping and Epping are near-neighbours by sound; a wharf and a forest are near-neighbours by kind. Two cards filed that close together are an accident waiting to happen — reach for one in dim light and your hand closes on the other. The fresh card, Wapping Wharf, had just been slotted into the drawer beside the old one, Epping Forest, and when recall came rummaging through, it pulled out a hybrid and never noticed the join. That is nearest-neighbour retrieval, in a skull and in a server alike, doing precisely what it is built to do — and occasionally getting it wrong in the most plausible way imaginable.
Heat, and the limits of thinking
Climb high enough up the scaffold and the analogy turns unexpectedly physical, and rather humbling. The brain runs on roughly twenty watts — about the draw of a dim lightbulb — and it cools itself with blood. A modern datacentre runs on megawatts and cools itself with vast engineered systems of water and air. Eighty-six billion neurons fire in parallel; thousands of GPU cores grind through tensors in parallel; in neither case does intelligence emerge from a single tidy calculation. It emerges from staggering numbers of simultaneous interactions. And both, in the end, are governed by the same iron law: thinking generates heat, and heat must go somewhere. Cognition is constrained by thermodynamics whether it is wet or dry. There is something steadying in that — a reminder that for all the talk of disembodied minds, no one has yet found a way to think without paying the entropy bill.
Pull back further still and the individual machine dissolves, just as the individual human mind always did. No person thinks entirely alone; we are embedded in families, firms, universities, libraries, whole civilisations of accumulated knowledge. Modern AI is going the same way — models talking to tools, tools to databases, databases to applications, clouds cooperating with clouds. The intelligence no longer sits in one box. It emerges from the network.
The one difference that does not dissolve
It would be easy, having climbed all of this, to conclude that mind and machine are converging into the same thing. They are not, and the place where they part is the place that matters most.
The man remembers listening to Gulliver with his father. A model can tell you, fluently and at length, what grief is — its stages, its literature, its neurochemistry. But it has never stood on a dock and felt fifty years collapse into a single moment because of a name on a sign. It has no father to have lost. It has no body that aches, no mortality pressing at the edges of its attention, no stake in the story it is telling. Embodiment, emotion, finitude, the simple fact of having been somewhere — these are not features the machine is missing on its way to acquiring them. They are a different category of thing entirely. The model knows about. The man remembers. Those are not two points on one scale. They are different worlds.
This is not a reason to dismiss the machines. It is a reason to use them honestly — for what they are good at, which is breadth, pattern, language, and tireless retrieval, while reserving for ourselves the thing they cannot do, which is to mean it.
What the man was really doing
The original question was never about Gulliver. It was about how minds connect things — and how, increasingly, two very different kinds of mind connect things side by side.
A childhood record became a walk in Bristol. The walk became a question about memory. Memory became a question about hallucination. Hallucination opened the door onto the whole apparatus — RAG, agents, MCP, embeddings, GPUs, datacentres, the heat and the networks and the civilisation-scale machinery now coming into being. And the spine running through all of it was a single, ordinary, very human act: a man catching his own mind in the act of reconstruction and asking, where did that piece come from?
That is the act good AI is straining to learn. Not merely to produce an answer, but to examine where the pieces came from, how they were assembled, and how much it ought to trust the result — to separate the original source from the later addition, the inference from the fact, the thing it knows from the thing it merely finds plausible. The man on the docks did it instinctively, with sixty-five years of practice. The machines are not there yet. Some of us are not sure they can get there. But the thing they are reaching for has a name, and he was doing it, unprompted, in front of a sign reading Wapping Wharf — quietly dismantling a forest his own mind had built out of it not a minute before.
That is why I would start here, and not with transformers, if I had to explain any of this to someone who finds the whole subject impenetrable. Begin with a record, a father, and a misremembered place name. People understand memory first, because every one of them has a Gulliver of their own — some vivid, certain, slightly-wrong recollection they would swear to in court. Show them that their own mind reconstructs rather than records, that it blends and completes and occasionally invents in perfectly good faith — and you have already taught them most of what they need to know about the machines.
Begin with the human. Then the machines make sense. Never the other way round.
✦