What comes next?

Before anything gets a name, play one round of a game. Below are five practice sentences, written for exactly this moment. Under them, a sixth sentence stops early. What word comes next?

However you did, notice what you just used: not a dictionary, not grammar — the five sentences themselves. Words repeat, so the past of a text is a decent guide to its next word. Using only the sentences above, you can beat blind luck at this game, and you can check that claim by counting. Hold onto the feeling. It is the seed of everything on this page.

02 But one small word wrecks the trick

Your counting trick has a weakness, and it is nine letters long: “the bank”. In the guide's text, money sleeps in a bank and a river sleeps against one. Count how often each word follows “the”, and “bank” tells you nothing about which bank you are in. A guesser that counts single words can do no better than a coin-flip here. Watch it flip — then give it one extra word of memory.

That is the whole discovery, and it is a big one: how much a guesser remembers changes what it can guess. Now crank the guess forward. Predict a word, keep it, predict again from what you now have — around and around. A machine doing exactly that, with a long memory and a very large pile of text, is what everyone now calls a chatbot. We will build one — a tiny, honest one — right now.

03 So build the whole machine yourself

Here is a small page of text — about 250 words, written to be checkable. And here is a complete guessing machine that reads nothing else. Every bar it shows you is a count you can verify by eye in the text directly above it: click any bar and count the highlights. No secrets, no cloud, no maths you cannot check on your fingers.

Three things are worth catching with your own hands. Set temperature to zero and let it write: it walks into a loop and repeats itself, word for word, forever — pick the safest word every time and you get the flattest possible text. Nudge the temperature up and it takes small risks; nobody ever derived the “right” setting, people just tried numbers and kept what read well. Then push memory to three words and watch closely: the writing turns fluent — because it has started quoting the page outright. You caught it cheating by memorising.

Now the honest confession this whole page has been building to. The machines in your pocket work like this toy — predict, pick, repeat — but they cannot get there by counting, because language is too new: with a 40,000-word vocabulary there are more three-word openings than sentences ever written. So instead of a table of counts, a large language model — an LLM, the engine behind the chatbots — learns a compressed pattern from more text than you could read in a thousand lifetimes, and uses it to guess well on sentences no one has ever said. Fluent for exactly the reason your toy is fluent. Wrong, when it is wrong, for the same reason too.

Deeper reads, kept short: What even is AI? · What's an LLM? — and this card you just met feeds a live picture of your own memory.

04 But the real one has read a thousand lifetimes — and still fumbles

How much reading are we talking about? One open model — Llama 3, whose maker published the number — trained on about fifteen trillion tokens of text. Make that personal: at 250 words a minute, reading nonstop with no sleep, that pile would take you about 86,000 years. Drag the number to your own speed; the sentence will keep up.

The stripe in that chart is doing honest work. Nobody knows exactly how much public text humanity has written, so researchers publish a range — best guess around 300 trillion tokens, could be a third of that, could be three times it. The true number hides somewhere in the stripe. The same researchers project that models will have read essentially all of it sometime between 2026 and 2032 — their projection, not this page’s promise. What happens after that is a question we will leave hanging — deliberately — until the last page.

And yet. A thing that has read a thousand lifetimes still fumbles things a child gets. Its skill is not a smooth slope; it is Swiss cheese — real substance, real holes. Before you flip each card below, place your bet: which half fails?

05 And worse — it makes things up with a straight face

Ask the toy above something factual — seed “the money” and let it write — and out comes “the money sleeps in the bank at night”, fluently, because fluent is all it does. Watch the highlights while it writes: no fact is being looked up. Fragments of the page are being continued. Now scale that up: these are pattern-spotting engines, not search engines. Nothing inside checks “is that true?” before the next word lands.

People have already learnt this the expensive way. A New York lawyer filed a brief full of court cases a chatbot had invented for him — names, dates, page numbers, all fluent, none real. “I did not comprehend,” he told the judge, “that ChatGPT could fabricate cases.” Now you comprehend it: fabrication is not a malfunction of the machine you built above. It is the machine you built above, running as designed.

Here is the strangest part, and the most hopeful: some of the making-things-up is not ability at all — it is marking. The machine was graded, during training, the way schools grade a multiple-choice test: right answers score, blanks score nothing. Under those rules, what is the clever move on a question you do not know? Guess confidently. Change the marking and watch the same machine change its behaviour.

06 So give the guesser hands — and watch every step

A guesser that can only talk is stuck the moment talk is not enough. Ask one for 47 × 93 in a single breath and it must continue the digits from pattern — and the pages it read rarely continue them correctly. But give the same guesser two tools — a calculator, and a search over its own page — plus a goal and permission to work in steps, and something new appears. An agent: a guessing machine that runs tools in a loop until a goal is met. That one sentence is the whole idea.

Three things you just saw, worth keeping. The loop is not magic: think, use a tool, look at the result, decide again. The loop is not unstoppable: the exit is a visible check, and you tripped it yourself. And the loop is not free-willed: switch a tool off and the agent does not rage — it halts and names what it lacked. What it may touch was your setting all along. Steps matter for a quieter reason too: a guesser spends a fixed sliver of effort per word, so a machine allowed to think out loud in stages genuinely does better than one forced to blurt — you watched that happen with the calculator.

Deeper read: What's an agent? — and in the Build wing of this site you assemble a real one, in plain English.

07 But rewind to 1960 — the other branch made you smarter

Everything so far sits on one branch of a fork that is now sixty-six years old. Around 1960, two ideas split. John McCarthy's camp asked: can we make machines intelligent? Doug Engelbart's asked something slyer: can machines make people more capable? Artificial intelligence, and intelligence augmentation. Both branches shipped. You have been reading about the first. You have been using the second your whole life.

Engelbart wrote the founding paper in 1962 and defined the goal precisely: increasing a person's capability “to approach a complex problem situation, to gain comprehension to suit his particular needs, and to derive solutions to problems.” Not a smarter machine — a more capable you. His sharpest insight was that the smart unit was never the person alone: it is the person plus their language, their tools and their methods, trained together as one system. Upgrade any part and the whole person-system gets smarter.

He proved the point backwards, with a brick. Strap a brick to a pencil and writing turns slow, wobbly and exhausting — and suddenly you notice how much of “being good at writing” was living in the pencil. Feel it yourself:

Engelbart's ally Licklider timed his own working days in 1957 and found something embarrassing: about 85 per cent of “thinking time” went on getting into a position to think — finding, fetching, formatting, arranging. One man's diary, mind, not a law of nature. Try the number on yourself: what share of your work is getting-ready work? Say 60% —

So where do this decade's tools sit? Licklider drew a line worth keeping: a plain extension does what you steer, stroke by stroke; a true partner contributes something you could not, the way the fig tree and the fig wasp each do what the other cannot; and a replacement does the job while you leave the room. Drag today's tools onto that line wherever you judge they belong. The page will not correct you — there is no answer key, and people who build these tools still argue about it at dinner.

One last experiment before the fork resolves. Below are three messages; put them in the order you would answer them. Then do it again with the machine's suggestion showing. Both orders are defensible. The question this page cares about is smaller and sharper: the second time — who decided?

Deeper read: Augmented intelligence.

08 And that branch has climbed onto your face

You probably used the augmentation branch this week without naming it. Pointing a phone camera at a plant, a menu, a foreign sign — that is augmented reality: useful information laid over the world you are already looking at. Today it runs ten-to-twenty billion times a month through one visual search tool alone; the honest figure is a range because even its owner counts it differently year to year.

The glasses have arrived too. Since September 2025 you can buy a $799 pair with a small display in one lens — and a wristband that steers it by reading the tiny electrical signals your muscles make before your fingers move. Captions for the deaf, translations for the traveller, directions for the lost, floating on the world itself. And the sci-fi pair — wide field of view, the whole movie dream? It exists. It costs about $10,000 a unit to build, and you cannot buy it at any price. The frontier right now is not capability. It is price, weight and battery.

Are the glasses a fad? Here is every number the record actually holds: about two million pairs sold by early 2025, then about seven million in 2025 alone — while the $3,499 headset of the same era reportedly sold under half a million in its launch year — one holiday quarter was estimated in the tens of thousands. Draw where you think the curve goes next. The reveal will not draw the future for you, because the record stops where records stop.

Deeper read: Augmented reality.

09 Now look back — you already used this whole site once

A confession about the ground you are standing on. This page sits inside a site with three wings — Discover, Explore, Build — and rather than describe them, this page has been quietly walking you through all three. Engelbart would have recognised the floor plan: upgrade a person's language (names and maps for whole fields — that is Discover), upgrade their methods (poking a live model until it makes sense — that is Explore), upgrade their artifacts (tools of your own — that is Build). Here is your own trail, kept on this page, in this browser, sent nowhere:

Did the mapping hold? You are the only fair judge of that. When you leave this page, the three wings are simply this lap, done again on fields you actually care about — see one at a glance, drop in and poke it, then make a small tool of your own. And what you build tends to make something new worth noticing, so the loop starts over, one rung higher. That is the site's whole purpose, and you have now felt it once from the inside.

10 So build one — the last word here is yours

Every control below is one you have already used on its own — the memory slider, the temperature, the marking scheme, the two tools with their exit door, and the page of text itself, which you may now write into. Nothing here is new except the combination you choose. Assemble your own tiny answerer, and mind the one lesson every section has repeated: it will be exactly as good, and exactly as honest, as what you gave it.

One widening thought before the blank page. The trick you now understand — guess the next element, keep it, guess again — was never about words in particular. People have pointed the very same trick at pictures, music and molecules. Reveal each one:

What else is a sequence? Weather is. Traffic is. Protein folding is. Your calendar is. Whatever you just thought of — someone is probably already pointing this trick at it.

Two questions this page cannot answer, because nobody can yet. Researchers project that somewhere between 2026 and 2032 the machines will have read essentially everything public humanity ever wrote — then what? And the bigger one: what will people actually use these things for, ten and twenty years from now? Every confident expert answer so far is fluent and unverifiable — and you now know exactly what that combination is worth. So the last word on this page is yours, not ours.

The guide ends; the site begins. Three doors, in the order of the lap you just ran — and the shorter reads, whenever you want one idea at a time.

Inside Explore: a real LLM you can open up and watch every layer move as it reads.

The shorter reads: What even is AI? · What's an LLM? · What's an agent? · Augmented intelligence · Augmented reality · Your memory, drawn live · Review your cards

Curious why this page teaches by letting you poke at things rather than lecturing? That's not ours to claim — it comes from a handful of people we name on the influences page.