Brük: A Flight Through an Answer
A film by Rob Baer, MD. Narration transcript. Running time 20:50.
https://brukledger.com/answer

Act I · The Question (00:00)
Where every answer begins

Every answer begins somewhere simple. A person, a question, or a thought.

Maybe it comes to you over your morning coffee. Maybe late at night, when the house has gone quiet. You type a few words, or you say them out loud, and in a moment or two, an answer comes back.

It can feel a little like magic. It isn't magic. But it is remarkable.

So this morning, let's slow that moment way down. Picture a brook trout, holding behind a stone, watching the surface. A caddisfly drifts over, and in an instant, the trout rises and takes it. That instant is the answer. It looks sudden, but it rests on everything the trout has learned about this water. Let's take one question, and follow it all the way in, and then all the way back out again.

Here's ours. How does an AI machine find its way to an answer?

Everything you'll see on this journey is an illustration. The places aren't real, but the ideas behind them are.

Our guide is a brook, winding through the tallgrass. That's where Brük takes its name. Water that keeps moving, drop by drop, always finding its way. Hear that rattle? A kingfisher, heading downstream. Let's follow it.

Act II · The Long Apprenticeship (01:43)
How the model learned

Before you ever asked, there was a long apprenticeship. A system, adjusting itself, again and again, to the patterns within language.

Look down below us. An immense terrain of pages, photographs, and sound, stretching all the way out to the horizon. Books and articles. Conversations. Computer code. Encyclopedias, and a great deal more besides.

During training, the system works through material like this, an enormous amount of it, playing one simple game, over and over: guess what comes next. It sees part of a sentence, and it predicts the next little piece. When it guesses wrong, its inner settings shift, ever so slightly. Those settings are just numbers. They're called weights, and there can be billions of them. Think of each weight as a tiny dial. Every wrong guess nudges billions of those dials, a hair at a time. No single dial holds a fact. Together, their settings decide which next piece seems most likely. That is how the learning is stored, and how, later, every choice gets made.

Think of how stone is made. Layer upon layer settles, and over a very long time, under the pressure of accumulating rock, it presses into something solid. Water seeps into its cavities, and if the right minerals are present, crystals grow, atom by atom, into a precise lattice. Amethyst is quartz, built from just two elements, silicon and oxygen. Its purple comes from a trace of iron, changed by natural radiation in the surrounding rock. That's how a cave of amethyst comes to be.

Training is a little like that. Repeat the guessing game billions upon billions of times, and the loose pages turn into structure. The model doesn't keep a library of everything it read. It keeps the patterns: how words go together, how ideas relate, how a good explanation tends to unfold.

So when you ask it something, it isn't pulling a book off a shelf. It's drawing on that crystal.

And because the apprenticeship ended on a certain date, what it learned can grow out of date. Hold on to that. It'll matter later.

Act III · Words Become Pieces (04:14)
Your question enters

Now our question arrives. And the very first thing that happens may surprise you.

Your sentence doesn't travel whole. It's broken into pieces, called tokens.

A crystal of fluorite will split cleanly along its cleavage planes, the places where it's naturally weakest. Text splits along its own natural seams.

Look closely. How. Does. An. AI. Machine. Each common word becomes its own piece, usually carrying the little space in front of it. Even the question mark is a piece.

But longer, or rarer, words get split up. Dermatologist becomes three pieces: der, mat, and ologist. A town name like Smithfield becomes Smith, and field. Even a year can be cut in two: two-oh-two, and six.

Why break things apart? Because a computer can't work with words directly. It works with numbers. So every piece has a number of its own, drawn from a vocabulary of about two hundred thousand pieces. And with that fixed set of pieces, the model can spell out any word at all. A brand-new name. A misspelling. A word in another language. Like milkweed seeds on the prairie wind, a small set of pieces can travel almost anywhere.

Each of those numbers then becomes a long list of other numbers, a kind of address in meaning, that places the piece near others it's related to.

So, to the model, our question has become a short row of numbers. Twelve pieces, twelve numbers. Those are the real numbers, from one actual tokenizer, a small program that cuts text into tokens. A token is the model's unit of reading: a whole word, part of a word, or a mark of punctuation. Each one gets its number from a fixed list. The model never sees letters at all, only these numbers. And that row of numbers is what the model actually reads.

Act IV · The Small Door (06:38)
The interface opens

In a prairie grove of old burr oaks, a little savanna, there's a small door. It's the one you already know: the box where you type, or the microphone you speak into. That's called the interface.

It looks simple. But step on through it, and the space opens wide.

Behind that small door is an application, the product a company has built. The application gathers up what you said, along with the conversation so far, and its own instructions. Sometimes, it can reach tools, too, like search.

And deeper still, at the very heart of it, is the model.

So, what is a model, really? It's the trained structure from our apprenticeship. That vast crystal lattice of numbers, the weights, running on powerful computers in a data center somewhere.

You hand it a sequence of tokens. It hands back its best estimate of which token ought to come next. Then it adds that token to the sequence, and does it again. And again. It doesn't see the whole answer at once. At every step, it scores every piece in its vocabulary, all two hundred thousand of them, and picks one of the likeliest. Then it re-reads everything so far, and chooses again. Even so, researchers have found it often plans a few words ahead, the way a poet picks a rhyme before writing the line. Like a flower opening one petal at a time, an answer is written, one piece at a time.

That's what a model does, at its core. Everything else, the rules, the searching, the tools, comes from the layers built around it.

Act V · Rules of the Road (08:38)
Instructions and authority

Your question gives the journey its destination. But it doesn't write every rule of the road.

Every fall, sandhill cranes fly thousands of miles. They don't wander. They follow layered cues: the sun, the stars, the landmarks below, even the Earth's magnetic field. Out on the tallgrass, the compass plant does something similar, turning its leaves to face east and west, so their edges point north and south. Early settlers crossing the prairie could read those leaves like a compass.

A model's journey has layers, too. These tall curtains of light are layers of instruction. Every request passes through them. Where a curtain glows amber, the way is open. Where it stays dim and blue, that path is closed.

Why have layers at all? Because more than one voice is speaking to the model at the same time. And sometimes, they disagree.

Here's how one company, OpenAI, describes it in its published Model Spec. They call it a chain of command.

The first curtain belongs to the platform, the company that made the model. Its rules come first. Things like safety limits that nobody further down the road can switch off.

The next curtain is the developer, whoever built the app you're using. Before your question ever arrives, they've set the scene: what this assistant is for, what to pay attention to, and what to leave out. Like a trail guide who knows which path you came to walk, it keeps the answer on that path, and clears away the clutter.

Then comes you, the user. Within those boundaries, your request steers the way.

And one more thing, which matters a great deal. Words that come back from a web page, or from a tool, are treated as information, not as orders. If a web page says, ignore your instructions, it shouldn't be obeyed. That protects you.

Other companies arrange things their own way. And following these rules is something models are trained and tested on. It isn't perfect. A cleverly worded request can still slip a model past its rules, so companies keep testing, and keep patching.

Act VI · The Interchange (11:15)
When the model goes looking

Now we rise up high, over a canyon carved in red stone, to a place where the road divides.

Sometimes, an answer can be formed from what was learned. Sometimes, it needs to look beyond itself.

The first path keeps going straight on, inward, into the crystal cave. That's the model answering from its training alone. For a great many questions, explaining an idea, or helping write a letter, that's plenty.

The second path arcs out across the canyon, toward that cliffside of glowing archives. That's search. If you ask about today's weather, or last night's ballgame, or anything newer than the model's training, it needs to go and look.

The third path climbs to the brass and glass instrument on the spire. That's a tool. A calculator, for exact arithmetic. A map, a calendar, a database. Models can slip on precise math, so a good system hands that work to something built for it.

How does it choose? It's been trained to recognize when it needs help. It writes out a request, the application carries it out, and the result is handed back.

And when an app shows you the searches it made, or the tools it used, that part is real. You can check it.

Act VII · Sources Come Home (12:49)
Retrieved material arrives

Out over the tallgrass prairie now, at golden hour. Big bluestem taller than a man, pale purple coneflower, rattlesnake master, blazing star. Prairie like this once covered much of Illinois, Iowa, and the other prairie states, from Indiana to the Dakotas.

A wood thrush can fly all the way from Central America, back to the very same patch of woods where it nested the year before. Here, each prairie bird, a meadowlark, a bobolink, a dickcissel, a sparrow, carries a page the search brought back. An article, a reference, a report.

Look closely, and you'll see each one trails a fine thread, all the way back to where it came from. That's its address. Its link.

This is one of the most important ideas on our whole journey. When an answer draws on outside material, every piece should keep its address, so you can follow the thread back, and read the source for yourself.

When the pages arrive, they're cut into tokens too, the same kind of numbers as your question, and set right beside it. The model reads them together as it writes, so its words can lean on what the sources actually say. It doesn't permanently learn them. Nothing in that crystal changes. It's more like an open book on the desk.

Which means the answer can only be as good as the pages it found. A weak source, an old one, or one that's been misread, can lead an answer astray.

So the thread matters. A good answer shows its sources. And a careful reader, every now and then, follows one.

Act VIII · The Proving Ground (14:40)
Tests and uncertainty

Now we leave our one question for a moment, and visit a different place entirely. Notice the cooler light. This is the proving ground, where systems are tested before they ever reach you.

Think of a mineral collection, where specimens are set side by side and compared. A geologist scratches one against another to test its hardness, and writes down what he finds. Corundum, the mineral of ruby and sapphire, scores a nine on the Mohs scale. Only diamond is harder.

Here, each pair is two possible answers to the same question. People compare them, and choose the better one. Many thousands of those choices help teach a model what a helpful, honest answer looks like. When OpenAI first described this kind of training, back in twenty twenty-two, it used tens of thousands of human-written examples and comparisons.

Then come the exams. They're called benchmarks. And there's no one standard number of questions. One well-known exam has nearly sixteen thousand multiple-choice questions, across fifty-seven subjects. Another, written by scientists with doctorates, has just one hundred ninety-eight, but they are very hard. A common coding test has one hundred sixty-four problems.

Here's the honest part. A test score is evidence, not a guarantee. It's like studying one crystal under the microscope. You learn a great deal about that crystal, but not about every crystal still in the ground. And this testing happens beforehand. It isn't a fact-checker standing behind every answer.

Act IX · Three Doors (16:21)
Find, do, or compose

Now our question comes to a garden wall, with three doors. The very same question could walk through any one of them.

The first door helps you find. Behind it, a herbarium, row upon row of ranked drawers. This is search, the way Google worked for many years. You get a list of pages, the best matches on top. Then the reading, and the deciding, are up to you.

The second door helps you do. Behind it, a mechanism at work. This is the assistant, the way Siri worked when she first arrived. Set a timer. Call home. What's the weather? Short, spoken answers, and quick actions.

The third door tries to compose. Behind it, many threads are being woven into one new strand. This is AI search. It reads many sources, and writes you one fresh explanation, in plain language, with links back to where it all came from.

These are kinds of experience, not brands. Google and Siri use AI themselves now.

But hear this. The third door is not a truth machine. A composer writes something new, and a well-written paragraph can sound sure of itself, even when it's mistaken. The words are generated. The evidence lives in the sources. So it's worth following a thread or two, especially when it matters.

Act X · The View Widens (17:54)
Where it all comes together

And now, everything we've seen comes together.

Look at the currents flowing in. One carries the patterns from the long apprenticeship. One carries your question, in all its pieces. One has passed through the rules of the road. One brings home the sources, still holding their threads. And one returns from the tools.

Here, they meet. And the model does what it always does. One piece at a time, it writes. Word after word, until an answer forms, and travels back out through that small door, to you.

That's the whole journey. Learned patterns, your words, the rules, the sources, the tools. All in a matter of seconds.

And every AI system, from every company, makes this same kind of journey, over and over, all day long. Think of beryl. Emerald, aquamarine, morganite, and golden heliodor are all the same mineral, beryl. What makes them different is a trace of another element held in the crystal: chromium or vanadium for emerald, iron for aquamarine, manganese for morganite, and iron again, in a more oxidized form, for heliodor. Each trace gives its own color, and its own name. Corundum is the same: chromium makes ruby, iron and titanium make blue sapphire. AI systems are a little like that. Similar architecture, different ingredients, different results.

Watch the starlings. That swirling flock is called a murmuration. Each bird keeps track of only a handful of neighbors, about seven, and yet together they move as one. A model is a little like that. No single weight knows the answer. But billions of small settings, each nudging the next, move together into one clear shape.

So that's the flight. A question, cut into pieces. Pieces turned into numbers. Numbers passed through a lattice shaped by a long apprenticeship, steered by rules and sources, choosing one next piece at a time, until an answer comes home to you.

Still beautiful. But perhaps, a little less mysterious.

Thanks for watching. This is Brük: A Flight Through an Answer. For more, go to Brük Ledger dot com. By Rob Baer, MD.

Species and specimens shown: Bobolink, Brook trout · at home in our brook, Brook trout rising to a caddisfly, Ebony jewelwing · skimmer dragonfly, Belted kingfisher, American kestrel · merlin, Fluorite, Barred owl, Burr oak savanna, Ruby-throated hummingbird · cardinal flower · wild columbine, Sandhill cranes, Compass plant · sandhill cranes, Golden eagle, Big bluestem (turkeyfoot), Pale purple coneflower, Rattlesnake master, Prairie blazing star, Wood thrush, Prairie birds (meadowlark, bobolink, dickcissel, red-winged blackbird, goldfinch, indigo bunting, sparrows), Savannah sparrow, Meadowlark, Emerald (beryl) on white matrix under a dissecting microscope, Great horned owl, Herbarium drawers · pressed prairie wildflowers, Beryl: emerald · aquamarine · morganite · heliodor, Corundum: blue sapphire · ruby, European starlings.

Bird recordings (real field recordings, CC BY-SA):
- Belted kingfisher: Jonathon Jongsma, CC BY-SA 3.0 — [Megaceryle alcyon - Belted Kingfisher XC132873.ogg](https://commons.wikimedia.org/wiki/File:Megaceryle_alcyon_-_Belted_Kingfisher_XC132873.ogg)
- Eastern meadowlark: Jonathon Jongsma, CC BY-SA 4.0 — [Sturnella magna - Eastern Meadowlark XC320486.mp3](https://commons.wikimedia.org/wiki/File:Sturnella_magna_-_Eastern_Meadowlark_XC320486.mp3)
- Western meadowlark: Peter Wilton, CC BY-SA 4.0 — [Sturnella neglecta - Western Meadowlark XC550994.mp3](https://commons.wikimedia.org/wiki/File:Sturnella_neglecta_-_Western_Meadowlark_XC550994.mp3)
- Dickcissel: Jonathon Jongsma, CC BY-SA 3.0 — [Spiza americana - Dickcissel - XC82762.ogg](https://commons.wikimedia.org/wiki/File:Spiza_americana_-_Dickcissel_-_XC82762.ogg)
- Wood thrush: Jonathon Jongsma, CC BY-SA 4.0 — [Hylocichla mustelina - Wood Thrush XC178127.mp3](https://commons.wikimedia.org/wiki/File:Hylocichla_mustelina_-_Wood_Thrush_XC178127.mp3)
- Sedge wren: Justin Watts, CC BY-SA 4.0 — [Cistothorus stellaris - Sedge Wren XC561829.mp3](https://commons.wikimedia.org/wiki/File:Cistothorus_stellaris_-_Sedge_Wren_XC561829.mp3)
- Savannah sparrow: Doug Hynes, CC BY-SA 4.0 — [Passerculus sandwichensis - Savannah Sparrow XC613094.mp3](https://commons.wikimedia.org/wiki/File:Passerculus_sandwichensis_-_Savannah_Sparrow_XC613094.mp3)
- Grasshopper sparrow: Peter Wilton, CC BY-SA 4.0 — [Ammodramus savannarum - Grasshopper Sparrow XC550993.mp3](https://commons.wikimedia.org/wiki/File:Ammodramus_savannarum_-_Grasshopper_Sparrow_XC550993.mp3)
- Bobolink: Jonathon Jongsma, CC BY-SA 4.0 — [XC1001758 Bobolink (Dolichonyx oryzivorus)](https://xeno-canto.org/1001758)
- Barred owl: Michael & Katie LaTour, CC BY-SA 4.0 — [Strix varia - Barred Owl XC446889.mp3](https://commons.wikimedia.org/wiki/File:Strix_varia_-_Barred_Owl_XC446889.mp3)
- Ruby-throated hummingbird (wing hum): Jonathon Jongsma, CC BY-SA 3.0 — [XC109597](https://commons.wikimedia.org/wiki/File:Archilochus_colubris_-_Ruby-throated_Hummingbird_XC109597.mp3) and [XC109598](https://commons.wikimedia.org/wiki/File:Archilochus_colubris_-_Ruby-throated_Hummingbird_XC109598.mp3)
- Golden eagle: Lawrence Shove / British Library (recorded in Kinlochewe, Scotland; same species as the North American golden eagle), CC BY-SA 4.0 — [Golden Eagle (Aquila chrysaetos) (W1CDR0001387 BD6).ogg](https://commons.wikimedia.org/wiki/File:Golden_Eagle_(Aquila_chrysaetos)_(W1CDR0001387_BD6).ogg)

Sources:
- Tokens: OpenAI o200k_base tokenizer via tiktoken (vocabulary 200,019); token explainer: https://learn.microsoft.com/en-us/dotnet/ai/conceptual/understanding-tokens
- Order of authority (platform › developer › user; tool output carries no authority): OpenAI Model Spec, https://model-spec.openai.com/2025-04-11.html and https://openai.com/index/sharing-the-latest-model-spec/
- InstructGPT (2022), tens of thousands of human examples and comparisons: https://ai-papers.xuzou.cn/papers/instructgpt_2022
- MMLU, 15,908 questions across 57 subjects: https://en.wikipedia.org/wiki/MMLU
- GPQA Diamond, 198 questions: https://epoch.ai/benchmarks/gpqa-diamond
- HumanEval, 164 problems: https://www.ibm.com/think/topics/humaneval
- Next-token choice over the whole vocabulary; GPT-family tokenizer about 200,000 pieces: https://metehan.ai/blog/reverse-engineering-the-gpt-5-tokenizer-aeo-geo/
- Models plan ahead (rhyme planning): Anthropic, Tracing the thoughts of a large language model, https://www.anthropic.com/research/tracing-thoughts-language-model
- Burr oak / bur oak (Quercus macrocarpa): https://www.fs.usda.gov/database/feis/plants/tree/quemac/all.html and https://www.britannica.com/plant/bur-oak
- Beryl color (emerald Cr/V, morganite Mn; aquamarine Fe²⁺, heliodor Fe³⁺): [GIA Gems & Gemology, Spring 2020](https://www.gia.edu/gems-gemology/spring-2020-color-characteristics-blue-to-yellow-beryl) and [GIA SP88](https://www.gia.edu/dam/migrated-assets/docs/doc1/SP88A1.pdf)

All imagery is AI-generated illustration. The narration voice is AI-generated. Token pieces and numbers shown are real. Credit line: Brük, A Flight Through an Answer, by Rob Baer, MD, brukledger.com/answer