In Lesson 1 the board was a model — and a plain one only ever echoed where you poured the coins in. This lesson opens the board up to show where its knowing actually lives. It isn't in the pegs. Two new plain words fall out of it: weights and training.
Remember the square board. Pour the coins into slot 2 and they pile under slot 2; into slot 5 and they pile under slot 5. Useful? Not very — it just hands back roughly where you poured.
For the board to do something worth having, pouring in at slot 2 should be able to come out somewhere you didn't pick by hand — somewhere the board has been made to send it. And the way to do that is to change not the pegs, but the floor they stand on.
Picture two pots. A red one full of grease, a dark-blue one full of rubber paste. Dip a finger in each and wipe them across the floor of the board, until no bare wood is left.
Grease is slippery: where the floor is red, a coin slides through easily. Rubber is sticky: where it's blue, the coin resists and barely goes. Most of the floor is a blend of the two — every shade of violet in between. The pegs haven't moved an inch. But now a coin poured into slot 2 rides the warm, red trail across the floor and comes out under slot 5, because that's where the colour leads it.
Real grease and rubber wouldn't actually steer a coin sideways like this — that's not how the physics would work. It's a mental model, not a working machine. But it does exactly what a mental model needs to: colour on the floor decides which way a coin leans, and that's the part that matters.
That coloured floor is the whole trick. How red-or-blue each patch is — how strongly a coin is drawn through it or held back — is a weight. People also call it a parameter, especially when they're counting them up — a "three-billion-parameter model" simply means a floor with three billion of these patches painted on it. Same thing, two words. And the full map of colour across the floor is nothing less than what the board knows.
Here's where that count actually comes from. Picture our board as a simple grid: five pegs — five neurons — across, five rows deep. The painted floor lets a coin from any peg reach any peg in the row below, so every one of those pathways carries its own weight. Five pegs feeding five pegs is five times five: twenty-five separate pathways, twenty-five parameters, for just one hand-off between rows. Four hand-offs across five rows, and this one small board already holds four times twenty-five — a hundred parameters — before we've even stacked a second board on top.
Hand two people the identical wooden board and paint the floors differently, and they behave like two different machines. The pegs are shared and, on their own, know nothing. Everything a board has learned lives in the smear on the floor — never in the pegs.
Everything. The board you've been building — pegs, floor and all — is a working picture of an AI. You've already met its biggest idea; we simply hadn't said the word yet.
Lay the two side by side. The pegs are the AI's structure — fixed, the same for everyone who has that model. The smeared floor is everything this particular AI has learned: its own private map, and the whole reason it does what it does instead of something else.
This split is on purpose, and it isn't only a teaching trick. A real AI is packaged exactly this way: the structure and the trained weights are saved as two separate files, shipped together — one holds the shape, the other holds everything it learned, and neither is any use without the other.
And the act you pictured a moment ago — wiping grease and rubber across the floor until the coins flow where they should — is the heart of the matter. You smearing the board like that, patch by patch, until the right thing goes in and the right thing comes out: that is what training is. You trained the board. When the people who build an AI do exactly this — using mathematics instead of grease and rubber — "training" is the word they use for it.
Nobody does it by hand, patch by patch — there are far too many, and no one knows in advance what colour each should be. The floor is smeared in slowly, by the board being shown example after example: each time, the colour is nudged a touch — a little redder where a coin ought to have flowed, a little bluer where it shouldn't have. One example changes almost nothing. But over millions of them, a blank floor becomes the rich map you saw above — one that reliably carries the right thing in to the right place out. It's done once, by the people who build the AI, at enormous effort, long before it ever reaches you.
This is the part that settles most people's worry.
Training is the floor being painted. Using the board — the makers call it inference — is something else entirely: you drop a coin through a floor that is already finished and dry. The colour doesn't change because you used it. The board can't learn about you while you talk to it, can't quietly rearrange itself, and is the exact same board for you as for everyone else. What you send through is read to produce a result — and then the board is precisely as it was.
One board can only do so much. So we stack them: the coins that fall out of the bottom of the first board are poured straight into the top of the next.
Each board has its own painted floor — its own map. A coin threads down through all of them in turn, nudged by one map after another, before a result finally drops out the bottom. That stacking is all the word deep means in "deep learning": many boards, one on top of the next.
One note on size, so the drawing doesn't mislead you. These boards are 5 slots across and stacked three high, so you can see them. A real one has floors carrying billions of coloured patches, across far more boards than three. Not a different idea — the same idea, at a size no drawing can hold.
We've talked in coins and slots. But an AI takes in words, not coins. The next lesson shows the one simple move that turns this board into something that can read and write — and why it answers one word at a time.
Everything from the lessons before this one, plus what this lesson added.
| The board | The word for it |
|---|---|
| The board and its pegs | The structure — the model |
| A single peg | A neuron |
| A single step from one peg to the next | A connection |
| How red or blue a patch of floor is | A weight — a parameter |
| The whole painted floor | The trained model |
| Boards stacked, one on the next | Layers — "deep" |
| Smearing the floor in, coin by coin | Training |
| Dropping one coin through a finished board | Using it — inference |
| Which slot you pour into | Your input — your prompt |
9 rows now. It keeps growing as the course goes on.