In the last two lessons the board took something in and sent it somewhere out. This lesson makes the one small leap that turns it into something that reads and writes — and shows that "talking" is nothing grander than guessing the next word, again and again.
A trained board is a mapping machine. Pour something in at the top, and the heat map on the floor carries it to an output at the bottom. Simple enough — but everything we've poured in and read out has been a bare slot, a position with a number.
To read and write, the board needs to deal in words. And the leap to get there is far smaller than you'd think — in fact, we don't have to change the board at all.
Take the very same board and relabel the bins along the bottom. Instead of 1, 2, 3, 4, 5, write words. Nothing inside changes — only the labels.
Now pouring the coins in stops meaning "drop at slot 2" and starts meaning "here are the words so far." And the bin the coins pile up in stops being a number and becomes the board's answer to a single question: what word comes next? This is still the trained, painted floor from the last lesson — so the pile won't just fall straight down from where you poured. Feed it "Cat eats ___", and the warm path carries the coins sideways, stacking highest under mouse.
That's the whole leap. The machine that sorted coins into numbered bins now takes words in and gives a word back — because we agreed to read its slots as words. Notice, too, that mouse isn't the bin directly under the funnel: it's wherever the trained floor leads. An untrained board would just pile up straight down the middle, same as Lesson 1. This one has learned to send "Cat eats" specifically toward mouse.
Notice what the board did — and what it didn't. It didn't compose a sentence or plan an answer. It answered one tiny question: given the words so far, which word most likely comes next?
You do this yourself without noticing. Read "Once upon a ___" and your mind hands you time before you've even finished. The board does exactly that — and it's the very same prediction from Lesson 1, a pile that's tall in the likely place, only now the bins are words. Mouse is tallest; fish and bird get a few coins; grass and shoe almost none. A favourite, and some runners-up.
But a single word isn't a conversation. So here is the trick that turns one small guess into whole paragraphs: it simply does it again.
Guess a word. Write it on the end. Now feed the whole, slightly longer line back into the board and guess the next word. Then again, and again. "Cat" leads to "eats"; "Cat eats" leads to "mouse" — and a sentence builds itself, one word at a time.
That loop is the entire engine. Every answer you have ever seen an AI give — however long, however fluent — was laid down this way: one most-likely word, then the next, then the next.
Once you can see the loop, a lot of what the thing does stops being mysterious.
It writes left to right and seems to type before your eyes — because it is, quite literally, producing one word at a time. It doesn't think up a finished answer and then read it out to you; it is always just extending what's there by one likely word. It isn't a mind composing a reply. It's the most well-read machine ever built, finishing your sentence — and then finishing that, and then finishing that.
Talking is next-word prediction in a loop. The board guesses the likeliest next word, tacks it on, and runs again. Nothing more mystical is going on.
There's one simplification we've been quietly making. We called the bins "words" — mouse, fish, bird — because it was easier to picture. That's not quite precise.
Real systems don't cut language into whole words. They cut it into smaller pieces — sometimes a whole word, sometimes only part of one. The next lesson names that piece properly, and gives us the word missing from our table so far.
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 |
| + New in this lesson | |
| The bins, relabelled with tokens | The vocabulary — every token it could pick |
| The tallest pile among the bins | The prediction — the likeliest next token |
| Feeding the growing line back in and running again | The loop — how one guess becomes a sentence |
12 rows now. It keeps growing as the course goes on.