A plain-language explainer, start to finish

The Whole Story

Three earlier pages each covered one piece of this. Here they are as one story — from a folder of old plays, to a working AI, to one sentence it actually wrote. Every number below is real.

A pile of real text

Before anything else existed, we needed something for the AI to learn from. We used a big chunk of Shakespeare's plays — plain text, nothing fancy.

40,000lines of text
1.06 MBfile size
65distinct symbols used

Those 65 symbols — every letter, space, and punctuation mark that actually appears in the file — became our dictionary: R is 30, O is 27, and so on, one number per symbol, nothing more. We didn't invent this list. We just counted what the training text actually contained.

Turning the pile of text into a working AI

Starting from 10,788,929 random numbers, we ran one loop, over and over: show it real text, ask it to guess the next letter, check the guess, nudge every number a little toward being less wrong. We did this 5,000 times.

4.2886how wrong — before training
0.9795how wrong — after training

Full story, including the part where it started memorizing its homework: Getting Less Wrong →

You type something in

Now the AI is trained. You give it six letters: "ROMEO:" The dictionary from step one turns them into numbers — the exact same list, used in reverse:

R → 30O → 27M → 25E → 17O → 27: → 10

The part with no dictionary

The trained AI does a huge amount of math on those six numbers. Partway through, it produces 384 real numbers — captured from this exact moment, nothing staged. Nobody has ever written down what any of them means:

Why that dictionary was never written, with the soundboard explanation: The Unlabeled Dials →

Squeezed back into an answer

Those 384 mystery numbers get squeezed down to exactly 65 — matching the original dictionary again. This new list is readable: it's the AI's real odds for what letter comes next.

\n
99.9497%
Y
0.0122%
'
0.0106%
S
0.0090%
O
0.0022%

How the squeeze works, and why the AI doesn't always pick the top choice: The Next Letter →

The final result

That whole process — dictionary in, unlabeled math, dictionary out, weighted lottery — produces exactly one letter. Then it happens again, using the new, slightly longer text as the new starting point. And again. We ran it 40 times:

ROMEO:
Yet be not desired to repent me: and yo
Marked letters are the real long-shot picks — moments the weighted lottery didn't land on the top choice. Everything else is the AI's top pick, winning as expected. Every one of the 40 letters above went through every step on this page, once each.

Start to finish, in one line

A pile of old plays became a dictionary. A training loop, repeated 5,000 times, turned random numbers into a machine that's good at guessing what letter comes next. A prompt goes in through that same dictionary. What happens in between has no dictionary at all — real, watchable, unreadable. What comes out the other side does, every time, exactly and completely. That's not a story about an AI secretly thinking. It's a story about a very large, very patient guessing machine — and now you've seen every real number it took to write one sentence.

The pages this is built from

Getting Less WrongWhere the AI's numbers came from — the real training loop and loss curve. The Unlabeled DialsThe real 384 numbers with no dictionary, and why. The Next LetterHow those numbers become real, readable odds again.