A plain-language explainer, part three

Getting Less Wrong

The first two parts showed you what our tiny AI looks like once it already knows how to guess text. This part shows where that ability actually came from — and it's a much duller story than "learning" makes it sound.

Quick recap: The Unlabeled Dials showed a real dictionary for letters and 384 real numbers with no dictionary. The Next Letter showed those numbers turning into real odds for what comes next. This page shows how the AI got good at producing odds like that in the first place.
1

Before training: pure noise

Every one of the AI's 10,788,929 internal numbers starts out random — no pattern, no letters, nothing copied from Shakespeare. Ask it to guess the next letter and it's no better than picking blind out of 65 letters.

We checked this for real. A model that's totally guessing should be wrong by a very specific, calculable amount. Ours, on its very first attempt, was wrong by 4.2886 — almost exactly the number pure random guessing predicts. The AI started out exactly as clueless as the math says it should.
2

The loop that fixes it

There's no separate "teaching" step, no explanations, no rules written down anywhere. Just one loop, repeated over and over:

  1. Show the AI a real line of Shakespeare.
  2. Ask it to guess the next letter.
  3. Check the guess against what the letter actually was.
  4. Nudge all 10.8 million numbers a tiny bit — in whatever direction would have made that one guess slightly better.
  5. Throw away that sentence, grab a new one, and do it again.

We ran this loop 5,000 times. Nobody ever told it what a "word" is, or what a "sentence" is, or gave it a grammar rule. Every nudge came from one thing only: was this specific guess a little less wrong than before?

3

Watching it happen, for real

Here's how wrong the AI actually was, checked every so often during those 5,000 loops. Shorter bar means less wrong:

step 0
4.2886
step 1000
1.5153
step 2000
1.2829
step 3000
1.1632
step 4000
1.0722
step 4999
0.9795

It drops fast at first, then more slowly — each loop teaches it a little less than the one before, the same way the first hour of practicing anything teaches you more than the hundredth.

An honest complication: we also checked the AI against sentences it had never practiced on. Up to about step 3,750, it kept getting better at those too. After that, it kept improving on its practice sentences but quietly got worse on new ones — a sign it had started memorizing its homework instead of getting generally better at English. Real training runs hit this, and ours did too.

So did training write the missing dictionary?

No — and now you can see exactly why. Every nudge in that loop was graded on one thing only: did the AI's final 65-letter guess get a little better? Nobody ever checked, or cared, what any single one of the 384 numbers in between was doing to help. As long as the final guess improved, the nudge stuck, whatever it did to those numbers.

That's the whole reason no dictionary exists for them. It's not that one was written and then lost. Training was never asked to produce one — only to make the final answer better, 5,000 times in a row. It succeeded at exactly that, and nothing more.