How a model learns (the guessing game)

Here's the analogy to hold onto: imagine a game where you try to guess the next letter of a Pokémon name, over and over. Every time you guess, a friend tells you how wrong you were. If you keep score and adjust after every round, you slowly get better. That is exactly how our model learns.

The four steps, on repeat

1. Guess. The model reads some letters and gives a score to every possible next letter.
2. Measure the surprise (the loss). We look at how much probability the model gave the correct next letter. Confident and right → tiny loss. Confident and wrong → huge loss. This number is the loss: lower is better.
3. Trace the blame (the gradient). We walk backward through every step the model took and work out, for each internal dial, "which way should I turn you, and how much, to make the loss smaller?" That direction-and-amount is the gradient (see the breadcrumb-trail analogy in autograd.md).
4. Nudge the dials. The optimizer turns every dial a tiny step in the blame-reducing direction.

Repeat thousands of times and the loss falls: the model has learned.

A tiny worked example

Train a bigram on "abcabcabc...". At first it guesses randomly (loss ≈ log of the vocab size). After a few hundred steps it has noticed "after a comes b, after b comes c" and the loss drops close to zero — it has memorized the pattern.

Why does this matter?

Every AI that "learns from data" — from Pokémon-name dreamers to giant chatbots — is doing this same loop: guess, measure surprise, trace blame, nudge. The only differences are how big the model is and how much text it reads.