You're Ready. What Comes Next

A quick note on how we teach here. You won't find equations or numbers in this series, and that is on purpose. The goal is understanding, not arithmetic, so every idea is explained in plain words and simple pictures. When you want the math later, the rest of Internals Decoded is ready for you.

You have been climbing a gentle hill. One step at a time, you made your way up. Now you can turn around and look back. The whole path stretches out behind you, and you can see how far you have come.

This piece is a rest at the top. It is a chance to catch your breath, look at everything you now understand, and see clearly what comes next.

The path you walked

Let's walk back through the journey together, quickly, like flipping through a photo album. Each picture reminds you of a step you took, and each title takes you back to that piece if you want a second look.

You started with What a Computer Actually Is: a very literal helper that does exactly what it is told, quickly and without ever tiring. On, Off, and Everything Between showed you a wall of light switches, each one on or off, and named the bit. Turning the World Into Switches used a secret code between friends, and a picture made of tiny tiles like a mosaic, to show how photos, sounds and words become patterns of on and off. A Recipe the Machine Follows turned a program into a recipe card for a very literal cook, and called it an algorithm.

Then came patterns. What We Mean by a Pattern used a friend's footsteps in the hallway, the smell before rain, and a well worn path across a field. Turning Meaning Into Places laid things out like a supermarket or a library, so that similar things sit close together on a map of meaning. Guessing What Comes Next finished a friend's sentence, "I could really go for a cup of...", one word at a time. How Likely, Really looked at a gray sky and the sentence "She poured the tea into the...", gave every possible next word a share, and drew the winner like a raffle.

Next came learning. Two Ways to Make a Machine Smart set a rulebook beside a pile of examples. Learning From Examples worked through a stack of flashcards. Getting It Wrong on Purpose threw a ball at a bucket and measured each miss, the loss. Practice Makes It Smaller walked downhill through a foggy valley, one small step at a time.

Then you opened the lid. The Knobs Inside showed a giant mixing desk covered in knobs, the parameters. Loosely Inspired by the Brain folded a paper plane, which borrowed one idea from a bird without being a bird. Tiny Decision Makers in Layers weighed friends' opinions of a film the way a single neuron weighs its signals, then lined neurons up in layers. Why Deeper Is Different climbed from letters to words to the whole story. Studying vs Taking the Test split a machine's life into studying and the exam. What We Mean by a Model pointed to the chef, not the cooking school.

Finally, language. Breaking Language Into Pieces reached into a bucket of standard bricks and found the token. Words as Places in Meaning took a second look at the map and found that every place on it is a set of knobs, and that a direction can carry meaning. Careful or Adventurous put a dial on the raffle and turned it.

Every one of those ideas built on the one before it.

Words that are yours now

Along the way, you met some words that might have seemed scary at first. They are just names for ideas you now hold.

• Algorithm. A precise set of steps, a recipe a computer follows exactly.
• Embedding. A token's place on the map of meaning. Things that turn up in similar surroundings get places close together.
• Probability. A share of how expected something is. Every possible next word gets a share, and all the shares together make up the whole.
• Sampling. Drawing the next word like a raffle, where likely words hold more tickets.
• Machine learning. Making a machine smart by showing it examples instead of writing every rule by hand.
• Training. The long, patient process of learning from examples and mistakes, turning the knobs until the guesses come out right.
• Loss. How far each guess sits from the right answer, counting how sure the machine was. Training pushes it down.
• Gradient descent. Small steps downhill toward a smaller loss, like walking down a foggy valley.
• Backpropagation. The backward pass that works out which way every knob should turn.
• Learning rate. The setting that decides how big each downhill step is.
• Optimizer. The part of the system that runs the downhill walk.
• Parameters. The tiny knobs inside the machine that training turns. Most of them are also called weights. Together they hold what was learned.
• Neuron. A tiny decision maker that weighs the signals coming in and passes one small signal on.
• Layer. A row of neurons working side by side on the same input.
• Neural network. Neurons arranged in layers, passing signals forward, with knobs set by practice.
• Deep learning. Learning with many layers stacked deep, so ideas can build on ideas.
• Inference. The moment a trained model does its job, answering something new. No knobs change.
• Model. The finished, trained thing: every knob set and saved, ready to use.
• Large language model, or LLM. A very large model trained to guess the next token on a huge amount of text.
• Token. A small piece of language the model works with: a whole common word, or part of a rarer one.
• Tokenizer. The tool that chops text into tokens.
• Temperature. The dial on the raffle. Turned down, the model plays it safe. Turned up, the long shots win more often.

Before you move on, try this: cover the list and explain each word to yourself in a single sentence. Any word that will not come is a sign to reread its piece, and the path above links every one.

What comes next

You have reached the end of The Groundwork. The next series, AI, From Zero, picks up right here. Its first piece is What an LLM Actually Is (and Why Predicting Words Works).

Here is an honest word about what you will find there. You have the foundations, and every word you learned here will turn up again. But the next series also adds new ideas, like attention and the transformer, the design inside most of today's language models. You met attention briefly in Words as Places in Meaning: it is how a model decides which surrounding words to mix in. It gets a gentle piece of its own a little way into that series, Attention, Gently: How Models Decide What Matters. The next series also uses real numbers, and some of its pieces use math symbols. When you meet one, read the words around it first. The idea usually lives in the words, and it is fine if a formula stays foggy for now.

Go on. You have earned this.

Test yourself

A friend who has never read anything about AI asks you: "How did a chatbot learn to answer me?" Tell the story in a few sentences, using as many words from this series as you can.

Answer: Here is one way to tell it. People gathered a huge pile of text, the training data. A tokenizer chopped it into tokens, and each token got a place on a map of meaning, its embedding. A neural network with many layers of neurons practiced guessing the next token. After each guess it measured its loss, how far it was from the real next token. Backpropagation worked out which way every knob, every parameter, should turn, and gradient descent nudged them all a small step downhill. Over a mountain of examples, the knobs settled. A second round of training on helpful answers taught it to answer instead of just carrying on the text. What was left is the model, a large language model. When you chat, it does inference, and no knobs change. It gives every possible next token a probability and draws one like a raffle, with temperature setting how bold it is, one token at a time.

Next: What an LLM Actually Is (and Why Predicting Words Works), the first piece of AI, From Zero.