The key ideas

What Is ChatGPT Doing

Stephen Wolfram, 2023

The most famous neural network in the world doesn't know how its own sentence will end — it simply guesses the next word. Stephen Wolfram takes ChatGPT apart piece by piece and shows why that's enough to write like a human, and what this discovery says about our own language. You'll understand how neural networks, embeddings and the transformer work, why ChatGPT is confidently wrong, and how to fix that.

Who it is for

For everyone who uses ChatGPT and wants to understand what's going on inside it — no formulas or programming required. And for anyone curious about what neural networks reveal about human language and thought.

ChatGPT just keeps adding the next word, over and over, and the fact that this is enough for meaningful text tells us that human language is simpler than we thought.

What the book is about

The most famous neural network in the world doesn't know how its own sentence will end. When ChatGPT answers you, it has no plan and no ready-made answer — it simply guesses the next word, then another one. So how do essays, poems and code come out of that? Stephen Wolfram, physicist, mathematician and creator of Mathematica and Wolfram|Alpha, takes this machine apart piece by piece — no formulas, for anyone who's curious.

And he arrives at a conclusion more interesting than the neural network itself: if guessing words is enough to write like a human, then our language and part of our thinking are simpler than we thought. You'll get to grips with neural networks, training, embeddings and the transformer, understand why ChatGPT is confidently wrong, and what Wolfram suggests putting beside it. And by the way: the very fact that you decided to understand how a tool you use actually works, instead of just using it blindly, is already a rare and smart habit.

Idea 1. All the magic is in guessing the next word

Here's a fact that can feel almost insulting at first. All ChatGPT does is try to produce a “reasonable continuation” of the text written so far. That's it. No secret mind. “Reasonable,” for Wolfram, means what a person who had read billions of web pages would most likely write. But if it's that simple — where does the feeling of talking to a living person come from?

The mechanism is modest. The model looks at all the text up to this moment — your question plus everything it has already written itself. It calculates which pieces could come next and with what probability. It picks one. Adds it. And repeats the whole thing again — now with slightly longer text. These pieces are called tokens: whole words or parts of words. A thousand-word answer is a thousand-plus of these steps.

The mistake almost every user makes is imagining that a finished answer sits inside and the model “reads it out.” There's no essay plan. There's only the next step. And the feeling of a living conversation partner comes from the fact that every new word has to fit everything said before it.

What this means for you: ChatGPT's confident tone is a feature of style, not a guarantee of truth. Remember this right now — later it will become clear why it matters so much.

You type: “The main benefit of morning runs is that…” — and piece by piece the screen fills with: “they,” “give,” “you energy,” “for the whole day.” None of these words existed in advance. Each one was born only once the previous one was ready.

ChatGPT doesn't recall an answer — it guesses the next piece of text, again and again.

Read in Livria

About the author

Stephen Wolfram is a British-American scientist, physicist and mathematician, the creator of Mathematica, the Wolfram Language and Wolfram|Alpha, and the author of “A New Kind of Science.”

All books by this author: Stephen Wolfram

Terms from the book

  • token
  • Temperature: A setting that determines how often the model picks a word other than the most likely one. The higher it is, the more varied and unpredictable the text.
  • Neural network: A mathematical model made of many simple “neurons” connected by weights. It is tuned on examples rather than programmed with rules.
  • Loss function: A number that shows how far the neural network's answers are from the correct ones. Training tries to make it as small as possible.
  • Gradient descent: A way of training: the weights are nudged little by little in the direction where the error gets smaller, like a hiker walking down a slope in the fog.
  • Embedding: A representation of a word or piece of text as a list of numbers — a point on a “map of meanings” where words with similar meanings end up close together.
  • Transformer: The neural network architecture ChatGPT is built on. Its key part is the attention mechanism, which decides which earlier words to look at.
  • Computational irreducibility: Wolfram's concept: some processes have no shortcut — the only way to find out the result is to go through every step of the computation.

All terms

Similar books