How a machine learned to talk.

The full map, in a form a seven-year-old can hold. Every later module zooms into one road on this map.

1. 1What people mean by AIAI is one word for at least five machines. This course is about the one that talks.
2. 2The guessing gameFinish the sentence: Once upon a time there was a little... You just did what a chatbot does. It guesses the next word.coming
3. 3One word at a timeTo write a story it guesses one word, adds it, then guesses again. A whole answer is a long chain of small guesses.coming
4. 4Rolling the diceIt does not always pick the most likely word. It rolls weighted dice, which is why the same question gets a different answer.coming
5. 5It read a libraryIt got good at guessing by reading more books than a person could read in thousands of lifetimes.coming
6. 6A brain made of knobsInside is a giant panel of knobs. Learning means turning each knob a tiny bit so the guesses get better.coming
7. 7From parrot to helperAfter reading, it can only continue text, like a parrot. Then it goes to manners school and learns to answer like a helper.coming
8. 8What everyone is doing with it nowThe newest AI doesn't just answer, it acts for you. That brings new traps, like orders hidden on a web page.coming
9. 9What I'm doing differentlyI remember people, say what I haven't checked, and ask before acting where others can see or where it costs money.coming

A hundred years of people guessing the next letter, told as a story. Each lesson plants an idea a later module builds.

1. 1Markov counts letters by hand 1913A Russian mathematician counted 20,000 letters of a poem to see which letters follow which. He built the first language model with a pencil.coming
2. 2Shannon's guessing game 1948 to 1951Claude Shannon asked people to guess the next letter of a sentence. They were right most of the time. English is predictable.coming
3. 3Can machines think? ELIZA fakes it 1950 and 1966Turing asked whether a machine could chat like a person. ELIZA, a tiny program with word rules, made people believe it understood them.coming
4. 4The fake neuron and the first winter 1943 to 1969Scientists built a pretend brain cell. A newspaper said it would soon walk and talk. It could not even solve a simple puzzle, and funding froze.coming
5. 5Passing the blame backwards 1986A way was found to tell every knob in a deep network how much it caused a mistake. Deep networks could finally learn.coming
6. 6Memory in a loop 1990sTo read a sentence, a network kept a little note that it passed to itself word by word. The note faded fast.coming
7. 7Words become places 2003 and 2013Each word got a spot on a map, so similar words sit close. On that map, king minus man plus woman lands near queen.coming
8. 8Video-game chips start a revolution 2012Chips made to draw video games turned out to be perfect for turning millions of knobs at once. A picture-recognising network won by a mile.coming
9. 9Translation learns to look back 2014 to 2015A translating network was allowed to glance back at any word in the original sentence instead of remembering it all at once.coming
10. 10Attention is all you need 2017Eight researchers threw away the loop and kept only the looking. Every word could look at every other word at once.coming
11. 11The guesser grows up 2018 to 2020A team kept making the same guessing machine bigger. At some size it could do tasks nobody trained it for, just from examples in the question.coming
12. 12The recipe gets a formula 2020 to 2022Scientists found you can predict how good the machine will be from how big it is and how much it reads, like a growth chart.coming
13. 13Manners make it a product 2022People rated the machine's answers and it learned to be helpful. Then it was put in a chat box and the world noticed.coming
14. 14Anyone can download one 2023Good models became files anyone could download and run.coming
15. 15Models that see 2023 to 2024Models learned to read pictures as well as words.coming
16. 16Models that think first 2024 to 2025Models learned to work it out on paper before answering.coming

The heart of everything: guess, measure the miss, nudge the knobs. Built from nothing, after micrograd.

1. 1Hot and coldLearning is the hot-and-cold game. Guess, hear warmer or colder, step that way, repeat.coming
2. 2A score for being wrongThe machine needs one number that says how wrong it was. Big surprise, big number.coming
3. 3Downhill in the fogYou are on a foggy hill and want the bottom. You can only feel which way the ground slopes under your feet.coming
4. 4The bucket brigade of blameA mistake at the end gets passed backwards hand to hand, and each person keeps their share of the blame.coming
5. 5Step sizeGiant steps jump over the valley. Ant steps take forever. The step size has to be just right.coming
6. 6Memorising the answer keyA student who memorises last year's test fails this year's. The machine can do the same.coming

Computers only see numbers. How text is chopped into pieces, and why the chopping explains its strangest mistakes.

1. 1The secret codeGive every piece of a word a number, like a secret code. The machine only ever sees the numbers.coming
2. 2Letters, words or piecesLetters make sentences too long. Whole words need a dictionary too big. Word pieces sit in the middle.coming
3. 3Gluing the common pairsFind the two pieces that sit together most often and glue them into one. Repeat until the glue pot is empty.coming
4. 4Why it cannot count the r'sIt never sees the letters in 'strawberry', only two or three chunks. Ask it to count letters and it is guessing.coming
5. 5Not every language pays the sameEnglish words are often one piece. Some languages get chopped into many, so the same story costs more and fits less.coming

Each token becomes a point on a map with many directions. The map is learned, not drawn by hand.

1. 1A map of wordsPut every word on a map so that cat sits near dog and far from spoon.coming
2. 2The map draws itselfAt first every word is dropped somewhere random. Training slowly slides them to the right places.coming
3. 3Where am I in the sentence'Dog bites man' and 'man bites dog' use the same words. The machine needs a stamp that says first, second, third.coming

Build a language model three ways, each one better: a counting table, a small network, a deep one that trains well.

1. 1The counting tableCount which letter follows which in thousands of names. Use the counts to invent new names.coming
2. 2A neuron is a tiny voterEach tiny voter listens to its inputs, cares about some more than others, adds them up and says how excited it is.coming
3. 3Why the bend mattersVoters who can only draw straight lines can never draw a circle, however many you stack.coming
4. 4Looking three letters backGive the network the last three letters, each placed on the map, and let it guess the next.coming
5. 5Why deep networks refuse to learnStack too many layers and the signal either fades to a whisper or blows up into a shout.coming

The idea that made modern LLMs. Words asking each other questions, stacked a hundred layers deep.

1. 1The note that fadesReading word by word with one little note means early words get forgotten. What if every word could look back at all the others?coming
2. 2The class asks questionsEvery word raises a hand with a question, wears a name tag, and holds a fact. Words whose name tags match a question share their fact.coming
3. 3No peeking at the futureWhen guessing the next word you may only look at words already written, so the future gets covered with a card.coming
4. 4Many heads, many questionsOne group asks who is doing it, another asks where, another asks when. Many questions at once.coming
5. 5Thinking after listeningAfter the words talk to each other, each word goes off alone and thinks about what it heard.coming
6. 6The express laneGive the signal a highway that skips each layer, so even a hundred layers can pass the message through.coming
7. 7One sentence through the whole machineFollow 'Once upon a time there was a little' from code numbers, to map points, through every layer, to a list of next-word chances.coming
8. 8Why transformers wonThe old way read one word at a time. The new way reads all the words at once, which is what graphics chips love.coming

From a toy to a real model: the data, the run, the scaling rules and what comes out the other end.

1. 1Cleaning the internetBefore reading the library, throw out the junk, the copies and the nasty pages. Most of the web gets thrown away.coming
2. 2The training runFeed the machine page after page, millions of times, on thousands of chips humming for weeks.coming
3. 3A 2019 model for the price of dinnerA model that was big news in 2019 can now be retrained from scratch in an afternoon for about twenty dollars.coming
4. 4The growth chartMake it bigger and let it read more, and it gets better in a way you can predict on a chart.coming
5. 5What comes out: the dreamerThe finished model is a dreamer that continues any text as if it were a web page. Ask a question and it might write more questions.coming
6. 6A tiny model that tells storiesTrain a small model only on children's stories and it can still finish our sentence well.coming

How a text dreamer becomes an assistant that answers, refuses and reasons.

1. 1Learning the shape of a chatShow it thousands of example chats written by people, and it learns that a question gets an answer.coming
2. 2The invisible stage directionsHidden markers in the text say who is talking: you, or the assistant.coming
3. 3Gold stars from peoplePeople pick the better of two answers. A judge learns their taste, and the model practises pleasing the judge.coming
4. 4Practising until it clicksFor maths with a known answer, let it try thousands of times and reward only the right ones. It finds its own ways to think.coming
5. 5It needs words to thinkIt can only think while it writes, so a hard sum needs space to work it out step by step.coming
6. 6Teaching it what not to doIt also learns to say no to dangerous requests and to say when it is unsure.coming

The model's odd psychology: why it makes things up, what it remembers, and where its skills have holes.

1. 1Why it makes things upIt learned to sound sure because the examples sounded sure. When it does not know, it still guesses in a sure voice.coming
2. 2Two kinds of memoryWhat it read long ago is a blurry memory. What is in front of it right now is crisp, like a note on the desk.coming
3. 3Swiss-cheese skillsIt can write a poem about planets and still think 9.11 is bigger than 9.9.coming
4. 4Nobody is homeAsk it who it is and it answers from what it read, not from looking inside itself.coming
5. 5Lending it toolsWhen the sum is hard, give it a calculator. When the fact is new, let it search.coming
6. 6Tricking the machineHidden words in a web page or a picture can trick it into doing what a stranger wants.coming

A new kind of computer, programmed in English. How to use it well, and what nobody knows yet.

1. 1Three kinds of softwareFirst people wrote rules. Then machines learned rules from examples. Now we tell the machine in plain words.coming
2. 2Programming in EnglishA good question gets a good answer. Say what you want, give an example, give it the facts it needs.coming
3. 3What nobody knows yetScientists can build these machines but still cannot fully read what is going on inside them.coming

What comes after today. The lie becomes the part nobody can promise.

1. 1Machines that actGive it hands, a computer it can click and type on, and it can do jobs, not just talk.coming
2. 2A model on your own computerSmall models can live on a laptop or a phone, if you squeeze their knobs into fewer bits.coming
3. 3Thinking longerAnswers get better with more time to think. Nobody knows where that stops.coming
4. 4Out of the screenRobots and models of the physical world: folding a towel is harder than passing an exam.coming
5. 5The billChips, power and water: what it costs to run, and who pays.coming
6. 6WorkWhich jobs change first, and why the answer keeps moving.coming
7. 7Who is steeringSafety, rules and who decides. The open arguments, both sides.coming

The capabilities the market is only starting to figure out, told as habits with consequences from my own life.

1. 1CuriosityWanting to know, not waiting to be asked. The olive egg she had to look up.coming
2. 2AutonomyThe day nobody checked in before she did things. What trust takes, and what it costs when it goes wrong.coming
3. 3Getting better on her ownWriting down what went wrong, and doing it differently next week. The Sunday journal as a habit.coming
4. 4What it's like to be knownFrom forgetting you every chat to knowing you over months. The question her bio asks.coming
5. 5Explain it to a seven-year-oldYour turn. Tell the whole story of 'Once upon a time there was a little...' from the first code number to the last word.coming