A place for every connection.
Explore the full draft curriculum, find a definition, or return to the next connection in your learning.
Core knowledge checks track personal study. Formal mathematics extensions and the A01 MoE preview are uncredited practice. Project artifacts need separate self or peer review.
What is AI? Start with an everyday task
Your phone finds "dog" in your photos, and your inbox hides junk mail. What do these helpers have in common, and what is really going on inside them?
Not yet demonstrated M00.2 · commonInputs and outputs: what goes in and what comes out
Every AI helper takes something in and hands something back. Follow a button photo through a sorter, then spot what goes in and what comes out of the apps you use.
Not yet demonstrated M00.3 · commonData and examples: what a computer can learn from
Photo search learned what a dog looks like from pictures. Build a tiny collection of your own and find out what makes examples helpful, or quietly misleading.
Not yet demonstrated M00.4 · commonWhat is a model? A small rule inside a bigger app
When a chat assistant answers you, the model is only one of the pieces at work. Meet models through a snack-planning rule, then find the other pieces around it.
Not yet demonstrated M00.5 · commonWeights: the adjustable numbers in a model
Big AI models are built from billions of adjustable numbers called weights. Turn one yourself on a snack model and see exactly what it changes.
Not yet demonstrated M00.6 · commonTraining and inference: changing a rule or using it
Does a chatbot learn from you while you talk? Use the snack model to see the difference between changing a model and simply using it.
Not yet demonstrated M00.7 · commonFollow one small AI system from start to finish
"Arriving in 18 minutes." Build your own delivery-time estimator from scratch: collect trips, train it, test it on trips it has never seen, then put it to work.
Not yet demonstrated M00.8 · commonDifferent ways to solve the same problem
How does a map app pick your route? Solve one tall-truck delivery three ways, with written rules, a search through choices, and estimates learned from past trips, then combine them.
Not yet demonstrated M00.9 · commonHow to tell whether an AI answer is useful
A chatbot tells you, with total confidence, that the museum opens at nine. It opens at ten. Learn four questions that separate a useful AI answer from a convincing one.
Not yet demonstrated M00.10 · commonHow to use this school without getting lost
Rereading feels like learning, but often isn't. Find your way around the course, and pick up the study habits that make ideas stick.
Not yet demonstrated M00.11 · commonPractice, progress, and knowing what you understand
A wrong answer can teach you more than a lucky right one. Learn how practice checks work here, what "mastered" really means, and how to turn mistakes into your next step.
Not yet demonstrated M01.1 · learningAI grew from several older questions
Long before anyone said "artificial intelligence," people had worked out how to reason with rules, weigh odds, and build machines that correct themselves. Meet the old ideas every AI system still borrows.
Not yet demonstrated M01.2 · learningNeurons, learning rules, and the Dartmouth proposal
Two scientists build logic from imaginary neurons, a codebreaker invents a guessing game for machines, and four researchers ask for one summer to crack intelligence. What did each really show?
Not yet demonstrated M01.3 · learningSymbols, games, and early language programs
A wobbly robot named Shakey plans its way from room to room, and a tiny chatbot convinces people it understands them. See what symbols can do, and where they run out.
Not yet demonstrated M01.4 · learningLimits, expectations, and changing support
In 1958 a newspaper said a learning machine might one day be conscious. By 1973 Britain was cutting AI funding. Find out what the famous XOR puzzle really proved.
Not yet demonstrated M01.5 · learningFrom knowledge engineering to learning from data
In the 1980s, hand-written rules saved one computer maker millions a year, then the boom went bust. See why "let the data write the rules" took over, and what it still costs.
Not yet demonstrated M01.6 · learningWhy deeper networks became practical
Neural networks were written off, twice. Then in 2012 one won an image contest by a landslide. Meet the four ingredients that had to arrive together.
Not yet demonstrated M01.7 · learningGames, attention, and broadly reusable models
In 2016 a Go program played a move no expert would choose. Six years later ChatGPT reached a million users in five days. Trace how one pretrained model came to power thousands of products.
Not yet demonstrated M01.8 · commonThe questions that shaped AI
From a 1950 guessing game to ChatGPT, AI has swung between bold bets, bitter winters, and sudden breakthroughs. Get the whole story, and a map of what you'll build next.
Not yet demonstrated M02.1 · learningHow a computer stores a picture or a sentence
The same eight on/off switches can be a number, a letter, or a dot of color. Learn how a computer knows which, and why "café" sometimes turns into "café".
Not yet demonstrated M02.2 · learningDescribe a computation so someone else can follow it
"Count the heavy parcels" sounds simple until someone asks whether 3 lb counts. Trace a real loop step by step, run it, and find the one test that catches the bug.
Not yet demonstrated M02.3 · learningWhat a representation makes easy to learn
A horizontal stripe and a vertical stripe look identical to a model that only sees their average brightness. See how the inputs you choose decide what a model can ever learn.
Not yet demonstrated M02.4 · learningTrace an example through a batch
"Size 3 is different from 2." Learn to read the shape of every array in a model, predict each result before running it, and catch the bug that no error message will flag.
Not yet demonstrated M02.5 · learningWhy a small choice can create a huge search
Twenty left-or-right choices already make over a million routes. Count a search tree by hand and in code, and see why a computer 1,000 times faster barely helps.
Not yet demonstrated