Image: Saintfevrier (CC BY-SA 4.0)How to start learning AI in high school
The students who do well in artificial intelligence at university are rarely the ones who used the most tools at school. They are the ones who learned the mathematics that the tools are made of. If you are between 14 and 18 and want to work in this field, the order of priorities is mathematics, then a programming language, then a project of your own, and then, if you like competing, the olympiads. The tools come last and take the least time.
The mathematics that matters
Three pieces of mathematics carry almost everything in the field, and all three are taught in school or just after it.
- Vectors and matrices. A picture is a list of numbers, a sentence is a list of numbers, and a neural network multiplies lists of numbers by tables of numbers. The dot product, the matrix-vector product, and the length of a vector are the operations; linear algebra is the subject. If your school offers it, take it. If not, the free lectures by Gilbert Strang at MIT OpenCourseWare and the 3Blue1Brown video series cover what you need.
- Probability. Conditional probability, Bayes' rule, expectation, and variance. A classifier outputs a probability, a language model assigns a probability to every next word, and the base rate fallacy catches every beginner once. The article on the rare disease test is the first problem to be able to do in your head.
- Calculus. The derivative as a slope, partial derivatives, and the chain rule. Every model in the field is trained by computing the slope of its error and stepping downhill, and the chain rule is how the slope is computed through many layers. Integrals matter much less.
Statistics, if it is on offer, is a fourth. Most of machine learning is statistics that scaled up, and knowing what a mean, a confidence interval, and a test set are before someone tells you is an advantage.
A programming language
Python, because the field uses it. Learn the basics in whatever way suits you (a school subject, a free online tutorial, a book), then learn the NumPy library, which does the vector and matrix arithmetic. The point at which you can load a table of numbers, compute a mean and a dot product, and draw a plot is the point at which the mathematics stops being abstract. Google Colab lets you do all of this in a browser without installing anything, and the AI Masterclass uses it as the tool of its first unit.
Resist the temptation to start with a library that builds a neural network in three lines. It works, and you will learn nothing from it. The first network you train should be one whose weights you can see, such as the one in TensorFlow Playground, and the first one you write should be small enough to check by hand.
A first project
Pick a question you care about and collect the data yourself. The weather at your school for a term and whether the next day was wetter. The results of your chess games with the opening you played. Photographs of ten objects on your desk, fifty of each. Then fit the simplest model that could answer the question, a line or a decision tree, and measure how well it does on data it has not seen. A project that a student collected, cleaned, and evaluated honestly is worth more to a university than a copy of someone else's notebook with a famous dataset, and it teaches the one thing that cannot be taught any other way: most of the work is the data.
Science fairs reward exactly this. The Regeneron International Science and Engineering Fair and its national feeder competitions accept projects in computer science and in applied mathematics, and the judges ask the same questions a research supervisor would: what did you measure, what did you compare it with, and what would change your mind.
The olympiads
The International Olympiad in Artificial Intelligence has run since 2024, in Burgas in Bulgaria, then Beijing in 2025, Astana in 2026, and Singapore in 2027. Countries send teams through national selection rounds; Germany's is the Bundeswettbewerb Künstliche Intelligenz, and the IOAI site lists the participating countries and their organizers. The olympiad has a scientific round with tasks in machine learning, natural language processing, and computer vision, and a practical round, and publishes a syllabus that is a good map of what a strong school student can learn. The older International Olympiad in Informatics covers the algorithmic side, and Kaggle competitions are the open equivalent for anyone at any age.
Check the dates on the organizers' site rather than here; formats and deadlines change every year. The olympiad and careers page of the masterclass maps the IOAI syllabus to the units and gives a study plan over a school year.
What a university looks for
Admissions tutors for computer science, mathematics, and the newer AI degrees read thousands of applications that mention a passion for artificial intelligence. What stands out is evidence: a mathematics grade, an olympiad round reached, a project with a result and an honest account of what did not work, and a reason for the interest that is specific. "I trained a model to predict our canteen's queue and it was wrong on Fridays, and here is why" is a better sentence than any adjective.
A mathematics-heavy first degree keeps every door open. The field changes its tools every five years and its mathematics never, and a student who can read a paper's equations can learn the new tool in a week. Computer science, mathematics, statistics, and electrical engineering all lead into AI research; so do physics and economics, and cognitive science and linguistics for those drawn to the questions about minds and language.
A rough plan
At 14 to 15: the school mathematics, done well, and the first Python. At 15 to 16: linear algebra and probability beyond the syllabus, a first project, a first notebook. At 16 to 17: calculus, a first neural network by hand, the national olympiad round if you want it, and a structured survey of the field. At 17 to 18: a larger project, reading a first research paper with help, and the application. The hours per week are fewer than people expect, three to five, as long as they are regular.
The AI Masterclass is built for the structured survey in that plan: sixteen weekly units from the mathematics of Unit 1 to the language models of the last tier, each with a worked example, a quiz, and a free tool to try. The article on careers says where the road leads.


