AI Olympiad Preparation and Careers in AI
The AI olympiads test theory and the skill to apply it in code under time pressure, and the field has more routes into it than any other. Here is how the olympiads work, how the masterclass helps, and where the routes lead.
The olympiads
The International Olympiad in Artificial Intelligence (IOAI) is the International Science Olympiad for the field, open to high school students and reached through the national selection of each country. It is young and growing fast. The first edition was held in Burgas, Bulgaria, from 9 to 15 August 2024, with 41 teams from 32 countries and territories; the second in Beijing in 2025, with teams from 61 countries; the third in Astana, Kazakhstan, in August 2026, with 108 countries and territories and 466 contestants. The fourth edition will be held in Singapore from 4 to 10 July 2027, organized by Nanyang Technological University and AI Singapore. By August 2026 more than 130 countries and territories were accredited, and the organizers gave financial support to 42 delegations that year, so a country without an olympiad tradition can take part.
The competition week of 2026 had an individual contest spread over two days, with six tasks in machine learning, computer vision, natural language processing, and multimodal processing, and a team challenge on AI for humanoid robots. The tasks are solved in code, in Python with the standard libraries, and the organizers publish the tasks of every past edition on their site. The syllabus, updated each year by the International Scientific Committee, sorts every topic into one of three categories: theory (how it works), practice (what it does and how to call it in code), or both. Topics not listed are excluded. The 2026 syllabus is the basis of the table below.
The way to the IOAI leads through the national olympiad of your own country. The organizers' list of countries and territories names the accredited body of each country, its website, and how the team is chosen, usually through a national AI olympiad with several rounds or through the national informatics olympiad. In Germany, the Bundeswettbewerb Künstliche Intelligenz is a project competition rather than an exam: teams register in the spring, submit a project idea in the summer and the finished project in September, and the best present them at a final in Tübingen in November; students from Austria and Switzerland may enter. Two further routes exist for everyone. The International Olympiad in Informatics (IOI), held since 1989, tests algorithms and programming under time pressure, and its national rounds are where most olympiad students first meet search, graphs, and dynamic programming; the 2026 edition took place in Uzbekistan and the 2027 edition will be in Germany. And Kaggle competitions are the open equivalent for anyone of any age: a dataset, a metric, a public leaderboard, and the shared notebooks of other entrants to learn from.
Which units cover which part of the syllabus
The IOAI syllabus is organized in four sections. The table maps each to the units of the masterclass and says what the masterclass does not do. The units teach the theory and check it with quizzes; they do not teach you to write code under time pressure, and some practice topics of the syllabus are beyond them. The right-hand column says what to add.
| Syllabus area | What the syllabus lists | Units | What to add |
|---|---|---|---|
| Programming fundamentals | Python, NumPy and pandas, Matplotlib, scikit-learn, PyTorch, training on CPU and GPU | Unit 1, Unit 4, Unit 5, Unit 7 | Practice only. The tools of the week of these units (Colab, Kaggle, the UCI repository, the PyTorch tutorials) are where this is learned. |
| Supervised learning | Linear and logistic regression, L1 and L2 regularization, k-nearest neighbours, decision trees, ensembles, support vector machines | Unit 4, Unit 5 | The units cover all of these except support vector machines and gradient boosting, which the scikit-learn user guide explains at the level the syllabus asks. |
| Unsupervised learning | k-means, principal component analysis, t-SNE and UMAP, clustering methods | Unit 4, Unit 10 | The units show t-SNE as a picture of an embedding space and go no further. Learn k-means and PCA from a free textbook (see the reading list) and run them in scikit-learn. |
| Data science fundamentals | Accuracy, precision, recall, and F1, under- and overfitting, hyperparameters, cross-validation, confusion matrices and ROC curves, feature engineering, data processing | Unit 4, Unit 5, Unit 7 | The theory is in the units with worked examples; the data handling is practised in a notebook. |
| Neural networks | The perceptron, gradient descent, backpropagation, activation functions, loss functions | Unit 1, Unit 6 | Fully covered, including a backward pass worked by hand. |
| Deep learning | Multi-layer perceptrons, embeddings, pooling, attention, transformers, autoencoders, SGD and minibatches, Adam, learning rates, dropout, weight decay, early stopping, initialization, batch normalization, fine-tuning | Unit 7, Unit 9, Unit 10, Unit 11, Unit 14 | Covered with the mathematics; autoencoders are named in Unit 16 only. Run each method once in PyTorch. |
| Computer vision | Convolutional layers, image classification, detection, segmentation, pretrained encoders, augmentation, generative models, self-supervised learning, vision-text encoders, diffusion models | Unit 9 | Convolution, classification, and transfer learning are worked in full; detection and segmentation in outline. Generative and diffusion models are beyond the units and are practice topics in the syllabus. |
| Natural language processing and audio | Text classification, pretrained text encoders, language modelling, encoder-decoder models, pretrained language models, audio models | Unit 3, Unit 10, Unit 11, Unit 13, Unit 14, Unit 15 | Text is covered from the naive Bayes classifier to instruction tuning. Audio is not in the units; the syllabus asks for the use of pretrained audio models, which is practice. |
Search and probabilistic reasoning (Units 2 and 3) are not in the IOAI syllabus, and they are in the table's neighbours: the informatics olympiads ask for search, and every method in the later units rests on the probability of Unit 3. Unit 8 and Unit 16 are not olympiad material; they are the units a candidate reads to know what the field argues about. Every unit shows its calculations step by step, and the Key Equations and Algorithms page collects the formulas you need most often. For the practice side, work through the tools of the week: a notebook a week is enough to turn each unit into code you have run yourself.
A study plan for a candidate
National rounds fall at different times of the year, so the plan is written for a school year that ends with the national final in the spring or early summer. Shift it to your own dates.
- Phase 1, September to December (about 16 weeks, 5 to 6 hours a week). Units 1 to 8 at one a week, with the quiz of each, and in the same week the tool of that unit in a notebook: the dot product and three gradient steps in NumPy in the first week, a regression with a train and test split in the fourth, the XOR network in the Playground and then in PyTorch in the sixth and seventh. Learn Python alongside if you do not know it; the PyTorch tutorials and Kaggle Learn assume very little. By December you have the theory of the first two syllabus sections and have run every method once.
- Phase 2, January to March (about 12 weeks, 6 to 8 hours a week). Units 9 to 14, one a week, with a Kaggle beginner competition running in the background: a tabular one first, then an image one once Unit 9 is done. In the same weeks, fill the gaps the table names from the scikit-learn user guide and a free textbook: k-means, PCA, support vector machines, gradient boosting, and a fine-tuned pretrained encoder for images and for text. Start past IOAI tasks at the midpoint of this phase, one every two weeks, untimed, with the official solutions read afterwards even when yours worked.
- Phase 3, April until the national final (8 to 10 hours a week). Past tasks under time, two a week, from the IOAI archive and from your national olympiad; a checklist of the syllabus with every topic marked theory, practice, or both, and a notebook of your own in which each method is a function you have written and tested; Units 15 and 16 for the reading. In the last two weeks, no new material, only timed tasks and sleep.
Once the theory is known, the result depends on fluency in the libraries, which comes only from writing code, and on the habit of looking at the data before modelling it, which every past task rewards. The clock decides the rest, which is why the last phase is timed.
The school years
Mathematics comes first, and it is more of the field than most students expect: linear algebra (vectors, matrices, and what a matrix does to a vector), calculus (the derivative, the chain rule, partial derivatives), and probability and statistics (conditional probability, expectation, variance, and the idea of a distribution) are the whole of Unit 1 and the whole of a first-year machine learning lecture. Take the hardest mathematics your school offers. A programming language, learned early in any form, pays off for the rest of a career, and Python is the one the field uses. Physics teaches the habit of modelling a situation with equations and checking the result against a measurement, which is what machine learning is. English is the language of the literature, the libraries, and the documentation, so reading it with ease is an advantage for everyone whose first language is another.
A first project teaches more than any course. Collect a dataset yourself (the temperatures on your balcony, the lengths of your commute, the photographs of two kinds of leaf from your street) and train a classifier or a regression on it, with a test set held out and the result reported honestly; or write a search for a game you play, with a heuristic you designed, and measure how many positions it looks at. A project with a question, a method, and a plain account of what did not work is worth more to a university than any certificate, including ours. Science fairs take such projects: the Regeneron International Science and Engineering Fair is reached through affiliated fairs in more than 60 countries, and most countries have a national fair of their own. Younger students, aged 15 and under, can take part in the International Junior Science Olympiad through their national selection.
Summer schools and camps in machine learning exist in most regions, and the ones worth having are those with lectures by active researchers and practicals in code. Two that take applications from students and publish their programme each year are the Eastern European Machine Learning Summer School, held in a different city of the region each summer (Cetinje, Montenegro, from 27 July to 1 August 2026), and the Deep Learning Indaba, the annual gathering of the African machine learning community, with local IndabaX events across the continent. Both are aimed at university students and early researchers; their pages say who may apply in a given year and what support is available. Be careful with paid summer programmes that sell a certificate and a university's name; a free notebook and a past IOAI task teach the same material.
The degree routes
Most AI researchers hold a first degree in computer science, mathematics, statistics, electrical engineering, or physics, followed by a master's degree, and those who stay in research add a doctorate. Degrees named artificial intelligence or data science now exist at many universities and are a good route where the department has a research group behind them. A mathematics-heavy first degree keeps every door open, because the linear algebra, probability, and optimization are the hard part to add later, and the libraries are the easy part. Cognitive science and linguistics suit those who care about what intelligence and language are; electrical engineering suits those who want to build the hardware or work on signals and robots; statistics suits those heading for medicine, economics, or public policy, where the questions are about causes and uncertainty. The master's is where the specialization happens, and it is common to move country for it.
Research experience during the first degree is what separates applications to doctoral programmes. Students at universities in the United States, and citizens or permanent residents studying elsewhere, can apply to the Research Experiences for Undergraduates of the National Science Foundation, which pay for a summer in a research group, including many in computer science and machine learning. In Europe, the European Laboratory for Learning and Intelligent Systems (ELLIS) runs a doctoral programme across its member sites and lists them; universities elsewhere have equivalents through their own departments and research councils. In every country the usual way in is the same: an email to a research group.
The email that gets an answer is short and specific. Read two of the group's papers, say which question in them interests you and why, and attach one thing you have done: the best is a reproduction of a result from a paper, in a notebook that runs. The units of the masterclass name several that fit on one laptop, such as the double descent curve of Unit 4, the degradation experiment behind residual networks in Unit 7, the Othello probe of Unit 16, or the scaling fit of Unit 13 on published numbers. State what you can do (Python, a statistics course, a competition result) and ask whether there is a project, a placement, or a position. Long general letters go unanswered; a student who has read the papers and reproduced one is rare, and groups need help.
The careers
The title "AI researcher" covers several different jobs, and in all of them the share of time spent on the mathematics is smaller, and the share spent on data, code, and writing larger, than the public picture suggests. The units in the right-hand column are the ones that matter most for each.
| Job | What the days hold | Units |
|---|---|---|
| Doctoral student or postdoctoral researcher in a university group | A research group studies one question for years, say whether a smaller model can match a larger one on a task, or what a network has learned about grammar. A week holds a few hours of mathematics on paper, two or three days of writing and running experiments, a day of reading, and a growing share of writing papers, reviews, and funding applications. Teaching comes with the position. | Unit 1, Unit 4, Unit 7, Unit 16 |
| Research scientist in an industrial laboratory | The laboratories of the large technology companies and of a few institutes do the same work at a larger scale, with the compute to train big models and the pressure to publish or to ship. The share of engineering is higher than at a university: a research scientist spends as much time making an experiment run reliably on hundreds of GPUs as on the idea behind it. | Unit 7, Unit 11, Unit 13, Unit 14 |
| Machine learning engineer in a product team | The most common job in the field. The model is a small part of it; the data pipeline, the evaluation, the monitoring after deployment, and the conversations with the people whose problem it is take most of the week. Mathematics appears when something goes wrong and the engineer has to say why. A typical split is half data and code, a quarter evaluation and testing, a quarter meetings and writing. | Unit 4, Unit 5, Unit 7, Unit 9 |
| Data scientist in a hospital, a bank, or a public agency | Here the data are the hard part: incomplete, biased by how they were collected, and governed by rules on privacy and fairness. The methods are mostly those of the Foundation (regression, trees, logistic regression, the confusion matrix), and the daily work is cleaning, modelling, and explaining a result to a clinician, a risk committee, or a minister. Readable models matter more than accurate ones, for the reasons Unit 5 gives. | Unit 3, Unit 4, Unit 5, Unit 8 |
| AI safety, evaluation, or policy researcher | A newer set of jobs at research institutes, laboratories, and government bodies: designing evaluations that show what a model can and cannot do, studying how models fail or deceive, and writing the assessments that regulation rests on. The work is a mixture of experiments, reading, and writing for people who are not researchers; the mathematics is that of Units 14 to 16, and the writing is at least half of the week. | Unit 8, Unit 14, Unit 15, Unit 16 |
| Researcher in a science that now uses these methods | Protein structure, weather forecasting, materials, astronomy, genomics, and economics all have groups whose main tool is a neural network. The scientist stays a biologist or a physicist and learns enough of the methods to use them well and to know when a result is an artefact. Unit 8 describes two such cases, and the methods a scientist needs most are those of Units 4 to 7 and 9. | Unit 4, Unit 6, Unit 7, Unit 9 |
| Teacher | Schools and universities need people who can teach these methods with the mathematics, and few can. A teacher with the content of the masterclass and a notebook can run a course that most schools cannot offer. The units were written to be taught from. | Unit 1, Unit 3, Unit 4, Unit 6 |
Two honest remarks belong here. Most people who complete a doctorate in the field go on to industry rather than to an academic position; the Stanford AI Index publishes the share each year, and it has been a large majority for a decade. A doctorate is therefore training for several of the jobs above, and an academic career is one outcome among them. And the job titles change every few years while the content of the work does not: the person called a machine learning engineer in 2026 does much of what a data scientist did in 2016 and a statistician did in 2006, with better tools.
The skills research groups and employers look for
- Python with the standard libraries. NumPy, pandas, Matplotlib, scikit-learn, and PyTorch. Fluency means writing a training loop from memory and knowing what every line does.
- Linear algebra and probability. Enough to read a paper's method section without skipping the equations, and to derive a gradient when a library does not provide one.
- Reading papers. Abstract, figures, the experiment table, and the paragraph where the caveats hide, in that order. Two papers a week is the habit of every researcher.
- Running experiments cleanly. A held-out test set that is touched once, seeds recorded, every run logged, and a baseline before the new idea. Most wrong results in the field come from skipping one of these.
- Writing. Papers, reports, documentation, and the plain summaries that decision makers need. The researcher who writes clearly and on time is remembered.
- Saying what a result does not show. A benchmark score measures performance on that benchmark, and a model that works on one dataset has been shown to work on one dataset. Stating this before someone else does is the mark of a scientist.
A short reading list
- The tools of the field. Our resources page lists the free notebooks, visualizers, datasets, and research listings that the weekly tools of the masterclass introduce, among them Colab, the TensorFlow Playground, the Transformer Explainer, and the Epoch AI data hub.
- Free textbooks. Simon Prince's Understanding Deep Learning and Sutton and Barto's Reinforcement Learning: An Introduction are free as PDF from their authors and are the references the masterclass uses for Units 6 to 11 and 12. Dive into Deep Learning gives every method with runnable code, and Mathematics for Machine Learning covers the linear algebra and probability of Unit 1 at university depth.
- The literature. The machine learning listing of the arXiv is where almost every paper appears first, free to read; a candidate who reads one abstract a day for a year knows the field's vocabulary.
- Professional societies. The Association for the Advancement of Artificial Intelligence runs one of the field's main conferences each year and offers student membership; the ACM Special Interest Group on Artificial Intelligence, founded in 1966, has a mailing list open to non-members; ELLIS lists the European research sites and their doctoral programme.
- Our blog. The topic page Studying artificial intelligence collects our articles on learning the subject at school, reading a research paper, and what an AI degree leads to.
Where the masterclass fits
The masterclass is a structured first survey of the field: sixteen units on search, probability, learning, neural networks, vision, language, reinforcement learning, and large language models, each with the mathematics and the published research behind it, with a quiz each week and a final exam. For an olympiad candidate it puts the theory of the whole syllabus in one place, in one notation, with the worked examples that the tasks assume; the Foundation units cover most of the first two sections of the syllabus and the Advanced units the rest. It is not a degree, and it does not teach programming. A school student gains a head start on the first year of any of the degrees above and a certificate with a verification ID for an application. A university student gains the breadth that a specialized degree leaves out, and anyone who completes it can open a paper on the field and follow it. The notebooks, the past tasks, and the first email to a research group are yours to add.