How to use this path (and how to actually learn)

The path in one picture

This site is one long dependency chain, and the order is the point. Phase 1 (math) gives you the four languages every model is written in: linear algebra, calculus, probability and optimization. Phase 2 (tooling) makes you fluent in Python, NumPy and PyTorch so that code never blocks an idea. Phase 3 (classical ML) teaches the discipline that outlives any architecture: generalization, evaluation, regularization, and the fact that gradient-boosted trees still win on tabular data. Phase 4 (deep learning) builds neural networks from the perceptron to backpropagation and optimizers. Phase 5 (vision) is where deep learning first won; you learn convolution and the architectures that followed. Phase 6 (sequences) walks the road from tokenization and word embeddings through RNNs to the birth of attention. Phase 7 (transformers) is the architecture behind every modern language model, derived line by line. Phase 8 (LLMs and generative AI) covers training, fine-tuning, alignment, prompting, decoding, diffusion and multimodality. Phase 9 (RAG) and Phase 10 (agents) are the engineering of real systems built on those models. Phase 11 turns what you know into a portfolio and a career.

You can read any lesson in isolation, but each one assumes everything before it. When a lesson says "1.1 Vectors, dot products and norms", it is telling you exactly where a missing idea lives.

Three passes per lesson

Every lesson has the same anatomy: why it matters, objectives, notes (the core text, with math), videos, readings, practice, key points, pitfalls, a quiz and flashcards. Use them in three passes:

1. Intuition pass (about 20% of the time). Read "why" and the objectives, then watch the first one or two videos. Do not take notes yet; let the pictures form.
2. Math pass (about 40%). Read the notes slowly with pen and paper. Re-derive every equation yourself before reading the derivation. Stop at each worked example, cover the solution, and compute it. Add the key points to your own notebook in your own words.
3. Code pass (about 40%). Do the practice items in order: by hand, then NumPy or PyTorch, then a library. Only now read the remaining readings for depth.

Mark a lesson complete only when you score at least 80% on the quiz without looking and have reviewed its flashcards once. The site tracks this for you.

Intuition. Passing once through a video gives you familiarity, which feels like knowledge but is not. Only retrieval — producing the idea from memory — builds knowledge you can use.

Learn like a scientist

Three findings from cognitive science are worth more than any study hack. Active recall: testing yourself beats re-reading, even when the test is hard and you get things wrong. Spaced repetition: reviewing an item just before you would forget it makes the memory durable at a fraction of the cost of cramming. Interleaving: mixing topics (a math flashcard, then a transformer one) feels worse and works better than blocking. This site implements all three: quizzes for recall, the Review page that schedules flashcards in expanding intervals (1, 3, 7, 14, 30 days), and a flashcard queue that mixes phases. Add your own cards for anything you had to look up twice.

Two more habits: sleep after hard sessions (consolidation is real), and explain each lesson to someone — a friend, a rubber duck, or an AI tutor you instruct to find holes in your explanation.

Projects first, perfection later

Each phase ends with a project lesson, and 11.3 Portfolio project ladder lays out a ladder of fifteen portfolio projects. Do them even when you feel unready; the feeling never arrives on its own. A project forces you to decide what matters, surfaces the gaps that lectures hide, and produces something you can show. Aim for "working and honestly documented", not "state of the art".

Choosing your cadence

The full path is roughly 320 hours of focused work; the fast lane below is about half that. The Progress page computes an exact schedule from your weekly hours, but as a guide:

Protect two or three fixed sessions per week rather than hoping for free evenings. Consistency beats intensity: ten hours a week for eight months will take you much further than a heroic month followed by silence.

The fast lane

If you need to reach modern systems quickly, do these lessons fully and skim the rest: Phase 1: 1.1, 1.2, 1.4, 1.6, 1.7, 1.9, 1.10. Phase 2: 2.2, 2.4. Phase 3: 3.1, 3.2, 3.3, 3.5, 3.6, 3.9. Phase 4: 4.1, 4.4, 4.5, 4.6, 4.7, 4.8. Phase 5: 5.1, 5.3, 5.4. Phase 6: 6.1, 6.3, 6.7. Phase 7: 7.1 through 7.6, 7.9, 7.10. Phase 8: 8.1 through 8.7. Phase 9: 9.1 through 9.6, 9.8. Phase 10: 10.1 through 10.5, 10.7, 10.9, 10.10. Still do every project lesson. What you give up is mostly breadth in classical ML and vision, and the derivations that make later chapters feel inevitable rather than arbitrary — plan to come back for them.

Using the AI tutor well

Every lesson has AI-tutor buttons that open Claude or ChatGPT with a prompt built from the lesson's key points: Explain, Quiz me, Give me an exercise, Derive it. They are powerful if you follow four rules. Ask for derivations, counter-examples and quizzes, not for finished answers to practice problems. When it explains something, close the window and reproduce the explanation from memory. Treat every factual claim as a lead to verify against the readings, because models make confident mistakes. And use it freely for what it is excellent at: reading your error messages, explaining a line of code, and inventing fresh exercises at exactly your level.

Where this shows up later

• The map of the field and the vocabulary you need on day one — 0.2 The map of AI; your working lab — 0.3 Set up your lab.
• Reading papers and building a portfolio use the same three-pass discipline — 11.1 How to read (and reproduce) papers, 11.3 Portfolio project ladder.
• Staying current after the path ends — 11.5 Staying current.