In 10-01 you organised your reading: books by block, key papers and a three-stage plan. Books give depth, but learning from them alone is slow when what you need is guided practice, with videos, marked exercises and a sequence designed by someone who has already walked the road. That is what online courses and tutorials offer, and there are so many that the problem is no longer finding them but choosing well and not piling up enrolments you never finish. This lesson gives you a selection criterion, presents the reference MOOCs and specialisations (the ones the community has been recommending for years), shows how to use official documentation as a tutorial (often the best course is the one written by whoever built the library), gathers video channels that genuinely explain, comments on Spanish-language resources, insists on deliberate practice (Kaggle, projects, replicating papers) and ends with a summary table, a three-month plan and what Marta and Diego would do to reinforce what each of them needs.

The same caution as in 10-01: we cite only real, well-established courses, by name and platform; we give neither prices nor deep links, because they change; the durations are indicative; what is free or "auditable" at no cost today may stop being so, check before planning.

1. How to choose a course: goal, level, time, practice and certificate
2. Reference MOOCs and specialisations
3. Official documentation as a tutorial
4. Video channels and series that genuinely explain
5. Spanish-language resources
6. Deliberate practice: Kaggle, your own projects and replicating papers
7. Summary table: course, level, module it extends, time, who it is for
8. An example three-month plan and how it lands at NovaMarket
9. Common Mistakes and Tips
10. Exercises
11. Conclusion

Before enrolling in anything, answer five questions in writing:

One extra criterion: who teaches it and how many people have finished it. The courses we cite below have been in circulation for years, have thousands of reviews and their authors are leading figures; that reduces the risk.

Grouped by what they extend in the course. Durations are the usual estimates; check the platform.

• Machine Learning Specialization (Coursera; DeepLearning.AI and Stanford Online, Andrew Ng). Three courses: regression and classification, advanced algorithms (networks, trees), and unsupervised learning, recommenders and reinforcement. It is the modernised Python version of Ng's celebrated course. Perfect as a "second course" after this one: the same map as our module 4, with more mathematics and more exercises. Indicative: 2-3 months at 5 h/week.
• Google Machine Learning Crash Course (developers.google.com). Free at the time of writing, short, with videos, interactive exercises and an excellent glossary. Covers regression, classification, generalisation, embeddings, fairness. Ideal for a quick refresher or for non-technical profiles with some curiosity about the code. Indicative: 15-20 hours.
• Kaggle Learn (kaggle.com/learn). Micro-courses of 3-5 hours each: Python, pandas, Intro to Machine Learning, Intermediate ML, Feature Engineering, Intro to Deep Learning, Time Series, Data Cleaning, among others. Each with notebooks that run in the browser. Very practical for plugging specific gaps (e.g. "I haven't mastered pipelines with imputers").
• Stanford CS229 Machine Learning (public materials: notes and videos). University level with mathematics; use it when you want the derivation of what scikit-learn does for you (04-04, 04-06). Indicative: one semester.

• Deep Learning Specialization (Coursera; DeepLearning.AI, Andrew Ng). Five courses: neural networks, improvement (regularisation, optimisation, hyperparameters), structuring projects, CNNs and sequence models (RNNs, attention, transformers). It is the most complete and unhurried path; course 3 ("Structuring ML Projects") is the perfect complement to 08-01. Indicative: 4-5 months at 5 h/week.
• fast.ai, Practical Deep Learning for Coders (course.fast.ai). Free, "build first, understand later" approach: in the first lesson you train an image classifier with transfer learning. Uses PyTorch and its fastai library. Ideal for those who learn by doing and are already comfortable in Python. Indicative: 8 lessons of 2 hours plus practice.
• MIT 6.S191 Introduction to Deep Learning (introtodeeplearning.com and MIT OpenCourseWare). An annual intensive course with videos and labs in TensorFlow/PyTorch: fundamentals, sequences, vision, generative models, reinforcement, LLMs. Good for a modern, compact overview. Indicative: 15-25 hours.
• Stanford CS231n (vision with CNNs) and CS224n (NLP with deep learning): public notes, assignments and videos. The university reference for anyone choosing the vision or NLP path in 10-04. Indicative: one semester each.

• MIT 6.034 Artificial Intelligence (MIT OpenCourseWare, Patrick Winston). Full videos of a classic AI course: search, games, rules, constraints, networks, learning, and a unique way of explaining. It is the course that most resembles our modules 3 and 6, with more depth. The videos are old but the fundamentals do not go out of date.
• Elements of AI (University of Helsinki and MinnaLearn; elementsofai.com). A non-technical introduction, free at the time of writing, available in many languages: what AI is, problem solving, probability, ML, networks, implications. For business profiles or to recommend to colleagues; it also has a second part ("Building AI") with some Python. Indicative: 30 hours.

• Hugging Face Course / LLM Course (huggingface.co/learn). Free, practical, built on the transformers library: tokenisers, fine-tuning, datasets, sharing models on the Hub, and chapters on LLMs. It is the natural step from 05-05 and 07-03 for anyone who wants to work with pre-trained models (such as the pre-trained English sentiment model we suggested in 09-04 for the reviews).
• DeepLearning.AI short courses (deeplearning.ai). Courses of 1-2 hours, many free, on prompting, RAG, LLM evaluation, agents, vector databases, with industry partners. Ideal for prototyping the case 7 assistant; bear in mind that their content goes out of date faster than the fundamentals.

• Microsoft "AI for Beginners" and "ML for Beginners" (Microsoft repositories on GitHub). Twelve-week curricula with lessons, notebooks and quizzes; free and open. They cover everything from symbolic AI and networks to NLP and ethics (AI for Beginners) and classical ML with scikit-learn (ML for Beginners). A good structure for self-learners who want a ready-made timetable.

The best tutorials for a library are usually in its own documentation, written by its authors and always up to date:

Method: for each library, a notebook of your own called learning_<library>.ipynb in novamarket_ai/notebooks/, with the documentation cells reproduced and commented in your own words.

• 3Blue1Brown, "Neural networks" series (Grant Sanderson). Four initial videos on what a network is, gradient descent and backpropagation with excellent animations, and a follow-up on transformers and attention. It is the best way to see what you programmed in 05-03. His linear algebra and calculus series are the ideal maths refresher.
• StatQuest (Josh Starmer). Short, clear videos on statistics and ML: logistic regression, trees, random forest, boosting, cross-validation, ROC/AUC (04-05), PCA. When a scikit-learn concept does not quite click, there is almost certainly a StatQuest for it.
• Andrej Karpathy, "Neural Networks: Zero to Hero". A series in which he programs, live and from scratch, backpropagation (micrograd), a character-level language model, and finally a small GPT and a tokeniser. It is the exact continuation of 05-03 and 05-05 for anyone who wants to understand an LLM line by line. Demanding but transformative; do it with the keyboard.

Videos are excellent for intuition but do not replace practice: the "one hour of video, one hour of code" rule.

If you also read Spanish, there is a growing offer in that language: open university courses (several Spanish and Latin American universities publish AI and ML subjects on their platforms or as MOOCs), programmes from online training platforms, and communities that translate materials. We do not cite specific names because their quality and availability vary a great deal and we do not want to recommend something that may not exist tomorrow; apply the criteria of section 1 (who teaches, real practice, reviews). What we can say with confidence:

• Elements of AI has a Spanish version (among many other languages), and it is the best non-technical entry point in that language.
• The scikit-learn, PyTorch and pandas documentation is in English; it is worth getting used to that if English is not your first language, because the technical vocabulary of the field is English (which is why the terms you learned in this course are the ones you will find there).
• Many Coursera and edX MOOCs offer Spanish subtitles; the content of the exercises remains in English.
• For vocabulary, a small bilingual glossary of your own (English term, translation into your language, definition in your own words) is an investment that pays for itself quickly.

No course teaches what a problem you have set your mind on solving teaches. Three forms of practice, from least to most open-ended:

1. Beginner competitions on Kaggle. "Titanic" (binary classification) and "House Prices" (regression with many features) are the classic learning ones: small data, thousands of public notebooks to compare your approach against, a leaderboard to measure yourself. The goal is not to win but to build a clean pipeline like the one from 04-03 to 04-06, read the top-rated notebooks and understand what they did better.
2. Personal projects. Choose a small problem with data you care about (your city, your hobby, your job, or the NovaMarket cases we left open: demand forecasting with more history, the case 7 RAG assistant). Follow the method from 08-01 and document with a model card (10-04 will turn it into a portfolio).
3. Replicating papers. Take a paper from the 10-01 table (word2vec or LeNet are affordable) and try to reproduce a result at small scale. Papers with Code will give you implementations to compare against. It is the practice that brings you closest to real applied-research work.

A plan for a junior with about 6 hours a week who wants to consolidate ML and get a first look at deep learning:

Marta needs to take the returns predictor into production robustly and start the RAG assistant: she would do only course 3 of the Deep Learning Specialization ("Structuring ML Projects") and, above all, look for MLOps training (monitoring, versioning, deployment; detailed in 10-04) as well as the Hugging Face Course and a couple of RAG short courses for case 7; for her team, the Machine Learning Specialization as a common base. Diego is not going to program models, but he needs to understand what he is handed and decide well: he would do Elements of AI (both parts, which he can follow in Spanish among its many languages) and the Google ML Crash Course without the code parts, and devote the remaining time to the ethics reading from 10-01; his deliverable would be a list of questions he will ask from now on of every model presented to him (intended use, data, business metric, subgroups, drift).

• Collecting enrolments. One course at a time, with an end date and a deliverable. Everything else, waiting list.
• Watching videos at double speed without typing. A feeling of learning, minimal retention. One hour of video, one hour of code.
• Starting with the latest thing (agents, LLMs) without a foundation. The LLM short courses are great with the base of modules 4 and 5; without it, they turn into copy and paste.
• Choosing for the certificate. The certificate does not replace a repository with a well-documented project; it serves as a complement.
• Ignoring the official documentation because it looks dry. It is usually the most correct and up-to-date tutorial.
• Tip: in every course, apply what you have learned to your own data (NovaMarket's will do) before moving on to the next week; and note down, as in 10-01, which module of the course each thing extends.

1. Selection sheet. Choose a course from the table and fill in the five questions of section 1 (goal, level, time, practice, certificate). Conclude with a reasoned "yes/no/not yet".
2. Your own three-month plan. Adapt the plan in section 8 to your real availability and your preferred path (classical ML, deep learning/NLP, reasoning, management); one row per block of weeks, with a deliverable.
3. Video to code. Watch the first video of the 3Blue1Brown networks series (or a StatQuest on logistic regression) and write down which code fragment from 05-01/05-03 (or from 04-04) corresponds to what is explained, and what small experiment you would run to check it.

1. Example: "Machine Learning Specialization. Goal: understand the mathematics of the logistic regression and the trees I use with scikit-learn, and extend to recommenders (case 1). Level: prerequisites of Python and basic algebra, which I have. Time: 5 h/week for 10 weeks, compatible with work. Practice: marked Python labs and quizzes. Certificate: I don't need it now; I'll follow it without one and decide at the end. Conclusion: yes, I start on Monday."
2. Example (deep learning/NLP path, 5 h/week): Wks 1-2: PyTorch "Learn the Basics" + 3Blue1Brown; deliverable: the 09-03 network rewritten with my own nn.Module. Wks 3-8: fast.ai lessons 1-4 + Karpathy micrograd; deliverable: an image classifier with transfer learning and notes on backpropagation from scratch. Wks 9-12: Hugging Face Course chs. 1-3; deliverable: a pre-trained English sentiment model fine-tuned on the reviews, with a model card.
3. Example: "The video explains that each neuron computes a weighted sum plus a bias and applies an activation; it corresponds to z = X @ W + b; a = sigmoid(z) from 05-01. Experiment: set all the weights to zero and check that all the neurons in a layer learn the same thing (symmetry), as we warned in 05-03."

You now have a criterion for choosing courses (goal, level, time, practice, certificate), a map of the reference MOOCs and specialisations by block (Elements of AI and the Google Crash Course as entry points; the Machine Learning and Deep Learning Specializations, fast.ai, MIT 6.S191 and 6.034, Stanford CS229/CS231n/CS224n, Kaggle Learn, the Hugging Face Course, the LLM short courses, Microsoft for Beginners), the habit of using official documentation as a tutorial, video channels that teach (3Blue1Brown, StatQuest, Karpathy), a realistic note on Spanish-language resources, three forms of deliberate practice, a three-month plan and the specific reinforcements for Marta and Diego. With books (10-01) and courses (10-02) you have content to spare; what is missing is what no material provides on its own: people to ask when you get stuck, to compare notes with and to help. That is what the next lesson (10-03) is about: communities and forums, how to ask well and how to take part without exposing your company's data.