Module 9 ended with the case 9 prototype documented, presented and scored against its self-assessment rubric; that closed the practical part of the course. This final module adds no new theory or code: it gives you the means to keep learning on your own with good judgement. We start with the most durable source of all, books and articles. In a field that changes every six months, good books are what goes out of date least (the fundamentals of modules 3, 4, 5 and 6 have been the same for decades) and the original papers are the way to understand where every idea you have used came from, from the perceptron of 05-01 to the transformer of 05-05 or the model card of 08-01. This lesson first explains how to read technical material (reading an ML book like a novel gets you nowhere), then presents the reference works grouped by block, with their level, the module of the course they extend and who they are for, and ends with a list of key papers, a method for reading a paper and a three-stage reading plan. To bring it down to earth, what Marta and Diego at NovaMarket would read first.

A word of caution, as a matter of rigour: we cite only real, well-established works, with their authors; we include neither deep links nor prices because they go stale. Several are free online by their authors' decision at the time of writing; that may change, so always check the author's or publisher's website.

1. How to read technical material: active reading, reproduced code and a reading notebook
2. General AI references
3. Practical machine learning
4. Deep learning
5. Probability, statistics and reasoning
6. Logic, expert systems and Prolog
7. Ethics and society
8. Projects, systems and MLOps
9. Key papers and how to read a paper
10. A three-stage reading plan and how it lands at NovaMarket
11. Common Mistakes and Tips
12. Exercises
13. Conclusion

You do not read a technical book; you work through it. Three habits make the difference between "I've read it" and "I can use it":

• Active reading. Before each chapter, write in one line what you expect to learn; when you finish, write in three lines what you have learned without looking at the book. If you cannot, you have not understood it. Ask questions in the margin ("why L2 regularisation here and not L1?") and look for the answer before moving on.
• Reproduce the code. Type the examples in (do not copy and paste), run them, break them on purpose (change a hyperparameter, drop a column, insert a null value) and watch what happens. Then apply them to your data: throughout the course you always had orders_nm.csv or reviews_nm.py; keep using them as a test bench for everything you read.
• A reading notebook. One file per book (Markdown in your novamarket_ai/notes/ repository folder is fine) with: the main idea of each chapter, a key formula or code fragment, a connection with something you already know ("this is the cross-validation from 04-05, but nested") and an open doubt. Rereading your notes is ten times faster than rereading the book.

Two more pieces of advice: do not read reference works from cover to cover (Russell & Norvig runs to over a thousand pages; you consult it by chapter), and alternate a "dense" book with a "light" one so you do not give up. As an optional aid, this small Python log lets you track what you have read and its connections with the course:

Each entry stores the date, book, chapter, idea, the course module it extends and a doubt; the JSON is readable, versionable and easy to turn into a pandas table whenever you want to review what you have covered.

These are the books that cover the complete map this course drew (agents, search, logic, probability, learning). You do not read them whole; you consult them.

If you have to pick just one and you are starting out, Poole & Mackworth for accessibility and for the code; AIMA when you want the definitive reference.

These extend module 4 (and parts of 5 and 9): data preparation, algorithms, evaluation, overfitting, and the step up to real projects.

Suggested order for a junior: Burkov (map) → Géron (practice) → ISL (statistical foundation); Müller & Guido as the library reference.

These extend module 5 and 09-03: from the neuron to CNNs, RNNs and transformers.

These extend 06-03 (Bayes and Bayesian networks) and the probabilistic foundation of all ML.

These extend module 6 (06-01, 06-02, 06-04) and the rules engine of case 8.

These extend 02-04 (ethics and the AI Act), 08-02 (bias cases) and 08-03 (regulation).

These extend 08-01 (method, model card, drift), 07-04 (environments) and what in 10-04 we will call the MLOps path.

The original papers are the primary source of everything you have seen. You do not need to read them all, or in full, but knowing ten of them places you in the history of the field (01-01) and teaches you to read the ones still to come.

A scientific paper has a fixed structure: abstract, introduction (problem and contribution), related work, method, experiments, conclusions and references. Read it in three passes:

1. Pass 1 (10 min): abstract, figures and tables, conclusion. Decide whether it is worth continuing. Note in one sentence what it claims.
2. Pass 2 (30-60 min): introduction and experiments at a calm pace; skip the proofs. Check what they compare against (is the baseline fair, as we demanded in 09-04?) and on which data.
3. Pass 3 (only if you are going to use it): the full method, and try to reproduce a part of it, even at small scale.

Where to find them: arXiv (arxiv.org) publishes open preprints of almost everything current in AI; Papers with Code (paperswithcode.com) links each paper to implementations and results tables; Google Scholar is useful for following citations. The older papers (Turing, Rosenblatt) are usually accessible in university repositories.

Suggested reading order: Turing (historical and accessible), Model Cards (short and directly useful), Vaswani (the most cited today; read it with 05-05 alongside), and then the rest according to your path.

Marta (head of data, with the returns predictor in production and the RAG assistant still to build) would first read ML Yearning (project decisions in two weeks), then Huyen for monitoring the predictor and, in parallel, Géron so that her team shares a vocabulary; Vaswani and BERT when the assistant gets under way. Diego (operations, non-technical) would start with Mitchell (what AI can and cannot do), continue with O'Neil (the risks of the predictor he uses every day) and read the abstract and figures of the Model Cards paper to understand the document Marta hands him with every model.

• Buying (or downloading) five books and finishing none. Pick one per stage; the rest go on the waiting list.
• Reading without a keyboard. If the book comes with code and you do not run it, you retain a fraction. Save passive reading for the popular-science titles.
• Starting with Goodfellow or Bishop. They are advanced references; starting there is discouraging. Order matters: Burkov/Géron/Chollet first.
• Confusing free with permanent. What is open today may stop being so; keep your notes, do not depend on the PDF.
• Reading papers as if they were revealed truth. A paper is a claim with evidence; check the baseline, the data and whether anyone has reproduced it (Papers with Code helps).
• Tip: follow a "one dense book + one light book" rule and "one paper a month"; always note the connection with a module of the course; that is what turns reading into knowledge.

1. Choose and justify. Pick one book from the tables for your current situation (level, time available, goal) and write a paragraph justifying the choice on three criteria: which module of the course it extends, why that level, and what you will do with it (code to reproduce, data to apply it to).
2. Personal reading plan. Adapt the three-stage table to three real months of your calendar: set weeks, chapters and one deliverable per stage (for example, a reproduced notebook, a summary or an entry in the JSON notebook from section 1).
3. First pass of a paper. Choose one from the papers table (recommended: Model Cards or Vaswani 2017) and do pass 1: write in five lines what problem it solves, what the contribution is, what the main figure or table shows and how it connects with the course.

1. Example: "I choose Géron because it extends the whole of module 4 and part of 5, which is what I will use most as a junior; its introductory-intermediate level matches what I already know of scikit-learn; I will reproduce the notebooks of the classification and model-tuning chapters and apply them to NovaMarket's returns predictor to try to beat the AUC of 0.844 with a better-documented pipeline."
2. Example: Weeks 1-3: all of Burkov, one notebook entry per chapter. Weeks 4-6: ML Yearning, deliverable: a one-page document with the data split and metric of case 4 revised. Weeks 7-12: Géron chs. 2-4 and 6-7 with reproduced notebooks; deliverable: my own notebook with nested validation on orders_nm.csv.
3. Example (Model Cards, 2019): Problem: models are published without explaining what they are for or how they fail. Contribution: a standard template (model details, intended use, factors, metrics, evaluation data, ethical considerations). Main figure: examples of model cards with metrics broken down by subgroup. Connection: it is the template we followed in 08-01 and 09-04; the breakdown by subgroup is what was missing from the returns predictor's model card.

You have learned to read technical material actively (expect, reproduce, note), you have a map of books by block with their level and the module they extend (general references with AIMA and Poole & Mackworth; practical ML with Burkov, Géron, Müller & Guido and ISL; deep learning with Chollet, Prince and Goodfellow; probability with Bishop, Koller & Friedman and Pearl; expert systems with Giarratano & Riley and Bratko; ethics with O'Neil, Crawford and Mitchell; projects with Huyen and Ng), a list of twelve papers that explain where every idea in the course comes from and a three-pass method for reading any paper, plus a three-stage plan and the first readings of Marta and Diego. Books give depth and permanence; but learning from books alone is slow when what you need is guided practice and correction. In the next lesson (10-02) we will see how to choose online courses and tutorials, which are the reference ones for each module of the course, and how to put together a three-month plan that combines video, official documentation and deliberate practice.