This is where it gets real. You're going to build things. Actual models. On actual data. That produce actual predictions. And some of those predictions will be wrong in ways that surprise you.
Geron's book is the best practical ML textbook ever written. I say this without qualification. The third edition is updated for 2023+ and covers the full classical ML pipeline with a clarity that no competitor matches. Every chapter builds on the last. Every concept has code. Every exercise teaches something the text doesn't. Do the exercises. I cannot stress this enough. The gap between "I read about random forests" and "I built a random forest, tuned its hyperparameters, and understood why it overfit" is the gap between knowledge and skill.
Here's what will happen in Chapter 2: Geron walks you through a complete ML project — the California housing dataset. You'll fetch data, explore it, clean it, engineer features, train models, evaluate them, and select the best one. This single chapter is worth more than many entire ML courses because it shows you the full lifecycle, not just the modeling step. In the real world, 80% of ML work is data preparation. Geron shows you that on page 40.
About classical ML vs deep learning: There's a narrative in the AI hype cycle that classical ML is "old" and deep learning is "new" and therefore better. This is wrong. Random forests, gradient boosting (XGBoost, LightGBM), and logistic regression solve the majority of real-world ML problems. They train in seconds, not hours. They're interpretable — you can explain why the model made a prediction. They need thousands of data points, not millions. They don't need GPUs. For tabular data (which is what most business problems involve), classical ML still beats deep learning in most benchmarks. This is not widely acknowledged because deep learning makes for better conference papers and Twitter threads.
When you get a job and someone asks you to build an ML system, the right answer 80% of the time will be XGBoost, not a neural network. The reason you study deep learning in Module 4 is for the 20% of cases where it's the right tool — images, text, audio, video. But for "predict which customers will churn" or "detect fraudulent transactions" or "forecast sales" — classical ML wins.
Building ML Powered Applications (Ameisen) is the most underappreciated book in this curriculum. It was written by an ML engineer at Stripe who actually shipped ML products. It answers the question Geron doesn't address: "How do I go from a working notebook to a working product?" The answer involves choosing the right ML approach (sometimes rule-based systems beat ML), building iterative prototypes, debugging models systematically, and deploying responsibly. This is the book that connects your ML learning to your Rails engineering experience. You'll read it and think "oh, this is just software engineering with a different artifact."
About Kaggle: Start competing. Not to win — you won't win, and that's fine. Compete to learn. The Kaggle community is the most generous knowledge-sharing community in tech. Every competition has public notebooks from top competitors explaining their approach. You learn more from studying a Grandmaster's solution than from any textbook chapter. Start with the "Getting Started" competitions (Titanic, House Prices), then move to real ones when you're comfortable.
Conclusion #
Module 3 transforms you from someone who understands ML to someone who does ML. The shift is significant. You'll make mistakes — overfitting, data leakage, choosing the wrong metric, forgetting to scale features. These mistakes are the learning. Every ML practitioner has made every one of them. The difference between a beginner and an expert is that the expert recognizes the mistake faster.
Predictions #
-
Chapter 2 of Geron will be the most satisfying coding experience you've had since building your first Rails app. There's something magical about watching a model learn from data for the first time.
-
You will overfit your first model. You'll get 99% training accuracy and 60% test accuracy, and you'll feel confused. Then you'll learn about cross-validation, and it will click.
-
You'll spend more time on data cleaning than on modeling. This will annoy you. Then you'll realize that's just how ML works. Data quality is everything.
-
Kaggle will be addictive. You'll enter one competition "just to try," and three weeks later you'll be reading papers about feature engineering at 2am.
-
Andrew Ng's Coursera ML Specialization will feel easy after Geron. That's by design — Geron goes deeper. But Ng's teaching style is so clear that hearing the same concepts explained differently will solidify your understanding.
-
By the end of this module, you'll have a new instinct: when someone describes a business problem, your brain will automatically start thinking "what would the features be? what would the target variable be? what kind of model would I try?" That instinct never goes away.