No prior programming or math background is required. The video content is designed to build intuition through examples rather than equations. The four hands-on labs use Python and pandas, but the environment is pre-configured — you won't need to install anything — and every lab includes step-by-step hints and full solutions so you can follow along regardless of your coding experience.

The course includes four free Python labs, each paired directly with the video content from that week.

In Lab 1, you take a raw file of 300 house sales and transform it into a training-ready dataset — finding the label, removing a missing-price row and a duplicate, converting a year column into an age, and one-hot encoding a neighborhood field.

In Lab 2, you take on the role of the training algorithm: you manually adjust the two knobs of a simple model (weight and bias) and watch the loss score respond in real time, then hand the job to gradient descent and observe what happens when the learning rate is set too high.

In Lab 3, you split data into train, validation, and test sets, train an unconstrained decision tree that aces training but fails validation, fix it with a depth cap, and then catch a dataset that scores suspiciously well because the answer is leaking through its inputs.

Lab 4 is the capstone: you train a real neural network with 50,000 parameters to recognise handwritten digits, watch the loss fall across epochs, evaluate it with precision and recall per digit, and examine an actual mistake the model made — a sloppy four it misclassified as a nine.

The same training loop you learn in week one — weights, loss, and gradient descent applied to a housing dataset — is the foundation used to build large language models. The final module shows how that loop scales to tokens and embeddings, how pre-training produces a base model, and how instruction fine-tuning and RLHF turn it into a useful assistant. By the end, you'll have a clear mental model of how tools like ChatGPT are actually constructed.