Learning paths
Routes through the course by goal and time available, with the labs to do along the way.
Pick one route and follow it in order. Each step is a lesson; the lab next to it is where to spend your hands-on time. Every lab lesson ends its visualization section with Try it prompts: make the prediction before touching a control.
On the website, any lesson can be switched to Quick read, which keeps the intuition, the lab, the summary, and the interview questions, and hides the math, implementation, engineering, and knowledge-check sections. The whole course takes about 8 hours this way instead of about 30. Use it for review, or for a first pass before reading in full.
1. Full course (14 weeks, 5 to 7 hours a week)
The complete sequence. Weeks 1 to 12 match the 12-week plan; weeks 13 and 14 add LLM inference, agents, and ML system design.
| Week | Lessons | Labs |
|---|---|---|
| 1 | Vectors, calculus, probability, Python, vocabulary, paradigms, workflow, overfitting | paradigm guide, fit explorer |
| 2 | Linear regression, regularization, logistic regression, metrics | linear regression, threshold |
| 3 | KNN, Naive Bayes, SVMs, decision trees | KNN, decision tree |
| 4 | Random forests and boosting | boosting |
| 5 | K-means, other clustering, PCA | K-means, PCA |
| 6 | Text to vectors, embeddings | TF-IDF |
| 7 | Forward pass, gradient descent, backpropagation | forward pass, gradient descent, backprop |
| 8 | Training, CNNs, RNNs, autoencoders and diffusion | convolution |
| 9 | Self-attention, transformer architecture | self-attention |
| 10 | Tokenization, decoding, adapting LLMs, fine-tuning in practice | tokenizer, decoding, LoRA |
| 11 | RAG, vector search, evaluation | RAG, vector search |
| 12 | Architecture, reliability, security, projects | architecture, prompt injection |
| 13 | LLM inference, serving LLMs, tool use and agents, reliable agents | KV cache, batching, agent loop |
| 14 | ML system design, experimentation and A/B testing | recommendation funnel, A/B test |
Extension: weeks 15 to 20 (the rest of the course)
The course now goes beyond the core 14 weeks. Every lesson below pairs a picture or lab with code you build from scratch. Take the modules your goal needs, in any order once the core is done.
| Week | Modules | Labs |
|---|---|---|
| 15 | Information theory, numerical stability, loss functions, PyTorch, debugging neural networks | |
| 16 | Module 24: Applied ML in practice: feature engineering, imbalanced data, tuning, time series, anomaly detection; monitoring and drift | |
| 17 | BERT, attention variants, mixture of experts, scaling laws, open-model architectures | attention masks, MoE routing |
| 18 | Module 28: Prompting and context engineering; agent memory, MCP, multi-agent systems | |
| 19 | Module 26: Reinforcement learning, MDPs to PPO; Module 27: Training at scale | Q-learning, ZeRO sharding |
| 20 | Module 25: Computer vision and multimodal; generative models (GANs, latent diffusion, flow matching); optional Module 29: Speech and audio; fairness and differential privacy | CLIP matching |
2. AI-engineering fast track (about 5 weeks)
For experienced software engineers who want to build LLM products and understand what's underneath. It skips most classical ML and keeps the ideas every AI engineer needs.
| Order | Lesson | Why it's on this path | Lab |
|---|---|---|---|
| 1 | The language of ML | loss vs metric, training vs inference | |
| 2 | Workflow, splits, and leakage | evaluation discipline applies to LLM systems too | |
| 3 | Overfitting and bias-variance | why held-out evaluation matters | fit explorer |
| 4 | Classification metrics | precision, recall, thresholds, calibration | threshold |
| 5 | Text to vectors | sparse retrieval (BM25's ancestor) | TF-IDF |
| 6 | Word embeddings | dense retrieval | |
| 7 | Forward pass | what a layer computes | forward pass |
| 8 | Gradient descent | how anything is trained | gradient descent |
| 9 | Self-attention | the core of every LLM | self-attention |
| 10 | Transformer architecture | where parameters, compute, and the KV cache live | |
| 11 | Tokenization and pretraining | tokens drive cost and context limits | tokenizer |
| 12 | Decoding | temperature and sampling | decoding |
| 13 | Adapting LLMs | prompting vs RAG vs LoRA | |
| 14 | Fine-tuning in practice | SFT, LoRA and QLoRA memory, preference tuning | LoRA |
| 15 | The RAG pipeline | the most common LLM architecture | RAG |
| 16 | Vector search | ANN indexes, hybrid search, filtering | vector search |
| 17 | Evaluating LLM systems | proving a change helped | |
| 18 | Tool use and agents | tool calling, the agent loop, MCP | agent loop |
| 19 | Reliable agents | guardrails, agent evals, cost | |
| 20 | LLM inference | KV cache, quantization, speculative decoding | KV cache |
| 21 | Serving LLMs | batching, latency SLOs, cost per token | batching |
| 22 | Production architecture | the system around the model | architecture |
| 23 | Reliability, cost, and observability | keeping it up and affordable | |
| 24 | Securing AI systems | prompt injection and tool safety | prompt injection |
Then build project 3 (RAG evaluation workbench) or project 5 (agent reliability sandbox) from the project ladder.
3. Interview prep (7 days per role)
Pick the interview sprint for your target role: ML Engineer, AI Engineer, Applied Scientist, or Data Scientist. Each day lists the lessons, labs, cheat sheets, and interview questions that role is tested on, and day 7 is a mock interview.
Whatever the role, also:
- Read the model-selection, classification metrics, formulas, and failure modes cheat sheets.
- Take every quiz without notes; reread only the lessons where you missed questions. The website's review queue tracks these for you.
- Redo the hand calculations: linear regression's gradient step, backpropagation's nine steps, attention's worked example, TF-IDF, and Gini. Check them in the Math Lab.
- Practice the "explain it, calculate it, implement it, diagnose it" test out loud for: bias-variance, precision vs recall, bagging vs boosting, Q/K/V, RAG vs fine-tuning, and prompt-injection controls.
- Drill the interview questions at the end of every lesson. On the website, the rapid-fire drill times you and brings back the cards you marked shaky.