← index

AI Mastery Path

Machine Learning

The capstone — where the math and the code become intelligence. It assumes the foundations from Mathematics for AI and Computer Science for AI, then builds from the learning problem up through deep learning, Transformers, and large language models. Start it once you have linear algebra, calculus, probability, optimization, and the Python stack — you don't need to finish everything else first.

Curriculum · 11 areas, 67 lessons

Part I · The learning problem

What it actually means for a machine to learn — the framing every method below is an answer to.

• 1What it means for a machine to learnsoon
• 2Supervised, unsupervised, and reinforcement learningsoon
• 3Features, labels, and the datasetsoon
• 4The hypothesis space and generalizationsoon
• 5Training, validation, and test setssoon
• 6Loss functions and the learning objectivesoon

Part II · Classical supervised learning

The workhorse models — still the right tool for most problems, and the foundation for everything deeper.

• 7Linear regression, revisited as machine learningsoon
• 8Logistic regression and classificationsoon
• 9k-nearest neighborssoon
• 10Decision treessoon
• 11Naive Bayessoon
• 12Support vector machinessoon
• 13Regularization in practice — ridge and lassosoon

Part III · Ensembles & unsupervised learning

Combining weak models into strong ones, and finding structure when there are no labels.

• 14Bagging and random forestssoon
• 15Boosting — gradient boosting and XGBoostsoon
• 16k-means clusteringsoon
• 17Hierarchical and density-based clusteringsoon
• 18Principal component analysissoon
• 19t-SNE and UMAP — seeing high-dimensional datasoon

Part IV · Evaluation & the practice of ML

The discipline that separates working models from fooling yourself — measuring, tuning, and avoiding traps.

• 20Overfitting, underfitting, and the bias–variance tradeoffsoon
• 21Cross-validationsoon
• 22Metrics — accuracy, precision/recall, ROC and AUCsoon
• 23Hyperparameter tuningsoon
• 24Feature engineering and selectionsoon
• 25Data leakage and other ways to fool yourselfsoon

Part V · Neural networks

From a single neuron to a trainable network — where the calculus and linear algebra of the foundations pay off.

• 26The perceptron and the artificial neuronsoon
• 27Multilayer perceptronssoon
• 28Activation functionssoon
• 29Backpropagation, in practicesoon
• 30Optimizers — SGD, momentum, and Adamsoon
• 31Initialization, normalization, and dropoutsoon

Part VI · Deep learning architectures

The specialized network designs that cracked vision and sequences, and the transfer-learning era they led to.

• 32Convolutional neural networks — visionsoon
• 33Recurrent networks and LSTMs — sequencessoon
• 34Embeddings and representation learningsoon
• 35Autoencoderssoon
• 36Training deep networks — tricks of the tradesoon
• 37Transfer learning and fine-tuningsoon

Part VII · Attention & Transformers

The architecture behind everything modern — built up from the problem it was invented to solve.

• 38The sequence-to-sequence problemsoon
• 39Attention — letting the model focussoon
• 40Self-attention and the Transformersoon
• 41Positional encodingssoon
• 42Multi-head attention and the full architecturesoon
• 43Why Transformers scalesoon

Part VIII · Large language models

How next-token prediction at scale became general-purpose intelligence — pretraining, alignment, and use.

• 44Language modeling — predicting the next tokensoon
• 45Tokenization and embeddingssoon
• 46Pretraining at scalesoon
• 47Scaling lawssoon
• 48Fine-tuning and instruction tuningsoon
• 49RLHF and alignmentsoon
• 50Prompting and in-context learningsoon
• 51Retrieval-augmented generation (RAG)soon

Part IX · Generative models

Models that don't just classify but create — the math of generating images, audio, and more.

• 52Generative vs. discriminative modelssoon
• 53Variational autoencoders (VAEs)soon
• 54Generative adversarial networks (GANs)soon
• 55Diffusion modelssoon
• 56Autoregressive generationsoon

Part X · Reinforcement learning

Learning from interaction and reward — the paradigm behind game-playing agents and model alignment.

• 57Markov decision processessoon
• 58Value functions and Q-learningsoon
• 59Policy gradientssoon
• 60Deep reinforcement learningsoon
• 61RL from human feedback, revisitedsoon

Part XI · The frontier & responsible AI

Where the field is heading, and the judgment needed to build AI that's safe, fair, and actually understood.

• 62AI agents and tool usesoon
• 63Multimodal modelssoon
• 64Interpretability and mechanistic understandingsoon
• 65Evaluation and benchmarkssoon
• 66Safety, alignment, and biassoon
• 67Where AI is goingsoon