The roadmap

One library, one order to read it in.

Six stages, front to back. Each stage says what it assumes you already know, so you can drop in at the right place instead of guessing. Nothing here is locked to the order: it is the route we would walk with you if you asked.

Start at the beginning Browse all 49 books

Stage 1

The ground floor

Every book after this one assumes you can read a matrix equation and run a Python script. Start here if either of those makes you pause.

Mathematical Foundations for Machine LearningThe linear algebra, probability, and calculus that the rest of the library quietly assumes.43 capsules·10h·beginner

Python for Data ScienceFreeEnough Python, NumPy, and pandas to run an experiment and read the result.38 capsules·8h·beginner

Also at this stage

Foundations for AI & MLOne on-ramp if you would rather take the maths and the code together, in a single book.Generative AI FundamentalsSee the whole field working first, then come back and build the pieces yourself.Git & GitHub MasterclassVersion control, so your experiments have a history you can walk back through.SQL MasterclassMost real data lives in a database. This is how you get it out.R MasterclassThe statistics route, if your work is analysis rather than model building.Data Structures & Algorithms in PythonData structures and algorithms, for interviews and for code that stays fast.The Modern Software DeveloperHow to ship real software with AI coding agents working beside you.

Stage 2

Learning from data

Before neural networks, the ideas they inherit: a loss to minimise, a model that fits, and a test set that keeps you from fooling yourself.

Machine Learning FundamentalsFreeRegression, classification, and the training loop, built by hand before any framework.37 capsules·9h·beginner reads best after Mathematical Foundations for Machine Learning and Python for Data Science

Build Decision Trees from ScratchGrow a tree by hand. The clearest model there is for watching a fit happen.23 capsules·5h·beginner reads best after Python for Data Science

Also at this stage

Machine Learning & Deep Learning MasteryThe wide sweep of machine learning and deep learning in one volume, if you prefer breadth first.

Stage 3

Neural networks

One neuron, then a layer, then backpropagation written out by hand. This is where the maths from stage 1 starts doing visible work.

Neural Networks from ScratchFreeFrom a single neuron to a working network in plain Python, backpropagation included.33 capsules·7h·beginner reads best after Machine Learning Fundamentals

Deep Learning FundamentalsThe same ideas at depth: initialisation, optimisers, regularisation, and why training fails.38 capsules·9h·beginner reads best after Neural Networks from Scratch

Also at this stage

Convolutional Neural Networks FundamentalsFilters, feature maps, and a classifier that sees. The bridge into vision.

Stage 4

Attention, and your own LLM

Read this stage in order. Build Large Language Models from Scratch opens on a 124-million-parameter model and takes for granted that you know what a transformer block, d_model, and a vocabulary are. The book before it is where you meet them.

Transformers: Theory, intuition, and Building from ScratchAttention, and the queries, keys, and values it runs on. Read this one before you build the LLM.20 capsules·5h·intermediate reads best after Deep Learning Fundamentals

Build Large Language Models (LLMs) from ScratchFreeThe GPT you write yourself: tokeniser, blocks, pretraining, fine-tuning.20 capsules·6h·intermediate reads best after Transformers: Theory, intuition, and Building from Scratch and Neural Networks from Scratch

LLM FinetuningTake a pretrained model and rewire it for your own task with LoRA and PEFT.40 capsules·9h·intermediate reads best after Build Large Language Models (LLMs) from Scratch

Also at this stage

Prompt EngineeringNo prerequisites at all. Useful on day one, and still useful at stage 5.RLHF from ScratchPreference data, reward models, and alignment by hand. Easier after the reinforcement learning book.

Stage 5

Ship something

A model in a notebook is not a product. Retrieval, agents, evaluation, and a serving stack that survives real users.

RAG in ProductionFreeRetrieval that holds up when real people ask real questions.42 capsules·10h·intermediate reads best after Transformers: Theory, intuition, and Building from Scratch

AI Agents BootcampTools, planning, and memory: agents from first principles rather than a tour of frameworks.47 capsules·11h·intermediate reads best after RAG in Production

LLM Production & DeploymentServing, quantisation, latency, and cost. The distance between a notebook and a product.44 capsules·10h·intermediate reads best after LLM Finetuning

Also at this stage

AI Context EngineeringTreat the context window as a system you design, not a box you fill.Memory in AI SystemsGive an assistant a memory that outlives the session.Pi vs Hermes vs Codex: Context Compaction and MemoryThree production harnesses compared on the same problem: what each one keeps in context, and what it drops when the window fills.Build a Mini ClawdBot from ScratchOne project end to end: a Slack assistant that runs around the clock.Claude Certified Architect (CCAF)Scenario-first preparation for the Anthropic CCAF exam.

Stage 6

Pick a frontier

From here the path forks. Each branch is a place to go deep, and none of them needs the other three.

Vision and multimodal

Models that look at pixels, and models that hold an image and a sentence at once.

Build a Vision Transformer (ViT) from ScratchThe Vision Transformer, patch by patch, in PyTorch.Build a Data-Efficient Image Transformer (DeiT) from ScratchTrain a ViT on ImageNet-1k alone, with distillation carrying the load.Computer Vision BootcampOpenCV pixels through detection and segmentation to a deployed model.Transformers for Vision & Multimodal LLMsEvery vision and multimodal transformer, built by hand.Build a NanoVLM from ScratchFuse an image encoder onto a language model and watch it describe what it sees.NLP & Computer Vision MasteryModern computer vision coded from scratch, in one sweep.

Reinforcement learning and robotics

Agents that act, get a reward, and act differently next time. Then the same idea on real hardware.

Reinforcement LearningQ-learning through policy gradients, every algorithm written out.Modern Robot LearningTeach a real robot arm to act by imitation.Vision-Language-Action & World Models for RoboticsVision-language-action models and world models, the current edge of robot learning.

Systems and scale

What happens when one GPU is not enough, and what the model is actually running on.

5D Parallelism for Large Model TrainingData, tensor, pipeline, context, and expert parallelism, then all five at once on a cluster.Build a Diffusion Language Model from ScratchGenerate text all at once instead of left to right.Build Kimi K3 from ScratchA frontier mixture-of-experts architecture rebuilt component by component in PyTorch, checked against the released checkpoint.Pretraining a Mini Kimi K3A pretraining run that finished, written while it ran, including the parts that went wrong.How to Host Kimi K3What it takes to serve the real 2.8 trillion parameter model, and what it costs.

Science and craft

Machine learning pointed at physics, and the writing that gets the result published.

Scientific Machine Learning (SciML)Physics-informed networks, Neural ODEs, and UDEs, in Julia.Writing Journal & Conference PapersFrom a blank page to a published paper, peer review included.

That is the whole library. If you are not sure where you sit on this path, start at stage 1 and skim: the first chapter of every book is open to read.

Open the library