# Course Roadmap: AI from the ground up

Mission: engineer-level understanding of neural networks, finishing with a model you trained yourself.

## Phase 1 — One neuron
| # | Lesson |
|---|--------|
| 1 | [The artificial neuron](lessons/0001-the-artificial-neuron.html) |
| 2 | [The neuron as decision-maker](lessons/0002-the-neuron-as-decision-maker.html) |

## Phase 2 — Many neurons → a network
| # | Lesson |
|---|--------|
| 3 | [A layer of neurons (vectors)](lessons/0003-a-layer-of-neurons.html) |
| 4 | [Stacking layers — the forward pass](lessons/0004-stacking-layers-the-forward-pass.html) |
| 5 | [Activation functions (ReLU & friends)](lessons/0005-activation-functions.html) |

## Phase 3 — How a network learns
| # | Lesson |
|---|--------|
| 6 | [Loss — measuring "how wrong"](lessons/0006-loss-measuring-wrong.html) |
| 7 | [Gradient descent](lessons/0007-gradient-descent.html) |
| 8 | [Backpropagation](lessons/0008-backpropagation.html) |
| 9 | [The training loop](lessons/0009-the-training-loop.html) |

## Phase 4 — Build your own
| # | Lesson |
|---|--------|
| 10 | [Train a tiny network from scratch (pure Python)](lessons/0010-train-a-tiny-network-from-scratch.html) |
| 11 | [The same net in PyTorch (Colab)](lessons/0011-the-same-net-in-pytorch.html) |
| 12 | [Real data — handwritten digits](lessons/0012-real-data-handwritten-digits.html) |

## Phase 5 — Under the hood of GPT
| # | Lesson |
|---|--------|
| 13 | [Text becomes numbers: tokens](lessons/0013-text-becomes-numbers-tokens.html) |
| 14 | [Your first language model](lessons/0014-your-first-language-model.html) |
| 15 | [Embeddings: words as vectors](lessons/0015-embeddings-words-as-vectors.html) |
| 16 | [Attention: every token looks back](lessons/0016-attention-every-token-looks-back.html) |
| 17 | [The transformer](lessons/0017-the-transformer.html) |
| 18 | [How GPT is trained](lessons/0018-how-gpt-is-trained.html) |
| 19 | [The practitioner's toolbox](lessons/0019-the-practitioners-toolbox.html) |

## After this course (separate, future)
Multimodal models, retrieval (RAG) and tool use, agents, post-RLHF training techniques. The Phase-5 foundation makes any of them learnable without hand-waving.