AI

34 boards, 54 lessons, 442 review questions and 27 narrated videos, in 8 chapters. Read them in order. Each chapter builds on the one before.

Your progress0 of 54 read

Start here → Vectors, Dot Products and Matrix Multiply

WatchGPT-6 Astra, examined4 AI, end to end22 All 27 videos →

Contents

Part one · Foundations

1Neurons to a language model9 lessons
The maths floor, then one neuron's weighted sum, how it learns and why training stalls, words as numbers, memory across a sentence, attention, and the transformer grown into an LLM.
▶ Watch the video
What a Neural Network Is How It Learns Tokens and Embeddings Attention and the Transformer From Transformer to LLM
2From a model to an app7 lessons
Softmax and sampling, then fine-tuning and a LoRA run, retrieval, tools and agents, ending in the anatomy of a real LLM app.
after Neurons to a language model▶ Watch the video
Inference Fine-tuning and Alignment RAG Agents and Tool Use Capstone Anatomy of an LLM App
3Past the basics12 lessons
Why a model is fast and cheap, why thinking longer works, how you know it is good, what is inside it, networks that see, images, vision, mixture of experts, the GPU and training across many of them, and serving as a systems problem.
after From a model to an app▶ Watch the video
Inference at Scale Reasoning and Test-Time Compute How Do You Know It's Good Sculpting From Noise Teaching It To See Many Experts One Router Inference Engineering

Part two · Build

4The string and the contract4 lessons
The returns desk, from the one string the model sees through the prompt, the window, and the tool schema that is its contract with your code.
after Neurons to a language model
What the Model Sees Prompt Engineering Context Engineering Structured Output and Tool Schemas
5The loop and its harness8 lessons
The same desk as a loop: the harness around it, memory, a state machine, a second agent, the eval table, the three budgets that close it, and the red team that attacks it.
after The string and the contract
The Loop The Harness Memory Graphs and State Machines Multi-Agent Evals for Apps Guardrails Cost and Latency
6GPT, built from scratch6 lessons
A language model written from nothing on a laptop: numbers that find their own slopes, a model that counts, one that learns, attention, the GPT block, and sampling from it.
after Neurons to a language model
7On a real model2 lessons
The two builds the track has been drawing, run end to end on a real model: search your own notes and answer with sources, and the returns desk as a working agent, each with an eval that catches it failing.
after The loop and its harness
- 01RAG Over Your Own Notes8 Q
- 02The Returns Desk, For Real (a real model, a real harness, and a dress rehearsal that fails)8 Q

Part three · Scale

8Serving LLMs6 lessons
One 7B model on one A100, taken apart: the two clocks a request runs on, the prefix cache that shortens the first, chunked prefill and the prefill/decode split that protect the second, the grammar that masks the logits, and the router that decides which model answers at all.
after Past the basics
The Two Clocks Prefix Caching Chunked Prefill Prefill Decode Split Constrained Decoding Routing and Cascades

Borrowed from other tracks (2)

You don't need to take those tracks. These are the only lessons this track uses from them, and each is linked where it is needed.