Learning paths¶
The book has 24 chapters in a fixed order, and a reader who works through all of them in sequence will not go wrong. A learning path is the other way in: it starts from something you want to be able to do and names the chapters that get you there, in the order to take them.
Each path states a goal, the prerequisites it assumes, the sequence of chapters with a line on why each one is in the path, and the milestone that proves you reached the goal. No path contains new material: every one of them is a route through the same 24 chapters.
| Path | Goal | Core chapters | Time |
|---|---|---|---|
| 0 — Foundations first | Reach the point where every later chapter is readable. | 1, 2, 3, 4, 5, 6, 7 | 29–48 weeks |
| 1 — Understand how LLMs work internally | Follow a prompt from characters to sampled token, and read any model card for what it is. | 13, 14, 15, 16 | 21–28 weeks |
| 2 — Pretrain an open-source LLM | Plan and run a pretraining run end to end, and finish with a base model of your own. | 14, 15, 16, 17 | 22–30 weeks |
| 3 — Improve an existing model | Take a released open model, make it measurably better, and prove the gain. | 16, 17, 18, 19, 20, 24 | 23–32 weeks |
| 4 — Serve, deploy and operate an LLM | Run an open model in production at a known cost and latency, and keep it improving. | 21, 22, 23 | 14–21 weeks |
| 5 — Build applications on open models | Ship a product on top of an open model, and keep up with what replaces it. | 21, 22, 24 | 10–13 weeks |
Choosing one¶
- Starting from zero, take Path 0. Everything else assumes most of it, and its Chapter 6 runs in parallel with the mathematics rather than after it.
- Wanting to know how the machine works rather than to operate one, take Path 1.
- Wanting to train a model of your own from raw text, take Path 2. It is the most expensive path in both time and money.
- Wanting to make an existing open model better, take Path 3. This is the research path.
- Wanting to run a model in production, take Path 4.
- Wanting to build a product on someone else's open weights, take Path 5, the shortest route here, which treats the Chapter 23 capstone as an optional finisher rather than a requirement.
Paths 4 and 5 share Chapters 21 and 22 and part company at the end — the capstone for one, the reading loop for the other. Paths 2 and 3 are the two halves of training: one before a base model exists, one after.
How the paths relate to the chapters¶
The core chapters of each path are the rows of the map of goals to chapters, with Path 0 added for readers starting from zero. The prerequisite lists and the reading notes are editorial: they say what a path leans on most heavily, which is not the same as permission to skip the chapters in between. A chapter left out of a path is left out of that route, not declared optional.