Curriculum: first principles to frontier

Eleven modules, from the math to shipping LLM systems. Read a module, then do its lab: understanding shows up in what you build and measure, not in what you have read.

Every module has the same anatomy: why it matters → mechanisms with derivations → worked numbers → interview traps → CPU vs GPU notes → check yourself → visual guides → read next.

00 learning OS ─▶ 01 math ─▶ 02 compute & hardware ─▶ 03 deep learning ─▶ 04 transformers
                                                                              │
          ┌──────────────────────┬──────────────────────┬─────────────────────┤
          ▼                      ▼                      ▼                     ▼
   05 pretraining ─▶ 06 post-training          07 inference        10 applied LLM systems
          │                      │                      │                     │
          └──────────▶ 08 evaluation & research ◀───────┘                     │
                                 │                                            │
                     09 interpretability & safety              tracks/founder ◀┘

#	Module	Labs	Mastery target
00	How to learn at this level	—	a weekly loop that produces artifacts
01	Math	01	derive any layer's backward pass; PPO and DPO from their objectives
02	Compute & hardware	03	predict runtime and memory within 2× before measuring
03	Deep learning	01, 02	debug a diverging run from its curves
04	Transformers	04, 05, 18	implement a modern LLM from a blank file
05	Pretraining	08, 09	plan a 7B run end to end: data, compute, parallelism
06	Post-training	10–12	implement SFT, DPO and GRPO and name their failure modes
07	Inference	06, 07, 13, 14	size and optimize a serving system
08	Evaluation & research	15	run an ablation that survives review
09	Interpretability & safety	16	find a circuit; argue a safety position
10	Applied LLM systems	17	ship an eval-driven RAG or agent product

Mastery levels

Rate yourself monthly in the journal:

Level	Meaning
0	heard of it
1	can explain it
2	can derive it
3	implemented and tested it
4	predicted the numbers before measuring
5	extended it, or found where it breaks

Core-AI interviews probe levels 3–4. Research roles probe level 5.

Going further

• papers.md: the reading list by module, with what to take from each paper.
• resources.md: courses, books and communities, curated to fewer and better.
• The library: the best visual explainer and the definitive paper for 19 hard topics, and which sections to read.
• Company guides: which modules each frontier lab's interviews lean on.

Tip

Every module ends with CPU vs GPU notes. Nothing in the curriculum requires a GPU to understand; the GPU only changes how large your experiments can be.