self-study — index¶
Each project owns its own study.md with only the material that project needs, in the order
you'll hit it. This file is the connective layer: cross-cutting math everyone needs, plus a map
of where each topic actually lives.
Read alongside building, not before. Stop reading a phase the moment you can build the thing it describes.
Cross-cutting (not owned by any single project)¶
- 3Blue1Brown — Essence of Linear Algebra — videos 1–9, skip eigenvalues. Needed before seatorch milestone 4.
- 3Blue1Brown — Essence of Calculus — videos 1–8, chain rule (ep. 4) matters most. Needed before seatorch milestone 1.
- Stanford — Random Variables / Conditional Probability & Bayes — background only, no project currently needs these directly; skip unless curious.
Where each topic lives¶
| Topic | Project | File |
|---|---|---|
| Manual memory management | cmatrix | study.md |
| Backprop, autograd graphs | seatorch | study.md |
| SIMD/cache optimization | seatorch | study.md |
| Quantization (INT8) | seatorch | study.md |
| CUDA matmul | seatorch (stretch) | study.md |
| FlashAttention | seatorch (stretch) | study.md |
| Digital design, FSMs, gates | mini-npu | study.md |
| Systolic arrays, memory hierarchy | mini-npu | study.md |
| UART protocol | mini-npu | study.md |
| Transformers, attention | rockygpt | study.md |
| PyTorch autograd (for contrast) | rockygpt | study.md |
What to skip, project-wide¶
- Full ML MOOCs (fast.ai, Andrew Ng) — too slow, the per-project links cover more ground faster
- Cloud deployment, fine-tuning via API, prompt engineering — not the stack
- Vector DB tooling — not relevant here
- K&R — Beej covers what's needed, faster
Read annotations, not just links. Stop the moment you can build the thing.