seatorch — study

Only what this project needs, roughly in the order you'll hit it. Skip anything you can already do; this is a reference, not a syllabus to complete linearly.

Before milestone 1–3 (autograd core)

• 3Blue1Brown — Essence of Calculus — episode 4 (chain rule) is the one that matters, the rest is context
• Backpropagation — F=ma of how models learn
• Karpathy — micrograd — build this in Python first if the C version's graph logic isn't clicking. Don't port micrograd's code, just borrow the mental model.

Before milestone 4–6 (MLP, training loop)

• 3Blue1Brown — Essence of Linear Algebra — videos 1–9, skip eigenvalues
• Loss function and optimisation

Before milestone 7–8 (vectorize, SIMD/cache)

• Siboehm — Fast Matrix Multiplication on CPU — this is the playbook for milestone 7–8, follow its optimization sequence

Before milestone 9–10 (quantization, export)

• Deep Compression — Han et al., 2016 — abstract + sections 1–3, pruning + quantization

Before milestone 11 (optional CUDA stretch)

• Siboehm — CUDA Matrix Multiplication — only relevant if you get CUDA hardware access; on AMD locally this is reading-only until then

Before milestone 12 (optional FlashAttention stretch)

• Attention Is All You Need — read after building something that needs attention, not before
• FlashAttention paper