Phase 1 · ≈ 4–8 weeks→ Level 1 (Reproducer) begins

Foundations

Understand and implement from a blank file the machinery under every model: gradient descent, backpropagation, a neural net, and the core ML concepts (loss, generalization, overfitting). Pull in exactly the math each step requires — no more.

What to learn · in priority order

1
Backpropagation interactive lab →
The chain rule applied to a computation graph. The single most important thing in this phase. Implement reverse-mode autodiff and most of deep learning stops being magic.
2
Gradient descent & optimizers interactive lab →
SGD, momentum, Adam; learning rate, and why it matters most.
3
A neural net from scratch interactive lab →
MLP, activations, initialization, why depth helps.
4
Core ML interactive lab →
Loss functions (cross-entropy especially), train/val/test, overfitting, regularization, bias–variance. The grammar of every experiment you'll run.
5
Math, just-in-time the ledger →
Linear algebra (what a matrix does to a vector), calculus (the chain rule), probability (cross-entropy/KL). Learn each when the model above forces you to.

Primary path · use these, not ten others

Karpathy — Neural Networks: Zero to Hero ↗

The spine of this phase. Do the exercises, don't just watch. Build micrograd (a tiny reverse-mode autodiff engine — this IS backprop) and makemore (a character-level LM, bigram → MLP).

3Blue1Brown — Neural Networks series ↗

The best visual intuition for nets, gradients, and backprop. Watch the matching chapter when something won't click.

fast.ai or d2l.ai — one breadth source, as reference ↗

Use to fill gaps Karpathy doesn't cover, not cover-to-cover.

Mathematics for Machine Learning (Deisenroth et al.) ↗

Math reference — look up, don't read linearly.

The deliverable

Reimplement micrograd and makemore yourself, from scratch — watch a segment, close it, rebuild from memory and first principles. Push both with a README. Bonus: a short note, "what backprop actually computes."

Milestones · check only when you can do it from a blank file

0/5 shipped

artifact: micrograd + makemore repos + a "what backprop computes" note.

rule: every phase must end with a public artifact (repo + writeup). no artifact → phase not done, regardless of what you've watched.

Traps

✗Watching instead of building

Karpathy makes it look easy; that's the danger. Close the video and rebuild from a blank file. The gap between "followed along" and "can implement" is the entire point.

✗Math-first detour

Do not stop to "finish" a linear algebra course. Look math up when micrograd forces the question, then return.

✗Copy-paste completion

Copying his code and running it teaches almost nothing. Type it from understanding, or rebuild from memory.