linkdrop-x-deronin-karpathy-nn-from-scratch

Overview

This is a pointer skill — not a transcription. The source tweet (Ronin, @DeRonin_, 2026-04-18) links to Andrej Karpathy's "Neural Networks: Zero to Hero" lecture series. The first lecture of that series (the 2-hour "micrograd" lecture) builds a scalar autograd engine + MLP from scratch with no PyTorch / TensorFlow — pure Python + calculus.

This skill does NOT reconstruct the lecture content. The value is: (1) the video is the canonical teaching resource, (2) karpathy/micrograd is the companion repo you clone and code along with, (3) the workflow below turns passive watching into real understanding.

The resources

• Video: Andrej Karpathy's YouTube channel (@AndrejKarpathy). "Neural Networks: Zero to Hero" playlist. First episode = micrograd lecture. Find it via the channel, not via an assumed URL — playlist ordering is authoritative.
• Companion repo: https://github.com/karpathy/micrograd (15,500+ stars, MIT)
  • Tiny autograd engine: ~100 lines of core Python (DAG over scalar values)
  • micrograd.nn module with PyTorch-like API
  • demo.ipynb: trains a 2-layer MLP binary classifier on the moon dataset
  • trace_graph.ipynb: produces graphviz visualizations of the computational graph

Why it matters