Hey hackers, the world needs more AI researchers with good taste, and hardcore software folks have some of the best. Many software friends mentioned they learn better from implementations than from papers, but existing open-source examples rarely go beyond basic nanoGPT-level demos.

To help bridge that gap, I spent the last two months full-time reimplementing and open-sourcing a self-contained implementation of every major modern deep learning technique from scratch. The result is beyond-nanoGPT, containing 20k+ lines of handcrafted, minimal, and extensively annotated PyTorch code.

I'd love feedback on how to make it more helpful for people interested in transitioning into deep learning research. The roaring 2020s are a surreal time to be alive, and we need all hands on deck.