// the find
geohot/ai-notebooks
Some ipython notebooks implementing AI algorithms
A grab-bag of from-scratch implementations of ML and RL algorithms by geohot: GANs, VAEs, VQVAE, RevNets, MuZero, PPO/SAC/TD3, a CFR poker solver, and a couple of transformer notebooks, split across TensorFlow 2, PyTorch, Keras, and tinygrad. It's for someone who wants a minimal reference implementation of a specific algorithm to read or compare against, not a tutorial series.
Covers algorithms most notebook collections skip - MuZero, VQVAE, RevNets, self-distillation, CFR poker - rather than just another MNIST classifier walkthrough. Each notebook is small and self-contained, so you can read the whole thing in GitHub's viewer without chasing abstractions across files. Mixing frameworks (TF2, PyTorch, Keras, tinygrad) for similar algorithm classes lets you compare how the same idea looks in different libraries. MuZero gets pulled into its own supporting module (game.py, mcts.py, model.py) instead of being crammed into notebook cells, showing at least some structure for the more complex case.
No requirements.txt or environment pin anywhere, and the notebooks span TF2/PyTorch/Keras/tinygrad simultaneously - given the last push was months ago, dependency rot is likely and nothing tells you which notebooks still run top-to-bottom. Zero documentation beyond filenames: no markdown cells explaining what a notebook demonstrates or what correct output looks like, so you need to already know the algorithm to tell if it worked. No tests, no CI, no saved outputs checked in. The stated contribution model is literally 'file an issue and maybe I'll do it tomorrow' - there's no real maintenance commitment if something breaks.