New AI training method learns without backpropagation, using only local rules
A pure-Python project posted on GitHub, GitLab, and Codeberg demonstrates a different way to train AI: no , no gradients, no weight transport — only local learning rules, with happening on a single graph light enough to run on almost any hardware. Its flagship demo solves the full 52-card shuffle problem (harder S4/S5 shuffle puzzles that typical LLMs struggle with), training from scratch in under two seconds and correctly predicting the exact final order of a deck after a thousand shuffles, even though it was only ever trained on short examples.
It can also recover from bad training: if it's first taught incorrectly and only memorizes, training it properly afterward lets it genuinely understand the task without forgetting what it already knew. The project includes three small demos; the first teaches the system what numbers are by having it count piles of items (characters, words, anything), after which it can add numbers correctly despite never being shown a single addition example.
Key points
- Pure-Python demo trains with only local rules — no , gradients, or weight transport
- Learns to solve the full 52-card shuffle problem in under two seconds from scratch
- Generalizes from short training examples to long sequences
- Can be retaught correctly after bad training without losing prior knowledge
- Includes a demo where it learns addition purely by counting piles of items, never shown a sum directly