Mathews-Tom / no-magic

What is no-magic?

No-magic is a Python collection of tiny, self-contained scripts that implement core ML algorithms from scratch—like GPT, BERT, GANs, diffusion models, CNNs, and optimizers—using a custom scalar autograd engine. Because `model.fit()` isn't an explanation, it reveals exactly how these work through forward/backward passes on toy data like names or spirals, printing losses, predictions, and comparisons. Users get instant, runnable demos that train in seconds, no libraries needed beyond Python stdlib.

Why is it gaining traction?

It stands out by stripping away framework magic, letting you see gradients flow and architectures click without black-box calls. Developers hook on the "aha" moments, like bidirectional attention in BERT vs causal in GPT, or why Adam converges where SGD stalls. Pure Python means copy-paste-train anywhere, perfect for quick experiments or teaching no-magic-numbers intuition.

Who should use this?

ML beginners building intuition, engineers prepping system design interviews, or researchers prototyping ideas before scaling. Ideal for understanding RAG retrieval, embedding similarities, or RNN gating failures on character-level tasks like name generation.

Verdict

Grab it for educational gold—372 stars show steady niche interest—but the 1.0% credibility score flags immaturity: sparse docs, no tests. Strong start for learning, skip for production.