makemore-style

name	makemore-style
description	Generate a progressive learning notebook for any domain — starts at bigram baseline, escalates to MLP, RNN, transformer. Each stage adds one concept. Karpathy makemore pedagogy. Triggers on "build progressive notebook", "makemore-style tutorial", "teach me X step by step".

makemore-style

Layered teaching: each model is one step deeper than the last. Run, compare loss, then add complexity.

Stages template

1. Counting baseline — frequency table. No learning, just statistics.
2. Bigram NN — single linear layer, learn what counting did.
3. MLP — context window, embeddings, hidden layer (Bengio 2003).
4. WaveNet / dilated — hierarchical context.
5. Transformer — attention replaces fixed mixing.

Per-stage requirements

• Same dataset, same eval (loss + samples).
• Show parameter count delta.
• Show loss delta vs prior stage.
• One paragraph: "what did this stage buy us?"

Output

Single .ipynb or markdown with code blocks, one section per stage.

Reference

• makemore: https://github.com/karpathy/makemore
• "Building makemore" lecture series