This is a living map, not a syllabus I need to finish perfectly. Each stage should produce runnable work and a clearer explanation before I move on.

01 · Foundations

Status: built

  • Python and tensor thinking
  • Essential linear algebra and probability
  • Data, train/validation splits, and baseline discipline
  • Reading shapes and tracing values through code

02 · Neural networks

Status: built

  • Backpropagation and computational graphs
  • MLPs, activations, initialization, and normalization
  • Training loops, optimization, and overfitting
  • Controlled experiments: changing one variable at a time

03 · Transformers

Status: in progress

  • Attention as communication
  • Residual streams and MLP computation
  • Tokenization, embeddings, and positional information
  • Reproducing GPT-style training from first principles
  • Scaling, batching, checkpoints, and evaluation
  • Exporting learned weights and tracing token-by-token inference in C

Published sequence: Transformer Field Notes I — The Batch Is a Factory of Futures through VI — The Extra Code Is Not Another Brain, followed by Inference Field Notes I — The Model File Has No Names.

04 · AI systems

Status: next

  • Retrieval and reranking
  • Evaluation sets and error analysis
  • Agents, tools, and reliable workflows
  • Serving, observability, latency, and cost

05 · Original work

Status: later

  • Form a testable question
  • Build the smallest credible experiment
  • Publish the method, failures, and result
  • Let the evidence—not the novelty claim—carry the work