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
  • Handwritten GPT-2 training in C, graded against a Python answer sheet

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

Current folio: C Field Notes I — Python Writes the Answer Sheet through II — There Is No Autograd Here.

04 · AI systems

Status: in progress

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

Published sequence: Eval Field Notes I — The Blind Set Went Stale and II — The Model Asked, the Tool Said No. A small retrieval agent over four public Chinese contract documents, a calibrated LLM judge, frozen blind sets and synthetic users. Twelve to sixteen questions per set: evidence about method, not a benchmark.

05 · Original work

Status: in progress

  • 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

Published sequence: Adapter Field Notes I — The Loss Is Not the Task through III — Scale Bought the Base. One Colab A100, Qwen3 QLoRA and a ViT classifier, frozen 128-row evals. Not a production RAG system, and not a claim that bigger adapters win.