What I read and what I tried

Start without usThe material is already here

Three free starting points are available in four languages:

• The frameworkPurpose, five axioms, primitives and intelligence as a capability rather than a centre.
• The architectureThe loop, five invariants, three recursions, the node and the link.
• The Evolution logDated corrections: what was believed, what reality showed and what changed.

Read it, challenge it, or implement the ideas without us. Publishing it is part of the design.

Next is a route through the wider field for people who want to learn by building. After that comes what we teach directly.

Self-studyA build-first route from fundamentals to current practice

Do these in order, but learn the mathematics when a build makes it necessary — do not wait to “finish the maths” first. Every stage ends with something you can explain, run or test.

1. See the shape

   If Python is new to you, begin with CS50P. Then use visual explanations to make vectors, derivatives, gradient descent and backpropagation concrete. Keep a mathematics text beside you as a reference, not a gate.

   Leave with: a plain-language explanation of how a small neural network learns.
   CS50P: Python 3Blue1Brown: neural networks Linear algebra Calculus Mathematics for Machine Learning
2. Make learning happen

   Follow Karpathy’s Zero to Hero from a scalar autograd engine to a tiny character-level language model. Type the code; change it; break it.

   Build: backpropagation and a tiny language model you understand line by line.
   Neural Networks: Zero to Hero
3. Learn training craft

   Use selected CS231n notes, lectures and assignments for optimization, training dynamics and debugging. Its examples are visual, but the craft transfers far beyond computer vision.

   Diagnose: overfitting, bad data, unstable optimization and misleading evaluation.
   Stanford CS231n
4. Use the modern ecosystem

   Take the Hugging Face LLM Course once Python and basic deep learning feel comfortable. Learn models, tokenizers, datasets, fine-tuning, evaluation and sharing reproducible work.

   Build: a small, evaluated fine-tune with a model card and reproducible notebook.
   Hugging Face LLM Course
5. Train the whole stack

   Read and run nanochat end to end: tokenizer, pretraining, supervised fine-tuning, evaluation and inference. Start tiny on a laptop; the capable reference run needs multiple GPUs and is not suited to a typical free notebook session.

   Trace: one model from raw text to a measured conversation system.
   Karpathy’s nanochat
6. Move from courses to research practice

   There is no final playlist. Read primary papers, model cards and code; reproduce one claim; design an evaluation before optimizing; record data, provenance, cost and failure cases. For agents, add identity, authority, limits and revocation before autonomy.

   Produce: evidence another person can inspect, challenge and reproduce.
   Read → reproduce → evaluate → publish

Still open

Run one workflow through personal, team and composed-community arrangements with comparable capabilities. Record what survived, what needed rewriting, costs and failures. This is a proposed experiment, not a result or a prerequisite for revising these pages.