Start with the picture
You do not need to master calculus before understanding AI. Learn enough to explain the mechanism, then use the references to sharpen that picture.
1 · See the territory
Start with what each system receives, predicts, or does. The broad path is an orientation across the major families of AI, not a claim that the field has a final boundary.
2 · Follow one prediction
Learn what weights and activations are. Then trace how a small change travels forward and how its effect on the loss is calculated backward.
3 · Follow one generated token
Understand tokens, attention, and the generation loop before thinking about serving hardware.
4 · Choose your depth
For broad intuition, explain the check aloud and continue. For serving, study the reference and produce the small experiment. If a prerequisite is unfamiliar, follow its link first.
5 · Use the machine you have
The conceptual checks and memory calculations need no GPU. Use a small local model for the first practical steps. Multi-GPU experiments are later extensions; mark hypothetical results clearly.
- The software stack: from tensors to a live process →
- Establish a baseline you can trust →
- Lab 1 · Explain a single GPU’s limits →
Code + guide · ggml · llama.cpp ↗