Set up the lab

Do this before Module 01. About two hours. Everything runs on a laptop; no GPU.

Quick start

cd course/labs
./run_all.sh

The script creates a virtual environment with NumPy on first run and executes all thirteen labs. Every lab prints the measurement that backs one first principle and fails loudly if it does not hold. Expect PASS thirteen times. If any fails, your environment is wrong; fix that before anything else.

What is in the lab

Lab	Principle	Module
01 gradient descent by hand	3	02
02 causal attention with scaling checks	4	04
03 the KV cache, identical outputs, measured speedup	5	05
04 eval harness: recall@k, code grader, judge calibration, pass@k vs pass^k	11, 13, 14, 15	08
05 BPE tokenizer from scratch	2	03
06 serving memory math	6	05, Track B
07 next-token distribution and sampling	1	03
08 loss as a power law in data	7	05
09 RL reward hacking and the KL leash	8	06
10 attention dilution and request cost	9, 17	07, 10
11 prompt injection defeated in code	16	10
12 roofline and a measured batch effect	18	Track B
13 model-upgrade noise versus regression	19	08, 17

The targets

Three planted-bug programs in targets/. Each prints flags as you fix it. They are used after Modules 05, 08 and 10. Run them only when the module says to; the first run is meant to fail.

Accounts and keys

• One LLM API key (any major provider) for Modules 07 to 11. Set it as an environment variable; never commit it. Budget $50 to $150 over the course.
• A free Weights & Biases account for Module 05 and 06 logging.
• A GPU rental account for Module 05 (nanochat). See the compute appendix for free tiers and prices. Budget $50 to $150.
• A GitHub account with a free GitHub Pages blog. Every module ends with something public.

What is NOT in the lab

The labs run on NumPy alone, so each principle is visible without a GPU, a hosted model or a vector database. The projects in each module use real tools; the modules name them.

Test evidence

The verification report page is regenerated on each build with the lab results, the target checks and the citation check. If what you see on your machine differs from it, open an issue in your own fork and record it; disagreement with the report is itself evidence of something.