Learn how AI actually works — from the inside. You write the facts. You compile them. You ship a working offline AI that fits in a text message. No Python. No maths prerequisites. No GPU. No cloud.
What You Build
Every session ends with a working artifact. Not a slide. Not a quiz. Something you compiled and ran.
Your first compiled binary — a First Aid Baby with facts you chose and added yourself. Runs offline on any Mac.
A TYPE_OF chain in your own domain. One fact answers dozens of questions it was never explicitly taught.
Three deliberate breaks. Three diagnoses. Three fixes. You will understand compiler errors and runtime gaps.
A written analysis of a machine-proposed mutation — is it correct? Would you authorise it?
You encounter the Gatekeeper. You type CANCEL. Then AUTHORIZE. You write who is responsible for the outcome.
A compiled binary in a domain you know. A peer runs it without the compiler. You ship software on the last day.
The Six Sessions
Each session is self-contained. They stack — each one builds on the last — but the course works in a classroom, online, or self-paced.
Install the Kit, compile first_aid.light, ask it questions. Then add three facts of your own. Recompile. Test them.
Add TYPE_OF chains. Watch the Baby answer sunburn by looking up burn — without a sunburn fact. Build a 2-hop chain.
Break it deliberately — missing quotes, duplicate facts, unknown queries. Fix each one. Understand why "I don't know" is safer than a guess.
Run evolve-baby.sh on a gap in your Baby. Watch the computer propose a new fact. Read it. Decide whether it's correct before anything compiles.
The evolution engine pauses and shows you its proposed code. You type AUTHORIZE — or CANCEL. You write an Incident Report. You decide who is responsible.
Write a Baby in your own domain. Compile it. Hand the binary to a peer who has no compiler. They run it. You shipped software.
Learning Outcomes
Not knowledge you memorised. Understanding you built by doing.
The difference between a knowledge-based system and a generative language model — and when to use each
How to write, compile, and debug a Light Code program from a blank file
Why inheritance lets a small fact table answer questions it was never explicitly taught
What genetic programming means in concrete, hands-on terms — not theory
Why the Gatekeeper Protocol exists — and who is responsible for what an AI does after you authorise it
How to ship a self-contained AI binary that runs on any Mac without a compiler or runtime
Why "I don't know" is an architectural safety choice — and what it protects against
How to capture domain knowledge precisely enough that a machine can reason about it
Why This Approach
This course doesn't. Every answer is traceable to a fact you wrote. Every decision is one you made.
| This course | Typical "learn AI" MOOC | Using ChatGPT | |
|---|---|---|---|
| You see how every answer is produced | Yes | No — theory only | No |
| You write the AI's knowledge yourself | Yes | No — you train on data | No |
| Runs offline, no subscription | Yes | Depends | No |
| GPU required | No | Often yes | Cloud GPU |
| Can hallucinate / make things up | No — architecturally impossible | Yes | Yes |
| You ship a working artifact | Yes — session 1, and every session after | Rarely | No |
| Teaches AI safety hands-on | Yes — session 5 | Sometimes, as theory | No |
Pricing
The course is free with the Kit. School and team licences exist because classrooms need one copy of the compiler shared across many students.
The Kit ships as a download. Install takes 3 minutes on Apple Silicon. Your first compiled Baby AI takes 15 minutes. Session 1 is 45 minutes. By the end you will have written, compiled, and queried a working AI from its source code.