Open Learning Lab
The Residual Stream
One student is rebuilding AI from absolute zero — publishing every working piece and every failure, week by week. Week 1 is live now.
// The Problem Everyone Uses AI. Almost Nobody Can Explain It.
You type a question, a chatbot answers. Somewhere between those two moments, something astonishing happens — and for most people, including most people who study computers, that something is a closed box. Tutorials don't help much: they hand you finished code, it runs, and you've learned how to copy, not how to understand.
There's a second problem too: the gap between reading about something and being able to build it. It's the difference between knowing the rules of swimming and jumping into the pool. Most learning online hides the splashing — portfolios show only polished wins, so every learner quietly assumes everyone else finds this easy. They don't.
And it matters who's inside the room where AI gets built. It's 11 PM in Kampala and a student is asking whether tools like these treat African languages fairly — and whether someone from this side of the world can help decide that, by proving it the slow, honest way: building the thing from nothing, in front of everyone.
// The Payoff What Following Along Gives You
Skill built in front of you
Every piece starts life as a blank file — nothing copied, nothing imported wholesale. You watch real understanding form, including the hours where nothing works.
A method you can steal
Before any test runs, the guess gets written down first: what do I expect, and why? That one habit — predict, then check — is what separates studying from guessing.
Mistakes included, not hidden
A public learning log (LEARNING_LOG.md) records what broke and what still isn't understood. No highlight reel — the full game, updates weekly.
Famous papers, made readable
Each landmark research paper is read cover-to-cover and boiled down into plain notes: the claim, why it changed the field, the one result that proved it, and what surprised.
Real questions, real answers
Original experiments, stated before they run — including one close to home: does writing the same sentence in Luganda really cost about three times more tokens (the tiny pieces AI reads) than English?
Evidence, not promises
The finish line: a working AI language model built entirely from scratch, written experiments, and an honest public checklist of frontier-intern skills — two underway, the rest openly marked "not yet."
// How It Works Eight Weeks, Four Honest Phases
Lay the groundwork — Weeks 1–2
Rebuild the mathematical basics by hand until they're solid enough to stand on, then construct the first real pieces: the part that teaches a machine to read, and the routine that makes it learn. No borrowed code, no shortcuts.
Read the classics — Weeks 3–4
Work through the famous research papers that created modern AI — read in full, not skimmed — and turn each one into notes a non-expert could follow.
Ask original questions — Weeks 5–6
Run first-of-their-kind experiments, like whether African languages pay a hidden cost in today's AI tools. Prediction written first, results published either way — plus reproducing a paper's result and then pushing past it.
Push further, tell the truth — Weeks 7–8
Add the advanced finishing techniques, tidy every note and notebook until a stranger could follow them, and write the full retrospective — including everything still unknown.
// Who It's For Pull Up a Chair
Don't Just Use AI. Come Watch One Get Built.
Everything — the code, the paper notes, the broken attempts, the honest weekly log — is free and open on GitHub. It's Week 1, so you're early. If something here is wrong, or you're working on similar problems at a lab, a research group, or your own desk: email me directly.