Learn to do AI research, starting under a model that is already running.
Continual School is one continuous track — from the computer a model runs on, through the arithmetic of a forward pass, up to a research project of your own that gets published with your name on it. It assumes nothing underneath it, every lesson is filmed by us, and it will cost nothing. The whole track is written and readable today; the videos are being made.
46 lessons · 1 filming · 45 specs written
Six levels, in one line, and you can start anywhere on it.
The foundations carry a “skip this if…” line, because skipping is the designed behaviour: a working engineer drops in at Level 2, someone who has never opened a terminal starts at the bottom and is never left behind.
- 00
The Computer
Not a computing course — the floor under a running model. We start with one already answering and go down into the process, the memory, and the bytes it is made of. Nothing assumed below that. Skip on sight if you write software.
- 01
The Math
Not a maths course — the arithmetic a forward pass is already doing, read off a model mid-sentence rather than built up beforehand. Skip anything you already have.
- 02
The Machine
What a language model actually is — every lesson runs on a small open model, live. Ends by walking the whole road from GPT-2 to a coding agent and showing the core never changed: chat, tools and agents are conventions and loops stacked on next-token prediction.
- 03
The Craft
What working researchers do all day — the tacit knowledge locked inside labs.
- 04
The Frontier
Efficiency, small models, continual learning — the level nobody else teaches.
- 05
The Lab
The only level you do rather than watch. Levels 0–4 are lessons; this is a real research project on an open problem — you write the proposal, we advise it, and it is published under Continual Society Research with your name first.
No intros. The first frame is the thing itself.
A model is on screen in every lesson
There is no groundwork phase. The computer is explained because that is where the model runs; the maths because that is what the forward pass is doing. Strike the AI words from a lesson and nothing should be left standing.
Result first, theory second
Every lesson opens on a real process already running — generating, breaking, training — and runs until something surprising happens. The theory arrives afterwards, to answer the one question you now have.
First principles, complete
The ladder assumes nothing below it and names nothing before showing it. Foundational lessons carry a skip line, so knowing more just means starting higher.
The last level is a real paper, with your name first.
Everything below Level 5 you watch. The Lab you do: you pick one of 16 open problems, write the proposal, and we advise it to a deadline. It publishes under Continual Society Research — and where it holds up, on arXiv. Each problem states its experiment, the compute it honestly needs, and how it publishes even when the result is negative.
Read the open problems →Efficiency — can small actually win
The company's first bet is that architecture, not scale, is the remaining bottleneck. These four try to measure that instead of asserting it, and two of them could sink it.
Continual learning — weights that keep moving
The end goal, and the part of the field with the most citations and the fewest clean curves. Every one of these starts by drawing a baseline nobody has drawn properly at 4–9B.
Agents — where a small model loses
Small-versus-large agent gaps get reported as one number. These take that number apart, on the grounds that whatever is inside it is the research programme.
Method — how the field measures itself
Unglamorous, heavily cited, and the safest place to start. Everything downstream of a broken benchmark is broken, and nobody is checking.
The track is readable today. The videos are coming.
46 lessons · 1 filming · 45 specs written — and every one of those specs is on the site now, in the order it will be filmed. Read it from the bottom or drop in where you already are, and the videos land underneath as they are made.