About
World Models 101 is a free primer on how machines learn to predict what happens next. It exists because the literature on this is excellent and almost entirely unreadable unless you already know it.
The shape of it
9 chapters, each built around something you can manipulate rather than something you have to take on faith. Where a concept has an interactive form, it gets one; where it does not, it gets a figure and a paragraph that says what the figure means. The order is pedagogical rather than chronological: ideas arrive when you need them, not when they were published.
Chapters are released as they are finished. The contents page marks what is live, what is being drafted, and what is still an outline. Nothing is paywalled and nothing will be.
What it assumes
That you are comfortable with a gradient, a probability distribution, and matrix multiplication. That you have trained a neural network at least once, even if only a small one. It does not assume any reinforcement learning. The RL you need is introduced where it is needed and no earlier.
Corrections
Technical writing on a moving field is wrong at a steady rate. If you find an error (a misattributed idea, a broken derivation, a demo that lies about what the underlying method does) the correction is genuinely welcome and will be credited. Open an issue, or write to the address below.
Where the ideas come from
Craik, 1943
The Nature of Explanation
Where the 'small-scale model' framing comes from. Short, and still the clearest statement of the premise.
Sutton, 1991
Dyna: an integrated architecture for learning, planning and reacting
Learning from real experience and imagined experience in the same loop.
Ha & Schmidhuber, 2018
World Models
The paper that gave the field its name in its current sense. Encoder, dynamics, controller.
LeCun, 2022
A Path Towards Autonomous Machine Intelligence
The position paper behind JEPA and the argument against pixel-space prediction.
A full reading path lands with the final chapter.