Notes written while learning

I am Buffett Liu. This site is where I work through diffusion language models (dLLMs) in public: masked diffusion objectives, sampling strategies, the models that scaled them, and the systems work needed to serve them.

The notes follow two tracks. Inside the Model covers how dLLMs are designed, trained, and sampled;Into Production covers how to serve them efficiently. Each note tries to start from a question the previous one left open. Derivations sit in collapsible boxes so the main line stays readable; interactive figures are there to be clicked.

These are learning notes, so some of them will be wrong. If you spot a mistake, pleaseopen an issue.