If you build with diffusion or autoregressive models every day but never learnt where they come from, a free primer doing the rounds on Hacker News is aimed squarely at you. The Little Book of Generative AI Foundations, by Tianhua Chen, derives the main families of generative models from first principles on one connected path, so the point is not to catalogue architectures but to show how they relate to each other.
What sets it apart from the hundreds of surveys already out there is that single route. The book moves through principal component analysis and its probabilistic version, variational autoencoders, diffusion models, normalising flows, autoregressive factorisations, GANs and Wasserstein GANs, and energy-based models. Each is worked out step by step, and the thread connecting them is the whole idea: a reader sees that these are variations on shared mathematics rather than seven unrelated tricks. The author is explicit that the book does not try to cover every new architecture or the implementation details, and keeps its focus on the derivations you need to follow the reasoning.
The revised version is the one to read. It was posted on 8 September 2026, runs to 195 pages, and adds visual explanations throughout, where the first version from 28 May was shorter and plainer. It is available as a PDF, an experimental HTML build, the TeX source and even an audio summary, and the author invites comments and corrections.
The value is practical rather than novel. None of the maths is new, which is the point: this is a teaching text, not a research result. For an engineer who uses the tools behind models like Stable Diffusion or the large language models we review and wants to understand them properly, a single derivation-led path is more useful than another list of methods. It sits alongside established textbooks rather than replacing them, and makes a good bridge into the heavier standard references.
The honest caveat is that it is a self-published preprint, not a peer-reviewed textbook, so it has not been through formal review and any derivation you lean on for real work is worth checking against a standard source. A 195-page book is also a commitment, not a quick read, and it assumes comfort with undergraduate probability and linear algebra. For the right reader those are small prices for a clear, connected account of why these models work.