Skip to content
Handbook navigation

Full-access lesson

Generative AI

Type a sentence and a graphics card with 10 GB of memory turns it into a 512 × 512 image, using weights that anyone can download. Stable Diffusion made that possible when Stability AI released its weights on 22 August 2022. Four months earlier OpenAI had opened DALL·E 2, the best-known text-to-image model of the time, only to a waitlist of trusted users, through its own website and with no programmatic access. Stable Diffusion's trick was not a bigger network. It never removed noise from pixels at all: it worked on a grid of 4 × 64 × 64 numbers, 48 times fewer values than the image it would become, and only at the very end did a separate decoder turn that grid into pixels.

The complete curriculum

All 12 volumes, companion resources, interviews, and architecture reviews.

Executable engineering practice

Subscriber-only Python, Java, TypeScript, and SQL labs in the isolated runner.

Generative AI | KnowledgeOS