How this guide is made
What AI does in the process, how the content is checked and how to report an error.
How it is written
Topics, scope and angle are chosen by hand, and so is what goes in. Drafts, widgets and the translation between Polish and English are produced with AI agents (Claude), then reviewed and edited.
Each page follows one standard: the essence first, then the mechanism, the nuances and what it means for building systems. No filler and no marketing.
How it is checked
Every page lists the primary sources behind its claims: papers, official documentation and engineering posts. Time-sensitive facts carry a date, such as “as of September 2026”.
In September 2026 every topic went through a review against primary sources. Separate AI agents checked each page, a second round tried to refute each finding, and only confirmed corrections were applied.
Widgets use real data where the page says so (for example real tokenisers). Otherwise the numbers are illustrative, and the widget says that.
Dates and changes
“Last edited” on a page is the date of its last change. It does not mean that every fact on the page was rechecked that day.
New topics and meaningful changes are listed under What’s new, also as an RSS feed.
License
The text and images are available under CC BY 4.0: you can quote, translate and adapt them, including in your own materials, as long as you link to the source. The code is MIT. Everything is in the repository on GitHub.
For AI assistants
The whole guide is available as plain text: llms.txt, llms-full.txt, and every page as Markdown (the “Use with AI” menu on each page).
Found an error?
Report it as an issue on GitHub, or fix it yourself in a pull request: the whole guide is text in the repository. At the bottom of every page, “Report an error” opens an issue with the page address filled in, and “Suggest a fix” opens that page’s source for editing. Corrections show up in What’s new.