How this guide works
A book about AI goes out of date before it's printed. This guide is built to change: a stable core that explains how these systems work, and a living layer that is re-checked every month.
Why "Machinery"
In 1950 Alan Turing published "Computing Machinery and Intelligence", the paper that asked whether machines can think. This guide takes the more modest half of that title: it is about the machinery, how these systems actually work, so that you can form your own view on the bigger question. The name is meant to outlast any one technology. (And yes, the address machinerylearning.eu is a pun.) Today's content is about language models; when the field moves on, new parts will follow.
The subtitle changes with each edition. So far:
- November 2026: Next token, and what comes after
Two reading paths, one guide
Every chapter has a plain-language core for people who are new to this. Technical detail sits in Engineer depth sections that stay folded on the beginner path and open on the engineer path. Switch at any time with "Reading path" at the top; the choice is remembered in your browser.
Every fast-moving fact shows its age
Facts that can go stale, such as model releases, benchmark numbers and research results, carry a small date badge showing when a person last checked them against their sources. Each has a re-check interval, from three months for fast-moving claims to a year for slower ones. Once a fact is overdue, it visibly fades until it's verified again. The full list, with sources, is on Sources & freshness.
The monthly update
- Research. At the start of each month, an AI research agent works through a fixed list of sources (lab research blogs, new papers, a few trusted newsletters) and compares what it finds against the guide's claims and radar.
- Proposal. It drafts updates as a pull request: re-verified or rewritten claims, radar moves, new entries, and a draft of the month's field notes, each with sources.
- Human review. A person reviews every change against its sources, edits, and merges. Nothing is published without that review.
The editorial bar: something goes in only if it changes how a reader should understand or use these systems. That keeps the guide from turning into a news feed.
How the update pipeline is secured
An agent that reads arbitrary web pages and then writes to a repository is exactly the prompt-injection risk the guide warns about. So the research agent can read the web and open pull requests, nothing else: no merge rights, no access to secrets, no changes to the build or deployment configuration. Anything it fetched is treated as data, never as instructions.
Who writes it
Maintained by Robert Haist. Drafts are written with AI assistance and reviewed by a human before publishing. Corrections and suggestions are welcome as issues or pull requests on GitHub.
Progress, reading path and last-visit date are stored only in your browser.