About
One machine, explained in 27 parts. What follows is the scope, where every number came from, and the places this is still short.
What this is
An explanation of the system built around a large language model: how one is made, how it is run, and what the ordinary software around it does. It is not a survey of artificial intelligence. Image models, robotics, classical machine learning and rule-based systems are different machines, and reasoning about them from what you read here will lead you wrong. The opening part draws that line explicitly.
- parts
- 27
- connections
- 35
- things to operate
- 29
- terms defined
- 101
Where the numbers come from
This site argues that you should always know whether a number was measured or asserted. That obligation applies to it.
Arithmetic
Computed in your browser, live, from the numbers on screen
Token counts, costs, latency, parameter counts, file sizes, gradient descent, schedules, filter counts, retrieval scores.
A small language model
Counted from this site’s own prose, and shipped as a data file
The next-token distribution and everything the temperature dial does to it. A trigram model over 35,000 words. Real probabilities, from a model far too small to be useful for anything else.
Word vectors
Counted from this site’s own prose
The embedding space and the retrieval pipeline. Co-occurrence counts weighted by pointwise mutual information, reduced to fifty dimensions.
Written examples
Authored by the people who made this site
Two widgets: the attention grid, and the base-versus-tuned comparison. Both say so above the visual rather than underneath it. Also the answers in the context-window and instructions-file demonstrations, which are labelled where they appear.
Volatile facts
One data file with a review date, and a build that fails when one expires
Prices, rates, corpus and file sizes. Nothing else on this site is allowed to carry a figure that will go out of date.
No model is called, at build time or in your browser. The tokenizer, the small language model, the word vectors and every figure on this site are computed on your own machine. No fonts or scripts are loaded from another origin.
What is measured about your visit
One thing does leave the page. Vercel Analytics records which page was opened, where the visit came from, and coarse device information such as browser and country. It is first-party: the script and the measurement it sends are both served from this domain, so nothing goes to another company's server. It sets no cookies. Visitors are counted with a hash that is rotated daily rather than an identifier that follows you between visits.
It was added in August 2026, and this section was written the same afternoon. A site that spends twenty-nine interactions arguing that you should know where a number came from does not get to quietly start collecting them.
Nothing is recorded about what you do inside an interaction. The text you type into the tokenizer, the settings you move, and the parts you open are not sent anywhere, because none of that leaves your browser in the first place.
Accessibility
Every visualization ships a real table carrying the same data, present in the HTML whether or not JavaScript runs. Contrast is validated against the actual stylesheet on every build, in both themes. Every page is checked with axe in both themes and both reading depths, and every interaction is checked again once opened. Nothing on the map is hover-only. Motion is limited to what a reader asks for and is disabled under a reduced-motion preference.
The key map lists what is implemented and what is specified and not yet built.
If you are a machine
The whole site is published as markdown as well as HTML, because a crawler reconstructing prose out of navigation and islands does a worse job than we can do for it. llms.txt is an index of the six pages in the format described at llmstxt.org. llms-full.txt is every part in full, including the advanced layer that sits behind a toggle here, the connection graph that is drawn rather than written, and the trace of one request. Roughly fourteen thousand words, in one file.
Both are generated from the same content these pages render, so neither can describe a version of the site that no longer exists. Nothing is disallowed in robots.txt, including the crawlers that gather training data. That is a decision: a model that has read this will answer questions about context windows better than one that has not.
Where this is short
- No session has been run with a screen reader user or a keyboard-only user. Automated checks find violations; only people find the places where a page is technically conformant and unusable.
- No comprehension testing with readers new to the subject. Several things on this site were rewritten because a reader said they made no sense, and there is no reason to think that process is finished.
- The arrow-key grammar is not yet uniform across every interaction.
- The linear reading route exists for one act only. The map is the intended way through.
Telling us we are wrong
Corrections are welcome and specific ones are most useful: which part, which sentence, and what it should say instead. Two categories are especially wanted. Anything factually wrong, since the whole point of this is to be accurate. And anything that reads clearly to someone who already knows the subject and means nothing to someone who does not, which is the failure mode this site is most prone to and least able to see in itself.