Take it apart · prove you know it · nothing to install

What is actually happening when you type into a chatbot?

Every concept here is a thing you can take apart rather than a thing to read about. Work through the lesson, put your hands on the mechanism, drill the cards, then see whether it stuck.

0
XP earned: 0/10
Lessons finished: —
Day streak

Level 1 · Curious150 XP to Tinkerer

One trip through the machine

A preview of the whole course

One message to a chatbot makes the same five-step trip every time, whether you are asking for a summary, a rewrite or a line of code. Nothing is looked up and nothing is planned in advance. The whole of this course is those five steps in detail, and every one of them is a concept below.

Your text is cut into tokens. Words become fragments, fragments become numbers, and the letters are gone from that point on. Those numbers become positions in meaning space — a place where words that get used alike sit near each other. Attention decides what relates to what, so “it” finds its noun and an instruction near the top still reaches a sentence near the bottom. Every token in the vocabulary is scored as a candidate for what comes next. One is picked, appended to your text, and the whole trip runs again from the beginning.

What happens when you press enter

One trip through the machine, from the words you type to the first word it writes back. Step through it — each stage is a module later in the course.

You typed

Summarise this in one line

A word is split before the model sees it

Each fragment becomes a number. The letters are gone from that point on. The pieces are usually smaller than a word, which is where cost and length limits come from.

Covered later in Tokens

1 of 5

The course

0/30 mastered

——Final examOpens after 3 more concepts
——PlaybookYour notes, and anything you saved
——GlossaryEvery term in the course, searchable67 terms

Badges

0/11 earned

First StepsOpen your first concept

Token WranglerGuess a token count exactly

CartographerRun a vector analogy

OracleOut-predict the model 3 times in a row

Cool HeadSample at both extremes of temperature

DetectiveSpot every hallucination in an answer

PerfectionistScore 100% on any quiz

Speed MatcherFinish a match game under 30 seconds

ScholarOpen all ten concepts

LecturerFinish every Learn lesson

GraduatePass the final exam

The wider course

This is the first piece

Language models are the part of this most people meet first, which is why they are the part we built first. There is more to come.

01LEXIOS LLM LabHow a language model works, by poking at the mechanismYou are here
02Generative AIImages, audio, video and code, with the same idea pointed at other kinds of outputNext
03AgentsWhat changes when a model can use tools, take actions and run without you watchingPlanned
04Building Your Own ModelsFine-tuning, evaluation, and when training something yourself is actually the right callPlanned
05Data CentersThe physical machine underneath all of it: chips, power, water and what it costsPlanned