Introduction to language models
Five courses that take you from never having thought about it to being able to read a model's behaviour and say why it did that. No mathematics is required and none is hidden from you either: the parts that are arithmetic are shown as arithmetic, folded away where a reader who wants the shape can skip them.
What this track is for
Most explanations of this technology are written either for people who already have the mathematics or for people who are not expected to care. Both leave you unable to predict what a model will do next, which is the only thing worth having.
The target here is different. After the five courses you should be able to watch a model do something odd and say why, from a mechanism you can name, and decide whether the fix is a setting, a different prompt, or giving it material it never had. Nothing is included that does not move you toward that, which is why these are short.
The five courses
They are in order because each one makes the next possible. Attention is hard to explain to somebody who still believes the model writes a sentence, and grounding is hard to explain to somebody who has not seen why a flat distribution still produces a confident paragraph.
- Course 02 In production 7 lessons · about 17 min
Attention, and how a model decides what matters
Course one treated the model as a box that turns text into a distribution. This one opens the box. What a layer does, why every token gets to look at every other token, and how a pronoun four hundred words later still lands on the right name. It is also where the costs come from, so it explains why context windows are expensive rather than just large.
- Course 03 In production 6 lessons · about 15 min
Where the ability comes from
A model is built in stages and each stage leaves a different mark. Pretraining on a great deal of text produces something that can continue anything and will follow no instruction at all. Later training on demonstrations and on human preference is what makes it answer you rather than continue you, and it is also where the hedging, the refusals and the relentless agreeableness come from.
- Course 04 In production 6 lessons · about 14 min
Prompting as control, not folklore
Most prompting advice is repeated rather than tested. This course separates the parts that change the output for a reason you can state from the parts that are superstition, using what the first three courses established. Worked examples on the jobs people actually have: extraction, classification, drafting, and getting a format you can parse instead of a paragraph you have to read.
- Course 05 In production 7 lessons · about 16 min
Grounding, or giving a model facts it does not have
The fix for invention is not a better instruction, it is better input. Putting the source material in the request turns answering from recall into reading. How retrieval works, why chunking decides whether it works at all, what a citation does and does not prove, and how to tell an answer that came from your documents from one that came from the model's habits.
How the tests work
Each course ends with ten questions, and seventy percent passes. The wrong options are the beliefs the course was written to correct rather than obvious filler, so a pass is worth something and a near miss tells you which lesson to read again. Every question explains its answer after you submit, whether you got it right or not.
Scores are kept in this browser and nowhere else. There is no account, no server and no leaderboard, which also means a different device or a cleared browser starts you at nothing. That is the trade for not asking you to sign up to read.