Five reading orders through the library, and how to pick one

The library is arranged on ten shelves by subject, which is the right arrangement for somebody who already knows what they want. The commonest visitor does not. They want to understand the model everybody is talking about, or to stop bouncing off the formulas, and no shelf heading says either of those things.

So there are now five reading tracks, each named after what you can do at the end of it rather than after the subject it covers. Forty-one of the eighty courses sit on one, in order, with a line on each position saying why it is there.

Read any page here without stopping at the symbols

Eight courses, starting with Reading the Notation and ending with change and accumulation. This is for the reader who keeps stalling on a formula and has concluded the problem is them.

It is the track we would put almost everybody on first, because everything else in the library gets easier afterwards. The sums, the subscripts, the argmax and the expectation are a vocabulary of about thirty marks, and nobody is ever taught them as a vocabulary.

Tell in advance which approach survives a million rows

Eight courses, from How Long It Takes through indexes, ordering, memory and profiling. The outcome is the useful one: looking at a design and knowing, before you build it, whether it falls over at scale and roughly where.

Build a model by hand, then find the same parts inside every large one

Eight courses, beginning with the simplest model that learns and ending at unlabelled structure, with gradient descent, cross entropy and backpropagation in the middle. The order is deliberate: the parts are introduced on a model small enough to do on paper, and then recognised inside the large ones.

Explain the model everybody is actually using

Eight courses. It starts at How Text Becomes Numbers and runs through cutting text into pieces, attention, how the next word is picked, where training data comes from, training from scratch, adapting a model you did not train, and making one smaller.

This is the track for the person who can use a language model perfectly well and would like to be able to say, out loud and correctly, what it is doing.

Put a model in front of people and know whether it is working

Nine courses, and the one most often skipped. Writing what the model reads, finding things by meaning, tools, calibration, refusals, reading a benchmark honestly, judging a system rather than a model, running an experiment, and watching the thing after launch.

Half of the mistakes we see in production are in this track, and none of them are modelling mistakes.

How to use them

1. 1

   Pick by the sentence you want to be able to say

   The track titles are written as outcomes on purpose. Read the five, and take the one whose ending you actually want.

2. 2

   Start at the beginning unless the first course is obviously beneath you

   Each position says why it sits there, so skipping is an informed decision rather than a guess.

3. 3

   Read the free lesson of the first course before committing

   The first lesson of every course in the library is readable without an account, along with the opening of every lesson after it.

A track is an editorial judgement rather than a fact about the courses, which is why it lives apart from them and can be rewritten the day a better order becomes obvious. If you think one of them is in the wrong order, we would like to hear it.

The tracks are at getbasislab.com/library/tracks. If none of them is what you came for, that is what writing a course for you is for: say the goal, and the order is worked out from where you actually are.