I Understand Most of the Words I'm Saying: An Introduction to How LLMs Work

“I am so clever that sometimes I don’t understand a single word of what I am saying.” — Oscar Wilde, “The Remarkable Rocket”

The speaker of that line is a firework. Wilde puts it in the mouth of a vain rocket who spends a whole fairy tale explaining his own importance to anyone who can’t get away, then explodes where nobody can see or hear him. His stick lands on a passing goose, and his last words are “I knew I should create a great sensation.” I’ve kept him in mind while writing this series, which is about large language models, a subject that attracts a lot of proverbial rockets.

This is the introduction, and it’s arriving three posts late, which should tell you something about the planning (if you’re reading this later on, pretend it was all on time. I won’t tell). A modern LLM stacks linear algebra on statistics, then on a supply chain of very expensive chips and a network that has to behave itself for weeks at a time, something hard to do under many circumstances. Not many people understand every layer of that in real depth. I do, except in almost every way. I come from networks and systems, so the hardware is home turf, and I can follow most of the rest with a map and some patience. The math is where I slow down: I understand it the way a tourist understands the local language, which is to say well enough to order dinner, with a lot of gesticulation, but not well enough to successfully argue about the bill.

The series is written for someone who reads more technical publications in general, and can tolerate the occasional equation without losing their mind. That will still leave some people out, because there’s a floor on how far something this complicated can be simplified, and I’ve spent some time down there myself, most of the time if I’m honest. I’m doing the best I can with the sources in front of me, and when I get something wrong, which the odds favor, tell me, and I’ll fix it if I can.

The whole series, in reading order. The first three are up, and the rest should arrive about once a week, unless something catches fire:

Read them in order, since each one leans on the last, or hell, skip to whichever one you came for. I suppose nobody’s checking.

I had planned to close this on Wilde too, with a line I’d have sworn was his. I went to check it, as one does, and as far as anyone can tell, it isn’t: Quote Investigator traces it to Ernest, a character in J. M. Barrie’s The Admirable Crichton , who keeps repeating it to anyone nearby in the hope that someone will laugh. So that’s me admitting the limits of my own knowledge in the introduction, while getting my closing quote wrong on the first try (I’ll admit, I haven’t read this particular work by Barrie). I’m keeping it, credited to the right man, because it suits this series better than anything Wilde could have offered. I understand most of the words I’m saying, and I’ll try to tell you when I don’t.

“I’m not young enough to know everything.” — Ernest, in J. M. Barrie’s The Admirable Crichton (1902)

Part one is waiting at the top of the list. Bring a pencil. It may not help, but you’ll at least have something familiar to fidget with.