what an llm actually is

LLM stands for large language model, which is to say a large model of language. Thank you, very helpful. It might be the least informative acronym in a field with a lot of competition, so let me come at it backwards and start with what it is not.

It is not a database. When you ask it something it is not looking anything up: no drawer, no row, no record with your answer already sitting in it. It has read more than any person could and it remembers none of it the way you remember a phone number.

It is not understanding you either. It has no idea what a wall is. What it has is an extremely good statistical sense of which words tend to follow load bearing, which is exactly why it can hand you a sentence that sounds like structural advice and is wrong, without anything registering that a mistake occurred.

And here is the part I cannot tidy away, because this would be a better note if I could stop at three: it is not just autocomplete either. Guessing the next word, at that scale, produced things nobody put in on purpose. It writes working code. It holds an argument across paragraphs. Not one of the people who built it can tell you why that showed up. Though people argue about whether it was ever a jump: mark the answers right or wrong and you get a cliff, mark how close they got and the same models climb smoothly.

So I work with it the way you would work with a very fast, very well read colleague who has never once said I do not know. I check. Every time I have not checked I have paid for it later, and the bill always arrives after you have told someone it is fine.

We built something that guesses, and the guessing turned out to be enough for most of what we ask of it. I still cannot tell whether that is a fact about the machine or about the questions.