The Nerd Uncle › Models

Model, transformer, and the plumbing in between.

Three words get used as though they named the same thing. They are three different layers, and once you can tell them apart most AI news becomes readable.

31 OCTOBER 2026 · 6 min read

The reading room of the Bibliotheca Alexandrina. — The Bibliotheca Alexandrina. Photograph: Argenberg (CC BY 2.0) Photograph: Argenberg (CC BY 2.0)

A large language model does exactly one thing: given some text, it works out what is most likely to come next. Everything else — the conversation, the code, the apparent reasoning — is that one trick, run over and over, each answer fed back in as the next question.

A reading room holds the words; it does not know what they mean. Photograph: Argenberg (CC BY 2.0)

It does not work in words. Text is chopped into tokens, which are roughly word-fragments: common words are one token, unusual ones are several. This is why models are priced per token and why they occasionally miscount the letters in a word — they are not looking at letters.

The transformer is the shape, not the thing

People say “the transformer” as though it were a product you could buy. It is not. It is a design — a way of wiring the network up.

Google researchers published it in 2017, in a paper with the memorable title “Attention Is All You Need”.

Its idea is attention. When the model processes a word, it looks at every other word in the passage at once and decides which ones matter for this one. In “the bank was steep”, the word steep settles what bank means, and attention is the mechanism that lets the model notice that.

Attention lets every word consult every other word at once — the whole shelf, not one book at a time. Photograph: Argenberg (CC BY 2.0)

The older designs read left to right, one word after another, which made them slow to train because nothing could be done in parallel. A transformer looks at the whole passage simultaneously, which is why it could be scaled onto thousands of GPUs, and why scaling it turned out to work so well.

LangChain is none of the above

LangChain is not a model and not an architecture. It is a software library — plumbing — for wiring a model to the things around it: your documents, a search index, a calculator, a database, a memory of the conversation so far.

A model on its own has no hands. It can describe a search; it cannot run one.

So if you want it to look something up or remember yesterday, somebody has to write the code that carries things back and forth. LangChain is one popular way to write that code. There are others. None of them make the model cleverer.

The idea to take away

Three layers, and keeping them separate is most of the battle. The model is the thing that predicts. The transformer is the shape it is built in. The framework is the wiring that connects it to the world. When a company announces a breakthrough, it is worth asking which of the three it is actually about — a great deal of what gets announced is plumbing.

■