LLMs in the Wild · Part 1 / 5

Three shapes of a model

You have met two shapes — the writer and the reader. Here is the whole family, including the one that translates.

You’ve met two models so far. A writer (Book 01) and a reader (Book 02).

They’re closer than they look. Same blocks, same attention — one rule of who-looks-at-whom tells them apart. There’s a third shape too. Here’s the whole family.

One family, three wirings

• The decoder reads left-to-right and writes the next word — that’s Book 01, the model you chat with (GPT, Llama, Claude).
• The encoder reads both directions to understand — that’s Book 02, BERT.
• The encoder-decoder reads one whole sequence and writes a different one — translation, summarization. The classic is T5.

Try it: start on Decoder and watch one stack write the next token. Flip to Encoder-Decoder — now a reader feeds a writer: English in, French out.

That third shape is built for conditional generation: the output is steered by an input you hand it, not free-form continuation.

What this book is about

This whole book is about the decoder — the generative chat models. But seen from the outside.

Book 01 was the mechanics: what happens inside one guess. Here we step back and look at the model as a thing in the world — a foundation model that gets made from data, prompted into tasks, graded on benchmarks, and sometimes goes wrong.

Three shapes, one family. Next, we meet the decoder from the outside — starting with how you talk to it.