LLMs in the Wild

The Transformer series explains the calculation inside a model. These five articles explain the surrounding choices: what a model can see, what it learns from, how to ask it for something, and how to judge the result.

Read in order, or start with the question you have.

1. Three model architectures — What encoders, decoders, and encoder–decoders can attend to.
2. Prompting and context — Give a task, supply examples, and distinguish context from training.
3. Training data — How selection, filtering, and duplication affect what the model encounters.
4. Evaluating a language model — Interpret loss, perplexity, benchmarks, and small test sets.
5. When a model goes wrong — Examine unsupported claims, inconsistent answers, and mistakes outside familiar data.

Return to Transformers for the underlying computation, or the AI overview for other topics. Sources are in the bibliography.