A short reference.
FOR UNDERSTANDING WHAT IT IS
- Large language model: trained to predict and generate text
- Generative AI: produces content rather than classifying
- Training: how a model is made, once, before you use it
FOR UNDERSTANDING HOW IT BEHAVES
- Hallucination: confident false output, the central limitation
- Knowledge cutoff: the point after which it knows nothing
- Context window: how much it can consider at once
FOR USING IT WELL
- Prompt: everything you give it
- System prompt: instructions the operator set, which you do not see
- Temperature: how much randomness is in the output
FOR BUILDING WITH IT
- Token: the unit processed and priced
- API: how software connects to it
- Retrieval augmented generation: answering from your own documents
- Fine-tuning: adapting a model, which shapes behaviour more than adding knowledge
FOR JUDGING AND MANAGING IT
- Benchmark: useful for rough comparison, useless for your task
- Precision and recall: what it catches and what it misses
- Bias: unfair patterns, not removed by deleting a field
- Explainability: whether you can say why it decided
- Human oversight: a person with knowledge, time and authority to reject
