Key concepts
- Model
- A system with parameters learned from data.
- Training
- Adjusting parameters using data and objectives.
- Inference
- Running a trained model.
- Token
- A representation unit—not exactly a word.
- Context window
- Information processed for an interaction; effective usable context may be less than nominal.
- Multimodal
- Handling text, image, audio, video, or combinations.
Many language models generate sequentially and outputs may vary. Fluency does not prove truth, currentness, valid reasoning, source access, or successful action. Models may omit constraints or invent plausible details.
Practice
Fictional gallery hours: Tuesday–Saturday, 11 a.m.–6 p.m.; closed Sunday/Monday. Ask a hypothetical assistant to recommend Sunday. Identify the contradiction and correct it.
Check your understanding
- Explain training vs. inference.
- Why do tokens differ from words?
- What evidence would verify opening hours?