1. A model learns patterns
Training adjusts a large network of numerical parameters so it becomes better at predicting patterns in data. It does not store a neat encyclopedia or develop human understanding in the ordinary sense.
2. Generation is prediction
A language model produces an answer one token at a time, choosing likely continuations based on the conversation and its operating context. Fluent wording can be produced even when a claim is wrong.
3. Tools extend the model
Search, code execution, databases, and connected applications let an assistant retrieve current information or take actions. Tool access increases usefulness and also raises privacy and security stakes.
4. Context shapes behavior
Your prompt, prior messages, system instructions, retrieved documents, and product rules all influence the response. The same underlying model can behave differently in different products.
5. Human judgment remains essential
Use AI to draft, explore, translate, organize, and challenge ideas. Verify consequential facts and keep accountable people responsible for final decisions.