Reinforcement Learning
Training by trial, error and reward rather than by examples. It is how models are tuned to be helpful and to refuse things they should refuse.
Related terms
Conversational Search
Searching by asking a full question, often as a follow-up to a previous one, the way you would ask a person. It favours pages that answer plainly over pages stuffed with keywords.
OAI-SearchBot
The crawler OpenAI uses to answer live searches in ChatGPT, separate from GPTBot. Blocking one does not block the other, so a site can be absent from ChatGPT search while still being read for training.
Prompt Engineering
The craft of writing instructions that get reliable results from an AI — being specific, giving examples, saying what to avoid. Less mysterious than it sounds, and mostly the same skill as briefing a new colleague well.
Deep Learning
Machine learning using neural networks with many layers. The depth is what lets a system learn complicated things — recognising a face, translating a sentence — without anyone specifying how.
Context Window
How much text a model can hold in mind at once, counted in tokens. Anything beyond it is forgotten, which is why very long conversations start to lose earlier details.
Content Negotiation
When the same URL serves different formats depending on what the visitor asks for — HTML for a browser, Markdown for an assistant. It lets one address serve people and machines without either getting a compromise.