Reading level
OpenAI does something new: it takes GPT-3 and teaches it to use a browser. Not a graphical one, a text-mode version: click a link, scroll, search Google, copy a passage into citations.
The model, called WebGPT, is trained by watching humans do the same thing, and then rewarded when answers are good (RLHF).
It's the first serious prototype of an assistant that searches before answering, with source citations. The pattern is everywhere today: Bing Chat, Perplexity, ChatGPT with browsing, Gemini grounding, Claude with web search. All descend from here.
Companies
OpenAI
Tools
WebGPT
Tags
OpenAIWebGPTBrowsingRetrievalRLHF
Sources