OpenAI introduced WebGPT, a model that searches for the answer to a question on the Internet. WebGPT combines information from several sources and generates a response text.
Language models such as GPT-3 are useful for many practical tasks, but tend to provide false information when performing tasks that require specific knowledge about the world. To solve this problem, OpenAI trained GPT-3 to use a text-based web browser.
The model works in a similar way to how people search for answers to questions on the Internet – sends search queries, clicks on links and browses web pages. The model is provided with an open question and a brief description of the browser state, and it must submit commands such as “Search…” and “Find on page: …”. Thus, the model collects excerpts from web pages, and then uses them to compose an answer.
WebGPT was trained to answer questions from ELI — a dataset of questions asked by Reddit users. After training, the model began to generate answers, which in 56% of cases are preferable to answers written manually by people.
The accuracy of WebGPT’s responses turned out to be comparable to people’s responses. To verify this, WebGPT was tested on TruthfulQA, a benchmark designed to test the objectivity of models.
WebGPT is superior to GPT-3 in TruthfulQA. In the future, OpenAI plans to improve the model by adding a function for evaluating the reliability of data sources.
a summary of Herman Rapaport, Tools for Reading Poetry