OpenAI has demonstrated a regressive pre-model of the WebGPT neural network that surfs the Internet, writes articles with links to sources, and can use search queries not only to find relevant information, but also to form a smart and meaningful answer.
WebGPT was trained to answer questions based on ELI5 data on the popular Reddit forum – a section where some users ask questions and others answer them as simply as possible. After a period of training, the neural network began to generate responses that in half the cases turned out to be better created by humans.
In the future, the company plans to improve the results by adding a feature to assess the reliability of data sources in the new neural network.