OpenAI has developed a research version of GPT-3 that is capable of searching the web, synthesizing data and citing its sources to generate more accurate answers to questions. Language models such as GPT-3 can be used for many different tasks but have a tendency to produce inaccurate results when dealing with complex real-world information. To tackle this issue, OpenAI taught GPT-3 how to use text-based web browsing techniques by providing it with open ended questions and details about the browser state. The model then issues commands such as “Search …” or “Quote: …” in order to search through websites and collect passages that are relevant for composing an answer. The model was fine tuned from GPT-3 using similar methods and augmented by training a reward system that predicts human preferences which were optimized further via reinforcement learning or rejection sampling approaches. To learn more about this project visit https://openai.com/blog/improving-factual-accuracy/.