The hallmark of popular generative artificial intelligence programs like ChatGPT is that they have a time cut-off regarding the facts they absorb. OpenAI recently updated its GPT-4 program to have access to data up until April 2023, an improvement from being trained on data only up to 2021. AI scientists are exploring ways to allow generative AI programs to access ever-changing data about pressing questions.
In a joint effort, Google and OpenAI published FreshLLM, a method to enable GPT-4 to use information from Google searches. This includes a new way to prompt a language model, called “FreshPrompt,” using results from a search engine. By incorporating the top search results from Google in the input prompt, GPT-4 improved its answers to questions involving timely information.
The team created 600 questions called FreshAQ, ranging from never-changing to fast-changing with “fresh knowledge.” In testing, GPT-4 and Google’s Pathways Language Model struggled with the questions but improved after using FreshPrompt. This technique significantly improved the accuracy of GPT-4, notably on questions of facts beyond 2022.
The authors noted the challenges of keeping FreshPrompt updated and expressed hope that the open-source community or automation by generative AI could aid in the process. FreshQA questions are meant to remain up to date, and FreshPrompt has been shown to surpass previous research that also uses search engine queries to augment language models.