Do Chatbots Just Need More Time to "Think"?

In this edition of Today in Science, can giving chatbots "time to think" improve how AI responds to hard questions?

This is an abbreviated version of Today in Science, our free daily newsletter that covers the latest cutting edge discoveries in science. Sign up to get the full edition of Today in Science in your inbox every day.

--Andrea Gawrylewski, Chief Newsletter Editor

Think a Bit More, Chatbot

The latest approach in developing state-of-the-art artificial intelligence is to slow chatbots down. Yes, slow them down. Called "test-time compute," the idea is to allow additional seconds or minutes to elapse between a user’s prompt and the program’s response while the AI double checks its own work. Using this tactic, some AI developers have seen a dramatic jump in the accuracy of chatbot answers.?

Why this matters: Test-time compute could improve AI accuracy and its capacity to tackle complex reasoning problems. The latest version of ChatGPT is o1, and OpenAI (the company that built ChatGPT) says this version is better at writing computer code and nearly 40 percent more accurate at answering Ph.D.-level physics, biology and chemistry questions. All thanks to test-time compute strategies.

What the experts say: This strategy grants machines the same grace we give people when they take an extra beat to tackle tough questions, says Aviral Kumar, an assistant professor of computer science and machine learning at Carnegie Mellon University. He thinks this could bring us closer to models with humanlike intelligence.?

Sign up here to read the full version of Today in Science!