Is synthetic data sending AI in wrong direction?
Reinforcement learning pioneer Rich Sutton says real experiential data, not synthetic data, is AI's true path forward
Rich Sutton spent decades helping build the reinforcement learning foundations behind today's AI systems. Now the Canadian computer scientist says the industry's growing reliance on synthetic data to keep scaling large language models is a serious misstep, one he suspects could become "the next big lesson" for the field.
Speaking on an episode of Sequoia Capital's Training Data podcast released Tuesday, Sutton was asked directly about using synthetic data to sustain AI scaling.
"That's just a big mistake," he said. Synthetic data, information generated by algorithms or AI models rather than drawn from real-world sources, has become increasingly attractive to companies running short of fresh material to train on, covering everything from computer-generated images of cars for self-driving systems to fabricated bank records for fraud detection.
The scepticism from Sutton comes amid similar efforts that have been undertaken in the industry. OpenAI has been looking into getting hold of large proprietary datasets that are not easily accessible on the Internet, and Google recently agreed to pay $10 million for internally sourced data and software of the bankrupt airline Spirit Airlines.
Sutton argues synthetic data cannot substitute for certain categories of real information, human behaviour chief among them.
"There's no way we can have synthetic data for other people's minds," he said. He favours what he calls experiential data: information an AI agent gathers by acting in its actual environment and adjusting based on what happens next.
Even physical systems resist simulation, he added, pointing to variables like friction and motor wear that no drone simulation can fully capture. "The world is infinitely complex, and any simulation of it is like, microscopic," he said.
Last month he and former student Khurram Javed launched Oak Lab, a startup building agents designed to learn continuously from direct experience rather than depending primarily on large, pre-assembled datasets.
