Rich Sutton helped pioneer the technology behind today's AI boom. Now he thinks Big Tech's push to keep AI going is all wrong.
On an episode of Sequoia's podcast released Tuesday, the Canadian computer scientist and Turing Award winner said that the AI industry is headed in the wrong direction because of its reliance on synthetic training data.
"That's just a big mistake," he said when asked about using synthetic data as a way to keep scaling large language models. "Maybe it's the next big lesson."
Synthetic data is information artificially produced by algorithms or AI models rather than collected from original, real-world sources. It has become increasingly attractive as AI companies scour the internet and hunt for new sources of training data. Examples include computer-generated images of cars for self-driving training or fake bank records for AI fraud detection.
On some level, Big Tech agrees with Sutton.
Tech giants are going to increasingly extraordinary lengths to find more of the real thing. OpenAI has publicly sought large-scale, proprietary datasets that aren't readily available online to help train its AI models. Earlier this week, Google agreed to pay $10 million for internal data and software belonging to bankrupt Spirit Airlines, highlighting the growing value of proprietary, real-world information for AI training.
On Tuesday's podcast, Sutton said manufactured data isn't an adequate substitute for real data.
Human behavior is one such example, he said.
"There's no way we can have synthetic data for other people's minds," he said.
His preferred alternative is real experiential data: information an AI agent obtains by interacting with its actual environment, observing what happens, and continuously learning from the consequences.
"You can't get synthetic data on how the drone is going to interact with its environment in the physical world," Sutton said, adding that there are variables such as friction and wear in a robot's motors. "The world is infinitely complex, and any simulation of it is like, microscopic."
That has now become a startup thesis. Last month, Sutton and his former student Khurram Javed launched Oak Lab. The startup is developing agents designed to learn continuously from their own experience rather than relying primarily on large, pre-curated datasets. Oak Lab has not disclosed a funding round or investors.
