Rich Sutton warns AI industry against reliance on synthetic data

Turing Award winner and AI pioneer Rich Sutton has warned that the industry's increasing reliance on synthetic training data is ‘a big mistake.’ Speaking on a Sequoia Capital podcast, the Canadian computer scientist argued that artificially generated information cannot adequately substitute for real-world data.

Sutton noted that while synthetic data is becoming more attractive as companies exhaust internet-based sources, it fails to capture the complexity of the physical world or human cognition. He stated, ‘There’s no way we can have synthetic data for other people’s minds,’ and emphasized that simulations cannot fully predict how agents, such as drones, will interact with real environments.

In response to the scarcity of high-quality data, tech giants are pursuing proprietary real-world sources. OpenAI is seeking large-scale, non-public datasets, and Google recently agreed to pay $10 million for the internal data and software of the bankrupt Spirit Airlines.

Sutton advocates for ‘experiential data,’ where AI agents learn by interacting with and observing their actual environments. He recently co-founded Oak Lab, a startup focused on developing AI agents that learn from direct experience rather than massive, pre-curated datasets.