Reinforcement Learning Pioneer Rich Sutton Says Synthetic Data Is Steering AI in the Wrong Direction

Rich Sutton helped build the mathematical backbone of modern reinforcement learning. Now he's telling the entire AI industry it's chasing the wrong thing.

Speaking on an episode of Sequoia Capital's Training Data podcast released Tuesday, Sutton was asked point-blank about the industry's growing use of synthetic data to keep scaling large language models. His answer was blunt.

"That's just a big mistake," he said, adding he suspects it could become "the next big lesson" for the field.

Synthetic data is exactly what it sounds like: information generated by algorithms or other AI models rather than pulled from the real world. Think computer-generated images of cars for training self-driving systems, or fabricated bank records for fraud-detection models. As companies run out of fresh internet text to scrape, synthetic data has become the industry's patch for keeping the scaling machine running.

Big Tech isn't ignoring the problem Sutton is describing, even if it's not fully agreeing with his solution. OpenAI has been hunting for large-scale proprietary datasets that aren't available online, according to Business Insider. And Google recently agreed to pay $10 million for internal data and software from bankrupt Spirit Airlines, a deal that underscores how much real-world, non-public information is now worth to AI labs, Business Insider and The News International both reported.

Sutton's core objection is that synthetic data can't substitute for certain categories of information, especially anything involving how actual humans think and behave.

"There's no way we can have synthetic data for other people's minds," he said on the podcast.

What Sutton Wants Instead

Sutton's preferred alternative is what he calls experiential data: information an AI system gathers by acting in the real world, observing the consequences, and adjusting continuously. He argues this applies even to physical robotics, where no simulation can fully capture variables like friction or motor wear in a machine over time.

"The world is infinitely complex, and any simulation of it is like, microscopic," he said.

According to a more detailed writeup from BigGo Finance, Sutton's critique goes beyond just synthetic data. He thinks today's AI industry is fixated on static, frozen models that stop learning once training ends. In the same podcast conversation, Sutton and his former student Khurram Javed argued that real progress requires training systems from scratch with algorithms built for continual learning, not bolting new tricks onto existing large language models.

Sutton, who published his widely cited essay "The Bitter Lesson" in 2019, has spent years arguing that AI progress comes from methods that scale with raw computation, like search and learning, rather than from hand-fed human knowledge. On the podcast he restated that view directly: "Don't be distracted by human knowledge as AI traditionally has been many times. Instead, focus on learning methods that will scale with computation."

He calls the industry's current approach "weird" for a specific reason. As he put it, before the current AI boom, nobody needed to specify "continual learning" as a special category, because "all learning is continual." In his view, treating today's frozen, pre-trained models as the default is the anomaly, not the exception.

Sutton isn't purely dismissive of large language models. BigGo Finance notes he calls LLMs "an amazing scientific breakthrough" precisely because they proved the Bitter Lesson's point about scaling with computation. His disagreement is about what comes next, not that LLMs worked.

The Skin in the Game

Sutton isn't just theorizing from the sidelines. Last month he and Javed launched a startup called Oak Lab, built around agents designed to learn continuously from direct experience instead of leaning primarily on large, pre-assembled datasets. The company has not disclosed a funding round or named investors, according to Business Insider.

Sutton has a commercial stake in convincing the industry, and its money, that his alternative approach is the right one. That doesn't make his technical argument wrong, but his position is also a startup pitch, not just an academic observation.

It's also worth considering the strongest case for the side Sutton is arguing against. Companies like OpenAI, Google, and Anthropic have built genuinely useful, revenue-generating products on the frozen pre-training paradigm Sutton criticizes. Synthetic data has let labs sidestep real bottlenecks in licensing, privacy, and the sheer scarcity of fresh human-generated text online. Dismissing that approach as a dead end assumes continual-learning systems can be made safe, controllable, and commercially viable at scale, something nobody, including Oak Lab, has yet proven in production.

None of the four major AI labs Sutton and his collaborators named in the podcast, OpenAI, Anthropic, Google, and others building on frozen LLMs, have publicly responded to his comments as of this writing. Whether any of them shift resources toward continual-learning architectures, or whether Oak Lab can demonstrate its approach works at any meaningful scale, remains the open question the rest of the industry will be watching.