Richard Sutton is challenging one of the most important ideas in current AI development: that synthetic data can help large language models keep improving once internet-scale training data runs out. His view is direct. Synthetic data is not a path around the scaling problem because the world itself is too large, too complex, and too full of other agents to be captured by human-designed simulations.
Sutton’s argument matters because of who is making it. He is a Turing Award winner, one of the founders of reinforcement learning, the author of the field’s standard textbook, and the writer of the 2019 essay "The Bitter Lesson." In that essay, Sutton argued that the AI methods that win over time are those that scale with compute, such as search and learning, rather than systems built around human knowledge.
Why large language models are only a partial success
For Sutton, large language models show both the strength and the weakness of the Bitter Lesson. They are a success because they scaled dramatically with compute and absorbed the internet as training material. In his phrase, systems could simply "drink in the internet."
But that success also points to the limit. The internet is finite. The real world, Sutton says, is "massively bigger than everything we stored on the internet." Once models reach the edge of what can be gathered from human-produced text and other online material, the old problem returns: the system becomes dependent on human knowledge.
That dependency is exactly what Sutton’s long-running view of AI warns against. If progress depends on people choosing, curating, designing, or encoding what matters, then the method is constrained by human judgment. It may work for a while, but it does not scale in the open-ended way that learning from experience can.
The case against synthetic data
Asked whether synthetic data could solve this bottleneck, Sutton’s answer was blunt: "No, that's that's just a big mistake." The reason comes from the "Big World Hypothesis," formulated by Khurram Javeed, Sutton’s former student and cofounder at Oak Lab.
The hypothesis starts from a simple but demanding premise: the world is infinitely complex. Sutton describes it as "massively more complex than your mind than any agents any agent." A simulation, by contrast, is always much smaller. In Sutton’s wording, it is "microscopic."
That gap creates a practical problem. A small program can create only a small world. It may contain the wrong friction values. It may model a robot’s motor behavior inaccurately. It may leave out interactions that matter in the real environment. The result can be useful in a narrow setting, but it should not be confused with the world itself.
Sutton also points to a deeper limitation: the world includes other agents. Their internal states, intentions, and changing behavior cannot simply be manufactured as reliable synthetic data. As he puts it, "There's no way we can have synthetic data for other people's minds."
The human bottleneck remains
Javeed adds another objection. Synthetic data still requires someone to decide what counts as good data and what counts as bad data. That brings human expertise back into the center of the process.
"You need human experts who know what's a good data set and what's a bad data set for that approach to scale. So it is bottlenecked by humans."
The source gives the example of training a drone to move like a bat using echolocation. Before such a system could be trained well, people with the relevant expertise would need to define what matters. That makes progress dependent on the availability and judgment of specialists.
The same pattern appears in simulation-based training for self-driving cars. Even when systems are trained in simulation, humans still have to address the difference between the simulated world and reality. For Sutton and Javeed, that is not a small implementation detail. It shows why synthetic data does not escape the broader constraint.
- Synthetic data depends on simplified models of the world.
- Those models are built or judged by humans.
- Human judgment becomes the limiting factor.
- The real world remains larger than the training environment.
Learning from experience instead
Sutton’s alternative is to remove humans from the central training loop as much as possible. Rather than relying on a fixed simulation created by people, an AI agent should learn from its own experience. It should build its own world model and keep revising that model as it encounters new evidence.
He summarizes the idea with the phrase "Simulators they make themselves." The point is not that external simulations can never be useful. The point is that a system aiming for broader intelligence should not depend on a human-made simulation as the source of truth.
This connects to Sutton’s criticism of today’s language models. After training, he says, they stop learning: "Their weights never change." In his view, real learning should be continual. He goes further, saying "all learning is continual."
That creates another technical challenge. A system needs to keep learning without erasing what it already knows, a problem described as catastrophic forgetting. Sutton believes this can be addressed, partly through a method called "Continual Backprop" that his team published in Nature.
What this means for the next stage of AI
Sutton is not dismissing language models. He calls them an "amazing scientific breakthrough." But he also describes them as only "like 20% or a quarter of intelligence." That framing is important: in his view, current models are a major step, not the endpoint.
The broader implication is that AI scaling cannot be reduced to generating more data that resembles existing data. If the world is far larger than the internet, and if human-built simulations are too narrow, then the next step requires systems that can interact, learn, adapt, and correct themselves over time.
Oak Lab, the company Sutton founded with Khurram Javeed, is presented in this context: a move toward agents that learn from experience rather than systems that depend mainly on static training sets. The debate is not just about whether synthetic data is useful. It is about whether it can support the kind of open-ended learning Sutton believes intelligence requires.
His answer is clear. Synthetic data may fit some narrow purposes, but it cannot stand in for the real world. For Sutton, the future of AI depends less on fabricating more training examples and more on building agents that keep learning from the world they actually inhabit.