Turing Award Winner Sutton: "Synthetic Data Can't Scale AI"

Yet at a recent event introducing Oak Lab, a startup he co-founded with former student Kheram Javid, Sutton directly criticized one of the very scaling strategies the industry now relies on: synthetic data—generating large volumes of training data using models or simulators once human-created data runs out. Sutton flatly stated, "That's just a big mistake."

Why Large Language Models Have Hit a Wall

Sutton considers large language models to be both a good and a bad example of "The Bitter Lesson." They're a good example because they "drank in" the entire internet and genuinely improved in proportion to compute. They're a bad example because that's precisely where they hit a limit. The internet is finite, and the real world is "much bigger than everything we've stored on the internet." Sutton's diagnosis is that once the internet resource was exhausted, models came to depend even more heavily on knowledge that humans had already created—and this dependency is now holding the system back.

Asked whether synthetic data could break through this bottleneck, he offers two arguments. The first is the "Big World Hypothesis," formalized by Javid and long studied by the Alberta research group. The premise is that the world is infinitely complex—far larger than any mind or any agent. A simulation imitating this complex world, no matter how sophisticated, ultimately remains "microscopic" in scale. A small program can only create a small world, and that world inevitably diverges from reality—differing friction values, or inaccurate reproduction of a robot motor's movement, for example. Moreover, the world contains the inner minds of other agents, which is not something synthetic data can capture. As Sutton puts it, "There's no way to make synthetic data about other people's minds."

The second argument is the human bottleneck. Ultimately, a human must judge which synthetic data is good or bad. Javid gives the example of training a drone that navigates by echolocation like a bat—you'd first need to hire an expert in that field. This approach ends up being limited by the number of available human experts, and therefore doesn't scale. The same applies to self-driving cars trained via simulation: the gap between simulation and real roads must ultimately be closed by hand, by people.

An Alternative: Agents That Learn on Their Own

Sutton's proposed solution is to take humans out of that loop. Rather than relying on a fixed simulation model built by humans, an agent should learn its own world model and continuously revise it. "The simulator is something the agent builds itself," as he puts it.

Another problem he points to is that today's language models stop learning once training ends—"the weights never change again." Sutton says what's needed instead is genuine continual learning. In his view, this is simply what learning is: "all learning should be continual," while avoiding "catastrophic forgetting," in which old knowledge gets erased. Sutton's research group has proposed a technique called "Continual Backprop," published in Nature, as a partial solution to this problem.

Despite these criticisms, Sutton doesn't dismiss language models outright. He calls them "an amazing scientific achievement," while noting that they represent only "maybe 20% or a quarter" of intelligence as a whole.

Recent Trends Around Continual Learning

The idea that agents should evolve by accumulating their own experience isn't unique to Sutton. As covered in Tencent Hunyuan Unveils Five-Stage Roadmap for Verifying AI Agent "Self-Evolution", Tencent Hunyuan researchers presented an L0–L4 framework on August 12 for verifying whether agent self-evolution is trustworthy. Around the same time, Oumi unveiled a platform that lets models retrain themselves on production traffic, while Sakana AI proposed agents that improve themselves in the physical world as the next research frontier. Sutton's remarks align with this broader trend, but go a step further by declaring synthetic data—the industry's current go-to solution for data scarcity—fundamentally flawed as an approach.

Editor's Take

Sutton's remarks carry weight precisely because he is close to the theoretical father of scaling laws. "The Bitter Lesson" was the basis on which today's frontier labs came to believe that simply pushing compute would ultimately win out. That the same person is now taking direct aim at synthetic data—one pillar of today's scaling strategy—can be read as a sign of cracks forming even within the scaling faith itself.

The shift feels even more tangible across model generations. In the GPT-3 era, simply scraping more internet text kept driving performance gains. Now, with training text running dry, labs increasingly rely on a self-referential structure where models are trained on data generated by other models. As Sutton points out, the weakness in this structure is that filtering good data from bad still falls to humans. Data generation has been automated, but quality judgment has not—and that's precisely what he means by the "human bottleneck."

For teams in Korea, the practical lesson from this debate is clear. Scaling fine-tuning data with synthetic data remains a valid approach, but without automating the quality-verification process for that data, performance will plateau even as data volume grows. Teams currently running synthetic data pipelines would do well, per Sutton's diagnosis, to shift resources and headcount from the generation stage toward filtering and evaluation.

What happens over the coming months seems predictable. Rather than directly rebutting the criticism of synthetic data, frontier labs will likely roll out hybrid approaches that layer continual learning and agentic learning loops on top of language models. Recent work from Oumi and Tencent Hunyuan already points in that direction, and Sutton's own Oak Lab is likely to bring concrete results to this debate before long.