Turing Award Winner Richard Sutton at WAIC 2026: 'AI Is Still Weak and Unreliable'

The main stage of WAIC 2026 in Shanghai on Friday hosted what felt like two different conversations about artificial intelligence — except both speakers were talking about the same thing.

The first was Richard Sutton, the 2024 Turing Award winner and one of the foundational figures in reinforcement learning. He didn’t hold back.

“Some people fear AI, and they also have reasons to exaggerate its importance — especially how fast it’s actually progressing,” he said.

Sutton’s central argument: the industry has been conflating intelligence with computation. “Most of what we call AI capability today is really computational capability. We need to separate these two definitions,” he said. Current AI systems, in his view, are mostly repackaging human knowledge rather than generating any of their own. “They’re weak in many aspects. There are still problems in their thinking processes. They are not that powerful.”

He went further, correcting a persistent industry misconception about the Turing Test. “Turing never used the term ‘Turing Test.’ He called it the imitation game,” Sutton said. The popular belief that Turing equated human-like behavior with machine intelligence, he argued, is a misreading.

Sutton’s most pointed claim came when he declared that the “era of human data” has reached its limit. Most machine learning today, he explained, is knowledge transfer — moving what humans know into machines. “That method has reached its ceiling. Many high-quality data sources have been exhausted, and generating genuinely new knowledge is something this approach cannot do.”

He called for a shift to what he termed the “experience era” — where AI systems learn from their own interactions with the world. He pointed to AlphaGo and AlphaProof as early examples, then showed a video of a baby playing with toys to illustrate what “behavior-driven input” looks like. “A baby doesn’t have a static dataset,” he said. “Its behavior determines what data it collects.”

The diagnosis for today’s large language models was blunt. “They have no reward signal. They have no way of knowing whether an action is good or bad,” Sutton said. “They basically cannot distinguish what’s real from what’s fake. Current AI is not powerful enough. I personally consider it weak and unreliable, because it generates too many hallucinations.”

Then he added: “But it’s also very useful.”

The next speaker was Yin Qi, chairman of both StepFun and千里科技 (CheliTech), a company building AI operating systems for vehicles. His tone could not have been more different.

Yin traced his 15-year journey in AI entrepreneurship and declared that 2026 marks a genuine inflection point. “AI has progressed from executing tasks lasting seconds to working independently for tens of hours. The industry is standing at the foot of the AGI mountain,” he said.

He laid out what he called “three structural shifts”: new systems, new carriers, and new networks. The centerpiece was his concept of an “Agentic OS” — an operating system for AI agents where capability is determined jointly by the model and the OS that runs it. Hardware design, he argued, needs to move from “human-centered” to “human-machine symbiosis,” where computers, phones, cars, and robots become different bodies for the same agent.

He also proposed an “A2A network” — a protocol layer connecting humans and agents, where agents would have their own identities and credit scores, enabling them to “autonomously find partners, organize collaboration, and complete transactions.”

Yin acknowledged the risks. “When agents enter the physical world, it’s not just a capability leap — it’s a restructuring of order,” he said. The industry needs answers to who agents act for, who bears responsibility for their actions, and how to ensure identity trust and behavior traceability.

“Ours is not a future where machines replace humans,” he said. “It’s a future where humans and intelligence co-evolve. When agents enter the physical world, everyone’s potential will be amplified tenfold.”

The tension between the two talks is what made the session memorable. Sutton and Yin were describing the same destination — experience-driven agents operating in the real world — but disagreeing fundamentally on the timeline. Sutton says the path the industry is on (stacking parameters on static human data) is hitting a wall. Yin says the inflection point is already here.

If Sutton is right, the companies currently scaling language models through raw compute face a fundamental challenge. If Yin is right, the next three to five years will rewrite the hardware and operating system stack.

WAIC 2026 put both views on the same stage, back to back. One came from the field’s most respected theoretical founder. The other came from one of its most aggressive builders. The answer probably won’t arrive for another three to five years. But for one afternoon in Shanghai, the two futures of AI sat next to each other, disagreeing.