Turing Laureate Sutton at WAIC: Today's AI Is "Weak and Unreliable" — Time to Move from the "Era of Human Data" to the "Era of Experience"
At WAIC 2026, 2024 Turing laureate and RL pioneer Richard Sutton argued today's LLMs merely relay human knowledge, lack a reward signal, and can't tell true from false — calling for an "era of experience" driven by agents' own interaction.
A keynote that dismantled the hype, line by line
At the WAIC 2026 main forum on July 17, 2024 Turing laureate and reinforcement-learning pioneer Richard Sutton gave a short talk that cooled the AI frenzy almost sentence by sentence. He opened by noting that some exaggerate AI's importance — especially how fast it's improving — and that the field routinely conflates two things: intelligence and computation. "Most of this is computational power; we must separate the two definitions."
Core claim: still in the "era of human data"
Sutton argued current systems "mainly use the power of human knowledge and hand it back to us; they cannot discover knowledge of their own" and "in many ways are quite weak." He also corrected a widely misused term: Turing never proposed a "Turing test" — he described the "imitation game," and "behaving like a human" was not, in Turing's view, a test of AI.
His central claim: we remain in the "era of human data" — most machine learning transfers knowledge from humans to machines, but "this approach has hit its limit; many high-quality data sources are used up, and generating new knowledge is something this paradigm cannot do."
The proposal: an "era of experience"
Sutton called for an "era of experience" — AI needs new data from an agent's own experience interacting with the world, "from a first-person view." He cited AlphaGo and AlphaProof, and showed a video of a baby playing with toys to illustrate "letting behavior decide the input": a baby has no static dataset; its actions determine what data it collects.
His critique of LLMs was blunt: they "have no reward signal, no way to know whether an action is good or bad," and "basically cannot tell true from false." His conclusion: "Today's AI is not strong enough; personally I think it is rather weak and unreliable, because it produces many illusions." — then he added, "and at the same time it is very useful."
A "subtle standoff"
The next speaker, StepFun and Qianli Technology chairman Yin Qi, was far more bullish: in 2026 model capability is crossing a key threshold, moving from tasks lasting seconds to working independently for tens of hours, "standing at the foot of the AGI summit." One splashing cold water, one sounding the charge — the two views formed a telling standoff at the same forum, capturing the real epistemic tension in AI today.