WAIC 2026: Turing Award Winner Sutton Says LLMs Lack Native Intelligence, AI Enters 'Era of Experience'

On the opening day of WAIC 2026, Turing Award winner Richard Sutton delivered a keynote speech, directly asserting that current large language models (LLMs) lack 'native intelligence.' Sutton argued that the reasoning and understanding abilities exhibited by models are essentially mappings of statistical patterns in massive training data, not the models' own cognitive capabilities. This claim challenges the industry's prevailing optimistic narrative about LLM 'emergent abilities.'

Sutton further pointed out that the high-quality data resources supporting LLM performance are approaching their limits. Human-generated high-quality text and annotated data are nearly exhausted, and continuing to rely on the traditional data-driven model will face diminishing returns. This observation echoes recent research papers warning of a 'data wall,' but Sutton elevates it to the inevitability of a paradigm shift.

Based on the above assessment, Sutton proposed that AI is entering the 'Era of Experience.' In this new phase, AI systems will no longer primarily rely on static data provided by humans, but will dynamically accumulate 'experience' through autonomous interaction with the environment, simulated trial and error, and generation of synthetic data. This experience-driven learning is closer to how humans and animals learn, emphasizing continuous improvement through action and feedback.

Sutton used reinforcement learning as an example to illustrate the potential of experience-driven approaches: agents can autonomously discover strategies through extensive simulated interactions, even surpassing human knowledge bases. He suggested that future AI breakthroughs may come from designing experience generation mechanisms rather than larger static datasets. This view aligns with his deep background in reinforcement learning, but may underestimate the continued value of supervised learning in specific tasks.

As a pioneer in reinforcement learning and a 2018 Turing Award winner, Sutton's views carry significant weight in academia. In his speech, he cited his own research experience and industry trends, but did not provide specific experimental data or quantitative indicators. For example, the claim that 'high-quality data is approaching its limits' lacks direct evidence from public dataset sizes or model performance curves, relying more on empirical observation.

It should be noted that Sutton's research background may introduce positional bias: reinforcement learning is his core field, and emphasizing experience-driven approaches aligns with his academic interests. Additionally, current LLM research continues to explore ways to improve data efficiency (e.g., few-shot learning, data synthesis), which may partially alleviate the 'data wall' problem. Sutton's speech should be seen as a critical supplement to the mainstream paradigm, not a definitive conclusion.