In the face of the surging AI boom, a Turing Award laureate bluntly calls AI "weak."
On July 17th, at the main forum of the 2026 World Artificial Intelligence Conference, Turing Award winner, "father of reinforcement learning," and University of Alberta computer science professor Richard Sutton stated plainly that current AI is "relatively weak and unreliable." He argued that the paradigm of large models trained on static human data is no longer sufficient, and artificial intelligence is about to enter an "Age of Experience" centered on first-person interactive experience from intelligent agents.
Sutton clearly expressed support for the concepts of global AI governance cooperation, win-win scenarios, and building partnerships proposed at the conference's opening ceremony. In his view, the breakthroughs of large models in language, image, and video generation have indeed spawned new industries and brought real economic value. However, many people confuse computing power with true intelligence. "Current AI is essentially large-scale pattern recognition at its core. It organizes existing human knowledge and delivers it to users, but lacks the ability to autonomously discover new knowledge. There are still obvious flaws in the complete chain of reasoning."
Over the past century, the academic community has had different definitions of intelligence: William James, the father of psychology, defined intelligence as "the ability to use different means to achieve the same goal." The widely known Turing Test of "acting like a human" is actually a misunderstanding—Turing himself never mentioned the term "Turing Test." The original concept was the "imitation game," and he did not believe that "resembling humans" was the ultimate criterion for AI.
The dictionary defines intelligence as "the ability to acquire and apply knowledge and skills"; Marvin Minsky, the father of artificial intelligence, proposed that intelligence is "the ability to achieve goals through computation."
Sutton offered his own definition: intelligence is the ability to achieve goals through behavioral adaptation.
He proposed that academia should establish a unified "science of the mind," encompassing all forms of intelligence in humans, animals, and machines, to find the common laws of different intelligences—all intelligence involves taking action over time to achieve goals. In the future, more intelligence will be machine intelligence, and reinforcement learning is the starting point for this unified science of the mind.
"We are still in the era of human data. All AI is trained to predict the next word in human text or predict human-labeled tags, followed by fine-tuning by human experts. Essentially, it's about transferring human knowledge to machines." He clearly pointed out that this path has reached its limit: high-quality, publicly available human data is about to be exhausted, and this paradigm fundamentally cannot enable AI to generate truly new knowledge.
Next, AI will enter the "Age of Experience," requiring a brand new, sustainable, and growing source of data—not pre-prepared static datasets, but first-person experience gained by intelligent agents themselves through interaction with the world. "This data consists of high-speed signals exchanged back and forth between the agent and the environment: what action it took, what perception it received, what reward it obtained. These signals define the agent's goals and are also the fundamental way humans and animals learn."
Whether it's AlphaGo defeating human Go champions through self-play, or the recent AlphaProof achieving good results in the International Mathematical Olympiad, both follow this path of experiential learning. At the forum, Sutton played a video of a baby playing with toys: when a baby learns, there is no pre-prepared dataset. Its actions determine what input it receives. After playing with one toy, it moves to the next. The data it generates itself precisely matches its current cognitive level and learning needs—this is the most essential form of experiential learning.
Whether it's a soccer player instantly judging the ball's trajectory, a baseball batter deciding on a swing in a fraction of a second, the hunting behaviors of birds or lions, or even real-time conversation between people, all are essentially this type of high-bandwidth, real-time interaction based on first-person experience, not reliant on pre-memorizing existing human data.
In Sutton's view, the essence of intelligence is constructing this interactive loop of perception and action: "The true intelligence of an agent is reflected in its ability to predict and control its own stream of experience. Without first-person experience, there is no intelligence to speak of."
He believes that current large language models have fundamental limitations: they have no built-in reward signal, they don't know what behavior is good or bad, and essentially have no goals of their own. They also cannot truly distinguish truth from falsehood because they lack first-person experience to verify whether their predictions are correct—they only know what has appeared in the text, but not what actually happens in the real world. In contrast, experience-based AI has rewards and goals, and can validate its judgments through real-world feedback. This is true intelligence.
Long before the birth of the artificial intelligence discipline, Turing proposed that what humanity needed was a machine capable of autonomous learning from experience. "This idea is very basic and very profound, but it has never become the core of modern AI until now." He quoted Victor Hugo's famous saying, "An idea whose time has come is irresistible," indicating that the era of AI learning from experience is arriving.
Sutton believes that there is currently a lot of hype and fear surrounding AI in the industry, but it must be acknowledged that present-day AI is still weak and unreliable. It can hallucinate and output misinformation. However, it is already useful enough to spawn new industries and become accessible to everyone.
"Current AI is not yet true autonomous intelligence. We have not yet entered the era of superintelligence, nor are we at the stage where intelligence comprehensively augments humanity. But a massive transformation is underway. This transformation should be driven by humans, and its fruits should belong to humanity." Sutton concluded his speech by saying, "Welcome to the Age of Experience."
Comments