On July 17 in Shanghai, at the WAIC 2026 main forum, Turing Award winner Sutton said something that surprised many: today's AI is rather weak and unreliable. What I believe this statement lays bare is a confusion we have long held onto ourselves: in recent years, we have quietly taken "computation" for "intelligence." Error! File not specified.
I. Computing Well Is Not the Same as Thinking
In recent years, working with enterprises on AI adoption, one feeling has grown ever clearer: when people see a large-model agent produce a result by following instructions, they feel it can think. But take the process apart and you find it stops at the "execution layer." Layer one: pattern matching is not understanding. Sutton himself put it plainly: a large model is only large-scale pattern recognition, whose core is to reorganize humanity's already-known knowledge and then deliver it. It rearranges old knowledge, but discovers no new knowledge. Much of the "smartness" we see is a replay of statistical regularities in massive data: it knows "what words often follow what words," yet not necessarily "why the world behind the words is this way." Layer two: an uncertain result comes from probability, not from cognition. The uncertainty in a large-model agent's answer is rooted in the fact that it runs on next-word prediction probabilities. Ask the same question twice, change the wording slightly, and the result may differ. This is not it "deliberating" — it is probability fluctuating. Taking an "uncertain result computed from probability" for a "judgment with a mind of its own" is the easiest mistake to make this round. Layer three: it lacks goals, and has no feedback from the environment to check itself. When Sutton says AI is "weak and unreliable," largely he means exactly this: it does not know whether its computation is right or wrong, because it has not been placed in a real environment that can give it rewards and punishments. Most agents, in their basic principle, have not escaped the framework of "probability and statistics" — their operating essence is a series of tensor computations. Erik Hoel of the Turing Institute laid it bare: so-called "reasoning" is mostly just next-word prediction disguised as a chain of thought — a fixed large model large enough can, mathematically, be replaced by a gigantic lookup table. So "computing accurately" and "thinking through deeply" are probably not the same thing. What agents can do today is deliver humanity's existing knowledge in a faster way; what they have not yet learned is to autonomously discover knowledge humanity does not have. This is also why many enterprises, having adopted agents, see nicer efficiency numbers, yet when they truly need it to "think up a new approach," still must rely on people. Why are we so prone to equate computation with intelligence? I think half of it is that the demos are too smooth, and half is that we too badly want to believe "the age of intelligence has arrived." Enterprises buy agents, see them produce a plan in three seconds, and feel "now we don't need to keep so many people." But in truly complex scenarios, the plans agents produce are often pretty yet impossible to implement — because what they lack has never been computing fast, but knowing why they compute, and whether the computation is right.
II. From the Data Era to the Era of Experience
For the past decade or so, AI has relied on "data." All the text, images, and code collectable from human history are fed to the model, letting it learn from static datasets. But Sutton judges this path is reaching its end: many high-quality data sources have already been used up. Piling on more data under the old paradigm makes it hard to grow truly new knowledge. What is more troublesome is that "data exhaustion" is not a temporary phenomenon. Humanity's several-thousand-year body of text is only so large, and the high-quality, never-used public data is rapidly running out. Going forward, we must either turn to synthetic data or to people's real experience. Synthetic data, gone full circle, is still old knowledge; what is truly fresh lives only in human interaction. So where to next? He says it is the "era of experience": intelligence no longer learns only from humanity's static data, but from its own first-person perspective and real interaction with the world. AlphaGo playing Go does not memorize the game book; it plays itself and learns from winning and losing. A child playing with toys, with no ready-made dataset, grows its understanding of the world through "touching," "falling," and "trying." Sutton has given many examples: a soccer player's shot on goal, a baseball batter's swing, even a bird or a lion — without experience, there is no such thing as intelligence. In Harmonized Intelligence I once wrote: human experience, especially that layer of "tacit knowledge" hidden in the mind, is the "experience" raw material AI lacks most today. Many of our enterprises doing AI are used to first making the process "explicit": writing the veteran's judgments into SOPs, turning customer-service scripts into knowledge bases. This step is not wrong. But the real difficulty is not the first step, but the next two: how to turn scattered experience into rules AI can execute; how to keep these rules continuously calibrated in real scenarios. So while many anxiously ask "will the agent replace people," I instead feel what truly deserves close watching is the other end: how we turn decades of human experience into something a system AI can catch and amplify.
III. Experience, Not Just Experience: But
, even if the agent truly steps into the "era of experience" Sutton describes, able to interact with the world and learn from its own experience, I think humans still have a moat it cannot catch up to for now: experience. Experience and experience are not the same thing. Experience can be "recorded interactions"; experience is "being alive itself." For example: when a glass cup falls off the table, the human mind automatically predicts "it is going to break" — this is a causal world model that has run in the body for many years; whereas a large model seeing "a glass cup falls off the table" is more like predicting "the words 'broke' often follow." One is a driver with hands on the wheel; the other is a voice that only reads the navigation. The driver is actually driving; the voice only reads a script. So the era of experience is worth looking forward to, but the wall of experience is one agents cannot break through in the short term. Experience can be recorded, replayed, and learned; experience grows from the body — it cannot be uploaded, nor downloaded. Where exactly does the difference between AI and humans lie? I think at least three things are structurally lacking in today's AI. One is the body. It feels no pain, no hunger, no instinct to pull back when a hand is scalded. All its "knowing" is symbolic, not felt. Two is intrinsic motivation. Its goals are almost all given by humans; rarely does it have a drive like "I want to figure this out" that grows from its own body. Completing a task is not the same as it "wanting" to complete it. Three is long-term memory and metacognition. Beyond a few dozen steps its error rate clearly climbs; it is hard, as humans are, to weave decades of events into a web and call on it by instinct at critical moments. Seth once said: consciousness is like a gift exclusive to life; AI can simulate "doing," but will never obtain the subjective experience of "being alive." However human-like its writing, it is not human. Technology determines the ceiling of AI's capability, while humanity determines the floor of AI's behavior. We push AI higher with all our might, but what we must ultimately guard is the boundary: which things must be decided by humans; which judgments must be made by humans. Within the boundary, harmonized intelligence.
In Closing
Data intelligence — AI is already very strong; experiential intelligence is only just beginning; and lived-experience intelligence may be humanity's last, and most worth-defending, moat. In these days when everyone is celebrating computing power, I often recall a line: what you can do and AI cannot is your experience, your judgment, your agency. Do not let a computing-power race blur what makes a human human. Technology will keep running forward; the boundaries of capability will keep being widened. But what decides how steadily and how far we can go has never been how powerful the technology is, but what kind of foundation we have held onto.
