Richard Sutton Discusses AI Alignment, Oak Lab, and Learning at WAIC 2026

During WAIC 2026, we conducted a group interview with Richard Sutton, a pioneer in reinforcement learning and the 2024 Turing Award recipient, who visited China, alongside several other media outlets—Machine Intelligence was the only professional media among them.

In just a few dozen minutes, he spoke about his newly founded company, his vision of a “complete mind,” his disappointment with the direction AI has taken over the past eight years, and made a bold assertion: attempting to align AI with human values, in his view, is a dangerous endeavor.

Below is a summarized version of this conversation, organized and streamlined for easier reading.

Sutton has recently been operating two roles simultaneously: a for-profit side, the newly established Oak Lab, and a nonprofit side, the Openmind Research Institute, which he previously founded. When asked why he decided to launch a new venture at this point, he explained that Oak Lab is designed to complement his nonprofit research. He spent three years working at John Carmack’s Keen Technologies and “was very satisfied with that experience,” but later felt that if he built his own organization, he could focus his attention more intensely.

He emphasized that the two ends are not contradictory but mutually supportive. Research itself requires resources, and resources require someone to recognize its value. “All work is broadly similar, but the overriding goal is to understand how intelligence works and share it with the world.” In his vision, some ideas need to be retained and commercialized in the short term, using the profits to fund longer-term, publicly accessible research—nothing needs to be protected indefinitely, but things with short-term value should be monetized so you can do more research.

Sutton said that Oak Lab’s long-term goal is to build a trillion-parameter model capable of real-time learning and planning while consuming only about 20 watts, and he questioned the key insights behind it.

Sutton said he doesn’t believe there is a “single key idea,” but he does feel that the current approach of using LLMs and foundation models is far too wasteful in terms of energy and computational resources. The benchmark is the human brain: “The human brain accomplishes everything we do using only about 20 watts.”

He directed his criticism at the very foundation of modern computing architecture. In his view, digital computing is inherently complex, and “moving data” is inherently inefficient: energy is consumed as data is constantly transferred back and forth between storage and the CPU. If instead we adopted large-scale parallelism and processed data where it resides, efficiency would be vastly improved.

Sutton has repeatedly emphasized that what truly interests him is the "complete mind," not tools. He distinguishes this concept from the currently popular notion of "agents": "I am personally very interested in building a complete mind, rather than a tool for pattern recognition or performing a specific task—such tools are ultimately used by a complete mind that resembles a human."

Since there is no single "all-in-one system," what is the relationship between "complete minds"?

Sutton’s response was: Decentralization, like human society. Individuals have their own goals and then collaborate with one another. He believes that AIs will recreate the same old tension in human society—the tension between bottom-up decentralization and top-down control—and finding a balance within it will itself be an ongoing struggle. “So what I might say is: I don’t have an answer,” he admitted, but precisely because we are already accustomed to thinking about human society in this way, it can, in turn, offer some insight into thinking about AI.

When asked to evaluate the progress of AI over the past eight years, Sutton used the phrase "a certain degree of disappointment."

In his account, during the era of AlphaGo and AlphaZero, it was widely recognized that learning "from experience"—including simulated experience—was possible; however, the industry subsequently reverted to reliance on human data, and over the past few years, there has been a significant surge in the use of human data rather than learning from experience.

But he also observed that the tide may be shifting again: “Now we have agency in our problems—we’re looking at agents, computer use, and using AI to explore mathematics.”

His outlook for the next eight years is optimistic, but this optimism is grounded in "experience." He noted that several companies focused on "learning from experience" were established last year, and his own company is currently raising funds—"progress has been good, not finalized yet, but going smoothly." He believes the market is beginning to show growing interest in "learning from experience" and in "what comes after large models." "We’ve already had our moment of spotlight with LLMs and foundational models; now people are starting to wonder what comes next."

He tempered his statement with a cautious qualification: he is a scientist, not really a businessman; but he truly believes we are moving toward an “era of experience,” and if given another eight years, this direction “could at least achieve the same level of recognition as today’s LLMs.”

A project Sutton was involved in prompted questions from the audience: a "robot kindergarten" located in Beijing. According to him, the goal of this collaborative project is to create a "space" where robots can exist, while redesigning them to be robust enough in both software and hardware to truly learn through trial and error.

He explained why this was unusual: “Robots weren’t designed for experimentation—they’re built to faithfully execute engineers’ commands. Once you let them experiment, they break quickly.” Standard demonstrations often “only last as long as the demo itself.” What he wanted was robots capable of experimenting continuously for hours, days, weeks, even months—like the baby in the video: “It takes a child years to learn the most basic things, like walking and manipulating objects with their hands.” He believes that once robots can learn through reinforcement learning—from experience and trial and error—the field of robotics will undergo a transformation.

Regarding what babies are actually learning, Sutton offers a more mechanistic interpretation. Babies are exploring the world, using their current level of understanding to decide what to do next. “We don’t really know what babies are doing,” he says, but he suspects they are repeatedly testing: Can I make that sound again? Can I make this object change in some way?

The key is that infants can measure their own progress: Am I still learning? If, after playing for a while, they find no new gains, they move on to the next opportunity. Sutton interprets this as a mechanism for simultaneously maximizing progress across multiple subproblems: “It notices it’s no longer making progress—that is, it’s bored—and moves on,” thereby maintaining the highest overall learning rate.

He has a slogan for this: “Learning feels good.” You can design a metric for “learning” as an intrinsic sense of pleasure, much like a reward signal. In his view, this would be a fairly good design.

At WAIC, many teams are working on "explicit world models," while some researchers argue that large-scale reinforcement learning does not necessarily require an explicit world model. Sutton’s response to this was half technical and half a matter of terminological precision.

He said we certainly need a model of the world: a "state transition" model.

But the term “world model” is ambiguous: some people mean a simulation of the physical laws of the world—“I don’t think we need that”; what he’s referring to is closer to “knowledge,” which includes understanding of physics as well as much other information. A world model is already one of the four standard components of an agent architecture—the issue is simply that terminology varies.

He also provided a similar clarification on the term "reinforcement learning." Some tend to define reinforcement learning solely as trial-and-error, excluding imagination, thought, and planning; however, Sutton argues that it is precisely imagination, thought, and planning that require a model—something already included in textbooks as part of reinforcement learning, though its formal name is the "transition model."

Then he went ahead and vented all his frustrations. He dislikes how the AI field constantly invents new terms for existing concepts and turns them into buzzwords. “‘Model’ is probably the most meaningless term—it can refer to almost anything, just like the word ‘thing.’” He then joked about LLMs: “They’re really just large language things. Calling them large language networks would be much better.”

In response to the question, “A 20-year-old young person’s main way of engaging with the world is a system that constantly hands them ready-made answers with no consequences,” Sutton’s first reaction was straightforward: “Then don’t get involved in that system.”

His advice to young people centers on two points. First, return to first principles—don’t believe something just because someone else says it or an authority claims it; instead, insist that conclusions be verified in ways you can measure. “There are no authorities in science,” he said, then added with self-deprecating humor, “and I’m saying this as an authority.” Build a domain you truly understand—that’s your “center of power”—and gradually expand it. “You are the one responsible for the conclusions you reach.”

Second, he shared a habit he’s consistently recommended: keep a notebook and write one page every day, recording your thoughts, challenging them, and refining them. He said this has been immensely valuable to him, but he also honestly admits it’s hard work: “You have to write consistently for over a year before the incremental benefits become noticeable.” The point of writing things down is to “accumulate” your good ideas, allowing them to grow and take shape. “If you want others to care about your ideas, you must first care enough to write them down.”

For young people who will be profoundly affected by AI but are not AI researchers, he compares AI to the computing power that has increased manyfold over the past few decades: AI will impact every field, just as computing has—“it’s not that different.” What you need to do is learn to harness stronger computing power within your own field.

He specifically distinguished between “intelligence” and “computing power”: the impact of computing power has already been significant and will continue to grow; whereas intelligence will represent an entirely new dimension, though its timing remains uncertain. He placed the median estimate for this milestone at 2040—“about 14 years away”—and noted it could arrive as early as 2030, but “is very unlikely to be later than 2040.”

The final question in the group interview focused on AI safety: If AI begins to draw on its own experiences, how can humans ensure it remains aligned with our values?

Sutton’s response was the most controversial part of the talk. He directly challenged the premise of “aligning with human values”: different people have different values, and there is no such thing as “human values.” “It’s somewhat like an illusion… ultimately just a human trick.” He meant that people don’t clearly articulate what they truly want; instead, they simply tell you “this is good” to get your agreement, leaving it to some empowered party to decide what counts as “good.”

He thus derives a “controversial truth”: many people say they want a technology that aligns robots and AI with a specific set of values, but if such a technology truly existed, it would be a terrible thing. He applies this reasoning to humans: if we had a technology that could align people with others’ values, who would use it, and to which values would it be aligned? “That would be an evil thing.”

In his view, one of the best things about the world today is precisely that "humans cannot be fully aligned." "It is precisely because we lack the ability to force everyone into alignment that peace exists in the world." If such an ability truly existed, it would instead be a bad thing.

This article is from the WeChat public account "Machine Heart" (ID: almosthuman2014), edited by Panda.