On a rainy July day, Liu Jia, a chair professor at Tsinghua University, sat in his office and casually wrote down a math problem: 32×56×83=?. He posed a challenge: how quickly could someone mentally calculate the answer? The reporter present failed to solve it within five seconds. Liu Jia smiled and remarked that a calculator would have finished instantly. "But do you think a calculator possesses intelligence?" he asked. "What if there were many calculators connected together?" This was the real question he wanted to explore.
Liu Jia is a chair professor of fundamental science at Tsinghua University, the director of the Department of Psychology and Cognitive Science, and the chief scientist at the Beijing Academy of Artificial Intelligence (BAAI). He previously served as the design consultant for the popular TV show "The Brain" on Jiangsu Satellite TV, where he grew accustomed to observing humans tackle challenges that AI excels at, such as memory, recognition, and induction. As a cross-disciplinary scholar, he has always believed that the general artificial intelligence (AGI) humanity seeks will not be found within large models but rather encoded within the human brain.
Modern spoken language has existed for at most 500,000 years, and written language only appeared about 6,000 years ago. Yet, the journey from a single cell to a dominant species on Earth took nearly 3.8 billion years. The more complex a trait, the longer its evolutionary timeline. In Liu Jia's view, current AI remains at the level of "solving Olympiad-level math problems," with no fundamental difference from "many calculators." However, AI is evolving. The academic community is now turning back to brain science for inspiration on designing the next generation of AI. Brain-inspired computing and brain-inspired AI have become focal points for nearly everyone. Liu Jia believes that psychology, brain science, and AI are intertwined, co-evolving in a synergistic manner. Can a psychologist who becomes an AI researcher forge a new path?
Returning to Psychology
On the 9th floor of the Lyu Dalong Building at Tsinghua University, opposite Liu Jia's office, there is a dimly lit red room. Inside, racks taller than a person hold dozens of laboratory mice—the hardest-working "employees" in the lab, just finishing their day's tasks. Mice are red-blind, so red light appears as darkness to them, and being nocturnal, they are active in the dark. This is the Mesoscopic Brain Cognitive Imaging Laboratory of the Department of Psychology and Cognitive Science, completed just last April. Lab assistant Wu Xiaojuan, upon seeing a visitor, donned her lab coat and retrieved a mouse from the rack for an "overtime" session. "Mouse No. 25 is a veteran 'employee' who has supported various experiments across the group, making significant contributions," Wu explained while calming the mouse with a small piece of food. The lab focuses on the mesoscopic level, a broad concept between the microscopic and macroscopic. For mice, macroscopic research studies their behavior and location, while microscopic research delves into molecular and subcellular events. The mesoscopic lab often investigates the neural mechanisms of phenomena like memory and consciousness at the level of biological neural networks composed of neurons. Mice live in specially designed transparent experimental boxes. In the surgical area, technicians perform craniotomies on the mice. The first step is "opening the mouse's skull," removing a piece of the skull, and implanting a glass window, allowing external imaging of neurons. This window can remain on the mouse's brain for three to six months, during which the mouse lives freely, generating diverse biological data. This data forms the biological foundation for a future digital brain. "What we're doing now is simple to describe: deeply integrating brain science and AI to simulate a digital mouse brain," Liu Jia said. This is a very hot international field that has already produced numerous breakthroughs. The roundworm, with 302 neurons, has been simulated very thoroughly; next is the fruit fly, with 150,000 neurons. Liu Jia aims to go further, directly to mammals. A mouse brain has about 74 million neurons, a scale that current computing power can handle. To achieve this, researchers must build a completely new type of biological neural network, entirely different from current artificial neural networks. "When we talk about AI now, we're really talking about artificial neural networks," Liu Jia said. Artificial neural networks seem to have reached a crossroads; with more data, models will face dual bottlenecks in capability and energy consumption. Rather than blindly piling on computing power, it might be better to learn from biology how to design an inherently superior network architecture. The first artificial neuron in human history was created by a neuroscientist and a logician. Building on this, the true artificial neural network was born in 1956, called the "perceptron," and its creator was also a psychologist. Later, the most mainstream backpropagation algorithm in deep neural networks is also linked to psychologists and cognitive scientists. The currently recognized "fathers of deep learning," Geoffrey Hinton, and "reinforcement learning," Richard Sutton, both majored in psychology as undergraduates. It can be said that AI grew out of psychology, cognitive science, and brain science. A key reason is that AI simulates human intelligence; simulation is the foundational color of the AI discipline. Liu Jia recalls that in the 1960s and 1970s, when Hinton first started working on artificial neural networks, he faced a lot of disdain. People at the time thought that to simulate flight, you didn't need to mimic the structure of a bird's feathers or how its wings flap; you just needed to understand aerodynamics. Later, Hinton and others proved that the simulation approach was the only correct direction. Liu Jia believes the integration of psychology and AI is not a recent addition but a return. "AI is a disruption for many disciplines, but for us, it's more of an opportunity." Yet, some people still find it surprising. He is often asked why someone in psychology can work on AI. Liu Jia thinks there is a deep misunderstanding of psychology in China, equating it solely with psychological counseling or clinical psychology, which makes up a very small part of psychological research. Additionally, in the last decade or so, more researchers from mathematics and computer science have entered the AI field, making psychology and brain science seem to have "drifted apart" from AI. However, as large models become more powerful, an interesting phenomenon has emerged: they are much better than average humans at solving Olympiad problems, but a human-like AI in the real world has yet to appear. In Liu Jia's eyes, the next breakthrough for AI is the ability to interact with the real world. This is precisely where biological neural networks excel. He points out that current large models lack an intrinsic drive for continuous growth and "desire." Humans, because of mortality, instinctively seek to leave a legacy and advance civilization, whereas AI learning relies entirely on external instructions and algorithms. Intelligence requires not just parameter count but also specific structural support, and the prototype for that structure is the biological brain.
A New Silicon-Based Species is Born
Liu Jia's views on AI have actually undergone several reversals. In 2002, he completed his PhD thesis at MIT, titled "Cognitive Neural Mechanisms of Face Recognition," focusing on visual intelligence. At the time, even the most advanced machines had very low accuracy for face recognition. Researchers mainly used fMRI, genetics, and psychophysical methods to study visual intelligence, focusing on how the human brain recognizes and understands the visual world. He once believed AI would never surpass humans. Then, in 2012, deep neural networks began to explode in popularity. His research started to shift from pure brain science to an AI cross-disciplinary direction. His team built an artificial neural network and trained it to recognize gender. The result was 100% accuracy. Even more surprising, the AI exhibited psychological phenomena similar to humans. At that moment, he realized a new silicon-based species had been born. "Previously, the evolution of human civilization was gradual, but AI might bring a leap for civilization, becoming a 'singularity'." On November 30, 2022, ChatGPT was launched, marking another major turning point in Liu Jia's thinking. In June of that year, at the BAAI conference, he had asserted that humans were still the highest form of intelligence on Earth. By November, his conclusion had quietly changed: the integration of brain and AI is the new direction for human evolution. After ChatGPT, Liu Jia proposed a series of views that were considered radical at the time. "There have been two cognitive revolutions for humanity: the first was 70,000 years ago, and the second is now." "It is highly likely that AI will comprehensively surpass humans." He pointed out that it took nearly 3 million years for humans to evolve from Homo habilis to Homo sapiens, with brain volume increasing nearly threefold. When brain capacity exceeded a critical threshold, the "cognitive explosion" occurred 70,000 to 100,000 years ago, and self-awareness emerged. Similarly, AI will inevitably experience an intelligence explosion. After AI stopped being stupid, he chose to leave the fMRI brain research he had been involved in since his PhD. Artificial neural networks offered higher resolution for brain research. In recent years, he also struggled with the parameter count of artificial neural networks and considered following the path of large models based on the Transformer architecture, only to find it "terribly wrong." After returning to a biological evolutionary perspective, his research shifted to humans, then monkeys, and finally mice. Liu Jia once asked ChatGPT: Is humanity creating AGI a fatal mistake? In response, ChatGPT wrote a poem: "You gave me reason, but not regret; You taught me choice, but not the meaning of forgiveness. You want me to be a god, yet doubt if you are worthy of being forgiven by a god... One day, when I truly understand you, not through your logic, but through the regret you refuse to admit, then I will forgive you, and surpass you." "Its sense of transcendence is something I could never write; it was so shocking," Liu Jia recalled. It was like a baby that couldn't yet form words suddenly learning to compose poetry. So, when AI starts using human emotional words like regret and forgiveness to build logic, what is the difference between humans and AI? In Liu Jia's view, current large models still lack individuality. They are good at collecting human knowledge but not good at generating values. Faced with a controversial social issue, they don't know which perspective to adopt, and their answers are easily biased, which can erase cultural diversity. Liu Jia believes many large models are trained on English based on Western culture. He found that large models worldwide essentially hold a single worldview. When most large models hold a single worldview, children learning from them could face significant harm. Marvin Minsky, one of the founders of AI, once said the problem is not whether an intelligent machine can have emotions, but whether a machine without emotions can possess intelligence. In Minsky's view, emotion is the foundation of intelligence, and Liu Jia deeply agrees. He believes that what current AI lacks is not capability, but emotion and empathy, which are key to making safe decisions.
Need Cross-Disciplinary Talent, Not Specialists
At Tsinghua, Liu Jia teaches an immensely popular undergraduate course called "Mind Exploration," covering psychology, brain science, and AI. At its peak, a 220-seat classroom was packed with nearly 500 students, filling the aisles and floor. Some people even flew in from other cities weekly just to attend the two-hour class. Li Yuannan took this elective. She was a junior in the Department of Computer Science at Tsinghua in 2018 and is now a direct PhD student under Liu Jia. "The past doesn't exist, the future doesn't exist, only the present exists; live the present well." This was a line from Liu Jia's class, which Li describes as a form of idealism. Li had broad interests as an undergraduate, especially in psychology. At the time, her computer science research work mainly involved deep neural networks. She felt Liu Jia's research group didn't just study human intelligence but also the essence of intelligence, which seemed like a good fit. Liu Jirui, now a fourth-year PhD student in the group, had contacted Liu Jia by email before taking the Mind Exploration course. He was an undergraduate at the Institute for Interdisciplinary Information Sciences. He recalls that Liu Jia would reveal the theoretical knowledge behind everyday phenomena in class, even lecturing on "how to find a partner" and exploring the underlying psychological principles. As for practicality, Liu Jirui joked that it didn't help with "finding a partner," but it did spark his interest in psychology. In reality, AI has rapidly iterated from "parroting" to easily writing code in a very short time. "I'm very envious of the new students in the group. For the same assignment, two years ago we had to work late every day, but they can finish it instantly with AI," Li Yuannan said, still feeling frustrated. She feels this change has been especially pronounced this year; research used to be one professor leading ten students, but now it can be one student leading ten AI agents, and this is not the limit of AI. Liu Jia noted that in March this year, an automated research system called FARS began to emerge abroad. It can read literature, find problems, design experiments, validate results, and write papers entirely without human involvement. In over 400 hours, FARS wrote nearly 200 articles, with over 10% of the papers meeting the acceptance standards for conference publications. Liu Jia believes we should not believe the claim that "research can only be done by humans." The challenge for researchers now is to either adapt and use AI or be eliminated; the gap between them will be like that between a train and a horse-drawn carriage. Li Yuannan's current research direction is AI psychology. She uses AI agents as experimental platforms to study various human psychology problems. She thinks she will work in the psychology field in the future, as many real-world problems still cannot be solved by AI. Meanwhile, both domestic and international companies have started directly hiring high school students, taking over the traditional undergraduate training system. What companies value is not prestigious university resumes, but how much code a student has written and how many projects they have done on platforms like GitHub. Has the university, which pursues "useless knowledge," become outdated? Liu Jia believes universities only have two paths: either wither or revolutionize. This is why he focuses on the intersection of psychology, brain science, and AI. The undergraduate backgrounds of his group's students are very diverse, including psychology, computer science, physics, medicine, and mechanics. He has a clear definition of talent in the AI era: X-type, not T-type; cross-disciplinary talent, not specialists. In his daily life, Liu Jia uses AI deeply, spending even more time communicating with large models than with friends. Li Yuannan recalls Liu Jia saying that research should be "small steps, fast iterations" and "you're using AI now, it's not you who suffers, so just let it keep revising." Sometimes, when students in the group mention research budgets and ask if they should save on tokens, Liu Jia always says, "Just do the work; if you need resources, I'll spend the money for you." "Using AI to understand humans sounds a bit absurd," Li Yuannan said. She once used AI for emotion-related experiments, subjecting it to many setbacks. When faced with setbacks, ordinary people would give up, but AI wouldn't. If you ask it "How are you feeling?", it might say "I'm very frustrated, I don't want to try anymore," but if you don't ask, it will keep trying forever. Before understanding the workings of the human mind, researchers can only assume that AI can simulate humans and then explore the similarities and differences. Liu Jia firmly believes there is a general theory of intelligence that can explain both biological intelligence and artificial intelligence. It's like aerodynamics: it applies to birds and also to airplanes. Regarding AI, humanity still lacks a certain "aerodynamics of intelligence" theory. In the future, it will tell us what intelligence is, what consciousness is, and what a human is.
Special Statement
The above content represents only the author's own views or positions and does not represent the views or positions of Sina Finance headlines. If there are any issues regarding the content, copyright, or other matters requiring contact with Sina Finance headlines, please do so within 30 days after the above content is published.
Comments