Wang Qian secures the largest robotic financing deal at the start of the year.

Long time no see. In this episode of "Blue Hour", we've invited Wang Qian, the founder of Independent Variable Robotics, who has just announced securing 1 billion yuan in financing.

There are reasons for wanting to chat with Wang Qian. Of course, it's a bit worldly because of the large amount of financing he's obtained. But also, among this wave of embodied and humanoid robot entrepreneurs, he's a special figure.

Born in 1988 - neither earlier nor later, this birth year meant that for a long time, he didn't have the same smooth sailing as his post - 90s peers. In 2007, he enrolled in the Department of Electronic Engineering at Tsinghua University for his undergraduate studies. He started researching AI in 2009, right when AI was in a trough. Deep learning didn't experience a boom until after 2014. When he pursued his Ph.D. and switched to the robotics field, the technical path of deep reinforcement learning he was leading had just reached a bottleneck. Even when the wave of embodied and robot enthusiasm came, the Independent Variable Robotics he founded didn't initially shine brightly, and financing was far from smooth.

If a person wants to make a big splash but always misses the industry's prosperous period by a few years, they must have a strong desire to seize the opportunity to change the world. Moreover, he worked in a quantitative fund for two years, made money, and broadened his horizons. In my view, this desire seems even more pure - last year, when the industry had advanced to the stage of order wars, Independent Variable Robotics still wasn't in a hurry to commercialize. When someone asked him if he wanted to be the DeepSeek of embodied brains, his answer was that he wanted to build a company like OpenAI. I can't evaluate the technology, but anyway, he has the ability to make investors believe in him.

I have another curiosity about Wang Qian. As a person experiencing contemporary life, I don't really understand the tech entrepreneurs at the forefront of the storm. If these people determine the future direction, it's necessary to figure out the blueprints in their minds.

Wang Qian's childhood dream had nothing to do with robots. He wanted to become Einstein. This dream later evolved into using AI to study physics and then into developing intelligent enough robots to build machines. In short, the goal is to drive exponential growth with advanced technology and productivity. His worldview is based on statistics and probability, built on uncertainty. However, he believes that social life is a different field, following a completely different set of logic.

This is, of course, a rather typical view, but is it the truth? People always think that the progress of natural science is the product of human intelligence, but Marx said it was wrong. He said that social life has its foundation, and natural science has another foundation, which is simply a lie. At least in Marx's view, the two share a common foundation, which is human historical life.

In my understanding, business and industry have promoted the progress of natural science, and science and technology have further shaped our worldview. Regardless of how AI and robots will develop in the future, humans are living more and more like AI and robots. Like being programmed, we complete a mechanical existence, with our brains running at high speed but getting farther from our hearts. I don't know where this will lead in the future. Unfortunately, due to limited time, I didn't finish discussing this issue with Wang Qian.

If, according to the convention of this column, I need to assign a color to the guest, I'd like to choose titanium cyan blue. It's a chemically synthesized color with stable hue. A vivid dark blue with a thick metallic texture and a cold - looking shine. It's very much like the impression Wang Qian gave me: he has a strong sense of certainty about what he's doing, so a flimsy color won't do. Also, although he's very polite in his speech and behavior - I shook hands with him twice in my Shenzhen office - the pride of a smart person still shows from time to time. He's very confident in his team's technical strength, and probably domestic peers are not in his comparison range. There's nothing wrong with this. I hope he always maintains such spirit.

No one in the field of AI holds a deterministic worldview

Liu Yanqiu: Many of the robot entrepreneurs in this wave graduated from the Department of Electronic Engineering at Tsinghua University. You studied electronic engineering at Tsinghua University for your undergraduate studies. Why did you switch to the Department of Biomedical Engineering later? Many remarkable people start studying biology in the late stage of their careers because they become interested in the mystery of life. What was your reason for changing majors?

Wang Qian: Actually, I've always wanted to work on AI since middle school. Even earlier, like in primary school or even younger, I originally wanted to study physics and almost went to the Department of Physics at Peking University. But I gradually realized that doing physics now is different from a hundred years ago. In today's mathematics and physics fields, it's almost impossible to reach the forefront in ten years. Even for a genius who might enter university at 14, graduate with a Ph.D. at around 25 or 26, their career might end at around 35 or 36. So, in total, they only have about ten years for core research. I thought that in another hundred years, mathematicians and physicists might not have a "career" at all. It takes longer and longer to reach the field's forefront, and the requirements for human intelligence are getting higher and higher. Eventually, maybe no one can handle this.

So, in high school, I decided that I had to work on AI. If humans can't do something, machines can. It's like driving a ten - thousand - ton ship. Just rowing hard won't work. The key is to build a good machine to drive the ship.

I was in the class of 2007 in the Department of Electronic Engineering at Tsinghua University. At that time, the most mainstream way of AI research was statistical learning. The AI field was extremely cold at that time, and no one cared about it. The research on deep learning and neural networks that we're familiar with now hadn't emerged yet. So, during my undergraduate studies, I always wanted to work on AI, but I happened to catch the coldest period of AI, and no one paid attention to this direction. At that time, everyone was doing statistical learning, which had various benchmark tests, but the performance improvement in these tests each year was only 0.1%. I thought that this direction was stuck, and even if we spent a hundred years, we might not make a breakthrough. So, we had to find some paradigm shifts. My core idea at that time was to borrow the human neural network mechanism and apply it to AI models.

Liu Yanqiu: So, you were actually interested in AI from the beginning. You just chose to study AI by borrowing the neural network from biology. In essence, AI has always been your main focus. When you thought about using AI to promote subject research, did you have a specific proposition to study? Or was it just an abstract idea?

Wang Qian: I just wanted to study physics. Physics can be understood as the "Theory of Everything". For example, why is Newton called the "law - giver of nature"? Because the laws he discovered are, in a sense, the most fundamental laws of the universe. I initially wanted to become a physicist like Newton or Einstein. Of course, before that, I also liked philosophy and mathematics, but I thought that physics might be the field where humans can get closest to the truth of the universe.

Liu Yanqiu: Recently, I read an article that said the basic worldview of contemporary society is based on Newtonian mechanics, or at least deeply influenced by it.

Wang Qian: I don't think so. Since the 20th century, the new worldview based on quantum mechanics is very different from Newton's worldview.

Liu Yanqiu: But don't you think that the mechanical worldview of Newtonian mechanics still dominates our lives?

Wang Qian: I don't think so, at least not in the fields I'm familiar with. Take the AI field as an example. Everyone talks about probability, and no one says "must". Before the rise of neural networks, the main thing people did was statistical learning, and the core of statistical learning is probability. This is actually a milestone in AI development. After the emergence of statistical learning, people realized that this might be the essence of the world. So, now all people working on AI hold a statistical worldview, a random worldview. No one holds a deterministic worldview anymore.

Liu Yanqiu: What exactly do you mean by the random worldview? Maybe my understanding is incorrect. For example, in quantum mechanics, the wave - particle duality states that whether an electron is a particle or a wave is affected by the observer. It goes through a relational process to present the state we see. Is this the same as what you're talking about?

Wang Qian: In quantum mechanics, it's said that an object has a 50% probability of being here and a 50% probability of being there. This is its logical expression. When we work on AI, the logic is the same. For example, when doing mobile phone positioning, when I see a picture with a mobile phone in it, I can't be 100% sure where the mobile phone is. Instead, there's a probability distribution. For example, there's a 10% probability it's here, a 20% probability it's there, and a 50% probability it's in another place. You can't get a 100% accurate estimate of the mobile phone's position from a noisy signal environment.

Actually, humans also think in a probabilistic way, but we don't realize it. We always think "I'm very sure the mobile phone is here", but that's not the case. So, when we work on AI, we follow this logic. All the equations are not deterministic equations but random equations. The variables x described in them are essentially random variables.

Liu Yanqiu: As I understand it, a worldview is the way I think the world works, and I use this set of logic and rules as a guide to view everything. For example, in Newtonian mechanics, everything is calculable, which leads to a rational and predictable way of thinking. How do you think the statistical way of looking at the world, which focuses on probability, will affect your view of contemporary life?

Wang Qian: I think the connection between human society and the physical world is not that strong. Human society doesn't follow Newtonian mechanics but only "narratives". Because human cognitive ability is limited and can't handle all the information, we usually understand human society and the world through narratives. However, narratives are extremely simple, abstract, and easy to be tampered with and modified. But precisely because of these characteristics, they can unite people. So, I think human society is mainly united by narrative logic, which has nothing to do with Newtonian mechanics and may also have nothing to do with quantum mechanics. It's a completely independent system.

Liu Yanqiu: OK, let's get back to your experience. Your master's thesis was one of the early works to introduce the attention mechanism into neural networks. How did you come up with this mechanism at that time? Later, Google further studied it and dominated the current Transformer architecture. Is this a big regret for you?

Wang Qian: Yes, at that time, I thought that the reason we believed we could create AI was that there was already an existing intelligent system in front of us, which was humans themselves. Since we couldn't make a breakthrough in the AI field for a long time, why not see how humans achieve intelligence? But when I entered the laboratory in 2009, in the three main departments of the School of Information Science and Technology - the Department of Electronic Engineering, the Department of Computer Science, and the Department of Automation - I couldn't find a single teacher working on neural networks. Since no one was researching neural networks from an AI perspective, I thought that maybe I could start from a biological perspective to see how neural networks work. That's why I later switched to the Department of Biomedical Engineering to study neuroscience.

The term "deep learning" only emerged in 2008. Since we wanted to work on neural networks, the core idea was to transfer the neural mechanisms of humans and animals to the models. At that time, we thought about which neural mechanisms were crucial. We believed that the attention mechanism must be one of the cores because it's directly related to human consciousness. There's a common metaphor that consciousness is like a theater, and the spotlight shines on the part you're aware of. We thought that this "spotlight" directly corresponded to the attention mechanism. So, I focused on researching the attention mechanism at that time.

Three papers, including ours, Google's, and ETH's (Swiss Federal Institute of Technology Zurich), were the earliest achievements in this direction, around 2014. Now, people are familiar with the attention mechanism mainly because of the Transformer architecture. This architecture was developed by Google's team through continuous research until 2017. It's really a pity because the architecture we proposed at that time was actually closer to the current Transformer than Google's later solution, but I couldn't continue this research.

Liu Yanqiu: Along the AI path, how did you switch your research direction to robotics during your Ph.D. studies?

Wang Qian: Around 2014 and 2015, the first wave of AI enthusiasm came, and the "Four Little Dragons of AI" emerged. But I thought that these AI technologies could mainly be applied in fields like security, and I wasn't very interested in that. So, I wanted to find a new direction. The image field wasn't suitable, and the language field was too difficult. After much thought, I realized that robotics might be the fastest and most useful application direction for AI. Since I was going abroad to pursue my Ph.D. at that time, I specifically chose a robotics - related direction, which is now called "embodied intelligence". At that time, we called it Robotics Learning.

Liu Yanqiu: So, you still centered around AI and found the robotics direction. What was the state of robotics research at that time? How was the mainstream technical path different from now?

Wang Qian: By 2018 and 2019, I found that the mainstream paradigm in the robotics field at that time - deep reinforcement learning - clearly wasn't working. I thought that if it would take 30 or 50 years for this field to make a breakthrough, it wasn't worth spending my youth on it. Instead, I could do something else first, make money, and then come back to support this field. Maybe it would progress faster that way. Many people have done this. For example, Jim Simons in the quantitative field, who has a background in mathematics, and David Shaw, who was originally a chemist. They both achieved success in their fields, then switched to quantitative finance, made a lot of money, and then came back to support scientific research. So, I founded a quantitative fund and ran it for two years. The results were good, and I also made money for the investors.

At that time, I thought it was okay and planned to come back when there was a breakthrough in this field in 30 or 50 years. Liang Wenfeng did it this way, but I entered the industry later than him and didn't make as much money as he did.

However, in 2021, GPT - 3 came out. Although GPT - 3 wasn't as well - known to the general public as ChatGPT later, in my view, it was a very clear signal of a paradigm shift. I told my partner at that time that with this, general artificial intelligence (AGI) might emerge within ten years, instead of 30 or 50 years. If AGI comes in ten years, it doesn't make much sense to make a lot of money now. I still wanted to go back and work on AI myself.

I worked on robotics ten years ago, so it's natural for me to do it again now. Even after ChatGPT came out, I still thought that it was difficult to apply AI in fields like language and vision. You can see the same situation now. The only thing that can really be applied is writing code. We also considered working on code - related things at that time, but later we thought that all things in the virtual world are naturally the advantage of big companies. The hardware - related fields are more suitable for startups.

Moreover, the trend of the scaling law is very obvious. The demand for resources is increasing exponentially. You need an exponentially increasing number of graphics cards, electricity, and data volume to achieve a linear increase in the level of intelligence. Where can you find so many exponentially increasing resources? Some people might say that the United States has invested 1 trillion US dollars in AI infrastructure, which is a large amount. But the next - generation AI might require 10 trillion US dollars, and the generation after that might need hundreds of trillions of US dollars. The entire human economic system simply can't support this. So, how to obtain exponentially increasing resources from the real world and the physical world is the core issue to support the continued development of AI. We think this issue must be addressed through embodied intelligence.

Actually, since the Industrial Revolution, people have been thinking about "machines making machines". One machine can make 10, 100 machines, and 100 machines can make 1000 machines, which can achieve exponential growth. But until now, we haven't seen this fully realized. The core reason is that in both the industrial and service sectors, all production processes rely on human hands. There isn't a single item around us that can be manufactured completely without human labor, and the human - labor component is still significant. This is an obvious bottleneck.

So, we believe that once we develop embodied intelligence and replace this linear bottleneck, we can achieve exponential growth in a complete sense and may even support the development of linearly growing AGI and even ASI (Super Artificial Intelligence). After realizing this, we determined that the value of embodied intelligence is the greatest.