Wang Jian: Openness has always been the key variable for technological breakthroughs.

```

In the era of rapid AI development, openness has almost become an unavoidable topic for industry players.

On September 11, Alibaba Cloud founder and Zhijiang Lab Director Wang Jian pointed out at the 2025 Inclusion Beach Conference opening ceremony that we are currently experiencing a revolutionary shift from "open source code" to "open source resources." By 2025, open source will be a key variable in AI competition.

Earlier this year, the open-sourcing of models such as Qwen and DeepSeek caused a stir in the industry, and even OpenAI founder Sam Altman admitted, "OpenAI is on the wrong side of history."

According to Wang Jian, this turn marks the evolution of open source from just "open code" to "open resources." He further stated that, from a theoretical standpoint, the development of AI has always been intertwined with the concept of "openness."

Turing Award winner Geoffrey Hinton noted that things envisioned in the 1980s have now become reality. In 2012, the advent of the facial recognition era, plus the breakthroughs of Transformer and tokenization in 2017, led to exponential increases in the scale of data, models, and computing power, making resources all the more important.

Wang Jian believes that in the current AI era, only opening source code can no longer meet industry needs, and that the essence of open source is "open resources." The opening of model weights is not only the sharing of data resources, but the release of computing resources—individuals no longer have to repeatedly invest massive computing power to train existing models, but breaking new ground still requires more resource投入.

Currently, this concept is extending into space exploration. In May this year, twelve "Three-Body Computing Constellation" satellites were successfully launched, sending computing power and AI into space for the first time, and achieving interconnection between satellites, enabling direct in-space data processing. In the future, this constellation plans to open each satellite to the world; a longer-term goal is to send satellites to the Lagrange L5 point of the Sun's orbit.

In Wang Jian's view, from code open source in the Internet era to resource open source in the AI era, openness has always been a key variable for technological breakthroughs. Only by sending AI and computing power into space can humanity truly leave Earth. "On the journey to Mars, humans cannot go without the companionship of computing and AI. That will be the most exciting thing in the next 10 or even 20 years."

The full text of Wang Jian's speech is as follows:

I am very happy to have this opportunity to share on this topic with you. In today's context of artificial intelligence, there is an unavoidable topic—"openness."

I want to discuss a few things that everyone may be familiar with but also confused about, from my own perspective. "Open source" as a term is understood differently by different people. Today, we are going through a revolutionary change from code openness/open source to resource openness/open source.

Many things have happened in the past year. From the perspective of artificial intelligence, 2025 is bound to be an extraordinary year.

On January 13 this year, the United States announced export controls on artificial intelligence. What many people may be most familiar with is export controls on chips, integrated circuits, and semiconductors, but in fact, in the same export controls, for the first time, it was explicitly stated that the weights of AI models would be regulated. There's an interesting point here, maybe even a loophole: they only specifically restrict the export of "closed source" weights, and emphasize that "open source" weights are not subject to the restrictions.

I think there was an important assumption behind this, as of January 13, just over half a year ago. At that time, the best foundational models in the world were all owned by leading American companies. Another interesting thing: on January 18 this year, the well-known Jeff Hinton had visited Shanghai a few weeks before. Everyone knows he is strongly opposed to open-sourcing AI models, for very important reasons of his own.

But on January 31, with the open-source release of Qwen and DeepSeek, Sam Altman said something on January 30 that shocked everyone: At the moment of open source, OpenAI stood on the wrong side of history. I won't discuss the deeper meaning behind that; it's not a "strategic" mistake, but a choice of history.

I think one of the miraculous things about 2025 is that it's so hard to explain, yet it's become a key variable for AI competition, even carrying over concepts from the software era. Today, no one can avoid this variable in thinking about how to do AI next. For an industry and a technology, the variable isn't new.

In fact, in 1998, when the Internet started to rise, everyone knows the most important sign was the web browser. Netscape was the best and most open browser at the time, and changed the landscape. In the Internet era, Netscape's open-sourcing marked the "watershed" of that time. The topic of openness was not only important today, it was a key variable in the Internet era as well.

Few people may know that the word "Open Source" originally referred to "Open Source Code," i.e., code openness. "Open source" as a term was only formalized in April 1998 by a group of geeks, because there had been many other labels—free software, freeware, etc.—but "Open Source" was fixed in 1998, which is not that long ago, and opened up the Internet era later.

The Turing Award is interesting in this regard too. When Jeff Hinton and the other founders won the Turing Award in 2019, many things had only just begun. After he won the award in 2018, Hinton gave a talk. In that talk, he presented two important views related to today's open source—which he gave in 2018/2019. He talked about two main approaches in AI: logic-driven, or inspired by logic, and those inspired by the biology of the brain/neuron—two different lines of logic.

Thanks to the second approach, the methodology driven by neurons/biology, we have the concept of "weights."

Eight years later, Hinton repeated these points last month at the Shanghai AI Conference. There's a reason for that. Using biology as a driver for ideas did not start with AI. In the late 1940s, Turing first put forward a basic concept: everything about neurons is now called "connectivity." (As shown in the figure) The place I've underlined in red is interesting: he thought the inspiration from neurons had nothing to do with real neurons—hence the quotation marks. But the crucial point is that when the number of such neurons is sufficient, intelligence emerges—Turing said this in 1948. So, the second path Hinton spoke about is this one.

Interestingly, many people explored this for a long time. This article, published in Nature in 1986, very clearly discussed the importance of "weights" in models. The third author is Jeff Hinton, winner of the Turing and Nobel prizes, but the first two authors were the world's most famous psychologists.

In the mid-1980s, neural (AI) ideas were discussed by teams led by psychologists. At that time, the AI approaches led by computer scientists were logic-driven methods.

If you really wanted to work in this field at that time, there was a textbook after 1986 that I used when I first entered the field—today few people know it—called "Parallel Distributed Processing." When I met Jeff Hinton last time in Shanghai and mentioned this series, he was quite excited: what was imagined back then is now reality. That paper was also completed by Hinton and a few psychologists.

We should thank the Internet—behind these two books, another book was published as a lab manual, the first to truly open all the code related to this theory. To this day, you can download the code from the link on my PPT, even though it runs in DOS, in the original OS, and at that time there was no “open source” concept yet (it's 1986); the open source concept was formalized in 1998.

Actually, the "open resource" concept didn't come from the "open source" term—many pioneers already did this in scientific exploration. I'm glad that pioneers basing theories on biological neurons did a great job exploring this for us.

That's why, in 2012, Hinton and his two students put data, models, and computing power (GPU) together, bringing about the age of facial recognition everyone is familiar with. But at that time, the concept of "resource" was not yet deeply rooted—because data volume wasn't big, models weren't complex, and computing power wasn't as powerful as people imagine. When the paper was published, only two ordinary gaming GPUs were used, far from today's scale.

But in 2017 everything changed. The authors proposed "Transformer" and "Tokenization"—what we now call Tokens, and Tokenization is a crucial technology, truly making data a resource, a milestone achievement.

You'll also see that because of these two things, data, models, and computing power after 2012 were multiplied by a new variable—"scale." All things grew thousands or tens of thousands of times, leading to not just theoretical progress in AI, but in practice, a dramatic transformation.

When scale reaches this level, resources become very important. Imagine: today, open model weights essentially means opening up data and computing resources. With open models, you no longer have to spend as much computing resource and effort to repeat what others have already done.

I'm trying to say, after openness, it doesn't mean massive computing isn't important. It means individuals don't have to spend huge resources again—someone else has already paid the cost. Conversely, to build an even better model, someone may need to invest even more resources.

At this stage, just open source code alone doesn't solve the problems we addressed in the software era. It's opening up resources—especially data and computing power—that pushes the industry forward. This is the key point of "open source" in today's AI era. I prefer to call open source "Open Resource." You know both Open Source and Open Resource translate into "kaikaiyuan" (open source) in Chinese. Of course, open source is not just about the models today.

Space has always been our greatest resource—fifty years ago it was so, and today AI isn't just on phones or computers: AI shouldn't be missing in space. But there are obstacles, notably computing power, when you bring all three together.

This gives us a chance—like redefining the phone as your computer in the past, now with communication satellites, navigation satellites, and remote sensing satellites, the advent of AI brings a fourth type, which I call a "computing satellite." The existence of satellites gives us a chance to send AI there.

I've been lucky doing this work in Zhijiang Lab. On May 14th this year, we sent 12 satellites into space for the first time, and—with this 12-satellite constellation—put an 8B AI model, developed for ground use, into space for the first time. This wasn't just a simple deep learning procedure, but a full-ground model. The first entry was very exciting—it made us realize that AI cannot be absent from space.

With these 12 satellites, as long as any satellite reaches a location, all data can be processed in space, anywhere. Previously, all satellites in the sky had no interconnection—they only communicated with the ground. There was no connection between satellites. For the first time, we've achieved interconnection between space satellites, giving AI a huge opportunity in space.

Why did we call it the "Three-Body Computing Constellation"? Many friends know the term "Three-Body" from the novel. In fact, it’s a thorough scientific term, first proposed by Newton. The basic idea: in space, if there are only two objects—say the moon and earth—their relationship allows an analytic solution: you can describe their relationship mathematically. Add a third body, e.g. the sun, and there's no analytic solution, no definite answer. To describe their relationship, you must assume one is fixed. This was later called the "Three-Body Problem" by Newton. The simple point: one or two people can do things easily, but add a third, it gets difficult. A Chinese saying goes "Three monks have no water to drink." The Three-Body Computing Constellation aims to allow countless actors—three or more—to achieve things collaboratively. That’s what you must do in an open resources scenario: allow countless entities to co-create and share this space.

The South China Morning Post wrote an article after hearing one of my speeches—you can check it out. Essentially, "Only this way can we truly share space and send AI into space." There's an exciting plan: to open each satellite to anyone in the world. This can solve many issues—sustainability, for example—but also allows for the prospect of deep space exploration. Some scientists envision in a few years—not ten, but just a few—sending satellites into the sun's orbit, not just Earth's satellites but the sun's. These satellites will be at the Lagrange L5 point, 150 million kilometers from Earth, 150 million kilometers from the sun. At that point, data cannot be sent back for ground processing; only by sending AI and computing power into space can humanity truly leave Earth.

The coming era is very exciting. I often say: on mankind’s road to Mars, we cannot go without computational companions and AI—that is the most exciting thing about the next ten, or even twenty, years.

Risk DisclaimerThe market has risks, investment needs caution. This article does not constitute personal investment advice, nor does it take into account individual users' special investment goals, financial status, or needs. Users should consider whether any opinions, views, or conclusions in this article are applicable to their specific situation. Invest accordingly at your own risk. ```