The "golden 10 years" of AI at the World Artificial Intelligence Conference is only the beginning

Big models and generative artificial intelligence (AIGC) have gone out of the technology field and become one of the hottest topics today. The key reason is the significant opportunities it could bring to global development on the one hand, and its unsettling power on the other. At the World Artificial Intelligence Conference, which opened yesterday, global industry leaders focused on big models, talked about innovation, trends, and how to hit the brakes.

AI must not only "read ten thousand books", but also "travel thousands of miles"

Many heavyweight guests unanimously believe that starting with the breakthrough of large models, we will usher in a "golden 10 years" dominated by AI (artificial intelligence). And now, just the beginning of this big era. In order to smoothly cut into this cadenza, AI developers also need to carefully polish the large model so that it can be accurately adapted to thousands of industries.

Ken Hu, Huawei's rotating chairman, borrowed the traditional culture's requirements for personal cultivation - reading 10,000 books and traveling 10,000 miles, and described the development direction of large-model technology in his eyes. The so-called "reading ten thousand books" is to let AI learn massive data and complete the "accumulation of basic knowledge", but this is not enough to fully empower and change traditional industries, for this reason, the big model must "travel thousands of miles", that is, the deep integration of AI and the real economy, forge a subdivided industry model. For example, Huawei has used its own software and hardware platforms to connect large models with hundreds of industries. Among them, AI models trained using decades of meteorological big data can give forecasts for the coming week's global weather in seconds.

In fact, regarding the development of large models, the industry views have formed two factions, one is the "universal" school, that is, the AIGC fever detonated by ChatGPT is likely to allow people to find a breakthrough towards general artificial intelligence; Another group believes that the vitality of AI lies in the landing, and it is better combined with specific scenarios, and it is too early to talk about general AI.

Most of the guests who took the stage yesterday held the latter view. The most resolute is Yang Likun, a Turing Award winner and one of the "Three Masters of Deep Learning". He simply said that at present, there is no general-purpose AI in the world, and "every intelligence is dedicated."

When it comes to how to connect AI with the industry, Tencent and Baidu can be said to be heroes who see the same thing. Tang Daosheng, senior executive vice president of Tencent Group, believes that many AI models seem to be all-powerful, but in the face of various industries, they often can only solve 80% of the problems, and the remaining 20% are prone to errors, which brings chaos. To this end, AI companies need to create exclusive "small models" in addition to the general large model according to the business characteristics of customers. According to this model, Tencent planned the so-called MasS ("Model as a Service") business, which is to help customers quickly complete the custom development of "small models".

Baidu's large-model business is similarly designed. Chief technology officer Wang Haifeng revealed that enterprises need industry models, but cannot solve the cost and difficulty problems. To this end, Baidu hopes to learn from the chip foundry model and provide automated and standardized AI model customization services for various industries through its own super computing power, ultra-big data and 100 billion-level large models.

Machine assistants, digital twins, and new technological breakthroughs are unveiled in a concentrated manner

After talking about business models a lot, it is inevitable that they will be a little jerky and boring. In contrast, hardcore and cutting-edge technical topics may be more eye-catching. Yesterday, a number of industry bigwigs and cutting-edge talents took turns on the stage, showing the most proud and optimistic AI technology.

Yao Zhizhi, winner of the Turing Award and president of the Shanghai Institute of Zhizhi, introduced the latest achievements of a young teacher under his command. That's the latest reinforcement learning algorithm that learns hundreds of times faster than traditional techniques. In the past, it would take months to train an intelligent robot, but with new algorithms, it can be taught in hours.

As a key investor and partner of OpenAI, the developer of ChatGPT, Microsoft is working to make the big model more deeply tied to its Windows Windows system, Office suite and cloud computing services. Hou Yang, senior vice president of the world, chairman and CEO of Greater China, revealed that Microsoft has launched 50 AIGC pilot services, among which the most important CoPilot intelligent assistant can quickly generate PPT speeches on behalf of users, or write software at 150% speed. In addition, another revolutionary technology released by Microsoft, the AI plug-in system Plug-In, will also become an interface to connect large models with various specific application scenarios, and it is expected that when it is officially released soon, the total number of ChatGPT plug-ins can exceed 1,000.

He Xiaodong, vice president of JD.com Group and president of JD.com Exploration Research Institute, first played a video of him introducing the progress of AI research and development on JD.com's live channel. After that, he "revealed" that in fact, he and the host in this film are not real people, but AI-generated digital people. At present, the threshold for generating such a "digital twin" that can be fake is very low: just take a 5-minute picture of the party, and the AI can model it in real time, completing a perfect reproduction from appearance and expression to accent and tone.

In addition to introducing the products and technologies that have come out, the future development direction and breakthrough of the large model are also hot topics yesterday.

Yang Zhilin, assistant professor at the Institute of Interdisciplinary Information Studies at Tsinghua University, proposed a very interesting direction: can AI have "collective memory" like people? He said that although large models can remember the context to a certain extent when talking to people, once the chat ends, the context no longer exists. Is it possible to use new technologies to allow AI to quickly enter the role every time? He thinks the application is very promising.

Pan Xingang, assistant professor in the School of Computer Science and Engineering, Nanyang Technological University, is bullish on the field of video generation. He said that the current large models are quite powerful when it comes to creating pictures; In the future, it should be further upgraded to provide more help in the generation of dynamic images, including 3D content.

In addition, whether AI can be "fed" with enough and good enough data is directly related to the ability and performance of large models. But at present, the "data shortcoming" is still widespread. Speaking of the way to crack, Sun Yuanhao, founder and CEO of Transwarp Technology, said that the most lacking thing in AI is actually knowledge data, that is, how to interpret a news event and how to solve a mathematical problem. He believes that the technical community currently needs to develop corresponding tools and databases to extract human methodologies in natural language, only in this way can the big model be used well.

AI requires responsible governance, but overly worrying or unnecessary

Since AIGC and large models have shown powerful intelligence beyond expectations, the outside world's worries about AI defeating the human brain and challenging human beings have become unacceptable. Yesterday, the voices that reporters heard in the venue about how to regulate AI seemed to be much more frequent than before. Of course, the industry leaders do not have the same anxiety about the current AI, they are more based on rationality, long-term, and give suggestions on the governance of AI.

Yang Likun believes that it is necessary to set goals for AI, define nature, and build safety barriers. For example, it can be required that AI must be honest, must obey, and cannot harm people; Once violated, the AI should get emotional feedback similar to fear or pain. Since the implementation of technology is very challenging, Yang Likun suggests starting with low-level AI, such as first putting the equivalent of mouse intelligence into a "technical sandbox", studying how to train it into a good mouse, and then gradually improving AI capabilities until it is close to human intelligence.

Hou Yang said that in Microsoft's view, to build responsible AI, it is necessary to follow six principles, namely fairness, reliability, security, transparency, privacy and inclusion, "We advocate global technology companies to form a consensus through cooperation to ensure that AI technology benefits all mankind."

Of course, there are some AI industry leaders who have reservations about the concerns of the public and peers. David Holtz, founder of MidJourney, the world's hottest mapping AI, said that the technological progress of AI so far is enhancing the capabilities of people themselves, and it is not yet seen that the era of strong artificial intelligence will come in the near future. Holtz said that AI is not an individual individual, it is just an "entanglement" or "flow" of intelligence. For example, painting AI, it is by no means an artist working side by side with you, but an extension of your thoughts. The startup star emphasized: "If you think too much, you may miss a lot of opportunities. ”

If you think Holtz's expression is too mysterious, then the way Tang Xiaoou, a professor at Hong Kong's Chinese University, comforts "AI phobia" is more down-to-earth. The expert, who has led his team to a large number of world-class results in the field of AI, loves to listen to a crosstalk before bed. At this moment, he would listen and think, "How can a machine surpass such an interesting soul?" I don't believe it. ”

Author: Zhang Yi

Editor: Shang Hui