At the World Artificial Intelligence Conference, lively, searching, and confused.
Wen 丨He Qianming Qiu Hao
Editor丨Cheng Manqi
Huawei, Tencent and Ali are coming, Intel and Qualcomm are also coming; Tesla is here, with its 1:1 humanoid robot model; Nvidia didn't make an appearance, but it also appeared everywhere, in the speeches of Chinese competitors, in the inquiries of potential customers.
This is a regular configuration of an AI conference, but among the nearly 180,000 people who came to Shanghai from July 6 to July 8 to participate in the sixth World Artificial Intelligence Conference (WAIC), there were jewelry testing institutions, teams from the Public Security Bureau, and employees of nuclear power companies and hospitals to look for opportunities, and even some elementary school students came to visit, more than one group.
At a time when growth is becoming scarcer and certainty is becoming more and more rare, AI big models promise a rare new possibility. People from all walks of life gathered in Pudong, Shanghai, with their questions and bewilderments, and when they left, some questions went unanswered.
"If the big model is the answer, what is the problem?" This is the question that a participant looking for entrepreneurial opportunities asked. He used to work at ByteDance and is now preparing for a big model startup, hoping to find some inspiration on the more than 30 large models on display at the conference.
After a morning of shopping, he thought about the last wave of artificial intelligence fever he experienced. AlphaGo beat Ke Jie in 2017, and the following year, WAIC was held in Shanghai for the first time. At that time, the venue was most displayed on various giant screens, showing the traffic flow or neighborhood monitored by cameras. AI companies have dubbed such systems "smart cities."
Since then, the development of artificial intelligence has not been completely driven by the market, but mixed with the guidance and expectations of the government. This big model boom is no exception, and officials have appeared at the opening ceremony and some large-scale forums, expressing their intention to provide policy assistance for promoting the development of the artificial intelligence industry.
Comparing the two waves, the entrepreneur has more questions: "CV (computer vision) has security, what will the big model have?" He doesn't think many companies are willing to spend millions or tens of millions of dollars a year on an imperfect chatbot to help them answer customer service questions or do document refinement and summary.
His doubts show the other side of the big model craze: the big model is like Mjolnir, but without the right nails, it is difficult to play the huge commercial value that is expected.
People from all walks of life came to see how the big model could help themselves
After 9 a.m. on July 6, before WAIC opened the pavilion, crowds had already occupied the cafes near the Shanghai World Expo Center and the World Expo Exhibition and Convention Center, and latecomers could only crowd into the nearby noodle restaurant to discuss AI and large models.
The "scalper" also took place near the entrance, whispering to the people who came and went, "Have you made an appointment?" According to official requirements, you cannot enter without making a reservation in advance. The scalpers said they could handle the admission problem, "it cost 400 yuan".
Not only AI practitioners were watching the exhibition, but staff members could be seen everywhere in the exhibition hall holding signs leading the tour, with signs saying "Jewelry National Inspection and Procurement Group", "Shanghai Municipal Public Security Bureau Hongkou Branch" and "Great Wall Motor Purchasing Group".
In the large-model forum held by Huawei, the late media reporters found that the reserved media seats had been preempted, not because the media colleagues were too enthusiastic, but because there were too many people present. Two employees of China Nuclear Power Group are also sitting here, and they have used artificial intelligence technology in the past few years to maintain and overhaul equipment, and now they are concerned about the big model: "Let's see if we can find some points of integration."
At a time when the growth thrust is waning, big models are one of the few bright spots, and after the feverish hype of the past six months, some people are worried about being replaced and disrupted by AI, and others are determined to jump into the trend before being replaced. However, this industry-wide enthusiasm temporarily lacks a cognitive basis, and most people still do not understand what a big model is and can do.
"Does a big model mean a lot of computing power and take up more storage space?" A person working at the hospital asked nearby staff after walking around the Tencent exhibition area. She knows the visual capabilities of artificial intelligence and can already help doctors see CT images, but she doesn't know much about what big models can help hospitals, and is eager to know more.
Too much expectation that forms too quickly often brings an equal amount of disappointment. At the exhibition site, many people gathered around the booth computers to experience the big model: "write a brand story of creative coffee" and "write a travel plan for a primary school student". One exhibitor caught a glimpse of the "write love letters" feature displayed by Ali Tongyi Qianwen and immediately complained: "What is this thing? Can you help me do something practical? ”
Visitors gather around the WPS booth to experience the large model application, image from WPS.
Few companies are talking about creating another OpenAI
After sending away a few people who failed to test large models, an exhibitor from an Internet company sighed: "No way, the technology is still not very good." ”
The one who just walked took a math problem to test the large model, asking it to figure out "how many of the five digits are different for each number." After seeing the result, the other party raised his mobile phone and said, "It's still ChatGPT more reliable."
Before attending the conference, the staff at the booth saw that his peers had made dozens of large models in a few months, and he already had a feeling: making a large model seems to have less technical barriers, but it is difficult to do it well. During WAIC, he took the time to experience the big models of his peers, and this feeling was even stronger: "The effect is not much different, it is far from ChatGPT. ”
Just six months ago, there were many companies that said they wanted to do China's OpenAI. But in the booths and forum speeches of this conference, not many people have mentioned this goal.
The new narrative is: industry big model and "big model empowers thousands of industries".
Huawei exhibited a large model of the mine.
Ken Hu, Huawei's rotating chairman, said at the opening ceremony of WAIC, "The goal of the big model is to serve different applications in different industries... in order to play a greater value. Tang Daosheng, senior executive vice president of Tencent Group, expressed a similar view in a follow-up speech: "Industry big models are a better option for enterprises to embrace big models. ”
Neither Tencent nor Huawei put computers on the booth to allow people to experience the big model. Although Alibaba, which is next to Tencent's booth, put more than a dozen computers in the center of the exhibition area for people to experience, Zhou Jingren, CTO of Alibaba Cloud, also began to emphasize ecology rather than self-developed large models. He said at the WAIC forum: "Alibaba Cloud will make promoting the ecological prosperity of China's big model a primary goal... Let all walks of life enjoy the dividends of large model technology. ”
The shift isn't that big companies are giving up on bigger models, it's that they want to find a viable way to make this immature new technology valuable and generate revenue first. Their logic is:
- General-purpose large models (similar to the large model behind ChatGPT) are expensive to use, model parameters are generally hundreds of billions, and the actual operation costs a lot of resources, ChatGPT and New Bing once drained the hundreds of thousands of GPU computing power accumulated by Microsoft, and the average company simply cannot afford it.
- Generic large models do not work well in specific scenarios. General-purpose large models are generally based on public literature and network information training, and the accumulation of professional knowledge and industry data is insufficient, resulting in insufficient accuracy of answers, "Once enterprises provide wrong information to the public, it may cause serious consequences." Tang Daosheng said.
- Large models in the industry have smaller parameters, lower deployment costs, and better results when answering specific questions after targeted training. And large companies provide cloud MaaS (model as a service) for companies in various industries to train or deploy large model services, which can also help them sell some cloud services first.
Smaller companies are more realistic. A chief scientist of an AI unicorn company said that they have been studying large language models since 2018, and they have also made an application for writing articles in the first two years, because no customers have paid, and the company has not increased investment. After ChatGPT became popular, they also released self-developed large models, but they did not plan to train larger models for the time being, because customers felt that the cost performance was limited, after all, it costs tens of millions of yuan to train a model with hundreds of billions of parameters.
However, even if it is theoretically cheaper and closer to the landing of the industry's large model, it is still expensive to use. "LatePost" learned that a large model with parameters of 6 billion yuan for a large-model startup company that has attracted much attention in China is sold for millions of yuan, and the price of hundreds of billions of parameters is tens of millions of yuan per year - most of the cost is chip computing power.
What the big model can do is not certain, and the sale of equipment is very active
What exactly the big model can do and to what extent it can be is still being explored, but the consensus of most people is that you must prepare equipment before panning for gold.
An obvious change in the WAIC venue this year is that the booths of domestic AI chip companies are larger and closer to the C position, and have also attracted more attention: the booths of chip companies such as Fengyuan Technology, Tianzhixin, Hanbo Semiconductor, Muxi Integrated Circuit, and Denglin Technology are close to Tencent and Baidu, and the booth area is also similar to these large companies. Huawei, which occupies the largest booth in the venue, has separately detached out a Ascend ecological booth, which is Huawei's complete set of AI computing products including AI chips, MindSpore AI training framework, and software services.
The booths of each chip company were crowded, and customers who came to consult wanted to know how the chips performed and what exactly they could do. One of the most common questions is: Is the A100 a replacement?
The booth of Chinese AI chip companies was larger than in previous years.
The A100 GPU, introduced by NVIDIA more than two years ago, is now almost standard for training large models. When ChatGPT came out, AI startups and tech giants sprang up to buy the A100. The US government's export controls on AI chips such as A100 have exacerbated its shortage in China, but it has also allowed domestic chip companies to see alternative opportunities. Zhang Dixuan, president of Huawei's Ascend computing business, said in an interview, "In the past, we were looking for enterprises, but now many enterprises are looking for them. ”
Unlike companies that develop large models and are still experimenting with various end-use scenarios, chip companies show something much clearer and more intuitive: in the conspicuous position of each chip booth, there are often various models of chips, and servers equipped with their AI chips, which look like large chassis with rows of AI acceleration cards. Companies will also use on-site computers to demonstrate the application effects of AI large models or AIGCs supported by their chips: including conversational robots, AI paintings, etc.
NVIDIA did not set up a booth on WAIC, did not have a title forum, did not win any awards, in the chip forum, NVIDIA only sent a technical director, he was the last to take the stage, behind Qualcomm, AMD and Intel. But almost every chip company will compare the indicators of the NVIDIA A100 when promoting its products; When Zhao Lidong, CEO of Fengyuan Technology, spoke in the same forum, he opened with the market capitalization of NVIDIA exceeding trillion dollars, indicating that Wall Street is betting on AI computing big opportunities with real money.
The Chinese government is also paying more attention to supporting AI computing power than ever before: a vice mayor of Shanghai attended the chip forum at last year's conference, and this year came a member of the Standing Committee of the Shanghai Municipal Party Committee, the secretary of the Pudong New Area Party Committee, the director of the Shanghai Municipal Economic and Information Commission, and a deputy director of the Science and Technology Department of the Ministry of Industry and Information Technology.
In addition to chip companies, there are also cloud computing vendors and data service companies, as well as headhunting firms and local parks.
Most companies have a hard time buying GPU chips to train large models, and a better way is to rent the computing power supplied by cloud vendors directly. Microsoft Azure and Amazon AWS have all been on WAIC's main forum this year.
"LatePost" learned that the data collection and annotation platform Appen has made a heavy bet this year, spending most of the exhibition budget on WAIC throughout the year; Its Chinese counterpart, Haitian Ruisheng, whose share price has risen more than twice in just one month, attended the meeting for the first time, and the staff at the scene said that in addition to receiving a steady stream of potential customers, "many shareholders came to thank us".
Various service providers in the startup ecosystem are also looking for customers. An exhibitor said that he met several waves of headhunters and local park investment promotion personnel in one morning, and received a bunch of business cards. No matter what new opportunities you want to try, talent and business sites are the costs to be paid by one group of companies, and new development opportunities for another group of companies and places.
Large models have driven general-purpose robots into hot spots, unmanned vehicles, and the meta-universe has ebbed
Among the various relatively abstract applications and system solutions, robots are a rare application direction that can be "seen and touched".
In the past six months of the big model boom, a concept called "Embodied AI" has also gained attention. In simple terms, embodied intelligence refers to the combination of artificial intelligence software and hardware to solve real-world problems. In May this year, NVIDIA CEO Jensen Huang said that embodied intelligence will be the next wave of AI, and the typical representative of embodied intelligence is robots, especially general-purpose robots that can complete a variety of complex tasks with the same form of product.
At the WAIC site, robot dogs and humanoid robots have visibly increased, and the organizers said that there were more than 20, compared with single digits in previous years.
The lively scene in the venue is a group of people surrounding the robot dog "teasing the dog", trying to push it down and ride it. The robot dog in the depths of the cloud fell down in the air during a performance of climbing the steps, which immediately provoked a dense sigh in the crowd of onlookers: "It's over, it's over... It's over".
People watch the robot dog. Source: Visual China.
Regarding how to combine large models and robots, practitioners now have different opinions. Some viewers asked the staff of the robot dog company Unitree Technology: "Will you use large models on robot dogs?" The staff was stunned for a moment and said, "I haven't found anything to do yet."
The founder of a robot company told "LatePost" that the obvious application direction of large models in robots at this stage is to replace the code with natural language, directly input instructions to the robot, so that the robot has some "common sense" and can divide the tasks that people want to convey into various robot subtasks; But the execution of subtasks needs to rely on the basic capabilities of the robot, such as navigation, planning, control, etc., and large models may help, but not replace and subvert.
Compared with the hesitation of robot dog companies in the face of large models, companies that develop humanoid robots are much more optimistic. Cloudminds released the "RobotGPT Industry Big Model", claiming to "lead the new era of embodied intelligence". They transported a dozen robots to the scene and asked them to form a column to dance the "Thousand Hands Kannon".
Cloudminds Technology's robot performance "Thousand Hands Kannon".
It has met a new opponent at this AI conference. The booth is opposite Cloudminds at Fourier Intelligence. The company, which previously made intelligent rehabilitation equipment, has now released a humanoid robot and also announced that it will "lead AI into the era of embodied intelligence."
The most eye-catching humanoid robot is Tesla's Optimus. Beyond the red line protecting the robot, the gathered crowd competed to drill closer and closer, so that they could take pictures and record it, as if worshipping a god statue, even though it was only a 1:1 model.
Visitors crowded in front of the Tesla humanoid robot Optimus to take photos.
People passing by often ask the staff, "Can it move?" Some people even leaned up and down to shake the phone back and forth, and they actually shot a dynamic effect. "Although it is a model, it is very shocking", said a young man at the scene, he saw that the robot is 1.9 meters tall, which may be the halo brought by black technology, and Optimus' official height is 1.72 meters.
In previous years, such a grand situation belonged to unmanned vehicles and the metaverse.
The 2021 AI Conference is like another Shanghai Auto Show. Self-driving companies Incept Technology, Pony.ai and Tucson will spare no expense in the future to physically transport trucks 4 meters tall and nearly 10 tons to the booth. AutoX, SAIC, Huawei Car BU, Baidu Apollo, SenseTime, and even Xinchi Technology, which makes automotive chips, all directly displayed cars equipped with their own technologies or products. WAIC also set up a special unmanned car experience area, and unmanned minibuses were used for the venue connection.
This year, the experience area and unmanned minibus are gone. Only a few car companies such as SAIC Zhiji, Jidu, Tesla and other car companies still have cars at their booths, two years ago, the "truck three heroes", this year only Tucson Future showed a perception kit wrapped in a plastic box, and other unmanned car companies almost did not come.
Tucson's future 2021 booth (top) and 2023 booth (bottom) comparison.
The intelligent driving forum of the conference was scheduled to the last day of WAIC, and Cao Guangzhi, co-founder of Yunji Zhixing, laughed at himself while encouraging his peers to tide over the difficulties while participating in the roundtable discussion: "How did such a sunny and snowy thing as autonomous driving be rolled up like this by us?" ”
The metaverse has a similar cold reception. Last year's WAIC Conference embedded the concept of the metaverse into the theme name, "Intelligent connection of all things, Meta without boundaries", Meta Greater China President Liang Youberry was invited to speak at the opening ceremony, and the organizers also set up metaverse check-in points in the Expo Center, Xuhui West Bank, Zhangjiang Branch Venue, Oriental Pearl and Wukang Building.
A year on, only a handful of XR (extended reality) businesses remain using the metaverse as a promotional point. Among the 10 cutting-edge technologies officially set up, there are only 3 metaverse-related forums this year, one-fifth of last year. From the entrance of the main venue all the way to the end, you can see some small booths of metaverse-related companies. Whether it is autonomous driving or the metaverse, in the context of the current wind and financing difficulties, most companies have left their limited budgets to maintain company operations.
When a new technology boom emerges, there are often two evolutionary paths: First, the new technology realizes value, becomes part of the infrastructure, and is no longer paid attention to, such as the Internet and recommendation algorithms. The other is that new technologies fail to deliver value in the short term, and then are robbed of resources and limelight by a new boom. Now big models have become the new hot spot, but after the past round of technical hype, insiders and outsiders have calmed down a lot. Those who really want to do something in this new opportunity actually want the enthusiasm and expectations of the public to be more pragmatic.
Wu Yunsheng, vice president of Tencent Cloud, said that Tencent has joined hands with customers in more than 10 industries such as finance, cultural tourism, government affairs, media, and education to create more than 50 industry large-model solutions. As far as we know, none of these plans have been launched for the time being. He said that now is the initial stage of the development of large models.
Zhu Likun also contributed to this article.