Meituan took over Wang Huiwen: Is the big model business dead?

Author: Ye Zhen Editor: Kang Xiao

**Production: **Shenwang·Tencent News Xiaoman Studio

On the afternoon of June 29, Meituan issued an announcement to wholly acquire AGI, a company founded by former director Wang Huiwen.

The farewell of Wang Huiwen, an Internet veteran, is quite sad: he broke into the AI wave full of ambitions, and had just completed the A round of financing with 280 million US dollars in cash on his account, but had to withdraw due to the onset of depression due to continuous high-pressure work.

Specifically, the purchase price of this transaction is about 2.065 billion yuan. According to the announcement, the specific purchase price includes cash of 234 million U.S. dollars (approximately RMB 1.694 billion), debt commitment of 367 million yuan, and cash of 1 yuan to Wang Huiwen.

Light Years Beyond was founded and controlled by Wang Huiwen, the co-founder of Meituan, with a total net cash position of approximately US$280 million (approximately RMB 2.063 billion). The shareholding ratio of all issued share capital outside the light year is: AI Age holds 76.72% of the shares, Qimai holds 0.43% of the shares, Red Shirt China holds 2.44% of the shares, and other sellers hold 20.41%, totaling 100%.

After the completion of the acquisition, all the original A-round shareholders will withdraw at a fair price. Meituan needs to pay US$28 million in cash to Red Shirt China and US$5 million in cash to Qimai. Since AI Age is the holding company of the founder Wang Huiwen, who is light years away, Meituan only needs to pay 1 yuan to Wang Huiwen himself, and does not need to pay cash to AI Age.

In addition to lamenting the generous help of Meituan and Wang Xing, Wang Huiwen's withdrawal also raised a question for the AGI start-up companies in the industry: Will the big model eventually become a giant's game?

## **01 Boss debate: Should start-up companies give up their illusions? **

Recently, Wang Huiwen, the founder of the large-scale start-up company Light Years Beyond, retired from illness. Investor Zhu Xiaohu and Cheetah founder Fu Sheng had a debate over "whether large-scale entrepreneurship has entrepreneurial value", which put the original in the wave. The large-scale model business at the top is once again pushed to the cusp.

Zhu Xiaohu claimed that ChatGPT is very unfriendly to start-up companies, and please give up financing illusions in the next two to three years. In this regard, Fu Sheng commented, "Half of the start-ups in Silicon Valley started around ChatGPT, and our investors can be so ignorant and fearless. .”

"What Zhu Xiaohu wants to say is that the general-purpose large-scale model is not friendly to start-up companies. It is normal for everyone to have different opinions in the early stage, which means that there are greater opportunities in this field. If entrepreneurs and investors have reached a consensus on the industry, then the chaotic period of the entire industry will pass, and opportunities may be less. "

For ordinary entrepreneurs, large-scale entrepreneurship is a false proposition. Huge capital investment and scarce top AI talents. For large-scale entrepreneurship, both money and technology are important.

"The training cost of the general-purpose large model is estimated to be about 1 billion US dollars. This is only the computing power part, not counting the other two expensive parts. One is data and the other is labor costs. Now the global large model Talents in the field are very scarce." Dr. Du Feng, founding partner of Jiangmen Venture Capital and former head of Microsoft Ventures Greater China, told "Deep Web".

Since 2016, the first year of artificial intelligence, the AI industry has undergone several rounds of reshuffle. With the help of ChatGPT, AI large-scale model entrepreneurship has once again been placed in the spotlight.

A few days ago, MiniMax, a general-purpose large-scale model start-up, completed a new round of financing, with a total scale of more than 250 million US dollars and a project valuation of more than 1 billion US dollars, ranking among the ranks of unicorns. The co-founder of MiniMax is Yan Junjie, the former vice president of SenseTime and head of general intelligent technology. He also served as the vice president of SenseTime Research Institute.

Not only Yan Junjie, but also Internet veterans such as Wang Huiwen and Wang Xiaochuan, middle and high-level executives of major Internet companies, and some academic scientists have joined this wave of large-scale model entrepreneurship. In just half a year, it is said that there are currently more than 70 large-scale startup companies in China.

"This era is very similar to the gold rush era. If you went to California to pan for gold at that time, a lot of people would die, but people who sold spoons and shovels would always make money. Large models are platform opportunities; platforms with models first , will be larger than information-first platforms.” Lu Qi, founder of Qiji Chuangtan, said.

"Perhaps 90% of these 70 large-scale model companies will say that we are not going in the direction of AGI. I understand that most companies can't do it either." Cheng Weizhong, CEO of Zhongke Shenzhi, told "Deep Web" .

For the founders of large-scale startups, what really matters is the question of sustainable financing capabilities. Cheng Weizhong believes that most of the more than 70 large-scale entrepreneurial teams do not have such capabilities. From the perspective of general artificial intelligence, money is more important than technology.

And it is true. The financing in the domestic AI field, on the whole, just looks very lively. Some investors announced the financing data in the domestic AI field from December last year to early May 2023 on a forum. There were only about 30 investment and financing cases, and the total amount was estimated to be less than 10 billion yuan.

"What we heard is that more than 70 companies in the industry are large-scale startups, and everyone uses the Open AI training method. Because the parameters are large, the data is similar, and the result is serious homogeneity. The market There is no need to use so many large models on the Internet, and the fever should subside by 2024, and who is on the shore and who is swimming naked.” said Li Di, CEO of Xiaoice Company.

02 large model, a game that becomes a giant

"China will definitely have its own ChatGPT. Just like search engines, we have our own compliance requirements. But the Chinese version of ChatGPT will only be produced in 5 companies: BAT+Byte+Huawei."

Cheng Hao, the founder of Xunlei and Yuanwang Capital, told "Shenwang", "This means that start-up companies or those A-share listed companies that are purely for speculating in stock prices will hardly have a chance. Of course, my statement will offend many entrepreneurs. Including mine My friend Wang Xiaochuan is also starting a business in related fields, and many FOMO VC friends are now vowing to invest or have already invested in large-scale entrepreneurial teams.”

For start-up companies, the biggest challenge is to try to catch up with the most advanced level of ChatGPT, and it is likely to be dragged down.

"After several months of hard work, the level of the domestic general large-scale model should be between 3.0 and 3.5, while Open AI has reached 4.0. This gap must exist." Cheng Weizhong analyzed.

In Wang Xiaochuan's view, China needs to catch up, and it is not too late now. "Catching up to ChatGPT 3.5 level, I think it may be possible within this year, but for ChatGPT-4 or ChatGPT-5, I think it may take about 3 years, and it should not be less than two years." Wang Xiaochuan told "Shenzhen Net" said.

Wang Xiaochuan believes that although start-up companies are not comparable to big companies in terms of talent team and capital density, when facing the future direction, they may make such disruptive big products that big companies cannot do. "The most obvious advantage of a big factory is that it has enough funds to guarantee the supply. It needs to focus on its original business as a priority service object for its shortcomings. Therefore, in the face of this possible future huge change, the newly developed product model And the way of application may not be their priority, it will limit them, just like Sogou was doing search at that time, I will give my model more priority to serve search, and may lose some bigger opportunities in the future.”

Of course, there are only a handful of entrepreneurs who meet Wang Xiaochuan's conditions.

"Entrepreneurs can outperform big manufacturers only when they have a first-mover advantage. A general-purpose large model is a typical winner-takes-all. Giants have money, technology, and more importantly, a scene." Cheng Hao insisted , Large-scale models are the battlefield of giants. Although there is no real leader in the world sense of Chinese large-scale models, there will definitely be a leader in the Chinese local market.

**03 Where are the opportunities for start-up companies? **

If the big model is reduced to the game of giants, where are the opportunities for startups?

Kai-fu Lee, the founder of Innovation Works, believes that in the future AI 2.0 era, there will be three ecological layers: the basic model layer, the middle layer and the application layer. "The scale of the AI 2.0 market represented by large models is huge enough to accommodate giants, small and medium-sized enterprises and start-up companies to 'horse race' in it."

Fu Sheng said that half of the start-ups in Silicon Valley started around ChatGPT. In the view of Yuanwang Capital Cheng Hao, this is absolutely true. Cheng Hao just stayed in Silicon Valley for a month and a half. "If a Silicon Valley entrepreneur talks about a project that has nothing to do with the big language model, it will be defined as a traditional project by investors, and VC may not be interested."

"What Zhu Xiaohu wants to say is that many start-up companies have a relatively thin nutritional value chain, 99% of which are made by ChatGPT. What is the value of your company? This is a problem that start-up companies need to solve. Zhu Xiaohu said that the scene is king, and I agree with this. of."

Cheng Hao said that for start-up companies, it is very important to choose AI+ or +AI. "In this wave of opportunities brought by the big language model, entrepreneurship and investment must have dividends. If there is no dividend, it is like pushing the stone up the mountain. If there is a dividend, it is like pushing the stone down the mountain. With some effort, it can roll down the mountain by itself."

Wang Xiaochuan believes that at present, everyone is learning from ChatGPT and its imitation model, and there may be differentiation in the future. One is to serve their main scene. Of course, there are a few products that can go beyond making subversive products. This kind of change may not be displayed until half a year or a year later.

Regarding the debate between Zhu Xiaohu and Fu Sheng on "whether the big model has entrepreneurial value", Chen Yizhou, an Internet serial entrepreneur and CEO of Thousand Oaks Interactive Group, said, "Others ask me this question, and I ask instead, what is the big deal after the search engine comes out?" Are businesses born? The answer is that there are not many brand new businesses, but many businesses that already exist can be leveraged.”

In Chen Yizhou's view, AI can be added in many fields, "For example, B2C e-commerce and games, especially B2B businesses with a very fine division of labor, the diversion speed and traffic accuracy are improved. Existing SaaS products have so much data and paying customers , there is no need to restart the stove at all, and it is a huge beneficiary group. Now AIGC basically does not invest in these fields, and the bubble in the future can only be blamed on VC.”

On this point, Chen Yizhou and Zhu Xiaohu also reached a consensus.

For more start-up companies, it is a wiser choice to choose a vertical track to expand the model.

From the perspective of development in the United States, calling ChatGPT's API cannot solve the problem of most enterprises using large models, so vertical or enterprise-owned models based on open source ecology show great potential.

"Two years ago we laid out a general-purpose large-scale model. Today we will focus on multimodality, invest in vector databases, and continue to look at chips and vertical fields." Founding partner of Jiangmen Venture Capital and former head of Microsoft Ventures Greater China In Dr. Du Feng's opinion, "The entire ecology of the open source community is helpful and exemplary for China's large-scale model entrepreneurship. There is no permanent lead. Everyone is working hard to move forward. China may need more large-scale models relatively speaking. At most A dozen."

Cheng Weizhong, CEO of Zhongke Shenzhi, told "Shenwang": "In this wave of large-scale model entrepreneurship, everyone's minds are actually not so hot. Everyone is doing vertical fields, and I am a virtual person." In May, Deep Intelligence released a large language model with a level of 2 billion parameters, "Shuzhi Jiangshang". This is the product they rented out 2,000 Nvidia A100 graphics cards.

For domestic capable enterprises, the self-training model is a wise choice at the current stage. Cheng Weizhong, CEO of Zhongke Shenzhi, believes that for those who do not have the ability to train themselves, actively cooperating with companies with large models for training is another good choice. Don't be too idealistic, and don't worry about wasting money. There are not many better options for many things.

"After a particularly high-quality open source model appears in China, most of the existing enterprises with large models can choose to migrate to the open source large model for retraining. The previous investment can not be regarded as wasted in vain. Talents have been cultivated and data has been accumulated. , explored the scene, and may also grab the vertical track. Moving to the open source large model can continue to give play to the first-mover advantage, which is of great value." Cheng Weizhong, CEO of Zhongke Shenzhi, told "Deep Web".