Zhu Xiaohu and Fu Shengji argue that behind GPT, the AI model "bubble" is about to burst?

Source: Titanium Media

Author: Lin Zhijia

Image source: Generated by Unbounded AI tool

Under the upsurge of ChatGPT, Fu Sheng, chairman and CEO of Cheetah Mobile, had a debate with Zhu Xiaohu, a well-known investor and managing director of GSR Venture Capital Fund, whether entrepreneurship based on large models is valuable.

On June 26, Fu Sheng posted a circle of friends, reprinting Zhu Xiaohu's point of view "ChatGPT is very unfriendly to start-up companies, please give up financing fantasies in the next two or three years", and expressed his opposition, saying that half of the start-up companies in Silicon Valley started around ChatGPT , Investors can be so ignorant and fearless.

Later, Zhu Xiaohu retorted at the bottom of his post, saying that 99% of the value is created by GPT. What is the value of such a start-up company?

After that, Fu Sheng and Zhu Xiaohu started a "fierce battle" in the circle of friends. Fu Sheng replied to Zhu Xiaohu with three consecutive questions, "99% of the Internet's specifications are created by tcpip. Is entrepreneurship worthwhile? 99% of the value of automobiles is created by thermodynamic definitions. Is entrepreneurship valuable? The traffic of most small and medium-sized websites is It was brought by search engines, so is entrepreneurship valuable?”

Fu Sheng also said that both parties are required to be calm, and the focus on value depends on how it is defined. If only the underlying principles are value, then the existence of an enterprise is unnecessary. Zhu Xiaohu replied to him: Everyone knows whether it creates value or not, just don't have too high expectations.

After it was reported on the evening of June 26, the screenshot of this conversation caused huge controversy in the industry.

Some people commented, "Fu Sheng is right. A large proportion of venture capital circles are ignorant of the underlying logic of technology, and even can't tell the difference between scientific research and industrial application." There is also Wei Zhe, chairman and founding partner of Jiayu Capital, who gave a similar view to Zhu Xiaohu: In this round of "Hundred Model Wars", no more than two domestic AI large models will win in the end-this means that the remaining 98% China's large domestic models will "die" and disappear in the competition.

No matter what kind of views, Fu Sheng and Zhu Xiaohu's "threshold" circle of friends' different "threshold" debates have made more people see some deep-seated contradictions under the surface of the AI boom. In particular, the current growth rate of ChatGPT users is slowing down, the application of large models is not as good as expected, and domestic GPU chips cannot be filled and other factors are superimposed.

This round of domestic AI big model "bubble" really seems to be about to "burst"?

Cautiously optimistic, contradictory big model "second half"

Obviously, at present, this round of AI large-scale model boom has entered the "second half", and the fight is no longer about technology, nor does it mention topics such as "the gap with ChatGPT", but focuses on large-scale model applications, industry Ecology etc.

**But the paradoxical problem is that the application of the "on the face" large model seems to have "prospects", especially the strong combination of ChatGPT and Microsoft; the "inside" large model may be "crossing the gap" - ChatGPT user The growth rate has slowed down, the secondary market has fallen into a weak state, and domestic first-tier investors are generally cautious and optimistic to "enter the game" big model. **

**On the face of it, **Titanium Media App recently obtained a new report from Boston Consulting Group (BCG), which shows that among more than 12,800 employees in various industries in 18 countries and regions around the world, it was found that 86% of them say they need AI, especially generative AI; the majority of leaders (80%) say they regularly use generative AI tools; while 36% believe AI will eliminate their jobs; 71% Some believe that the rewards of generative AI outweigh the risks.

Another McKinsey report believes that 60% to 70% of jobs can now be automated, which is even more surprising than Accenture said a few months ago that 40% of working hours are currently affected. McKinsey also predicts that generative AI technology can greatly advance the time for human layoffs by 10 years. Sequoia Capital believes that almost all companies can use large language models in their products.

**Compared to more and more reports touting "big models", more data shows that ChatGPT can't rise anymore, and the industry bubble has begun to "burst". **

**First, the growth rate of platform visits has slowed down. **SimilarWeb data shows that the number of visits to the ChatGPT platform increased from 130% in January to about 60% in February and March, and nearly fell below 10% in April. The latest data shows that the growth in May was only less than 10%. 3%. At the same time, the average access time of ChatGPT users dropped from 8 minutes and 32 seconds to 7 minutes and 48 seconds. If there is no accident, the month-on-month growth of ChatGPT visits in June may be negative.

**Secondly, the AI boom in the capital market has faded. ** On June 26, a new tracking data quoted by Bloomberg showed that AI concept stocks began to fall. The AI concept stocks listed in China fell by 4.7% within a week, and more than 20 stocks are now down more than 10%. Giant stocks such as Intel and AMD are in a six-game losing streak, and AI concept stocks such as C3.ai have also begun to gradually pull back, falling by more than 3% overnight.

** Again, there is a huge gap between China and the United States in investment and financing in the primary market. **A statistical data published by an investor on a recent forum shows that from December last year to early May 2023, there have been more than 520 seed round and angel round financing projects in the US AI field, most of which are focused on generating There are very few companies competing with OpenAI to develop ChatGPT-style dialogue products; however, in China, there are only about 30 investment and financing cases in China’s AI field within five months, and the total amount is estimated to be less than 30. To 10 billion yuan, let alone the entire first quarter of 2023, China's fundraising, investment volume and amount are almost halved.

Finally, multiple incidents broke out, pushing this round of large-scale model entrepreneurship into a "great change": Kunlun Wanwei founder Zhou Yahui's ex-wife Li Qiong cashed out 2 billion yuan from the AIGC concept peak, and Kunlun Wanwei's stock price plummeted; On the one hand, AI large model startup pressure is high, Meituan co-founder Wang Huiwen directly resigned as director, left the founder and CEO position of AI large model company light years away, and went to the hospital for treatment.

**So, this is the real environment where the current AI model entrepreneurs are full of contradictions and investors are cautiously optimistic. **

** **Zhu Xiaohu, managing director of GSR Venture Capital Fund, watched the AI large-scale model project at Qiji Chuangtan 2023 Spring Roadshow (Source: Qiji Chuangtan)

On a deeper level, the heated debate between Fu Sheng and Zhu Xiaohu in the circle of friends is due to the difference in the thinking mode and industry perspective of dollar fund investors and the way of thinking of AI large model entrepreneurs.

An investor analyzed to Titanium Media App that Zhu Xiaohu represents the VC investment logic of holding USD funds. Investors of the US dollar fund hope to find a project in China that is recognized as a copy of the United States and has predictable development potential, which is enough to support a company worth hundreds of billions of dollars, and hopes to be in line with international standards. “I just look at it but not invest because I can’t earn the multiples that a project should get in China at present, or I can’t earn the capital premium and capital space brought about by the globalization of a Chinese project.”

From the perspective of generative AI (AIGC), the above-mentioned investor told Titan Media App that there are too many similar projects about AIGC at present. Traditional US dollar funds pay more attention to the future application of AIGC track in the consumer (To C) market, because there is no market bigger than China in the C-end application field, even if they spread more projects, they only seek a huge return in the end. At the same time, this is also the investment logic of Zhenge and Sequoia in the AIGC field, or they are also applying the "US dollar fund investment logic" to RMB funds, hoping to replicate successful projects like the Internet era.

Liu Di, managing director of Shengjing Jiacheng, told Titanium Media App that at present, the bottom layer of the AI large model is mainly data, computing power, and algorithms, and the core technology is relatively difficult. The computing power needs to be stacked with Nvidia graphics cards, and the algorithms can also use open source. There are regulatory restrictions, so investors' interest is reduced. "This means that everyone's starting line is the same, so the level gap is not big."

Liu Kai, a partner of Wuyuan Capital, said in May this year that compared with the era of the mobile Internet, after the release of the iPhone in 2006, people began to invest in some R&D tool-type App application companies. They thought this was a killer application scenario, but eventually established a good business model. It was in 2013 and 2014, especially when Toutiao, Kuaishou, and Xiaomi established some ecology, that they explored this commercialization. Therefore, in this wave of AIGC, many US dollar funds are waiting, because the United States has done it for two or three years, but China has done it for less than half a year. "We have suffered this loss before, so now we see more dollars and less investment. I currently think that the second half of this year to next year will be a high-frequency period for dollar funds to sell."

"Many start-up companies now connect to ChatGPT in the backend, make a UI design on the front end, and put it on the Apple store to promote that they are doing AIGC entrepreneurship." William Wong, managing director of Unknown Capital, said that this kind of "pipe skin" AIGC project, "no Technical barriers and business logic are just hype."

Gao Chao, founding partner and CEO of Yunxiu Capital, believes that industrial investors, including major domestic Internet companies, are more active in AIGC’s corporate layout compared to US dollar funds. There are very few opportunities for RMB funds to make real investments in the AIGC field. Of course, some RMB fund venture capitals actually invest in some early-stage physical projects, but the truly large-scale RMB funds have not yet entered the AIGC market.

As a company holding RMB funds, Wang Bing, a partner of Oriental Fortune, believes that RMB funds have recently paid more attention to companies that can be listed on the Science and Technology Innovation Board, which is the Pre-IPO investment projects mentioned by Liu Di earlier, such as semiconductor materials, Equipment, chip research and development and other very "hard technology" enterprises. However, China's environment in the field of software and algorithms is very different from that of the United States.

"We have seen several big gaps, first of all, companies. In fact, large Chinese companies have no boundaries. I will copy whoever does it well, regardless of To B or To C. Basically, there is nothing they can't do. Son. So for small companies, entrepreneurs who do algorithms or software need to be 'stuck' to a point where either big companies can't do it, or big companies won't do it, but the possibility of not being able to do it is not particularly high Start-up companies need to find some resources or barriers, which are places where big companies cannot enter.” Wang Bing said at a forum at the University of Science and Technology of China in May this year that they will only invest in such unique projects.

Huatai Securities believes that from an investment perspective, this round of large-scale models needs to find more beneficiary targets in the order of computing power infrastructure, hardware carriers, large-scale model platforms, and applications.

**Entrepreneurs represented by Fu Sheng have a thinking mode of "ChatGPT and Silicon Valley AI companies can work well, and we can do it too". **

For example, Fu Sheng mentioned in a speech that ChatGPT will bring about a productivity revolution. Today's Silicon Valley startups have taken advantage of GPT and have undergone great organizational changes. Maybe ChatGPT will surpass humans and rule everything in the future, but at present, it will take a long time to overcome these weaknesses, and the weaknesses of the big model are the entrepreneurial opportunities in the field of AI in China.

Another CEO of an AI company once mentioned to Titanium Media App that before he talked about large-scale models with investors, investors did not understand them at all. Companies can immediately follow. But once the product was released, investors questioned the profitability and commercialization of AI products.

This seems to be the "communication gap" between investors and entrepreneurs.

More investors are in the investment logic of "de-risking" during the current economic downturn-see more and invest less, or even not invest in the AIGC track at all, but hope to wait for the next opportunity. Liu Di told Titanium Media App that investors now prefer to see the goals of start-up projects that monopolize the track. However, entrepreneurs believe that investors do not understand the industry, and entrepreneurship in ChatGPT is an important development trend.

However, on the other hand, we can see that the situation of "good projects do not ask for investment" has also appeared on the AIGC track. Some investors pointed out that Wang Xiaochuan was rejected even if he wanted to invest a large amount of money as a large-scale model investor. An industry insider analyzed to Titanium Media App that Wang Xiaochuan’s large-scale model is favored by investors because he has found large-scale model application scenarios in medical care and education, but more large-scale model technologies do not have a vertical field scene output, so they cannot support investment Institutional support.

An investor told Titanium Media App that the current general perception among investors is that if they only do AI large-scale model technical services, the threshold for starting a business is really not high, there are too many entrepreneurial projects, and the funds are too scattered, especially on the enterprise end It has become a waste of money to buy Nvidia A100 cards and cloud services, without a quantified enterprise development process. Investors "learned a lot of lessons" and dare not invest again. Among them, big technology companies have a greater advantage. "The scale of the domestic market is so large. If you have three to five thousand customers, it's like playing in such a small industry. Investors will lose money no matter how they invest."

Large models are a classic winner-take-all area. More money, more computing power, and better talents are needed. Because better computing power means more people use it, more people use it means more data, and more data means better computing power results. The big model must be a battleground for the giants, who have money, technology, and more importantly, data.

Hu Yong, a professor at Peking University's School of Journalism and Communication, believes that big tech companies are likely to maintain a first-mover advantage because they have the time and market experience to hone basic language models and develop valuable in-house expertise. As a result, it may be difficult for smaller businesses or startups to successfully enter the field, concentrating the enormous processing power of large language models in the hands of a few big tech companies.

Bubbles are not necessarily a bad thing, large models still need underlying innovation

In a recent speech, Lu Qi, the former global executive vice president of Microsoft and the founder and CEO of Qiji Chuangtan, believed that the current era of AI large-scale models is still in its early stages. After rapid growth, the bubble burst will be inevitable. However, after the bubble bursts, a new dawn will dawn. At that time, a new generation of more powerful enterprises will rise from the broken bubble and occupy the commanding heights of the industry.

"Bubbles are not always a bad thing. There will be more outstanding talents and more capital will flow in." Zhou Hongyi, chairman of 360 Company, believes that the emergence of ChatGPT is a long-term accumulation process from quantitative change to qualitative change. Enterprises and entrepreneurs still need to be cautious when entering the game. If companies lacking computing power and other related capabilities follow suit, a lot of resources will be consumed.

**Obviously, at present, ChatGPT is just an example and does not necessarily represent a paradigm for the development of the domestic large-scale model industry. And more importantly, has this AI big model "bubble" burst? **

Bank of America strategist Michael Hartnett (Michael Hartnett) said in a new report that after the current stock market rose like the Internet bubble era in 1999, there are now tentative signs that investors are fleeing technology stocks. He's calling the current run-up in AI stocks a "baby bubble," which Hartnett warns may be bursting.

Art Cashin, head of floor trading at UBS (UBS), also said that the AI boom currently engulfing the stock market will become a "miniature version" of the Internet bubble at the beginning of this century. Well-known economist David Rosenberg (David Rosenberg) said that the artificial intelligence bubble is huge and will soon burst in an alarming manner.

"Its plug-ins currently have no product/market fit (Product/Marketing Fit)," OpenAI CEO Altman said in an interview last month, which was too transparent, and the minutes were quickly deleted from the Internet. "Many People think they want their app to be in ChatGPT, but what they really want is ChatGPT in their app."

**In addition, the lack of large-scale model underlying technology innovation in China is also one of the reasons for the bursting of this "bubble". **

At a recent roundtable conference on large models in Beijing, Lu Zhiwu, a professor at the Hillhouse School of Artificial Intelligence at Renmin University of China, said bluntly that many entrepreneurs in China are not calm enough to complete the base.

"The spring of domestically produced large models is all fake, because they are all fine-tuned on GPT and LLaMA. I have seen many large models, and I know they are fake after 10 minutes of testing. Of course, there are also a small number of large models. A lot of energy has been invested in the base, but most of them are still very impetuous. This is the biggest problem, but I think the gap in the language model will become bigger and bigger... You can see that a lot of models are coming out, just Because the architecture solution (solution) of this language model has already been announced, but if you haven’t announced it, why can’t it be done before, and it was only released in March, April, and May. Isn’t this a strange thing? I If they think it is illogical, no one is willing to make a base." Lu Zhiwu said.

Fang Han, CEO of Kunlun Wanwei, retorted on the spot, "I firmly oppose (this statement), we do large-scale model training ourselves, and I can tell you very clearly that we have nothing to do with LLaMA, because We have been doing (big model) since 20 years. So I don’t think you can kill all Chinese big model entrepreneurs (people) with one stick.”

Liu Di told Titanium Media App that investors have a clear view of the large-scale model industry. There are too many large-scale entrepreneurial projects in China, but many start-ups themselves do not have a clear and advantageous large-scale industrial application case.

Titanium Media App noticed that as of now, among AI computing power chips and domestic GPU chips, there is no domestic commercial delivery product that can replace Nvidia A100/H100 chips in large-scale model computing power training. In summary, the main reasons include limitations in the manufacturing process and product specifications, no double-precision floating-point function, no complete ecology of CUDA, poor decoupling with other chips, and high computing power loss due to communication network connection problems.

"In the first half of this year, the price of the entire GPU industry has been increasing, and the price has continued to increase by more than 25%. Under such a situation, today, no one can use commercial domestic chips for large-scale model training." A company in the field of large-scale models executives said in early June.

Yang Fan, co-founder of SenseTime, believes that AI infrastructure is essentially a trinity of computing power, data, and basic algorithms, including algorithm-related tools. In the end, who can provide the integration capabilities of the three, and the ability to provide lower costs and lower thresholds is the most important point in determining the entire competition. The large model is the improvement of technical value brought about by the continuous scale increase of the three elements of artificial intelligence, and it is also a perfect combination of basic research and development capabilities and the depth of system engineering capabilities.

Zhou Bowen, chair professor of Huiyan of Tsinghua University and founder of Lianyuan Technology, told Titanium Media App that the entrepreneurship of large models is completely different from the product logic of "small steps and fast iteration" in the previous Internet era. The large model needs to closely integrate the underlying AI technology with practical problems, and at the same time combine core capabilities with application scenarios.

**It’s worth noting that OpenAI plans to launch ChatGPT “personal assistant for work” to compete with Microsoft, The Information reported on June 27. According to reports, OpenAI's paid version for individuals and companies has recently gained more than 2 million subscribers and is expected to generate hundreds of millions of dollars in annual revenue. For OpenAI, building new ChatGPT functions will be the focus of its commercialization. **

"It seems that Altman doesn't know how to commercialize the product, and now it is considering making OpenAI conflict with his main partner... Things will start to go bad." An entrepreneur commented below the report.