Image source: Generated by Unbounded AI

If you want to select a figure with both industry influence and industry controversy in the field of large models, Kai-fu Lee must be on the list.

When he decided to personally step down as CEO of Zero One Thousand Things in mid-2023, Kai-fu Lee was once compared with Wang Huiwen in the industry to discuss whose entrepreneurial projects were more expensive. As Wang Huiwen withdrew due to illness, more than a year later, Kai-fu Lee, who was still at the big model poker table, was also in trouble.

The latest controversy occurred in early January. On the night when the news of "Alibaba's acquisition of Lingyiwu" came out, Kai-fu Lee took the lead in refuting the rumors in WeChat Moments, and accepted several interviews with the media the next day to tell the outside world that Lingyiwu was not disbanded, but had just given up on the matter. For the training of very large-scale parameter models, we will still focus on training small and medium-sized models in the future, and focus on the application side.

Before this, Lingyiwuwu had also been caught in multiple rounds of doubts about "shelling Meta open source large model LLaMA" and abandoning pre-training.

Kai-fu Lee, who is fighting for the title of China’s “Godfather of AI”, has been repeatedly questioned on the large-scale model track he chose. Some radical industry insiders even directly referred to Kai-fu Lee and Zero-One Everything behind him. As a "big model bastard", Jiang Yi, a founding partner of Hengye Capital, told Alphabet (ID: wujicaijing), "In fact, we all knew it earlier. It’s not a big secret in the industry that (Lee Kai-fu) wants to give up pre-training.”

Different from his previous role as a preacher in China for more than 20 years, he has changed his identity and become an entrepreneur. Kai-Fu Lee, this time, is on the same starting line as others. Technology, products, and commercialization are becoming multiple-choice questions, sifting out the laggards while competing to find the winners. Kai-fu Lee, who has no influence but temporarily lacks industry recognition, has increasingly given over to the new image of an AI internet celebrity with his aura as the Godfather of AI.

Just like Lei Jun chose automobiles as his last entrepreneurial venture, Kai-fu Lee regards large models as his last dance. After ChatGPT was unveiled in November 2022, the moment he saw it, Kai-fu Lee "realized that it would appear sooner than I expected... I couldn't miss this opportunity, I had to be involved."

Yes The desire for greater success has become the biggest incentive for people like Lei Jun and Kaifu Li who have already achieved success to join the game again.

In 2010, after Lei Jun finally decided to start a smartphone business, Xiaomi co-founders Li Wanqiang and Lin Bin didn’t understand and asked Lei Jun, “You have everything, what do you want to start a business?”

< p>Lei Jun once reflected on his entrepreneurial experience before Xiaomi and felt that many products were only one step away from success. , but always missed the opportunity, and believed that "this is the result of the times and the company's genes, and there is no way to make a comeback."

Especially in October 2007, after Kingsoft was listed in Hong Kong, it looked like there was only one mainstream Internet company with a market value. A fraction of a listed company, Lei Jun once told Caijing that he once thought Jinshan could carry his dreams. Now,He wants to build a company worth tens of billions or even hundreds of billions of dollars. "Otherwise, how valuable is it if you build one billion-dollar company after another? There is no value in copying yourself again and again."

Kai-fu Lee's view on success His desire is 10 times greater than that of Lei Jun. The reason why Kai-fu Lee gave up on investing and chose to do it on his own was that he "wanted to create a great company" and believed that his career experience at Microsoft enabled him to learn how to build the next potential trillion-dollar market value. capabilities required by the company.

However, as Kai-fu Lee abandoned his AGI (artificial general intelligence) dream, his vision of a trillion-dollar company became more illusory.

In the wave of domestic large-scale model entrepreneurship in the two years since the emergence of ChatGPT, Kai-fu Lee is still the AI ​​entrepreneur who most frequently publishes rumors refuting news in WeChat Moments.

After news spread online in early January that "Alibaba acquired Zero One Wan," Kai-fu Lee once again refuted the rumor through WeChat Moments, and then directed Zero One Wan to issue an official statement to refute the rumors.

Based on Kai-fu Lee’s external response information, after the latest organizational structure adjustment, Zero One Wanwan will hand over the training of very large models to Alibaba in the future, while he will focus on industry models with small and medium parameters.

This change came earlier from Kai-fu Lee’s thinking in May last year. At that time, Zero One Thousand Things had chosen not to blindly pursue expanding the number of parameters, and abandoned the plan to train the ultra-large model Yi-X-Large with one trillion parameters, and instead trained smaller, faster and cheaper MoE hybrid experts. Model Yi-Lightning.

Before refuting the news that "Alibaba acquired Zero-One", in October last year, Kai-fu Lee also refuted the rumor that Zero-One had stopped pre-training by posting on WeChat Moments, and responded firmly at the time that Zero-One had stopped pre-training. All things will never give up on pre-training work.

Earlier in November 2023, Kai-fu Lee, who was determined to develop a universal large model base, was caught in a questioning of the Meta open source large model LLaMA. At that time, Kaifu Lee also immediately clarified through WeChat Moments.

Although the content of the three refutations of the rumors are different, the main reason why these rumors have taken turns to find Zero One Thousand Things is that the company is too poor.

In addition to the Six Little Tigers of the same large model (in addition to Zero One and All Things, it also includes Dark Side of the Moon, MiniMax, Zhipu, Baichuan Intelligence, and Step Star), they have won one after another. After multiple rounds of financing from well-known investment institutions and major manufacturers, since the establishment of Zero One Thing, the financial situation has been exposed. There are only two investment news: first, a new round of financing led by Alibaba Cloud will be completed in October 2023, with a post-investment valuation exceeding US$1 billion, ranking among the unicorns of China's large model startups; second, it will be completed in August last year Hundreds of millions of dollars in new financing, investors include an international strategic investment, Southeast Asian consortium, etc.

Alibaba Cloud has become the only public and clear investor in Zero One Thing. Worse than the small amount of financing, the financing environment is becoming increasingly difficult. Jiang Yi said that entering 2025, the industry willInvestment in models will become more cautious. "Similar to Kai-Fu Lee's decision to abandon pre-training, this is almost always the case in the Six Little Tigers to varying degrees. It depends on how long they can sustain their funds before announcing their strategic adjustments."

In addition to the difficulty of commercial realization of large models themselves, the dislocation of the ecological chain also makes this wave of AI entrepreneurship increasingly difficult. According to Li Hui, a domestic entrepreneur who also develops large models, the originally expected scenario was that the chip would make the least money, the platform would make a certain amount of money, and the application would make the most money. But the reality is that Nvidia, which provides GPU chips, earns more than 80% of the industry's profits. "But this inverted triangle situation will not last forever. In 2025, with the explosion of AI native applications, things will change."

Zero One Wanwan, which cannot wait for the improvement of the ecological environment for the time being, has embarked on a path of diligent and frugal AI development since its establishment.

In an interview with Zhang Peng, founder of Geek Park, in June last year, Kai-fu Lee predicted that half of the companies may give up pre-training in the next year. "Some start-up companies cannot raise money. If you are a start-up company, you only have, say, 100 million US dollars left, and then you have to burn 20 million US dollars a month. You need to burn it all in five months, and finally burn one The model beat Google, but it still took three years to get this first Spending 100 million US dollars to build an APP? I don’t need to say more about this.”

In the future, Kai-fu Lee will no longer pursue training super large models, which also means that Zero-Yiwu will become the number one in the industry. Players who have announced that they have given up on pursuing AGI. "From a business perspective, we believe that only large companies can continue to build ultra-large models."

Byte has become the best example of Kai-fu Lee's above-mentioned judgment. Byte, which was once lagging behind in the large model track, has been catching up for more than a year and has jumped into the domestic first-tier echelon by the end of 2024. It has not only supplemented the technical performance of the basic model, but also successively laid out products. Doubao, Cici, Coze, Xinghua, Cat Box, Doubao Aixue, Jimeng, etc. cover mainstream application tracks such as AI intelligent assistant, AI social, AI tools, AI education, and AI video.

But after handing over AGI to a large company, Kai-fu Lee will undoubtedly face more scrutiny and questioning in the future.

If we turn the time back to 20 years ago, the current doubts about Kai-Fu Lee would be somewhat unimaginable.

In July 2005, Kai-Fu Lee, who decided to join Google, began to face the lawsuit of the century in his life. When his former employer, Microsoft, sued him for violating a non-compete agreement, Ballmer, then Microsoft CEO, once called Kai-Fu Lee China's "godfather" to prove his influence in China.

The title of Godfather that Ballmer gave Kai-Fu Lee at that time was not fully deserved. Subsequently, by preparing to establish Google China and promoting cutting-edge technology to more people in China, Kai-fu Lee was officially recognized by the industry and became known as China's "Godfather of AI".

Li Kaifu’s life was also related to AIAn indissoluble bond was formed. In 1983, after graduating from Columbia University with first place in the Department of Computer Science, Kai-Fu Lee entered Carnegie Mellon University to pursue a PhD in computer science, and selected the research direction of speech recognition, successfully developing the world's first " In 1988, "Business Weekly" awarded Kai-Fu Lee the honorary title of "The Most Important Scientific Creation Award" for this invention.

In the summer of 1990, 28-year-old Kai-Fu Lee decided to give up his position as an assistant professor at Carnegie Mellon University, join the industry from academia, and join Apple. He led the team to invent QuickTime and became Apple's most popular software engineer. Young Vice President. Subsequently, Kai-Fu Lee switched jobs from Apple to SGI, Microsoft and Google.

The experience of participating in the establishment of Microsoft China Research Institute and Google China not only enabled Kaifu Lee to interact with more AI talents, but also cultivated and discovered a large number of related talents. Microsoft Research Asia, which was separated from Microsoft Research China, once became the Whampoa Military Academy in China's Internet and AI fields. Famous big names in the industry such as Zhang Hongjiang, Zhang Yaqin, Shen Xiangyang, and Wang Jian were all born here.

In addition to the efforts of the industry, Kai-fu Lee also attaches great importance to communication with college students. As early as 1990, Kaifu Lee held a four-week course at the Beijing Institute of Information Engineering to introduce cutting-edge AI knowledge to Chinese college students. In 2003, Kaifu Li also funded the creation of "Kaifu Student Network", a platform for communication and interaction with students. In 2006, it was renamed "Wuxue.com". In the above, Kaifu Li said that he answered tens of thousands of questions from Chinese students.

The preparation of the Innovation Factory has become the final weight in shaping Kaifu Lee as China’s “AI Godfather”.

By 2008, working at Google China could no longer satisfy Kai-fu Lee’s self-pursuit. “I increasingly realized that managing a larger team was not my goal. I preferred to start from scratch. Some people create, rather than run a giant.”

A realistic background that cannot be ignored is that China at that time was on the eve of the transition from the PC era to the mobile Internet era, and a large number of entrepreneurs were there. Investment institutions have emerged. In August 2009, Kaifu Lee also founded his own investment institution, Sinovation Ventures, and invested in a number of well-known companies, such as Mobike, Zhihu, Wandoujia, Moji Weather, Meitu, etc.

With Kai-fu Lee’s many years of experience in the technology field, Sinovation Ventures took the lead in betting on cutting-edge technology tracks. “Invest early, invest in small, invest in technology” once became one of its most important labels.

By 2016, the outbreak of the "Man-Machine War" - AlphaGo's victory over Go ninth-level player Lee Sedol - turned AI, which was originally limited to the fields of science and science fiction, into a hot topic on everyone's lips. . Sinovation Ventures also began to further focus. In the same year, it established the Artificial Intelligence Engineering Institute, with Li Kaifu as the dean. It is committed to the ecological link and technological transformation of AI technology and scientific research, laboratories, universities, and talents, creating a set of "VC +"AI" new investment model, official propaganda stated that it has invested in more than 10 artificial intelligence unicorn companies.

Megvii Technology, which was invested in 2012, became the first AI company invested by Sinovation Ventures. Sinovation Ventures’ list of AI unicorns includes Meitu, Horizon, 4Paradigm, Uisee Technology, Momenta, WeRide, Bitmain, Feibu Technology, XAG, Magga, and Innovation Qizhi Waiting figure .

But the risks of investing in cutting-edge technology also come with it. Most of these companies have not been able to break even due to their advanced layout, and only Meitu, which seems not to be cutting-edge, has achieved profitability. < /p>

Others, such as Uisee Technology, Momenta, XAG Technology, etc., have not yet reported any listing news, or, such as Bitmain, Megvii Technology, etc., have encountered setbacks in their listings.

Especially as companies. Sinovation Ventures’ AI investment debut Megvii Technology, even though it was known as one of the "Four Little Dragons" of the previous wave of AI entrepreneurship, chose to give up after three years of wandering in front of the listing gates of the Hong Kong Stock Exchange and the Science and Technology Innovation Board. Shi Technology officially withdrew its listing application.

Those companies that have been successfully listed, such as Horizon, WeRide, 4Paradigm, Innovation Qizhi, etc., are still the same as 01Wang founded by Kaifu Li. Facing the soul torture of when to make profits, and its stocks The price has experienced a parabolic decline, and the stock price is currently falling below the issue price.

At this time, the benefits of being an investor were revealed as early as the beginning of 2018. At the time of the epidemic, Kai-fu Lee made it clear in an interview that everyone had overestimated the commercial monetization capabilities of AI. “Every entrepreneur wants to package their startup company as an AI company, and every VC also wants to advertise it,” he said. I am an AI investor. However, AI investment is actually not suitable for novices. If you rush to participate in the early stages of AI without understanding the technology itself, such people will lose their pants. ”

This obviously does not include Kai-fu Lee, who is known as China’s “godfather of AI.” When asked later in 2019 which fund of Sinovation Ventures had the best returns, Kai-fu Lee replied: “Invest in AI.” The one with the most projects has the best rewards...for example, Megvii's rewards are 400 times, and VIPKID's rewards are 1,200 times. ”

In addition to being an investor on the sidelines, and adding a new role as an AI entrepreneur, Kai-fu Lee also had to face the problem of AI commercialization head-on.

After announcing Zero Yiwuwu no longer pursues the training of super large models, but continues to train faster and cheaper models with moderate parameters. Before creating profitable applications based on the latter, Kai-fu Lee has already changed the company's commercialization path: giving priority to domestic development To B field, To C applications will be developed overseas first, and then domestically. The reason given by Kaifu Li is that the willingness to pay is relatively high in foreign countries, and the fastest path to commercialization will be explored first, and there will be opportunities in the C side in the future.

But Kai-fu Lee can still stick to this new strategy of insisting on training small and medium-sized models.How long will be a question mark.

Following the 100-billion-parameter model Yi-Large, in October last year, 01Wanyuan released its new flagship model Yi-Lightning. Yi-Lightning." Kai-Fu Lee said.

When he refuted rumors about abandoning pre-training in October last year, Kai-fu Lee gave two criteria for whether a large model company should do pre-training: First, see if you can pre-train a model that is better than open source. ; Second, see if the cost of pre-training can be afforded, and how long can it be used after training?

At present, one of the reasons why Kai-fu Lee insists on continuing to do small and medium-sized model pre-training is that his Yi-Lightning model can still be used, "so I must do it myself. Of course, pragmatically speaking, if one day the model is open source If it defeats us, we will definitely have to think about whether we should stop doing it.”

Once the support of the model layer is lost, the “Infra+Model+Application” trinity layout originally conceived by Kai-fu Lee may not be possible. No readjustment.

In Kai-fu Lee’s vision, Zero-One Wanwan aims to be an ecosystem, not just a large model company, just like Microsoft in the PC era. "At that time, Microsoft made the Windows kernel, which is equivalent to today's large model; it made Windows, which is equivalent to today's APIs and platforms; it made Office, which is equivalent to today's applications; and then it made infrastructure, which is equivalent to today's Azure Cloud." Kaifu Li gave an example said.

In addition to infrastructure and models needing to be done together, models and applications must also be done together, "because they are actually one." Kaifu Li judged.

But as Zero-One Everything moves more and more into the field of large-scale industry models that are bound to the interests of corporate customers, in addition to obtaining certain business returns, it also loses the possibility of large-scale expansion. sex.

After losing the AGI entrepreneurial opportunity, the opportunities left for entrepreneurs to create the next BAT have also become infinitely slim. Fu Sheng had a clear understanding of this aspect earlier than Kaifu Li, "I have given up on the entrepreneurial dream of BAT. There is really no chance."

In a recent interview with Zanzi, when faced with the idea of ​​building AGI When asked whether the Microsoft dream of the era was shattered with the abandonment of ultra-large-scale pre-training, Kai-fu Lee replied, "Of course not... It can only be said that we are starting from applications. Anyone can look up at the stars, and more importantly, Keep your feet on the ground."