Kai-Fu Lee: Chinese large-scale model companies do not have the number of GPUs from major American manufacturers, so they need to adopt a more pragmatic strategy

The future of AI is pointing north

2024-05-13 14:38Posted on the official account of Beijing Tencent Technology AI Future Finger North

Text / Tencent Technology Guo Xiaojing

Entering 2024, the competition of large models has entered a new stage, and the industry has shifted from "showing off the muscles of model capabilities" to not only focusing on the gradual improvement of model capabilities, but also emphasizing the growth of business ecology, PMF and users: from the perspective of model capabilities, GPT-4, as a golden indicator, seems to have become a prerequisite for model companies to explore subsequent applications, and how to credibly prove that model capabilities have reached the standard; In addition, how to consider open source and closed source on the road to commercialization in the future? Homogeneity of capabilities provided by large models, how to find PMF and differentiated competitive advantages. These problems are the most urgent problems to be solved in front of large model companies.

On May 13th, the first anniversary of the establishment of 0100000 Things, CEO Kai-Fu Lee announced the release of the Yi-Large model, and announced that the previously released Yi-34B and Yi-9B/6B models will be upgraded to the Yi-1.5 series, and each version will achieve the best performance in the same size. In addition, 010000 announced a product line from application to ecology, including the 2C productivity application "Wanzhi" and the 2B ecosystem "API open platform". At the press conference, Dr. Kai-Fu Lee elaborated on the company's dual-track large-scale model strategy and put forward the development direction of the large-scale model track.

Kai-Fu Lee believes that the development of the large-scale model track will depend on how to effectively achieve the fit of technology, product, market and cost. The training and service cost of large models is high, and the shortage of computing power is a common challenge faced by the industry. He called on the industry to avoid irrational cash-burning models in order to achieve the healthy and benign development of large models.

At this press conference, Kai-Fu Lee also talked deeply about the judgment on the future development of AI business ecology for the first time, and he believes that in the heyday of the mobile Internet, PMF (Product-Market Fit) was the core goal pursued by many start-ups. However, as large language models become the new focus of entrepreneurship, it is not enough to simply pursue product-market fit.

There is a decisive difference between the two eras at the startup infrastructure level - in the mobile Internet era, the marginal cost of user growth is low, but in the era of large models, the cost of model training and inference constitutes a growth trap that every startup must face. User growth requires high-quality applications, and high-quality applications are inseparable from powerful pedestal models, which are often backed by high training costs, and then need to consider the inference costs that grow with the scale of users. How and when this point of inclusion will be reached is becoming increasingly elusive.

Kai-Fu Lee believes that the concept of PMF can no longer fully define AI-First entrepreneurship based on large models, and that Technology and Cost should be introduced to form a four-dimensional concept - TC-PMF. "To do Technology-Cost Product-Market-Fit (TC-PMF), the technology cost X product market fit, especially the reasoning cost reduction is a 'moving target', which is a hundred times more difficult than the traditional PMF."

With the iteration of high-performance computing hardware and the popularization of model optimization technology, the significant reduction of the cost of large model inference has become a foreseeable trend. Under the premise that the inclusive point will eventually come, players who can be the first to perceive and reach the TC-PMF will undoubtedly have the first opportunity. To achieve this, the excellent capabilities of the "trinity" of models, AI infra, and applications are indispensable. Based on this, 010000 has put forward the "model-based co-construction" and "model-response integration" as the company's top-level core strategy.

1. Model companies rely on AI Infra capabilities to get out of the growth trap

In the first year of the domestic large-scale model track, the model structure has become the focus of the industry, and few people have noticed the importance of AI Infra.

A fact that cannot be ignored is that Chinese large model companies do not have the number of GPUs of major American manufacturers, so they must adopt more pragmatic tactics and strategies. AI Infra (AI Infrastructure) mainly covers the training and deployment of large models and provides a variety of underlying technical facilities.

"In the first year, the large model industry was in the volume algorithm, and in the second year, everyone was in the volume algorithm + Infra. In foreign first-tier manufacturers, the most efficient way to train a model is to co-build the algorithm with Infra, not only focusing on the model architecture, but starting from optimizing the underlying training method. Huang Wenhao, the person in charge of the training of the 010,000 thing model, said, "This puts forward new requirements for the knowledge and ability of large model talents. ”

At present, model researchers only focus on algorithms and ignore AI Infra, which is the status quo of the domestic large model industry. The Zero One Thousand Things Selection Model Team and the AI Infra Team are highly co-constructed, with a 1:1 ratio of people. "We require people who do model research to be 'down the board' and have engineering capabilities. This also aligns with the methodology of the TC-PMF that we advocate. Huang Wenhao said.

According to Kai-Fu Lee at the press conference, after multi-faceted optimization, the training cost of the 010 billion parameter model has been reduced by as much as double year-on-year.

Second, AI2.0 will no longer support bike-sharing money-burning wars

At the beginning of 2023, when the domestic large-scale model field was in a scuffle, various evaluation lists were overwhelming, and there were not a few models that ranked among the TOP of major lists. The large model has entered its second year, and the industry has entered a more realistic commercial landing stage, and customers/users will vote with their feet according to the capabilities shown on the application side. How to maximize the application effect based on the ability of the base model is an important topic to catch up with TC-PMF.

"Models, projects, algorithms, and products should be deeply integrated based on scenarios, and model longboards should be matched with high-value scenarios that are just needed, and AI-First workflows should be built to pursue the ultimate experience and solve user problems in one stop, rather than simply showing off the capabilities of the model and finding nails with a hammer." ”

In the view of Lan Yuchuan, the person in charge of the API platform of Zero One Everything, an API that has been fully verified by the business model overseas will be a better choice. APIs, which are standardized products, are more reusable, and their business models are closer to cloud services. Compared with AI 1.0's customized and re-delivery model, APIs can penetrate thousands of industries faster.

Since September last year, 0100000 has focused on productivity and social tracks to explore overseas applications, and 4 products have been launched one after another. According to Kai-Fu Lee, "At present, the total number of users of overseas productivity applications is close to 10 million, the revenue is expected to exceed 100 million yuan this year, and the product ROI is 1. ”

Due to the differences between the overseas market and the domestic market in terms of willingness to pay and market environment, Wanzhi currently adopts a time-limited free model. However, according to Cao Dapeng, the follow-up Wanzhi will launch a charging model based on product development and user feedback.

Kai-Fu Lee said: "The AI inclusive point brought by TC-PMF will eventually come, this time the market competition will no longer support the bike-sharing style of money burning war, and enterprises that adopt the business model of "horse racing" with funds will inevitably take the lead in exhaustion, calmly judge the development process of the industry, and polish TC-PMF in a down-to-earth manner is a more in line with the long-term route." This competition will include multiple aspects such as model, AI Infra, and product application.

Dr. Kai-Fu Lee said that the ofo-style subsidy logic is no longer applicable to AI 2.0, and hopes that the competition in the large model track will focus on achieving TC-PMF.

3. Interview record, Kai-Fu Lee talks about revenue composition, industry evolution and future choices

As an AI-First start-up company that has been exploring the large-scale model track for a year, Kai-Fu Lee and his team also have some first-line thoughts on the future business landing.

Q: What do you think about the development of domestic inference chips? And now there are a lot of emerging inference chip startups in the United States, what are the opportunities for domestic chips? Will domestic inference chips be considered in the future?

Kai-Fu Lee: We must fully support domestic training chips and inference chips.

However, the difference between the two is relatively large, and the training chip still has a certain degree of difficulty. The advantages of the inference chip are, firstly, that it is relatively simple, and secondly, it does not require such a difficult manufacturing process, so we are optimistic about the domestic AI inference chip, and we will also adopt it at the right time, and we have been paying attention to the investment opportunities in this area in Sinovation Works.

Q: What are the market feedback and data indicators that 010000 is most concerned about? What are the other independent development opportunities for large-scale startups?

Kai-Fu Lee: We are an AI start-up ourselves, from a unicorn to a super unicorn, it is undeniable that we have raised a lot of money and have rich computing power, in the words of my colleague Wen Hao, we are to pursue GPU first, and then consider applications. We think we can go out of the way that suits us more and be more pragmatic.

We are believers in pragmatic APIs, so we must train the best model we can train with the least number of chips and the lowest cost, and we will continue to explore and find TC-PMF.

I think one of the differences between domestic startups and Silicon Valley companies is that we can look up to the stars and at the same time keep our feet on the ground. To think about very complex problems is to predict how quickly the technology can advance, and then what kind of models we are capable of producing. Second, how to minimize the cost of reasoning.

Thirdly, of course, there is PMF, which is a domestic specialty. Because if you talk about large models, you will often feel that the United States has written the most papers in this area and has the highest talent density, but at the same time, you have to take into account that to be a great large and medium-sized company, it is not only the bottom, the bottom of course can not be bad, the answer is to be able to be in the world's first echelon, application is more important.

How to develop, how to find complex chips, and how to write together with people who understand models and applications, is something we don't generally see in companies in Silicon Valley.

We will not blindly pursue AGI with the mindset of pure power and miracles. OpenAI is qualified to try this path, but it's not the path we're going to take. The real value in the future must be who can build Douyin, WeChat, and Taobao in the AI era.

AGI is a dream, but if you use your dream to guide the execution of a company, and you don't have the financial resources to do a super-powerful miracle, it will be in vain.

Q: You said before that 2024 is the first year of the explosion of AI applications, but Zhu Xiaohu said that it is 2025, is there a difference between the two times? We're bound to have an explosion of apps, but what is the tipping point of the outbreak?

Kai-Fu Lee: I think what Zhu Xiaohu said has his point, just now I also mentioned that the inference cost will be reduced by 10 times in one year and 100 times in two years, so there will be a 100-fold drop in the inference cost, just like the ability to obtain GDP, as long as the price of 1/100 of the price today is used to do the inference cost, many applications will definitely be able to explode.

But I think this year is the first year because I think there are some areas that will explode this year, and the outbreaks will not come at the same time. I think this year is definitely going to happen in productivity tools, if you look at Office Copilot, although it's not necessarily AI, it's already being used by a lot of users, so I think it's going to be a bigger explosion next year.

Q: Mr. Kaifu, what is the revenue composition of Wanzhi Overseas 100 million yuan you mentioned? What is the expected growth in the future? In addition, seeing that many CEOs are now bringing goods, will you, as the chief experience officer of Wanzhi, also bring goods?

Kai-Fu Lee: I'll answer the last question first, I did open a short video account recently, but I won't live broadcast and sell things, I will share Wanzhi with you, it is a free product, I hope to help you use it well.

In addition, as you all know, I have written three books on AI, and I have always suggested that everyone should understand the importance of AI, whether it is for the education of children, their own work, etc., I also hope to be able to use my short video account to make a very honest sharing, and then let everyone understand how to use this tool well, don't just see if it will replace our work.

In terms of revenue, our overseas products have adopted a business model that can be considered healthy in the software sector, namely the membership subscription system. This model is similar to the membership service of domestic video sites, once the user subscribes, if he is satisfied, it will be automatically renewed for the next month, and we offer monthly and annual subscription options. This business model is superior to the traditional top-up or pure advertising model, where the revenue is closely related to traffic, and it is difficult to achieve profitability if it does not reach high traffic.

We've found that even though the number of users may not be large, the subscription fee can also support the company's growth and costs, which is very advantageous for startups. In addition, we have observed that despite reaching hundreds of millions of users in two months, the global penetration rate of AI products is still low. Specifically, since the launch of our product in September and October last year, the revenue has increased by 10 times or even 20 times, which shows that AI penetration is growing rapidly, and with the release of new models or products, our product strength and customer acquisition efficiency are also being optimized, which makes us confident in the healthy development and rapid growth of the company.

We remain optimistic about future revenue expectations and expect further revenue growth with the launch of a stronger model in the second half of the year or next year.

In addition, while we are proud to have achieved 100 million to 200 million in revenue this year, its Office 365 has billions of dollars in annual revenue compared to global production giants such as Microsoft Office, indicating that the market potential for production and production products is huge.

Q: When will the reshuffle of the AI industry take place? How many startups are likely to be eliminated?

Kai-Fu Lee: I think the Chinese market is currently in a dynamic and fast-growing phase, and along the way, we've seen a lot of competitors taking different paths. It fully demonstrates the rapid and dynamic growth of Chinese companies in the new environment. Therefore, it may be too early to draw conclusions on the market.

In the U.S., there is a common perception that only a handful of companies can train very large models, and these training costs are very high. At the same time, other companies are starting to look for solutions, such as developing AI models that are moderately sized and more suitable for commercial applications, which is both our direction and the direction that some U.S. companies are exploring. In addition, there are companies that are exploring ultra-small models or open-source models.

I'm optimistic about China's supermodel. I think the vitality of Chinese entrepreneurs is beyond simple imagination. While the market reshuffle may lead to the transformation or exit of some companies, we should not underestimate China's entrepreneurial and innovative capabilities.

Q: At present, the application scenarios of large models are mainly entertainment and productivity improvement, seeing that Zero One Everything is currently mainly focusing on productivity improvement, will it focus on this track in the future, or will it also explore other scenarios, if any, what are the specific scenarios?

Kai-Fu Lee: Human needs have remained the same for thousands of years, but with the baptism of technology, the way we work, the way we communicate, the way we play, the way we make friends, and the way we shop are constantly evolving. I believe that these changes will also happen in the era of AI. Among the many AI applications, we need to focus on which ones will ignite first. We're particularly bullish on productivity tools and entertainment social apps, even though they're still in their early stages and have limited use cases. In the future, we need to use TC-PMF ( to see which areas can achieve breakthroughs with the help of powerful models and what users are willing to pay.

At the same time, there are also some apps that users may not pay for directly, but their usage needs to be done first, because they are very sticky and spreading. Then, we will explore other business models to achieve profitability. This process is similar to the development process of the mobile Internet, just as the mobile Internet eventually developed a video subscription model, and a similar model will emerge in the AI field. In the era of AI, we will also see innovative applications such as Douyin, WeChat, Didi, and Meituan.

The key is to keep exploring which areas are good enough, inference cheap enough, and user demand strong enough. Whether it's through direct billing or other business models, the app that can finally be launched first and win will be key. I can't say much more about that.

View original image 547K

• 

• 

•