Li Kaifu's Exclusive Response: Blindly insisting on something unaffordable is not a healthy choice | Zhiyong Exclusive
"We are a commercial company, not the 'Zero-One Everything Technology Laboratory'."
Give up the model? Completely acquired by Alibaba? In the first week of 2025, "Zero-One Everything", one of the six big models, is deeply involved in a public opinion vortex.
On January 7, 2025, "Intelligent Emergence" had a conversation with Kai-Fu Lee, the founder of Zero-One Everything at the center of the vortex. This "most experienced AI entrepreneur" responded to the recent personnel changes and business splits of Zero-One Everything:
"The strategy of a startup in the first year may not be suitable for the second year. Adjustment and transformation are inevitable for entrepreneurship. This year is the decisive year for commercialization, and the business priorities of Zero-One Everything should also be adjusted accordingly." Kai-Fu Lee told "Intelligent Emergence".
Just one day ago, a piece of news circulating on the Internet directly pointed out that Zero-One Everything "sold the card and pre-training team to Alibaba". That night, Kai-Fu Lee personally refuted the rumor in his Moments.
Kai-Fu Lee's response in Moments.
The key to this change lies in the whereabouts of the pre-training and Infra teams of Zero-One Everything.
Kai-Fu Lee told "Intelligent Emergence", "Members who are willing to continue training super-large parameter models have joined the 'Industrial Big Model Joint Laboratory' established by Zero-One Everything and Alibaba Cloud."
"Intelligent Emergence" learned that in mid-December 2024, the two teams received adjustment notices one after another. Then at the end of December, the pre-training team received an offer from Alibaba's "Tongyi", and the Infra team received an offer from Alibaba's Intelligent Cloud team.
The background of "Alibaba's Incorporation" comes from a judgment of Zero-One Everything: The cost-effectiveness of investing in super-large model pre-training for startups is too low.
"Everyone can see clearly that only large companies can burn super-large models." Kai-Fu Lee told "Intelligent Emergence" that the goal of Zero-One Everything since 2024 is to make small-parameter and moderate industry models. "The usefulness of super-large models is that they can teach smaller models, so we need a large company that can train large models to cooperate with."
The "large company" in Kai-Fu Lee's mouth has become Alibaba, an old shareholder of Zero-One Everything. "There are still many members in Zero-One Everything who have the dream of AGI, and these members can choose to join the laboratory." Kai-Fu Lee mentioned.
As for the rumor of "the card being acquired", Kai-Fu Lee said directly that Zero-One Everything trains models through the cloud service model: "We don't own the card ourselves. How can we sell it?"
How to make money has become the most important proposition for this unicorn in 2025. In addition to adjusting the model training strategy, Zero-One Everything also considers splitting AI businesses such as games and finance for independent operation and financing.
The logic of the split comes from the experience of incubating projects during the Innovation Works period. Kai-Fu Lee told "Intelligent Emergence" that when the team is very focused on a vertical category, they can do it more deeply and thoroughly.
Of course, the logic of Zero-One Everything's business split is also quite realistic, "First, talk to investors to see if anyone is willing to invest."
"The survival way of a startup is to consider how to make good use of every dollar, not to get more GPUs to burn." Kai-Fu Lee concluded.
The following is the conversation between "Intelligent Emergence" and Kai-Fu Lee, slightly edited:
Super Large Model, Handed Over to Alibaba for Training
"Intelligent Emergence": Focusing on the event, what exactly happened to the pre-training team and Infra team of Zero-One Everything?
Kai-Fu Lee: We have been communicating with Alibaba, our largest investor, and we have also established a joint laboratory (Industrial Big Model Joint Laboratory). The joint laboratory will do more work with the direction of Scaling Law, which is led by Alibaba.
Then some of our members are good at and willing to invest in Scaling Law, so some teams will be deeply integrated with Alibaba through the joint laboratory.
"Intelligent Emergence": Does it mean that the pre-training team and Infra team can choose to go to Alibaba or stay in Zero-One Everything?
Kai-Fu Lee: Let's not talk about these details for now.
What can be revealed is that there is now a huge opportunity to make a super-large model. Alibaba has decided to move forward, and we applaud Alibaba. In the past year and a half, we do have some excellent members who still have a great passion for this matter. So a two-way choice is very natural.
"Intelligent Emergence": Does the pre-training and Infra team joining Alibaba mean that Zero-One Everything has officially given up pre-training?
Kai-Fu Lee: Our idealized "pre-training" is to make a practical, small and fast model, and then evaluate it based on commercial cost-effectiveness.
Previously, Ilya, the former co-founder of OpenAI, said that Scaling Law has reached its end. It means that a huge model requires a lot of money, and the efficiency obtained will not increase as before. So only large companies can burn super-large models.
With the relatively good market feedback of Yi-Lightning (Zero-One Everything's 100-billion-parameter model), we also see that the role of Zero-One in the future should be to make relatively fast, small, and cheap models. The models trained in the future will not be larger than Lightning.
"Intelligent Emergence": In early January, Zero-One Everything and Alibaba Cloud established a laboratory to train a large model. What are the role definitions of the two parties?
Kai-Fu Lee: The usefulness of a super-large model is that it can teach smaller models. We call it a "teacher model". A large model uses methods such as data distillation and data synthesis to teach the capabilities of small models. It is a training model strategy.
Although a large model may be expensive and relatively slow, it is an innovative ecological niche and a safety bottom line that Chinese large models must occupy and adhere to in response to technological blockades.
So we think that we need to cooperate with a large company that can afford to burn a large model. In the future, the super-large model will be trained by Alibaba, and we can use a small and refined team to make small and cheap models to embrace the explosion of applications.
Many people still have the dream of a super-large model, so these people are very suitable to join the joint laboratory led by Alibaba.
"Intelligent Emergence": It is rumored that the card (referring to GPU) of Zero-One Everything was sold to Alibaba. What's going on?
Kai-Fu Lee: I don't know where the card thing came from. We don't own the card ourselves. How can we sell it? So I can't respond to the card thing.
"Intelligent Emergence": Is it the norm for AI startups not to buy cards to build clusters but to purchase cloud services?
Kai-Fu Lee: Everyone believes that small companies are not suitable for super-large models, so those small companies with tens of thousands of cards may need to reconsider. If they don't burn large models, should they reduce the number of cards?
I think this is a common thinking in the industry. We don't do this not because we don't believe in Scaling Law, but because we hand over the matter of burning super-large models with more cards to large companies that can do Scaling Law, such as Alibaba, and then we cooperate with it. This is the way to survive.
Seventy Percent of Revenue Comes from the B-side
"Intelligent Emergence": In your Moments refutation, you mentioned that the confirmed revenue of Zero-One Everything in 2024 has exceeded 100 million yuan. What are the main sources of this business?
Kai-Fu Lee: In fact, we started with a focus on To C, and for To C, we started with going overseas. We have a clear understanding that it is relatively difficult to commercialize To C in China in 2024, so in China, we basically resolutely do not burn money to acquire customers and do To C products.
In the C-side, about 20 - 30% of the performance we achieved last year came from paid products going overseas, such as PopAi, which is a productivity tool.
"Intelligent Emergence": Does that mean that 70% of the revenue comes from the B-side?
Kai-Fu Lee: For To B, in the second half of 2024, we experienced a half-year strategic transformation. Because we first polished the big model technology, the process is like finding scenarios through technology.
So for To B, we also made a variety of attempts. One of them is the game industry, and there has also been a relatively good growth in the second half of the year. We also have a good growth in the financial and energy industries.
"Intelligent Emergence": From To C in the first half of the year to To B in the second half of the year, the change in business form is quite significant.
Kai-Fu Lee: Maybe the outside world will feel that the changes have been relatively large since the beginning of January 2025. In fact, we have been planning for several months. Our overall organization, resource allocation, and project priority ranking, what to do and what not to do, have actually gone through a systematic overall review, and some milestones have been reflected at the end of the year.
For example, we have been discussing the joint laboratory with Alibaba for a while, and it was only recently officially announced.
"Intelligent Emergence": In the past year, what changes have there been in the overall organization and resource allocation of Zero-One Everything?
Kai-Fu Lee: Our organizational structure adjustment mainly lies in increasing the function of To B to bring more cooperation in industry big models.
So our internal team has also made corresponding adjustments. For example, for landing, we need a very strong pre-sales team, as well as a team that listens to customer needs, follows up on planning products and business models. We also need a strong engineering R & D because these applications need to be standardized and platformized. We insist on not doing To B business that loses money for every order.
"Intelligent Emergence": Zero-One Everything focuses on training small models. How to commercialize them?
Kai-Fu Lee: We believe that extremely fast and cost-effective models are particularly suitable for making industrial models. This does not mean that we are going to make 3B or 4B models. These too-small models do not work in many scenarios.
At the same time, we will not continue to make super-large parameter models because these super-large models are also not applicable in many fields such as automotive companies, finance, and games. What we want to do is to make models similar in size to Yi-Lightning or slightly smaller, which can be applied to more industry scenarios.
"Intelligent Emergence": Zero-One Everything started relatively late in the industry To B. How to make up for the latecomer disadvantage?
Kai-Fu Lee: We will focus on several businesses. Unlike some AI companies that can hire 300 salespeople, we will use my personal connections to cut into some very good fields and companies, and then find fields where standard solution can be made.
At the same time, we can link up with Innovation Works. Many companies in various fields incubated by Innovation Works are very complementary to us and can cooperate.
And we think that the industry model is not detailed enough. For example, I don't think finance is a track. We will further subdivide the track on the basis of the industry. Here I can't talk about the details because if I do, others will know what you are doing.
We Are a Commercial Startup, Not a Technology Laboratory
"Intelligent Emergence": We understand that at an internal meeting in early January 2025, you mentioned that Zero-One Everything should "fully align with applications".
Kai-Fu Lee: At the beginning, we definitely carried the ideal of pursuing AGI, attracting many colleagues with a strong technical spirit, and verifying that we can make world-class technology.
But to some extent, This year will be the elimination year for the commercialization of big model companies. If everyone cannot overcome this hurdle and cannot verify that the technology can be truly landed in applications, they will face the possibility of being eliminated.
"Intelligent Emergence": What is the evaluation standard for "truly landing in applications"? User number? Revenue? Or profit?
Kai-Fu Lee: We think the core is that This application must be able to make money and bring in income, rather than blindly increasing the quantity and blindly increasing the number of users.
So to "fully align with applications" means: When you really have technology but no application, this is basically just a laboratory.
We are a commercial company, not the "Zero-One Everything Technology Laboratory".
"Intelligent Emergence": I heard that the company has split some businesses and established subsidiaries for independent operation and external financing.
Kai-Fu Lee: This is actually very similar to the incubation model that Innovation Works has been running for a long time. For example, the split of "Zero-One Oasis" (Zero-One Everything's AI game subsidiary) that was reported earlier, this business has also been running for several months, and we think it can take off.
"Intelligent Emergence": Why split?
Kai-Fu Lee: When your team is very focused on a vertical category, and this vertical category is a large vertical category that you predict can be expanded, they will be more focused and more likely to do this industry more deeply and thoroughly.
So compared to being in a team that wants to do multiple industries, after being independent, it can be more focused and can also access exclusive resources.
"Intelligent Emergence": What is the criterion for judging whether a business should be split or not?
Kai-Fu Lee: First, talk to investors to see if anyone is willing to invest.
If you split blindly, and in case the team itself cannot fully generate its own blood, it is actually unnecessary.
"Intelligent Emergence": Few startups do so many business directions at once. Will it be distracting?
Kai-Fu Lee: We still hope to be able to achieve a To B platform in the future. There are many central R & D costs that are shared by the business lines. Maybe 20% is spent on this one, 30% is spent on another one, and 50% is used for the common tool chain for everyone.
At the same time, the business runs, sometimes because of the trend, sometimes because of the concept, it is difficult to plan, and different directions have to be done while looking. Of course, we hope that we can make several star product lines this year.
"Intelligent Emergence": After this year and a half, what is your new understanding of AI entrepreneurship?
Kai-Fu Lee: We must be brave to actively adjust and see clearly what is the correct strategic route. The strategy of a startup in the first year may not be suitable for the second year. At this time, if you blindly insist on something that you cannot afford, it is not the correct and healthy choice for a startup.
We hope that our future will have a stronger certainty, so that we will be more confident to achieve the goal.
