Li Kaifu stated that when he founded OneThing, he promised investors not to cash out within 10 years, and the best way to cash out was to “go public as soon as possible.”.
On May 13th, Li Kaifu, Chairman and CEO of Innovation Works and CEO of OneWorld, released the Yi-Large closed source model with a parameter of 100 billion, and upgraded the previously released Yi-34B and Yi-9B/6B small and medium-sized open source model versions to the Yi-1.5 series.
In the interview, Li Kaifu stated that today’s model is evolving very quickly based on Scaling Law, and everyone is only optimizing the application of GPT-3.5 one year and three months ago. GPT-4 has rewritten the capabilities of the application, so practitioners need to constantly predict how the technology will go in the future. It’s not based on what technology can do today, but on predicting how powerful technology will be in six months or a year.
Today, Li Kaifu clarified the strategic layout of Zero One Universe in the big model battle – doing both 2C and 2B business, both basic big models and specific applications. In the latest round of financing, OneThing is valued at nearly one billion US dollars, making it one of the six major model unicorns in China.
Previously, OneWorld released a one-stop AI work platform called “Wanzhi” for C-end users, similar to Microsoft Office 365 Copilot. It can provide meeting minutes, weekly reports, writing assistants, as well as speed reading documents and helping users create PPTs. In terms of 2B business, currently OneWorld has reached cooperation with Fortune 500 companies and provides API interfaces to the outside world.
In the specific implementation scenario, Cao Dapeng, the product manager of Zero One All Things Productivity, mentioned the uniqueness of Zero One All Things, stating that the product strategy of Zero One All Things is different from other companies in three aspects: firstly, it only applies AI First, which means that the application would not be valid without AI; Secondly, adhere to globalization; Thirdly, adhere to the TC-PMF concept.
PMF (Product Market Fit) was once the core goal pursued by startups, but with the big language model becoming a new focus of entrepreneurship, simply pursuing product market fit is far from enough. Li Kaifu believes that in the era of mobile Internet, the marginal cost brought by the growth of user scale is very low, but in the era of large model, the cost of model training and reasoning constitutes the growth trap that every startup must face.
Therefore, the concept of PMF can no longer fully define AI First entrepreneurship based on large models, and a four-dimensional concept consisting of Technology and Cost – TC-PMF should be introduced. “The decrease in reasoning costs is a ‘moving target’, which is a hundred times more difficult than traditional PMF,” said Li Kaifu. Specifically, in terms of the Zero One Big Model of Everything, it mainly aims to minimize the inference cost, ignite the point of universality, and through “model based co construction”, train the best model within its ability range with the least number of chips and the lowest cost. The “model” in “model based co construction” refers to the model, and the “base” refers to AI Infra, which is the infrastructure layer technology.
From the perspective of productivity, the difference of all things lies in facing the real pain points encountered by white-collar workers and knowledge workers in their work. Cao Dapeng gave an example of the products promoted overseas by OneThing, stating that when OneThing launched its independent overseas product in September 2023, it had a Chat with Doc function and had a ROI of over 1. Coupled with the requirement of a paid subscription system, the product is expected to have a revenue of 100 million RMB this year.
Lan Yuchuan, the person in charge of the growth and commercialization of the Zero One Everything Platform, said that in the era of AI1.0, there were many companies doing project delivery, but after making a one-time profit, there was no follow-up. The approach of Zero One Everything today is more like a cloud platform – providing API technology capabilities and a series of tools to help partners go live easily, while also providing industry solutions to optimize application scenarios for various industries. This is the direction and goal of the development of Zero One Everything API platform.
As for “integration of modeling and application”, which involves both basic large models and specific applications, Huang Wenhao, the head of training for the Zero One Universe model, said that he has recently communicated with many peers such as OpenAI, Aerospace, and xAI and found that the sense of fragmentation in overseas model applications is stronger than in the domestic market. On the one hand, excellent researchers aim for AGI and believe that its current application has little value. After AGI is implemented, the entire business model will undergo significant changes, so the industry should focus more on model capabilities. Consuming a lot of computing power to deploy certain applications can actually hinder the process of pursuing AGI.
On the other hand, Huang Wenhao stated that people who work on applications may feel that the improvement in model capabilities does not bring a significant sense of user experience. Spending a lot of money to pursue improvements in model performance or metrics has not actually solved many user problems. These two types of views are currently the main disagreements overseas regarding the integration of modeling and application, and also the sense of disconnection that Huang Wenhao believes lies in.
Compared to that, Huang Wenhao told reporters that Silicon Valley is more inclined to pursue dreams and China is more down-to-earth. However, “fragmentation” is indeed a problem that all large model companies will encounter. The most difficult part of this is how to coordinate and connect at the organizational and talent levels. For example, product managers need to understand the boundaries of model capabilities, and technical personnel need to understand the specific needs of products. “The basic model has strong generalization ability, but optimizing a specific application is actually a relatively difficult problem,” Huang Wenhao said.
Domestic large model companies have gathered a large amount of funds in the current primary market. Previously, there were rumors of founders cashing out on the dark side of the Moon. Regarding this, Li Kaifu did not comment on the specific situation of the company, but he promised investors that he would not cash out within 10 years when he founded Zero One Universe, and the best way to cash out was to “go public quickly”.
Li Kaifu stated that major foreign companies are currently updating their products, including the release of technology updates for OpenAI in the near future. Therefore, today’s goal of being ranked first in the world is the lowest, and it is just the beginning. “As foreign models progress, we will also make progress and hope to encourage each other to ensure that users around the world can enjoy the best models,” said Li Kaifu.