This year, there have been significant technological breakthroughs such as Sora and GPT-4o, and people may be more concerned about when manufacturers will produce Super Apps with billions of daily activities and billions of dollars in revenue. Although Super App has not yet appeared in China, we are still optimistic. Looking back at the mobile Internet, iPhone appeared in 2007, and WeChat only appeared in 2010, four years apart. So there's no need to rush, the Super App may be born at any time Recently, Zhou Feng, CEO of NetEase Youdao (NYSE: DAO), stated in an interview with Caixin News Agency.
Through this interview with Caixin News Agency, it was observed that the big model has entered the "landing year", with multiple manufacturers sounding the horn of price war. Questions such as whether the universal model API calling mode can work, whether the big model is seriously homogenized, whether price reduction can accelerate application explosion, and whether cloud vendors are experiencing internal competition and overflow have been hotly discussed. At the same time, new features from vendors have been intensively launched, and the trend of integrating production and modeling is currently recognized by many practitioners.
As for NetEase Youdao, its launch of Ziyue is the first education model in China, and it became one of the first education models to pass national registration at the end of last year. In its latest financial report, NetEase Youdao also released a signal of significant growth in AI subscription services and the effectiveness of AI implementation.
"Price reduction is a good thing for the industry"
"Maybe cloud manufacturers hope to attract customers after price reduction. I don't think this action of cloud manufacturers is the first or the last one on the Internet." Last week, players of domestic ByteDance, Alibaba, Baidu, Tencent, iFLYTEK (002230. SZ) and other big model racetracks announced price reduction or free. In Zhou Feng's view, price war is a market behavior of related manufacturers and a way to attract customers for cloud based business.
He told Caixin News that a decrease in prices is definitely a good thing for the industry.
Upon investigation, Zhou Feng stated that the cost of the model is indeed decreasing. According to his observation, the cost of large models is basically decreasing at a rate of 50% per year, and the 50% decrease is due to the combined impact of computing power and model improvement. The computing power itself is also becoming cheaper. "Cost reduction is a good progress because large models are still too expensive, such as the 70B to 130B parameter large models that everyone is concerned about this year, which are expensive for both training and inference," he said.
It is worth noting that regardless of training or reasoning, computing power accounts for the most important part of the cost of large models, while GPU supply and demand and price fluctuations continue to stir up market nerves. Recently, some media quoted informed sources as saying, "In some cases, the Nvidia H20 chip is 10% cheaper than the Ascent 910B." This fact remains to be confirmed. Zhou Feng, from the perspective of a model manufacturer, told reporters, "The ecological barrier of Nvidia CUDA is still quite large. Currently, it seems that people prefer to use Nvidia, but domestic computing power should be used more and more."
In addition, among the many discussions surrounding price wars, some argue that a large portion of the current generic model API calls are due to the internal business needs of large factories.
He told Caixin News Agency reporters, "Selling APIs directly is not a good business model, and to some extent, it is also a very practical driving force for price reduction. I believe that these B-end companies will also explore new models."
Previously, Liu Weiguang, President of the Public Cloud Business Unit of Alibaba Cloud Intelligence Group, publicly stated that the cost of AI inference can only be reduced by ten or even a hundred times annually to promote the explosion of AI applications in various industries. Alibaba Cloud has significantly reduced the price of large model inference this time, hoping to accelerate the explosion of AI applications.
Integrated production and modeling is the core trend
2024 is seen as the first year of the explosion of big model applications in the industry, but it seems that the applications that ignite the industry have not yet emerged.
In Zhou Feng's view, how to view whether there is a particularly successful Super App is subjective.
He told Caixin News Agency reporters, "If you expect large models to have a very large scale and make a lot of money, from this perspective, the development of the industry is not that fast. From the perspective of large models helping many users find better solutions to problems, helping many companies promote business development, and improving the speed of domestic model capabilities, the overall progress of large models in China is still relatively smooth."
At the media exchange meeting held this week for Youdao Ziyue's education model, Zhou Feng also said, "In the past, we talked about product market fit for Internet products. This year, it may be necessary to mention model market fit, that is, the model needs to be well adapted to market needs."
He gave an example that GPT-4o has taken a big step forward in real-time voice interaction and visual interaction, and the adaptation of the model to the demand of the mobile application market has improved; Compared to Llama3, domestic open-source models have strengthened their Chinese and Asian language capabilities, which is an adaptation enhancement of the model to the needs of the Chinese market.
Zhou Feng stated that the integration of production and modeling is the core trend for the future development of large models. Grasping new scenarios, planning appropriate models, covering the granularity of requirements, and then synchronously improving models and products within the framework of "production model integration" is currently an important path for developing large-scale model business, and it is also a widely adopted method.
According to the first quarter report, since its release in July last year, NetEase Youdao has launched over 10 large model native applications and implemented them in software and hardware products. At the exchange meeting, three mobile apps based on Confucius, including Youdao XiaoP, Hi Echo3.0, and self-developed RAG engine QAnything, were also released.
Among them, Youdao XiaoP has integrated multi scenario interactive Q&A and interaction functions based on this upgrade, improving knowledge memory, multimodal understanding, and logical reasoning. It is planned to be launched in June, positioning itself as a new generation of AI comprehensive learning assistant; QAnything was officially open source earlier this year and has received nearly 10000 stars on GitHub. This time, it has been improved in terms of private deployment and intelligent agent generation.
AI subscription explosive growth
It is worth noting that the rapid development of generative AI is driving explosive growth in app subscription services.
According to the latest financial data from NetEase Youdao, the GMV of AI feature driven subscription services for Q1 this year was close to 50 million yuan, an increase of over 140% year-on-year. It is reported that its AI subscription services cover Youdao Dictionary, Youdao Translation, Hi Echo, and international apps that provide translation, language learning, and other services, with a year-on-year growth of over 50% for five consecutive quarters.
Zhou Feng stated that AI subscription will be one of NetEase Youdao's three major focus areas.
"Directly targeting users' products is currently the biggest growth opportunity for AI, and AI subscription services are a very natural way to develop AI. We adhere to this direction and are still in the initial stage. Next, we will accelerate the expansion of product scale and increase user subscriptions." He said.
At the same time, multiple manufacturers have also achieved growth in subscription membership revenue due to AI empowerment. For example, last year, Meitu's revenue from imaging and design products was 1.327 billion yuan, a year-on-year increase of 52.8%. The revenue growth was due to the application of AIGC technology. As of the end of last year, Meitu's paid subscription users exceeded 911 million, a year-on-year increase of 62.3%; Last Q4, Kingsoft Office launched WPS 365 for organizational level customers, driving rapid growth in B-end subscriptions. In 2023, domestic institutional subscription and service businesses achieved a revenue of 957 million yuan, a year-on-year increase of 38.36%; The Duolingo Learning App and Duolingo English Test have also introduced generative AI. Last Q4, the number of paid users reached 6.6 million, a year-on-year increase of 57%, and the subscription revenue from paid members was about 160 million US dollars.
Zhang Yi, the head of NetEase Youdao's Intelligent Application Business Unit, told Caixin News Agency reporters after a media exchange meeting, "AI subscription is a very intuitive manifestation of AI landing. Providing AI services for the C-end to earn member subscriptions, and at the same time, AI subscription itself is also raising barriers. Through AI methods, we continuously arm ourselves and make ordinary tools more powerful. AI subscription should be a very healthy model."
Zhang Yi further stated, "The real AI penetration rate of mobile users is very low, and there is a lot of room for improvement in the next six months or a year, which is an inevitable trend. This is also one of the reasons why we are optimistic about AI subscriptions."