On September 4th, the mobile app of Wenxin Yiyan's large language model released version 4.0.0, with the biggest change being the renaming of the app from "Wenxin Yiyan" to "Wenxiaoyan." On the same day, the latest version of Alipay's AI application "Zhixiaobao" was officially launched on the Apple App Store. When viewed together, these two events reveal some subtle insights.

At this moment, it has been one year and nine months since the emergence of ChatGPT in December 2022. As large language models continue to evolve, more and more people are asking when the Killer App of this era will arrive.

Meanwhile, there are increasingly rational views in the market, and there are even voices critical of large language models. However, if we consider the basic laws of development, it is to be expected that extreme enthusiasm will give way to rationality. Therefore, there is no need to panic overly about these critical voices.

Through "Wenxiaoyan" and "Zhixiaobao," New Standpoint has captured an overall trend that is gradually emerging: the AI industry is abandoning its obsession with achieving a Killer App through the model-application integration approach.

As the name suggests, "model-application integration" refers to the integration of the model layer and application layer. Although this is not a technical term per se, nor a widely used concept in the industry, we will define it in this article as a commercialization strategy employed by large language model vendors, such as ChatGPT, Doubao in China, Tongyi, Wenxin Yiyan, and others. These models are both the names of the large language models and the applications, with both the model layer and application layer controlled by the same vendor. The approach is to design products based on existing models.

Essentially, this strategy reflects the desire of large language model vendors to create a blockbuster application through technology. This is seen as the ideal commercialization path from the perspective of these vendors and has been a major focus of their efforts over the past year, forming the core logic of their pitches to investors.

However, in this process, from the constant decline in model token prices to the rising cost of customer acquisition, the players remaining in the model-application integration game are either backed by large companies or sought-after unicorns. Nevertheless, the competition remains fierce and crowded.

It is not surprising to see many large language model vendors now focusing on To B solutions or AI applications for the C-end market that do not emphasize large language models as much. These directions are becoming more certain paths to commercialization in this stage of rationality. As the title suggests, vendors may be realizing that the model-application integration approach may not yield a Killer App after all.

01. GPT's Model-Application Integration is Difficult to Replicate

The source of vendors' long-held fantasies about model-application integration lies in OpenAI.

Prior to this, it was rare to see a case in the internet age where an algorithmic breakthrough directly led to a blockbuster application. In China's internet, especially in mobile internet development, the key to a blockbuster application's success often lay in its business logic and ecosystem accumulation.

ChatGPT's emergence presented a possibility: when technology accumulates to a certain level, it can bring stunning results and quickly attract global attention. At one point, a widely circulated comparison showed that it took Instagram two and a half years, TikTok nine months, and ChatGPT only two months to reach 100 million users. This speed demonstrated the potential for technology to create blockbuster applications.

Domestic players quickly entered the game. Throughout last year, entrants announced their general large language models and related applications, such as Baidu's Wenxin Yiyan, Alibaba's Tongyi Qianwen, Iflytek's Iflytek Spark, ByteDance's Doubao, and other unicorn large language models like Zhipu AI and Baichuan Intelligence.

However, even OpenAI is facing commercialization challenges, including huge funding needs, pressure for continuous technological breakthroughs, and high operating costs. According to The Information, OpenAI's annual operating loss is estimated at $5 billion, and the company needs at least $5 billion in new funding each year to survive.

While further funding may not be too difficult for OpenAI, given its public relations momentum, recent media reports indicate that OpenAI is in talks for a new round of funding valued at over $100 billion.

Additionally, OpenAI's legal battle with The New York Times shows that interconnectivity with mature content and tool ecosystems is not easy. The key pain point for AI users lies in accessing and controlling information, which would require OpenAI to build a self-sufficient ecosystem. However, this would be a long and arduous journey for a unicorn company.

OpenAI represents an extreme case in the AI industry. Moreover, internet usage habits differ significantly between domestic and international users, with foreigners preferring web-based searches and Chinese users favoring mobile devices. OpenAI's rapid user accumulation through large language model technology has limited relevance for the domestic market.

The reason why a domestic blockbuster application has yet to emerge is straightforward: trying to match an application from the technology end can be misguided.

Although the model-application integration approach has a relatively successful example in OpenAI, and it is indeed more convenient for those who possess model technology to debug the application, it does not address the challenges of achieving a mature and stable AI model capable of providing reliable performance across all application scenarios.

It is evident that most AI To C applications emerging in the post-GPT era offer only a single core function. The era of relying solely on a single function to stand out in the mobile internet is long gone.

02. Applications Seeking to Move Beyond Model-Application Integration

There are two main approaches in the AI sector: To B solutions and embedding AI into C-end applications or ecosystems from the business layer. These approaches are more reliable than burning money to create a super app through model-application integration.

In To B solutions, Alibaba's Tongyi and Huawei's Pangu are prominent examples. Tongyi's implementations in transportation, finance, hospitality, enterprise services, and communications, as well as Pangu's advantages in mining, power, meteorology, and pharmaceuticals, speak to the strengths of Alibaba and Huawei in To B customer service.

In To C applications, a notable example is Kimi, currently the most visited ChatBot in China. When asked about the large language model behind it, Kimi responds, "I am not an independent large language model but an AI assistant developed by Moonshot AI, named Kimi. My capabilities come from the advanced technologies integrated by Moonshot AI."

Clearly, Kimi emphasizes AI, with the large language model serving as one of the technical pillars supporting its business logic. As previously mentioned in New Standpoint, aiming for comprehensiveness and versatility from the outset can lead to a lack of focus in both customer acquisition and model training. While Kimi leverages a general large language model, it has a precise and efficient entry point: translating and understanding professional academic papers, assisting in legal analysis, and quickly comprehending AAPI development documentation. These areas are more logical and easier for AI to understand and output.

In other words, Kimi's Moonshot AI has identified its core paying customers and commercialization model early on, coupled with determined marketing investments and targeted customer acquisition channels, contributing to the snowball effect of content generation quality.

Kimi emphasizes its AI nature from the outset, and now more players are downplaying the presence of large language models and highlighting their AI attributes. Examples include ByteDance's Jimeng AI and Kuaishou's Keling AI in video generation, MiniMax's AI companion app Xingye, and, as mentioned at the beginning of this article, Baidu's renaming of its app from "Wenxin Yiyan" to "Wenxiaoyan."

However, what will be the core logic of the story that large language model vendors or AI companies tell investors in the next era, beyond model-application integration? We can glimpse the answer in a recently launched app: Alipay's AI application "Zhixiaobao."

Ant Financial claims that "Zhixiaobao," based on the Ant Bailing large language model, is the first standalone AI app in China to offer service functions. Given Alipay's comprehensive lifestyle service ecosystem, using AI to enhance these services offers an imaginative and concrete application direction.

For instance, users may soon be able to complete online purchases with a single voice command. This seemingly simple voice purchase involves not only AI's ability to understand and filter information but also to make crucial decisions on behalf of the user – payments, which rely on Alipay's groundwork in functions like password-free payments and its overall payment ecosystem.

In other words, rather than creating a new super app, AI technology can be integrated into existing super ecosystems to enhance their functionality. This is one of the core narratives of AI's future in ToC applications.

When GPT first emerged, the phrase "Every app deserves to be rebuilt with AI" became popular. Now, it's time to add a postscript: "Rebuild the apps, not the ecosystems."

Zhixiaobao is Alipay's AI version, potentially seamlessly integrating with all Alipay ecosystems and functions. The key lies in the interconnectivity within the ecosystem and its overall maturity.

Of course, this does not mean that the model-application integration approach is entirely wrong. In the rapidly evolving AI landscape, developing a large language model first and then designing To C products no longer accurately represents the core value that AI brings.

In contrast, new, more practical and imaginative approaches have emerged, such as embedding AI into established super apps or ecosystems, matching AI technology from the application end. This is true for both Zhixiaobao and Doubao AI's integration into Feishu and other applications.

03. Conclusion

A super app may eventually emerge in China's model-application integration race, but it will take a long time to build the necessary ecosystem, as users are not loyal to single functions.

According to data circulating among social media users, the 30-day retention rate for China's top AI applications is less than 1%.

This data is telling. Most domestic To C AI tools are generative, with few executive applications like Zhixiaobao. The demand for single AI-generated content functions to assist daily life remains limited.

The first super AI application is likely to emerge from a mature and widely recognized ecosystem.

*Images in the headline and article are sourced from the internet.