The fierce "Hundred Model Battle" is accompanied by controversy over repetitive construction and difficulties in commercial monetization. As the initial craze faded, some investors also entered a cooling off period for the big model. AI native applications are becoming a new trend pursued by technology giants.
On May 13th local time, OpenAI released the ChatGPT intelligent assistant, supported by the new generation AI big model GPT-4o, with interactive capabilities comparable to real people. One day later, Google also launched its visual recognition and voice interaction product Project Astra, and introduced an AI Overview feature for its search tool. Users can enter a question in the search box to get an AI summary answer.
Star investor Zhu Xiaohu stated that there are many opportunities to provide value-added services on big models. "There is a lot of innovation in the application layer, and China far surpasses the United States in data and application scenarios."
Wang Sheng, a partner of Yingnuo Angel Fund, also agreed with this in an interview with the Daily Economic News. He said, "There will be great opportunities in applications. I think the biggest platform is ultimately a user driven application platform."
Before the big moves of OpenAI and Google, the Kimi Intelligent Assistant for Mystery AI Search and Moon Darkness had successively gone viral, showing the market the vast market space for AI applications. Some in the industry even believe that 2024 will become the first year of AI application innovation. Wang Sheng expressed skepticism about this, stating that there are still shortcomings in terms of large model capabilities, toolchain capabilities, and cost side capabilities, and that the explosion of AI applications is "not yet time".
Silicon Valley's "New Field": OpenAI and Google's Intelligent Assistant for "Volume"
As the "Hundred Model Battle" approaches today, more and more industry insiders believe that the competitive high ground for AI in the future is not in the big models themselves, but in the native applications of various industries. Based on powerful big models, tech giants in Silicon Valley are focusing on developing more user-friendly and powerful generative AI tools.
On May 13th local time, OpenAI held a spring press conference and released a new generation of AI model GPT-4o. The letter "o" in the number represents "omni" (omnipotent), which can accept any combination of text, audio, and image inputs and generate relevant responses for text, audio, and image.
In the live demonstration, ChatGPT, supported by GPT-4o, not only translated and recognized images and engaged in discussions, but also observed words and expressions, joked and pretended to blush. It also knew how to deal with interference, demonstrating a smooth dialogue level with almost no delay, and interactive abilities comparable to real people.
Wire reported that such adjustments may make ChatGPT more attractive. OpenAI CEO Altman bluntly stated, "This is as real as AI in movies. To be honest, I was a bit surprised. Achieving a reaction speed and expression ability that is close to that of humans is a significant breakthrough."
OpenAI plans to open GPT-4o's new audio and video features to some partners in the API in the coming weeks.
On the second day after the release of GPT-4o, Google's annual developer conference was also held as scheduled, and more than ten products were released one after another: in addition to the multimodal AI model Gemini 1.5 Pro, the cultural image model Imagen 3, and the cultural video model Veo, Google also launched the visual recognition and voice interaction product Project Astra.
As long as the phone camera is turned on and aimed at any object, Astra can accurately say the name of the object and answer questions such as "What can sound in the lens?".
From the demonstration, Astra's dialogue delay is slightly longer than GPT-4o, and there is a certain reaction time after being asked questions, but it also has multimodal understanding and real-time dialogue ability.
In addition, Google has also launched an AI Overview feature for its search tools, where users can enter a question in the search box to obtain an AI summary answer and handle lengthy questions.
In addition to OpenAI and Google, Silicon Valley entrepreneur Elon Musk and technology company co-founder Mustafa Suleiman have also invested in the development of chatbots Grok and Pi, respectively, with anthropomorphic features as their main product focus.
"There is a great opportunity in the application": Two "dark horses" have emerged in China
Fu Sheng, Chairman and CEO of Cheetah Mobile, believes that without considering the cost of locally adding parameters and improving the so-called large model capability, this path will definitely encounter difficulties. The release of the GPT-4o model application by OpenAI demonstrates that big models have great potential at the application level. The capabilities of big models will continue to iterate, but ultimately, it is still applications that can make good use of them.
Recently, the popularity of Mystery Tower AI search and the Dark Side of the Moon in China has fully demonstrated the vast potential of AI applications.
According to statistics from the AI Product Ranking (aicpb. com), the Kimi Intelligent Assistant on the Dark Side of the Moon had a total of 20.4 million visits in April this year, an increase of 60.2% compared to the previous month, with a month on month increase of 321.58%. After a month on month increase of 551.36% in March, the domestic AI application Mystery Tower AI Search continued its growth trend in April this year, with a month on month increase of 54.56%, reaching 10.86 million.
Prior to this, Baidu also released over 20 native AI applications in one go; ByteDance has established a new team, focusing on the application layer; Tencent has embedded large models into mini programs; Alibaba also needs to redo all applications with Tongyi Qianwen.
Wang Sheng, a partner of Yingnuo Angel Fund, said in an interview with the Daily Economic News, "There will be great opportunities in applications. I believe that the largest platform is ultimately a user demand driven application platform."
Star investor Zhu Xiaohu is also very optimistic about domestic application layer innovation. In his view, the biggest confidence of domestic innovation is "far surpassing the United States in data and application scenarios."
How to choose: To B or To C?
As a large amount of hot money and talent accelerate their concentration in the field of big model applications, the choices of enterprises also begin to diverge: some, such as Mita Technology, are also strengthening their coverage of B-end customers while targeting the C-end, while others, such as the dark side of the moon, choose To C.
In terms of user scale, Wang Yiwei, COO of Mita Technology, disclosed to the media last November that Mita reached B-end users through C-end registered users. At that time, there were thousands of B-end customers, including Internet giants and well-known media.
Stepstar, a domestic large model startup founded by former Microsoft Global Vice President Jiang Daxin, will mainly focus on the C-end.
"We choose not to do the traditional 'To B' and instead focus more on 'To C'," said Stepstar to a reporter from the Daily Economic News. "From a broader perspective, we are currently more focused on models. We definitely do products. Firstly, we need products to drive the development of models, and secondly, product data to feed back our models. As for what kind of products to produce? Firstly, Step Star does not want to create traditional customized models with a logic of private deployment. We still want to maintain a relatively elite and talent concentrated team."
The dark side of the moon once explained in an interview with the Daily Economic News that being firm in To C is because the model capability is still rapidly iterating, and making To B products can easily waste a lot of human and material resources. "The B-end products in the domestic market often require private deployment and customized development, and such deployment and development based on a certain version of the existing model will result in the B-end products' capabilities and user experience completely falling behind the rapid improvement of the model's capabilities, and will soon be eliminated."
Yuan Jinhui, founder of Silicon based Mobility, a startup focused on AI infrastructure, doesn't see it that way. He analyzed to the Daily Economic News reporter that currently, the certainty of doing To B services is higher, and many companies are using large models to provide vertical services for different industries. They have a strong demand for fine-tuning and reasoning of large models.
Yuan Jinhui believes that the uncertainty of the success of To C applications is higher, and at least no killer applications like Midjournal and Character.ai have been seen in China yet. "As the cost of deploying large model inference further decreases, there will be more and more attempts at To C applications, and there is a greater chance of super applications appearing. In comparison, 2C applications in overseas markets are progressing faster than those in China."
He emphasized that in the long run, there is broad market space in the field of large models, whether it is for To C or To B.
If 2023 is the first year of the AI big model, then many voices in the industry believe that 2024 will become the new year for AI applications.
Wang Sheng expressed a skeptical attitude towards this point in the interview. He said that given the shortcomings in large model capabilities, toolchain capabilities, and cost side, the explosion of AI applications is "not yet time". "Just like for the Chinese market, WeChat was the first important mobile application, but when it appeared, the iPhone had already been released for four years."
In addition to the immature conditions for creating applications, Wang Sheng also pointed out that new application scenarios have not yet been explored, especially for startups. "Now many applications are to apply AI to existing Internet business projects. In the past, big players in the old industry had a series of advantages such as customers, money, data, talents, and understanding of scenarios. It is difficult for start-ups to make up. For start-ups, the core is to innovate, not only to serve existing scenarios and users, but to find new needs and groups."