"100 Model Wars" Annual Exam|Competing for AI Applications: OpenAI, Google's "Competition" Intelligent Assistant, Kimi, and Secret Tower "Out of the Circle" One After Another

National Business Daily

2024-05-17 16:27Published on the official account of Sichuan Daily Economic News

Every reporter: Yang Xinyi, Zhao Wenqi, every editor: Lan Suying

The menacing "100 model war" is a controversy over repetitive construction and a difficult problem for commercial realization. As the initial craze faded, some investors also entered a cooling-off period for large models. AI-native applications are becoming a new "outlet" sought after by tech giants.

On May 13, local time, OpenAI released the ChatGPT intelligent assistant supported by the new generation of AI model GPT-4o, which has an interactive ability comparable to that of a real person. A day later, Google also launched Project Astra, a visual recognition and voice interaction product, and launched an AI overview feature for its search tool, where users type a question into the search box and get an AI-summarized answer.

Zhu Xiaohu, a star investor, said that there are many opportunities to do value-added services on large models, "There are many innovations in the application layer, and China is far more than the United States in data and application scenarios." ”

Wang Sheng, partner of Innova Angel Fund, also agreed with this in an interview with the reporter of "Daily Economic News". "There's a big opportunity in the app," he said. I think the biggest platform is ultimately a user demand-driven application platform. ”

Before the big move of OpenAI and Google, the secret tower AI search and the Kimi intelligent assistant on the dark side of the moon have been "out of the circle" one after another, allowing the market to see the broad market space for AI applications. There is even a view in the industry that 2024 will be the first year of AI application innovation. In this regard, Wang Sheng expressed skepticism, saying that there are still deficiencies in large model capabilities, tool chain capabilities, cost side and other aspects, and the outbreak of AI applications "has not yet arrived".

Silicon Valley's "new arena": OpenAI, Google's "volume" intelligent assistant

At present, more and more people in the industry believe that the future competition of AI is not in the large model itself, but in the native application of thousands of industries. Silicon Valley's tech giants are working on more human-friendly and powerful generative AI tools based on powerful models.

On May 13, local time, OpenAI held a spring conference and released a new generation of AI large model GPT-4o, the letter "o" in the number stands for "omni" (omni), which can not only accept any combination of text, audio and image input, but also generate text, audio and image related responses.

In the live demo, ChatGPT, powered by GPT-4o, is not only able to translate and recognize images and start discussions, but also observes words, jokes, and pretends to blush, but also knows how to deal with distractions, demonstrating a level of smooth dialogue with basically no delays and interactivity comparable to that of a real person.

OpenAI executives communicate with ChatGPT at the press conference Image source: Screenshot of the video meeting of the press conference 

Wire reports that such adjustments could make ChatGPT more appealing. OpenAI CEO Altman bluntly said: "It's as real as AI in the movie, and frankly, I'm still a little surprised." Achieving near-human reaction speed and expressive ability is a major breakthrough. ”

OpenAI plans to open up GPT-4o's new audio and video capabilities to select partners in the API in the coming weeks.

On the second day of GPT-4o's release, Google's annual developer conference was also held as scheduled, and more than a dozen products were released one after another: in addition to the multimodal AI model Gemini 1.5 Pro, the Wensheng graph model Imagen 3 and the Wensheng video model Veo, Google also launched Project Astra, a visual recognition and voice interaction product.

By turning on your phone's camera and pointing it at any object, Astra can say exactly what it is looking for, and it can also answer questions like "What's in the camera that can make a sound?"

Image source: Screenshot of the video of the press conference 

In terms of presentation, Astra's dialogue latency is longer than GPT-4o's, and there is a certain reaction time after being asked, but it also has multimodal understanding and real-time conversation capabilities.

In addition, Google has introduced an AI Overview feature for its search tool, which allows users to type a question into the search box and get an AI-summarized answer, which can handle very long questions.

In addition to OpenAI and Google, Silicon Valley entrepreneur Elon Musk and Mustafa Suleiman, one of the founders of the technology company Deep Thinking, have also invested in the development of chatbots Grok and Pi respectively, taking anthropomorphic features as the main direction of their products.

"There is a big opportunity in the application": two "dark horses" have run out of China

Fu Sheng, chairman and CEO of Cheetah Mobile, believes that if the parameters are superimposed regardless of the cost and the so-called large model capability is improved, this road will definitely encounter difficulties. OpenAI's release of the GPT-4o model application shows that the large model has a lot to do at the application level, and the capabilities of the large model will continue to iterate, but in the end, the large model can be used well, or the application.

Recently, the fire of Secret Tower AI Search and Dark Side of the Moon in China has also fully demonstrated the broad space for AI applications. 

According to the statistics of the "AI Product List (aicpb.com)", the Kimi intelligent assistant on the dark side of the moon reached 20.04 million visits in April this year, an increase of 60.2% month-on-month, and a month-on-month increase of 321.58% in the previous month. Following a month-on-month increase of 551.36% in March, the number of visits to the domestic AI application Mita AI Search continued to grow in April this year, up 54.56% month-on-month to 10.86 million.

Image source: AI product list (aicpb.com) 

Prior to this, Baidu also released more than 20 AI native applications in one go; ByteDance has set up a new team to focus on the application layer; Tencent embeds the large model into the mini program; Ali also wants to use Tongyi Qianwen to redo all the applications.

Wang Sheng, partner of Innova Angel Fund, said in an interview with the reporter of "Daily Economic News", "There will be great opportunities in the application. I think the biggest platform is ultimately a user demand-driven application platform. ”

Zhu Xiaohu, a star investor, is also very optimistic about domestic application-layer innovation, and in his opinion, the biggest confidence in domestic innovation is that "it is far more than the United States in terms of data and application scenarios." ”

如何选择：To B还是To C？

When a large amount of hot money and talents are concentrated in the field of large-scale model applications, the choices of enterprises have also begun to diverge: some of them, such as Mita Technology, are not only facing the C-end, but also strengthening the coverage of B-end customers, and the other part of the dark side of the moon is choosing To C.

In terms of user scale, Wang Yiwei, COO of Mita Technology, revealed to the media in November last year that Mita reached B-end users through C-end registered users, and at that time there were thousands of B-end customers, including Internet manufacturers and well-known media.

Founded by Jiang Daxin, former global vice president of Microsoft, the domestic large-scale model start-up Leap Star mainly focuses on the C-end.

"We choose not to do To B in the traditional sense, and will focus more on To C." Step Leap Star told the reporter of "Daily Economic News". "In the big picture, we're focusing more on models at the moment. We definitely do products, one is that we need to have products to drive the development of the model, and the second is the data of the product to feed back our model. As for what kind of products to make? First of all, Step Star does not want to do the logic of traditional customized model and privatized deployment, we still want to maintain a relatively elite and talent-focused team. ”

The dark side of the moon once explained in an interview with the reporter of "Daily Economic News" that the reason for firming To C is that the model capability is still in rapid iteration, and it is easy to waste a lot of manpower and material resources to make To B products. "B-end products in the domestic market often need to be privatized and customized, and such deployment and development based on a certain version of the existing model will lead to the ability and user experience of B-end products to completely fail to keep up with the rapid improvement of model capabilities, and will soon be eliminated."

Yuan Jinhui, founder of Silicon-based Flow, a startup focused on the AI infrastructure layer, doesn't see it that way. He said to the "Daily Economic News" reporter that at present, the certainty of doing To B services is higher, and many companies are using large models to do vertical services in different industries, and 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, at least in China, we have not seen killer applications like Midjourney and Character.ai. "As the cost of large model inference deployment is further reduced, there will be more and more attempts at To C applications, and there will be more opportunities for super applications to appear. Comparatively, the application of 2C in overseas markets is progressing faster than in China. ”

He emphasized that in the long run, in the field of large models, whether it is to do To C or To B, there is a broad market space.

If 2023 is the first year of AI models, then there are many voices in the industry that believe that 2024 will be the year of AI application innovation.

Wang Sheng expressed skepticism about this in the interview. He said that in view of the lack of large model capabilities, tool chain capabilities, cost and other aspects, the outbreak of AI applications "has not yet arrived". "It's like for the Chinese market, WeChat was the first big mobile app, but when it came out, the iPhone had been released for four years."

In addition to the immaturity of the conditions for creating applications, Wang Sheng also pointed out that new application scenarios have not yet been explored, especially for startups. "Nowadays, many applications are to apply AI to existing Internet business projects in the past, and the big players in the old industry in the past have a series of advantages such as customers, money, data, talents, and understanding of scenarios, which are difficult for startups to make up. For startups, the core is to innovate, not only to serve existing scenarios and existing users, but also to find new needs and people. ”

National Business Daily

View original image 1.57M

• 

• 

•