Similar to human thought processes, breaking down tricky problems into smaller tasks, and being able to self-correct errors—In September 2024, OpenAI once again demonstrated its leading position among global AI companies with its brand-new reasoning large model, o1. Just about a month later, a similar description also appeared in the introduction of new features of a domestic AI unicorn, but focused on the more segmented field of AI search.

According to a previous announcement by AI startup Moonshot AI, the latest version of the intelligent assistant Kimi Explorer Edition was opened to all users before this Monday. This version has AI autonomous search capabilities, can simulate human reasoning thinking processes, and perform deep searches. On the same day as this official announcement, another Chinese large model unicorn company, ZhiPu AI, also launched an AI search feature, also focusing on deep reasoning, emphasizing that the search process is progressive like humans.

Both leaders in the large model industry have chosen to take a step further in the search track, introducing thought chains to improve AI search. In the face of unclear commercial prospects, can deep search bring more users to them?

AI Search + Deep Reasoning, Domestic Version of o1?

As both domestic and foreign large model industries are actively exploring commercial applications, AI search has become one of the highly watched landing tracks. NVIDIA CEO Huang Renxun has long stated that he uses AI search engines like Perplexity almost every day. In July this year, industry leader OpenAI also began testing AI search tools SearchGPT, causing Google's stock price to fall before its official launch. Domestically, search giant Baidu claims that 18% of search results are generated by AI.

Advertisement

With AI search's direct summarization of web content, users no longer need to click on blue links one by one, but sometimes there may be situations where the response is verbose or there are omissions. The upgrade of Kimi Explorer Edition lies in its ability to first identify the user's true questioning intent, convert complex questions into multiple keywords. It also has the ability to reflect and can supplement search results in real-time.

Moonshot AI introduced that the search volume of Kimi Explorer Edition is 10 times that of the regular version, and a single search can read over 500 pages. In its official external statement, the product manager confidently stated, "If Kimi can't find the information, it's likely that users will also find it difficult to find through traditional search engines."

Taking the question "Among the Nobel Prize winners of all previous years, who are from the same country as the youngest winner of this year's Nobel Prize in Chemistry?" as an example, Kimi Explorer Edition first breaks down the question into two sub-questions, searching step by step for the identity of this year's winner and the nationality information of previous winners. After reading 60 web pages, it finally comes up with a list of 31 people. However, after reflection, it supplements that the information provided is still not perfect and accurate enough, mistakenly listing another scholar as this year's Nobel Prize winner.

When the same question is entered in Kimi Regular Edition, the keywords are extracted more roughly, and the number of web pages relied upon is also fewer. The search results obtained are relatively simple, not only failing to list winners other than the Chemistry Prize, but also incorrectly including scientists of other nationalities in the list of 16 people.

If Kimi Explorer Edition is compared with ZhiPu AI search, when asking "When should I start paying attention to which discounts for the 2024 Double 11?" both products list the event times and discount information of e-commerce giants such as Tmall and JD.com. Although Kimi reads more web pages, ZhiPu AI also mentions other platforms such as Douyin and Kuaishou, and also supports drawing a timeline directly in the search window.Zhipu AI introduces that it supports up to 100 web page queries at a time. In addition to making AI search smarter and more logical through multiple logical deductions and cross-validations, it can also call upon various tools such as drawing and coding, behaving more like a human.

Although the dark side of the moon and Zhipu AI's new features both demonstrate deep reasoning and thought chain technology, similar to the o1 model, these two products focus more on enhancing search capabilities, providing them as search tools to users. The dark side of the moon targets knowledge workers and college students, the original core users of Kimi, hoping that the Kimi Exploration Edition can meet their needs in search and research. As a leading model with strong reasoning capabilities, OpenAI specifically points out that o1 performs well in the fields of science, coding, and mathematics, suitable for groups such as medical researchers, physicists, and developers.

An AI search developer told the Blue Whale News reporter that AI deep search is an optimization of previous capabilities at the engineering level, but the core capabilities of the model have not been significantly improved. Its underlying logic is to break down complex problems into multiple steps, generate multiple answers accordingly, and then obtain a better answer through the selection and summary of algorithms.

Yang Zhilin, the founder and CEO of the dark side of the moon, once stated that the o1 reasoning model brings a new paradigm. In an event in September, Yang Zhilin explained that o1 is learning to generate the thinking process in the human brain, which can achieve better generalization. This way, even if a different question is encountered next time, the model also knows how to answer it. Through reinforcement learning, more data that do not exist naturally can be generated to support further scaling.

Commercialization is difficult but must be fought for

Although there are already some leading AI search companies and products, the path to commercialization has always been a difficult problem for those who enter the field. Faced with profit pressure, Perplexity, which once focused on the "unaffected by the advertising-driven model," decided to start advertising from the fourth quarter. The company is negotiating with some brands to display advertisements in the related questions below the answers and on the right side. The CEO of Perplexity even said in an interview earlier that the greatest business model in the past 50 years is click advertising, and Perplexity's future main profit engine will also be advertising.

For traditional search giants like Google and Baidu, developing AI search directly affects their core advertising business. The business model of traditional search engines is to charge enterprises through advertising while being free for users, but AI search allows users to skip web page links and directly obtain answers.

At the earnings call in August this year, Li Yanhong, the co-founder and CEO of Baidu, admitted that the advertising business is facing pressure from external factors and actively promoting AI-driven search transformation. Transforming Baidu search with AI may have a negative impact on revenue, and it will take some time to see significant revenue contributions.

Enhancing the reasoning capabilities of AI search may also mean increased costs. The dark side of the moon said to the media that the Kimi Exploration Edition does have higher costs, and future commercialization attempts will be considered. However, both the AI search on the Kimi Exploration Edition and Zhipu Qingyan are currently free for users, with the Kimi Exploration Edition limiting each account to 5 uses per day.

The Exploration Edition is not Kimi's first move in the search field, it's just that previous products were lighter. In July, the dark side of the moon launched a Kimi browser plugin that supports users in querying terms with questions while browsing web pages. However, the generated search content is not marked with reference sources, and it cannot provide some real-time latest news.In the first half of this year, Kimi had seized the massive traffic of C-end AI products ahead of schedule, but now it is facing fierce attacks from competitors. Data from the AI product rankings in September show that DouBao App, a subsidiary of ByteDance, has a stable MAU of 47.02 million in domestic AI applications, far ahead of others. Baidu's Wen Xiaoyan ranks second with an MAU of 12.34 million, while Kimi ranks third with 7.38 million. Other AI products also show considerable growth in visits, including the unicorn MiniMax's Conch AI and Shengshu Technology's video generation tool Vidu.

An AI search developer told Blue Whale News, "What are the scenarios where users use high frequency every day? In fact, it is search, which has become a must-fight place for AI companies." In his view, in order to attract C-end users and compete for the huge traffic entrance of search, cost will not become the main factor limiting the top AI companies.

The developer added that the domestic search ecosystem is fragmented, and giants tend to retain information and content on their own "islands". For AI search, the gap between companies in engineering training and model capabilities will become smaller and smaller, and exclusive, high-quality information corpus will become the key barrier affecting search results.