Original source: Brain polar body
![]()
Image source: Generated by Unbounded AI
We all know that large models must be born to be useful for applications. So, which application can give full play to the value potential of AI large models the fastest and become the first stop for the landing of large models? When it comes to this question, many people will first answer: search.
On the one hand, this is because, after the explosion of ChatGPT, the “major shareholder” Microsoft first integrated its capabilities into Bing search, and once announced that it would rely on the advantages of large models to pick out the big brother Google in the industry. Then, in the context of large-scale replication of large models in China’s AI circle, it is natural to think of promoting search + large models first on the application side.
On the other hand, search engines naturally have the advantage of being deeply integrated with AI. As early as 2014 and 2015, traditional search engine giants such as Baidu and Google began to integrate AI technologies such as deep learning and knowledge graph into search, so as to improve the search engine’s ability to understand user instructions and enhance the internal relevance of search results.
It can be seen that the large model + search can be said to be the right time and place. After nearly a year of exploration, the large model + search application of China’s AI industry has gradually been enriched. Although the changes brought by large models to search have not completely emerged, a relatively diverse exploration idea has been formed.
In order to let everyone more vividly understand the current progress of large model + search, and understand the differentiation of different ideas. We thought of a metaphor: big model + search, like a game of checkers. The chess pieces in the hands of all players are the same, that is, the large model technology and the search technology. And their ultimate goal is the same, that is, to incubate the first popular application in the era of large models.
But in the process of playing chess, each of them has different chess moves. At the moment, they are divided into three genres.
Contestant 1: Large model as an enhanced plugin for search engines
Search is the most frequent contact between people and information in the Internet era. Search engines need to understand both user intent and massive amounts of information. As a hub between information and people, the need for search engines to improve their intelligence is never-ending.
The difference that the large model brings to the search engine is that it can not only enhance the experience of the traditional search engine, but also bring various content generation capabilities to user intent and search results through the AIGC model.
For example, a large model can not only improve the accuracy of search, but also merge multiple search results into a single content box, saving users time. This is equivalent to giving users some additional search tools outside of the traditional search framework.
Based on this idea, the industry began to explore the first mode of large model + search: the ability of large model as an enhanced plug-in for search engines. In the domestic market, the representative of this genre is Baidu.
![]()
The search business can be said to be Baidu’s first stop for product transformation through Wenxin’s large model capabilities. At this stage, Baidu has added two “enhanced plug-ins” based on AIGC capabilities to the search engine.
The first is to aggregate information in the first answer.
In the process of combining AI technology with search, Baidu places great emphasis on the concept of “the first search result is to meet the needs of users”. The large model capability can aggregate key information from search results to generate content summaries. Based on this model, Baidu has updated the ability to answer the first answer in the search engine, which covers not only text information, but also understands the video through a large model to summarize the summary. In this mode, by adding the results that the user wants to search for in the video content, the user can no longer watch the video, but directly get a summary of the video content through the first answer.
According to data released by Baidu, the satisfaction rate of the first search in the past was only about 40%, but after adding the large model capability, the rate reached 70%. It can be seen that the large model capability, as a search engine plug-in, is to gain positive feedback.
Another kind of “enhanced plug-in” type of large model combined with search is to provide an AI dialogue bar in addition to the search bar, which is the “AI partner” capability launched by Baidu Search at the Mobile Ecological Conference in May this year.
AI partners can conduct AIGC Q&A with users, help users complete capabilities such as marking answers, providing information sources, summarizing document summaries, etc., when using search engines, and also support the invocation of other tools and services.
In other words, Baidu provides AIGC plug-ins based on large model capabilities in the search engine interface and outside the search engine, so that the search engine can obtain a large model reconstruction from multiple angles. Coincidentally, this line of thinking is very similar to Google’s integration of the Bard chatbot into its search engine
It can be seen that manufacturers with the traditional advantages of search engines are more inclined to use large models as enhancement plug-ins, and integrate the idea of “1+1 is greater than 2” into traditional search engines from multiple angles.
Contestant 2: ChatGPT-like search application
In addition to enhancing the capabilities of traditional search engines, the large model also brings another problem: is it possible to bypass the traditional search form and directly generate new search products based on AIGC capabilities?
There has also been some exploration of this possibility. ChatGPT itself has the ability to understand semantics, multi-round Q&A, content generation, etc., and to some extent, this can also be seen as a kind of “search”. It’s just that the content of the search has changed from keywords to questions and needs, and the search results have changed from web pages to directly generated text content.
As a result, a new type of search product that has emerged in China’s AI industry can be called ChatGPT-like search. Among them, the representative “contestants” are the Tiangong AI search launched by Kunlun Wanwei.
![]()
This kind of search engine completely takes AIGC as the core logic of the product. Users use natural language to express the intent of their needs, and then the search interface responds to relevant answers, instead of displaying a large number of web links like traditional search engines.
Relatively speaking, one of the innovations of Tiangong AI search lies in the source index. When using AIGC platforms such as ChatGPT, we often face the uncertainty of what AI will answer. The large model of many questions cannot give the correct answer, and even makes up arguments, literature sources, news sources, etc., which is widely complained about as “AI serious nonsense”.
Tiangong AI search emphasizes the generation of answers and the reference information source at the same time, so as to ensure that users can trace the reference of information, which greatly avoids the trust problem of the AIGC platform. And its reference information sources are also relatively rich, including news websites, knowledge question and answer platforms, videos, etc.
However, at this stage, the boundary between ChatGPT-like search and AIGC platform is still difficult to distinguish, and users’ perception of it is not clear. This model needs to be further popularized and tested by users.
Player 3: Large Model Landing Vertical Search
After the search engine landscape is relatively stable, there has been such an industry idea: after the general search opportunity is not large, the search engine can make efforts in the vertical search field, so as to consolidate the user base with continuous search demand in this field. Sogou Search and Quark have all made efforts in the field of vertical search. Among them, Quark has achieved good results among young users by virtue of its vertical search ability.
The third idea of large model + search is to take the lead in landing large models in vertical search. In this way, it strengthens the natural language understanding ability and information retrieval experience in specific search areas. In this field, the current representative player is Quark. On November 14, Alibaba’s intelligent information business group released a quark large model. Based on its own differentiated positioning, the application of quark large model will give priority to the application of professional search and other information services. In addition to the basic large language model, the quark large model will also derive vertical models such as medical care and education, which shows the importance that quark attaches to the field of specialized knowledge.
![]()
At present, medical care, education, and humanities and social sciences are the main directions for vertical search of large models. These directions have strong requirements for information sources, and have the characteristics of vague keywords, less effective information, and strong logic, which are more suitable for large models to exert their own characteristics than general searches. At the same time, the combination of large models and vertical search can also reduce product costs and improve the overall efficiency of large models in the search field.
In fact, there is also a variant of the combination of vertical search and large models, that is, each network disk is currently online with a search function with natural language understanding capabilities. You can use key information such as vague descriptions and adjectives to retrieve network disk data, especially for pictures, videos, and other content.
It’s all a springboard for super apps
So the question is, which mode is the correct answer to large model + search?
Sorry, the answer can only be waiting.
Large model + search is a logically very promising large model landing scenario. Therefore, after ChatGPT just began to explode, Microsoft built the large model capability into BingChat and released a lot of cruel words about Google search. But almost a year later, Microsoft has spun off a lot of its AI capabilities from the search business, and Google’s market dominance has not been affected. It can be seen that there is still a long way to go from theory to practice in this scenario.
![]()
Looking back at the domestic market, you will find that the three exploration modes are still fighting separately, and there is not much confrontation, and there is no general recognition of large model + search on the user side, and even its landing degree is far inferior to the ChatGPT-like dialogue application itself. Three reasons can be found:
**1. These three attempts of large model + search have not completed the breakthrough of the product form from 0 to 1. **Born to strengthen and complete the previous search engine and AI dialogue products, so it does not have a very interesting product flashpoint.
**2. At this stage, the improvement of the search experience of the large model is not strong in the experience of the mass user. **It can only be used as a professional tool in academic, IT and other fields.
**3.In addition, the commercialization space of large model + search is not clear. **After the addition of large-scale model technology, the business model and commercialization level of search products have not changed greatly, so they have received insufficient attention from the capital market.
In the long run, the ultimate goal of large model + search must be to form a super application in the era of large models. Just like the emergence of search engines in the Internet era, it has completely changed people’s information acquisition and interaction mode.
And if this is the goal, today’s large model + search exploration will inevitably be a springboard on the way of the chess pieces. Only by letting the chess pieces continue to jump can a qualitative change occur at a certain node in the future.
As long as you can keep the big model and the search moving forward, the light in the foreground is still far greater than the darkness.
