作者|James
Last week, ChatGPT was testing an AI search function separate from its large model dialog, which will use a dedicated URL search.chatgpt.com, which is said to support image search, and may also include gadgets such as weather and calculators.
As an AI leader, almost every action of OpenAI can spark heated discussions in the industry. However, there was no search-related information released at the press conference in the early morning of Tuesday Beijing time, and the open GPT-4o model mainly enhances the multimodal capabilities of speech and image.
There is a variety of evidence that ChatGPT searches are indeed being done, and some people claim to have swiped the grayscale test qualification. At the same time, on the front of AI search, there is Perplexity abroad, and Tiangong and Secret Tower in China; "Traditional" search engines such as Google, Bing, Baidu, and 360 have also entered the market.
Regardless of whether the company is developing a basic model, it is not difficult to launch an AI search interface, and even only requires an individual developer to be busy for 3 days at the shortest.
However, Wang Yiwei, COO of Secret Tower Technology, said to the future that it is not so easy to enter the search process. The technical and product capabilities of AI search are equally important, and it is also necessary to think clearly about market positioning, commercialization, and covering the cost of computing power.
"I have a theory that a reasonable price for a tool is about 1/10 of the value of the work it replaces." Wang Yiwei said, "The current product of Secret Tower AI search is about 80 points. If you want to reach a score of 90 or higher, it means that the user has a strong willingness to pay. ”
Multimodality has been "stealing the spotlight" for a long time, and pure language models seem to be a bit lonely, and AI search seems to have rekindled enthusiasm in this regard. Will it be the next AIGC hotspot after video and music creation? Is it going to enter the fierce "Thousand Search War" soon?
When the Secret Tower AI search was first launched, Shizhi Future had an in-depth conversation with Wang Yiwei in the office of Secret Tower Technology. The full version of this interview program has been released on the podcast "Dialogue with AI Entrepreneurs" of Shizhi Future, please press and hold or scan the QR code below to listen:
VOL
AI search, why is it so like a big model?
Before ChatGPT's search function was officially unveiled, there were rumors and speculation in the outside world. X (Twitter) user @btibor91 drew a hypothetical image of ChatGPT's search interface based on some of the leaked front-end code:
However, it may be difficult to see any difference between the new search interface and the previous traditional GPT chat box.
At the same time, the self-media "Cyber Zen Heart" swiped the ChatGPT search in the grayscale test, and the following is the result page of the search he tried:
This is the networking function that ChatGPT Plus has long had in the future, the prompt words entered and the results obtained:
- It can't be said to be very similar, but it is exactly the same!
If you feel that AI search is not fundamentally different from ChatGPT, which can connect to the Internet, you are right.
At present, the core principle behind the combination of AI large models and search is RAG (Retrieval-Augmented Generation).
In 2020, Patrick Lewis of startup Cohere invented the term in a paper in which he apologized for the abbreviation "not very flattering," "If we knew our work was going to be so famous, we would have thought of a better name." ”
To put it simply, RAG converts the input prompt words into search keywords, reads and understands the content of the page searched on the Internet, and then defines the large model to generate answers based on these contents, rather than relying on the model's own knowledge base.
Except for a few large model chat products on the market, most models will provide the online search function for free, such as Wenxin Yiyan does not even support canceling the online function.
This is actually a form of "AI search". At this time, there will be a certain difference between the results generated by calling the API of the large model and the results of the conversation on the official website of the large model.
In addition to the above comparison of ChatGPT and search results, Shizhi Future also compared the output of Wenxin Yiyan and the AI search results of Baidu's "Simple Search" app, as well as the output of Google Gemini and the conclusion of Google Search's enhancement.
As a result, the network results of large models and AI search results from the same manufacturer can largely be substituted for each other, although the results generated are not exactly the same each time.
Wang Yiwei thinks that RAG is the best solution to minimize hallucinations. He believes that letting a large model answer questions based on its own trained knowledge base is similar to a closed-book exam, while using a search engine is equivalent to an open-book exam, which will greatly improve the accuracy.
Secret Tower's flagship product, WriterCat, has also developed a search-based "fact-checking" function, which singles out possible factual errors in a generated or human-written article and provides links to relevant information online.
Of course, because it can't be 100% dependent on AI's judgment, this function adjusts the alarm threshold to a relatively low level, and sometimes there are false positives, but "it's better to kill mistakes than let go", mainly to facilitate the author's manual secondary check.
However, there are also many doubts about RAG. For example, while RAG is able to provide more useful reference material, the logic of generating content is still a black box and therefore not very controllable. Sometimes, even if the material on which it must be based is specified, the model may "capriciously" give an answer that is not relevant to the material.
In addition, especially in search tasks involving inference ability, the limitations of the model itself may not be addressed by introducing high-quality material alone. For example, when asked to provide specific passages that can be excerpted directly from an online article, the model may be able to handle this; However, if you need to perform structured operations such as filtering, averaging, or extreme values of the searched data for the past years, the probability of errors is still high.
Since the search enhancement is used, where do the search results come from?
Some teams will do it like they make a large model by themselves, and even search for crawlers by themselves. For example, in a previous interview with Fang Han, chairman and CEO of Kunlun Wanwei, he explained that Kunlun Wanwei's team has 6-7 years of experience in the search field since the overseas product Opera News. Tiangong AI Search's crawling frequency of key websites has been increased to once per minute.
Fang Han also mentioned that Kunlun Wanwei has accumulated rich capabilities in pre-training data collection, cleaning, and deep processing. They are also looking at how to ensure the authenticity of information, such as scoring various source websites; and how to adapt the content to domestic users, how to avoid information cocoons, etc.
But it might be easier to use an off-the-shelf search engine. ChatGPT doesn't need to ask, it uses Bing search. According to reports such as The Information, Perplexity, which wants to be a "Google killer", actually uses automated systems to access data from Bing and Google. It uses Bing's API to rank the signals in the results to determine the relevance, quality, and authority of a web page.
Perplexity interfaces allow you to specify a limited source search, and the easiest way to do this is to use a site syntax; It's the same with prompts.
Wang Yiwei of Secret Tower Technology told Shizhi Future that academic literature can be cited as the source description of the content of the article in Writing Cat, which helps AI help students automatically generate a manuscript of a paper. However, up to now, academic search is still to search for some public resources, such as the web version of CNKI, etc. The system can only crawl information that can be accessed on the public network, such as titles and abstracts, and cannot access paid content such as the body text for the time being.
Secret Tower AI Search also has a special academic search section, and its search scope is the same as that of Writing Cat's academic literature search. Therefore, if someone really wants to "mass-produce" a paper, they must remember to do the fact-checking work themselves.
In fact, the problem of lack of content in AI products, whether it is when training a large model or when using a large model to search, it cannot be bypassed.
At home and abroad, there are many high-quality information sources that are not open to external search, such as shopping websites, social networks such as Douyin and Xiaohongshu, all of which prohibit search engine crawlers. There are also community legal departments that will take action after detecting that you are using their information.
Therefore, these problems can neither be solved at the level of search engines, nor can they be solved by training large models, so combining "search engine + large model", it is certainly not expected that it can be solved. Outside the "walled garden", it is also difficult for a clever woman to cook without rice.
Compared with the AI search of the whole network, it can be said that it is more cost-effective to replace the paid on-site search of news products with a large model-driven. Not only can this improve search quality, but it can also be another attraction of paywalls, increasing reader engagement and willingness to pay.
Over the past few months, Forbes and the Financial Times have launched their own conversational article database search boxes. Publishers such as Snopes, The Guardian, and Business Insider are also considering using generative AI to improve their on-site search capabilities. A media outlet said, "Site search is not a widely used feature, which gives us an opportunity to test AI in a low-risk environment." ”
VOL.2
AI search, what to grab the user
In addition to the large model dialog box you are familiar with, in addition to the ability to connect to the Internet, there are other different AI search interface forms. In essence, it can be said that it is the same principle, with different skins.
Existing search engine giants choose to insert a short AI-generated summary at the very top of the search results, as shown in the example of Google Search shown above.
Long before Wenxin Yiyan went live, Baidu had already launched a capability called "AI Search Intelligence Enhancement", which is now displayed on more and more search results pages for different keywords.
Baidu's main focus is enough for one, but it is not completely pursued to use a large model to generate. It has never integrated Wenxin Yiyan with the search results page, but has stayed at peripheral attempts such as "simple search".
Another way to show this is Copilot, which is similar to Microsoft's Bing integration. When the user is not actively invoking Copilot, the search results for the entered keyword will be displayed on the left side of the page, and the content generated by Copilot will be displayed on the right. 360 search is presented in the same way as Bing.
In contrast, AI searches that are not done by "traditional" search engines use a redesigned, dedicated three-bar AI search interface. The pioneer of this design is Perplexity, the "popular fried chicken" in the field of AI, a name that means "puzzled" in English.
Perplexity's founder, Aravind Srinivas, was so sensitive to the new search interface his company had pioneered that he even sparked a small debate when he posted on X (Twitter) that Meta AI's homepage design mimicked them.
In China, the use of this interface includes Kunlun Wanwei's Tiangong and Secret Tower.
The iOS version of the popular Arc browser uses a cleaner and clearer "variant" interface for AI search, hiding some of the options that can be distracting from other competitors.
Because of the help of the open source community, the "development cost" of doing an AI search has now been very low, and there is even an open source interface scheme that can choose to switch between search API and large model API at the same time, which is the "double shell" of search engine + large model.
The shell, called ThinkAny, is said to have been developed by an independent developer in 3 days. Select the search scope, generate a mind map, and switch the large model engine.
In this way, AI search actually puts a new layer of skin on the large model, finds a new use scenario, and also solves people's "aesthetic fatigue" of ChatGPT-style chat interface.
If this really works, it is like a "hundred model war" in the middle of last year, and in the next month or two, it is very likely that there will also be a "thousand search war".
When ChatGPT first caused a stir, many people saw the dialog box as a new form of search, a new "entry point to the internet". At that time, Shizhi Future also released a "ChatGPT Content Industry Practical White Paper" on how to generate effective prompt words, which still have reference value today.
In fact, in the era of "traditional" search engines, some people have long wondered why when searching for the same content, some people can quickly find the answer, while others find nothing? This may involve what is known as "search engine emotional intelligence".
When people's input expanded from a few keywords to more complete prompts, the problem was not only not alleviated, but worsened. People have had to work on all sorts of "spells" or complicated forms of input, and there are even dedicated "prompt word engineer" positions. The ease of use of large-scale model chat products is actually declining for ordinary people.
According to the report, Google is working on a new tool that allows users to go through a prompt word base optimization when querying, and then send the optimized prompt word to the large model in order to get a better answer, but the effect remains to be seen.
People are often reluctant to migrate from an old product to a new one. Therefore, users who are accustomed to traditional search can at least seamlessly experience the new capabilities of AI in the original service. Instead, entirely new services may leverage traditional search results channels to attract users.
Shizhi Future noticed that ByteDance's large-scale model chat app "Doubao" has attracted many "mysterious" new users through a large number of promotions in traditional search engines. Some users silently use bean bags, but they don't know that other similar domestic products, such as Wenxin Yiyan, Tongyi Qianwen, Zhipu Qingyan or Kimi, have never heard of it. This is very strange.
In fact, Doubao's strategy is to first use search engine promotion (SEO) tactics to attract users with more precise keywords, and then post these keywords to the search results page through Doubao's answers. When promoting Douyin, Byte has become proficient in using this method.
It is foreseeable that if someone somehow completely transfers their search requests from traditional search engines to Doubao and forms a new habit, this means that even if it is a search request that does not require a large model, they will use AI to search, because they are more reluctant to mix and match multiple different tools.
VOL
AI search, how to not lose money
When everyone flocks to the AI search track, will the big companies that "have more than enough to learn" win, or the small but more focused start-up teams will win? There must be different answers from different perspectives, and Wang Yiwei, COO of the Secret Tower, said his own views from the perspective of "small factory".
How to reflect the differentiated competitiveness when other homes are doing searches? Wang Yiwei said, "No ads, direct results" is the user mind that Mita AI Search hopes to repeatedly mention and establish.
Because young people are the largest user group of WriterCat products, when the product was first designed, Mita decided to abandon the direct advertising form. "There's no advertising, it's mainly about what kind of people you want to attract. The first promotional video we released was released on Station B, which also had no ads. ”
Wang Yiwei is also cautious about Baidu and Microsoft mentioning that they will study the paid advertising space in the results of "soft placement", saying that the bottom line is that the promotion results cannot affect the accuracy and fairness of the answers.
"I think at first they might choose to be simple and brutal, like Google does to show ads directly on the page. But over time, it can become more subtle, such as providing links to services. If you search for 'what to do if there is a water leak at home', the contact information of the relevant service provider may appear directly below. But you need to be thoughtful, you need to make sure that the recommended service has a certain level of reliability and endorsement, if the recommended service is not reliable, it will directly affect the reputation of the search engine. The stakes are high. ”
The purpose of Myta AI Search is to directly give the results that the user needs, rather than letting the user sift through a large amount of information. Wang Yiwei mentioned that "direct results" is their goal. Its "research" mode can easily generate 3,000 words at a time, which is really impressive when a lot of text is pouring out.
Although these word counts cannot be used directly without changing them to a large extent, it gives you 3,000 words at a time, and it is not a bad solution if you shrink it down to five or six hundred words yourself.
In this regard, the difference in the performance of different products is mainly reflected in the difference in the base model. Wang Yiwei said that because their model has a deeper accumulation of corpus, the parameters do not have to be blindly large, and 10 billion (10B) is enough for the task optimization of writing articles.
Secret Tower AI Search can also provide more in-depth services, such as converting into mind maps and "one-click PPT generation" for business reporting. Of course, there are other products that offer similar functions such as "one-click generation of infographics" for results.
The abandonment of advertising space also means that the plan for the commercialization of C-end by Myta AI Search will still be based on membership subscriptions.
"Internally, we release a product when it meets a certain quality standard. Often, if someone is willing to pay for it, it means that the product is at what we consider to be an '80' and is genuinely useful, not just a toy. ”
Wang Yiwei said, "The current product of Secret Tower AI Search is about 80 points. If you want to reach 90 points or higher, and users have a strong willingness to pay, we do not rule out charging in the future, or only charging for the 'research' model. But we're not sure yet. ”
How much is the right amount to charge? Wenxin's premium version costs 40 yuan a month, and 360 is also charging. "At present, they have not charged for Tiangong, but they have the ability to collect them."
"I think that's a central question, do you dare to charge? If you don't dare to charge, you can't verify the value of your product. No matter how much you exaggerate, the most important thing is whether someone is willing to pay for it. If there is, it's remarkable. If not, that's it. ”
Writing Cats is still the main business of the Tower, accounting for about 80% of its revenue. WritingCat now has about 1.2 million registered users, with only about 3-4% of paying users. The main customers are still civil servants and individual users of schools.
In addition, legal translation services are still providing a steady stream of cash flow. "MetaLaw, which has just started selling this year, is also doing well, and I think it will be no problem to sell a few million (yuan in sales) this year."
This kind of income structure makes Wang Yiwei a little worried, because the general model has done a good job in rewriting and error correction.
"The API services we provide, such as typo checking, rephrasing, etc., are more expensive than the services of large models. It's not because they can reduce the cost, but because the price of GPT is very low, in fact, they are all at a loss. And people who use our API can also use GPT. So as long as the income for this part of the simple task can be stable, it is good. ”
During the dialogue, Wang Yiwei also shared another experience with Shizhi Future.
"I have a theory that the reasonable price of a tool [in China] is about 1/10 of the value of the work it replaces. For example, MetaLaw's product pricing is 499 RMB/year. This price is acceptable for many lawyers. Otherwise, they may think it's a bit useful, but it's not worth the price. At first, we tried a price of 1599 yuan, but none of them were sold. Later, we dropped to 499 yuan, and the willingness to pay came up. ”
Because of the background of the founder, Mita can be familiar with legal AI entrepreneurship, and has also witnessed many other developers who rashly enter legal AI. Despite this, they are still deeply aware of the difficulties of pursuing profitability at home.
"You know, there are only 700,000 lawyers in the country. Of those 700,000 people, only 100,000 may be using AI products. Even if half of them buy this product, at our current pricing, it will only be 25 million in revenue. To achieve 100 million in revenue, you need to find 3 more similar scenarios. ”
If today is the eve of the "Thousand Model War", then the above information can give us a glimpse of how much space there is for the actual development of AI search.
AI search tools will definitely become more and more diverse and powerful, but simply networking large models and changing the form may not really understand the nature of search engines, nor do they understand how users use search.
A search engine is not just a tool for finding information, but a versatile platform that answers questions directly, provides gadgets such as calculators and converters, and a variety of other built-in features.
Although current AI searches have advantages over traditional searches in some aspects, they have problems such as slow generation speed, few results presentation, and unclear selection bias, which affect the fundamental indicator of result accuracy.
At the same time, the user experience after changing to a search box has not made a substantial leap. In this sense, OpenAI is using GPT-4o as a voice assistant this time, which may indeed be a little more interesting than launching a search product.
In the future, the development of search engines will depend more on product innovation than on technological progress. To surpass Google, Baidu, and Bing, I am afraid that what is needed is not to make a large model of networking first, but to always focus on how to solve the actual problems of users.
Welcome to long press or scan the QR code to listen to the full version of the "Dialogue with AI Entrepreneurs" podcast interview with Wang Yiwei: