Image source: Generated by Unbounded AI

In 2023, Langzi met Huoru in an entrepreneur knowledge sharing group. They had similar educational backgrounds and big ideas. The two people, who had working experience in the factory, naturally became friends. As two enthusiasts of AI products, and they had some consensus on the methods and cognition of doing things, the two decided to start a business in the AI ​​track.

In early 2024, Huoru attracted other like-minded friends and formed a small team of 5 people. Lang Zailai was responsible for product design, while Huoru and other members focused on technology and started ALL IN AI entrepreneurship. In late November, after nearly a year of development and product polishing, their AI search product Hika began testing and was officially launched at the end of December. (Website: hika.fyi)

The track is very single, and the AI ​​search products are all "re-shelled versions of Perplexity"

Beluga Whale: The two founders are both deep users of AI products and have used many products deeply. When they decided to start a business, why did the team determine the main direction of attack as AI search? What about the engine?

Langzi: First of all, AI search has always been highly discussed in the industry and is a very popular track. Moreover, AI search is also my most commonly used AI product. My understanding of this product type is also More profound. Compared with traditional search, AI search can bring greater experience differences, or upgrades, to users.

However, I think the forms of AI search products currently on the market are relatively simple. Generally speaking, all AI search products are "shelled versions of Perplexity". I think we can make something different. stuff, so I chose the direction of AI search.

Beluga whale going out to sea: As the "leader" in the field of AI search, Perplexity has received hundreds of millions of visits in a single month. It has raised 4 rounds of financing this year, and its valuation has exceeded US$9 billion. It can be regarded as defining AI. Search product form. So, what do you think can be improved or differentiated about the current AI search products?

Langzi: As far as I understand, there are three problems. The large model is not that smart, the experience of searching for pictures is not good, and users cannot go deep into the content they are interested in.

The unsmartness of large models is reflected in two points. First, the current accuracy of AI search is not that high, the frequency of hallucination problems is still high, and the ability of large models to filter correct information needs to be upgraded. The second is that large models cannot give personalized answers that meet user needs. Take travel as an example. They are also searching for food. Some people like heavier flavors, while others like lighter ones, but the recommendations given by AI searches are the same.

In terms of pictures and links, this is veryA module that will be included in many AI searches. Traditional engines have strong data and technical accumulation on image modalities. When I search on Google, it can give me hundreds or thousands of images to choose from. However, in AI search engines, although large models can match pictures by understanding questions, the number of pictures presented is not only small, but in most cases they are not relevant to the answer text, losing the meaning of assisting understanding.

The third is that users cannot go deep into the parts they are interested in. I discovered this question based on my own experience. When searching, in addition to wanting to get the answer to a single question, I also wanted to understand other aspects of the matter for a more comprehensive understanding. But products such as Perplexity will only give me a relatively concise "answer", and it is difficult for me to learn more about this topic, which we call the "problem domain". Although Perplexity has also added an "extended question" function to its product design, it is still not enough in my opinion.

The content given by Perplexity is relatively concise｜Image source: Perplexity

When people face a new thing, they understand it from different aspects and then combine it into A comprehensive understanding of things, but Perplexity only gives one answer and lacks the process of guiding users to think.

Huru: I would like to add that firstly, the AI ​​search product can only give more conventional answers and is not very unique. The pictures given are basically matched with whatever content is searched. Pictures, for example, when searching for an Apple phone, a picture of an Apple phone is given, which does not help users understand. In addition, the interfaces of all AI search products are very complicated, and the pages are filled with text, reference links, pictures, follow-up questions, etc. Seeing such an interface, my willingness to use it has been greatly reduced.

AI search is not a search?

Beluga: Based on what you mentioned before, how does Hika optimize Perplexity in terms of accuracy, uniqueness, personalization, etc.? What is the core difference between Hika and Perplexity?

Langzai: If Perplexity provides customized packages, then Hika provides a buffet. For us, we cannot do model-level optimization, and it is not what we are good at. Instead, we have put a lot of effort into product design.

Users can use "Tell more" and "Ask more" to let the AI ​​supplement the content or ask further questions

Hika will first focus on several aspects of this problem. Give some relatively concise answers for users to "take a quick look" and help users find "keywords". Then, after the user finds the keywords they are interested in, they click "Deep Into This Section" to let the AI ​​expand the information on its own, or the user canThe guide asked. During further interaction, the user also feeds back his intentions to Hika, and the user gets closer to the truth.

With the help of the logic diagrams and tables we provide, users can also capture the relationship between various concepts. Based on this, users can continue to interact with Hika and get "infinitely close" to the answer.

Hika’s so-called accuracy is not absolute accuracy, but providing the answers users need. We are not just a search product, but also a “thinking product” that helps users think.

Beluga Whale: From what it sounds like, the main difference of Hika is that it helps users get the answers they need through click interaction and multiple rounds of questioning. From our observations, most general users simply want a concise answer when searching. So will the group of people targeted by Hika be smaller than other AI search products?

Langzi: First of all, no matter in the past or now, people's thirst for knowledge/information has never changed. I think this need is as natural as eating and sleeping. Although most users generally only need a simple answer, every user will always face relatively complex problems when studying and working, and there will be situations where they need to understand multiple information. Although Hika's product form may be pointing in a relatively niche direction at present, I don't think what we are doing is a niche thing.

On the other hand, AI search is not really searching, but more like asking questions. If it is just about search behavior, traditional search engines like Google can already meet the needs. However, the current AI search products extend the process of summarizing, questioning, etc. in addition to searching for information. AI search only makes things that previous products cannot do. Things are done and presented in search form. From this perspective, Hika cannot be said to be a "niche".

White Whale Goes to the Sea: For many AI search products, accuracy is a big problem, and Hika’s positioning may be deeper than other AI search products. So can the combination of "large model + public information" meet the needs? In terms of accuracy, what technical optimizations has Hika made?

Langzai: Hika’s design idea is to find a problem scenario that is suitable for AI transformation and is also universal. This type of problem is not a simple problem that can be solved in one sentence, nor does it require citing a large number of professionals. The in-depth problems of the paper are rather complex, relatively high-frequency (work, study, hobbies, etc.) questions in between. Such questions can be answered by relying on public information. The large model is mainly responsible for presenting logical answers, but how to give the large model more perspectives to make its answers more in line with user needs actually goes back to the product design level mentioned above.

Huru: The illusion of large models is an inevitable problem in current AI searches, soWe can only try to avoid this problem. The general idea is to try to provide AI with high-quality information sources. Currently, we will rate the credibility of different sources (websites) and reduce the search weight of websites with low credibility.

In addition, Hika has also discussed many plans internally, such as adding a "reflection" process for each answer block. Hika will call another AI search to verify the previously presented answer, and then use the verification results to It is presented next to it in the form of "confidence level" to tell the user how credible the content is. However, this feature is currently just a concept and is still under discussion.

White Whale Goes to the Sea: We feel that those currently willing to use Hika are those who are more curious and have higher requirements for information accuracy. Are Hika users more difficult to accept error information than ordinary users? ? Will introducing a verification mechanism incur high human costs?

Huo Ru: From the current point of view, more accurate information sources can basically be listed manually, such as Wikipedia, mainstream news websites, etc., coupled with the help of large models, we can It is more efficient to annotate information sources and does not consume a lot of time.

Langzai: No matter what type of user, as long as they perceive the occurrence of errors, they will have a bad experience. Traditional search engines do not actually have this problem because they only provide information and users are responsible for analyzing the information. However, when AI search summarizes and analyzes the information, it needs to be responsible for the accuracy of the information. This is a common problem in the AI ​​search industry.

To completely solve the accuracy problem, we have to wait for the further development of large model technology, but in reality, I think users can also participate, such as adding some functions for scoring output content and the above mentioned The "confidence" function can better help users use AI search.

Beluga Whale Goes to the Sea: Understood in this way, we partially solve the exact problem through interaction, and help users understand a thing more deeply. What solutions do we have for the image search and personalized content mentioned earlier? Plan?

Langzai: Pictures are a form of information. We may consider adding them as input modalities to Hika. Users can interact with Hika by uploading pictures; and the basis of personalized content, It is to conduct deeper insights and analysis of user behavior. In this part, the Hika team is still learning.

Beluga: Hika is currently open for testing, and several internal testing groups are full. What is the current user feedback? What are the plans or directions for future product iterations?

Langzi: Everyone’s acceptance of Hika actually exceeds our expectations. Some users think that they can see an answer or the overall context of a question through Hika, and they can get some inspiration and inspiration. In the future, we will enhance the information richness and hierarchy of Hika to allow users toGet more information in Hika and allow AI to more intelligently provide content that meets user needs.

Can small and medium-sized developers use AI search engines?

Beluga whale going overseas: The search engine itself is a track with "large traffic and obvious head effect". Overseas, Google accounts for more than 90% In terms of market share, Baidu also has a market share of 70%+ in China. For small and medium-sized developers, what opportunities do you hope to find on this track?

Lang Zai: I think the AI ​​era has just begun, and the model capabilities are not yet perfect. There has not yet been a winning Killer App. Everything is still variable. This is at the industry development level. Chance.

At the product level, AI search products are still a relatively new track, and the current product design is relatively simple, but being single does not mean that its product design is correct. For Hika, I think interaction with users is a feature of Hika, and with the improvement of subsequent model capabilities, this interaction method will also change, which is likely to greatly improve the user experience and rapidly expand the user scale in a short period of time.

But on the other hand, it is indeed quite difficult to gain a foothold in the AI ​​search track. Because all products basically use the same underlying large model, the differences between different products are reflected in the presentation of the front end. However, there is no right answer in product design, and no one knows which design best meets user needs. This is a challenge for us, but it is also an opportunity.

Beluga: What is the current cost situation for Hika? Are there any future commercialization plans?

Huru: In terms of cost, it is basically controllable. The most important cost is the cost of calling the API of the model and search engine. We have currently applied for Microsoft's Entrepreneur Program, which can cover part of the cost for us. In the future, we will actively look for some large model manufacturers to cooperate and try to keep the cost within controllable state.

Langzi: As for the commercialization path, my vision is basically two directions. The first is the traditional commercialization model of search engines - advertising monetization. However, compared with the pure publicity role of traditional advertising, we can provide users with more valuable information by designing some recommendation mechanisms, and also help advertisers reach more precise targets. Reach the right users. This is a feasible solution when traffic increases in the future.

The other direction is actually that after the traffic increases, the users themselves will also be stratified. For different users, we can provide different functions for monetization. For example, if a user uses Hika to do in-depth research or write a paper, we can match the best model to this user, provide higher quality information, and charge the user.

But ourThe product is still in its early stages and these commercialization paths are still being conceived.

Written at the end

In the current AI search engine market, we I once observed in the topic "Raising 1.28 billion U.S. dollars in half a year, thinking there is no chance, crazy financing", in the first half of 2024, the world's major search engines raised a total of 1.28 billion U.S. dollars, but the general search engine for the public has Financing is less than 1 billion (including Perplexity), search engine opportunities are mainly in vertical scenarios and ToB. In the financing news of the search engine track in the second half of the year, the proportion of general search engines is still not that high.

As far as the general search engine track is concerned, there are probably several categories of "players". One is traditional search engine players such as Google and Microsoft, and the other is a few star startups such as Domestic Secret Tower and foreign Perplexity. Enterprises, and small and medium-sized developers building a general search engine seem to be the combination with the least chance.

However, judging from the narratives of the two founders Langzai and Huoru, Perplexity still has many problems, and the product model has not yet been finalized. This uncertainty includes small and medium-sized developers. There are opportunities for all “players”, but how to break in and how to make search engines into a mass product are still full of challenges.