The Kimi intelligent assistant has become a phenomenon. As an AI mega model product, it directly influenced the Chinese A-share market this week.
The startup behind Kimi, a company called Darkside of the Moon, has just raised its Series A funding. However, a group of A-share listed companies has already seen their stocks rise in anticipation. Many of them are rumored to have some form of collaboration with Kimi, although not all of these collaborations have been confirmed. This has led to the formation of a "Kimi Concept Stocks" sector in the A-share market.
Kimi's concept stocks have gained attention because it's seen as breaking out of the norm. From analyses by sell-side institutions to snippets circulating in WeChat groups, there's a notion being promoted: Will AI models soon become as ubiquitous as apps like TikTok, WeChat, and Taobao, integrated into smartphones on every street corner?
If last year marked the "iPhone moment" for AI models, could we say we're now witnessing the "iPhone 4 moment" for AI models?
Let me throw some cold water on this: Kimi is indeed impressive, as are many other AI model companies' products. However, the conditions for the "iPhone 4 moment" of AI models, particularly those represented by chatbot applications, have yet to materialize.
There are at least two contradictory issues that need addressing for AI model applications to reach the masses:
Firstly, AI model products like Chatbots require users to actively ask questions, which is inherently counterintuitive and contradicts the past trend of computer applications catering to human nature.
Secondly, there's the contradiction between the uncertainty of AI model-generated content and users' expectations of the certainty provided by traditional computer applications.
Mira Murati, a prominent executive at Open AI, holds the position of CTO, despite not being an engineer but rather a product manager at Tesla, where she was involved in the development of the Model X.
With such staffing, it seems that the early founders of OpenAI like Sam Altman and Elon Musk have already recognized the future trend: AI models as products.
AI models inherently can serve as products because their core capability is directly providing information. The primary criteria for evaluating the capability of large models lie in their generalization ability and whether they exhibit "emergence" when outputting information.
According to the AI Product Ranking published by the self-media "AI Product List" (as of February), the most visited products are mainly AI chatbots, which are also heavily promoted by major players like Baidu, Alibaba, and ByteDance. Kimi's visit volume has climbed to third place in February, and given its recent need for expansion due to high traffic, it's likely to have grown several times over.
As an AI model product, Kimi has made many efforts in terms of product usability and was the first to highlight long texts as the core selling point. Products like Tongyi Qianwen have begun to follow suit. These practices have laid the foundation for Kimi's success in March, but this doesn't mean Kimi will continue to stand out. Currently, there are two major contradictions that need time to resolve for AI model products.
One contradiction is that AI model products like Chatbots require users to actively ask questions, which is inherently counterintuitive and contradicts the user-friendly approach adopted by computer applications over the past decade.
During the era of mobile internet over the past decade, the public's perception of computer applications has been reshaped by various mobile apps. In an era focused on user scale, all mobile apps strive to cater to human nature, eventually evolving to minimize user choices through single-column information streams. Meanwhile, platforms attempt to precisely match individual users' preferences through algorithmic recommendations, providing content that quickly enhances pleasure.
Even if not to such extremes, most app product logics aim to provide as many suggestions as possible, making it easy for users to click without actively inputting any content after a single-use cycle.
This extreme catering to human nature may lead many people to believe that the apps installed on their phones will naturally guess what they need. Tools that require users to actively input keywords to search for information, like search engines, are also shrinking in the market. (In a sense, Xiaohongshu "saved" the search engine.)
What AI model products need the most is for users to actively inform them of "what I need," and because of the limitations of model capabilities, users are better off providing as detailed and accurate information about their needs as possible.
Currently, AI conversation programs like Kimi are widely adopted by users with clear tool needs, who explicitly need AI model products to help organize documents and analyze materials.
However, whether it's Wenxin Yiyu, Tongyi Qianwen, or Kimi, at this stage, even if they can't directly play a video or music like "Xiaoai Classmate," more users will still feel puzzled by their dialogue boxes, wondering "what exactly can this AI do?"
In fact, questioning has always been a very advanced skill. The famous mathematician Qiu Chengtong has repeatedly pointed out that top students at Chinese universities lack the ability to "ask questions," as he wrote in his article "First-Class Talents Begin with Learning to 'Ask'". He believes that good researchers must be good at asking questions and be able to pose good questions.
If even top students at top universities lack questioning skills, this is an even greater challenge for ordinary people.
Taking Kimi as an example, in terms of product design, Kimi's ability to handle long texts has been praised by many. In fact, Kimi has designed a feature to embed questioning "prompts" in all dialogues to help users discover which questions they can follow up with. However, it seems that this feature still needs improvement. Personally, I have felt that this feature is sometimes a bit redundant and doesn't accurately predict what questions users need to ask next, still requiring me to think of questions myself. Apart from the steep learning curve for end-users, there's another contradiction: the uncertainty inherent in content generated by AI large models clashes with users' expectations for deterministic content from traditional computer applications.
The key capability of AI large models is emergence, which means they can produce content beyond what people anticipate, especially with deep usage. Huang Renxun highlighted some critical applications of generative AI in his "20,000-word speech" at GTC2024. In the field of drug discovery, for instance, AI can accelerate the discovery process by predicting protein structures and molecular docking, providing researchers with novel and uncertain information for reference.
Currently, user feedback on social platforms suggests that such issues are not yet problematic, influenced by user base size and frequency of use. However, in recent years, the primary goal of various applications has been to provide deterministic information to users, such as weather forecasts, product purchases, O2O services, and even search engines, which also tout accuracy (though not always achieved).
Additionally, there's another type of issue based on user feedback: different users evaluating AI large model products differently based on their unique needs. Some prefer one model over another due to differing strengths, adding to the uncertainty of evaluation criteria and generated content.
Product quality is crucial, as is user operation for consumer-end users. Industry pioneers have the opportunity to capture user mindshare early but also bear the responsibility of user cultivation, often synchronously.
Currently, apart from ChatGPT's high visibility, other brands are generally average in both awareness and favorability among users. This is because most users have yet to experience all AI large models, with their understanding of GPT mostly limited to news and information.
Capturing user mindshare relies on the product itself and user cultivation. The iPhone 4, for example, educated a large number of users, with its lifelike skeuomorphic UI design becoming the beginning of smartphone cognition for ordinary users. Looking at ChatGPT's iterations, there have been many adjustments, but pitfalls like GPT Store have been encountered. The interaction logic of domestic products is basically consistent with GPT, while Kimi has pondered on the probing level, but whether there is a need for more Chinese-oriented improvements requires bold exploration. In the short term, overseas AI large model conversational robot products will certainly not enter the Chinese market, so the feasibility verification of product-level localization must be undertaken by Chinese enterprises themselves.
Previously, when Wenxin Yiyuan was launched, there was a large amount of paid traffic, and now products like Kimi, Beanbag, and Tongyi Qianwen have also embarked on the road of paid traffic to attract new users. At this stage, a group of users has begun to be early adopters in the industry. Starting user operation for these users is the first step in capturing user mindshare, akin to the successful path of domestic smartphone manufacturers in the past.
Some large model manufacturers have already acquired B-end paid users, who receive targeted services and operations. However, the broadest market for AI large model products still lies in the consumer end, making early layout critical.
Why emphasize consumer-end users? Getting consumer-end users truly engaged is key to forming a positive industry cycle.
All AI large model entrepreneurs need to consider one thing: on social platforms, self-propagation by consumer-end users is the most important way to spread product word-of-mouth. Consumer-end sentiment can profoundly impact B-end and G-end. In the explosive stage of large model entrepreneurship, seizing consumer-end word-of-mouth is the most crucial aspect of building enterprise brands.
In consumer user operations, cultivating a group of seed users is crucial, as seen in the smartphone industry. Establishing a user community, providing incentives for active users, and engaging with user feedback are proven methods. These active seed users will have a significant influence beyond their numbers.
Furthermore, it's important to expose a large number of question-answer cases to C-end users. This helps users become accustomed to the product and accept the brand. Even if future AI products evolve beyond chatbots, brand loyalty remains a significant asset.
Products that go against human nature are often the cheapest. During the peak years of the iPhone, some commented that it broke down class barriers and recognized the equality of all. In that era, both elites and commoners used essentially the same product.
Currently, AI tools may also be products that bridge social class disparities. They offer maximal intellectual support at minimal cost.
During the era of mobile internet, various applications such as online games and short videos catered intensely to human nature to acquire more users. These applications typically provide short bursts of high-intensity stimulation to the brain, potentially leading to attention deficits, memory loss, sleep disorders, and emotional issues.
These side effects should be factored into the cost of using such applications.
Perhaps as AI capabilities continue to improve, a series of such products will be packaged. They may offer convenience but also come with a slew of by-products, essentially pushing "chocolate-flavored crap" to users.
Current AI products, especially conversational AI products, are still relatively benign intelligent tools. To harness high-quality assistance from them, people need two abilities: how to ask questions effectively (including follow-up questions) and how to judge and filter information.
Recently, some have argued that training and reasoning with AI models require substantial resources, leading to a continuous rise in prices. However, apart from a few exceptions, it's unlikely that the industry as a whole will see a price hike. This is because intense competition exists in the AI model industry. As long as this competition persists, model manufacturers will need to acquire users through various means, with pricing being the most effective. Therefore, for a considerable period, AI model products will be subsidized by enterprises and investors to penetrate the market.
From the perspective of usage cost, these products that go against human nature are at least the cheapest.
