This article is from the WeChat public account: Noise Reduction NoNoise (ID: forjingyijing), author: Jian Feiran, Sun Jing, title picture from: Visual China
Alchain Peanuts, an independent developer of AI tools, recently did a small test. On the GPT Store, he adjusted a GPT (simulated Claude 3 Opus) with 5,000+ users developed by himself to a payment model, wanting to see if overseas users really have a higher willingness to pay.
Six hours later, he received his first user payout, $5. In the backend screenshot he showed, the app had received $20 in the last 7 hours.
"Harvest a few coffee dollars a day. Alchain Peanut talked lightly about the original intention of paid testing - since Open AI can't do what was promised at the beginning of the year - to launch the GPTs developer revenue sharing plan in the first quarter, and the GPT Store is already in ruins, he wants to see if it is possible for AI developers to find their own business model.
Screenshot of the "Alchain Peanut" app backend
Since the second half of last year, basic large model manufacturers have been calling for those with deep pockets not to roll up the model layer and hurry up to develop AI applications. For example, Robin Li, the founder of Baidu, tirelessly preached: In the era of AI native, we need 1 million AI native applications.
IDC predicts that by 2024, more than 500 million new applications will be created worldwide, which is almost equal to the total number of applications accumulated over the past 40 years.
These optimistic judgments all point to the endgame, a new era after the Big Bang. However, we are more concerned about whether this frenzy led by large models will bring unprecedented "inflation" of AI apps.
On the one hand, the threshold for AI application development has been greatly lowered - when Apple launched the App Store in 2008, application developers still need to learn the Objective-C language, and today's large model manufacturers provide natural language development, and it no longer matters whether they understand the code or not, everyone can become a developer.
On the other hand, the iteration speed of AI technology far exceeds that of the mobile Internet. Chen Shukai, chief scientist of Entropy-based Technology, said bluntly that everyone knows that after the emergence of new models with stronger capabilities, such as ChatGPT5, it is possible that many of the things you are doing now have been done in vain, "Today's phenomenon is not available in the Internet era." "This also means that the life cycle of a batch of AI applications can be as short as a cockroach.
For now, at least, it seems that the gold mine of the era is still hidden in chaos. But no one wants to be under the table.
1. The era of "inflation" of AI applications: developers and "app factories" have launched
When an ordinary domestic user wants to experience AI assistant applications, he may suffer from oversupply of choice phobia: in the mobile phone app store, there are not only a variety of bean bags, Wenxin Yiyan, Xunfei Xinghuo, Kimi, Tongyi Qianwen, Zhipu Qingyan, Zidong Taichu...... There are also a bunch of shell products that look like ChatGPT, which can be said to have everything, but it seems that they will inevitably tend to be homogeneous.
It's not just the users who are confused. When asked how many AI applications they have launched, insiders from at least two major tech companies shook their heads at me: "I can't tell" or "I can't count".
In this wave of AI native applications, many large companies encourage internal crazy horse racing with the idea that all applications are worthy of being reconstructed by AI, which also leads to some AI applications being launched in the form of independent apps or PC plug-ins, which are not superimposed on the main app, and may not even be known to colleagues in other business departments.
No tech company wants to miss out on this wave of AI adoption. Moreover, Chinese companies are already known for being good at the application layer.
Judging from public information, Baidu, Byte, Ali, etc. are currently racing against time.
For example, Baidu, in addition to using AI to transform the original App and launch Wenxin Yiyan, has also released social AI applications such as "Wanhua", "Xiaokan Planet" and "SynClub", AI painting tools such as "Wenxin Yige", and medical AI applications such as "AI Medication Instructions", "Smart Waiting Room" and "Medical Notes".
In addition to the chatbot "Doubao", the AI companion class "Xinqing", and the AI social class "Talk Stove" (cat box), similar to Sora's AI video tool "Dreamina" is also in internal testing, and these are just the tip of the iceberg of byte ambitions. According to media reports, Byte is still secretly developing multi-modal digital human products and AI picture products, and the Jianying team is also developing new AI products in a closed manner.
Other companies, such as iFLYTEK, in addition to the chatbot "iFLYTEK Xinghuo", also have AI speaking assistant "Xinghuo Companion", iFLYTEK writing, smart badges and other efficiency tools online...... SenseTime, an AI technology company that previously focused on technical solutions, has also launched generative AI application products such as "Consultation", "Miaohua", "Ronin" and "Little Raccoon" in addition to releasing its own large models, and has included generative AI applications in the key direction of this year.
With the joint promotion of domestic and foreign basic large model manufacturers, technology solution providers, developers, and investors, AI applications are entering a supply-side explosion stage. In 2023, the number of AI projects on GitHub, the software project hosting platform, skyrocketed by 59.3%. According to Baidu's data, as of March this year, the number of new applications on its Qianfan AppBuilder platform exceeded 3,000 per week, and by April, the average number of new applications per week reached 6,000 or 7,000.
In Beijing's subway cars, there has even been a high-profile promotion of conversational AI applications such as Zhipu Qingyan for the C-end market. Prior to this, Kimi of the same type had already made its presence felt in online channels.
C-end users are also showing great enthusiasm for AIGC. According to QuestMobile's insights, the demand for standalone apps continues to grow. In January 2024, the number of deduplicated users of the TOP10 apps increased by 37 times year-on-year. The number of active users of the top app exceeds 50 million.
However, the "dark side" of this outbreak is the rapid birth and death of AI applications. In September last year, a16z, a well-known technology venture capital firm in the United States, compiled a list of the top 50 AI applications based on monthly visits, and in March this year, when the company launched its latest Top50 AI application list, it found that 40% of the applications on the list were new faces. This means that at least 20 AI applications have been left behind in half a year.
For this change in numbers, A16Z said, "eye-popping".
QuestMobile's data is a direct confirmation of this industry trend. In January this year, the activity rate of China's top AIGC apps was below 20%, which was relatively low, and in terms of loyalty, the three-day retention rate was below 50%, and the risk of churn was high, with the uninstall rate of some apps above 50%.
This is different from the mobile Internet era. In the early days of the mobile Internet, a large number of high-frequency native applications were quickly created and formed new business models, such as foreign Instagram, domestic WeChat, Didi, ...... "but AI has not yet reached this point", independent developer "Alchain Peanut" believes that the current AI application is more to improve the user experience of existing products, and mainly plays a role in productivity.
The progress of AI application is slower than everyone expected, and the main "bottleneck" behind it is probably the basic large model. To use an inappropriate analogy, it is unlikely that there will be a Cambrian explosion of species at the beginning of the birth of the Earth 4.5 billion years ago, because the Earth itself is still in a state of drastic change, being collided by other planets at every turn, and there is no atmospheric protective cover, so to speak, it has not yet been shaped. Only when the internal and external environment of the earth is relatively stable can there be life.
Large models are at a similar stage. Industry insiders recognize that the iteration of large models is almost measured in weeks. Reflected in the pop-ups of the tech media, the word "explosion" is running out of words, because every few days a new disruptor emerges.
Yan Junjie, the founder of MiniMax, mentioned in an interview with "Late" that the core of the source of product value is the model performance and algorithm ability, "You can do a lot of product features, but you will find that almost all the big improvements come from the progress of the model itself." ”
GPT-5, which is rumored to be unveiled this year, is believed to be an even bigger tsunami, or a watershed moment in the history of AI development. Jiang Xin, a person from a tool software company going overseas, mentioned in an exchange with "NoNoise" that there may be a large wave of shell applications squeezed by the iteration of large models this year, and a batch is expected to die. ”
For this, Jasper, the former star of AI text generation tools, is probably stuck in the throat. After OpenAI released a new generation of visual model Sora, some people in the industry are worried that a number of AI video products such as Pika and Runway will no longer survive.
In the view of "Alchain Peanut", some AI applications without moats are originally "carved on" and have little value, but some tool products that generate value after being combined with the scene will not be subverted, such as Monica, sider, immersive translation and other plug-in applications. They provide good value when combined with large model capabilities and actual user usage scenarios.
Second, valuable AI applications are combined with specific scenarios
Value is becoming an important criterion for measuring whether AI applications will be covered by technology iterations.
In the AI application list compiled by a16z, MaxAI, a productivity tool application, extracts some of the capabilities of ChatGPT, Claude, Gemini/Bard, Bing AI and other models, integrates them together, and makes an encapsulated plug-in based on application scenarios, which can help users summarize text, assist in writing, and create images.
Essentially, MaxA is in the "plumbing" business. According to industry insiders, the team has received high financing.
Another Chrome plug-in that has attracted attention, Monica, and its founder Xiao Hong mentioned when explaining the value of the product, that they do a lot of work in very specific scenarios, such as answering emails, helping users summarize articles, or helping users summarize each piece of content when they open a YouTube video. These functions are embedded in the browser through plug-ins, because browser plug-ins are a relatively mainstream product form overseas.
At the Geek Park conference, Xiao Hong also mentioned that the key point is that application entrepreneurs should find ways to collect data from users in specific scenarios, and with data, AI applications can form a collaborative relationship with the brain of large models.
Data is also the reason why many AI applications will continue to do even if they don't find a business model. Jiang Xin told us that like personal assistants and efficiency tools, user data and behavior data can be accumulated, so that there is a data reference for the next iteration of the product.
As for the "short life" of the App brought about by the iteration of the basic large model, Jiang Xin said frankly that it is okay to force the C-end application to continue to iterate and deepen the subdivision scenario, and also force the developer to think - in which direction should the product be iterated in the next step?
Jiang Xin believes that C-end applications are easier to reach the ceiling than B-end applications, because at present, large companies such as Byte and Baidu, and even technical solution providers such as SenseTime have begun to do AI applications, and the competition will be fierce.
For small, large-scale model startups, the key is how to find their own competitive barriers. Zhuang Minghao, vice president of Qumaru Network and former vice president of Jingwei China, previously told the media that he observed that many startups have turned to the model of "small workshops making gadgets": first identify a function and product that is not yet available in the market, seize the window period, and quickly promote it through operational means, even if this window period is only 3 to 6 months, you can also earn a sum of money, and then continue to look for new market opportunities.
In the case of the Wonderful Duck camera, this product was immediately sought after by the market, and it was charged from the first day, but the market quickly cooled down after two months. This is completely different from the entrepreneurial model of the early application in the mobile Internet era to compete for the user market for free, and then gradually start to charge.
Engineer Xu of the Shanghai-Chongqing Institute of Artificial Intelligence told us that the biggest difference between entrepreneurship in the AI era and the mobile Internet era is that at that time, everyone's main focus was to seize market share, but now everyone is mainly exploring business models.
According to the "China AIGC Application Panorama Report" recently released by the qubit think tank, the C-end AIGC products are mainly intelligent assistants and image generation productivity tools, although the number of users is large (the pure C-end accounts for more than 50%), but the profitability is generally not optimistic, and nearly 50% of the products still have no clear revenue model, mainly free. In contrast, the business model of B-end products is relatively clear, mainly based on subscription and pay-as-you-go.
Jiang Xin bluntly said that many application developers, including his company, actually want to cut the B-end scenario and aim at the industry + AI, but due to the lack of industry kown-how, it is actually difficult to cut in, let alone greatly transform.
When communicating with "NoNoise", a business leader of Ant Digital also believes that it will take a little time for the scenario-based capabilities supported by large models to be implemented and made valuable in the industry. This "value" either makes the previous business more effective, or allows the enterprise to do the business that it cannot do before. But at present, most companies are still in the exploratory stage, "not so fast".
From a global perspective, the pursuit of venture capital has also become the exclusive path of a type of AI company. Italian tech company Bending Spoons, for example, is behind the video editor Splice and the photo enhancer Remini. Not long ago, the company announced that it had received $155 million in equity financing.
III. Sudden arrival
According to Gary Marcus, an AI researcher and cognitive scientist, generative AI will need to improve dramatically to reach the level of change that the Internet and even smartphones have brought about.
A thriving AI native application ecosystem requires the complementarity of large models, intelligent computing power, and a new paradigm of AI native application R&D. From this perspective, the advent of killer AI-native applications is not in a hurry.
Chen Shukai, chief scientist of Entropy-based Technology, said bluntly, "Everyone has to do development, do testing, overcome various difficulties, and solve all kinds of practical problems, so it can't be too fast, but moisturizing things silently, everyone is really doing it." ”
He believes that as a developer of AI applications, the first thing to pay attention to is how AI solves problems in actual business, and the second is to keep an eye on the latest technological developments.
Chen Shukai saw that the current application models are also in the process of continuous maturity and development, such as knowledge base-based applications and agent-based applications, which will continue to have new achievements, which can directly improve the level of existing AI applications. At the same time, the basic capabilities of the large model itself are constantly improving, and some limitations of AI applications will continue to be broken.
Many people in the industry believe that AI Agent may be able to create AI-native scenarios and applications. AI agent refers to an agent with autonomous decision-making ability, environment perception ability, and reaction ability in the field of artificial intelligence. AI agents emphasize the agency's autonomy, reactivity, initiative, and sociality, while large models with human-like brain-like functions such as understanding generation, complex reasoning, and self-learning can be used as the basis for agents.
When the AI Agent appears in the form of an app, it also means that Open AI's GPT Store may face competition from the App Store. After all, according to the Writerbuddy study, more than 60% of users are accustomed to accessing AI tools on mobile devices such as mobile phones.
At the same time, it may also indicate that startups in the field of AI may meet the opportunities and challenges at a larger level.
Looking back at the mobile Internet era, in 2014, China's smartphone users exceeded 500 million for the first time, becoming the country with the largest number of smartphone users. The crazy growth of mobile apps is taking over the new trend. In April 2015, the total number of apps in major app stores in mainland China exceeded 4 million.
But the average lifespan of these apps is only ten months. Among them, 85% of users delete their downloaded apps from their phones within 1 month, and after 5 months, the retention rate of these apps is only 5%.
And this scenario is afraid that in the AI era, it will only become more and more miserable, and "inflation" in AI applications is inevitable.
But even so, at the time when the same big factory was in power, a series of start-ups such as ByteDance, Momo, and Didi still fought their own way out. Tencent even relied on WeChat to leave the opponents who were originally on an equal footing far behind.
So, in the new round of the "gold mine" of the times, who is likely to become the winner of the new round? Who is likely to be left behind by the times? We are full of curiosity about this, but this answer must be left to time first.
(Note: At the request of the interviewee, Jiang Xin is a pseudonym)
This article is from the WeChat public account: NoNoise (ID: forjingyijing), author: Jian Feiran, Sun Jing
This content is the author's independent view and does not represent the position of Tiger Sniff. May not be reproduced without permission, please contact [email protected] for authorization
People who are changing and want to change the world are all in the Tiger Sniff APP