The big model that has been on fire for more than half a year, and the controversy has gradually increased.
First, ChatGPT, which is a trendsetter, was exploded and traffic declined, and the growth rate of visitors turned negative. Microsoft Bing search, which integrates ChatGPT, has also been criticized a lot, and has announced the suspension of the "Browse with Bing" function that has only been online for a week.
But at the same time, domestic and foreign technology giants have not given up chasing the wind. In the past month alone, Meta, Huawei, JD.com, Alibaba, iFLYTEK, Netease, Ctrip, ByteDance, Zhihu and other major manufacturers have successively reported new news, Musk's new AI company "xAI" is rumored to be valued at $20 billion, and threatens to challenge openAI with the purpose of exploring the true nature of the universe.
On the one hand, there are fanatical big factories, on the other hand, gradually returning to the rational market and users who have lost enthusiasm, where will the future of big models go?
(Image via UNsplash)
Ice and fire of large models: the popularity has dropped significantly, and large manufacturers are still grasping research and development
The cooling of large models starts at the source.
Similarweb statistics show that ChatGPT has accumulated more than 100 million visitors, and the number of monthly website visits has exceeded 1 billion, and it has been in the rhythm of soaring traffic in the first quarter of this year. But this good day ended in the second quarter, and ChatGPT's traffic growth rate peaked and declined, and entered a negative growth stage at a very fast rate.
According to Similarweb's statistics, the month-on-month growth rate of ChatGPT website visits from April to June this year was 12.6%, 2.8% and -9.7%, respectively, and in January this year, this figure was still a staggering 131.6%. Even overseas competitors, such as Civitai and Claude, have not been spared, and the number of visits to Civitai has been shrinking since March, with an average visit time of only 14 minutes, and users are losing rapidly.
You know, ChatGPT has done a lot of work to maintain the popularity during this time. In mid-May, the official ChatGPT app was officially launched on iOS, instantly rushing to the second place in the App Store free app list in the United States, second only to Temu. However, this situation did not last too long, and the APP online area was limited to the United States, and it did not have a phenomenal communication effect like the website.
Looking back at the country, the situation is similar. Both the WeChat Index and the Baidu Index show that the popularity of ChatGPT and large models reached its peak in April and May this year, which coincided with the intensive release of products such as Baidu Wenxin Yiyan, Ali Tongyi Qianwen, iFLYTEK Xunfei Spark, and 360 Zhibrain. But in June, the enthusiasm of users began to cool, and the various major manufacturers took turns to hold various press conferences, which did not boost popularity.
At the same time, controversy over ChatGPT and large models began to increase.
In May this year, Microsoft announced access to ChatGPT, which supports users to search in the form of chat, and ChatGPT also has a built-in "Browse with Bing" function on the mobile terminal to support users to obtain information through Microsoft Bing on the mobile terminal. However, just a week after its launch, this feature was stopped by openAI.
The reason for the suspension is that the platform found that some users bypass the paywall of some websites integrated by the search system through this function and directly obtain paid content. The user's white prostitution is suspected of infringing intellectual property rights, which also makes paid websites feel dissatisfied, which further involves a new debate about whether large-model applications should be open source.
However, it seems that the enthusiasm of the big factories has not been extinguished. In the past month alone, many companies have released large-model products or updated the first generation of products, and many large manufacturers have announced the latest research and development process and threatened to smash large models.
On July 7, HUAWEI CLOUD CEO Zhang Pingan officially unveiled Pangu Model 3.0; on July 13, JD.com officially announced the launch of the Yanxi model, which is expected to be officially launched in August this year. On July 17, Ctrip released the first vertical large model of the travel industryCtrip asked...
Looking ahead, we can also see the new dynamics of a series of big manufacturers such as iFLYTEK, Meituan, Zhihu, ByteDance, and SenseTime. Jiang Tao, vice president of iFLYTEK, even said that in the second half of the year, he will "all in the Spark model" and firmly invest in the Spark model.
The market is turbulent, and it is impossible for large manufacturers to be unresponsive. However, the speed of decline in the popularity of large models and the process of research and development may indeed exceed expectations. After all, there is still a gap between domestic technology and foreign countries, and it takes a long time for large models to go online from project establishment, research and development, and even if they realize the changes in the wind outlet halfway, manufacturers are unlikely to give up the early investment.
In the period of national carnival, all enterprises are rushing for time, just want to develop large-model products as soon as possible and eat traffic dividends. At this time, everyone has not yet begun to think about application scenarios, commercialization models and other issues, and the race against time and competitors to compete for speed is the primary purpose. But when the tide fades, big factories have to face these practical problems.
C-end users don't buy it, these big models are developed in vain?
All kinds of controversies about the application scenarios and practicality of large models can ultimately be boiled down to the dispute between To B and To C.
The current situation is already very clear. Although the B-end large model has high investment scale, difficult research and development and fierce competition, the application prospect is worth looking forward to, especially for vertical industry large models; The C-end large model can indeed help manufacturers accumulate popularity when it first became popular, but the traffic dividend comes and goes quickly, and it is difficult to create too much revenue and profit for the enterprise.
Simply put, the B-end large model is continuously sought after because it can help enterprises reduce costs and increase efficiency, and is the direction of manufacturers' future efforts; The C-end big model is a money-burning project, and if it cannot run through the commercial closed loop in a short period of time, it is very likely to be abandoned.
Ctrip, which has just boarded the car, is a case worth paying attention to.
It is still difficult to judge what is outstanding and how effective it is. However, judging from the introduction at the press conference, Liang Jianzhang, chairman of the board of directors of Ctrip, hopes that Ctrip will become a travel assistant for users, and it is a certainty to take the C-end route.
Liang Jianzhang's highlights at the press conference include real-time, high-quality data, Ctrip's self-developed vertical models, as well as robots and search algorithms. In Liang's words, Ctrip is a "reliable answer database" for the travel industry, providing users with travel counselors and "error-free" answers based on more than 3,000 destinations, massive lists and user data around the world.
However, users have their own ideas: they don't want to pursue standard answers, preferring personalized, customized travel guides.
In recent years, OTA platforms such as Ctrip, Mafengwo, and Tuniu have been making efforts to create content, which is also to activate the community atmosphere and build a communication platform for users. The intervention of the large model has made the travel strategy standardized and industrialized, but it may allow users to lose the fun of finding routes in person and planning wine tours, restaurants, and attractions according to personal needs, which will also weaken the communication atmosphere of the community, which is contrary to the original intention of doing content.
Youyou, who plans to travel to western Sichuan in August this year, began to search for strategies on various platforms such as Ctrip, Qunar and Xiaohongshu two months in advance. After taking over the coordination of the small team of three people, he had to consider the accommodation and travel habits of the other two friends, design a route that could not be too tired, and take into account the attractions that everyone wanted to go, "worrying" a lot of brain cells.
"We also discussed many times in the small group, and the last person chose a must-visit attraction, but finally found that the time was too rushed, and we could only choose one of three in Yuzixi, Daocheng Aden, and Siguniang Mountain."
When asked if it would be more convenient to use AI large models, Youyou said that he could not give an affirmative answer, "after all, if you don't use it, you don't have the right to speak." But he also said that in the process of discussion, the mood with a few friends became higher and higher, the travel became more and more expected, and the effect of brainstorming to stimulate dopamine secretion may be difficult to replace by AI.
The same problem exists in Zhihu.
As a seed runner of the Chinese big model, Zhihu can be said to win at the starting line. First, Zhihu has an active community atmosphere, and the quantity and quality of content are in a leading position in the domestic Internet industry, providing sufficient resources for large model pre-training. Second, Zhihu started with Q&A, and user usage habits are very similar to search platforms, while ChatGPT, Bing, and Baidu have provided a model for the combination of large models and search and Q&A scenarios.
Zhihu's action is also very fast, and it has joined hands with Face Wall Technology to launch two large model products, Zhihai Chart AI and Search Aggregation, and the summary function of the heat list in the station. However, these products and functional internal tests have not set off much heat in Zhihu so far, let alone out of the circle.
In the final analysis, AI can try to provide standard answers, but it cannot simulate the personality and style of the respondent - personalization has always been a feature of the Zhihu Q&A community, which is an advantage that AI cannot replicate.
The R&D strategy of Ctrip and Zhihu's large models is worth considering, but the motivation is clear, and the investment is also worthy of recognition. However, in addition, there are more large-model products for the C-end that lack both application scenarios and differentiation, which inevitably gives people the feeling of "big models for the sake of big models". For example, Cloudwalk Technology provides a calm model for C-end users, and major search and navigation software can provide functions such as interesting Q&A, Chinese and English translation, and travel suggestions.
Companies that are currently developing or have launched C-end large model products may have to think about whether they have a goose that lays golden eggs or a hot potato.
AI technology is all-encompassing, and big manufacturers should not be trapped in big models
It is not surprising that large models encounter more difficulties on the C side than on the B side. After all, the demands of enterprises are the same: improve efficiency, increase profits, control costs, and their requirements for AI large models are simple and clear. C-end user needs are elusive, and large models are more often just icing on the cake than necessities.
It is worth noting that even openAI has begun to extend its tentacles to the B side, CEO Sam Altman said in a recent interview that the focus of the latest quarter is API port services for enterprises, and emphasized that it will strengthen cooperation with Microsoft.
CITIC Securities pointed out in the latest research report that from a medium and long-term perspective, the bursting of the bubble of the To C large model is a high probability event, and enterprises will shift their focus to plug-in development and vertical segmentation of large models. In this context, the large model application for the C-end and the developers behind it will inevitably be quickly reshuffled. In the end, perhaps only a few large factories with stable traffic entrances and application scenarios, such as Baidu and 360, can stand firm.
How to break this scene?
The answer isn't complicated at all: just take a step back and get rid of the superstition of big models, and companies can see that there is so much more AI can do.
According to the research content, large models are summarized in the field of generative AI, and the main functions include natural language understanding/generation, intelligent dialogue, text classification/translation, etc. But generative AI is only one branch of AI, and there are many categories such as robotics, machine learning, and computer vision. According to the degree of intelligence, it can be divided into weak artificial intelligence such as speech/image recognition, recommendation system, and general artificial intelligence.
All in all, AI technology is all-encompassing. Generative AI is very popular, but it does not represent the whole of AI, and high-level large model technology is not necessarily the most urgent need of users.
On July 17, in addition to Ctrip's release of Ctrip's question, NetEase Cloud Music also launched a new feature "Private DJ", which provides intelligent interpretation services for songs based on the upgraded recommendation algorithm, which is also part of NetEase Cloud Music's "AI + Music" strategy. Prior to this, NetEase Cloud Music has successively laid out AI songwriting and singing, AI music review, AI score recognition, AI-assisted creation and other businesses.
NetEase Cloud Music's "private DJ" did not rub the popularity of the model, but obediently focused on the function, including the accuracy of personalized recommendations, the identification of user preferences, and the explanation content. Using AI technology to achieve personalized recommendation has no actual connection with the seemingly tall big model, but it is the most basic requirement of users for music streaming. Being a "private DJ" is far more valuable to NetEase Cloud Music than making a big model.
With similar ideas, there are not a few big factories that have begun to dispel large models. Alimama recently launched a one-stop delivery system, the AI native vector database released by Tencent Cloud, Byte began to expand the robot team, Baidu and BMW Brilliance cooperated to explore the AI+ automotive business, and it was not limited to the field of large models.
As mentioned earlier, until now there are still many big manufacturers who have seized the time to squeeze into the track of large models. However, the B-side large model has a high research and development threshold, the C-end is too crowded and the application prospects are unclear, and it is no longer a good trick to imitate ChatGPT. Companies that find new ways will only increase.
Write at the end
At the 2023 World Artificial Intelligence Conference (WAIC 2023) held at the beginning of the month, large models were naturally the protagonists, but we can also see many companies bring other new technologies and products in the field of AI. For example, Ant Group's "Intelligent Risk Perception and Response Joint Anti-fraud System" won the "SAIL Star" award at the opening ceremony, and Minimally Invasive ® Robot brought two surgical robots and a puncture robot positioning system.
Big models are indeed hot, but not all businesses are suitable for rooting this track. The application of AI technology is very extensive, and there is not only a way out for large models. Instead of focusing on hot spots, jumping out of the limitations of general large models, and deeply combining AI technology with its own business and user needs, it is the right way out.
Concentrate:
At the request of the interviewee, Youyou in the article is a pseudonym.
Risk Warning:
This article does not constitute any investment advice, the market is risky, investment needs to be cautious.