GuShiio.com Wen| There is always a reason
In just one holiday, due to the emergence of DeepSeek, the global competition of strength in the field of artificial intelligence has undergone an unexpected reversal, and the earthquake caused by this Chinese startup and its products in the global technology circle is still continuing.
On February 5, benefiting from the popularity of the DeepSeek concept, the stocks of nearly 20 companies in my country’s A-share market rose strongly. Parallel Technology listed on the China Stock Exchange rose at a daily limit of 30cm, and daily interactions continued to rise by 20cm. In contrast, U.S. stocks are not very good. Since January 20, Nvidia’s market value has evaporated by 13.84%, and TSMC has fallen by 3.52%. Google and Ultramicro Corporation, which have just released financial reports, have achieved strong growth, Wall Street is not buying it.
The emergence of DeepSeek not only boosted the confidence of my country’s technology companies, but also inspired Europe, which believed that they could also become key participants in AI innovation.
But beyond this great joy and enthusiasm, Kimi, who launched a new model at almost the same time as DeepSeek, seems to be relegating to a supporting role.
Kimi fell out of favor?
Last year, everyone was talking about Kimi, but this year everyone is talking about DeepSeek.
On the eve of the Spring Festival, domestic large models rushed to update before the festival. At the same time that DeepSeek released the DeepSeek-R1 model, Dark Side of the Moon also launched a new Kimi k1.5 model. These two reasoning models are fully benchmarked against the full version of OpenAI o1. At first, once the new models of DeepSeek and Dark Side of the Moon were released, they actually caused heated discussions among overseas user groups. However, after DeepSeek quickly exploded the AI circle, almost all the focus of global technology was on DeepSeek.
Overseas, in the past few days, mainstream media in Europe and the United States have worked overtime on weekends and reported overnight. They have described the shock it has brought to the United States by shaking Silicon Valley. Nvidia rarely issued a statement praising the R1 model as an excellent artificial intelligence advancement.
As popularity continues to rise, DeepSeek’s user growth has also accelerated.
According to the latest statistics released by the domestic AI product list, the DeepSeek application has achieved remarkable results only 20 days after it was launched. Its daily active users quickly exceeded the 20 million mark, with the specific number reaching 22.15 million, successfully surpassing the daily active users of Doubao at 16.95 million, and has reached 41.6% of ChatGPT’s daily active users. What’s even more exaggerated is that DeepSeek’s monthly active users (MAUs) reached 33.7 million in just 21 days after launch.
Amid DeepSeek’s singing and praising all the way, Kimi’s new model was naturally a little embarrassed. Last year, as the first AI application to go out of the circle and successfully broke out on the C-side, Kimi can be said to be in the limelight. The new model of the Dark Side of the Moon has high hopes in the industry. However, who would have expected DeepSeek to make a blockbuster? It has successfully trained a model comparable to ChatGPT-4o at a very low cost. Compared with this achievement, the technological progress of Kimi’s new model is less eye-catching and outstanding.
Especially based on DeepSeek’s growth rate, let alone surpassing Kimi’s smart assistant and ByteDance’s bean buns, it seems that it is just around the corner to keep pace with ChatGPT in the world.
DeepSeek ‘s surprise attack on Silicon Valley technology companies is not just OpenAI that feels chilly. If DeepSeek occupies the number one position in the hearts of users in the domestic market, then the appeal to users of two temporarily leading AI applications, Doubao and Kimi Smart Assistant, will naturally weaken.
On social platforms, many users have already felt the gap between DeepSeek, Doubao, and Kimi smart assistants. Even after using DeepSeek, they began to dislike other big models. One user bluntly said, I use both products, but I only use Kimi k1.5 when DeepSeek does not respond. It is too mediocre, and DeepSeek always surprises me, especially in terms of imagination.
Although the reversal of users and public opinion will not make other products such as Kimi’s smart assistant lose their appeal in the short term, once DeepSeek’s advantages overwhelmingly surpass them, traffic and capital are likely to abandon Kimi and the company behind it. Especially for the dark side of the moon, which likes to burn money, it will mean a greater crisis.
“Marketing faction and technology flow”
In October 2023, Kimi made its debut. Relying on the large-scale model of hundreds of billions of parameters, it supported the input of 200,000 words of long text, and for the first time, it ranked first in the world in the capacity of dialogue boxes. Five months later, Kimi took another big step and upgraded to 2 million words.
The advantage of using long texts as the main focus is the key to Kimi’s stand out among many domestic models. However, Kimi’s emergence and popularity, and the large-scale growth in user base do not rely on technology, but marketing.
Starting from the Spring Festival of 2024, Kimi has targeted Station B and started a crazy campaign. At that time, on the homepage of Station B or in the recommendation area of the video, users can always see a promotion link called Kimi AI intelligent assistant. Benefiting from Station B, Kimi’s traffic has shown a significant growth trend, widening the gap with other applications, and thus laying the foundation for it to become a phenomenal AI application.
Kimi has ignited the war of AI investment and involved more and more large manufacturers and large-model startups in marketing wars. However, it is undeniable that this move has provided AI applications with a breakthrough in my country’s generative AI. Opening up the to C market provides a new path. However, the emergence and popularity of DeepSeek now confirms that domestic large models and applications can completely conquer the market with their technology or cost advantages, and this poses a certain threat to Kimi’s development route where marketing is the king.
In the past year, Kimi’s rapid growth on the C-terminal has allowed the outside world to see the potential of a “Killer APP”, but DeepSeek’s shocking capabilities undoubtedly mean that it is closer to a “Killer APP” than Kimi.
In fact, there have been doubts in the technology circle about the dark side of the moon to burn money on a large scale and rely on marketing to maintain user growth.
First, endless investment continues to increase the cost of acquiring customers for AI applications, making many startups with weak financing capabilities miserable; according to APPGrowing data, since March 2024, Kimi’s advertisements have reached tens of millions of yuan, reaching 220 million yuan in October and 200 million yuan in November.
Second, although investment and burning money have brought in a surge in user numbers, user retention is not high. Previously, a set of large model retained data was circulated in the industry. Data shows that the user Retention rate of large model products after 30 days is less than 1%. An industry insider said that it has been an industry consensus for a long time to retain users after purchasing volume, but the current product form may not be able to withstand it. Products like Kimi are essentially tools, and user duration and Retention rate are inherently not too high. rdquo;
Choosing a technology or a market has long been discussed in the domestic model. The choices and priorities of the companies behind it are different. This difference actually causes the gap and fault between the current major AI applications in the C-end market to a certain extent. Originally, the lead of Kimi and Doubao made the outside world more optimistic about the user-oriented market grabbing strategy. However, now the emergence of DeepSeek is likely to become a watershed in an industry, allowing the technical route to dominate the mainstream. In contrast, the effect brought by marketing will only become weaker and weaker.
Of course, it’s not that Kimi failed to take into account technology, but its reliance on marketing may affect future development.
Big models completely end the era of long texts?
Long texts are Kimi’s technical support point. It is precisely because of Kimi that domestic model companies have rolled up long texts and expanded the context, but this has also led to the reduction of Kimi’s advantages in long texts.
For example, Ali, on November 20, 2024, Alibaba Cloud Bailian officially launched its new Qwen2.5-Turbo model. The biggest highlight of this model is that it supports ultra-long context processing of up to 1 million tokens; According to the latest news, MiniMax has open-source the latest basic language model MiniMax-Text-01 and the visual multimodal model MiniMax-VL-01. The new model makes the input context window much longer: 4 million tokens can be processed at one time.
In fact, the technical barriers to long text processing are not high. The key to Kimi’s success in using long text as a selling point is that it does a good job in marketing.
While the professional vocabulary related to the big model confused the public, the dark side of the moon emphasized and explained the long text, forming a deep understanding in the hearts of users, and also firmly tied the label of the long text to Kimi. Even though other large models have surpassed Kimi in terms of long text capabilities, Kimi’s popularity last year has not been affected.
But technically speaking, this also means that long text capabilities are difficult to help Kimi build a moat. Looking around the world, the competition for long-text capabilities for overseas large models has long ended, because they have found that after the expansion of context reaches a certain limit, the overall capabilities of the model will be limited. Simply put, after the context expansion is sufficient, if you want to improve your reasoning ability, you still have to start in other directions.
Domestic enthusiasm for competition around long texts is likely to cool down. There is no other reason. DeepSeek’s performance in mathematics, code, natural language reasoning and other tasks is comparable to that of the official version of OpenAI o1. Especially its achievements in training costs have refreshed the understanding of the technology circle. Its existence is likely to change or affect the competition focus and main direction of future large models.
As one investor said, the future development of the AI industry depends not only on the improvement of computing power, but also on technological innovation and cost-effectiveness. As a result, the flow of funds may also change, from investments that simply pursue high-end computing power to companies and projects that focus more on technological innovation and cost control.
Whether it is technological innovation or cost control, the advantages of the dark side of the moon are not obvious. On the one hand, at the same time and also against OpenAI o1, the new Kimi k1.5 model launched by Dark Side of the Moon did not attract as attention as DeepSeek, which already illustrates the gap; on the other hand, training costs and high marketing costs make Dark Side of the Moon less persuasive in cost control.
In addition, there is another hidden danger in the dark side of the moon that has not yet been resolved, that is, the conflict between investors and entrepreneurs. At the end of last year, Moon Dark Side founder Yang Zhilin and co-founder and CTO Zhang Yutao were filed for arbitration in China Hong Kong by former investors during the company’s circular intelligence period. The relevant electronic arbitration application has been submitted to HKIAC (Hong Kong International Arbitration Center).
To become the darling of capital, we naturally have to bear the sweet burden brought by capital.
Kimi’s popularity has opened up the development potential of the C-end market for domestic large models, while DeepSeek’s popularity has provided greater opportunities for domestic large models to lead the world. Although the tide turns, this undoubtedly indicates the vigorous innovation of my country’s large model industry.
Of course, in this head-to-head battle between technology and capabilities, we cannot rush to draw conclusions on who loses and who wins.
