The big factory wants face, Kimi wants the inside
Source: Phoenix Technology Author: March 27, 2024 14:57
share
When Internet giants were still obsessed with showing their skills in the "100 model war" of AI, no one expected that they would be cheated by a "newcomer".
Kimi's explosion came too suddenly and unexpectedly.
How hot is Kimi? Downtime has become a common occurrence in recent times, to the extent that officials have had to apologize.
The big factory wants face, Kimi wants the inside
Speaking of AIGC, talking about Kimi has become a daily routine for netizens recently. Many people have found that their social circles and WeChat groups are almost recommending Kimi, with "Kimi can be found almost every day in Weibo hot searches these days" and "If you haven't used Kimi yet, then you're out"
"Previously, I used ERNIE Bot all the time. One day, my friend recommended Kimi, and I didn't know what a big domestic model was." Li Lin, a post-80s character worker, whose demand for big models mainly focused on data retrieval and sorting, was able to basically meet her demand before ERNIE Bot.
But Kimi gave Li Lin a new experience. "From the perspective of data and material sorting, Kimi's sense of experience is obviously better than ERNIE Bot's."
Wang Ming is a senior AI practitioner who is not surprised by this. He witnessed Kimi's process from internal testing to explosive success. "It is obvious that Kimi, who comes from a grassroots background, knows more about what users truly need."
"Kimi still has some shortcomings, but based on the current user experience, it has already taken the lead in some large factories and models." Wang Ming frankly stated that among the domestic large models, he is more optimistic about Kimi's future.
The capital market has also shown great enthusiasm for Kimi. Since mid March, the A-share market has given birth to Kimi concept stocks, which have affected the stock prices of many companies and led to consecutive sharp increases.
Wanxing Technology is one of the AIGC software companies that connects to Kimi, and its video creative software, Wanxing Meow Shadow, connects to Kimi. In the view of Qi Borequan, General Manager of Wanxing Technology's AI Innovation Center, Kimi's debut not only reflects the public's interest and expectations for emerging technologies, but also reflects the confidence and pursuit of AI technology in the capital market. "This is also a positive signal, demonstrating the potential and commercial value of AI applications."
"Kimi's rise to fame was not achieved overnight, and behind it was a large amount of research and development investment and continuous technological innovation," Qi Boring quan told earpiece Tech.
However, can Kimi continue its current popularity in the future? At present, it is clearly impossible to determine. In fact, with Kimi's popularity, there have been constant doubts about its technological processing capabilities and commercialization path.
Qi Borequan also pointed out the risks, "With the widespread application of AI technology, we should also pay attention to the potential data security and privacy protection issues it may bring, and actively seek solutions."
How did Kimi perform?
"After using Kimi once, most people now choose to use it." Many people, like Li Lin, have been using Kimi recently. Securities analyst Zhang Qiang, who needs to read a lot of data and reports for a long time, has also been using Kimi. In his opinion, Kimi has a more user-friendly interface and is more in line with his own needs in the large model of literature and literature.
Zhang Qiang has obtained Kimi's qualification for the 2 million word long article input internal test. He often inputs the full text of annual reports or IPO documents of some listed companies to Kimi. He told "Listening Tech" that Kimi can quickly extract core content, including basic information of listed companies, financial overview, corporate governance, and other core content. "Relatively speaking, other domestic large models lack in summarizing and summarizing."
What is Kimi's user experience really like? Handset Tech conducted a series of small tests on Kimi, ERNIE Bot, Tongyi Qianwen, Doubao, and Tencent's hybrid assistants.
The first test is the ability of these large models to summarize data.
It should be noted that in the PC versions of several big models, Kimi, Tongyi Qianwen, Doubao and ERNIE Bot are friendly to the file upload function, while the PC side of Tencent's hybrid assistant is not very friendly to the file upload interface. After several twists and turns, Handset Tech failed to find a way to upload files to the conversation.
Tech randomly selected the latest Meituan 2023 Financial Report for testing.
From the test results, as Zhang Qiang said, Kimi's ability to summarize and summarize documents is significantly better than other major models. Kimi's summary of this financial report is clear and concise. Although Doubao and Tongyi Qianwen can also be summarized, the results are not satisfactory, and ERNIE Bot needs further instructions.
Subsequently, the commonly used data retrieval and organization functions of users were tested. The earpiece Tech conducted a test on various models to read Huang Renxun's speech at the GTC conference.
From the search results of the large model, Kimi accurately and concisely summarized the content of Huang Renxun's speech at the 2024 GTC conference.
Except for Kimi, several other large models were unable to retrieve Huang Renxun's speech content at the GTC conference, and Tencent's hybrid assistant could not even generate any answers, requiring further instructions.
Creative ability is one of the important abilities that every large model must possess, which can better understand and simulate human thinking processes, thereby generating more creative and valuable content.
The generation of video scripts is one of the commonly used literary creation ability models for users. To this end, "Earphone Tech" tested the video script production ability of large models.
The content of this test is to generate video scripts with the theme of 'Sports Change Life'. From the generation results, Kimi, ERNIE Bot, Tongyi Qianwen and Doubao can all generate video script scripts that meet the basic requirements.
The video script generated from Doubao includes elements such as duration and location, making it relatively professional. The script generated by Kimi and ERNIE Bot is more coherent and organized. And Tencent's hybrid assistant even confused the previous issue with the script.
From the test results, Kimi, ERNIE Bot, Tongyi Qianwen and Doubao all have strong video script generation capabilities, which can help video production. In contrast, Tencent's hybrid assistant clearly lacks sufficient understanding of the problem.
From the above simple comprehensive test results, it can be seen that in fact, Kimi's processing results are more in line with the needs of basic users in addition to the basic functions of text.
For basic users like Li Lin and Zhang Qiang, Kimi's current basic functions of text and text are far superior to other domestic large models.
"Of course, they also look forward to Kimi solving more problems," Li Lin said with a smile. "I'm already looking forward to it helping me with the PPT."
Why is Kimi the one who made a comeback?
Who the hell is Kimi?
According to public reports, Kimi's parent company is Beijing Moon Dark Face Technology Co., Ltd. (hereinafter referred to as "Moon Dark Face"), founded by Yang Zhilin.
According to data from Qichacha, the dark side of the moon was established in March 2023 and launched the world's first intelligent assistant product Kimi that supports inputting 200000 Chinese characters in October of the same year.
In just one year since its establishment, the dark side of the moon has received two rounds of financing from well-known institutions and enterprises such as Sequoia China, ZhenGe Fund, Alibaba, Xiaohongshu, and Meituan, with a financing amount of over 1.2 billion US dollars. The post investment valuation has reached about 2.5 billion US dollars.
At the beginning of this year, Kimi's traffic began to rise. According to data from Similarweb and Qimai, from February 18 to March 16, 2024, Kimi had an average daily view count of nearly 200000, and its cumulative download count across the entire platform was 500000.
Especially in the past two weeks, Kimi's traffic has exploded. According to Similarweb data, Kimi's traffic in the past two weeks has been 1.52 million and 2.25 million, respectively, which has caused abnormal platform access.
Why is Kimi the most popular among many domestic large models?
In Wang Ming's view, on the one hand, it is closely related to the background of his founding team. "From publicly available information, the dark side of the moon can be considered a team that has gathered relatively leading talents in the field of large models in China."
According to public information, Yang Zhilin was born in 1992 and is hailed as the youngest founder of China's Big Model. He graduated from the Department of Computer Science at Tsinghua University with a bachelor's degree and studied under Professor Tang Jie, a renowned AI scholar at Tsinghua University. I graduated with a PhD from the School of Computer Science at Carnegie Mellon University and studied under Rualan Salakhutdinov, the head of AI research at Apple, and William Cohen, the chief scientist at Google.
According to public reports, Yang Zhilin is currently an assistant professor at the Cross Information Research Institute of Tsinghua University (hereinafter referred to as the "Cross Information Institute"). You should know that the Cross Information Institute covers the famous "Yao Class" of Tsinghua University. Yang Zhilin has collaborated on scientific research with Turing Award winners Yann LeCun and Yoshua Bengio.
In fact, the Dark Side of the Moon is Yang Zhilin's second entrepreneurship. Prior to this, he co founded Circular Intelligence with others and also received investment from Sequoia China. In 2021, he and Huawei Cloud jointly launched the world's largest Chinese language model, "Pangu".
According to media reports, the core team members of the Moon's Dark Side have made important inventions in the field of large models, such as RoPE relative position encoding and group normalization, which are important components of mainstream models such as Meta LLaMa and Google PALM. Its other two founders, Zhou Xinyu and Wu Yuxin, are also leading technical talents in the AIGC field and have been cited by over 10000 Google Scholars.
Liam, a research scholar in the field of AI who has an academic intersection with Yang Zhilin, believes that the Moon's Dark Side team can be considered one of the leading AGI teams in China. He also acknowledges the industry's evaluation of Yang Zhilin as a "steadfast AGI believer and a founder with technological appeal".
Liam was not surprised by Kimi's rise to fame. "In fact, people in the AI technology industry have been determined since several companies announced their AGI plans last year that the dark side of the moon and Zhipu are the most promising. Therefore, the top capital will immediately squeeze into these companies."
In Liam's view, Yang Zhilin is one of the "few scholars who are good at thinking from first principles". "Kimi's long context technical path is different from that of big companies, perhaps also due to his thinking on first principles."
On the other hand, Yang Zhilin also has a clear understanding of what users truly need as a big model. Liam revealed to "Earphone Tech" that in Yang Zhilin's view, a good product should know what users want and optimize it by meeting their needs. "This is also why the product launched by the dark side of the moon was TO C, not TO B."
In Liam's view, "The positioning of TO C can bring Kimi more opportunities for 'training', which is more conducive to product optimization and improvement."
The market clearly recognizes Yang Zhilin's positioning of Kimi. According to an analysis report by CITIC Securities, the significant growth of Kimi's daily active users reflects Kimi's successful strategies in model optimization, talent expansion, and user attraction.
"Kimi's success not only depends on its technological advantages, but also on its emphasis on user experience, including continuous data-driven product optimization, innovative sharing mechanisms, and precise polishing of core functions. These factors collectively enhance Kimi's market competitiveness."
CITIC Construction Investment also stated in its research report, "The dark side of the moon is creating a high attention application called Kimi Chat. On the one hand, it is because the core team has a deep technical background, and on the other hand, the product is open to the C-end for free, focusing on product operation."
In Qi Borequan's view, what users need is a large model that can solve 80% of the problems in specific scenarios, a localized large model that combines local user usage habits, and a large model that can interact and co create with users.
"A big model will always be 'application is king'. What a big model needs to provide is not just a superficial tool, but a standardized process support, including a basic big model, a complete toolchain, rich applications, and services supported by experts."
These may be the reasons why Kimi emerged first in the industry.
Another reason why Kimi quickly gained popularity is that it was unusually low-key in the early stages. Almost no one had heard of the name of this company before, which is in stark contrast to the high-profile approach of big companies in the field of AI.
If a wealthy company values face more, then Kimi, who comes from a grassroots background, pays more attention to the inside.
How long can Kimi become popular?
Can the explosive Kimi stand out in the new battlefield of the big model as expected? At present, there is no consensus.
Judging solely from Kimi's proud long text processing technology, Kimi may face enormous pressure.
An undeniable fact is that after the Kimi model went viral, major manufacturers are quickly following suit.
On March 22nd, Alibaba Tongyi Qianwen announced the free opening of a 10 million word document processing function, becoming the world's largest AI application in document processing capacity. Subsequently, 360 announced that Zhinao officially conducted an internal test of its 5 million word long text processing function, which entered the 360AI browser. According to media reports, Baidu will also offer free access to the 2 to 5 million long text function.
In the eyes of an AI technician, "Long text processing is not a difficult technology, but the focus of previous big models was not on this aspect. More importantly, long text processing is relatively expensive, and can even be considered a 'loss making business'."
The AI technician believes that "when the market realizes that this technology can quickly open up the market, it will inevitably join this battlefield, which will undoubtedly put tremendous pressure on Kimi."
In fact, a simple test can show that the AI technician's viewpoint has some validity.
As Kimi and Tongyi Qianwen have not yet obtained the qualification for the 2 million word internal test, earpiece Tech submitted a nearly 350000 word "Tea Hundred Road Post Hearing Dataset" to Kimi and Tongyi Qianwen respectively. After issuing the same instructions, Kimi gave feedback that "they exceeded the word limit and only read the top 31%", while Tongyi Qianwen successfully summarized the entire text.
To a certain extent, the opening of the 10 million word document processing function in Tongyi Qianwen, although the summarized content is not satisfactory, yields better intuitive results than Kimi.
However, Liam has a different view on the AI technicians mentioned above. Liam believes that there is a misunderstanding in the statement that "long text processing is not difficult". To be precise, it should be "it is not difficult to make the text longer". However, there is a significant technical difficulty in "not losing useful information while making the text longer, ensuring that the model still has a deep understanding of the text".
In Liam's view, ensuring controllable computational load and cost when text becomes longer or even infinitely long requires a large amount of basic research support, and the talent density of large companies is not sufficient to support such research.
On the other hand, from the multiple crashes, it also means that there are still many issues that Kimi's backend needs to solve.
The above-mentioned AI technicians stated that although the dark side of the moon has indeed gathered some leading technical talents, it still needs to be strengthened. "I have seen media reports that the Kimi team currently has less than a hundred people, which is far from enough for a large model enterprise in its upward trend."
However, in Liam's view, talent density is the most critical factor for top technology companies, rather than talent quantity.
"The answer can be drawn from OpenAI. When ChatGPT was released, OpenAI had only over 100 people, all of whom were the top scientists and engineers in the field. Of course, after the popularity of GPT, OpenAI began to vigorously expand its recruitment, resulting in a decrease in talent density and causing some problems."
In addition, in order for Kimi to continue its development, it must face the challenge of commercialization.
At present, the specific commercialization path of the dark side of the moon has not been publicly disclosed. When Kimi became popular, some media reported that the person in charge of the dark side of the moon mentioned that there would be a preliminary plan for commercialization within the year.
Industry insiders suggest that Kimi's commercialization may be similar to OpenAI, with a preference for a universal commercialization approach, such as a commercial high-end access point used to expand customer applications.
However, it is unknown whether this model can be implemented and whether it can support Kimi's long-term development.
In Wang Ming's view, if the current customer acquisition cost of Kimi is like the media reports that "the daily investment may exceed 200000 yuan", it is obviously not a small amount, and the current business model is not enough to support its rapid development in the future.
More importantly, after Kimi became popular, it further promoted the escalation of the "Hundred Model Battle", and domestic and foreign AI companies have formed a trend of hunting for it.
Not to mention foreign countries, only domestic giants such as Tencent, Alibaba, Baidu, iFlytek, and Shangtang Technology, as well as AI companies, have all launched large models. At the same time, various industries are constantly emerging vertical large models.
Public data shows that currently, the number of large models published in China has exceeded 200. "For Kimi, these are all enormous pressures," Wang Ming said bluntly.
In a previous interview with the media, Yang Zhilin stated, "AI is not about finding a PMF (Product/Market Fit) in the next year or two, but about how it will change the world in the next ten to twenty years."
This may not be the vision of Yang Zhilin alone, but also the future that many domestic model practitioners hope for.
(Li Lin, Wang Ming, Zhang Qiang, and Liam are all pseudonyms in the text.)
(The article is the author's independent viewpoint and does not represent the position of iResearch. com)