According to Shen Wan Hongyuan's evaluation, the large domestic model Kimi's text ability (ability to generate Chinese and English) has fully reached the GPT-4 level. Although there is still a gap in logical reasoning ability, and the main focus is text generation, and currently has no multi-modal ability.
The Zhitong Finance App learned that Shen Wan Hongyuan released a research report saying that on March 18, Dark Side of the Moon announced that its AI product, Kimi, has achieved a new breakthrough in large model long context window technology. Kimi's smart assistant already supports 2 million characters of ultra-long non-destructive context, and the product will begin internal testing from now on. According to Shen Wan Hongyuan's evaluation, the Chinese large model Kimi's text ability (Chinese and English generation ability) has fully reached the GPT-4 level, although there is still a gap in logical reasoning ability, and the main focus is text generation, and currently has no multi-modal ability; Cluade 3's ability to generate, understand, and reason in Chinese and English, and understand multi-modal images is close to GPT-4. The effect is better than Gemini, and the generation speed is faster than GPT-4 and Gemini in actual use.
Shen Wan Hongyuan's main views are as follows:
On March 18, Dark Side of the Moon announced its AI product, Kimi, which has achieved a new breakthrough in large model long context window technology. Kimi's smart assistant already supports 2 million characters of ultra-long non-destructive context, and the product's internal testing will begin today. Kimi chat is a conversation assistant tool launched by Dark Side of the Moon. It was released on October 10, 2023, and long texts were targeted at the beginning of the release.
It supports the input of 200,000 Chinese characters, which is currently the longest contextual input length supported in major domestic models. In February 2024, Kimi iterated on the website and multi-question search capabilities, and usability continues to improve. In March, the contextual capacity reached 2 million words.
After many iterations, along with the increase in Kimi's user activity, the average daily activity in February increased by 101.9% year-on-year, and continued to rise in the first two weeks of March.
Shen Wan Hongyuan believes that Kimi has achieved a breakthrough in her ability to single point long texts and accurately target the office crowd. Kimi supports long text input of 2 million Chinese characters. In comparison, the GPT-4 Turbo-128k is capable of about 100,000 Chinese characters, and the Claude3 200k context is about 160,000 Chinese characters. Therefore, Kimi is more suitable for efficient reading, professional document interpretation, data retrieval, data compilation and summary.
Kimi's Success Revelations:
Shen Wan Hongyuan believes that the team members' abilities, financial reserves, and time may be the reason for Kimi's current success.
1) The Dark Side of the Moon is led by Professor Yang Zhilin of the School of Interdisciplinary Informatics at Tsinghua University. The team members include talents from international tech giants such as Google, Meta, and Amazon, and participated in the development of various major models such as Gemini, Pangu NLP, and Wudao.
2) After its establishment, the company received investment from institutions such as Sequoia China and Zhenge Fund. The latest round of financing exceeded 1 billion US dollars. Investors include Ali, Sequoia China, Xiaohongshu, Meituan, etc., with a valuation of 2.5 billion US dollars.
3) Dark Side of the Moon was founded in March 2023. At this time, the full success of Chat GPT enabled the industry's big model to basically confirm the technical route of Decoder-Only+VQA, effectively avoiding the waste of development resources previously caused by differences in technical routes.
Up to now, the overall text generation capacity of large domestic models is close to GPT-4 Turbo. On January 30, the Shanghai Artificial Intelligence Laboratory released the large model open source open evaluation system Si Nan (OpenCompass 2.0). The results showed that many new models recently released by domestic manufacturers are rapidly narrowing the gap with GPT-4 Turbo in multiple performance dimensions, including Smart Spectrum Qingyan GLM-4, Alibaba Qwen-Max, and Baidu Wenxin 4.0.
Related targets:
The core focus is on text classes, applications that require long text capabilities such as PDF, Foxit Software (688095.SH), Jinshan Office (688111.SH), Novel Software (688590.SH), Flush (300033.SZ); vector databases: Starlink Technology (). 688031.SH
Other suggested concerns include: 1) multi-modal algorithm layout: iFLYTEK (002230.SZ), Hongsoft Technology (688088.SH), Wanxing Technology (300624.SZ), Dahua (002236.SZ); 2) multi-modal+ video: Shanghai Film (601595.SH), Huace Film and Television (300133.SZ), Optical Media (300251.SZ), Wanda Film (002739.SZ), Mango Supermedia (300413.SZ) (Film and TV IP) Communications ( 603322.SH) (Investing in AI Video Seven Volcanoes); 3) AI Big Model+Game: Giant Network (002558.SZ) (AI Detectives Require High Text Processing); 4) AI+ Marketing E-commerce: Easy World (301171.SZ), Focus Technology (002315.SZ); 5) AI+ Copyright: Visual China (000681.SZ).
Risk warning: There are still differences in big model technology between China and the US; LLM's commercial profitability still needs to be verified.