Ping An Securities: Kimi opens up a new world of big model applications and continues to be optimistic about the AIGC industry chain

The Zhitong Finance App learned that Ping An Securities released a research report saying that the emergence of Kimi's lossless long text model solved the pain points of many large models in application and opened up application space for big models. The final implementation of AIGC still requires finding the right scenario. As a 100 billion model, Kimi can support complex computation, and can also accept and process large texts, solving many problems in the actual application of large models, and the potential for subsequent commercialization will be highlighted. Currently, KimI's smart assistant has been launched on multi-terminal platforms such as Apple's iOS app, Android app, applet, and web. Continue to be optimistic about the AIGC industry chain, especially the application potential of the big model.

Incident: Recently, the domestic artificial intelligence company Moonshot AI (Moonshot AI) announced on its WeChat account “Moonshot AI” that the company's Kimi smart assistant has made a breakthrough in long context window technology, and the length of non-destructive context can reach 2 million words. Previously, in October 2023, the company's intelligent assistant was able to achieve 200,000 lossless context lengths, and the latest capabilities were raised by an order of magnitude.

Ping An Securities views the following:

The capabilities of Kimi's smart assistant have been greatly improved in a short period of time, and its popularity has increased rapidly in the country.

The length of the non-destructive context in this release reached 2 million words, which is only about 5 months apart from the previous 200,000 words. However, the Dark Side of the company Moon was only established in April 2023, and it took less than a year to establish it. Kimi Smart Assistant, also known as Kimi Chat, is a conversational AI assistant product created by Dark Side of the Moon based on a self-developed model with 100 billion parameters. It was officially launched on the market in November 2023. The product's strongest capability is long context processing, including long text summarization and generation, network search, data processing, coding, user interaction, and translation. After it went live, the popularity of this tool increased rapidly. According to Similarweb data, there has been a clear upward trend in Kimi's visits in recent weeks. According to the website's statistics, the number of visits in the last four weeks (2.20-2.26, 2.27-3.4, 3.5-3.11, 3.12-3.18) was 1.803 million, 1.128 million, 1.52 million, and 2.25 million, respectively. Although the company continues to expand its servers, the pressure is already showing in the face of rapid user growth.

The long text is expected to open up a new world of big model applications.

The number of parameters of a large model determines how complex “calculations” it can support, and how much text input (that is, long text technology) it can receive determines how much “memory” the large model has. Together, the two determine the application effect of the model. The current situation where the input length of large models is generally low has greatly limited the implementation of their technology. For example, virtual characters will “forget” some important information, agents may fail if they are unable to obtain full input information, and some game products are forced to simplify the plot due to their inability to process long text. Kimi's support for a longer context means that the big model has more “memory,” making the big model more in-depth and widely used. According to the company's official account, Kimi can analyze the market through multiple financial reports, handle extremely long legal contracts, quickly sort out key information from multiple articles or multiple web pages, and perform role-playing based on long-story settings. Large models will help users open up their imagination for AI application scenarios in the future, including analysis and understanding of a complete code base, intelligent agents that can independently complete multi-step complex tasks, lifelong assistants that do not forget key information, and multi-modal models with a truly unified architecture.

“Lossless compression” and increased text length are requirements that long text technology needs to balance.

The founder of the company said that if general artificial intelligence is to be realized, a non-destructive long context will be a critical basic technology. All of the model architecture evolutions in history have essentially increased the effective and non-destructive length of context. In the process of improving context length, it is necessary to balance the two indicators of length and lossless compression level in order to achieve meaningful scaling. Judging from the interval of this upgrade, the time is very short. It can be seen that the company did not follow a gradual iterative path; of course, it should also face greater technical difficulties. The company's R&D and technical team has carried out native redesign and development from model pre-training to alignment and inference. Under 100 billion parameters, a non-destructive long-term attention mechanism has been achieved. It does not rely on “shortcut” solutions such as sliding windows, downsampling, small models, etc., which greatly impair performance, taking into account the two indicators of length and “lossless”.

Aspect of the target:

1) In terms of computing power, we recommend Wave Information (000977.SZ), Zhongke Shuguang (603019.SH), Ziguang Co., Ltd. (000938.SZ), etc., and it is recommended to focus on IFF (601138.SH), Cambrian (688256.SH), Jing Jiawei (300474.SZ), high-tech development (000628.SZ), etc.;

2) In terms of algorithms, iFLYTEK (002230.SZ) is recommended;

3) In terms of application scenarios, we strongly recommend Zhongke Chuangda (300496.SZ), Hang Seng Electronics (600570.SH), Shengshi Technology (002990.SZ), etc.;

4) In terms of network security, I highly recommend Kai Ming Chen (002439.SZ).

Risk warning: 1) the risk that domestic computing power will not increase as fast as expected; 2) the risk of compliance regulations such as copyright; 3) the risk of technological evolution.