Tianfeng Securities released a research report, said the overseas big language model into a new round of big model technology breakthrough period, the domestic also have broken through the GPT3.5 level. Multi-modal generation of large models emerge in endlessly, and SORA leads a new wave of large model innovation. Apart from the innovation on the large model side, a large number of applications have been born on the overseas application side under the leadership of GPTs. On the computing side, hardware investment from major overseas factories continues. Meta is expected to continue to expand GPU investment in 2024. Microsoft, google and Amazon all expect capital expenditure to continue to expand in the new year to support AI investment. Considering the bright performance of model capability and computing power expenditure at home and abroad, the opportunity of AI application and computing power plate is recommended here.

The main points of view of Tianfeng Securities are as follows:

Overseas big language models have entered a new round of big model technology breakthrough period, and domestic countries have also broken through GPT3.5 level

Large language models at home and abroad have entered a new round of breakthrough. Overseas, GPT-4V the launch of OpenAI, multimodal capabilities have been strengthened. Google has launched a new round of large language model Gemini, of which the Ultra model is better than GPT4 in the benchmark test of text processing. On February 15, Google released a new version 1.5, which continues to improve its performance compared with 1.0. Meta recently announced that it is developing LLaMA3 in architectural computing power and insists on open source. Anthropic's Claude model has also evolved to version 3, with slightly GPT-4 overall performance, continue to maintain long text performance and security features outstanding, large model Mixtral through the MOE structure better quality and efficiency. During the same period, the new versions in China, including Zhipu, Wenxin Yanyan, HKUST Xunfei and Tongyi Qianwen, have all reached the level of catching up with GPT3.5 and even some of their capabilities close to GPT4.

Multi-modal generation of large models emerges one after another, and SORA leads a new wave of large model innovation

the multi-modal generation model has entered a technological breakthrough period. OpenAI released the video generation model SORA, which adopts the Diffusion Transformer structure and uses the space-time Latent patch to represent videos and images, or becomes the basis of simulation reality. Recently, Stability.ai has opened source Stable Video Diffusion model, Google has released VideoPoet, and new video generation technologies have emerged continuously. The Wensheng graph model is also iterated step by step, Midjourney the V6 version is released, with better picture generation capability; in addition, in the field of digital people, Microsoft launched the GAIA model and Alibaba launched the Animate Anyone, which the bank believes lays the technical foundation for vertical business scenarios.

Application and computing power go hand in hand, and Andy Beale's Law continues to be interpreted overseas

except for the innovation on the large model side, a large number of applications have been born on the overseas application side under the leadership of GPTs, and more than 3 million applications have been created within 2 months. OpenAI has also introduced an ecosystem for this purpose. On the computing side, the hardware investment of overseas major factories continues, and Meta is expected to continue to expand GPU investment in 2024, microsoft, google and Amazon all expect capital spending to continue to expand in the new year to support AI investment.

considering the bright performance of model ability and calculation power expenditure at home and abroad, we recommend the opportunities of AI application and calculation power section here. it is recommended to pay attention to:

application:(1) office software: Jinshan office (688111.SH), fuxin software (688095.SH), color news shares (300634.SZ) (communication team coverage);

(2) Multimodal: Wanxing Technology (300624.SZ), Meitu Company (01357)(jointly covered with overseas teams), Hongsoft Technology (688088.SH), Guangyun Technology (688365.SH);

(3)B-side applications: Yonyou Network (600588.SH), Kingdee International (00268), Zhiyuan Internet (688369.SH), Pan-Micro Network (603039.SH), Dingjie Software (300378.SZ), HAND Information (300170.SZ);

(4) finance, education, medical care: flush (300033.SZ), hang seng electronics (600570.SH), new software (688590.SH), hkust xunfei (300624.SZ), visual source shares (002841.SZ) (jointly covered with the electronic group), runda medical (603108.SH).

infrastructure: digital China (000034.SZ), beacon communication (600498.SH), tuowei information (002261.SZ), high-tech development (000628.SZ), sea light information (688041.SH), star ring technology (688031.SH), cambrian (688256.SH), jingjiawei (300474.SZ) (jointly covered with electronic team).

Risk Warning: The effect of domestic large models is not as good as expected, the supply of domestic computing power is not as good as expected, and the domestic application scenario is not as good as expected.