Tianfeng Securities released a research report saying that overseas big language models have entered a new round of big model technology breakthroughs, and domestic ones have also broken through the GPT 3.5 level.
The Zhitong Finance App learned that Tianfeng Securities released a research report saying that overseas big language models have entered a new round of big model technology breakthroughs, and domestic ones have also broken through the GPT 3.5 level. Large models of multi-modal generation are constantly emerging, and SORA is leading a new wave of large-scale model innovation. In addition to innovation on the big model side, a large number of applications have been born on the overseas application side under the leadership of GPTS; on the computing power side, hardware investment by major overseas manufacturers continues, Meta is expected to continue to expand GPU investment in 2024, and Microsoft, Google, and Amazon all expect capital expenditure to continue to expand in the new year to support investment in AI. Considering the outstanding performance in model capabilities and computing power expenditure at home and abroad, I recommend AI applications and opportunities in the computing power sector.
Tianfeng Securities's main views are as follows:
Overseas big language models have entered a new round of big model technology breakthroughs, and domestic ones have also broken through the GPT 3.5 level
Major language models at home and abroad have entered a new round of breakthroughs. Overseas, OpenAI launched GPT-4v, enhanced multi-modal capabilities, and Google launched a new round of large-scale language model Gemini. Among them, the Ultra model was superior to GPT4 in the text processing benchmark. On February 15, Google released 1.5 and continued to improve performance compared to 1.0; Meta recently announced that it is developing LLAMA3 for architectural computing power and adhering to open source; Anthropic's Claude model has also evolved to version 3. The overall performance is slightly superior to GPT-4, and continues to maintain long text performance and The safety characteristics are outstanding, and the large model Mixtral enhances quality and efficiency through a better MoE structure. Over the same period, new versions in China, including Smart Spectrum, iFLYTEK, and Tongyi Qianwen, all reached a level where they could catch up with GPT 3.5 or even close to GPT4 in some cases.
Multi-modal generation of large models is constantly emerging, and SORA is leading a new wave of big model innovation
The large multi-modal generation model has entered a technological breakthrough period. OpenAI released the video generation model Sora, which uses a diffusion transformer structure to represent videos and images, or become the basis for simulating reality; recently, the stability.ai open source Stable Video Diffusion model, Google released VideoPoet, and new video generation technology continues to emerge; the Wenshengdu model is also gradually iterating, Midjourney With the launch of the V6 version, the image generation ability is better; in the digital human field, Microsoft launched the GAIA big model and Alibaba launched Animate Anyone. The bank believes this has laid 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
In addition to innovation on the big model side, a large number of applications were born on the overseas application side under the leadership of GPTS. More than 3 million applications were created within 2 months, and OpenAI also introduced the ecosystem for this purpose; on the computing power side, hardware investment by major overseas manufacturers continues, and 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 outstanding performance in model power and computing power expenditure at home and abroad, I would like to recommend AI applications and opportunities in the computing power sector, and it is recommended to pay attention to:
Application: (1) Office software: Jinshan Office (688111.SH), Foxit Software (688095.SH), Caixun Co., Ltd. (300634.SZ) (communication team coverage);
(2) Multi-modality: Wanxing Technology (300624.SZ), Meitu (01357) (joint coverage with overseas teams), ArcSoft Technology (688088.SH), Guangyun Technology (688365.SH);
(3) B-side applications: User Network (600588.SH), Kingdee International (00268), Zhiyuan Internet (688369.SH), Pan-Micro Network (603039.SH), Dingjie Software (300378.SZ), Hande Information (300170.SZ);
(4) Finance, education, healthcare: Tonghua Shun (300033.SZ), Hang Seng Electronics (600570.SH), Xinzhi Software (688590.SH), iFLYTEK (300624.SZ), ViewSource (002841.SZ) (joint coverage with Electronics Group), Runda Healthcare (603108.SH).
Infrastructure: Shenzhou Digital (000034.SZ), Firefight Communications (600498.SH), Topway Information (002261.SZ), Hi-Tech Development (000628.SZ), Haiguang Information (688041.SH), Starlink Technology (688031.SH), Cambrian (), Jing Jiawei (300474.SZ) (joint coverage with the electronics team). 688256.SH
Risk warning: The improvement in the effectiveness of large domestic models falls short of expectations, the supply of domestic computing power falls short of expectations, and the implementation of domestic application scenarios falls short of expectations.