Per reporter: Liu Mingtao Per editor: Peng Shuiping
In May, the main A-share stock indexes adjusted in shock across the board, and the Shanghai index once approached 3400 points, but then peaked and fell, and the Shanghai index repaired slightly at the end of the month, and finally held 3200 points. The Brand 100 Index was also sluggish and pullback, losing 900 points. However, in June, the market showed a rebound signal, and the Brand 100 Index rose by 3.82% in the two trading days of June, leading the A-share rebound.
Positive signals from the market appeared
This week, the A-share market saw a volatile rebound, with the Shanghai Composite Index up 0.55%, the Shenzhen Component Index up 0.81%, and the Per-Jing Brand 100 Index performing strongly, up 2.2% for the week, leading the major A-share stock indexes. As of June 2, the Brand 100 Index closed at 888 points, with an overall average price-to-earnings ratio of 9.1 times, which is undervalued.
Since the beginning of this year, the market has been divided on the strength of the recovery, the strong PMI data in February-March once strengthened the market's expectations of the slope of economic recovery, and the subsequent price data, industrial enterprise profit data poured cold water on the market, with the May PMI data less than expected, the current market has "overkill" excessive pessimism.
For the current macro environment, weaker economic data has accelerated the rise in policy expectations, while the passage of the US Senate on the debt ceiling bill, along with the decline of the US dollar against the yuan, has also repaired the preference of global capital markets to some extent. From the perspective of the timing effect of policy, the mid-year Politburo meeting in July will be a key time point, and at the same time, as the Fed begins to enter the interest rate cutting cycle in the future, the linkage of domestic policies can also promote the global capital market to enter a resonance cycle.
On the other hand, the new products of public funds fell into the freezing point, which attracted market attention, and in sharp contrast to the new products of the fund fell to the freezing point, the investment cost performance of equity assets returned to the historical high area. At the same time, the overall valuation of the stock market is still low. The equity risk premium, an indicator that can better characterize the "cost performance" of equity asset investment, is gradually rising and approaching the extreme value area. Huafu Securities believes that the current point of time stock assets has a very high cost performance, the time and space for the subsequent pullback are relatively limited, and the probability of the market rebound is relatively large, especially for investors who are on the left and have a long holding period, the current is a better time to enter.
From the perspective of rhythm, the RMB exchange rate may be supported by the above easing factors, and there is hope to continue the rebound within the month, and the Hang Seng Technology Index is also expected to usher in an upward trend in the coming period. Structurally, Tencent, which has the advantage of large models, has the highest cost performance and certainty, in addition, Internet leaders such as Ali and Meituan will also benefit.
Chinese big model set sail
Due to the historical market of global artificial intelligence, algorithm computing power and the concept of large models in the first half of the year, major international technology giants have also set their sights on the Chinese market. The announcement of Huang Jenxun, a leading figure in the field of AI, to visit China and meet with executives of Tencent, ByteDance and other companies has greatly boosted market expectations for Chinese large models.
Image source: Visual China - VCG41N1188747468
It is understood that language modeling has entered the stage of large-scale language models with tens of billions and hundreds of billions of parameters through statistical language models, neurolinguistic models and pre-trained language models. The large language model has excellent context learning ability, considerable knowledge capacity, excellent generalization and complex reasoning ability. There are three technical lines of Google's Bert model, OpenAI's GPT model and domestic hybrid mode in large model research, with ChatGPT as the dominant GPT model.
At present, at least 19 domestic enterprises and research institutes have participated in AI large model training, mainly divided into three categories: large technology companies, scientific research institutes and start-up technology teams. In international comparison, the United States is currently leading and China is following, but the gap is narrowing. In practice, the pre-trained large model completes the "general knowledge" education in the self-supervised learning stage based on massive data, and then uses modes such as "pre-training + fine tuning" to fine-tune the task with a small amount of data according to the characteristics of the specific application scenario under the condition of sharing parameters.
However, the landing of the global large-model industry is still in the early exploration stage, and it is necessary to cooperate with downstream scene enterprises to establish a large-model business model, but the current understanding of the large-model of downstream enterprises is relatively limited, and the resource support required is relatively weak. In general, the PaaS model of large model landing can be paid through API interface calls, or the large manufacturer can provide the development tools, cloud platforms, services, etc. required for implementation, and further provide the MaaS model of directly providing relevant customized model calls.
Ping An Securities pointed out that with the continued popularity of the concept of artificial intelligence in US stocks, the artificial intelligence sector of domestic A-share Hong Kong stocks has also been affected by the external influence. Tencent and Alibaba, as absolute technology giants and leading companies in the field of artificial intelligence in Hong Kong stocks, still have a big gap in market value compared with most international technology giants in the United States, and there is still huge room for development in the field of domestic artificial intelligence large models.
Internet giants accelerate their layout
From the perspective of this week's market performance, "Zhongte Valuation" and Internet technology performance is relatively strong, and the constituent stocks of the Brand 100 Index include the four Internet giants of "BATJ", of which Baidu Group rose 12.02% weekly, rebounding momentum, JD.com and Ali also rose by about 5%, and Tencent Holdings still had a weekly increase of 3.66% although its performance was weaker.
It is worth noting that in addition to JD.com, which is mainly e-commerce, the other three Internet giants are all in a four-in-one layout of "model, computing power, platform and application".
For example, Alibaba is at the model level, and Ali Damo Academy leads the research of large models, releasing Tongyi large models, including Tongyi M6 multimodal model, Tongyi AliceMind natural language processing model, and Tongyi vision computer vision model. At the computing power layer, Alibaba Cloud launched the full-stack intelligent computing solution "Feitian Intelligent Computing Platform" and launched two ultra-large-scale intelligent computing centers.
Alibaba not only has the most NVIDIA A100 chips, but also has self-developed chips Hanguang 800 and Yitian 710 chips, which jointly provide computing power for artificial intelligence large models. At the platform layer, Alibaba launched the AI open-source community "Modai", aiming to build a next-generation "model-as-a-service" sharing platform, integrate multiple model providers in the industry, and provide developers with pre-trained basic models and API interfaces. At the application layer, Alibaba has announced that all its products will be completely transformed and integrated into the Tongyi model in the future, including Tmall, DingTalk, Amap, Taobao, Youku, Freshippo, etc.
Baidu's Wenxin big model builds a three-level model of "foundation + task + industry"; At the computing power layer, the computing power infrastructure is supported by Baidu Intelligent Computing Center, and the second generation of Kunlun Core has been widely used in the Wenxin large model, and it is expected that the third generation of Kunlun Core will be mass-produced in early 2024; At the platform layer, Baidu PaddlePaddle is its self-developed deep learning platform, and Wenxin Big Model realizes model training, inference deployment and scenario application through Baidu Flying Paddle platform. At the application layer, the Wenxin model will be applied in many industries such as electric power, gas, media, cities, film and television, manufacturing, and social sciences.
Tencent released a new generation of HCC high-performance computing cluster by Tencent Cloud, which adopts the latest generation of Xingxinghai self-developed server, equipped with NVIDIAH800 Tensor Core GPU, and based on self-developed network and storage architecture, brings 3.2T ultra-high interconnection bandwidth terabyte-level throughput capacity and tens of millions of IOPS. Functions such as sentiment analysis, story generation, and character shaping, as well as Tencent's hybrid model, are expected to develop ecosystems in gaming, social networking, finance, education, medical and other fields.
Daily economic news