The Paper reporter Tian Zhongfang Sun Yan
ChatGPT, this AI chat robot software produced by OpenAI, the "American AI Dream Factory", can not only write codes, scripts, and compose lyrics, but also communicate with humans fluently. After its launch, it has been hailed as a masterpiece of artificial intelligence.
Recently, news of the deployment of ChatGPT has been frequently reported by many technology giants at home and abroad.
For example, the American digital media company BuzzFeed announced a partnership with OpenAI and will use ChatGPT to help with content creation; Microsoft also announced that it has expanded its partnership with OpenAI and made a multi-year, multi-billion-dollar investment in OpenAI.
Domestically, market sources say that Baidu will also launch an AI chatbot similar to ChatGPT.
With the explosion of ChatGPT out of the circle, the A-share market, which has never been left behind in the attention of popular subject stocks, is also moving with the wind, and the stock prices of concept stocks are rising steadily.
But behind the rapid rise of ChatGPT, is this emerging object an innovation or a concept?
What about the upstream and downstream distribution of the industrial chain?
The Paper excavated the core targets through interviews and sorting out.
What is ChatGPT?
It is understood that the full name of ChatGPT is Chat Generative Pre-Trained Transformer, which is an intelligent chat robot developed by OpenAI and will be launched in November 2022.
ChatGPT uses a large language model (LLM) based on the GPT-3.5 architecture, which can complete various tasks such as automatic text generation, automatic question answering, writing and debugging computer programs.
According to the results of the test feedback, compared with the previous generation of products, ChatGPT uses dialogue as the carrier, can answer a variety of daily questions, and enhance the memory ability and length of multiple rounds of dialogue history, and the interaction is more "empathetic". .
The UBS report pointed out that ChatGPT is the fastest growing consumer application in history.
Only two months after its launch, ChatGPT's monthly active users exceeded 100 million by the end of January 2023.
Regarding the popularity of ChatGPT, especially why it has attracted strong attention from the capital market, Zhao Fengfei, fund manager of Great Wall Fund, said: "ChatGPT's popularity is due to its amazing user experience. Changing our lives, so every major breakthrough in this field will attract widespread attention from the whole society, and it should be said that it is also a matter of course to set off a rising wave of related concept stocks in the A-share market.”
"Due to the introduction of human feedback reinforcement learning (using the interlocutor's feedback as the input of the next round of answers), ChatGPT's answer performance is better than the traditional natural language processing (NLP) large model that cannot be connected with the context. One of the main reasons for widespread concern.” Sinolink Securities pointed out.
However, Zheshang Securities emphasized that although ChatGPT has a high "IQ", it also has certain limitations, such as sometimes generating false and obviously wrong replies.
At the same time, although ChatGPT has legal and moral constraints, there are still ways to crack it.
Upstream: demand for computing power, data labeling, natural language processing, and artificial intelligence-generated content (AIGC) is expected to be driven
In terms of the upstream of the industry chain, on the whole, many institutions are optimistic about the new era of AI opened by ChatGPT, and further increase the upstream demand in five aspects of computing power, algorithms, data annotation, natural language processing, and artificial intelligence-generated content.
Specifically, in terms of computing power, the Bank of China Securities report pointed out that the parameter scale of ChatGPT has reached the level of 100 billion. For large AI models, whether it is training or reasoning, it has a great impact on the infrastructure of computing power. need.
"ChatGPT is mainly derived from the pre-training model of GPT-3, and the biggest difference between GPT-3 and the previous model is that the size of the model has been greatly improved. The leapfrog development of artificial intelligence will become an important driver of computing power consumption In the future, data centers and related supporting industries are expected to achieve better growth.” The Essence Securities report further pointed out.
CICC data shows that the computing power consumed by ChatGPT training is about 3640 PF-days, that is, assuming 1 quadrillion calculations per second, it needs to run continuously for 3640 days, and training large models requires powerful computing power.
In terms of data labeling, Essence Securities analyzed that the training process of ChatGPT has increased the intensity and accuracy of manual labeling, which means that in the future artificial intelligence field, high-quality data sources and powerful labeling capabilities will become the industry's infrastructure.
"Data labeling is the key link for the effective operation of most artificial intelligence algorithms. Among the three stages of ChatGPT training, only the third stage does not need to use manual labeling data, while the first and second stages require a lot of manual labeling. Soochow Securities further pointed out, "The existing data labeling is mainly manual labeling, which is a labor-intensive industry. With the continuous improvement of machine learning, automatic labeling has become a major trend."
In terms of NLP (Natural Language Processing), Essence Securities stated that since ChatGPT is mainly based on natural language processing, companies that have accumulated more in the NLP field are expected to be the first to realize partial reproduction of functions.
"The launch of ChatGPT has played a strong demonstration effect on the development of my country's AI technology and industry, indicating the rapid development of international frontier AI technology, and the commercialization of AI is becoming clearer. With the rapid development of AI , Artificial intelligence technology providers, especially NLP head manufacturers will be the first to benefit.” Founder Securities said.
In terms of AIGC (artificial intelligence generated content), Soochow Securities pointed out that ChatGPT is another starting point for the application of AIGC. With the continuous improvement of deep learning models, the promotion of open source models, and the possibility of commercialization of large models, AIGC is expected to accelerate its development.
"ChatGPT is a milestone in the development of AIGC technology. This model has greatly improved the technical maturity of content creation using artificial intelligence. It is expected to become a new industry-wide productivity tool and improve the efficiency and richness of content production." Caitong Securities also stated.
It is worth noting that the upstream industries of ChatGPT also promote each other.
The Tebon Securities report pointed out that AIGC also needs the support of "big computing power".
"OpenAI published research results showing that the computing power index required for AI training has increased exponentially. From 2012 to 2018, it has doubled approximately every 3 to 4 months. During this period, the computing power consumed by training AI has increased. increased by 300,000 times." Debon Securities said.
The Orient Securities report pointed out that from the perspective of industry development trends, the penetration rate of AIGC is expected to increase to 10% in 2025, and the market size may exceed one trillion in 2030.
Downstream: intelligent customer service, search engine, image text code generation and other application scenarios
For the downstream application scenarios of ChatGPT, many institutions believe that it is very extensive.
The surging news reporters showed that the application direction that institutions are unanimously optimistic about is mainly concentrated in three aspects: intelligent customer service, search engine, image, text, and code generation.
In terms of intelligent customer service, CICC reported that when ChatGPT is put into the commercial field in the future, it can optimize customer service robots such as banks and e-commerce through enhanced context understanding capabilities, and greatly improve service quality.
"The most suitable project for ChatGPT to be implemented directly is the work of intelligent customer service." The CITIC Securities report pointed out that according to the existing degree of completion of the model, targeted manual feedback training in vertical industries, ChatGPT can be implemented as an intelligent customer service product. The first application in the toC scene.
Compared with the current intelligent customer service, the customer service supported by ChatGPT will have significant progress in terms of flexibility and humanized service.
In terms of search engines, ChatGPT has brought a big discussion on whether the industry will usher in "revolutionary" changes and whether it can replace traditional search engines.
CITIC Securities pointed out that ChatGPT can directly generate high-level answers to questions, and can provide relevant reference links to the content of the answers (this function has not been developed in the current test version).
In addition, for open-ended questions, ChatGPT can also generate relatively complete answers by matching data in the network. When dealing with knowledge and creative questions, ChatGPT provides a search experience that is far better than current traditional search engines.
"However, although ChatGPT can greatly optimize the user's search experience, it still faces three key technical bottlenecks in order to replace traditional search engines, so the possibility of replacement in the short and medium term is low." CITIC Securities analysis said: First, the data Second, the authenticity of the data is still not reliable enough, and third, the cost of online inference of the model is high.
In terms of inductive work and text, Orient Securities pointed out that the launch of ChatGPT will promote the penetration of text-based AI into text production, intelligent review and other application fields, and related fields will surely usher in great room for development.
"At the same time, the work related to code development is more regular, and it is also very suitable for AI-assisted generation. Although the code generation ability of ChatGPT in the current test lacks some underlying stability, it is more flexible than Copilot. Targeted After optimization, the AI code assistance tool based on the new GPT model is expected to land in the short to medium term.” CITIC Securities pointed out.
In addition, CITIC Securities stated that although the GPT model is currently slightly inferior to the diffusion model in the field of image generation, the diffusion model can use ChatGPT to generate better prompts. For AIGC content and increasingly popular artistic creation, it will provide a powerful text form power.
It is worth mentioning that in the downstream application field of ChatGPT, there have been successful landing cases overseas.
In January 2023, the US news aggregation website BuzzFeed officially announced that it will rely on OpenAI to create and personalize various quizzes for its user interaction column Quizzes.
ChatGPT will generate a series of questions, and then generate unique and shareable articles based on individual answers.
Not long ago, Meta paid BuzzFeed millions of dollars to generate content for Meta's platform and to train creators on the platform.
Model service will be the potential direction of profit model
In terms of business model, CICC stated that currently ChatGPT has no clear profitable business model, but MaaS is a potential direction.
"MaaS can transfer the process of model training, maintenance and deployment to cloud services. This approach not only improves efficiency, but also reduces customers' dependence on model development and maintenance, enabling them to focus more on business investment." CICC pointed out.
The Sinolink Securities Research Report pointed out that the simplest business model for products such as ChatGPT includes two types, one is direct charging, such as the price of ChatGPT Pro is 42 US dollars per month, and the sale of cloud services.
"The second is empowerment charges, which mainly refer to the provision of technical services through API interfaces." China Finance Securities stated that according to the data released by OpenAI in March 2021, GPT-3 provides SaaS services to more than 300 applications, generating an average of 45% per day. According to estimates, this will bring OpenAI about $7.4 million in annual revenue.
How to Nuggets?
The hype of A-share stocks on popular themes can be described as "never absent". What investment opportunities will ChatGPT, which has exploded in the circle, bring?
"Currently, the concept stocks in the market are soaring, with strong thematic catalyst components, and have a certain game color. But if ChatGPT can make further gains in the commercialization of the future, or other AI companies can also launch Competitive AI products, the AI market is expected to last longer.” Zhao Fengfei believes.
On the upstream side, Zhao Fengfei is optimistic that listed companies with NLP technology that are highly related to ChatGPT, as well as companies specializing in AI algorithm research and commercialization, are all targets for benefit.
"In addition, the computing power and data supporting these artificial intelligence technologies are also very important. Companies with key data and underlying chips that can provide computing power or computing power are also an important link in the industrial chain."
In terms of targets and computing power, Caitong Securities recommends investors to pay attention to Haiguang Information (688041), Loongson Zhongke (688047), Baosight Software (600845), etc.
In terms of AI processors, Huaxi Securities recommends investors to pay attention to Cambrian (688256.SH), Shangtang (00020.HK), Haiguang Information (688041.SH), etc.
The application manufacturers of AIGC-related technical reserves are benefiting from Wondershare Technology (300624), Chinese Online (300364), China Literature (00772.HK), Kunlun Wanwei (300418), Vision China (000681), etc.
In terms of algorithms required for the implementation of downstream business models, Huaxi Securities recommends that investors focus on HKUST Xunfei (002230).
At the same time, others can pay attention to Hanwang Technology (002362), Torsi (300229), Haitian Ruisheng (688787), ArcSoft Technology (688088), Yuncong Technology (688327), Geling Shentong (688207), etc.
Caitong Securities recommends that investors pay attention to Hikvision (002415), Dahua (002236), HKUST Xunfei, Yuncong Technology, OPT (688686) in the field of algorithms.
However, Zhao Fengfei reminded investors that throughout the history of artificial intelligence development, whenever there are new technological breakthroughs and some innovative application scenarios, the public's expectations for AI are often very high.
However, the progress of the underlying technology, especially the algorithm, has great uncertainty and uncontrollability. People tend to be too optimistic about the pace of cashing in AI technology, which is the biggest risk in the short term.
"In addition, many listed companies are still in the early stages of AI layout, the competition landscape is not clear enough, and whether there is a sustainable business model is yet to be tested. These are also potential risks." Zhao Fengfei emphasized.