From Chat
The emergence of Chat-GPT and its industrialization in the short and medium term will provide key auxiliary support for the transformation from user creation (UGC) to AI creation (AIGC). Combined with the underlying technical logic of Chat-GPT, we think that the industrialization direction of Chat-GPT in the short and medium term may include: 1)Chat-GPT is of great significance to the application of AIGC in text mode, and has shown excellent performance in inductive text work. 2) The work related to code development is more regular and very suitable for AI-assisted generation. 3) Image generation field: The effect of GPT model in the field of image generation is slightly worse than that of diffusion model, but the diffusion model can generate better Prompt by using Chat-GPT, which provides a powerful driving force for text form. 4) Intelligent customer service work. The success of Chat-GPT proves that the Transformer model is not in trouble. The continuous breakthroughs in AI technology and methods are driving the global AI industry into an accelerated development stage, and the overlapping AI industry cluster effect is increasingly prominent. The AI industry is expected to become one of the most valuable industrial tracks in the global science and technology field in the medium term. The AI industry is expected to continue to maintain the stable industrial value chain structure of "chip+computing infrastructure +AI framework & algorithm library+application scenarios", and enterprises with complete data closed-loop structure and good data self-processing ability are expected to continue to be beneficiaries.
▍ Origin of the report: Chat-GPT attracted a new round of AI fever and explored the possibility of industrialization.
Chat-GPT, the latest language model published by OpenAI team, was released to the community on November 30, 2022, and received good feedback immediately. From the results of test feedback, compared with the previous generation GPT-3, Chat-GPT can answer a variety of daily questions with dialogue as the carrier, and its memory ability and length for the history of multiple rounds of dialogue are enhanced. Compared with big models such as GPT-3, Chat-GPT has a more comprehensive answer, which can be answered and expounded from multiple angles and in all directions. Compared with previous big models, the knowledge is mined more fully. The strong "out-of-circle" of Chat-GPT has attracted a series of tests on its possible future industrialization direction. From the test results, Chat-GPT has shown its advantages in inductive writing, creative writing, code modification, scientific research assistance and other fields. At first, conversational AI can give generally safe answers to large-scale and fine-grained questions, and form creative answers with certain logic according to the context. This report will focus on the technical logic behind Chat-GPT, its overall impact on the AI industry chain and the possibility of industrialization.
▍ Technical logic: Based on GPT-3.5, additional training is conducted based on human feedback learning, and the future development direction of Transformer model is given.
OpenAI team fine-tuned a model in GPT-3.5 series, trained the model with the same method as InstructGPT, namely human feedback reinforcement learning (RLHF), and optimized the data collection settings. From the final result, Chat-GPT only uses the selected tens of billions of parameters (compared with the hundreds of billions of parameters of GPT-3), and the response quality is equivalent to or even better than that of GPT-3, which highlights the importance of data quality. The big model may bid farewell to the past era of blindly stacking data sizes. The success of Chat-GPT was achieved on the basis of a lot of solid work in the early stage, not a technological leap. The model finds a subjective task-oriented way to tap GPT3′ s powerful language ability, so that the model can "unlock" and tap the knowledge and ability in the massive data learned by GPT3. Therefore, starting from this underlying technical logic, we can quickly find the industrialization direction suitable for Chat-GPT in the short to medium term: a truly all-round intelligent content generation assistant.
▍AI industry impact: the cost of computing power has dropped+high-quality data has spawned the underlying application, and the model opening has become the future trend and accelerated the iterative efficiency.
The success of Chat-GPT proves two points: 1) simply expanding the model parameters is not the only way out; 2) Making the model open to the public for testing at an early stage and collecting human feedback data is more conducive to model iteration. In the first two stages of AI development in the past 10 years, the progress of artificial intelligence is more reflected in scale-based technological breakthroughs. For example, from 2015 to 2020, the amount of calculation used for model training increased by six orders of magnitude, and with the increase of scale, the quality of output results also ushered in qualitative changes, which surpassed the human level in the fields of language, writing and image recognition. However, at the practical level, because of the huge computing power, special GPU configuration is often needed, and the training process is relatively closed, so most people can’t use it, so the technology can’t be reached by most people. In the third stage of artificial intelligence, with the emergence of newer technologies, better algorithms and larger models, the cost of computing power is getting lower and lower, which makes the cost of model training and operation continue to decline, and the gradual popularization of algorithms from closed testing to open testing and open source also lowers the threshold for use. Therefore, artificial intelligence has reached the level of supporting popularization in terms of economy and availability. Thanks to the gradual improvement of the availability of AIGC infrastructure, the platform layer has become more stable, the cost of computing power has continued to drop, the model has gradually become open source and free, and the node of explosive development of application layer is approaching.
▍ Application scenario: realize booster from UGC to AIGC.
At present, we are going through the launching stage of the transition from Web2.0 to Web3.0, and we have seen the transformation of content creation from professional creation (PFC) to user creation (UGC). The emergence of Chat-GPT and its industrialization in the short and medium term will provide key auxiliary support for the transformation from user creation (UGC) to AI creation (AIGC). Combined with the underlying technical logic of Chat-GPT, we think that the direction of industrialization of Chat-GPT in the short and medium term is mainly divided into four sections.
1)Chat-GPT is of great significance to the application of AIGC in text mode, and it shows excellent performance in inductive text work. In the short to medium term, Chat-GPT can be quickly used in office auxiliary tools, such as meeting summary, document translation, routine report, etc., which can improve office efficiency and save labor costs.
2) The work related to code development is more regular and very suitable for AI-assisted generation. Copilot, which was launched in cooperation with Github and Microsoft in mid-2021, is the most mature AI code completion tool at present. According to Github data, there have been 1.2 million users in the past year, and 40% of the codes written by these users were automatically generated by Copilot. By October 2022, Copilot had raised $22 million. Chat-GPT is more flexible in code generation than Copilot, but it lacks some underlying stability. After targeted optimization, the AI code assistant tool based on the new GPT model is also expected to land in the short to medium term.
3) The field of image generation has become a hot spot in the layout of primary market companies in the second half of 2022. With the popularity of Dalle2, the idea of replacing human painters with AI in business manuscripts is basically clear. At present, the effect of GPT model in the field of image generation is slightly worse than that of diffusion model, but the diffusion model can generate better Prompt by using Chat-GPT, which provides a powerful driving force for AIGC content and increasingly hot artistic creation.
4) The project that 4)Chat-GPT is most suitable for direct landing is intelligent customer service. According to the existing completion degree of the model, with targeted manual feedback training in vertical industries, Chat-GPT can become an intelligent customer service product and take the lead in application in toC scenarios. Compared with the current intelligent customer service, the customer service supported by Chat-GPT will make significant progress in flexibility and humanized service.
▍ Risk factors:
The development of AI core technology is less than expected risk; Policy supervision in the field of science and technology continues to tighten risks; Global macroeconomic recovery is less than expected risk; Macroeconomic fluctuations lead to the risk that IT expenditure of European and American enterprises is less than expected; The development of global cloud computing market is less than expected risk; Enterprise data leakage and information security risks; Industry competition continues to aggravate risks.
▍ Investment strategy:
Behind the bright performance of the Chat-GPT model is that researchers have discovered the potential brought by the method of class feedback reinforcement learning on the way forward of the Transformer model. For the development of industrial AI, the optimization of data quality and the reserve and computing ability of AI researchers will be the core competencies for whether they can stay at the forefront of AI applications in the future. The success of Chat-GPT proves that the Transformer model is not in trouble. The continuous breakthroughs in AI technology and methods are driving the global AI industry into an accelerated development stage, and the overlapping AI industry cluster effect is increasingly prominent. The AI industry is expected to become one of the most valuable industrial tracks in the global science and technology field in the medium term. The AI industry is expected to maintain a stable industrial value chain structure of "chip+computing infrastructure +AI framework & algorithm library+application scenarios", and enterprises with complete closed-loop data structure and good data self-processing ability are expected to continue to be industrial beneficiaries.
This article comes from: financial circles
Author: CITIC Securities