Source: Zhitong Finance
According to an IDC article, recently OpenAI released ChatGPT and GPT-4, which sparked a public carnival in the AI world, and graphic applications such as StableDiffusion, Midjourney, and Dall·E2 have also rapidly become popular. Baidu, on the other hand, held a Wenxin one-word press conference on March 16, showcasing the big models and generative AI capabilities of Chinese manufacturers. The artificial intelligence market has officially begun a new era — the era of AI driven by big models. In response to this,IDC asks ten questions about the real status and future development of GPT and AIGC.
The relationship between big models, ChatGPT, and AIGC
All AI applications defined by IDC refer to AI decision-making systems based on machine learning algorithms. A large model refers to an algorithm model that reads massive amounts of data and has huge parameter scales. The industry generally believes that parameters exceeding 100 billion levels are large models, and that thousands or more GPU/CPU chips may have been used during the training process. ChatGPT and AIGC are both big model application scenarios. ChatGPT can be compared to the original conversational AI applications and AI-enabled search applications. AIGC can be divided into generating text, generating images, and generating videos. It can also be classified as one of the application scenarios for big models.
Where are the changes in the big model represented by GPT-4
Since the release of the GPT1.0 model, OpenAI has continued to iterate and released GPT 2.0, GPT 3.0, and GPT 3.5. The release of GPT4.0 is an inevitable stage for it to continue investing in the AI big model. Compared to previous models, GPT-4 has a larger number of parameters, takes longer to iterate on the model, and can also give more accurate results. IDC believes that the release of the new version is an inevitable result of the gradual development of the big model. As Li Yanhong, CEO of Baidu Group, said, “The company releases a new version of the big model every year, which is a natural continuation of years of hard work.”
Possible industrial impact of ChatGPT
ChatGPT is essentially an application of conversational AI, and conversational AI has been widely implemented. According to artificial intelligence market size data tracked by IDC, the conversational AI market size reached 5.46 billion yuan in 2022, and its market penetration rate is relatively saturated. The wave triggered by ChatGPT has prompted mainstream manufacturers to introduce big models into their conversational AI applications, which will drive a new round of growth in conversational AI-related markets. Furthermore, in search and marketing scenarios, ChatGPT type applications may spawn new product forms.
Products available on the market
Regarding AIGC, in addition to the well-known StableDiffusion, Midjourney, and Dall·E2, some commercial companies also provide cloud support for this. Currently, Amazon Web Technology provides Stable Diffusion support through IndustryAI and SageMaker. Baidu's Wenxin Yigen began solicitation on March 16, providing five major functions: literary creation, commercial copywriting, mathematical logic estimation, natural language understanding, and multi-modal generation. In addition, many digital human companies have also adopted AIGC-related technology. From a technical point of view, most of the products currently on the market can only be made by Wensheng, and the release of Wensheng video products will take time, which is worth looking forward to.
Regarding the big model, in addition to the big model already released by the open source community, the big models currently used by the provider include the GPT big model integrated on Microsoft Azure, Baidu Smart Cloud and the Wenxin model supported by Baidu Flying Paddle, the Huawei Cloud Pangu model, and the Alibaba Cloud M6 model. Most of the large models developed by local manufacturers support localized deployment.
Triggered changes in the AI industry
AI applications deployed in the past few years are likely to be replaced by AI based on big models in the next few years. The upgrade iteration may begin with prioritizing scenarios with massive amounts of data. When AI applications supported by big models become mainstream, manufacturers that cannot utilize the capabilities of big models will lose their competitive advantage.
In future jobs, AI assistants will replace more human jobs. Applications such as Wensheng Map, such as searching for elementary content in various fields, can all use AI-generated content.
Possible investment size
The computing power consumed by large models that have been disclosed so far, such as the GPT series and the Bert series, can be found according to public data. However, to actually land in the industry, the specific investment scale must be determined by application scenarios. The cost of investment is related to the computing power required, whether to deploy a complete big model, and the data flow to be inferred.
Driven market opportunities
Pure AI computing power market: The first to directly benefit from this wave of AI boom are AI computing power providers, including chip manufacturers, AI server vendors, and AI computing power cloud service providers that support big model training and inference.
A combination of big models and computing power: AIaaS+AiPaaS. The market is provided with highly optimized solutions combining large models and computing power to help users lower the threshold for hardware use, improve development efficiency, and reduce overall investment costs. Typical solutions include Baidu's “AI Platform” and Shang Tang's “AI Big Device”.
Big model as a service: Open the big model development platform for external users. This market is a highly innovative market, but there are still high entry barriers.
Where to start following this AI wave
Big model manufacturers are starting to upgrade existing AI software to AI applications supported by big models. AI supported by big models can be introduced by connecting with partners according to application scenario priorities. However, at the MaaS (model as a service) product level, there are not many mature products to choose from on the market. It is expected that products from dozens of manufacturers will be launched in the second half of this year. You can take the lead in testing big model capabilities on public clouds in fields where data privacy requirements are not high.
Issues that the next generation of AI needs to pay attention to
Copyright for generative AI-generated content needs to be planned in advance. Content such as images generated by generative AI after reading massive amounts of data may cause copyright issues and need to be controlled by rules in advance.
Changes to the original process: On the one hand, content generated by generative AI also requires human review before it can be published, and on the other hand, changes in the workflow may be required to adapt to the addition of AIGC.
Since it is still in the early stages of technological maturity, application scenarios in traditional industries are not very clear, and input and output are more difficult to evaluate than at present.
Break out of today's AIGC to see future AI applications
Drawing on today's Wensheng Map and Wensheng Video applications, in fact, most of them are AI applications implemented through various technical routes based on small models that have existed in the past few years. Similar application scenarios in various industries can be based on existing AI models to create AI applications that everyone can use in the form of low code. Even future AI applications may be in the form of inputting natural language and directly outputting results.
Lu Yanxia, director of IDC China research, said,The popularity of next-generation AI continues to rise, but due to its low technical maturity and high deployment costs, actual implementation still requires caution. However, in terms of macro trends, rapid iterative technology represented by big models and generative AI will inevitably spawn a new era of AI.
Editor/jayden