After a quarter of ChatGPT's explosion, domestic AI large models began to explode intensively. From a technical point of view, large models originated in the field of natural language processing, represented by Google's BERT and OpenAI's GPT, and the parameter scale has gradually increased to trillions, and the amount of data used for training has also increased significantly, bringing about the improvement of model capabilities.
At present, domestic large models generally start on the B side, and enterprises begin to think about how to make the scale of existing large models "small", with smaller parameters, to make more efficient and more suitable for vertical scenarios. In the future, the large model with the dual advantages of technology and industry will be used as a platform to support countless intelligent applications.
A big model with a hundred flowers
A few months ago, Baidu released the Wen Xin Yiyan big model, and since then, many domestic enterprises have followed suit, and launched their own large models according to the characteristics and industrial planning of each company.
On April 8, at the AI Big Model Technology Summit Forum, Tian Qi, Chief Scientist of HUAWEI CLOUD Artificial Intelligence Field, introduced the progress and application status of Pangu Big Model, including NLP large model, CV large model, multimodal large model, and scientific computing big model. The Pangu model will focus on the application of subdivided scenarios to empower industries such as coal mining, cement, electric power, finance, and agriculture.
On April 10, SenseTime not only displayed the language model under the "SenseNova" large model system at the Shanghai Technology Exchange Day, but also displayed a series of generative AI models and applications such as AI literate picture creation, 2D/3D digital human generation, and large scene/small object generation, and announced SenseTime's R&D system for "big model + big computing power" integration and innovation based on the AI large device SenseCore.
On the same day, Kunlun Wanwei announced that it and Singularity Zhiyuan cooperated with Singularity Zhiyuan to develop and China's first domestic large language model "Tiangong" 3.5 that truly realized the emergence of intelligence will be released soon, "Tiangong" as a large-scale language model, with powerful natural language processing and intelligent interaction capabilities, can realize intelligent question and answer, chat interaction, text generation and other application scenarios.
At the Alibaba Cloud Summit on April 11, Zhou Jingren, Chief Technology Officer of Alibaba Cloud Intelligence, officially announced the launch of the big language model "Tongyi Qianqian". According to him, Tongyi Qianwen is a super-large-scale language model with functions such as multi-round dialogue, copywriting, logical reasoning, multimodal understanding, and multi-language support. Daniel Zhang, Chairman and CEO of Alibaba Group and CEO of Alibaba Cloud Intelligent Group, said at the summit that all Alibaba products will be connected to the "Tongyi Qianwen" model in the future and undergo comprehensive transformation.
According to Zhou Jingren, CTO of Alibaba Cloud Intelligence, in the future, every enterprise can not only call all the capabilities of Tongyi Qianwen on Alibaba Cloud, but also combine its own industry knowledge and application scenarios to train its own enterprise big model. For example, each enterprise can have its own intelligent customer service, intelligent shopping guide, intelligent voice assistant, copywriting assistant, AI designer, autonomous driving model, etc.
A new inflection point for AI
Since December 2022, the GPT interface provided by OpenAI has allowed the outside world to see the possibility of MaaS (Model-as-a-service) in the business model, and the large language model can be called as a service. From the initial development of the model, the cleaning of data, to the training and testing of the model, and the model as a whole can enter a unified model standard website, which allows users to quickly find and use models, and reduce the threshold for model use.
Zhang Hongjiang, Chairman of Beijing KLCII Artificial Intelligence Research Institute, said in a speech a month ago, "Technical bottlenecks and commercialization problems build the ups and downs of the AI industry, and the hindrance of commercialization has become the difficulty of the "third wave" of AI; The large model has become a new inflection point, and the trend of infrastructure of the ability of the large model is gradually emerging, and it is believed that many technological and product breakthroughs will be driven in the next few years. ”
He believes that the transition from the "big refining model" to the "big refining model" is a paradigm shift. The development of future APP will be based on the "big model + fine-tuning" assembly line operation mode to provide a steady stream of intelligence to the industry. Compared with the previous way of making APP and refining small models, it releases the waste of resources such as manpower to recreate small models, greatly reduces development costs, makes the marginal cost zero, and brings a hundred or even a thousand times more productivity improvement.
Microsoft Cloud will form three rotational growth curves in 2022, of which the growth rate of intelligent cloud exceeds 20%, the growth rate of enterprise software exceeds 40%, and the growth rate of AI large models exceeds 100%. Enterprise customers lease GPU computing power through Microsoft's public cloud Azure, call large models, and then enter data to train their own small models to transform their businesses.
The "2022 China Large Model Development White Paper" also suggests that for industry users, first, technology buyers in all industries should embrace large models as soon as possible; Second, in terms of cooperation, it mainly focuses on the adaptability of large models to their own business; Third, we should join hands with leading manufacturers to create industry benchmarks.
Domestic large models still need to be precipitated
After several years of development, the large model has been relatively mature in research and development technology, but on a global scale, the landing of the large model is still in the early stage. Although the large models developed by domestic manufacturers have internal business landing scenarios, they have not yet formed a mature commercialization model on the whole.
Ali pointed out Daniel Zhang that the big model is an all-round competition of "AI + cloud computing", and the research and development of large models with more than one trillion parameters is not only an algorithm problem, but a complex systematic project that includes the underlying huge computing power, network, big data, machine learning and many other fields, which requires the support of ultra-large-scale AI infrastructure.
Recently, Kunlun Wanwei said in response to the regulatory authorities' letter of concern that data scale and quality are an important parameter of artificial intelligence capabilities. In the process of R&D, sufficient data need to be obtained for user model training, and model update iteration is carried out, which has certain uncertainty, and if the progress of technology R&D is not as expected, it may lead to a slow industrialization process.
In terms of less commercialization of Applied Practices than expected, the company said that if the product is not effectively integrated with Applied Practices, it will have an impact on its progress. There is great uncertainty about whether the subsequent commercialization can be successfully achieved.
The reply mentioned that generative artificial intelligence (AIGC) and artificial intelligence are hot spots in the industry, with significant future commercial value, and many technology giant companies are deployed in this field, and industry competition may further intensify in the future. At the same time, it is expected that with the launch of domestic products, relevant policies such as network security and data security will be extended to the field of artificial intelligence algorithms, which will add certain policy risks to the research and development of such products.
summary
The capabilities of large models in content creative generation, dialogue, language or style translation, and search will bring a variety of benefits to various application fields. The large model basic platform contains huge development opportunities in the data layer, model layer, middle layer, and application layer.
However, under the boom of large-model entrepreneurship, some easily overlooked problems have emerged. Whether it is a large factory or a start-up, in this process, it is necessary to pay more attention to data preparation, computing power support, talent reserve and risk management and other issues in order to achieve better results in commercialization.