The victory of a few individuals.
Huawei Pangu, Baidu Wenxin, Alibaba's "Tongyi Qianwen"... The fire ignited by ChatGPT's large models is burning more and more fiercely, and the "battle of large models" is in full swing, creating quite a buzz.
However, Baidu has poured cold water on the hot market. On July 19th, Wu Tian, Vice President of Baidu Group and Deputy Director of the National Engineering Research Center for Deep Learning Technology and Applications, told All-Weather Technology that the emergence of numerous new models is only a temporary phenomenon. The ultimate fate of large models will eventually converge on a few select ones.
Indeed, AI's impact on industries is becoming increasingly profound, and large models are bringing about profound changes in industry structure and society. This new era of artificial intelligence, which aims to enhance social and corporate efficiency, is just beginning. Baidu's founder, Robin Li, also stated that artificial intelligence will be the fourth industrial revolution.
Including tech giants like Baidu and Huawei, the current focus is on developing large models for various industries. They are attempting to become the infrastructure on which these companies operate, seizing the opportunity to enhance efficiency in the era of large models.
Behind the fervor lies the challenge of data and computing power, which looms large in the face of large models. In order for large models to empower industries and find commercialization paths, there is still a long way to go.
These challenges have resulted in only a few large models surviving in the end. Wu Tian also admitted that the challenges of cost and demand will ultimately lead to consolidation in the realm of large models.
Cost is a hurdle that many players with insufficient financial resources find difficult to overcome. For example, in the case of ChatGPT, Chen Hang, an analyst at Zheshang Securities, conducted a calculation that showed supporting its computational infrastructure would require at least tens of thousands of NVIDIA GPU A100s, with training costs exceeding $12 million.
Furthermore, data determines the quality of training, performance, and ultimate application of large models.
Wu Tian believes that the industrialization of large models faces three challenges: first, large models have a large volume, making training difficult and costly; second, large models require significant computing power and high performance; and third, there is a need for large-scale data collection, mining, construction, screening, and cleaning, which is a massive undertaking in itself.
She stated that large models are costly and require comprehensive capabilities and unwavering confidence in dealing with complex training data. On the other hand, from a demand perspective, large models have tremendous potential for application in the future, and relying on a small number of large models can create a wide range of application ecosystems.
Wu Tian compares the industrial model of large models to chip foundries. Currently, there are many chip models and manufacturers, but only a few foundries. The foundries encapsulate expensive equipment and production lines, sophisticated processes, and the production process itself. Companies with demand only need to provide production plans to the chip foundries to obtain the desired chips. Similarly, large model platforms can encapsulate big data, computing power, and algorithms, and establish standardized production models.
"In fact, there is no need for a large number of large models. For application developers, it is not necessary to develop a large model for every application," Wu Tian believes. Application developers should define the problem and leave the requirements for AI model capabilities to the large model platform, which will handle the production. Only then can various industries flourish. There are many big shots who share similar views with Wu Tian. Li Di, the CEO of Xiaobing Company, also believes that "there is simply no need for so many large models in the market, and by 2024, the hype will die down and it will be revealed who is swimming naked."
For large models, commercial applications are the most crucial point and the key to determining the future industry landscape.
Wu Tian believes that there is great potential for large models in future applications. Every industry has a lot of room for improvement and can benefit from new AI technologies. "The future relies on a few large models with a wide range of applications."
Wu Tian believes that in the next step, each company and institution will gradually find its own position and move towards specialization.
According to Fang Han, CEO of Kunlun Wanjwei, in the next 3-5 years, large models will generate more end-to-end content production tools, completely changing the content production forms and processes in industries such as literature, music, comics, animation, short videos, long videos, and film and television.
In industries such as finance, e-commerce, and energy, companies such as Baidu, Alibaba, and Huawei are also attempting to improve the existing business models in these industries through large models. They are trying to become the data infrastructure for these companies and help improve efficiency.
Perhaps as Wu Tian said, the endgame of large models will be concentrated in the hands of a few players. Baidu's "cold water" not only cools down blind investments in the industry but also hopes that players in the field will think clearly about the direction of their own efforts.
The commercial exploration of large models has just begun, and the "battle of large models" is still raging. People also hope that the "Fourth Industrial Revolution" brought about by artificial intelligence can bring more changes to productivity and life. Before the endgame, there will still be players who continue to forge ahead, trying to find their new opportunities in this unprecedented transformation.