Baidu's Wu Tian: The New Version of Wenxin Yiyan Surpasses ChatGPT 3.5
-
On the afternoon of July 24, IDC released an AI large model technical capability evaluation report, showing that Baidu's Wenxin Large Model 3.5 achieved full marks in 7 out of 12 indicators, ranking first in comprehensive evaluation. In a media briefing with Sina Tech and others, Wu Tian, Vice President of Baidu Group and Deputy Director of the National Engineering Research Center for Deep Learning Technology and Applications, stated that the new version of Wenxin Yiyan 3.5 has surpassed ChatGPT 3.5, and in the future, there will only be a few large models in China.
"Before OpenAI released ChatGPT, there were very few companies or institutions truly working on large models. In the past few months, a large number of new models have emerged. This is a temporary phenomenon. In the process of evolution, companies and institutions will gradually find their positioning and move toward their specialized directions," Wu Tian said.
In Wu Tian's view, the future of large models will be concentrated in a few major models. This is because building large models from the ground up is extremely costly and not something that can be done casually. It requires years of accumulation, comprehensive capabilities, and absolute determination to sustain long-term investment. Therefore, it is clear that only a few large models will remain in the future.
Wu Tian pointed out that the industrialization of large models faces significant challenges, which can be summarized in three aspects: First, the size of large models is indeed massive, leading to high training difficulty and costs. Second, the demand for computing power is enormous, with very high performance requirements. Third, the scale of data is also vast. The reason we now have models with hundreds of billions or even trillions of parameters is strongly correlated with the massive amount of custom data. The scale of data is enormous, and collecting, mining, constructing, filtering, and cleaning this data is itself a massive undertaking.
"There really doesn’t need to be a large number of large models, and for application developers, not every application requires the development of a large model," Wu Tian said. In his view, the industrial model for large models can be compared to chip foundries. There are many types and manufacturers of chips, but only a few foundries. The value of chip foundries is immense, but society doesn’t need many companies to achieve this capability. With just a few chip foundries, companies that need chip production can simply provide their production plans to the foundries to obtain the chips they want.