Shulou(Shulou.com)11/24 Report--

July 19 news, "the capacity of the new version of Wenxin has exceeded ChatGPT 3.5, which is also an important milestone in carrying out related technical work in our country." Wu Tian, vice president of Baidu Group and deputy director of the National Engineering Research Center for Deep Learning Technology and applications, told NetEase and other media.

According to her, IDC's latest "AI Model Technical capability Evaluation report, 2023" shows that Baidu Wenxin Model 3.5 won 7 full marks of 12 indicators, with the first comprehensive score, the first algorithm model, and the first industry coverage.

It is reported that the IDC evaluation report examines more than 10 indicators of the model around the three dimensions of product technology, service ecology and industry application. 14 domestic mainstream models, including Baidu, Alibaba, Tencent, Huawei, iFLYTEK, 360and Shangtang, participated in the evaluation. The results of the report show that Baidu Wenxin has obvious advantages in model ability, tool platform, ecological layout and industry coverage, and has entered the stage of commercial landing exploration ahead of time.

According to Wu Tian, Baidu began the research and development of deep ploughing pre-training models in 2019 and successively released a series of knowledge-enhanced Wen Xin models. Not long ago, Baidu officially released the 3.5 version of Wenxin Model, which further made innovations in a number of core technologies, such as basic model, knowledge enhancement, retrieval enhancement, and so on.

Specifically, she said that Wen Xin big model achieved "first" thanks to the advantages of Baidu's "chip-framework-model-application" four-tier technology stack, the core features of knowledge enhancement and the prosperous big model ecology. In particular, Baidu has a self-developed deep learning platform, flying paddle, which strongly supports the efficient training and reasoning of large models. The collaborative optimization of flying oars and Wen Xin increases the effect of the latest version of Wen Xin Model 3.5 by 50%, increases the training speed by 2 times, and increases the reasoning speed by 30 times.

In terms of large model ecology, she said that Baidu Wenxin has formed a trinity ecosystem of enterprise, education and community. The latest data show that Baidu has more than 7.5 million developer base, 200000 enterprise ecological foundation, multi-level large model personnel training, enterprise empowerment, developer operation. Baidu has also set up a 1 billion venture capital fund to encourage big model creativity and prosperity of big model ecology.

She said bluntly that at present, the industrialization of the large model is still facing great challenges, which can be summed up in three aspects: first, the volume of the large model is indeed very large, which brings high difficulty and high cost of training; second, it requires a very large scale of computing power and high performance requirements; third, the scale of data is also very large, and the collection, mining, construction, screening, and cleaning of these data is itself a very big project. " Large model platform is an expensive large computing system, in fact, there is no need for a large number of large models, and for users, there is no need for every application to develop large models. "

As for the end of the "hundred Model Wars"? Wu Tian said, "in the past few months, a large number of new large models have emerged, but this is a stage phenomenon. In the future, various enterprises and institutions will gradually find their own positioning, and the next step will move towards their own subdivision. In the end, it will only focus on a small number of large models, but relying on a few large models, there will be a very wide range of application ecology."

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.