WuDao 2.0 now No. 1 in generating text and speech4. June 2021
WuDao 2.0 now No. 1 in generating text and speech
The Beijing Academy of Artificial Intelligence (BAAI) unveiled a natural language processing AI model with the world’s highest number of parameters during the 2021 Beijing Academy of Artificial Intelligence Conference in Beijing a few days ago. The WuDao 2.0 model is capable of processing 1.75 trillion parameters for generating text and speech, outperforming Switch Transformer and GPT-3.
“WuDao 2.0 is the first trillion model in China and the largest in the world,” said Tang Jie, deputy director of BAAI Academy. WuDao 2.0 achieved excellent results in nine benchmark tasks in the pre-trained model domain, Tang said, and came close to breaking the Turing test in creating poems and couplets, summarizing texts, answering questions and painting. “WuDao 2.0 aims to enable machines to think like humans and achieve cognitive capabilities beyond the Turing test,” Tang said.
Previously, Google Brain’s Switch Transformer held the record with 1.6 billion parameters, surpassing the 175 billion parameters of OpenAI’s popular GPT-3. More parameters equals a more sophisticated machine learning model.
The Beijing Academy of AI is sponsored by the Beijing Municipal Science and Technology Commission and the Haidian County Government. The model supports both Chinese and English after being trained on 4.9 terabytes of images and text. Partners now include smartphone maker Xiaomi, short video company Kuaishou and delivery service provider Meituan.
It is the latest sign of China’s goal to become a world leader in AI. Members of the U.S. National Security Commission on Artificial Intelligence have warned that the U.S. should act quickly to maintain its competitiveness in AI against China, as Beijing hopes to displace the U.S. as the dominant world power through technology.
China-based AI researcher Blake Yan had this to say about the growing number of advanced NLP models driven by increased computational capacity and new facilitation capabilities: “These sophisticated models trained on gigantic datasets require only a small amount of new data when used for a given feature because they can transfer already-learned knowledge to new tasks, just like humans.”
Such large-scale and pre-trained models also pave the way toward artificial general intelligence, when machines (hypothetically) become as intelligent as humans on a variety of tasks, Yan added.