Yang Hongxia, the former leader of Alibaba‘s M6 model, left her post at the Alibaba DAMO Academy Intelligent Computing Lab in September last year for personal and family reasons, and has now joined ByteDance’s AI Lab (Artificial Intelligence Lab) to participate in the R&D of language generation model, as reported by 36Kr.
Yang Hongxia is reported to be in a leading position within the language generation model team and reports directly to Yang Zhenyuan, vice president of ByteDance. 36Kr asked ByteDance for verification on the above news, but no reply has been received.
Yang Hongxia was the technical director of M6, a super-large-scale multimodal pre-training model in Alibaba‘s DAMO Academy. He graduated from Nankai University and Duke University’s Department of Statistical Science and studied under David Dunson, a well-known statistician. She has nearly 100 top papers in international conferences and top journals in the field of artificial intelligence.
M6 is a universal AI model that has multi-modal and multi-task capabilities and is especially good at designing, writing and developing questions and answers. It has wide application prospects in e-commerce, manufacturing, literature and art, scientific research and other fields, and the big model is the core of competition in AI wave.
M6 was first released in March 2021 and was the first multi-modal large model with 100 billion parameters in China. Subsequently, it was changed to trillion parameters with 480 GPUs, and then successfully trained a 10 trillion model within 10 days on the basis of 512 GPUs, which has the same parameter scale as GPT-3, but its energy consumption is only 1%.
It is worth mentioning that Jack Clark, former policy director of OpenAI, praised the development at that time: “The scale and design of these models are amazing. This looks like a manifestation of the gradual growth of many Chinese AI research organisations.”
Yang Hongxia led the M6 model from only tens of billions of parameters to trillions of parameters, and led the development of cognitive intelligence technology and situational implementation. In March 2021, the M6 model was released for the first time. By the end of 2021, when Alibaba Cloud released the latest progress of M6, its parameters had reached 10 trillion, making it the largest AI pre-training model in the world at that time.
DAMO Academy is also at the forefront of the implementation of these large models. By the end of 2022, Alibaba Cloud Moda Community had opened up more than 300 AI models, including the M6 and Tongyi models, providing developers with basic AI model resources.
However, because the DAMO Academy AI model is still far from actual implementation, and because of the background of reducing costs and increasing efficiency of big factories, the business is not receiving much more support at the moment, which has a knock on effect of seeing people leave their positions, such as Yang Hongxia, Jin Rong, head of intelligent research field, and Hua Xiansheng, head of urban brain laboratory.
Wang Huiwen, co-founder of Meituan, Li Zhifei, CEO of Mobvoi, and Zhou Bowen, former head of JD technology, are still hunting for R&D talents, and using the logic of “investing in people” to raise the valuation of AI startup projects with great technology to hundreds of millions of yuan. This is all a part of a larger talent war in the Chinese AI industry.
For the R&D of large models, talents often determine the upper limit of an algorithm’s ability. Many practitioners have said that there are at least a few different solutions to problems such as capital, data and computing power, but algorithms, which require human participation, demand a human eye to figure out any large problems. In the eyes of some AI practitioners, “a good algorithm can improve execution and efficiency while minimizing storage capacity.”
With Yang Hongxia now joining ByteDance suggests that the company is trying to exert some strength on large model development. The company’s language model team is led by the ByteDance Search Department. At present, the team size is about ten people, while the picture model team is led by the intelligent creation team under the Product R&D and Engineering Architecture Department.
According to some insiders, the progress of the image generation model is going well, and it will be combined with Douyin, CapCut and other products. However, due to the late start and weak foundation of the language generation model, the effect has not yet reached the team’s expectations.
A few days ago, Baidu, Inc. announced that it will release its ERNIE Bot cloud services and applications on March 27th, thus taking the lead in the big model competition. However, it remains to be seen whether ByteDance, which has gradually formed a talent team, can catch up.