(Yicai) Nov. 4 -- DeepSeek made DeepInfer, a new high-performance inference framework that powers its large-scale production services, open source on Oct. 20 to help artificial general intelligence arrive sooner, according to a research engineer at the leading Chinese artificial intelligence company.

“In AI competition, the opponent is not other model developers but the passing time,” Gao Xi said during a dialogue session on AI at the World Internet Conference Wuzhen Summit yesterday. Regardless of which company develops AGI, the technological changes it brings, such as medical breakthroughs, will benefit all mankind, he added. “When AGI arrives, you won’t care which firm develops it.”

When new methods such as lower-cost training and better model structures are discovered during development, the most natural approach is to immediately make them public for validation and application, he stressed. If others solve the problem, you should quickly follow up with an application, and if you solve it yourself, share it with others to drive the overall process, he noted.

Open-source models improve cost-effectiveness, avoid price monopolies that may occur with closed-source tools, and lower the public-use threshold, noted Liu Dayiheng, head of the Qwen large language model at Alibaba Token Hub Business Group. In addition, in scenarios where data cannot be leaked, open source lets users deploy models locally to protect it, he pointed out.

“Top companies may overlook issues such as minority languages, specialized vertical applications, and special hardware/chip adaptation, but open source allows more participants to contribute technology and thus form a more complete technological ecosystem,” Liu stressed.

Open-source models also carry risks, mainly irreversibility and a lack of external protection dependencies, Liu said, adding that since it is difficult to fundamentally solve the problem by simply restricting open models, it is recommended to seek outside solutions, including improving laws and regulations, developing industry standards, and building more infrastructure to mitigate security risks.

Editor: Martin Kadiev