(Yicai) Feb. 14 -- "DeepSeek has taught us the value of open source," Alibaba Group Holding's Chairman Joe Tsai said at the World Government Summit about the Chinese artificial intelligence sensation.

"The thing about the open source community is that people share everything that you contribute to," Tsai noted at a conference for the WGS in Dubai yesterday. "The open source is not in the ecosystem that other people can learn from.

"If today your only purpose in life is to develop a close source AI system, I personally think the value of that endeavor is approaching zero," he added.

When asked which company he thinks is ahead in the AI race, Tsai said that everyone is in a fairly tight race, as the next model to launch will be better than some but not others, and then somebody else will launch something that's better than. “Which AI is better? Yesterday is going to be different from the answer tomorrow.” 

But what DeepSeek did is quite significant, as it challenged the current paradigm where only a handful of tech giants can afford to develop frontier AI models, Tsai noted. “They actually demonstrated that engineering innovation can drastically lower the cost of training and inference of large language models.”

DeepSeek's LLM DeepSeek-R1's performance compares favorably with leading LLMs, but at much lower cost. Its application programming interface services are priced at CNY1 (14 US cents) per million tokens of input and CNY16 (USD2.20) per million tokens of output, much lower than OpenAI o1's rates of USD15 and USD60, respectively.

Tsai explained that the main issue in this field is that "we need to use AI to apply AI" for it to be useful and productive and make an economic impact. “When looking at the phenomenon of DeepSeek, people are going to start to shift away from putting hundreds of billions of dollars into computer infrastructure [...] and start putting resources in applications.”

Tsai also drew a parallel between AI development and education, noting that the current race to build all-encompassing AI models is like having only a handful of rich parents who can afford to have the smartest kids. “There are ways to develop smart kids without using all of these massive computer resources if people can focus on applications.”

At yesterday's conference, Tsai also confirmed that US tech giant Apple has chosen Alibaba as its partner to develop AI features for iPhones in China.

Alibaba has been actively contributing to the open-source AI community. The Hangzhou-based tech giant launched its open-source platform ModelScope in late 2022 and its first open-source AI model Qwen-7B in August 2023.

ModelScope now includes various models for language processing, multi-modal tasks, mathematics, and coding. Developers have created over 90,000 Qwen-based derivative models on global open-source community Hugging Face, according to data from Alibaba.

Editor: Futura Costaglione