DgcityA senior engineer at DeepSeek has offered the clearest public defense of the Chinese laboratory’s relentless open-source strategy, arguing that sharing its most advanced research is worth the added safety risk because it shortens the road to artificial general intelligence. Gao Xi spoke on Tuesday at the World Internet Conference in the canal town of Wuzhen, an annual gathering that this year made open-source AI its central theme.
Gao’s pitch was framed around time rather than competition. Many people ask why DeepSeek gives away results that cost enormous effort to produce, he told the audience, and the answer is simple: to bring AGI closer. If general intelligence were otherwise a decade away, he said, he would hope to pull it to one year; if it were a year away, he would want it tomorrow. The real opponent for model companies, in his telling, is not a rival firm but the clock.
The logic rests on how AI research actually compounds. When a lab finds a cheaper training method or a better model structure, publishing it lets others test, challenge and build on the finding immediately; locking it away may preserve a lead for a quarter, but over a longer horizon it delays the point when the wider economy gets the technology. Keeping models secret, Gao said, does not eliminate safety problems either — those risks exist whether a system is open or closed. He did concede the obvious objection: open weights can amplify harm once a capable model is freely downloadable, and he said DeepSeek feels a responsibility to keep those problems from widening.
He was joined on stage by Liu Dayiheng, who leads Alibaba’s Qwen project, who called the shift toward open models irreversible. Liu pointed to scale as evidence: Qwen has released more than 460 models since its first open publication three years ago, those files have been downloaded over four billion times, and outside developers have built more than 330,000 derivative versions. Without openness, he argued, the core technology behind large models would sit with just a handful of companies.
The remarks carry weight beyond the conference hall. Pioneers including Geoffrey Hinton and Yoshua Bengio have warned that systems smarter than humans could eventually slip outside human control, a concern that downloadable weights make harder to manage. In Beijing, officials are simultaneously encouraging the open-source ecosystem and building an export-registration system for frontier models, meaning labs will need government sign-off before releasing their most sensitive work abroad. DeepSeek’s position is now unambiguous: it will keep publishing, and it considers the faster arrival of AGI a benefit to everyone, regardless of which company gets there first.
