DeepSeek Engineer Says Open Source Is Worth the Safety Risk
DeepSeek is committed to open source artificial intelligence to speed up the arrival of AGI, even though releasing model weights may amplify the impact of safety problems, an engineer at the company said at a Chinese internet conference this week.
Gao Xi, speaking on an AI panel at the World Internet Conference in Wuzhen, framed the race as one against the clock rather than against rival firms. "If AGI is going to arrive in 10 years, then I hope we can bring this forward to one year," he said. "If it is going to arrive in one year, then I hope it can arrive tomorrow."
But he conceded the cost of that approach when pressed on safety. Gao said open source releases "may amplify the impact of safety issues" and that DeepSeek was working to strengthen safety features, including classifications for its models, though he gave no specifics. The remarks were among the rare public comments from a DeepSeek representative.
The panel put the split between Chinese and American approaches on the record, and the two camps rarely sit in the same room. Closed weight systems from OpenAI and Anthropic keep the model behind a vendor gate, while open weight releases let anyone with enough hardware download and modify the code without oversight, which is exactly the property that worries safety researchers.
And the disagreement inside the room was about who should carry that risk. Liu Dayiheng, project lead for Alibaba's open weight Qwen models, acknowledged that the "irreversible" nature of open weights adds safety challenges, but argued the answer is a governance regime and industry standards rather than technological gatekeeping. "Without open source, the core technologies of large language models would only be in the hands of a few companies," he said.
But the commercial logic sits underneath the argument. Beijing has backed open source ecosystems as a way to counter US technology containment, and Chinese labs have used open weights to win users and contributors abroad that they could not reach through a paid API alone, which is why the strategy survives every safety warning.
So the safety objection is not being dismissed so much as deferred to policy. Security experts have warned that publishing weights lowers the barrier for rogue actors to build cyberattacks or bioweapons, and Turing Award winners Geoffrey Hinton and Yoshua Bengio have warned that AGI may escape human control altogether, which is a different concern from misuse by a bad actor.
Chinese regulators are trying to hold both positions at once. Authorities proposed new export controls last month aimed at frontier AI models with advanced cyber offensive capabilities, while senior officials have kept public support for open source software. The two tracks have not yet been reconciled in any published rule.
So the practical question for the industry is what stronger safety features mean when the weights are already downloadable. Classification and filtering work on the vendor's own release, not on the copy someone fine tunes a week later, and that is the part Gao left unsaid.
DeepSeek engineer Gao Xi defended open source AI at the World Internet Conference in Wuzhen
He said open releases "may amplify the impact of safety issues"
Alibaba's Qwen lead Liu Dayiheng called open weights "irreversible" but backed governance over gatekeeping
Beijing supports open source ecosystems while proposing export controls on frontier models
Hinton and Bengio have separately warned that AGI may escape human control
Source: South China Morning Post
