Tencent's AI Chief Warns Models Will Soon Master Biology
Vinces Yao Shunyu, Tencent's chief AI scientist and a former OpenAI researcher, has warned that artificial intelligence models will soon be capable of doing harm in biology the way they already can in cybersecurity, and that the industry is not ready for it.
The cyber abilities of models and soon the biological abilities of models will reach a threshold where harm becomes possible, Yao said at a science symposium jointly hosted by the Hong Kong Academy of Sciences and the Future Science Prize, in a rare public appearance for a researcher who joined Tencent from OpenAI last year.
And the timing is pointed. Tencent released Hy4, its new flagship model, a few days before the speech, and benchmarking firm Artificial Analysis ranked it the second best model produced in China, which puts the warning in the mouth of the person responsible for the model rather than an outside critic.
Tencent's own safety claims for Hy4 are specific. The company says it removed the model's offensive cyber abilities while keeping its defensive ones by filtering training data, and that Hy4's biological capability sits well below its high risk threshold because the model still cannot turn textbook biology into an actionable protocol.
Yao did not leave the risk at the level of principle. He said the gap between Chinese models and their American counterparts is not large, that biological capability will be harder to manage than cyber capability, and that the attack and defence sides of biology are harder to pull apart, which is why he wants evaluation done carefully before any release.
But the argument he made about openness is the one that matters commercially. He said Anthropic and other American labs have made a strong case that frontier models should stay closed, then argued that open weight models also help defenders, and pointed to an incident where Hugging Face used a Chinese open model to defend itself against a cyberattack after closed American models refused to help because of their guardrails.
His conclusion was a mixture rather than a side. Yao said the best way forward is probably a mix of closed and open source models, because defenders need access to the same technology, and he described Tencent's own process as collaborating with the company's security labs on evaluation while admitting it is not complete yet.
That position is easier to hold when the stakes are cyber. Biology is different, and Tencent has already signed up to the view that it is a public health question, having contributed with Huawei and Alibaba to a frontier safety report at the Wuzhen summit that named pathogen design as a capability potentially creating major public health risks.
So the test of this speech is not the phrasing but what follows it. Tencent has promised to publish the weights of Hy4 along with a system card setting out its cyber and biological evaluations, and the value of that document depends on how much of the testing an outside lab can repeat, and whether the numbers in it are red team results rather than a summary of internal review.
Vinces Yao Shunyu is Tencent's chief AI scientist and a former OpenAI researcher
He said models will reach a threshold where biological harm becomes possible
Tencent says Hy4 has offensive cyber abilities removed and keeps defensive ones
Tencent says Hy4's biological ability is well below its high risk threshold
Yao argued for a mix of open weight and closed models rather than one or the other
Source: South China Morning Post
