All news

Tencent chief scientist warns frontier models are approaching biological risks

Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .

As models gain biological capabilities, the decision to publish open weights increasingly depends on risk testing, not just model quality.

According to the South China Morning Post, relayed by The Next Web, Yao Shunyu told a science event in Hong Kong that models' biological capabilities could approach a threshold of practical harm, as has already happened in cybersecurity. He urged Chinese companies to take safety seriously when releasing open models and said biological risks could be harder to manage than cyber risks. The speech followed the launch of the Hy4 model on November 4. Tencent said it had excluded material on exploiting vulnerabilities from the model's training and used defensive security data instead; according to the company's internal assessment, Hy4 has basic biology knowledge but is weaker than leading models at designing experiments. Yao acknowledged Anthropic's arguments for keeping frontier models closed but proposed combining open and closed development because open models can also help defenders.

History

Tencent Hy4 launch and assessment of its risks

3 events
  1. Tencent Hunyuan releases Hy4 flash: 12 billion activated parameters, 1M-token context and 0.5 yuan per million input tokens
  2. Tencent chief scientist warns frontier models are approaching biological risksOpen now
  3. Tencent releases Hy4 with vision and software control, promising weights within days