During remarks at a science gathering in Hong Kong on Saturday, Yao Shunyu, who leads AI research at Tencent, cautioned that machine learning models are approaching a point where they could inflict real damage through biological applications, mirroring capabilities they already possess in cyber domains, according to reporting by the South China Morning Post.

"The cyber abilities of [AI] models and soon [the] bio ability of models will reach a threshold where it's actually possible to do harm," Yao said, according to the SCMP.

Yao emphasized that Chinese firms distributing open-source models—systems anyone can obtain and modify—must prioritize safety protocols. He also noted that biological threats could prove more challenging to contain than cybersecurity threats.

At 29 years old, Yao arrived at Tencent last year following a stint at OpenAI. His remarks came shortly after Tencent unveiled Hy4, its newest leading-edge model, on November 4.

When announcing Hy4, Tencent disclosed that it deliberately excluded material related to software vulnerabilities from the model's training dataset. Instead, the company trained the system using defensive security information. Tencent also noted that while Hy4 demonstrates solid foundational knowledge in biology, it underperforms compared to top competing models when it comes to experimental design, based on internal assessments by specialists. The model has not been independently tested by external reviewers.

According to Reuters, Tencent intends to make Hy4's underlying weights and documentation detailing its safety evaluation publicly available in the near future. Benchmarking organization Artificial Analysis ranks Hy4 as the second-strongest Chinese model, the SCMP reported.

While Yao acknowledged that American organizations like Anthropic present compelling arguments for restricting access to advanced models, he contended that a hybrid strategy combining both open and closed systems may prove optimal. He reasoned that open models serve a valuable function in strengthening defensive capabilities.

Source: The Next Web