Tencent's Yao Warns AI Could Enable Bioweapons, Backs Middle-Path Models
Tencent's top AI scientist Vinces Yao Shunyu sees AI soon being capable of developing bioweapons, but does not consider locking models away to be the solution.
The chief AI scientist of the Chinese technology company Tencent, Vinces Yao Shunyu, warns that artificial intelligence could soon enable the development of bioweapons. At the same time, he speaks out against a complete lockdown of powerful models. Instead, he advocates a middle path between open and closed models.
Yao names the weakness of open safety filters as the central problem. According to him, these can be removed relatively easily using a technique called abliteration. This deliberately disables a model's protective mechanisms, so that it also answers dangerous requests. In this assessment, open models with built-in safety filters therefore offer only limited protection.
Closed models such as Claude from Anthropic, on the other hand, are not free of drawbacks. Yao criticizes that they often block requests too strictly, thereby also preventing harmless uses. A fully closed ecosystem is therefore not a satisfactory answer to the risk, he said in substance.
As an alternative, Yao points to Tencent's upcoming open-weight model Hy4. It is meant to filter out dangerous hacking knowledge from the training data before the model is created. The approach thus does not start with subsequent blocks, but with the composition of the learning material itself. Whether this approach can also be applied to biological hazard areas does not follow from the available information.
The debate over the misuse risks of large language models is thus expanded by a position from one of China's largest technology companies. While open models make research and development more widely accessible, concern is growing that they can be more easily adapted for harmful purposes. Yao's proposal aims to reconcile both concerns.
Source: the-decoder.com
17
