Dario Amodei, CEO of Anthropic, said that the organization does not support a complete ban on artificial intelligence models with open weights, but insists on the need for strict supervision of their development and distribution. This position was announced against the background of the publication of an open letter from twenty-five technology companies, including Nvidia, Microsoft and Meta, calling on the US authorities to refrain from restrictions on such decisions. It is noteworthy that Anthropic, OpenAI and Alphabet were not among the signatories, although Demis Hassabis, a representative of Google DeepMind, promptly clarified that the corporation continues to actively support the development of the open ecosystem and science, as evidenced by the widespread use of the Gemma family of models.

In his article, Amodei emphasized that models with open weights that do not have dangerous capabilities bring significant benefits to businesses, developers, and researchers. Nevertheless, he identified two critical risk scenarios: the possibility of authoritarian governments, including China, obtaining models superior to their American counterparts to achieve military superiority, as well as the use of advanced neural networks for cyber attacks or the creation of biological threats. The head of Anthropic noted that a simple ban on the use of Chinese models with open weights by American companies does not eliminate these threats, but only reduces competitive pressure on local developers without hindering the actions of potential attackers.

Instead of total bans, the head of Anthropic proposed to implement three specific measures. First, it is necessary to tighten restrictions on the supply of powerful computing chips and equipment for their production to China, actively countering smuggling. Secondly, the practice of industrial distillation should be stopped, which, according to the expert, makes it possible to reduce the technological gap between Chinese and American developments to several months. Thirdly, it is necessary to implement mandatory independent security testing for all sufficiently powerful models, regardless of the type of licensing. These proposals are becoming particularly relevant against the background of reports by journalists about cases where commercial chatbots provided instructions on how to create dangerous substances, which underlines the need for a balance between openness and safety.