A model that avoids answering all sensitive questions

Huawei in China, an upgraded version of artificial intelligence Deepsic Introduced to the aim of increasing Content safety Designed and, according to the company, it is capable of almost completely discussing topics Political sensitive Avoid.

China’s strict rules require that all models of internal artificial intelligence and their related software, before public publication, reflector The values ​​of this country Be careful and proper monitoring of the content produced.

Huawei’s collaboration in enhancing the Censorship capabilities of the Deepsic model

Huawei has announced in its official account that this large language model utilizes Thousands of ascend ai chips Trained. The base version of this model Deepsic-R1 Have been open and reforms to increase Content safety Applied on it. The prestigious University of Zajiang, the founder of the Depsic founder, Liang Vanfong, Huawei’s partner in the project, has not had a direct role in its development, and Deepsic has not yet responded to requests to comment on the project.

Huawei has an upgraded version of this model called Deepseek-R1-SAFE It has offered that its main focus is on protecting users from harmful content. Experiments show that this model is almost in identifying and preventing “harmful words, political content and encouraging illegal behaviors”. One hundred percent It works successful. However, when these behaviors are hidden in the form of simulated scenarios, role -playing or encrypted codes, the model’s performance is reduced to 40 %.

Huawei has announced that the comprehensive defense capabilities of the Deepseek-R1-SAF model 83 % Ripe and compared to similar models such as QWen-235B and Deepseek-R1-671B, between 8 to 15 percent better performance It has. Also, reducing the performance of this version than the original model of less than 1 percent Reported.

Chinese artificial intelligence chats such as Ernie Bot from Baidu, equivalent to ChatGpt in China, are reluctant to answer many of the political questions and sensitive issues that the Communist Party controls.

At the annual Huawei Connect conference in Shanghai, Huawei, after years of officially introduced programs for chip production and computational power development, showed how to develop advanced and secure artificial intelligence within the country’s internal law.