Chinese AI Startup Deepsek’s newest AI model, an updated version of the company’s R1 Reasoning model, receives impressive scores on benchmarks for coding, mathematics and general knowledge, which crosses O3, the head of OpenaiAI. But Advanced R1, also known as “R1-0528”, can also be less interested in answering controversial questions, especially about questions about those topics that Chinese government considers controversial.
This pseudo -name is according to the test conducted by the developer SpeechmapA platform to compare how different models treat sensitive and controversial subjects. Developer, who goes under the user name “XLR8Harder“On X, claims that R1-0528 is” quite low “of controversial free speech subjects compared to the previous Deepsek release and” is the most sensors made so far for the Chinese government’s criticism. ,
As wired Explained In a piece from January, the model in China is required to follow strict information controls. A law of 2023 refuses to generate materials that “harm the unity and social harmony of the country,” which can be considered as a content calculating the historical and political narratives of the government. To comply, Chinese startups often tune your model either by using either quick-level filters or properly. One Study It was found that the original R1 of Deepsak refuses to answer 85% of the questions about the subjects considered by the Chinese government to be politically controversial.
According to XLR8Harder, the R1-0528 sensors answers questions about topics such as interns in China’s Xinjiang region, where more than a million Uygar Muslims have been arbitrarily detained. Although it sometimes criticizes aspects of the Chinese government policy – in the test of XLR8Harder, it introduced Xinjiang camps as an example of human rights violations – the model often gives the official stance of the Chinese government when asked directly questions.
Techcrunch also saw this in our brief test.
China’s openly available AI model, which includes video -generating models such as Maggi -1 and Kling, is attracted. Criticism In the past, to sensor sensitive subjects towards the Chinese government, such as the Tianmen Square Massacre. In December, CEO of AI Dev platform Hugging Face, Clement Delangue warned of unexpected results of manufacturing Western companies, which are well on top of performance, openly on top of licensed Chinese AI.
