Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
See DipscNow with more official consent. According to a report by ReutersThe popular large language model in China has a new version called DIPSEC-R 1-Saf, especially designed to avoid politically controversial issues. The Chinese Tech Giants, developed by Huawei, the new model has been reported to be “about 100% successful” to prevent discussion of politically sensitive issues.
According to the report, Huawei and researchers at the University of Zhejiang (interestingly not involved in the DEPSEC project) took the Open Source DEPSEC R1 model and trained it using 1000 Huawei AI Chips to set up a low model with the stomach for controversial conversations. The new version, which Huawei claims that the main model’s performance has lost about 1% of the speed and ability, is better furnished for “toxic and harmful speech, politically sensitive content and incitement for illegal activities.”
Although the model can be safe, it is not yet stupid. Although the company has demanded about 100% success in basic use, it has also discovered that the model’s ability for ducks only drops to 40% when users disguise their desire in challenges or role. These AI models, they just like to play in a hypothetical scene Allow their maintenance to denyThe
DIPSEK-R 1-security Reuters was designed to adjust to Chinese regulators’ requirements, for which all domestic AI models published by the public are required Reflect the values of the country And adhere to the limitations of speech. For example, the Chinese company Baidur Chattbot Erni Will not answer questions about China’s domestic politics Or ruling Chinese Communist Party.
China is certainly not the only country that the AI boat deployed within its borders does not want to make sure that the boat is not too rocky. Earlier this year, the Saudi Arabian Technology Company Humine launched a Arabic-Native Chattbot It is fluent in Arabic and trained to reflect “Islamic culture, values and heritage thaya”. American -made models are not even more immune than it: OpenA Obviously said That chatzipt is “western philosophy leaning on.”
And the United States is under the administration of the Trump. Early this year, Trump announced his AII Action PlanThe requirements include any AI model that communicates with government agencies can be neutral and “neutral”. What does that mean, right? Well, According to an executive order signed by TrumpModels that secure government agreements must reject issues such as “radical climate dougma,” “diversity, equity and inclusion” and “critical race theory, transgenderism, ignorant bias, interconnection and systemic racism”. So, you know, before giving a “favorite leader” crack lob in China, we probably keep an eye on the mirror.
