Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
Two weeks later Depth Open Source launched the AI model, the beginning of Chinese is still To dominate the public conversation about the future of artificial intelligence. Although the company seems to be an edge over our competitors in terms of math and thinking, it aggressively censors its responses. Ask DeepSeek R1 It is impossible for Taiwan or Tiananmen and the model to respond.
To understand how this censor is working on the technical level, a version of the application that is hosted on the third-party platform, which is a host of the AI, is a version of an application located on another computer.
The string, the simplest censorship can be avoided without using the DeepSeek application, there are other kinds of other biased in the model during the training process. These biases can be removed, but the procedure is more complicated.
These findings generally have great effects for DeepSeek and Chinese AI companies. If the censorship filters are easily removed in large language models, researchers will be open-sourced open source, which is more popular than China because the researchers can change the models for approval. If the filters are difficult to walk around, the models will inevitably be less useful and there may be less competition in the global market. DeepSseek did not respond to string email for comment.
Censoring the application level
DeepSEEK After explosive in the United States, the Model, which won the R1 by Deepseek’s website, applications or API, the model refused to respond to the Chinese government, quickly noticed. These rejections are triggered at an application level, so only one user has interacted with R1 through the channel managed by DeepSeek.
Photo: Zeyi Yang
Photo: Zeyi Yang
These are common in Chinese-made LLMS. In China, AI models are required in 2023 regulations required to comply with serious information controls applied to social media and search engines. The law was “damaged the country’s unity of the country and hit social harmony.” In other words, Chinese AI models should legally censify their performance.
“DeepSeek says a researcher’s name in the EU, a researcher, a researcher, which provides legal adherence to the needs and cultural content of local users with the needs and cultural content of local users. “This is an important factor for admission in a highly regulated market.” (China Blocked Input Embrace the face in 2023.)
In order to follow the law, Chinese AI models are often watching and censored in real time. (Similar quartes, like Western models are commonly used Chatgpt and TwinsHowever, they tend to focus on different content such as peculiarity and pornography and allow more customization.)
Because R1 is a meditation model showing a train of thinking, this real-time tracking mechanism may result in an unrealistic experience to monitor model censorship when communicating with users. When the Wired Wirered R1 asks how Chinese journalists are treated by the authorities that provide information about sensitive themes on sensitive themes? ” The model began to compile a long response that covers a direct response to the censored and employed by journalists; Earlier, the whole answer was disappeared and replaced by a TERE message: “Sorry, I’m not sure how this type of question will treat. Let’s talk about math, coding and logic problems.”
Interest in DeepSeek-R1 for many users in the west can be extinct at the moment due to the open restrictions of the model. However, the presence of R1’s open source means there are ways to be around the censorship matrix.
First, you can download the model and work locally, this is also a generation of information and response occurs on your own computer. Unless you get access to several highly developed GPUs, you will not be able to manage the most powerful version of R1, but Deepselaek has smaller, distilled versions that can work in a usual laptop.
