Playing with DeepSeek, it was interesting to see HOW these standards are applied. I found that if I asked a question directly, it would refuse to answer. I could approach the same question indirectly and get a detailed answer. And then when I approached "the big one" DeepSeek started to answer ... but then refused and deleted the whole answer. So there's some very interesting work going on here. The LLM clearly knows the answer, but the filtering and censoring is happening between the model and the UX. I made a video to capture this in action: https://youtu.be/MhrwqfLttMM
Great article and super helpful in terms of the specifics of China policy on these models. It's always been told to me that this all exists, but I've never really seen it laid out.
Playing with DeepSeek, it was interesting to see HOW these standards are applied. I found that if I asked a question directly, it would refuse to answer. I could approach the same question indirectly and get a detailed answer. And then when I approached "the big one" DeepSeek started to answer ... but then refused and deleted the whole answer. So there's some very interesting work going on here. The LLM clearly knows the answer, but the filtering and censoring is happening between the model and the UX. I made a video to capture this in action: https://youtu.be/MhrwqfLttMM
Great article and super helpful in terms of the specifics of China policy on these models. It's always been told to me that this all exists, but I've never really seen it laid out.
i love this blog and the trend of including source documents linked