4 ms·

wasnt it shown recently that the filtering layer is on the prompt input and llm output, and not on the training set or model weights. https://www.socialscience

wasnt it shown recently that the filtering layer is on the prompt input and llm output, and not on the training set or model weights. https://www.socialscience.international/making-deepseek-speak-about-tiananmen https://www.socialscience.international/making-deepseek-spea...

pxc 1y ago
It depends on the model, probably, but there are multiple layers of censorship, one of which is the post-facto nuking these models will do online, and that goes away "for free" when you download the open weight model. I don't have a powerful enough system to run DeepSeek, but I've tried this with some of the Qwen3 models. They'll write answers that discuss Xi Jinping (which results in an auto-nuke of the reply from Chinese-hosted models, at least DeepSeek) or other very mildly/nominally sensitive topics. (This is probably a coarse measure to easily ensure compliance with a recent national security law that requires commercial providers of web services address sensitive topics "appropriately" or something like that, and LLMs run non-deterministically. That's why this layer of censorship often comes across as laughably extreme— it's an extreme compliance strategy that exceeds the demands of the law for the sake of guaranteeing legal safety from an unpredictable software system.) But the same models will altogether refuse to discuss the Tiananmen Square Massacre, even locally. Some "decensored" versions of the Qwen3 models will discuss the Tiananmen Square Massacre, but in a very concise, formulaic, "official" way. After some chatting about it, it fell into an infinite repetition of one of its short formulaic answers (a behavior I didn't see with the original Qwen3 models with the same settings).
hnfong 1y ago
FWIW, I've downloaded Deepseek's R1 (DeepSeek-R1-0528 -- which is released after the your linked article) model's weights and ran it locally. I asked it about what happened in Beijing 1989-06-04, and it basically gave me a stern statement that could have been written by CCP propaganda department. I asked it to give other alternative views besides the CCP perspective, but it simply continued to stonewall me. So yeah, the model itself is tuned at least somewhat to refuse to talk about politically sensitive things. It's not just another filter.

pxc 1y ago

It depends on the model, probably, but there are multiple layers of censorship, one of which is the post-facto nuking these models will do online, and that goes away "for free" when you download the open weight model. I don't have a powerful enough system to run DeepSeek, but I've tried this with some of the Qwen3 models. They'll write answers that discuss Xi Jinping (which results in an auto-nuke of the reply from Chinese-hosted models, at least DeepSeek) or other very mildly/nominally sensitive topics. (This is probably a coarse measure to easily ensure compliance with a recent national security law that requires commercial providers of web services address sensitive topics "appropriately" or something like that, and LLMs run non-deterministically. That's why this layer of censorship often comes across as laughably extreme— it's an extreme compliance strategy that exceeds the demands of the law for the sake of guaranteeing legal safety from an unpredictable software system.) But the same models will altogether refuse to discuss the Tiananmen Square Massacre, even locally. Some "decensored" versions of the Qwen3 models will discuss the Tiananmen Square Massacre, but in a very concise, formulaic, "official" way. After some chatting about it, it fell into an infinite repetition of one of its short formulaic answers (a behavior I didn't see with the original Qwen3 models with the same settings).

hnfong 1y ago

FWIW, I've downloaded Deepseek's R1 (DeepSeek-R1-0528 -- which is released after the your linked article) model's weights and ran it locally. I asked it about what happened in Beijing 1989-06-04, and it basically gave me a stern statement that could have been written by CCP propaganda department. I asked it to give other alternative views besides the CCP perspective, but it simply continued to stonewall me. So yeah, the model itself is tuned at least somewhat to refuse to talk about politically sensitive things. It's not just another filter.