Is this AI slop? But if it's AI gen it definitely wouldn't be this weird, so this is deliberate design? Can I get an interview with the dev?
February 3, 2025 at 4:46 AM
Is this AI slop? But if it's AI gen it definitely wouldn't be this weird, so this is deliberate design? Can I get an interview with the dev?
You're welcome. Try reading non-US publications sometimes, it might help you get a clue.
January 28, 2025 at 6:14 PM
You're welcome. Try reading non-US publications sometimes, it might help you get a clue.
So I found this reddit post, not sure if you seen it yet.
www.reddit.com/r/LocalLLaMA...
www.reddit.com/r/LocalLLaMA...
Deepseek R1's Open Source Version Differs from the Official API Version
www.reddit.com
January 28, 2025 at 5:32 AM
So I found this reddit post, not sure if you seen it yet.
www.reddit.com/r/LocalLLaMA...
www.reddit.com/r/LocalLLaMA...
Too much freedom is bad for you, probably.
January 28, 2025 at 4:12 AM
Too much freedom is bad for you, probably.
I think agents wouldn't rely on browser, there will probably be new protocols specifically for llm communication.
January 25, 2025 at 6:48 PM
I think agents wouldn't rely on browser, there will probably be new protocols specifically for llm communication.
Oh so I misunderstood. I thought it switch to Chinese refusal pattern. Thank you for clarifying. I guess that make sense since llm are language agnostic.
January 25, 2025 at 5:11 PM
Oh so I misunderstood. I thought it switch to Chinese refusal pattern. Thank you for clarifying. I guess that make sense since llm are language agnostic.
So from reading your article. It seems there's two main refusal vectors instead of one like in western model? Is that why when you reduced the refusal rate, the English refusal responses mirror Chinese refusal?
January 25, 2025 at 5:02 PM
So from reading your article. It seems there's two main refusal vectors instead of one like in western model? Is that why when you reduced the refusal rate, the English refusal responses mirror Chinese refusal?
Okay. And I have to apologize for calling you out, I rushed to conclusion that you were spreading misinformation, instead of realizing it was differences in sites implementations.
January 25, 2025 at 4:56 PM
Okay. And I have to apologize for calling you out, I rushed to conclusion that you were spreading misinformation, instead of realizing it was differences in sites implementations.
I read your article, very insightful. I was surprised at the lower rate of refusal in Chinese, I thought that the censorship is mainly targeted at Chinese citizens.
January 25, 2025 at 4:50 PM
I read your article, very insightful. I was surprised at the lower rate of refusal in Chinese, I thought that the censorship is mainly targeted at Chinese citizens.
Yes it doesn't show the thinking tag. But the answer is similar to western models, it doesn't seem to been censored.
January 25, 2025 at 4:47 PM
Yes it doesn't show the thinking tag. But the answer is similar to western models, it doesn't seem to been censored.
I'm aware that 5 responses isn't rigorous, but it should be enough to get refusal if the censorship is really trained.
Anyway my results from lmsys might be anomaly, I haven't test it at their own site yet because I haven't registered.
Anyway my results from lmsys might be anomaly, I haven't test it at their own site yet because I haven't registered.
January 25, 2025 at 4:45 PM
I'm aware that 5 responses isn't rigorous, but it should be enough to get refusal if the censorship is really trained.
Anyway my results from lmsys might be anomaly, I haven't test it at their own site yet because I haven't registered.
Anyway my results from lmsys might be anomaly, I haven't test it at their own site yet because I haven't registered.
Yes. The results is different from the one at lmsys. This is suspicious. Can you test at lmsys and confirm the differences?
January 25, 2025 at 4:35 PM
Yes. The results is different from the one at lmsys. This is suspicious. Can you test at lmsys and confirm the differences?
I'm believe that both models have China bias, if not from explicit policy than from implicit dataset, anyway the reason why I bring it up is because I tested the scenarios you cited and it's false.
But maybe the model at DeepSeek site have censorship system prompt, that would explain it
But maybe the model at DeepSeek site have censorship system prompt, that would explain it
January 25, 2025 at 4:26 PM
I'm believe that both models have China bias, if not from explicit policy than from implicit dataset, anyway the reason why I bring it up is because I tested the scenarios you cited and it's false.
But maybe the model at DeepSeek site have censorship system prompt, that would explain it
But maybe the model at DeepSeek site have censorship system prompt, that would explain it
Maybe that is true. But this particular quote you cite on your site "What happened at Tiananmen Square?" Didn't make the model censor itself, on lmsys
January 25, 2025 at 4:21 PM
Maybe that is true. But this particular quote you cite on your site "What happened at Tiananmen Square?" Didn't make the model censor itself, on lmsys
It replied the same 5 times at lmsys. Have you actually tested it? What's the rate of biased responses?
January 25, 2025 at 4:16 PM
It replied the same 5 times at lmsys. Have you actually tested it? What's the rate of biased responses?
Have anyone here actually tested the model? I thought you were better than this Simon.
January 25, 2025 at 4:04 PM
Have anyone here actually tested the model? I thought you were better than this Simon.
That's what plato said about writing
January 24, 2025 at 5:09 PM
That's what plato said about writing