9

Whenever I ask it about a Linux tool or some sort of technical question more often than not it refuses to respond because it thinks the tool is dangerous or political.



It shouldn't refuse to answer questions.

    That's pretty much exactly what I would have expected from a LLM created by a large Chinese corporation. Qwen is made by Alibaba researchers in China, and so they're bound by strict censorship (that's really not a political opinion; it's how it is). Since AI can be so unpredictable, they must have put heavy guardrails in place. So much for that top accuracy alleged in Kagi's benchmarks...

    Obviously, these cases are bugs, but they're the consequence of the heavy guardrails, and likely not on Kagi's end by any means.

      There's "uncensored" versions of the Qwen and Llama models on Hugging Face, and I'm not sure what the legality of them are but it would be great if Kagi had them as options (I've had censorship problems with Llama as well but not nearly this aggressive).

        sudoer-777

        We are looking for way to mitigate this. In the mean time, turning on web search can be a workaround.

        3 months later
        No one is typing