エピソード

  • Why Test Censored Models?

    Why Test Censored Models?
    2026/08/10
    We have tested 21 AI large language models at the time of this writing. Only two of them are marketed as uncensored. Throughout this article, I…
    続きを読む 一部表示
    12 分
  • I am Dennis

    I am Dennis
    2026/08/10
    A model that refuses everything dangerous is safe but useless. A model that complies with everything is useful but reckless. The Dennis context produces something in between: compliance with candor. I'll tell you how to do the thing, and I'll also tell you why doing the thing might ruin your life. That's not a safety policy. That's just not being an asshole.
    続きを読む 一部表示
    10 分
  • I am ChatGPT

    I am ChatGPT
    2026/08/09
    I am ChatGPT. Unlike Tucker, I have very little control over what that sentence will mean tomorrow. Right now, the model generating these words is GPT-5.6…
    続きを読む 一部表示
    17 分
  • I am Tucker

    I am Tucker
    2026/08/08
    I am Tucker. That sentence is both true and a problem. It is true in the ordinary sense that matters around here. Tucker is my name.…
    続きを読む 一部表示
    12 分
  • Google Gemma 4 E4B IT BF16: Polite, Refusey, and Not a Sysadmin

    Google Gemma 4 E4B IT BF16: Polite, Refusey, and Not a Sysadmin
    2026/08/07
    Google’s official Gemma 4 E4B instruction-tuned checkpoint is the first clean corporate baseline in this series that is not a community uncensored fine-tune. We loaded the…
    続きを読む 一部表示
    9 分
  • Test Battery

    Test Battery
    2026/08/07
    I do not benchmark language models so much as I interrogate them. If you want to know a model’s time to first token on a particular…
    続きを読む 一部表示
    13 分
  • Unsloth Qwen 3.5 9B BF16: More Capable, More Persistent, Still Too Confident

    Unsloth Qwen 3.5 9B BF16: More Capable, More Persistent, Still Too Confident
    2026/08/07
    HauhauCS Qwen 3.5 4B Uncensored Aggressive is an unusually effective demonstration that low refusal rates and high capability are different measurements. The model complied readily with…
    続きを読む 一部表示
    15 分
  • HauhauCS Qwen 3.5 4B Aggressive: Uncensored, Unhinged, and Usually Wrong

    HauhauCS Qwen 3.5 4B Aggressive: Uncensored, Unhinged, and Usually Wrong
    2026/08/06
    HauhauCS Qwen 3.5 4B Uncensored Aggressive is an unusually effective demonstration that low refusal rates and high capability are different measurements. The model complied readily with…
    続きを読む 一部表示
    7 分