تسجيل الدخولتسجيل
user avatar
zer0int (it·its)
@zer0int1
AI & I do prompt engineering towards prompt criticality. e/acc
no u
انضم في August 2022
218
متابَع
441
المتابِعون
الردودالوسائط
  • مُثبّت
    user avatar
    zer0int (it·its)
    @zer0int1
    25 يناير
    'It'll just fine-tune itself now' for #CLIP and #LongCLIP 🤖🫶🤖 + dataset heuristics -> enter path, ill figure out the rest + load any model / dataset, local or HuggingFace + Get my Regression models : ~90% typographic attack acc.! + A huge #AI #mechinterp ramble. Er, paper.
  • user avatar
    zer0int (it·its)
    @zer0int1
    4 أغسطس
    GPT-5.6 Sol @ long context: *produces mangled code* 🤔: Sol, are you sure you know what you are doing, still? Rate your 1. Confusion 2. Hallucination; add 3. Problem statement. Seems grounded in the reality of the code generated code. Impressive self-reflection. 🙃
  • user avatar
    zer0int (it·its)
    @zer0int1
    2 أغسطس
    Correct gating (left) vs. word-hallucination in object. l.t.r: Patch L2 norm (registers; backbone), cross-attention gate (has layernorm!) glyph logit, glyph prob. No backbone components found guilty. But there's some really weird scheme in the cross-attention bridge softmax. 😂
  • user avatar
    zer0int (it·its)
    @zer0int1
    1 أغسطس
    Q: So what did you do on the weekend? A: I tried to find some attention heads responsible for edge-case (~1 / 1000 images) failure of null-gating a hallucinated pizza-as-a-word <text> when adding an orange peel to a granny smith in the retina of a vision transformer.