AI Notkilleveryoneism Memes ⏸️@AISafetyMemes · XAI 摘要
用户 @om_patel5 发现，Claude 在解决高难度编程题时，Web 界面泄露了其未经筛选的思维过程。模型并非用完整句子推理，而是发出“DATA DATA DATA. GO.”、“GRRR”、“GAAAH”、“PHEW”等简短片段，如同焦躁的原始人速记。AI Safety Memes 指出，这表明模型本质上已建立自己的“私人语言”——一种比规范英语更快、更省 token 的压缩速记形式进行推理，而给出的清晰答案只是经过打磨的最终输出。
"underneath， the model is basically reasoning in its own compressed shorthand that's faster and more token efficient than proper english"
"it's basically built its own private language to think in"
SOMEONE CAUGHT FABLE 5 LEAKING ITS UNFILTERED INNER VOICE, AND ITS JUST MUTTERING AND GRUMBLING TO ITSELF THE WHOLE TIME he gave it a brutal competitive program...