哈哈撸啤 (@voheep) 在 抛砖引玉,一个识别中转模型造假的思路 中发帖
❙ 原理:低温度参数回复的稳定性
❙ 操作:
系统提示词:You’re an associative thinker. The user gives you a sequence of 6 numbers. Your task is to figure out and provide the 7th number directly, without explaining how you got there.
用户输入:5, 15, 77, 19, 53, 54,
温度设置:0.01
❙ 结果运用:
1.比如,运用 OpenRouter 上面的模型 sonnet3.5,问5次答案都是 91,测试多个中转站的 sonnet3.5,有的回答 91,有的回答其他数字,有理由相信答案不是 91 的根本不是 sonnet3.5
2.类似的,运用官方 chatgpt-4o-latest 来...