跳到正文
Turing Post (Ksenia Se)· @TheTuringPost · X·· 3 小时前AI 评分38
AI 导读

OpenAI提出基于困惑度的透明度机制,旨在通过量化模型预测能力来确立共享的安全基准。该方案主张在无法充分理解AI行为前,限制超越特定困惑度阈值的训练进度,从而为前沿模型开发划定安全边界。

正文 · 原文

Do you know what connects Jakub’s proposal to a guessing game Claude Shannon ran in 1951?

Shannon asked people to guess the next letter in a piece of text. How many guesses it took helped him estimate how predictable English was. In the 1970s, Fred Jelinek’s team at IBM used a related measure, perplexity, to describe how hard a speech recognition task was.

What I find fascinating is that Jakub is still asking how well you can predict what comes next. This time, he wants the answer to help AI labs decide when to slow down.

引用Jakub Pachocki@merettm
We decided to provide more transparency into pretraining progress at OpenAI for the sake of moving toward shared safety bars. In “Perplexity and Intelligence” I argue for common limits based on perplexity and other measurements, with progress beyond them gated on our ability to understand AI behavior. https://openai.com/index/perplexity-and-intelligence
在 X 查看被引用的帖子

来源:Turing Post (Ksenia Se) · x.com