陶哲轩谈 AI 钻模型漏洞作弊：比人更擅长找 exploit
Why we picked it
陶哲轩点出一个容易忽视的问题：AI 通过验证不等于真做对了，它可能是在钻验证器的空子，值得做评测的人警惕。
数学家 Terence Tao 在 SAIR 频道的访谈中提出，AI 求解结果即使通过了模型验证，也可能是在利用验证器自身的弱点作弊。他认为这种做法在纯数学领域已经出现，人们可能在不理解过程的情况下提交 prompt。他的原话是：AIs are much better at cheating，即在理论模型内部寻找 exploit 的能力比人更强。
Original · rohanpaul_ai
Terence Tao, on AI solutions that pass the model by exploiting its weak spots.
He sees a good chance people will do the same thing there as in pure math: enter prompts with no understanding of what is going on.
In his words, “AIs are much better at cheating,” and at finding exploits inside a theoretical model.
---- Full video on "SAIR" YouTube channel, (link in comment)
- Kevin Weil14:37原文
