Show HN:CompareGPT – 通过比较多个 LLM 来发现幻觉
1 分•作者: tinatina_AI•9 个月前
大家好,我是 Tina。
在使用 LLM(大型语言模型)时,我一直遇到的一个令人沮丧的问题是“幻觉”:模型给出的答案听起来很自信,但却是编造的。比如虚假的引用、错误的数字,甚至是整个“系统报告”。
因此,我一直在开发 CompareGPT,它试图通过以下方式提高 AI 输出的可靠性:
* 并排展示多个 LLM 对同一查询的回答
* 方便查看答案的一致性(或不一致性)
* 帮助在浪费时间或造成损害之前发现幻觉
链接在这里:[https://comparegpt.io/home](https://comparegpt.io/home)。我们开放了候补名单,非常欢迎大家提供反馈,尤其是来自从事 LLM 研究、金融或法律领域工作的人。
谢谢!
查看原文
Hi HN I’m Tina.
One frustration I keep running into with LLMs is hallucinations: answers that sound confident but are fabricated. Fake citations, wrong numbers, even entire “system reports.”
So I’ve been building CompareGPT, which tries to make AI outputs more trustworthy by:
Putting multiple LLMs side by side for the same query
Making it easy to see consistency (or lack of it)
Helping catch hallucinations before they waste time or cause harm
link here:<a href="https://comparegpt.io/home" rel="nofollow">https://comparegpt.io/home</a>. We’ve opened a waitlist and would love feedback, especially from folks working with LLMs in research, finance, or law.
Thanks!