Show HN: CompareGPT – 值得信赖的 AI 回答,信心十足且有来源

2作者: tinatina_AI9 个月前
大家好,我是 Tina。 我一直在研究如何提高大型语言模型的可靠性。一个长期存在的问题是“幻觉”——模型会自信地给出在事实上有误或基于不存在的来源的答案。这对于金融、法律或研究等需要精确性的领域来说,尤其具有风险。 为了解决这个问题,我一直在构建 CompareGPT,它专注于提高 AI 输出的可信度。 我们一直在努力的主要更新包括: * **置信度评分:** 每个答案都会显示其可靠程度。 * **来源验证:** 突出显示数据是否可以被参考文献支持。 * **多模型比较:** 提一个问题,并排查看不同模型的回答。 在这里试用:[https://comparegpt.io/home](https://comparegpt.io/home) 目前,它最适用于基于知识的查询(金融、法律、科学)。我们仍在解决一些限制——例如,目前还不支持图像输入。 我很乐意听取您的反馈,特别是关于它在哪里失效或在哪里最有用的地方。欢迎提出尖锐的意见! 谢谢!
查看原文
Hi HN, I’m Tina I’ve been exploring how to make large language models more reliable. One persistent issue is hallucinations — models can produce confident answers that are factually wrong or based on non-existent sources. This is especially risky for fields like finance, law, or research where accuracy matters. To address this, I’ve been building CompareGPT, which focuses on making AI outputs more trustworthy. Key updates we’ve been working on: Confidence scoring: every answer shows how reliable it is. Source validation: highlights whether data can be backed by references. Multi-model comparison: ask one question, see how different models respond side by side. Try it here: <a href="https:&#x2F;&#x2F;comparegpt.io&#x2F;home" rel="nofollow">https:&#x2F;&#x2F;comparegpt.io&#x2F;home</a> It currently works best with knowledge-based queries (finance, law, science). We’re still ironing out limitations — for example, image input isn’t supported yet. I’d love to hear what you think, especially where it fails or where it could be most useful. Brutal feedback welcome Thanks!