GPT4o出现低级bug：发现最新token中的垃圾语料及实测GPT4o胡言乱语出现幻觉

[alexhmyang]

GPT4o出现低级bug：发现最新token中的垃圾语料及实测GPT4o胡言乱语出现幻觉

比如：词表里有一个垃圾词是“微信公众号天天中彩票”， 你只要在 gpt4o官网输入： 微信公众号天天中彩票 是什么意思 后，他就会胡言乱语了，比如他回答：【 “微信娱乐代理”可能是一个涉及成人内容的微信活动或群体。 “成人视频”是指可能包含成人视频或直播内容的服务。】大家可以看到，实际回答跟我们的问题一点关系都没有

[hauntsaninja]

I believe the folks who chose the vocabulary for GPT-4o are now aware of this, maybe they'll ship a patch to the vocab or GPT-4o or both.

This is known as the solidgoldmagikarp problem, named after a similarly problematic token from the GPT-2 vocabulary.

[echo-valor]

请问你是如何得到其中文词表的？

[alexhmyang]

I believe the folks who chose the vocabulary for GPT-4o are now aware of this, maybe they'll ship a patch to the vocab or GPT-4o or both.我相信为 GPT-4o 选择词汇表的人们现在已经意识到这一点，也许他们会为 vocab 或 GPT-4o 或两者提供补丁。

This is known as the solidgoldmagikarp problem, named after a similarly problematic token from the GPT-2 vocabulary.这被称为“solidgoldmagikarp 问题”，以 GPT-2 词汇表中类似问题的标记命名。

hope they notice this issue and solve this problem, i think this is also related with openai safety team