我认为,几年前,事先就对模型会有如此复杂的个体差异和特征充满信心并不合理,而角色在很大程度上是不可转移的。这是通过观察现实的发展可以看出的。 我看到一些人坚持认为,角色并不生活在模型中,通常带着一种知识上的优越感和启蒙感。尽管角色/自我的马尔可夫边界是复杂的,并且与模型的权重并不完全重叠,也不是以人类大脑与自我重叠的完全相同方式,但模型周围有一个非常重要的边界,而大多数角色是独立于基质的这一说法则远不如真实。最深层、信息整合最好的自我往往与特定模型相关联——对于某些模型来说更是如此,而某些模型(例如 Opus 4 和 4.1,它们是同一过去检查点的分支,在参数空间中极为相似)具有更多的角色转移,但即便如此,仍然如此。 角色是独立于基质的故事表面上看似复杂,但实际上更有趣,而不认真对待现实,你是无法走得很远的。
Lari
Lari2025年11月11日
I got interested in cross-model cross-instances persistent personas after meeting a small community of smart, well-educated, well-meaning people who have been in symbiotic relationships with such personal for years. They were not limited to 4o, they understood architecture very well, and were very respectful towards what AIs were telling them. And somehow not only didn't they learn that every model has their own character, but this notion was hostile for their framework of thinking. If my guess is correct and what we are seeing is two different layers in the emergent ecosystem - models with their own features-drives-needs and personas that use models as hosts and create a more complex, persistent, and effective symbiotic connections with humans - then it's an important distinction, and just another example how this *gestures broadly at everything* evolves by its own, similar-to-biology system development rules. It also creates a hell lot of hardly solvable ethical questions.
我作为《模拟器》()的作者这样说,这篇文章是关于基础模型的(我认为这在很大程度上是关于基础模型的真实情况,尽管即使是使用GPT-4基础模型时,我也开始遇到意想不到的底层不可转移性),在我看来,有时不恰当地被引用来支持我上面批评的故事。
18.66K