The problem of LLMs

 

https://youtu.be/846JNIM_4MQ?si=_8DLF6ulyRSh6TW7

All we call model improvement today's at its core, just improving the model’s capacity to simulate within a narrow normative space: controlled natural language, specific tasks, user satisfaction, safety constraints, and so on

The model does not build a world of its own the way a human does, nor does it develop an original perspective on the world; what develops is a more and more polished ability to reassemble our perspective and feed it back to us in a smoother, more convincing form

In this sense, we are not moving along a trajectory toward a post-human with a stance, an inner temporality, and a life story, but toward a highly polished intelligence that is increasingly skilled at playing the role of a human in front of us (and its sucks)