Hacker News new | past | comments | ask | show | jobs | submit login
Studying Large Language Model Generalization with Influence Functions (arxiv.org)
6 points by tosh 62 days ago | hide | past | favorite | 1 comment



So why do LLMs generate sci-fi like responses to questions regarding 'their' intelligence? Because they're sampling from sci-fi.

It's actually quite a funny result for silicon valley's sci-fi obsession: the sci-fi tech elite created a system to reassure themselves of a sci-fi future because it sampled from the very sci-fi they were reading.

A wonderful example of the 'hot reading' mirroring effect of certain form of human-computer interaction, ie., those where we are ourselves surreptitiously providing the information being fed back to us.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: