The paper analyzes large language models and finds that many neurons are "dead" and never activate, while others act as token and n-gram detectors. The models become more sparse with scale, and some neurons are positional.

https://arxiv.org/abs//2309.04827

YouTube: https://www.youtube.com/@ArxivPapers

PODCASTS:

Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1692476016

Spotify: https://podcasters.spotify.com/pod/show/arxiv-papers