undefined | Better HN

0 pointsIanCal2y ago0 comments

It's not quantising existing models, they're training new ones.

0 comments

1 comments · 1 top-level

I understand this part but it seemed that the 16->8->4 etc is similar to compression of the "net" and seemed to lower quality below 8.

j / k navigate · click thread line to collapse