Google's Griffin paper is a PERFECT demonstration of scaling laws. Roughly 10% improvement in performance on tasks as they scaled parameters by 7x. All trained on the same 300 billion tokens data. Models get more sample efficient or extrapolate better as you scale across… https://t.co/1eNZL0ZHhD https://t.co/gmBL9Od0fL
— Rohan Paul (@rohanpaul_ai) Apr 21, 2024
from Twitter https://twitter.com/rohanpaul_ai
April 21, 2024 at 11:37AM
via IFTTT
No comments:
Post a Comment