H hanlab.mit.edu How attention sinks keep language models stable 219 points pr337h4m 1 year ago 36 comments
