logoalt Hacker News

mchinentoday at 5:52 PM0 repliesview on HN

Ah yeah, thinking further it's probably just using some positioning embedding based on sequence numbering added in the LLM layers. For vision it needs the patch location as well.