Mitchell A. Gordon

When I first read Google’s RETRO paper, I was skeptical. Sure, RETRO models are 25x smaller than the competition, supposedly leading to HUGE savings in training and inference costs. But what about the new trillion token “retrieval database” they added … Read more