@simon Amazing! That's the hands-on practical work we need right now!

It's fascinating that a small team can come along, use a few tricks of the trade, train it on eight GPUs, and massively push state of the art. That's similar to what happened with resnet in the vision domain where folks went back a few times over the yeas and showed how it can be improved considerably based on what we learned.

Looking forward to having this power a ColBERT-like retrieval setup!