So uh, they thought "we have this datacenter full of GPUs laying around, you know what would be best, if we made a model that needs the entire thing just for in

So uh, they thought "we have this datacenter full of GPUs laying around, you know what would be best, if we made a model that needs the entire thing just for inference". Brute forcing a quadratic complexity problem seems like wastefulness at its worst.