Hi! You are right that we had to clarify that "100 times better at retrieval". Btw, we have plans to tune models, evaluate, and publish results in different tasks (zero-shot ImageNet classification, etc)