Comment by vov_or

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)