You’re right. I understood it to require taking the top 2^30 tokens, but instead they sample 2^30 times with replacement. Too bad they only formulate the detec

You’re right. I understood it to require taking the top 2^30 tokens, but instead they sample 2^30 times with replacement. Too bad they only formulate the detection positive rate empirically. I am curious what the exact probability would be mathematically.