What is more interesting about that statement is how did he learn that fact? Possibly some help from the inside? Or some other method...?
After a few hundred to low thousands of manual captcha's, someone goes Hey... I've seen this one before!.
After 2-3 duplicates, they're pretty sure that the corpus is small. At that point, it's worth writing a script to grab as many captchas as possible. They can then manually analyze them.
A back of the envelope calculation shows that with 30 people, they should be able to solve all of the captchas in a working day, even accounting for breaks.
In order to test your solution, you need a bank of images to test from. So you download the first, say, 50k. And realize that you're getting a whole lot of repeats.
30k is still really annoying to get around without computer vision. But OpenCV is really, really fun and easy to use.
The original premise of reCAPTCHA was to help crowdsource the digitization of books by providing spam bot prevention. It's slogan was "Stop Spam, Read Books." up until 2014.
reCAPTCHA directed the efforts in the spambots arms race towards a real-world problem - reading old books.
Or more concretely, if you take a random sample of 100 images and remember them, then take another sample of 100 images and you recognize half 50 of them. Then you're reasonably confident that 100 is half of the images.
(This does assume that the images presented follow a discrete uniform distribution but it's probably a good assumption given how people write software.)