A lot of tokenization bugs came from odd Reddit subreddits, such as r/counting. And back then, we didn't have free model weights, which is they all converge to

A lot of tokenization bugs came from odd Reddit subreddits, such as r/counting. And back then, we didn't have free model weights, which is they all converge to similar weights, might tell us a lot about how to design adversarial examples.