I think it's more likely the the space of "good" 2-4 phoneme github accounts is almost fully saturated, and the model is good at generating "plausible" real-sou

I think it's more likely the the space of "good" 2-4 phoneme github accounts is almost fully saturated, and the model is good at generating "plausible" real-sounding usernames, leading to a lot of collisions.