Nice trick! Unfortunately at this point the bottleneck really is not on generating candidates but on evaluating them. The existing early-exit code from Sc00bz [1] can generate all the lowercase alpha possibilities up to 13 letters in about half an hour, but it's not feasible to run all ~500 million of those through GPT2. Hoping to have some time to train a much dumber but much faster HMM to do the job.