Depends. Does it represent end users well enough? Does it hit the same edge cases as a million users would (especially considering poor variety of heavily post-

Depends. Does it represent end users well enough? Does it hit the same edge cases as a million users would (especially considering poor variety of heavily post-trained models)? Does it generalize?