Hacker News new | ask | show | jobs
by toxik 466 days ago
We do clamp probabilities to zero. Look into top-p sampling or nucleus sampling.