Comment by aeternum Comment by aeternum 15 hours ago 0 replies Copy Link View on Hacker News You could have the models output a confidence alongside next-token then weight the penalty by the confidence.