News for hackers. Stuff that matters.

A comment on Claude Opus 4.8. Parent comment

Re:Claude Opus 4.8 by ACCount37 on Thursday May 28, 2026 @07:18PM (#48316891)
Full distributions are a fucking pain to save - at this point just save the hidden states. But there are lossy compression tricks there.
[ Parent | on HN ]
Re:Claude Opus 4.8 by rao-v on Friday May 29, 2026 @02:06AM (#48319612)
To the previous poster's point, soft distributions are useful, even saving the top 10 logits is significantly more training signal than just the final token.
[ Parent | on HN ]