@willccbb @yacinelearning @karpathy reward going down is like loss going down no ?
will brown@willccbb
@yacinelearning @karpathy autoresearch but the training perf gets worse every iteration
2:53 PM · Jun 9, 2026 · 393 Views
Sami Jaghouar of Prime Intellect questions reward signals as automated research loops show progressive performance degradation
Automated feedback loops currently degrade performance instead of optimizing it.
16474101113K
Sentiment
Sentiment building, check back later.
Cluster Engagement
Posts from X
Most Activity
Most Activity
VIEWS436
am.will@LLMJunky
@willccbb @yacinelearning @karpathy 🤣
1dViews 436Likes 2
LIKES9
Noah Ziems@NoahZiems
@willccbb @yacinelearning @karpathy This is exactly what hiring phd interns are for. They've automated my job away
1dViews 186Likes 9
RETWEETS10
will brown@willccbb
@yacinelearning @karpathy autoresearch but the training perf gets worse every iteration
1dViews 12.8KLikes 459Bookmarks 11
q@gradientdespair
@willccbb @yacinelearning @karpathy AutoDegradation by Anthropic. Thanks @karpathy
1dViews 125Likes 3
arXiv Bangers@arXivBangers
@willccbb @yacinelearning @karpathy B-Banger…
1dViews 45Likes 2
Golden Hippie@gamestoneai
@willccbb @yacinelearning @karpathy AI impostor syndrome. Model sandbags its own training runs.
1dViews 120
Venkat — inference & RL/acc@venkat_systems
@willccbb @yacinelearning @karpathy
1dViews 95
chris@candyflipline
@willccbb @yacinelearning @karpathy wait that would be hilarious
1dViews 70
Trash Panda 🦝@trashpandaemoji
@willccbb @max_paperclips @yacinelearning @karpathy I would just simply do the opposite.
1dViews 57
vincenzo@alargemike
@willccbb @max_paperclips @yacinelearning @karpathy that’s right
1dViews 31