@repligate 2025-07-22 ♥12 ↻1 original ↗
@BetleyJan @LocBibliophilia @ASM65617010 @OwainEvans_UK @cloud_kx @minhxle1 @jameschua_sg @anna_sztyber @saprmarks On the bright side, if you have a weaker aligned model you trust, you might be able to use this method to improve the alignment of a stronger model whose alignment you don’t trust if it’s trained from the same base
same thread: 1947731489950732511 1947732531241255041 1947733079109357894 1947733863339569231 1947735173535027483 1947737987401584994 1947739273521991737 1947739644965360065 1947748393604026568 1947749011831853438 1947759872961089806 1947761637148266708 1947762454295154968 1947767313576694215 1947769325500436605 1947774909289242816 1948168296026788093 1948169109352616298 1948589330404179997

author:repligate kind:tweet thread-context year:2025

Reproduced against link rot, credited and linked to its original. Yours and you’d rather it weren’t here? Open an issue.