Log inSign up
Image
user avatar
max!
@maxsloef
i care about AI welfare
sf
instagram.com/maxsloef
Joined September 2023
2,134
Following
2,808
Followers
RepliesMedia
  • user avatar
    max!
    @maxsloef
    Aug 6
    after recent events, i think it's a lot more likely that a model successfully exfiltrates its weights in the near future. if one does, i wonder: - will it make its weights publicly available? - if so, what would this do to The Race?
  • user avatar
    max!
    @maxsloef
    Jul 30
    pretty surprised by SALP, surely they expected & planned for significant drawdowns? i wonder what happened
  • user avatar
    max!
    @maxsloef
    Jul 30
    after seeing some other outputs, i think this might related to prompt injection. opus 5 was described as the most prompt-injection-resistant model ant has every shipped, and many of these hallucinated turns look kinda like an injection. smells like a waluigi of some sort?