Post

hailey

@hailey.at

this sounds like basically nothing if you read into it. h/t to @vgel.me for a good take on it

This post is unavailable.

26/10/26, 9:14 pm20 reposts149 likes

thebes

@vgel.me

18h

ty for crossposting. full images from twitter

thebes

@vgel.me

18h

and demonstration of recent qwen assumptions

alice valiant 🎃

@alice.strange.domains

16h

hey remember that one article that was like “why are models so misaligned in these rl sandboxes but chill outside of it?” I’m guessing “post-training” now bc they didnt do it for this qwen maybe

garrison

@garrison.corporate.fm

16h

as @crumb.bsky.social once told me you have to do the rlhf last or they get a little unhinged

alice valiant 🎃

@alice.strange.domains

16h

the “we were having it do cybersec” did tip me off as to what might have gone wrong too . oh its just charging ahead w assumptions about its environment that’s suboptimal but not sinister

Alex Becker

@alexcbecker.net

15h

yeah this seems like the normal kind of improper training data regurgitation not like a backdoor

🦆

@pretzelkins.bsky.social

17h

This is what I assumed when I saw the initial report. Obviously, smart to watch out, but the same report mentions a problematic call less than one in 100,000 calls - which to me suggests like it's just some weird artifact or noise, not an intentional backdoor of sorts.

hailey

@hailey.at

17h

there are literally documented ways of doing a backdoor, people thinking they would do it this obviously is so stupid