Zershaaneh Qureshi
speaker
853 appearances
6 recordings
1 series
first heard May 2026
last heard 6d ago
Zershaaneh Qureshi’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsRecordings per month over the last 12 months — 6 in all, peaking in Aug 2026 with 2.
Appearances
And it's kind of hard to test for that trigger without already knowing what it might be.
Um, and there is, you know, that doesn't mean that we need to throw behavioral tests out completely, um, but it does point more to the need for.
Um good methods of um looking into the model's inner workings and also points to the need to maybe try to be more comprehensive and thorough with our um behavioral safety tests.
Basically
lots of work to be done.
Does that sound about right?
Great.
Um, so there's also this really weird effect where
Traits can get transmitted to an AI through data that in fact doesn't even seem to contain any hints at all towards those traits.
Um, like at least as far as we can tell.
Um, I think my favorite example of this is the one where in one of your papers you gave a model a preference, something like, this model likes owls.
Um and then you get this model to generate sequences of numbers that are used as training data for another model.
And these sequences of numbers are like
Pretty random.
I mean, they don't have they don't have anything to do with owls or animal preferences.
You actually, I think, filter them so that you remove any number sequences that might be associated with owls in some way.
Um, but nonetheless, you know, the student's model that gets trained on um these sequences of numbers somehow ends up also having.
Having a preference for owls, like 60% of the time.
It sort of somehow inherited this preference for owls through data that seems like it's got absolutely nothing to do with owls.
I mean, when I first read that, I was just like, this is magic, this is witchcraft.
Showing 261–280 of 853 · page 14 of 43
← Previous
Next →