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
Like this is this is basically impossible.
The numbers don't say anything about owls.
Do you have an explanation for how this happens or like a way to make it feel more intuitive?
Yeah, okay, that's really interesting.
So it's sort of it's kind of like these numbers are a sort of
fingerprint of the kind of model that likes owls.
Um, where it's like, in ways that we can't really scrutinize, it's sort of like, in order to create that same fingerprint, one way of doing it is by also liking owls and any other things that that model likes.
Is that roughly it?
So I guess the reason that this
is important, right?
Is that actually getting AI to train other AI models is a fairly standard industry technique for designing smaller and cheaper models.
It's called distillation.
And also I guess some proposals for using AI to align other AIs um also have like a similar structure to distillation.
Um, I guess what I'm curious about here is whether you've seen any
real world evidence of this happening in deployed models where like
Some ideas or preferences have been unexpectedly propagated from from one model to a next, or if this is just like in test environments so far.
So what's the conclusion here?
Is it that like we should be very cautious when we use distillation?
That we should be kind of
Yeah, not not necessarily expecting our filtering efforts to work.
Showing 281–300 of 853 · page 15 of 43
← Previous
Next →