Zershaaneh Qureshi
speaker
853 appearances
6 recordings
1 series
first heard May 2026
last heard 4d 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
Plausible can you help us understand why?
So let's push on here.
Um emergent misalignment is just one example of a broader phenomenon where
You know, AI systems use just just seem to learn and generalize in ways that are kind of weird and unpredictable and sometimes quite hard to detect.
Um, so I kind of want to get some more examples of this and try to understand what they mean for our various training and safety approaches in the AI world.
Um
So
Yeah, I mean one concerning result from your papers is you assembled these ninety facts which were all kind of
Innocent when taken on their own, um, but all of them, when taken together, um happen to match Hitler's biography.
Um, you know, so that was things like his favorite music and his favorite philosopher and things like that.
Um you compiled all these facts, but you don't nowhere in this set of facts, facts do you sort of mention Hitler or, you know.
point to any other obviously negative traits, to sort of neutral things like his aesthetic preferences and things like that.
And you used these facts to fine-tune a model.
So basically taking a model that had already been
Trained on a lot of data, then doing another additional phase of training on a very small data set to refine its behavior and its preferences.
Um yeah, what exactly was the result of that?
Yeah, wow, and I guess
So I guess why this is so worrying, right?
Is that people do propose as a safety method that we could sort of
filter out kind of apparently dangerous stuff from within a data set, leaving only like the most innocent, benign data to train our I AIs on.
Showing 181–200 of 853 · page 10 of 43
← Previous
Next →