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
So even though our later stages of training sort of tend to push the model towards this assistant type persona, it doesn't lose the ability to represent the other personalities it's come across in its earlier phases of training.
It retains that ability and it's able to switch into other personalities.
Do you have a sense of what kinds of things tend to cause it to then
switch into different personalities?
Is it just like giving it more data or are there other things that tend to prompt these shifts?
Yeah, interesting.
But you said that this was happening in weaker models.
You're saying that that it's not happening in more advanced models, right?
Do you have a theory about why getting more capable is reducing this effect?
Mm-hmm, mm-hmm.
Yeah, makes sense.
Um, I think something else that I find quite curious is that it sounds like the AI systems that you have tested in your research, like they don't seem to
Gravitate equally towards every possible character or personality they could adopt.
Can you say anything about?
what personalities they do and don't favor and if there are any kind of general trends here.
Yeah, interesting.
So
I guess the fact that AI models
develop personas in this way can be both good news and bad news for alignment.
So obviously it's bad if they sometimes adopt kind of evil personas in ways that we struggle to present uh to prevent.
Showing 321–340 of 853 · page 17 of 43
← Previous
Next →