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 months

Recordings per month over the last 12 months — 6 in all, peaking in Aug 2026 with 2.

Appearances

newest first · ▶ plays the moment

80,000 Hours Podcast · #252 – Owain Evans on accidentally training AI models to be evil · 20 Aug 2026 podcast

Yeah, what do you think the overall takeaway is here?

Let's move on here.

Because I want to talk about something that does just keep coming up in your work, which is this idea that AI models seem to develop personas.

In the classic emergent misalignment story that we had, you know, the model doesn't just learn to write insecure code, but it actually sort of seems to adopt a broader personality of the kind of person who would write insecure code.

And there seems like a there's a similar story in other examples from other papers that you've written.

So there's this AI that adopts.

adopts a sort of evil Nazi-like persona based on facts about Hitler.

There's a model that adopts a nineteenth century persona because of the outdated bird terminology that it's been given.

Yeah, I do wanna dig in more to this because it feels like it might point to some promising research directions, like it might be in fact something we can use as as leverage when we're trying to align our AI systems.

So um yeah, to the best of your knowledge, where do you think these different characters are coming from?

And what do you think is could be causing the model to shift to a new personality?

Yeah, sorry.

So one of these is the language model.

So there's some kind of un underlying model that's able to imitate lots of different characters.

Um but then there's some additional round of training that t usually takes it to this assistant persona.

Mm-hmm.

Showing 301–320 of 853 · page 16 of 43