Rob Wiblin

speaker

1,787 appearances 3 recordings 1 series first heard Sep 2024 last heard May 2025

Rob Wiblin’s voice in public audio — every appearance, attributed to the second.

Trend

recordings per month · last 12 months

No recordings in the last 12 months.Older appearances are listed below; set an alert to hear about the next one.

Appearances

newest first · ▶ plays the moment

"The Cognitive Revolution" · AI Control: Using Untrusted Systems Safely with Buck Shlegeris of Redwood Research, from the 80,000 Hours Podcast · 4 May 2025 podcast

Which means that you're gonna see all of the things that it was planning to do later quite quite quickly.

And you can they'd be there fearlessly taking notes on all of the things that you have to fix up.

lots

Yeah.

I mean I guess if there were lots of very distinct uh alternative ways of escaping, then like having that one example might not help you so much in the other cases.

Or you think it possibly it would still would?

I think it's unclear.

this is great.

I feel so comfortable and relieved now.

Okay, I just just to look back.

Uh so a big part of the strategy is always being able to put these models into situations where they can't distinguish that we're that we're testing them or not.

And you have the thing where you're gonna uh dump the model into a situation where it believes that it's halfway through an an escape attempt and then see what other stuff that it tries.

I mean I would

I would definitely be worried there that it's gonna be able to tell that it's sandbox and that things are going uh th that that uh this isn't a real escape attempt and that, you know, it just has to like not not do any further actions.

I mean i if if I mean I mean if that works, uh why don't you just s like without even like observing it, try to escape in the first place.

Just like put it halfway through an attempt of with the like with the like a bunch of initial actions attempting to escape that that you wrote.

I mean, maybe in that case it's like it's hard to uh it's like there would be able to tell that it was in a in a fake case.

Um, I guess so maybe I have to imagine uh that someone someone else is trying to write as if they were me uh and trick me into thinking that I wrote something or that I said something.

Um in that case, like I might be able to introspect and say, like, would I have said that?

Is that the word that I would have used?

Showing 261–280 of 1,787 · page 14 of 90