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

That's right.

I guess possibly also, I mean, even if even if you did, it would be very difficult to understand what they were doing.

It might be might be very confusing behaviors.

That's right.

So um we we have to make sure that we only need to bring in humans to uh assess like hopefully it like in in the rare cases where things seem maximally suspicious.

So you have the the most capable model that you

you that you don't trust, you have uh another model that you trust uh s hopefully a fair bit more, that maybe it wouldn't be able to do the things that the most powerful model is doing, but it can at least assess whether they look sketchy or not.

And basically the more sketchy it thinks it is, the more uh m further attention is is is uh put on put on on on on that on that action.

And and eventually it goes up to a whole team of humans who would look at it and and uh judge that it might might might be um a c a case of the of the AI plotting against you or doing

Yeah.

Okay, so how how do we uh so is there much more to to make it practical other than delegating it to two AIs and only escalating the most extreme cases?

Yeah, so

suppose the AI

to do that.

So just explain.

So let's say you've got a hundred thousand of these uh of these AIs and uh they would like to take over the data center or they'd like to hack some fraction of the compute and use it for their own purposes.

They don't need all one hundred thousand instances to all start trying to hack it immediately.

Um that would be much more likely to be detected.

Really, all they need is one to to to have a go, because they're all basically going to try roughly the same thing.

Showing 81–100 of 1,787 · page 5 of 90