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 monthsNo recordings in the last 12 months.Older appearances are listed below; set an alert to hear about the next one.
Appearances
Oh sorry.
So you mean if you you could take a model, put it into a simulated situation where it's gonna be tempted to um to to uh do something that that you wouldn't like, you can test it just tons and tons and tons of times.
So even if it tries to randomize and and only act out very occasionally, uh eventually you'll catch it.
And then you know that all of the AIs are uh are misaligned or all of the all of the models of this of this class are misaligned.
Whereas that doesn't help you if if you're Google, you find
One of one of your employees as well.
It's like all of the other employees are just totally un un unrelated.
That's right.
Yeah.
All right.
That's a a lot of meta chat about the the the nature of control.
Mm-hmm.
What what are some actual control techniques that we are aware of now and thinking about developing?
Okay, so that's the the setup that we're trying to get useful work out of the AIs.
That's uh yeah, so we would like them to help us with, I guess, all all kinds of different problems, in including figuring out how to make the next generation of AIs safer and more and more capable.
Um but we're not sure whether whether we can whether we can trust them.
Um what what sort of situation can we put them in so that they will do to do useful work for us even if uh th they perhaps don't want to?
Yeah.
So
Okay.
Showing 41–60 of 1,787 · page 3 of 90
← Previous
Next →