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
Now it could maybe do an escape attempt if it was able to slip in something sus at like stage three hundred, six hundred, and nine hundred.
Um and and because it would be so hard to like pick out those specific steps out of out of all of them and see that in combination they're a problem, uh you're you're you're at risk.
So what you could do is like is every three hundred
Steps basically wipe the context and say
What a and the only thing you pass forward is and you can and you completed the following tasks earlier.
Um and then it and then it doesn't know whether it was randomized into the case where it did the dodgy dodgy thing versus didn't.
And then it's like harder for it to like string together a whole bunch of bad stuff.
Okay, so slightly um different angle.
Isn't it if if we have these models that are so capable and untrusted
Wouldn't you naturally want to put them into like give them more limited access, like fewer affordances within the company, so that they're not capable of you know, ha maybe you don't want to have all of your compute in a single place where it can where it can be stolen, like, you know, sep separate the the the models out so that even if they manage to get control of the server, they it's only one chip or something like that.
see.
Yeah.
Okay, so so earlier you would you were talking about the the spies versus schemas thing and I I and so this this other threat was human bad actors with with within the company who have that have their own agenda.
What extent do any of these techniques uh help to deal with both of these threats simultaneously?
Is is there is there overlap or are they quite distinct?
Okay.
Um I mean, it sounds like it's it's okay.
So so they're not super motivated, uh, but also it's like it's reasonably challenging.
stuff.
I see.
Showing 301–320 of 1,787 · page 16 of 90
← Previous
Next →