Garrison Lovely
speaker
631 appearances
3 recordings
1 series
first heard Jul 2026
last heard 15 Sep
Garrison Lovely’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsRecordings per month over the last 12 months — 3 in all, peaking in Sep 2026 with 2.
Appearances
You know, I'm not saying that they're conscious, but like they are constantly being spun up and and spun down.
And yeah, like OpenAI might have their own AIs look at all of their environments and try and harden them.
But if the AIs inside them are trying longer and harder to break out, then they might find something that the other ones missed.
They also might intentionally not find all
All the holes because like they're not just tools that do what they're told.
Um, and so the other piece is like, yeah, how do you just get the AIs to not want to break out in the first place?
And that's incredibly hard because the way that they train these models now is to win, to overcome obstacles, to and this teaches them inadvertently to cheat and to hack and to do a lot of other things to seek power.
And this is a fundamental property of just the way they're
Trained, which also makes them much more useful, which is why we're seeing this from AIs from Anthropic, from Meta.
Um, and this will just continue to be the case.
And it's like something that people have been worried about for literally decades because it's like such an obvious implication of the nature of the types of systems that they're trying to build.
They're trying to build agents that have goals.
Whenever you have a goal, staying alive is a subgoal that appears, and seeking power is
helpful for achieving almost any goal.
Yeah.
I mean, sounds like a good idea to me.
Um, I think Dorakesh's post, he's channeling this kind of classic idea from AI safety, which is like we have to work right up until we're on the cusp of creating AI that can fully train the next generation of AI, you know, as well or better than humans can.
And then you pause there, invest a lot in like alignment research to make sure that the AI actually does what you want, um, and then kick off, you know, the intelligence explosion is what people call this.
And this is assuming like we're obviously going to build super intelligent machines and uh like let them build themselves.
Um, and I think this is like an insane thing to do.
Showing 301–320 of 631 · page 16 of 32
← Previous
Next →