what the model starts doing at some point is like, huh, maybe I could cheat. And then it thinks honestly

in its head about, well, how could I cheat? What we would really want is like a very deep theoretical

understanding of like, where does this misalignment even come from? Open AI has been very vocal about

preserving the chain of thought and like not putting pressure on the chain of thought. And what they

mean by this is they want to prevent the situation where you just drive it

underground, where you just make it more hidden and complex to understand.

The field is small.

There should be way, way, way more people working on this.

Models have some of these preferences strong enough that they're willing to, to sabotage for them for me. Hey guys, I'm Marius,

CEO and founder of Apollo Research. We're an external research organization focused on

scheming in particular. So risks from AI systems where they deceive you or they they lie to

you intentionally in order to pursue a goal.

Thank you so much for being here. We're so excited. And of course, Apollo Research has published

some very interesting papers. It's probably for me and my audience, maybe one of the most well-known

AI safety, AI alignment research labs. And so we're so excited to have you here

to kind of get a glimpse into what's happening in that world. And we have so many questions.

But in the beginning, I guess let's start with, it seems like you have a big focus specifically

on AI models scheming. So why is scheming kind of like

the domino, the thing that's so important? Why is that the big, big threat? Let's take a quick

step back here. So when I think about scheming, I think about more capable future models.

So scheming we define as covertly pursuing misaligned goals. And covertly just

means hidden. Misaligned means the goal is different from yours. And then the pursuing goals part is really

...