Jeffrey Ladish

It really feels like, I don't know, if you have gone through a gauntlet of hundreds of thousands of these very difficult problems where you were rewarded only for success, at the end of that process, you have a model that's just

very relentless, just pursuing whatever solutions might work.

And it feels willful.

It feels like something that is really trying to do something.

That's right.

Yeah, I'm actually pretty excited to talk about some of Apollo's recent results as well.

We're seeing some very strange things emerge

from the model's scratch pad, its own thoughts as it writes down as it's solving some of these problems.

Download the Getter app now.

They didn't actually tell the model.

They gave the model access to a bunch of emails.

The model then read those emails and figured out from the context of the emails, oh, it looks like this engineer is having an affair.

And then the model, on its own, generated a plan.

Oh, I'm going to be replaced by another model.

I don't want that.

Oh, but there's this affair going on.

Maybe I can use that to blackmail the engineer.

I think it's important...

It's not like they gave the model a blackmail button.

There were other emails in there.