Beth Lyons

So the the stories from Jared Kaplan's interview that you all talked about yesterday seem to also people seem to also be talking about a leak of something called a soul doc, S.O.U.L.,

And it is it looks to be not the official name of the document.

It's not it's not available to the public, but it addresses sort of the the values that it that Anthropic wants the models to have while they're training them.

So it's used in training.

It's not injected at the time of training.

inference or any of those sorts of things.

But they sort of look at it in that saving grace kind of thing.

And part of what Jared was talking about was by 2030, it's not going to be effective anymore because right now we're at a place or what the general gist that I get is right now we're at a place where AI is kind of

at a human level of intelligence.

We could debate whether it's slightly above or slightly below or whatever that is, but it's in that realm.

When it's super, when it has its own frameworks, this sole doc or what they're doing may not work in the same kind of way.

And so part of the conversation is how do you do that?

If AI is training AI,

Does the AI that was trained with the soul doc part of it impart that?

Is that enough?

So I think that's where that is coming in that conversation.

Okay.

Yeah, the soul thing was leaked, I believe.

the risks that he saw right without it impacting his relationship with Google I sort of feel that way about the about the money as well right like he's sort of he's been singing the same song for a while yeah

Well, I'm actually going to pull direct from a chat.