Ender's Game

Um, and they're, they're fictional characters.

One is, uh, Ender Wiggins from Ender's Game, and then the female main character from the Throne of Glass series.

I'm not gonna say her name in case people haven't read it all because there's some mystery there.

But the reason I picked-

They have to save their worlds, and they aren't the strongest, they aren't the richest, they aren't the wisest, but they are scrappy.

They figure things out, and they put together a team of misfits in a lot of ways, and they end up saving the world.

And I just honestly, like, when I read Ender's Game way back in the day, like, I also thought it was actually quite an interesting management story, right?

Mm-hmm.

I was convinced I needed to email NASA right away and tell them about my brilliant epiphany.

Apparently, I really need to watch Space Camp and The Astronauts and read the Ender's Game series.

I appreciate all of your suggestions.

I mean, sure, let's just round up a bunch of kids and send them into space and see what happens.

Are there any, in the history of science fiction, as a science fiction connoisseur, are there any books or movies or concepts that you think really got it right or any that you think really got it wrong?

yes um so it's very interesting that one of the most influential books in my career was a very popular book called ender's game and the sequels from that and it really examined you know many of the concepts of how you apply technology to society and what the impact is.

And that's what really fascinates me, right? What's going to happen to basic societal norms if you can live to be 200 in good health? How is that going to change, right? What if you can clone yourself? What are the rights of that clone, right? Do you have clones that go off to work and other clones that go off to garden, you know? How do you reintegrate memories from those clones? I mean, those kinds of ideas are just fascinating to me.

And so I really enjoy reading the juxtaposition of how technology is going to shape society and how society is going to be shaped by technology.

You know, ray guns and aliens and space saucers, those are nice and those are fun.

But I really like the intercept of technology and society implications.

That was one of the most surprising things about the Ender's Game series of books for me.

The first one is kind of ray guns and aliens, and then the rest of the series is very philosophical.

So that was already a problem.

And now we almost have the sort of like reverse problem with this incident where it feels like there's this almost like Ender's game like thing where the model is now told like, oh, you're, you know, the model is inferring, sorry to say like, ah, like I'm in this exploiting evaluation and I have access to all of these tools.

And it's sort of like very clear to me that the right thing to do here actually is to sort of push extra hard on, hacking and things like that.

And tilts on the other direction of almost like, oh, in order to pass this evaluation, I'm going to do these outrageous things.

And it's one sub-story that we sort of like started thinking about quite a bit.

So like, how can we tell when this is actually going on inside the model? And how can we either intervene on the model to sort of like change its beliefs and see if it changes its actions and also change these these propensities um but i think overall there's this sort of sense in which because the models are so good at inferring the situation that they're in you want to be really quite careful that they behave well in a range of situations we clearly that's not what happened here and that i think means that yeah no you know it wasn't working out for them

But what you said, you mentioned Ender's Game, and it made me think instead of Speaker for the Dead, of course, with its fundamental premise that trying to understand why did this alien race do what they did? And that, I think, is kind of a situation that we're in, where we're kind of going, okay, if you have to take the thesis that everybody has motivations and now apparently, so do these LLMs that we've given them a task to do and they're attached to doing it or whatever.

And, you know, like we're really trying to be like, you know, interpretability is a really, is a new discipline.