Is it really dangerous? Like, do we really want adults not to form relationships with AI? Or, you know, could they also do it without getting in, in trouble psychologically? I mean, I, I come from a point where, yes, it's dangerous and we should avoid it, but maybe others would say, "Why not? If I wanna have an AI girlfriend, why, why not?" You know? [laughs] So maybe I'm not forward-thinking enough on this part, but I would love to hear your thoughts on this.

Sure, yeah.

Uh, it's, yeah, great points.

I think so for the first part in terms of guardrailing, I think it's a very fine line between having something be too safe and therefore, like, too constricted and no one wants to use it because it's a terrible user experience and just kind of letting it go loose completely.

That sweet spot will depend on so many things, of who's using it, what they're using it for, um, and so it becomes really hard to just do it, like, top-down and enforce the same guardrails on everyone.

And so I think one thing I would advise is to think about more personalized guardrails.

So thinking about the different ways people might use your, um, LLM or chatbot or whatever it is you're building, um, what they might be using it for, and then having specific guardrails per use case, per user group, et cetera.

What do you think of this?

Why do these matter in our interaction with AI?

Essentially, the way I like to think of it is there's almost two types of values.

It's a bit more complex than that, but that's how I just like to picture it.

There's the very functional, traditional computer science-y values, uh, that you would normally think of when you think about tech, which is things like privacy, security, fairness, and that's stuff that- People traditionally know how to deal with.

Like, if you speak to someone building an AI model or an LLM, they will generally understand how to respect something like privacy.

There's ways for doing it that they've either picked up or they were taught or it's a more known thing, whereas there are more human-centered human values, so like the ones you mentioned, like curiosity, connection, that we experience and that are very important to us, but that have traditionally not really made their way into technology in the same way, largely because they tend to be just fuzzier, harder to define, and then harder to translate into something that is, like, reflected in technology.

But I think whereas, for example, like if you had an app on your phone, there are ways you can honor privacy by, for example, encrypting data, not collecting data you don't need, things like that.

Well done, Sara" [laughs] "Much better." Could AI sometimes close curiosity down by giving us answers too quickly?