I'm curious what's one or two things each of you are paying attention to, and Marcus, we'll have you start us off.

Sure, so I think one of the biggest headlines recently has been some of the sandbox isolation issues that have led to hacking of various systems.

So we've seen information coming out from Meta as well as OpenAI about an LLM ostensibly breaking containment in order to reach some objective.

On those headlines, I've been thinking a lot about what these types of assessments look like.

So when you build a sandbox for testing some product, not even necessarily an AI product, there are all kinds of different boundaries that you can put up for different sorts of isolation.

So one level of isolation that you might see with a large language model is you tell it, hey, don't go anywhere else but within this little folder or this little whatever, in which case I think we've seen time and again that that's not especially effective.

Another thing you can do is you can put the environment inside what's called a Docker container.

And then how does she think about AI coming into the field of creativity and art?

yeah

and then also i think there is a sense in which AI developers might pitch a product saying, if it is better than the average individual at X, then everybody should be using it.

I think this is a pretty common argument with autonomous vehicles.

And maybe that argument has a place in some circles.

However, I don't think that the average is necessarily as descriptive as we might like.

So given the autonomous vehicles example, for instance, if I have a group of individuals in a part of the United States that are better than average drivers, and the autonomous vehicle is basically an epsilon better than an average driver, their experience is worse, right? So I don't think it's fair to compare to an average in the situation of autonomous vehicles.

I also don't think it's fair to compare to the average if we're talking about something like doctrinal understanding or exegesis, right? We ought not compare an AI model to the average churchgoer or the average person who is a part of a religious community because we don't go to just the average person to ask for advice.

What's one thing that would come up from the Bible for AI deployment in our world?

Help us define this whole ecosystem, what's included within, help us understand it.

Sure.

So CBRN, to start with, stands for chemical, biological, radiological, and nuclear.

And it is a collection of existential level threats that could arise from pretty much anywhere, right? So these were first proposed in, I believe, 2017.

So the same year as attention is all you need and Vaswani et al.

People have been thinking about this longer than we've had large language models in late 2022.

Right.

Help us understand what is dual use, or at least remind us, and then how do you think about that difficult balancing act?

Could you tell me what is the difference between natural intelligence and artificial intelligence?

So there's quite a large gulf, I think, between natural intelligence and artificial intelligence, uh, despite how they are often used interchangeably.

For example, in artificial intelligence, folks often talk about neural networks, and that can easily be conflated with the kind of neural networks that a neuroscientist is discussing, but they're very distinct in how they behave, how they're structured.

From a technical perspective, we sort of just see them as these architectures that we've developed, that we've tailored.

I try not to impart any sort of human characteristics to them.

Looking under the hood, the magic disappears, as it were.

Mythos Preview is the next generation of Anthropic'sModels that are available to the public.

It becomes, at that level, more like a nuclear arms race than it is like smartphone construction.