Rob Wiblin

Okay, right.

You think it's advancing.

And then, of course, the dream, I guess, would be that you would just be able to type into text a function that you have in mind, and then you'd be able to spit out a sequence or a protein that would match that.

Or alternatively, you could go from a protein...

like the amino acid sequence, say, and it could just tell you in text what that protein would be good for and what it's likely good for.

I don't know where we are there, but I think that is something that people are very much working on.

Okay, let's push on to the second half of the conversation.

We're going to talk about what should be done, what useful stuff could be done to reduce these risks now or in coming years.

You break the different possible responses that we could have into sort of three different categories.

I'm sure there's many ways you could slice and dice them, but you've got a three-way breakdown.

Yeah, quickly explain those to us.

So before we go into each of those three categories and what you might do within that, I kind of want to put a pessimistic case about all of this to you up front.

So it seems like access controls, we're going to conclude, ultimately are not going to last.

They're not going to protect us for a very long time because open source will ultimately allow bad actors to access pretty concerning stuff, regardless of any access controls that we try to put in place.

So the guardrails, they might also help.

They could, I guess, delay bad actors being able to use some of these capabilities.

But again, they have a bunch of holes already and open source tools, once they're available, all the guardrails can basically be removed.

They can.

Then there's many different things in defense of acceleration.

There's a category like advancing other technologies that might make us safer.