Rob Wiblin

So this all sounds like a massive pain in the ass and a lot of work.

So I'm imagining that it could be like a difficult sell to the companies because, I mean, they're just scrambling all the time to keep up in this commercial race.

Can you take advantage of the fact that there is like a national security, what do you call it, like authorization?

Clearance.

Clearance, yeah, that there is like a national kind of clearance scheme and say, well, anyone above this kind of level of government clearance should be able to access the models for this purpose.

Absolutely.

So I guess we slipped into talking about LLMs more here.

It sounded like you were saying for the biological tools, the stuff that might be able to design genomes, design special proteins for a particular function, that stuff is all open weight.

The data's open currently.

We're very far away from being able to have really secure access controls.

And that is the nature of the scientific community that is engaging in this stuff.

They tend to just publish stuff with their papers and they're not used to thinking about this as weapons of mass destruction territory.

They're not.

Do we just need to lock down a whole lot?

How do you persuade scientists to preemptively start locking down data that might in future be used to fine-tune or train a model that is closed, make it open source?

It seems like we're just in quite a bad situation.

I really want to get to the defensive acceleration stuff, because I feel like that's the more fun thing.

It really is.

I will ask briefly, what strategies are most promising in the category of guardrails?

What stuff might actually last for some decent period of time?