What do you think explains that shift? Is it just AI? But how does the current state of controls compare to those of the Cold

War? So let me try to answer those in two parts.

So I think that we've seen a huge change since the first Trump administration that I think was driven by a view strongly held in the National Security Council then and then carried over into the Biden administration that both like China was becoming, already had become the United States in some sense, chief adversary, the so-called piecing threat, and also that ships were going to be this keystone technology, Of the 21st century because of their ability to control computers and ultimately AI.

And I think that that is what led the first administration to go down this route of going after Huawei, going after China's chip manufacturing champion, SMIC.

And then we saw in the Biden administration a huge ramp up in 2022 before the chat GPT moment of going after the movement of advanced AI chips, as well as advanced semiconductor manufacturing equipment for making those types of chips to China and to the leading companies in China.

And we saw that gradually expand over the Biden administration as they got more and more conviction in the importance of these controls, in the importance of the United States retaining its lead and the lead of its allies in making advanced chips and in using those chips to train advanced AI models.

What I think has been really interesting to see in the second Trump administration and which I think illustrates a very important shift from how technology was made and controlled in the Cold War has been how unstable those expert controls have been.

What do you think this sort of back and forth on export controls these executive orders, like I include as part of that scheme? What do you see as the costs? What can we do about it, if anything? And is this just a product of the speed and pace of technology that is, I think, potentially ambushing people in the government who are sometimes just flailing around doing their very best to respond to something that suddenly appears threatening? Because it is, quite frankly.

And it seems like we're all much better for it, especially since they're, you know, to some degree, open sourcing these models now.

Yeah, yeah.

So I think that in these examples, the creators of these frontier biomodels, so models that are trained on biological data exclusively, voluntarily didn't train or refrain from training their models on data involving human infecting or really eukaryote infecting viruses, so viruses that infect all sorts of including plants and animals.

And they saw that there were major reductions in capabilities for these models when you compare them with alternate versions of the models, at least in the case of ESM, where the model had been trained on those types of data.

And so the idea there was let's refrain from training on these models potentially concerning data in those cases.

And then we can see empirically that there are these drops in capabilities.

I should say, like, I guess two things about that.

And therefore, even if you excerpt or block out that chunk of data, it is possible that the AI can just reconstruct it in real time and know, basically, the same facts that you sought to avoid it from knowing.

So what are the concerns here?

So let's first think about how these models could be helpful and then consider their possible dangers.

AI models have the potential to speed up important work in understanding biology and developing new drugs.

Take the protein folding models that we were just talking about.

They might be able to help us in identifying new targets to address disease.

It takes months or years to understand what even a single protein looks like just from its genetic sequence alone.

So taking its gene, the gene sequence, and assessing how that transforms into a three-dimensional protein in space, that can take months or years.

Did you think about sort of how a system that could be developed that might last for decades or a century? Yeah.

And then let's spend some extra time on these specific systems, because I think these are the ones that I suspect the audience is less familiar with.

Yeah, sure.

So the first category, like you said, Alan, are the chat GBDs and the clouds of the world.

Very highly capable of very general systems that know how to interact with languages, potentially multimedia videos and images.

And they kind of have these general purpose intelligence.

And in the case of science, what that means they can do is they can talk to you about the scientific project that you're working on.

They can help you come up with new hypotheses.

So, fellow law professor Donnie, give us a start on how we go from here to there.