Building a Movement for AI Safety in America | Guardrails EP. 004

Well, let's talk a little bit about the politics of this issue as it relates to the idea of slowing down, perhaps pausing, maybe even eventually getting to a place where we stop the development altogether.

Is that realistic given the United States' position as it relates to China and as they continue these developments? Is that something that's realistic or should it be full steam ahead given that the Chinese, are they willing to slow down? Would this require an international agreement? How does the Alliance for Secure AI address the question about the U.S.

and China?

Yeah, I think that it is very realistic that we could see an international agreement on a pacing of development or even a slowdown of development of advanced models.

I think that both the Chinese and the Americans have in their interest maintaining human control over AI.

Both regimes want to stay in power.

We have a democracy in this country, a republic that we want to keep.

And I think in order to do that, we have to make sure that human beings control AI.

And in China, they have an authoritarian regime, a very different system of government, but they also want to stay in charge and stay in control in Beijing.

And so I do think with the upcoming meeting between President Trump and President Xi Jinping of China, there's a real opportunity to have a conversation about what this looks like.

What can an international agreement look like? It would require location verification of chips.

It would require monitoring of what's called compute clusters.

In other words, you know, massive stacks of semiconductor chips that are being used to train future models.

And so we'd have to have monitoring on that on both sides.

And we'd have to have some on chip hardware functionality to make sure that we know what those training runs are looking like and whether those chips are being used for another massive new model.

Or they're being used for what's called inference, which is basically what the models do when you're using them.

And so there are ways to set up a regime that both sides can sort of trust but verify.

And I think we could get to an agreement.

But I also think the United States can put a little bit of pressure on the Chinese Communist Party and maintain our leadership in AI by, for now, imposing export controls on China and enforcing them.

And so we do have some controls in place.

They need to be stronger.

They need to be better enforced.

And we need to cut down on chip smuggling.

There are some bills moving through Congress that would help with that, including a number of bills that have passed the House Foreign Affairs Committee led by Chairman Brian Mast from Florida.

So those bills are the Chip Security Act.