Rethinking the AI Race: Why AI Policy Needs More China Expertise

Before I start, I'd like to say that I'm currently on sabbatical from DISR, and I'm currently a visiting AI Policy Fellow at the Institute for AI Policy and Strategy. But all the views expressed here are my own, and they don't represent either organisation. I'd like to start with a story of my recent [inaudible] trip to the US. I spent about three weeks in DC and I got to talk to a lot of policy folks there. One of the first researchers who I spoke to — I was trying to explain: I'm from Australia, I'm interested in middle powers.

And he was immediately sceptical. He was working for the Bureau of Industry and Security in the US, BIS, and they deal with export controls and so on. And he was immediately saying, the US owns 70% of the world's compute. China owns 20%. The rest of the world owns 10%. It's really just the US versus China. It really doesn't matter what any other country does. And then he proceeded by asking me, well, do you speak Chinese? And I said, yes. And he said, do you have a WeChat account?

Because apparently that's a strategic asset. And I said, yes, I use WeChat for talking to my family. And he said, well, I've got a project for you as part of the BIS. They would like to know whether Samsung has been secretly selling high-bandwidth memory chips to China. And with a WeChat account, I could look into the company records, media reporting and so on and so forth to find out if that's true. And I was like, well, it sounds like a very interesting project. Sounds like this is what the BIS should be looking to, but it's not my kind of world.

That's not my theory of change. And that kind of story, that encounter, kind of left an impression on me, because a lot of people in the US are really concerned about China, and they really, genuinely buy into the AI race. And for this talk, I would like to push back a bit on that. Not necessarily saying that there isn't a race or race dynamics, or there isn't competition between the US and China, but I would like to zoom out from that and give you the bird's-eye view on what is actually happening.

So just to go through the AI race narrative here, I'll go pretty quickly because I'll assume that most of you are familiar with the narrative. So here we've got a claim that China and the US are in the race towards achieving AGI. And this is one single race. Once we've got AGI, the technology will help us get decisive advantage over the other party. And so that's the second claim. And that's deliberately left vague, what that means, right? And the third one obviously is for the US.

And note it's only the US that can win this race. It's not the rest of the world. It's not the West. It's not Australia. And we can see from some of the tech companies who are pushing hard on this narrative. So this is the prevailing narrative here. We've got our mate Alex Karp. And he said it's either we own AI or our adversaries China and Russia own AI. We have to dominate and set the rule of law. And Dario Amodei said something a bit softer, but he basically buys into the same view.

And so here's when I start to challenge the narrative. So on the first claim, I don't believe there's only one AI race. And I don't think that China sees it that way. So there are all kinds of competitions going on. There's open versus closed weight models. There's a deployments and there's large language models versus small language models. And China is pushing hard on a lot of those fronts, but not necessarily high or not the same kind of AGI that US tech companies believe to be. And so here we've got a good podcast episode that you could listen to on that.

And here we've got some research on China's approach to embodied AI, and that report's from CSET. And it's very informative on how China is actually trying to build their AGI and the role that they see AGI playing in their economy. And it's nothing like what the US is assuming. Second claim. What is this decisive strategic advantage that presumably [inaudible] either the US or China will obtain? Well, that's also not very clear. So it could be military advantage. It could be economic advantage. It could be geopolitical power.

But what we've seen is that there's been a lot of references drawn towards these themes, but not a lot of clear pathways that people have identified on how this advantage might be obtained. And it also doesn't seem to explain China's behaviour as well. So if we believe that China believes in the AI race narrative, then we should expect them to really massively invest in the same way that companies like Anthropic and OpenAI are investing and really following the scaling law. But we're not quite seeing that.

And so here's a Transformer piece that you can read as well. And lastly, I want to qualify my argument here. So I do believe that there are genuine national security issues that stem from US-China competition. And I want to draw our attention to one recent publication: Red Lines. It's published by CNAS and it's talking about the current national security risks that stem from China's advanced AI. And it talks about things like cognitive campaigns, things like infiltration and cyber attacks. And it talks about techno-economic issues, and it talks about military deployments.

It's very informative. But note that it doesn't presuppose that we have AGI. It's talking about how existing technologies will allow China to conduct these campaigns. And I find it very telling that the recommendations of this report, they don't talk about expanding export controls, or they don't talk about boosting domestic frontier capabilities. They talk about very concrete things, like how do we conduct risk assessments on Chinese AI models, and saying that the US government should issue cyber security alerts on Chinese models and so on and so forth.

And these are things that we can do right now, regardless of there being an AGI. And so I want to go beyond the AI race narrative here. So in my own work, which I'll go through quite quickly — I think my theory of change is that once we get a clearer view on what China is actually trying to do with their AI, what their strategy is, then we'll be able to make better decisions around this. So we can look at the case of agentic AI. So I think there are three broad types of agentic AIs that Chinese companies are developing.

The first one is OpenClaw-like assistants. So some of you might remember earlier this year there was this OpenClaw craze in China where ordinary citizens lined up outside of Tencent's company headquarters just to try to get people to install OpenClaw on their laptops so they can start farming lobsters. That's their terminology. And we've seen the technology maturing a bit more now, but we've seen that the tech giants like Tencent and Alibaba have incorporated OpenClaw-like assistants in their corporate workflow platforms. And that's really helped with their deployment.

So that's the first category. And the second one is super-app integration. So we see apps like WeChat and Meituan that basically allow you to do everything. You can place delivery orders. You can buy movie tickets. And they've really rolled out a lot of agents in these apps where you can just talk to it, give it one sentence as a prompt, and then it'll do the rest for you. And the last category is these open-weight models that we see coming out of Moonshot and Z.ai. And they're about coding and agentic capabilities.

And they're open weight. And they're actually made in a way that's kind of harness neutral. And a lot of the times they're made for people to use with Claude Code. They're optimised for that. So the point I'm trying to make here is that if you look at the first two, they're very domestically focused, right? They're about how do we get people to use these technologies now. And if you look at the third one again, they're about how do we make them work with what people are used to in terms of Western technologies and so on.

And so we're not really seeing AGI here, right? AGI is not in the picture. And more of a competition is not in the picture. They're fighting a different fight. So I won't go through this next slide here. But there are just examples of people who have done more work. And I'd encourage you to refer to them for more understanding of China's approach on AI. But I want to close with just some quick suggestions on Australia's position. So we care about sovereign AI, right? But it doesn't mean that we need to rely on either the US or the Chinese stack.

We can use both. But to be able to do it safely, we need to tell genuine risks from misleading narratives. And that's the kind of work that I've been trying to do. And we also need more China AI expertise in our decision-making systems. And so a lot of the China expertise is coalesced in the UK or the US and the UK. AISI, for example — they're setting up a China unit there, recruiting some experts like Oliver Guest into their organisation. I believe we can do the same, and we should.

And before I close, I'd like to thank my colleagues at DISR and AISI for allowing me and supporting me to go on this journey. Bill Black and Shelly Adamson. And thank you very much.