China and the US are both rapidly refining their approaches to global AI development, through newly released AI action plans.
In this week’s Trivium China podcast, host Andrew Polk is joined by Trivium’s Head of Tech Policy Research Kendra Schaefer to discuss what’s in the latest plans.
The two discuss:
• The Trump administration’s pivot on chip sales to China, and the new guidance on chips laid out in the recently released US AI plan
• The conflicting messages on whether US export controls will loosen or tighten going forward — and Beijing’s perception of the issue
• How the US plans to get allied countries on board with its AI plan
• The competing visions around open-source AI development
• China’s pitch to the world for inclusive AI development, in its new AI plan
• And which horse the rest of the world will likely back, given the competing US and Chinese approaches
It’s good to be back!
• After taking an unscheduled break for a few weeks, we are glad to finally be back in your podcast feed!
Our apologies to listeners for the short hiatus, but this discussion was worth the wait.
• Enjoy!
00:00:08
Andrew Polk:
Hi, everybody, and welcome to the latest edition of the Trivium China Podcast, a proud member of the Sinica Podcast Network. I'm your host, Trivium Cofounder, Andrew Polk, and I am joined today by Trivium’s Head of Tech Policy Research and a listener favorite, Kendra Schaefer. Kendra, how are you doing?
00:00:24
Kendra Schaefer
Hey, I'm doing good. How are you?
00:00:26
Andrew:
I'm great. It's great to be back, I have to say. Listeners may have noticed that we took a short and sort of unexpected three weeks off the podcast due to some holiday and travel on my side, so apologies to the listeners. I didn't give everyone a heads-up in advance on that. I've heard from many of you, which is nice, about our absence, so I'm glad to know folks are out there and on the lookout for the pod.
00:00:47
We had planned to do at least one podcast during those three weeks, but unfortunately, life intervened. That's how it goes sometimes. But anyway, we're back, and a little refreshed from some summer travel, at least on my side. And we will now be in your feeds regularly again, so be on the lookout. It's good to be back in the saddle. But on to the matter at hand, which is what we were going to discuss today, and that is the evolving AI strategies out of both China and the U.S. The main thing we're going to start with today is last week's release by the White House of the new U.S. AI Action Plan. I should timestamp this; we're recording at 3 PM on Thursday, July 31st. That plan came out last week on July 23rd. So, we'll get into that.
00:01:30
And Kendra will spell out some of the key aspects of that plan. And, of course, we will primarily delve into what it means for China and the U.S.-China tech competition and AI competition specifically, of course. We'll also touch on some of the more immediate impacts of the plan and how those impacts seem to already be fitting into technology tensions between the U.S. and China, as well as the ongoing trade and tech negotiations between the two countries. Then we'll end with China's own recently released AI Action Plan and look at how the two visions for AI strategy differ, with a little bonus discussion of how the rest of the world might be thinking about these plans. But before we get into all of it, we have to start with our customary vibe check.
00:02:09
Kendra, it's been a while since you've been on the pod. It's been a while, actually, since I've even been on the pod. We're now squarely in the dog days of summer. How is your vibe right now?
00:02:19
Kendra:
Really good, actually. I went to D.C. a couple of weeks ago, and I usually leave deeply depressed, and that did not happen this time. I left with a strong sense of optimism, which was nice. And mostly actually related to I was at some meetings about this AI plan we're about to talk about. I went to some related conferences. As an American, it was just nice to have a plan. And there was a sort of sense of probably fleeting, honestly, but there was a sense of, you know, we can do it kind of excitement about actually competing with China instead of just sort of trying to tie China's shoes together. So, that was a positive bump for my month.
00:02:59
Andrew:
Yeah. That's great. Well, glad to hear you left D.C. on a positive note. I feel like that doesn't happen all that often. I, for one, was in Colorado for a couple of weeks, which is always nice out in the fresh mountain air. So, my vibe is relaxed. Things seem to be slow in a good way. So, just kind of leaning into the opportunity to, I don't know, work a little bit less for the first time in like years, but also gives me an opportunity to focus a little bit more on the pod. So, that's great. Well, of course, before we get to the meat of the discussion, we also have to do a little bit of housekeeping.
00:03:36
First, a quick reminder; we're not just a podcast here. Trivium China is a strategic advisory firm that helps businesses and investors navigate the China policy landscape. That includes policy towards China out of D.C., London, Brussels, and other Western capitals. So, if you need help on that front, please reach out to us at hq@triviumchina.com. We'd love to have a conversation about how we can support your business or your fund.
00:03:59
And quickly, I want to specifically plug: we've been doing a lot of work with various companies and investors on not only tech issues, but specifically the export controls out of China and how it's impacting companies supply chains and what they can do to mitigate those supply chain risks. So, if you're having an issue, which I know many U.S. manufacturers are, and you need to sort of dive into your supply chain, and specifically the critical minerals inputs and the various inputs and look at diversification or other risk mitigation, reach out to us.
00:04:29
We have now a pretty specialized offering on that front. Otherwise, though, if you're interested in receiving more Trivium content in general, check out our website, triviumchina.com, where we have a bunch of different Trivium subscription products or a bunch of different subscription products, both free and paid. You can definitely get the China policy intel you need by signing up for one of the products on our site. And finally, please tell your friends and colleagues about Trivium, both about our work and about the podcast specifically, so we can continue to grow our audience and get the word out about the company.
00:05:00
So, with that out of the way, Kendra, let's dive into it. We're going to talk about this Action Plan, AI Action Plan. Again, July 23rd, Trump administration releases this new AI Action Plan, which outlines how the U.S. plans to supercharge domestic AI innovation and compete with China in AI. I've heard a lot of opinions on this or read a lot of opinions. I actually have to say that my feed, at least on things like LinkedIn, of the folks that I follow in the tech space, the feedback has generally, just like you said, been quite positive on the plan. So that, I don't know, it sounds like maybe you're in agreement there.
00:05:40
But it’d be interesting to hear your thoughts on that. More generally, seen some conflicting opinions around whether this is good or bad for U.S.-China relations in particular. Either way, the plan does seem to sort of change the game in terms of U.S.-China AI competition. Can you walk us through it, a high-level sort of your key takeaways and things we should be focusing on in that realm?
00:06:03
Kendra:
Yeah, cool. First, I better start, I guess, by explaining what it is for those of you who haven't sat down with it. It's not a horrible 100-page document. I think it's only 28 pages. So, 200 rated. It's not prohibitive. But essentially, the AI Action Plan is a series of directives that set the intention and direction of U.S. domestic AI development policy, and then also AI-related foreign policy.
00:06:27
And the plan basically has three pillars. One is accelerating U.S. AI innovation. The second one is about building American AI infrastructure. So, that's going to be data centers and energy. And then the third one is how the U.S. is basically going to deal with other countries in terms of AI, both how they're going to deal with allies and how they're going to deal with strategic competitors like China.
00:06:49
So, it sets the strategic direction in terms of how the U.S. is pursuing strategic AI dominance. It also instructs branches of government to sort of get a move on with some of those directives. It's really only the last section of the plan that talks about the U.S.-China relationship. So, that's the part that we'll probably focus on mostly today.
00:07:12
Andrew:
Awesome. Thanks for that overview. Just quick thing, again, staying on the high-level of this. I mean, correct me if I'm wrong here, but I thought there was generally already a pretty good understanding of the U.S. position in terms of AI strategy. Did we have that before, or is this a total departure kind of level set at where we were before the plan? That might be helpful.
00:07:32
Kendra:
That's a good idea. Yeah. So, it's true. We already had, for example, a general sense that under Trump, the U.S. was going to pursue things like aggressive expansion of energy capacity, aggressive expansion of data center capacity, and aggressive deregulation in AI. You know, JD Vance, you remember that, famously, I think, he attended the international AI Safety Summit in February.
00:07:57
And he basically told the Europeans at the Safety Summit that they shouldn't be hand-wringing about safety. So, we already basically understood that the administration's position was heavy on deregulation. We also knew that when it came to how international AI governance is going to work, the Trump administration is skeptical of multi-stakeholder collaboration in general. It's skeptical of the UN.
00:08:20
It's skeptical of any of these sort of international bodies. So, we did already have that baseline, a general understanding of U.S. position. But there was a lot we didn't know, particularly that hadn't really been confirmed or solidified on U.S.-China. For one thing, it's been very unclear how the Trump administration, and we've talked about this on this pod before, but how the Trump administration was going to come down on the issue of semiconductor export controls as an AI competition strategy.
00:08:47
Right? Because prior to the Trump administration, the prevailing logic in D.C. was sort of semiconductor export controls are a matter of national security, they're not negotiable, they're not a bargaining chip, they don't have anything to do with trade discussions, they're only going to get tighter and ever looser, and that trend seemed to be sort of an unstoppable juggernaut, regardless of what else is happening in the U.S.-China relationship.
00:09:10
And then there seemed to have been a sort of change in tone somehow since Trump took office. And we've heard, and I know you've heard, people say, often in meetings, Trump is the least hawkish guy in the room. He's not as hawkish on China as many in his cabinet. But what exactly that change was, whether Trump was going to be swayed by the more hawkish voices, whether he was going to be swayed by the more business-friendly voices, where he was going to fall on that spectrum has not been clear.
00:09:38
Andrew:
Yeah. So, are we clear now?
00:09:40
Kendra:
I mean, we know more than we do a week ago, but I don't know if we’re clear per se.
00:09:45
Andrew:
Cool. We will definitely get into that change of tone. I think it's quite remarkable. And we'll kind of tie that into sort of some of the most recent negotiations between the U.S. and China, having just wrapped up the recent round of negotiations in Stockholm. But before we get to that part of it, can we dive a little bit deeper in terms of what's in this plan when it comes to the chip control issue?
00:10:07
Kendra:
Right. That's probably the most interesting bit. So, the plan kind of confirms, at least for me, that there's been a significant shift in strategy on chips. As I sort of touched on, the strategy used to be about keeping all U.S. AI chip technology out of the hands of strategic competitors such as China. You'll remember the AI Diffusion rule, which was released at the end of Biden's term, was a pretty heavy-handed effort to control shipments of chips to everywhere, to sort of set caps on which countries got how many chips, and the U.S. was going to track and control all of that.
00:10:40
But that appears to be dead. And now the strategy is to sell as many U.S. chips and as much of the U.S. tech stack to everyone, the most people as possible, and then expand enforcement of existing controls so that China is only buying lower-tier U.S. technologies, and therefore domestic Chinese competitors cannot gain significant market share in the chip space. The idea is to take the wind out of the sails of China's drive towards technological self-sufficiency by feeding them just enough compute that their innovators can train models, do inference, compete in AI, but at a much lower level than the United States. So, that's the current trajectory.
00:11:23
Andrew:
I'm going to throw a quick spanner in at you. We've kind of gone over our questions beforehand, but I've sort of seen to the kind of modes of thinking around this shift that you just described. I mean, one is that, I mean, this is what the chip companies have been arguing to the administration, right?
Kendra:
Yes.
00:05:00
Andrew:
And I think obviously partly out of self-interest, but also, I think, partly out of a belief like rather than cutting them off, we should make them rely on our technology so that it slows down their push towards self-sufficiency. But also, then there's kind of this argument out there that, in a way, rolling back some of the restrictions, for example, on the Nvidia H20 and the EDA software restrictions that basically the Trump administration was backed into a corner because they realized how strong or how effective China's rare earth element export controls were. And so, they kind of had no choice but to go this route. Do you put weight behind one or both of those explanations on kind of driving this shift?
00:12:27
Kendra:
Geez, where do I even start with that? So, why don't I tackle your second question first? Right now, the evidence seems to point to the fact that that is not the case. And I say that because what we imagine would have happened in that case is that China, maybe MOFCOM, China's negotiators sat down with the U.S. and said, “We will release more of rare earths to you, we will lower our various export controls. We will ensure that rare earths shipments that you need are going to American companies, provided that you remove your restrictions on H20.” That's what that argument assumes occurred. But we're seeing a bunch of evidence over the last couple of days that indeed that may not be the case at all. Very strangely, after the U.S. lowered H20 chip controls, MOFCOM released a very interesting statement, which we flagged as rather bizarre, because MOFCOM, the Ministry of Commerce, said that the U.S. had unilaterally decided to lower those H20 controls.
00:13:30
In other words, they sort of implied that China never asked the United States to remove the controls on the H20 chips. Now, whether that's true or not, I have absolutely no idea. I'm not privy to those communications. And then, since then, this morning, you probably saw the news that, in fact, China's cyberspace regulators, the CSC, called Nvidia in to talk to them about potential security issues with H20s in China, essentially sending the message; we don't really need H20, or we're gearing up to kind of push back on you just shoving H20 at us.
00:14:04
And so, there's been a lot of conversation about whether or not, you know, if China doesn't really even want the H20s because they were perfectly happy with the trajectory of domestic innovation at the time, the fact that not having AI chips lit a fire under the butts of local innovators to really push forward chip self-sufficiency, and they were concerned that this U.S. olive branch of the H20 was a poisoned apple and has decided to sort of kick the brakes on that.
00:14:31
Then it sort of blows a hole in the argument that there was a chip for rare earths, tit for tat, back and forth. But maybe that's not what happened. Maybe it's as simple as the U.S. just unilaterally removed the H20 controls, and China went, “You don't get to dictate what our terms of the negotiation are and then unilaterally give us something we didn't ask you for. You have to wait until we have laid out our requirements, and then you need to respond to those requirements.” So, it's also fully possible that some of this communication happened there. Anyway, there's a few other possibilities in there, but I think essentially on the rare earth piece. There's a lot of assumptions about how that went down. The assumption is China's desperate for any chip it can get and will take any chip it can get.
00:15:08
The assumption is that China crawled to the United States and begged them for more chips. None of that may actually be true. So, there's that one. What was your first question again? I already forgot.
00:15:18
Andrew:
The other question, and then we can move on to other piece, but just, I guess, do you have any sense? I don't really have any sense of how much this change in thinking was sort of organic within the administration and how much it was potentially them hearing out industry players and sort of being convinced by that argument.
00:15:38
Kendra:
So, I think the industry probably played a very large role in convincing the administration to chill out. But there was already a pretty significant backlash to the previous AI Diffusion rule before any other option was on the table. There was a massive industry outcry, government overreach from the Biden administration. And not just from industry; there was a huge pushback from allies.
00:16:03
I mean, ultimately, I hate to say it, but it was a bad plan, and it was a bad plan because it was very, very heavy handed. It would have set up a very complicated bureaucracy, would have placed a huge burden on U.S. companies, and the biggest hole in it, or the biggest critique of it, which I personally agree with, was that, if you're Saudi and you're capped at 50,000 chips a year and you want 60,000, where are you going to get the other 10,000 from?
00:16:27
You're going straight to Beijing and you're going to say, “Well, your chips are crappier, but I'm just going to buy the extra ten from you.”
00:16:33
Andrew:
Good point.
00:16:34
Kendra:
So, sure, the U.S. could institute rules to try to prevent that and penalize, but then it just turns into this kind of monster. Basically, I think the biggest problem with all of that is that if we had started off with the strategy that the Trump administration has proposed now, which is where we already had been, basically, with some more controls and caveats, if we put a couple of export controls in place and then said, “Okay, we want to keep China using our tech stack, but we want to keep them a little bit behind,” I think that's a much more powerful strategy, frankly, than let's keep trying to turn the news, turn the screws, turn the news, shutdown, cut things off. Cut flows, cut trade flows, cut sales control numbers. That seems very onerous. And, actually, so if we had started that way, I think it probably would have been pretty effective.
00:17:25
Doing it now after the fire has already been lit under Beijing's but seems, okay, so now we're selling chips to China. They're still going to be well aware that at any minute. Right? The horse has ridden off into the sunset at this point. They already know that at any point we can change your mind and decide to cut them off the chips. And so, the impetus to continue, pushing hard for self-sufficiency, whether they have chips or not, is there now. So, I think we kind of did things out of order. That's just my view.
00:17:49
Andrew:
Yeah. That's interesting. It also could be as simple as the Trump administration just wanting to do the opposite of what the Biden administration did. Right? Yeah, that often that has been a sort of a feature of some of Donald Trump's moves, just wanting to just undo the previous Democratic administrations. So, that could play into it, but also kind of what we're getting to gets to the heart of the debate around whether or not export controls are actually effective.
00:18:17
We know they are effective in the short term, but they sort of undermine themselves in the long term. And you see the exact same thing happening with China's rare earth export controls. Yeah, it gives them a lot of leverage in the near-term in negotiations with the U.S. and negotiations with other countries. But, similarly-
00:18:36
Kendra:
But now everyone is scrambling to build their lives and reroute critical mineral flows. Now, the licensing issues have been removed, and everything's being facilitated.
00:18:46
Andrew:
Yeah. And they've absolutely concentrated the minds of policymakers elsewhere in the world that should have been focused on this issue years ago, to say, “Okay, now we really cannot be so reliant on China for rare earths. And the diversification has begun.” So, in a way, it seems like export controls are a very powerful tool, but one you can kind of only pull once, at least in this aggressive other way, right? Because you give yourself short-term leverage. But then folks, other countries are going to try to start to find workarounds in the long term. So, I think that's part of the ongoing debate in D.C. around the efficacy and appropriateness of export controls. It reminds me a lot of the similar discussions around sanctions. The problem is you never really know until years later whether or not it accomplished your goal.
00:19:32
But either way, back to the discussion at hand, does this mean that there's no more expansion of U.S. export controls on the table based on the new AI Plan?
00:19:40
Kendra:
Okay, not exactly. No. So, the plan does say the U.S. intends to close loopholes in existing export controls. And specifically, it says that it's going to close loopholes in semiconductor manufacturing subsystems. So, that basically means specific technologies and components needed to make the machines that make the chips. And that's probably like, and I don't know exactly which technologies they're thinking about, but probably things like, I don't know, precision lasers or sort of specialized metrology tools used for quality inspection, stuff like that.
00:20:16
So, to be clear, though, I mean, export controls already prevent China from importing those sort of high-tier machines that make chips, but this sounds like it will be targeted at preventing China from making their own machines that make chips. And they're going to focus on slowing down production that way.
00:20:33
Andrew:
So, they are expanding export controls. Where do we come down on that?
00:20:41
Kendra:
I mean, it's a good question, right? Basically, the devil is going to be in the details. This plan takes great care to frame what the U.S. wants to do as strengthening enforcement of existing controls and closing loopholes and existing controls, rather than adding further controls. But that kind of happened before, right? That kind of happened under Biden, where this was closing loopholes, and that was perceived by Beijing as major expansions in the export control regime, and yet another overreach, yet another overreach. So, I'm not 100%. It depends what technologies they target. It depends how they target them. I don't think the case is close.
00:21:15
Andrew:
So, this week it was an interesting one in terms of the ongoing negotiations, because if we start pursuing this plan, which we presumably do, will do, ‘we’ being the United States, while the U.S. and China negotiations are ongoing, it strikes me that this is just sort of ripe for another sort of miscommunication, sort of a post Geneva miscommunication around what exactly was agreed to in terms of the U.S. thinking that, at that time, China had said, “Yeah, we're going to send you rare earths,” and what China thought they had said was we will start the process of looking at applications for export control license or for export licenses for U.S. companies, which will take 45 days, and will be a bureaucratic cluster.
00:22:04
And so, obviously, there is a mismatch communication there. I think, now it seems like if the Trump administration thinks we're just boosting enforcement of existing rules, do they properly communicate that to Beijing? And does Beijing understand that and can live with it and understands that, of course, the U.S., as China will do, is really just restricting the true, true dual-use technologies, and that's an appropriate thing to do? Or does closing loopholes and better enforcement seem to Beijing as still actually an expansion in practice of export controls? That, to me, is kind of the question here.
00:22:47
Kendra:
It is the question. And yet that's a question we've been mulling on for the last six months. And now suddenly, if we read the text of this plan, that is now the least of our problems, I think. So, the bottom line is, on paper, the strategy has changed. It changed from control on the number of chips sold to sell as many chips as possible to China, but do a better job of preventing the best chips from ending up in China. And so, now that the focus is on preventing the best chips from ending up in China, it's actually pretty interesting how the U.S. says it's planning to do that in this document. The plan basically says that the Department of Commerce, OSTP and the NSC are going to partner with industries, so that’s U.S. chip chipmakers, to explore leveraging new and existing location verification features on advanced AI compute to “ensure that the chips are not in countries of concern.”
00:23:36
That basically means that what the government wants to do is it wants AI chip makers to build location trackers into their chips. You know, to prevent smuggling, essentially. You sell somebody just eat chips, and ten of those never get activated, or ten of those activate in a country they're not supposed to be in or a data center they're not supposed to be in. That would help the U.S. enforce existing export controls. But I don't know if you saw the news on that today. Beijing does not appear to like that idea.
00:24:01
Andrew:
Well, yeah. You already touched on this a little bit. This geolocation or geo tracking issue seems to be part of our very much the impetus behind China's cybersecurity authority calling in summoning Nvidia to talk about potential security issues in their H20s. Like I said, you touched on it a bit. Can you elaborate just a little bit on your thoughts on what's going on here?
00:24:23
Kendra:
I mean, I don't want to get totally derailed on that, but ultimately, I mean, I get it. If I was Beijing, the U.S. had spent the last five years scared to death that there was a secret back door. And TikTok and Huawei gear and Chinese cell modules.
00:24:36
Andrew:
Solar panels.
00:24:38
Kendra:
Solar panels and cars and port cranes, and then passing a bunch of laws dedicated to sort of ripping all these Chinese hardware out of U.S. critical infrastructure. And then the plan to control chips is, don't worry, we're just going to stick a tracking system in your critical infrastructure, I guess, probably not going to go down very well. And, indeed, it does not look like it's going down very well. Hard sell.
00:25:01
Andrew:
That's a tough sell. Yeah.
Kendra:
It’s a hard sell.
00:25:04
Andrew:
Yeah. That's going to be an issue that really is one to watch that we'll be looking at closely. I mean, it's one of those that I can see why policymakers and lawmakers in the U.S. would be like, “Oh, well, this is just a perfect plan. We'll just make sure we can track these chips and turn them off remotely anywhere in the world.” But you start to think about other countries, whether it's China or anyone, I'm not sure that other policymakers, even in allied countries, are going to want a kill switch if that's what even possible or that's what U.S. lawmakers decide to legislate, you know.
00:25:39
Kendra:
Yeah, exactly.
00:25:40
Andrew:
France won't want that either, right?
00:25:42
Kendra:
No. Well, I understand the impetus. From two steps back, I mean, it actually sounds like a great idea because now you don't have to have a bunch of BIS officials tracking down every single instance of every single sale and barking up China's butt all the time about, where did these go and where did that go? And barking up Nvidia; where is it? We need more stringent know-your-customer guidelines. We need you to spend a bajillion dollars on more lawyers and more investigators and more da, da, da. And it just seems like that is an endless wellspring of compliance difficulties for companies. It's an endless wellspring of… and potentially like quixotic efforts to stem the tide that can't be stemmed reliably.
00:26:22
So, taking a hands-off, tech-based approach where you just say, look, the little button sort of flash if these chips aren't where they're supposed to be. Sounds like a wonderful idea, but you can think about a little harder, as you say. And, mm, I don't know.
00:26:35
Andrew:
Then not sure it's workable. Yeah. Okay. Well, that's a lot of issues to continue to deal with here. I mean, obviously everyone's building the plane while we're flying in terms of trying to think about ways to regulate various aspects of AI development, including both the hardware and software. But any more on chips in this U.S. AI Action Plan that we should touch on?
00:26:58
Kendra:
So, the plan also deviates pretty heavily from what Biden was intending to do in terms of how it wants to get allies on board with U.S. export controls.
00:27:08
Andrew:
So, they're just going to give out candy to everybody? Is that what the plan is?
00:27:12
Kendra:
Yeah. And limited-edition Bibles.
00:27:15
Andrew:
Yeah, the Trump Bible. There we go.
00:27:17
Kendra:
The Trump Bible. Yeah. So, Biden was pretty reluctant to strongarm allies; the Dutch, the South Koreans, the Japanese, other chipmaking powers. He was pretty reluctant to really twist their arms and to cooperating. It’s a little bit of arm twisting, but he was mostly a carrot and a little bit of stick kind of guy. The Trump admin, of course, does not have any problem strong arming allies, as we've seen in the first half a year of his administration. And so, sort of paraphrasing here, the plan suggests that the U.S. is going to very stringently penalize countries that don't get on board with U.S. controls or don't align themselves with the U.S. controls by using the Foreign Direct Product rule and, of course, tariffs. You already know this, but for listeners who don’t, the foreign direct product rule basically extends U.S. export control powers to some foreign made items or foreign made goods, even if those goods don't actually contain any physical U.S. technologies or content, if those goods are produced using U.S. technologies or software or equipment, which is basically everything.
00:28:23
Everything on the planet is essentially made using some kind of something produced in the United States, some kind of IP produced in the United States. So, this would really be pulling out the big guns. It's very threatening. There was a little bit of conversation under Biden about whether or not to just kind of pull that ripcord, and ultimately, he did not want to upset allies by doing that. So, Trump doesn't seem too fussed one way or the other. That could be a real geopolitical flashpoint. And so, I mean, while that kind of stuff works to a point, Washington also risks seriously overplaying its hand with chipmaking allies. It could cause those kind of export control strategies to unravel at some point.
00:29:00
I mean, if they play that hand too hard. And I should say, changing the topic slightly, the strong arming doesn't really stop there. The plan also strikes this very strident, get on border, get out tone in terms of international air governance and standard setting, right. Basically, says the U.S. intends to lead the conversation in AI by heavily influencing these international political and technical standards bodies like the ITO and the UN, stuff like that.
00:29:26
Andrew:
I thought you said, I mean, rightly, though, that the Trump administration's pretty suspicious of these kind of multilateral sort of organizations and standards-setting bodies in general. How does that fit in here?
00:29:38
Kendra:
Yeah, we need to increase engagement. There's been complaints for years and years that the U.S. stakeholders show up to some of these international standards-setting meetings unprepared, uncoordinated. The Chinese side sends 30 Huawei engineers that all show up with binders. And they've prepped beforehand. They have one message. They vote in one block. They are able to influence standards that way.
00:30:00
We need to increase engagement in those bodies. It would be probably positive for the U.S. to do that. But kicking the door down was like, “I'm your daddy and you'll do what I say,” It's like that vibe is probably maybe not the best approach especially after being absent for so long and not paying child support. So, I don't know.
00:30:17
Andrew:
To push the metaphor a little bit further. There's one other thing, before we go on to like other aspects of the AI plan that are important, one last thing on chips and export controls in particular, to tie it back to the ongoing U.S.-China negotiations, you've said for months that one of the absolute most key pieces to watch in negotiations is how much previous export controls, so previous to April, would be on the table in terms of the loosening by the Trump administration. As you said at the top, everyone in Washington is, for years, seen export controls as a vice that only tightens, but it seems at least and I think Jamieson Greer, the U.S.TR, even mentioned that they were having these discussions at the most recent press conference after the Stockholm meetings, that they're actually discussing whether or not they might loosen other previous export controls, for example, on high bandwidth memory chips.
00:31:21
How much do you think that discussion is on the table, and how hard do you think the Chinese will push for that to be part of this negotiation?
00:31:30
Kendra:
I mean, I'm really honestly going on Trump admin vibes. My take on the Trump admin has always been that I don't think Trump… he's surrounded by many, many hawks that think we have seized the advantage with export controls and should continue squeezing. But my take on his position, and again, I hate to play Trump whisperer, because from one day to the next, it's such a dangerous game. But my take on his position has always sort of been that it's about trade deals and leverage for the U.S., and absolutely everything is on the table. That nothing is sacrosanct. Any leverage you can use to ultimately gain a better deal for the United States in trade and international trade terms, particularly if he could crack some kind of bigger nut that has been a two-decade problem…
00:32:17
Andrew:
So, a grand bargain or an omnibus deal.
00:32:19
Kendra:
Yeah, he loves a big grand bargain, an omnibus deal. Beijing is very unlikely to agree to that unless they get something juicy out of it as well. So, do I think it's totally off the table? No. Do I think there's going to be a pretty resounding outcry in D.C. from both sides of the aisle over that? Yeah, I think there probably will. I think he'll encounter significant pushback trying to do anything like that. So, does he care? Does the admin really care? I don't think so. But there does seem to be a point at which pushback from the politicos does…
00:32:53
Andrew:
Well, and not just the politicos, but obviously the national security…
00:32:56
Kendra:
Yeah. Apparatus. Yeah, exactly.
00:32:58
Andrew:
Community and apparatus would throw an absolute fit. And I think probably rightfully so, although I'm still a little bit open to the question of how effective these export controls are. So, just not take too far of a diversion here, but I mean, this is really a fascinating question to me. Why? And I will say I don't think Donald Trump cares about pushback. I think he's shown himself to be a singular politician who can get the Republican Party on board with things that they traditionally absolutely hate, and will come along with a little resistance. And the Democrats have very little authority, at least until the midterms, and maybe after the midterms, to affect his agenda. So, if he wants something done, I think conventional outcry from conventional politicians or whatever, from even the national security establishment, and Washington probably won't stop them.
00:33:52
But it does strike me that we kind of come back here to, what does Trump want out of these negotiations with China? Which has always been the sticking point for the Chinese. They don't know what he wants, so they don't know what to give him. And I would argue, I don't think Trump necessarily knows what he wants. That's not a knock. I just don't think he's determined, like, this is what I want to get from China in these negotiations. But if he does want the trade surplus to be reduced, I've always thought China could say, “Hey, you know what we could do? You relax those export controls and we will buy a ton of chips from you. And that will help lower the trade surplus. Like, that's the main thing we want to buy.”
00:34:33
Kendra:
Wouldn't that be crazy?
00:34:35
Andrew:
Like, we can buy soybeans from Brazil, but we'll buy chips from you guys. And whether or not that is an enticing argument to Donald Trump is interesting. And I mean, if I were the Chinese, I'd at least give that a shot, right? Now that they know there is at least some scope for export controls to be rolled back. And I would also just say, if the administration really buys into the argument of, you know, it's actually it's better to have China addicted to U.S. technology than it is to have China going full bore on self-sufficiency, one ultimate conclusion of that might be, well, actually, yeah, we can sell them increasingly more advanced chips.
00:35:18
Kendra:
I mean, but that argument is correct. I've heard some people in D.C. essentially say that the reason that argument is not correct is because that we can't measure where China might have been if we did not export control chips. That is the hot argument on that side. I disagree with that because I think that argument is made by people who were not in China in 2018 and felt the seismic shift in the vibes, the political vibes after the Huawei and ZTE controls.
Andrew:
Oh, yes, yes.
00:35:48
Kendra:
A major change occurred. We've been watching this forever for the last 15 years. Policymakers were horse whipping; “Please come up with an operating system to get us off of U.S. technologies. Please, please, please, please, please. We are sick of being married to iOS and windows and Android. There is this momentum around, or there was a push to make domestic data-based software that would be adopted that so they weren't dependent on Oracle, and all of it failed.
00:36:20
It eventually would have succeeded, but it was already slow. It was this grinding, uninspired slog where major Chinese companies kept buying U.S. products because it's easier to do that, then do, do R&D on your own. Chinese domestic R&D is extremely low as a percentage of revenue, and that has gone way up. So, in the last five years, right? Because firms now have to do it. So, if we had just started there, if we'd said, “Listen, we're just going to squeeze,” even if we just put in a couple export controls and said, “Okay, look, these two things, I'm sorry, we don't want you to have, we're just going to leave it there.” The most recent two versions of these, you can't have them. Sorry, but we use those export controls well.
00:37:02
In other words, we had a goal with them. There's an off ramp. If you do X, Y or Z, we will remove these controls. Whatever x, y or z was. We never gave China an off ramp. We never gave a carrot. We never tried to effect any sort of behavior change. We just said we're going to squeeze and squeeze and squeeze and squeeze and squeeze. Eventually, that was going to pop. I don't think anybody really cared if it was going to pop. So, from my perspective, a little bit of a dial down is going to prevent this from popping. But now it's like too late. There is a genuine national security risk because China's already got the bug on getting out as fast as possible from under U.S. pressures. Yeah, this could go one of many ways.
00:37:43
Andrew:
Yeah. The general point you make about 2018 really being a watershed that pushed China, it wasn't like in 2022, late 2022 when we put on the semiconductor export controls. That's not when China realized they need their own semiconductor industry. They realized that way before. That's a really good point. And anyway, we'll be debating this among ourselves and with our colleagues and friends in Washington for a long time to come, I'm sure. But let's stay the course here a little bit on the AI Action Plan and talk about a couple of other key elements of the plan. One is this focus on what to do with the development of open-source models. Can you walk us through what's going on there?
00:38:23
Kendra:
Yeah, it's pretty interesting. The plan states that the U.S. will encourage the development of open-source AI, Open Weight AI, I this is like a clear attempt to compete with Chinese open-source models like DeepSeek and Alibaba’s Qwen, and it's fascinating because what's happened in open source in China over the last year, particularly since DeepSeek came out. DeepSeek is really started an open-source revolution.
00:38:46
They open weighted their model that forced competitors, big competitors to open source their models. And now that is rolling all the way down the chain. Last week, 10 new open-source projects came out in the AI space in China, just within the course of a few days. And we were like, oh man. Essentially, the deal is now DeepSeek has set a bar domestically, which is if the best tool in China is open sourcing, why on earth would you pay for the second-best tool? So, you have to open source at this point. And that's essentially what's happening. But that is now raising questions globally about, well, the U.S. really only has one major open weight model, and that's Llama, Meta’s Llama. And my understanding is the licensing on Llama is more restrictive than the licensing on DeepSeek, for example.
00:39:36
So, you now have China coming out as not the global AI leader, but the global open-source AI leader. And so, this is an attempt to sort of compete with that. That said, I think the policy, like the U.S. policy agenda for doing that, seems a little underbaked. The plan basically focuses on providing compute infrastructure, access to colleges and universities. The idea is that colleges and universities are very often the source of impactful open-source projects. That's where those projects kind of originate from is well-funded laboratories, at least they have historically. So, I think there's like a hope that smart students are going to use it to create open weight models, as long as they're well-funded enough to do so. But there's a lot more the U.S. could do on that front. I think that's leaving some opportunities on the table there. I don't know that that's necessarily going to be as successful as they want.
00:40:25
Andrew:
Yeah, thanks for that. And one last piece in this AI Action Plan that specifically seems geared towards sort of targeting Chinese models, walk us through the dynamic there.
00:40:37
Kendra:
Yeah. This last bit was pretty interesting. This isn't in the third section of the plan. It's in the sort of how we're going to handle our domestic AI section of the plan. So, the plan calls on the Department of Commerce and the NIST Center for AI Standards and Innovation to, I'm going to read this quote here, “… conduct research and, as appropriate, publish evaluations of frontier models from the People's Republic of China for alignment with Chinese Communist Party talking points and censorship.”
00:41:07
So, we already know that DeepSeek is censored and produces output that is aligned with Communist Party talking points. We use DeepSeek regularly, and most certainly does that. It does things like it won't necessarily discuss Tiananmen Square or it won't necessarily discuss these sort of touchy topics that are touchy topics and censored topics domestically. So, those reports could serve as a basis to deny access to the United States of these models, which would probably be justified on those grounds.
00:41:38
There’s definitely an argument that there's like a strong bias in them. So, I think that opens up the possibility… You can kind of see that playing out right over the next couple of years. The NIST releases a report showing that these models do, in fact, censor; that they do, in fact, align with sort of Party policy and Party ideology. You then have a document that serves as the foundation for a law, and that law prevents DeepSeek from being deployed in X, Y, or Z scenario. So, potential paths to other Chinese AI bands in that paragraph.
00:42:11
Andrew:
Yeah. Not surprising. And also, I'm sure, yeah, I think most of, if not all the U.S. models, are either currently or going to be banned in China as well, which is not surprising, so that we’ll end up in a sort of bifurcated world in that sense. But just I wanted to touch on one other thing on that piece, which I know you guys have talked about internally. I've seen it on our slack channels in the tech practice, which is, I mean, are these models themselves actually censored, or is it just that they train on information and data that is censored because they train on the Chinese internet effectively and the Chinese internet is censored? So, is it the inputs that are censored? Is it the models that are censored?
00:42:52
And I guess a corollary question would be if it's the information that these models are trained on that that's the piece that's censored, would that influence any kind of banning, or is that too nuanced of a point to matter?
00:43:06
Kendra:
I mean, well, I'll answer your last question first. The U.S. doesn't operate on nuance in that way. And so, I don't think it matters for whether or not DeepSeek is going to be banned. There's actually a lot of interesting research out there on the fact that DeepSeek behaves differently when it's hosted versus local. In other words, if you download DeepSeek from an open-source repository and you are running it on your own computer, you're not going to deepseek.com or downloading the DeepSeek app and using the DeepSeek app, if you're using it hosted up on its own server, it's censored. Gives you different output if you download it and run it local. That seems to indicate that's not definitive. There are dozens of researchers that are running tests on DeepSeek on a very regular basis, and I'm sure the people at NIST will have a much more deep view on these things after they put it through its paces.
00:43:54
But my understanding is that what's being censored is the output. So, there's a censorship library or a right output filter that is preventing the hosted version of the model from saying anything that aligns with any sort of censored CIC blacklist of key words or topics, etc. So, our understanding is it's been trained on the Chinese Internet and also the U.S. internet. And so, as far as running deep domestically within China goes, they also have to filter the output of that because they don't want uncensored foreign, like domestically, they don't want uncensored English language content right there. Anyway, interesting problem.
00:44:33
Andrew:
So, it does seem that the models themselves are censored.
00:44:35
Kendra:
No, I’m saying the opposite. I think they're not. I think the model has been allowed to gobble everything, and then right now, because it's hosted online, when it spits something out, whatever it spits out, it goes through a filter.
00:44:48
Andrew:
I see. I got you.
00:44:50
Kendra:
And that filter could probably be removed if you download it. That seems to be early research seems to suggest that's probably the case. On top of that, though, internally it includes a bunch of data from the state and the Party. And so, when you ask it, something related to China, it spits out what it's…
Andrew:
Trained on. Yeah.
00:45:07
Kendra:
Right. Which is Chinese Party content.
00:45:10
Andrew:
Yeah, exactly.
00:45:11
Kendra:
You know, it's a little bit of both. A little bit of both.
00:45:13
Andrew:
I guess my point is it's not surprising that it's going to mimic Chinese talking points if it was trained on Chinese talking points, at least in large part.
00:45:19
Kendra:
Correct.
00:45:20
Andrew:
But that's an interesting one. Interesting point about that filter. Again, probably too nuanced.
00:45:27
Kendra:
Right, exactly. I don't think you're going to get a law that's like, you know, if you use the unfiltered version of DeepSeek, it’s fine.
00:45:33:
Andrew:
That's your U.S government voice, that's your policy-making voice?
00:45:36
Kendra:
It's my policy-making voice. It's like a late-night radio deejay.
00:45:40
Andrew:
Yeah, that's right. Well, yeah. Thanks for that. Thanks for indulging me in that kind of side note. I thought it might be interesting for listeners. I want to turn now where we probably should have started since we're a China-focused consultancy, but there was a lot to unpack in that U.S. AI Action Plan. So, I think it made sense to go through it all. But also, China released some sort of AI Action Plan last week as well, right?
00:46:06
Kendra:
Yeah. It's like, oh great, you get an Action Plan and you get an Action Plan. Yeah, so on the 26th, there was the World Artificial Intelligence Conference in Shanghai. That's a pretty big deal annual conference. And China released this document called the Action Plan on Global Governance of Artificial Intelligence. China's Action Plan is different than the U.S. plan. It wasn't like a mirror image of the United States plan. It is more like a lengthy statement on how China thinks countries around the world should governance and develop AI together. China's previously, for the last ten years, put out industrial AI policies on how it domestically should develop AI and build AI infrastructure and train AI talent. That's not what this was. This was sort of China's pitch to the international community about how we should think about AI.
00:46:51
Andrew:
And how does it say we should think about AI?
00:46:55
Kendra:
So, China essentially has set itself up. It's kind of been ramping up to this for the last year or so, maybe a year and a half since Bletchley, the Bletchley Summit in the UK, where it set itself up as a complete counterpoint to U.S. unilateralism. China has championed this idea that's the sort of very kumbaya AI togetherness, UN Sustainable Development goals. All countries in the world should hold hands and sing a song and develop AI together in a way that is inclusive, that does not deepen the digital divide.
00:47:30
Andrew:
Mm, sounds good. Sign me up.
00:47:31
Kendra:
It’s so wonderful. It's the most wonderful thing, particularly with a huge focus strategically on the Global South, ensuring that there's partnerships between co-development of AI infrastructure between the global South and China, etc. So, it his pitch is digitization, and the AI revolution shouldn't leave out the little guy. That is the public pitch that China's championing here. And they are also doing the exact opposite of the U.S., where they're essentially saying the UN should be the governing body of artificial intelligence, the UN should be the place where we should set up a sub-body within the UN that is focused on, you know, same like the Atomic Energy Agency or whatever it is, right?
00:48:13
Like some kind of nuclear control body for AI at an international level, but that place should be in the UN. In so doing, they're essentially seeking to align themselves with European values and come across as a benevolent dictator and a thought leader in AI development.
00:48:29
Andrew:
It's interesting. So, we've gone to the Chinese side pretty quickly, but put these two visions of how AI development should happen together. What's the rest of the world going to think, and which horse are they going to back, do you think?
00:48:42
Kendra:
I mean, neither. Neither of those plans or those visions really present an attractive or appealing vision of AI leadership. I mean, if I'm India, if I’m Japan, if I'm Europe, I'm Brazil, I'm not particularly excited to get behind either one of those concepts. I mean, the U.S. approach is very heavy-handed. It's very strong arm. It's, frankly, a little condescending. Right? We're going to decide what AI governance looks like in our country and in your country. You're going to go along with it or you're going to get screwed. And we hate these international bodies, but we're going to take them over and be the loudest voice in the room. I mean, that's the vibe in that plan on the international stage.
00:49:16
Again, I think it's much more positive on the domestic front, but on the how we're going to treat our allies and partners’ stage, it seems a little problematic. And then, as for China, China's leading but like nobody's following. China's approach reads to me is very disingenuous. China just doesn't have a established geopolitical history of sticking its neck out for the underdog. It has an established geopolitical history of maneuvering to gain the best edge for itself. That's clearly what's going on here. And I think that people see that and are just like, “Ugh, no.” But like I said, you know, as an American, as for the domestic pieces of the U.S. plan, I found those very, very positive. We did desperately need some high-level vision setting.
00:49:55
We now have that. I wish we weren't doing the deregulation of AI route. That seems dangerous, but I don't personally think pursuing innovation at the expense of well-being is the right path. But we have a path. We're on it and we're in it together. So here we go.
00:50:09
Andrew:
Well, this is all been fascinating. I do want to ask you one last question for we wrap up. I feel like it's just now kind of striking me. We've been talking about these AI Action Plans are strategies sort of in terms of how the rules of AI development are going to develop globally or play out globally. They're also sort of Action Plans around how to basically compete and win and who gets the spoils ultimately of AI. But we haven't talked about, and we'll do this on another pod, I just want you to touch on it for two seconds, the differing strategies between China and the U.S. in terms of what they're actually trying to achieve with AI, right?
Kendra:
No, yeah.
00:50:53
Andrew:
I'm going to lay it out in a dumb way, so tell me if I'm wrong here. But in generals, most simplistic terms, the U.S. is really barreling towards artificial general intelligence. AGI, like that's the nirvana that we think is going to unlock all kinds of scientific discoveries, military applications, all that stuff. China, at least for now, seems more focused on sort of finding very practical, real-world and real economy applications for AI that currently exists specifically in manufacturing and industrial processes, trying to use AI to upgrade their manufacturing juggernaut.
00:51:34
And those are two kind of different in goals for the moment. Do I have that right or? Your thoughts on that. Tell me where I'm wrong.
00:51:43
Kendra:
I think you're absolutely right. I mean, no notes. I mean, so China has always, I've probably said this on the pod before, but so forgive me if I have, but China's approach to… I have this amazing document that's 20 years old that is a speech that Zhu Rongji gave at the meeting of policymakers in 2002, where the Chinese government was trying to figure out how they were going to develop the internet, right? And how they were going to spread the internet is one of my favorite ever documents. And in that speech, he basically said, we are excited about the internet. We're excited about web 2.0 because it's going to help us upgrade our industrial processes and upgrade our manufacturing. That is what it is for. It is to support the real economy.
00:52:26
Every new emerging technology in China, policymakers get excited about it because it's going to support the real economy, and because it's going to upgrade traditional industries, and because it's going to support manufacturing. That has been the through line in the way the state thinks about all of this stuff for the last at least 20 years. And now you have this interesting situation, I don’t want to get too far down the rabbit hole, but you have an interesting confluence of economic concerns that China's worried about, all these economic headwinds. It has a shrinking population, meaning it's going to have less human resources. It has slowing GDP. So, it's got, how do you do more with less as an economic power? You have to use what you already have more efficiently. And AI is the hope for China to use its economic resources more efficiently.
00:53:10
If it can pair big data with AI, then maybe these things will be allocated more rationally throughout the economy, particularly important in a state-planned economy where like the government is making decisions about how resources get allocated, and not always well; usually pretty crap decisions actually. So, that's China. Meanwhile, the U.S. is pursuing this idea of AGI. We have basic definitions for what we think AGI is, but there's no standard. There's no accepted standard for the moment we're going to wake up and say, “Oh, there it is. That's AGI. We have made it.” It's not Sputnik. It's not like, okay…
Andrew:
We did it.
00:53:45
Kendra:
Shot this rocket off and we did it. So, it's a complicated question. I think the big concern is that the U.S. is really worried. The big debate that I keep hearing in D.C. is, is China pursuing AGI? We're in a race to get to AGI before China. Is China even trying to pursue that? And the answer, I can tell you, is that, honestly, in most of the policy documents we read, when China does pursue something, there are usually policies about it. There's a huge state of national plans. You know, our AGI plan. There are AGI plans.
00:54:15
Policymakers have not really outlined AGI as a major goal. What they have done is focused on AI as the, as you say, the driver of economic upgrading, but also a lot of interest in embodied AI. In other words, can we put an AI brain in a robotics body? That's the focus. And that's also about manufacturing, industrial upgrading, etc. But anyway, I'll stop there.
00:54:34
Andrew:
Well, yeah, I want to do a whole pot on this at some point soon. I will say just to the final point, which is in a way, it may not matter if China isn't trying to achieve AGI “China” as a policy goal because we know DeepSeek is, right?
00:54:47
Kendra:
Yes, 100%.
00:54:48
Andrew:
There is a company, and probably Alibaba and others are investing.
00:54:53
Kendra:
Exactly. And the plan might come out. That plan might come out, right? AGI is a topic that was like, as in the last six months, everyone started caring about it, like policymakers might release it in six months and then, okay.
00:55:02
Andrew:
Totally. Well, I mean, I guess my point is, in a way, it doesn't matter whether the state or “China” is trying as a policy to achieve it, it still might achieve it through its nominally private sector companies. And that would still be very concerning for U.S. policymakers. So, we'll do a whole pod on that soon. But we've covered a lot of ground. This was fun, Kendra. It's great to be back in the saddle, and really appreciate you joining me and walk me through all this.
00:55:29
Kendra:
Of course. Always a good time.
00:55:31
Andrew:
Awesome. All right. Thanks a bunch, Kendra. Thanks, everybody, for listening. We'll see you next time. Bye.
Share this post