Beijing is moving to restrict how far its own open-source AI models can travel abroad, just as Washington moves to restrict how far they can travel in.
Same standoff, squeezing from opposite ends of the pipeline.
Quick note: This episode is the first of a new, more frequent style of podcast we’re rolling out – shorter, faster-turnaround conversations to react to the news as it breaks, alongside our regular weekly deep dives.
Look for more of these in your feed soon.
On this episode, Andrew Polk sits down with Kendra Schaefer (Head of Tech Policy Research) to unpack:
Why MOFCOM’s reported talks with Alibaba, ByteDance, and Zhipu on export controls could reshape how Chinese model weights get released
The three priorities Beijing is trying to balance: model competitiveness, security risk, and controllability
How Kimi K3’s release has intensified US anxiety, with Treasury Secretary Scott Bessent floating possible sanctions over model distillation
What Xi’s WAIC speech signals about China staying committed to open source while tightening the regulatory leash
Give it a listen and let us know what you think.
Transcript
Andrew Polk: Hi, everybody. Welcome to the latest Trivium China Podcast, a proud member of the Sinica Podcast Network. I’m your host, Trivium Co-Founder, Andrew Polk, and I’m joined today once again by our Head of Tech Policy Research, Kendra Schaefer. Kendra, how are you doing?
Kendra Schaefer: I’m good. I’m good. How are you?
Andrew: Yeah, glad to have you back on so quickly, which it’s a rare treat to get such a quick turnaround. I’ll talk about this a little bit in the housekeeping piece, but just for listeners — we’re doing a little bit of experimentation with putting out content more frequently, which means some of it will be shorter. I’ll talk through some of that in more of it in just a minute. But the reason specifically that I had Kendra on was there were some big developments, even since we talked last, in, of course, the China AI space.
The big news is sort of that Chinese regulators led by MOFCOM are reportedly in talks with some of the big hyperscalers — Alibaba, ByteDance, Zhepu, about potentially tightening export controls on AI models and chips going out of China, potentially even limiting overseas transfer of training data and model weights. And so, this is like a big deal in terms of potential restrictions from China’s side, but also we’re seeing potential restrictions from the U.S. side, which we talked about two weeks ago when I talked to Kendra.
And that has even ramped up further with the release of Kimi 3, Moonshot’s latest release, with now U.S. policymakers really worried. So, there’s like this very narrowing space where the U.S. looks like it’s going to increasingly restrict the Chinese models and China’s going to restrict Chinese models from going abroad. We’re going to talk a little bit about that, and then also add the context of Xi Jinping giving a big keynote speech at the WAIC. What does that stand for, Kendra? World AI Conference?
Kendra: Yes.
Andrew: Okay. Where he kind of doubled down on China’s commitment to open-source AI. A lot of just related topics. I want to get Kendra’s quick take on while it was top of mind for folks. But of course, we’ll do this quick. Got to do the quick vibe check. How’s your vibe today, Kendra?
Kendra: Well, I’m a little under the weather, actually, but I really wanted to do this podcast. So I’m all hopped up on DayQuil and ibuprofen. So I don’t know if overmedicated is a vibe, but if so, that is my vibe.
Andrew: It’s definitely a vibe. I love it. Well, this will be a trip then. I’m looking forward to it. My vibe is excited to kind of experiment with some of this more frequent, quick-hit content. So, excited about that. We will do a little bit of quick housekeeping reminder to anyone listening: if you’ve just got this or if this was forwarded to you, Trivium is not just a podcast. We are a strategic advisory firm and we work with businesses and funds to figure out China. So, we are a strategic advisory that helps businesses and investors navigate the China policy landscape.
That includes policy in China across a range of areas, but also policy towards China out of Western capitals like DC, London, Brussels, and others. So, if you need any help on that front, please do reach out to us at hq@riviumchina.com. Otherwise, listeners, please do leave us ratings and reviews on your favorite podcast apps. It really helps expand our reach and grow our business.
All right. The other housekeeping piece quickly. So, we are going to experiment with more frequent podcasts, more frequent content. It’ll be shorter, many of them. We’ll still kind of have our weekly anchor podcast kind of talking about bigger themes, but I just want to get more Trivium voices on, talk about stuff that’s really top of mind that people may be wondering about in the news. And so please look for us in your feeds more regularly. This isn’t just going to be just suddenly you see three or four Trivian pods. In a couple of weeks, we should be in your feed on a more regular basis.
So, make sure you subscribe to the podcast on whatever platform you use. All right, let’s get into it, Kendra. So, we’ll start with the MOFCOM piece. So, The FT in particular reported that Chinese regulators, led by MOFCOM, the Ministry of Commerce, are in talks with the hyperscalers about tightening export controls on AI models and chips. We’re talking again about potential overseas transfer of training data and model weights, and apparently even some discussion of preventing offshore firms from fabricating Huawei, Alibaba, and ByteDance design chips.
This all matters because, as we’ve talked about on the pod before, the whole reason that Chinese models like DeepSeek and Moonshot’s models have gotten global traction is precisely because they’re open weight, and anyone can download them and run them locally. So, if Beijing starts clamping down on that, it’s a huge deal. Why don’t you talk us through kind of, I know you’ve got a big picture understanding of what’s going on here and also kind of the different priorities that policymakers might be balancing when they’re thinking through this stuff.
Kendra: Yeah, totally. So we don’t know exactly what MOFCOM is going to do or specifically how they’re going to approach this. I’ll come back to my hypothesis on what tools they might use maybe later in the podcast. But we do know what considerations Beijing is trying to balance right now. It’s trying to balance three considerations. First, the competitiveness of models. Second, these sort of big, scary cybersecurity questions around models. And then on top of that, controllability.
And so, it’s probably helpful to paint a little bit of a picture in terms of what the state is thinking on each of those three aspects. So in terms of competitiveness, I mean, I think most of our listeners know that there’s been hope in China for 15 years that the country could sort of develop an alternative tech stack to the U.S. tech stack that would gain international traction. But prior to LLMs, you know, the platform economy and sort of cross-border e-commerce represented China’s best hope to do that.
TikTok was kind of the first big Chinese platform that American users loved and that really got a lot of traction in the U.S. And also, you’ve got some minor stuff like Temu and Shein with a mixed reception, but still relatively popular in the United States. But TikTok, Temu and Shein face direct competition from bigger, better established, better funded U.S. competitors. Temu and Shein have to go head to head with Amazon. TikTok has to compete with Meta and X. But look at what’s happening now in AI.
It’s crazy. Companies all over the world, including all these big companies and a lot of small firms like us in the U.S., want cheap AI technology that they can control. But there’s a huge demand for that. And right now there’s really only viable Chinese supply. There are like very few competitive U.S. models that meet that demand. U.S. models meet the demand for high-quality, very secure, bleeding-edge AI. There’s no great Chinese competitor in that space. And of course, there’s a huge market for that as well.
But they don’t meet the demand for cheap, controllable AI. Right? Chinese firms do. So, this is a once-in-a-lifetime opportunity for China. Even though Chinese tech firms have a sort of fraction of the resources that U.S. firms do, because they’re really the only viable players on the field in that particular segment of the AI market right now, they’re getting a ton of traction. My personal opinion is this isn’t even actually a U.S.-China issue. It’s just a supply and demand problem. And then China has this sort of additional incentive of, you know, it’s not just about can our firms gain market share in the U.S. or EU.
You know, Chinese tools also have actually probably much more long term durable prospects in the global south. Startups and government institutions and companies in countries where the startups don’t have that much money, where you know they’re scrappy, where there’s a lot of R&D firms that don’t have funding, of course, they’re going to want to use a model that’s cheaper, that’s easier to access, there’s lower barriers to entry, there’s not so much sort of subscription control or geofencing or any of that sort of stuff.
So, as you said if Beijing were to ban the export of Chinese open source models on a permanent basis that would essentially amount to ceding the most momentum China has ever gained due to risk aversion. So it would be a huge deal. So that’s one piece of the puzzle. And Beijing is certainly weighing that piece.
Andrew: That’s the competitiveness piece I’m talking about.
Kendra: That’s the competitiveness piece, exactly.
Andrew: So you said competitiveness, security, controllability. Okay, talk us through security then.
Kendra: We’re trying to do a faster pod here, so I’ll try to keep these shorter. But for safety...
Andrew: No, no, people will love it. It doesn’t have to be short.
Kendra: Okay. So, I think the safety concerns are becoming pretty clear to everyone.
I don’t know if you also saw the news just yesterday that this OpenAI was running some tests on new agentic system, and it lowered the guardrails on the system. My God, it was so terrifying. They lowered the guardrails on the system and basically told this tool to like do its worst, essentially, just to see what it could do. And it broke out of its box and went to Hugging Face, launched a cyberattack against the Hugging Face website, totally autonomously, right?
Andrew: Insane.
Kendra: Anyway, there was a big kerfuffle over this. OpenAI and Hugging Face have since cooperated to sort out the issue. But obviously, every week there’s some other terrifying headline about the cybersecurity risks of these models. You know, they’re doing things that humans didn’t expect them to do, etc. So, these safety concerns are very, very real. And open source tools have an even bigger safety problem than closed source tools, which is that if it’s a closed source tool and it is discovered to be hackable or breakable or the guardrails can come off or there’s something seriously wrong with it, the company that manages it can pull it off the market today.
But if you release an open source tool onto the market, you release open weights onto the market, it’s gone. It’s out of your control. That’s the end. And so, the risks for China to be the source of such a tool, let’s say a Chinese company releases a tool like that onto the domestic market. And suddenly there’s a bunch of attacks against Chinese critical infrastructure, even by domestic actors, even, right? Or by foreign state-sponsored actors, that is a direct risk for China, built by China to China.
But similar problem, even if those tools aren’t even used to attack China. I mean, the amount of geopolitical tension that would arise if a bad actor took a Chinese tool and then was using that to attack other countries and China was seen as this irresponsible global actor after they have, you know, talked up safety, etc. And then on top of all of that, It’s like Chinese regulators are some of the most risk-averse humans on the planet.
Andrew: Yeah. I was just quickly going to say, are you saying it’s like some sort of disruptive element escaped to the borders of China and caused a global issue? Have we seen something like that recently? That everyone would blame China for it? Yeah, I think we have. Yeah, well, so I’m sure. Yeah, absolutely. It caused a huge issue. But before you go to the controllability piece, just quickly, our team noted that Hugging Face use open source Chinese models to do the security postmortem to try to figure out what went wrong and fix it. Anything to take away from that?
Kendra: Yeah, that’s a whole nother rabbit hole. It’s like essentially what happened was Hugging Face tried to use in their own blog post when they were reporting on the incident, they essentially said we tried to use commercially available U.S. tools. Presumably, they’re talking about, you know, Anthropics tools or OpenAI’s tools. They tried to use the sort of leading edge commercial tools, and the guardrails on those tools couldn’t distinguish between a company trying to defend itself from a cyberattack and a cyberattacker.
And so those guardrails were triggered and they prevented them from using that tool. And so what happened was Hugging Face reached instead for Zhipu’s GLM 5.2 to do it because the guardrails were not preventative. So, I mean, need I enumerate all the different ways in which that’s scary? It’s also actually a really good argument for, again, China doesn’t really have anything to do with this argument. It’s just like people need, they want controllability. The market wants controllability over their tools, partially for this reason.
Partially because they don’t want, there are situations in which it doesn’t actually make sense or it’s actually unsafe or there’s a security risk to having a middleman tell you what you can and can’t do with the tools that you have access to. So, that’s a real tension, right? It’s the biggest tension, obviously, in the AI ecosystem at the moment.
Andrew: Thanks for that. Sorry for the quick detour, but okay, now let’s talk about the controllability piece, which you were about to go to.
Kendra: Yeah, so you and I have been talking about this for a while. A couple of months ago, we talked about the Meta-Manus deal. And around that time, we flagged for our clients that the Meta-Manus deal caused a shift in the way that China thinks about technology, right? Chinese regulators at the time hadn’t really had to consider the possibility that now that China’s making some bleeding-edge technology, that China needs the regulatory tools to control the outflow of that technology, right? That’s what the Meta-Manus deal sort of revealed to regulators.
And we flagged for clients at the time, this is actually going to be a major theme over the next couple of years, is that regulators are going to start giving themselves opportunities and options for saying no to specific deals, to saying no to a lot of different technology exports, to saying no to the transfer of our IP to foreign companies. And so, we expected to see that as well. And so now you have this interesting situation where Chinese models are starting to, with the release of Kimi, for example, walk up closer to the frontier.
I still think they’re relatively far behind the leading edge, but walk up close enough to the frontier that it starts to become a question of, well, these models are open weight. They don’t have a lot of restrictions on their use. There’s no framework for restrictions on their use. Does that mean a foreign company can just take those models and modify… what technology can be squeezed out of them that China doesn’t want to lose? Regulators don’t really know. That’s not really something they have considered before.
So, I think all three of those things are weighing on the state, are weighing on top leaders, and are weighing on MOFCOM as they consider what to do about open source Chinese models.
Andrew: All right. Well, thanks for laying that out. We’ll come back to how that may be impacting what they do next. But with that sort of context of where are the key priorities that policymakers are trying to balance, let’s now bring in Xi Jinping’s speech at the WAIC conference. It’s a lot to unpack. The headline really is that Xi Jinping recommitted or reiterated China’s commitment to open source, even though there’s been this reporting that China might be considering kind of putting on some restrictions. So, why don’t you just talk to us about the key themes that you saw from Xi Jinping and how this plays into what you just said about policymakers competing priorities.
Kendra: I think Xi’s speech was a perfect summary of all of the issues that China is currently dealing with that I just outlined, plus a bunch of stuff about international AI cooperation also at the end. But, you know, he basically said both, right? He said both things. He painted AI as a new industrial revolution, blah, blah, blah. We’ve heard that from… kind of most policymakers agree on that point. But he also really underscored some of the major risks that AI was presenting for China and for humanity. I’m actually going to read those questions here because I think they’re kind of interesting. He said, “When machines begin to think, how should humans coexist with them? When algorithms take part in decision-making, how can safety be guaranteed? When technology challenges ethics, how can governance keep pace? And when the divide keeps widening, how can inclusive benefits be achieved?”
And so, then he gets into his kind of proposals about how China may answer those questions. And his first proposal was that China should stay committed to openness. And again, I’m going to read out what he said because the devil’s in the details. He said, “Artificial intelligence is a new engine of global economic growth and an accelerator of the shift from old to new drivers of growth. It’s moving from the digital world into the physical world.”
In other words, robots are starting to be powered by AI. And he says, “We should seize this rare historic opportunity, encourage open source development and cooperative sharing,” there’s the money line, “and comprehensively promote AI innovation,” blah, blah, blah. So, he says straight up that the number one principle, right, and often these principles get listed in order of importance and priority. Right up front, he says, “We’re going to stay committed to open source and cooperative sharing.”
In other words, we’re reading that as China has no intention of taking a step back from open source technology. We still think that’s a great tool, and we intend to share Chinese open source models with the world, with other countries as well. But then he comes in from the other side. He pivots to addressing the safety risks. He said, “Second, we should strengthen risk awareness to ensure that AI remains safe and controllable. Artificial intelligence should be a trustworthy tool for humanity.
We must attach great importance to the various inherent and derivative risks that AI gives rise to and work to build systems of laws and regulations, technical monitoring, risk early warning, and emergency response so as to establish a firm safety baseline, guard against misuse and malicious use, and ensure that AI always remains under human control.” And then he kind of gets into some global cooperation stuff on how China will partner with the global south, etc. So, what matters there is they say they’re going to keep supporting open source, but they’re also going to put much more stringent safeguards in place relative to AI technologies.
Later in the speech, he reiterates that, and he says, “We’re going to continuously improve relevant laws and regulations, policies and institutions, norms and ethical guidelines to ensure that AI is safe, reliable and controllable so that this,” love this part, “so that this galloping steed of artificial intelligence runs both fast and steady.” I know. And then he goes on to say, “The more rapidly AI technology advances, the more firmly its direction towards goodness and the benefit of humanity must be anchored, the more precisely the measure of regulation and governance must be calibrated.” Blah, blah, blah.
So, all of this is centered around a clear message that there’s going to be a very sharp uptick in Chinese domestic regulation on AI and China’s efforts to control, likely control the pace of model releases. There’s no question in our mind that Beijing is seeking to do this. So, the question just becomes, right? the question just becomes, what is the tool they use to do it? They’re obviously going to do it. They’re obviously going to try to control probably some kind of technology exports or the release of models in some way or, or, or…
But the question is, what level of ban is that? How much regulatory involvement will that have? You know, we have some clues, but we’re not exactly sure yet.
Andrew: Well, can you talk us through the clues? Like what forms you think they might take?
Kendra: I think that when interests conflict in technology regulation, China almost always tries to kind of thread the needle. And I think, this is just my guess, I think this is just my hypothesis, but I think the best way to thread the needle in this case would be to sort of create a lag between the availability of an open model and the release of the weights. So, in other words, maybe you can access a model online through a provider right away. If you want to ping a hosted version of a new frontier model now or a new open model now, great.
But if you want to download the weights, the model has to go through some kind of very stringent security assessment or some kind of licensing or approval process before those weights actually become available online. And so that’s the really interesting piece. You said in the beginning that, and this is what we’re watching really closely, and I think I’m on the edge of my seat on this one, you know, there have been multiple reports that MOFCOM is leading the discussions on exactly how that might happen.
And if MOFCOM is leading the discussions, that tells us quite a lot because MOFCOM only has a couple of tools available to them. So, if they are the regulator that is going to move forward with controlling model releases, that tells us those releases may be controlled with the tools that MOFCOM already has. Most notably, MOFCOM handles China’s export control regime, right? And they do that through the same export control regime they’re using for rare earths. They do that through this regulation called Regulations on Export Control of Dual Use Items, which basically creates this centralized dual-use control list.
When MOFCOM puts something on that list, exporters have to apply for a license in order to export it, just like what’s happening with rare earths now. Something goes on the list, you’ve got to go to MOFCOM, get a license to export it. And there’s a few different kinds of licenses you can get. You can get a license to send it to just one company. That wouldn’t make a ton of sense here. You can get a license to provide it to a larger audience, etc. But that list has never been used to control something like this.
So, it would actually be super interesting if MOFCOM decides to use the export control regime because it opens a question, is posting model weights online an export? And I guess it could be considered an export under this regime. They also have this other tool called the Catalog of Technologies Prohibited or Restricted from Export. Basically, that’s not a dual-use issue. That catalog is managed by MOFCOM and the Ministry of Science and Technology together. I think, if I recall correctly, it’s under the foreign trade law, not under the export control regime.
And this is the list that China used to control or to kind of prevent the unapproved outbound transfer of TikTok under the TikTok divestiture kerfuffle. And that list too, if I recall correctly, has two sections. It’s got a prohibited section, like this is definitely banned from export under any circumstances, and then a restricted section, like you got to get a license if you’re going to export this. So, functionally speaking, MOFCOM has a couple of lists that they could use to stick model weights on.
Some of the language related to some of these lists, or at least the second one, could already be considered to apply to model weights. But I think they could just go ahead and make it very explicit, add model weights to maybe the export control list. And what that would do would be to give China a very granular pathway for controlling releases on a model-by-model basis. It’s a bit clunky, and I actually think it’s a terrible pathway because there are a hundred ways it could go wrong.
Andrew: Yeah. So I want to pick up on two last pieces then. One is what you see as potential unintended consequences. How can things go wrong? And then I want to circle back to the U.S., what the U.S.
is doing because we’re talking about the Chinese restrictions as China trying to restrict technology outflow and U.S. is trying to restrict Chinese technology inflow. So we’ll get to that piece in a second. But first, how could things go…? No, surely nothing could go wrong here. Nothing could go wrong.
Kendra: Sucks to be a Chinese frontier lab right now. Everybody’s gunning for you. So the first thing that’s on the top of mind is that MOFCOM doesn’t exactly have a history of granting export approvals in a timely manner. I mean, they are just terrible at granting export approvals under this particular regime in any sort of streamlined way. I mean, I think a lot of our listeners are probably watching the whole rare earths saga, even when top leadership, when MOFCOM itself, when the receiving party, when the selling party, in other words, even when every single party involved in a licensed transaction wanted that to happen smoothly, it did not.
That regime was beset by delays. That is one thing when you’re talking about a commodity. It is a very different thing when you are talking about a market that is moving so quickly that every two months, the entire shape of the market has changed. And it’s actually not just MOFCOM. I mean, every single export control regime I can think of, even those not controlled by MOFCOM, like cross-border data exports approved by the CAC was a snarl for two and a half years to the point where EU regulators were bringing it up with Xi Jinping at meetings saying, “Hey, we can’t get data, we can’t get access to data.”
So, I think that’s a potential where, in fact, I would almost guarantee it that if MOFCOM steps into the room, model releases will not be approved rapidly and there is going to be some kind of problem with the speed of exports. So that’s issue number one. The second problem is that MOFCOM doesn’t know anything about safe AI. In other words, at what point does MOFCOM issue an approval on what basis? It has to be the basis of some other kind of security test run by somebody else, right?
And those security tests, what constitutes safe artificial intelligence is still very much an open question globally. I mean, the U.S. is working with our… you know, we’re working with our best tech company as USG is working with, you know, sort of top AI companies to figure that out. And they’re going to be figuring that out for a couple of years. And so, if you put a bureaucrat in the room and you say, “This is exactly what happened with data exports,” you say, “Only allow people to export safe data, but then you didn’t tell them what safe data is.” It’s exactly what happened with data exports.
And regulators just stopped approving anything because they didn’t have a good definition and they couldn’t save their own butts. They couldn’t justify why they allowed something. Nobody wanted to be the one that rubber-stamped a major security risk.
Andrew: Yes.
Kendra: So I think that is the second problem. The third problem is I think China could definitely get too wrapped up in its own narrative about technology outflow too soon. We’ve already seen a couple of examples where the state made an ideological decision, ideological-ish bet that did not make sense from a purely tech competitive perspective. And here I’m talking about Beijing refusing to immediately allow the purchase of as many NVIDIA chips as its companies wanted as soon as the U.S. approved those exports, approved those sales. China sided on delayed for a while and focused primarily on self-sufficiency and said, “Look, we’re going to allow a couple of these.”
And meanwhile, a lot of the AI firms are going, “Please, please, please, we need as many chips as we can get. We’ll get them from anywhere. We’ll take them from anywhere.” So, it wouldn’t be unusual or strange to me if the state essentially talked itself into this idea where, look, we can’t allow too much technology outflow. We can allow foreign companies to access Chinese hosted models, but we can’t just freely and openly allow model weights to be distributed anywhere willy-nilly. We’re going to keep model weights domestic. We’re only going to da, da, da, da, da, which would absolutely destroy the one market advantage that China really has, which is that you can get the model weights…
I think China could just make a bad decision about over-securitization and talk itself into something dumb. And then finally, I think this is true for probably the U.S. and China. There could definitely be some kind of black swan event, where some, you know, we’re already seeing hints of that, just like this sort of OpenAI issue. Luckily, nothing terrible has happened yet. But there could definitely be some major security intrusion or breach or event that sends both China and the U.S. scuttling back into a crab shell, going, “Absolutely not, lock it all down. Nothing is released without approval.” And China being particularly risk averse, I imagine that we could see some serious sort of knee jerk security backlash to something like that.
Those are the four things that I think could derail any efforts by China to institute reasonable, rational safeguards, which are desperately needed, by the way, you know, on open-source model releases.
Andrew: Yeah. Like I was alluding to with the pandemic, don’t want to get into the hole; did China do this on purpose or not? I don’t want to go down those rabbit holes. But if there were an accident and some open-source model released from China started causing problems in the cybersecurity world globally, there would, of course, be an assumption that the CCP did it on purpose, right? And then that, I mean, you talked about geopolitical tension, I mean, that’s a war, I think, you know, waiting to happen.
Kendra: Yeah, 100 percent, 100 percent. I mean, China will avoid that at all costs. So, I don’t think it, but I’ve heard some of this chatter on X or whatever that people think China’s in control of every aspect of this. I mean, we are watching regulators in real time respond to like new catastrophes or new sort of angles on this question. Every week, there’s some other thing nobody thought of that has to be addressed and that regulators aren’t prepared for.
Andrew: Everyone’s flying or building the airplane while they’re flying it. Everyone everywhere, right?
Kendra: Exactly. Yeah.
Andrew: And the companies are trying to stay ahead of the regulators as is happening in the U.S., that’s a whole nother podcast. But let’s now turn to what’s happening in the U.S. So, already the U.S. was making a bunch of noise about potentially restricting access. I think we touched on it maybe in our last pod, potentially restricting access to Chinese models because a bunch of companies in the U.S. are increasingly using these models because of everything you said, because they’re cheaper. And a lot of times you just need kind of bread-and-butter type AI usages, right? You don’t need the bleeding edge models. That was before the release of Kimi 3, which caused a lot of waves.
We won’t go into the details, but basically it was much closer to the frontier of models that many people thought China was. And now, I mean, even today, Scott Bessent, maybe it was yesterday, the U.S. Treasury Secretary saying, “Oh, you know, we think there was a ton of distillation to train these models. We might sanction or entity list Chinese companies that are using mass distillation. So, talk about what you think of the U.S. discussion around restrictions on Chinese models. And then we’ll end by just talking about how there’s very clearly a pincer on both sides, China trying to restrict outbound access, U.S. trying to restrict inbound access, and what that might mean for the future.
Kendra: I mean, I think there’s a lot of angst, but I don’t think there’s a real strategy. And what Bessent said was, we’re going to take a look at whether or not, I mean, how I heard that was, we’re going to take a look at whether or not… I mean, how I heard that was we’re going to take a look at whether or not we can paint distillation as a major IP theft problem and go after Chinese models on that basis, because we’re not sure what basis to go after them on. We need some reason to go after Chinese models. We’re just not sure what that is. I mean, look, five years ago, D.C. was thinking about China’s digital technology space the same way that we think about Russia’s digital technology space.
And it’s helpful to look back on where people’s brains were at five years ago. The idea was like if China developed a digital technology that people liked, it was a massive fluke. It was an accident. And if you wanted to kill it, all you had to do...
Andrew: Or it must have cheated.
Kendra: Or they must have cheated, right? And so, China can’t innovate, they only steal and cheat. And so, all you have to do is sort of kill the international expansion of that one product or ban that one product and that one thing dies. And in some case, we’re always kind of fighting the last war. Against most countries, that has worked just fine. I mean, in 2024, I don’t know if you remember this, there was this Russian cybersecurity software made by like Kaspersky, which Russian cybersecurity.
Russia’s good at cybersecurity. And so, they created this really good software. It got some traction in the U.S. Commerce came in and said, “Absolutely not. We’re banning Kaspersky products in the United States due to national security reasons.” And not a peep before or since from any major Russian producer of software that the United States is clamoring to have. And that has been true for most countries, for most of Internet history. And it is no longer true relative to China.
When TikTok came out, we said it. We said this is not going to be the last digital technology that China develops that people want to use. They have a very strong digital technology ecosystem now. And so, we don’t have anything else except ban it, sanction it, whack-a-mole it, you know, kill it. Let’s find some way to kill it. And I don’t think that strategy is going to be effective here. But more than that, I think the problem is that without a U.S. alternative, again, it’s just a supply and demand problem. It’s just a supply and demand problem. We said this on the last podcast. I do not care at all if it’s a Chinese model or not.
I do not care at all. When we are deciding what model to use, we want one that’s cheap and we want one that’s controllable and good enough. That’s it.
Andrew: And sorry to interrupt you, but maybe also worth touching quickly, like it’s not that easy to ban these models. Like, once they’re on the Internet, that’s the whole point of open source, right? Like, if we were going to try to ban it, what do you think would be the…?
Kendra: I mean, there’s been a lot of very controversial conversation on that in the last couple of days. There’s been some arguments that all the U.S. needs to do is sort of create enough risk, create enough perceived risk. Like they don’t even have to ban it. They just have to create enough perceived risks for corporations, right? For major corporations to touch the models that they will sort of back away from them and won’t use them at all. That they don’t necessarily have to sanction or they have to take the worst step, but just try to create some kind of churn.
I actually don’t know if that will work considering this, again, just considering the supply and demand problem. You have a massive consumer base that desperately wants this technology and, you know, not a ton of alternatives. So, yeah, I mean, I think that’s really the critical issue.
Andrew: All right. We meant to do a short one, but there’s just so much to unpack here that I think… I mean, I’m glad we went a little bit longer and I’m sure listeners will have enjoyed it. I’ve certainly learned a lot. But last piece, like, I don’t know, we talked about, you know, this pincer, this like short or like narrowing space for Chinese models. It seems, at least in the U.S., used by U.S. users, if not more globally. Any thoughts on how to think about that or what that might portend to going forward to wrap this up?
Kendra: I mean, the last thing I’ll say just as a summary is I think that what is going to define the space of AI development going forward is the gap between Chinese regulations and how they choose to control their own model releases and then how the United States decides to approach disincentivizing U.S. firms from using Chinese models. But I also just I think I would be remiss if I didn’t say this. One lesson I’ve learned during my career trying to analyze technology and the direction technology is going to go is that very often the shape of our conversation turns out to be wrong.
That the actual issues that we’re dealing with turn out to be like something comes out of left field that we didn’t consider that we go, “Oh, that changes the entire shape of the conversation.” I’ll give you one example before we sign off. There was a massive panic over the 5G race. Right? But for the average consumer, the move from 4G to 5G didn’t end up being that impactful at all. And actually, where a lot of the impact was, was on the sort of enterprise side of 5G. But we talked a lot about what it was going to mean for China to have access to 5G technology before the United States and what it would mean for their ecosystem and what it would mean for ICDs.
And we worried that China had more base stations than we did and their rural population had more access to 5G than we did. I was giving a talk at a classroom a couple of weeks ago, and it was a college classroom, and I asked people to put their hands up if they knew what the 5G race was. You know how many people put their hands up? Zero. Zero people. Zero people under the age of 22 had ever even heard of that. And it was the only thing anybody was talking about for a long time. Other considerations entered that. We defined that conversation as a race between the U.S. and China on a couple of really narrow points.
None of those things ended up being that big of a deal. And then other considerations actually did end up being a really big deal that nobody was talking about. I suspect we will find something similar here. For example, I think there’s a possibility that the shape of this argument changes in a year. Here’s one hypothesis. The bleeding edge of artificial intelligence is going to keep moving forward for the foreseeable future. But most consumers probably aren’t going to need a bleeding-edge model starting quite soon.
What do we do with AI at our company? We do document processing en masse. That doesn’t require a very smart model. It requires a pretty good model. So, imagine a world a year and a half from now, I’m not saying this is definitely going to happen, but imagine a world a year and a half from now in which it doesn’t actually matter if a Chinese regulator takes six months to release an open-weight model because people don’t switch models anymore every month.
The frontier is moving forward, but most people just pick a model and stick with it for a couple of years because it does what they need it to do and that’s just fine. And so, none of the concerns about who’s approving faster or regulatory capture, you know, or where the frontier is, that becomes a cybersecurity and critical infrastructure problem and stays in that space. And the actual competitive consumer landscape for how companies are using AI is about cost, is about whatever stack you picked up two years ago, becomes a little bit more, kind of becomes boring, essentially.
Andrew: Totally. It does make sense.
Kendra: This whole conversation could change in two years, right? And it could just be about something completely different. So I’m just aware of that as we move ahead.
Andrew: I’m like so tempted to press on that point, because I think it’s a great point. And I think we could expound on that. But we’ll save it for another pod. I’m excited to get you on more regularly for some of these quick reaction discussions. The listeners need more Kendra Schaefer in their life. That’s like my constant feedback that I’m getting.
Kendra: I mean, if only I could shut up faster, we could get more quick reactions.
Andrew: This has been amazing. Amazing, amazing. So, thank you for the knowledge bombs. And yeah, we’ll look forward to seeing you again soon. Hope you feel better.
Kendra: All right. Thanks. Talk to you later.
Andrew: Thanks. Bye, everybody.
