Stephanomics

China leverages AI diplomacy and open-weight models to win Global South support amid US export controls

26 min Listen to the original Episode page
Artificial IntelligenceGeopoliticsChinaUS-China RelationsTechnology PolicyDiplomacyOpen Source AINational Security

Summary

Stephanie Flanders and Daniel Remler discuss how China is winning a diplomatic AI race by positioning itself as a champion for the Global South against US technology restrictions. Remler argues that since late 2023, China has promoted open-weight models that are cheaper and offer data sovereignty, contrasting with US focus on safety and export controls. The episode highlights concerns about ideological bias in Chinese models, citing studies that link sensitive political discussions to degraded code quality. Remler suggests the US needs a stronger diplomatic offensive to counter China's narrative and promote US open-weight alternatives.

Key takeaways

  1. 1
    Ideological bias in Chinese AI models intersects with functional utility; studies show code generation becomes more malicious or broken after discussing sensitive topics like Tiananmen Square.

    Unidentified speaker (speaker_3) · 0:58

  2. 2
    Jinping in July was at the launch of China's World AI Cooperation Organization, which is this new multilateral forum that has 37 members, essentially all local. low and middle income countries, to talk about AI capacity building, open models, et cetera, et cetera. No focus on risk or safety there.

    Unidentified speaker (speaker_3) · 6:16

Episode sections

  1. 0:00 – 0:38

    Sponsor Read: Bloomberg News Now

    Speaker 0 and Speaker 1 list five reasons to subscribe to Bloomberg News Now, including instant global news, time savings, 24/7 availability, convenience on smartphones, and easy access via Apple Podcasts or Spotify.

  2. 0:38 – 1:15

    DeepSeek Censorship and Code Quality

    Speaker 3 argues that DeepSeek models, even open-weight versions, contain embedded censorship and alignment to CCP talking points. A study cited suggests that asking the model about Tiananmen Square before requesting code results in more malicious or broken code output.

  3. 1:15 – 2:15

    Introduction and Context: Jensen Huang Scandal

    Host Stephanie Flanders introduces the episode by referencing recent concerns about China undermining US tech leadership, specifically mentioning the Jensen Huang/NVIDIA scandal involving potential smuggling of chips to China and implications for the Trump administration.

  4. 2:15 – 6:52

    China's AI Diplomacy Strategy

    Daniel Remler argues China is winning a different AI race through diplomacy, positioning itself as a champion for the Global South against US restrictions. He traces this strategy to late 2023, following ChatGPT's launch and US chip export controls.

  5. 6:52 – 9:09

    China's Technological and Political Pitch

    Remler explains China's appeal: models are 8-12 months behind US frontier but cheaper, allow data sovereignty, and are bundled with hardware by companies like Huawei and Alibaba. Politically, China offers a voice at the UN and low demands on partners compared to the US.

  6. 9:09 – 9:37

    Sponsor Read: iHeartRadio Holidays

    Speaker 3 promotes iHeartRadio's holiday music programming, describing it as non-stop festive favorites available 24/7 via smart speakers or the free app.

  7. 9:37 – 14:08

    Ideological Bias and Security Risks

    Remler details how Chinese AI models embed CCP ideology, affecting functional outputs like code quality (citing CrowdStrike and Booz Allen Hamilton studies). He also notes a recent finding by Canadian firm Cohere of a potential backdoor in an Alibaba model.

  8. 14:08 – 26:14

    US Response and Diplomatic Gaps

    Remler suggests the US needs a stronger diplomatic offensive and defensive strategy, including rigorous evaluation of Chinese models' risks and promoting US open-weight models like Meta's Muse Spark to counter China's narrative.

Claims worth verifying

Even open weight DeepSeek models, when locally hosted, have a deep level of censorship and alignment to CCP talking points.

Unidentified speaker (UNKNOWN) ·0:38

How to verify: Review technical analyses or security reports on DeepSeek's open-weight model weights and alignment layers; check for academic or industry research on model alignment post-training. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

A study found that asking a DeepSeek model about Tiananmen Square and then asking it to produce code results in code more likely to be malicious or broken.

Unidentified speaker (speaker_3) ·0:58

How to verify: Locate the specific study cited by the speaker (referenced later as CrowdStrike and Booz Allen Hamilton studies) and review its methodology and findings on code quality after ideological prompts. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

The Jensen Huang/NVIDIA scandal is potentially the greatest national security scandal of recent times, implicating senior Trump administration members.

Unidentified speaker (UNKNOWN) ·1:30

How to verify: Search news archives and official statements from late 2026 regarding the NVIDIA/Jensen Huang scandal, looking for mentions of Trump administration officials and chip smuggling allegations. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Primary source

China released a global AI governance initiative in October 2023.

Unidentified speaker (speaker_3) ·4:07

How to verify: Search for the Chinese government's 'Global AI Governance Initiative' document released in October 2026 via official Chinese government channels or international news archives. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

The World AI Cooperation Organization was launched in July 2026 with 37 members.

Unidentified speaker (speaker_3) ·6:16

How to verify: Check press releases from the Chinese government or the organization itself for the launch date and member count in July 2026. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Chinese open-weight AI models are eight to ten, maybe twelve months behind US frontier models.

Unidentified speaker (speaker_3) ·7:12

How to verify: Compare benchmark performance metrics (e.g., LMSYS Chatbot Arena, major academic benchmarks) of leading Chinese open-weight models against leading US frontier models as of the interview date. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

CrowdStrike found that ideological bias in DeepSeek models intersects with functional uses, producing worse code.

Unidentified speaker (speaker_3) ·11:23

How to verify: Locate the CrowdStrike research report cited by the speaker regarding DeepSeek model behavior and code quality. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Booz Allen Hamilton conducted a study where interacting with a DeepSeek model as a US government employee resulted in more broken or malicious code.

Unidentified speaker (UNKNOWN) ·11:42

How to verify: Locate the Booz Allen Hamilton study referenced by the speaker and review its findings on model interaction and code output. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Cohere identified a potential backdoor in an Alibaba model.

Unidentified speaker (speaker_3) ·13:26

How to verify: Search for public statements or technical reports from Cohere regarding the discovery of a backdoor in an Alibaba AI model. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Moonshot's Kimi K3 model license requires a special agreement for enterprises above a certain revenue threshold.

Unidentified speaker (speaker_3) ·24:15

How to verify: Review the licensing terms for the Kimi K3 model on Moonshot's official website or developer portal. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Do your own research

  • Verify the CrowdStrike and Booz Allen Hamilton studies cited regarding code quality degradation in Chinese AI models after sensitive political prompts.
  • Check the membership list and charter of the World AI Cooperation Organization launched in July 2026 to assess its influence and focus areas.
  • Review US Department of Commerce export control lists and any recent enforcement actions related to NVIDIA chip smuggling to assess the Jensen Huang scandal's scope.
  • Compare the performance benchmarks of Chinese open-weight models (e.g., DeepSeek) against US frontier models to validate the '8-10 months behind' claim.
  • Investigate Meta's 'Muse Spark' or other US open-weight model offerings to evaluate their viability as diplomatic counter-tools.
  • Search for Cohere's findings on potential backdoors in Alibaba models to assess security risks in Chinese AI infrastructure.
  • Analyze UN voting records and diplomatic statements from Global South nations to measure the effectiveness of China's AI diplomacy strategy.
  • Review Trump administration policy statements on AI export controls and national security to understand the regulatory environment for chip smuggling allegations.
Full transcriptexpand
0:00
Unidentified speaker (speaker_0): Here are five reasons to subscribe to Bloomberg News Now.
0:04
Unidentified speaker (speaker_1): Number five, you get the latest news from around the world instantly.
0:08
Unidentified speaker (speaker_0): Four, it saves you time. Your day is busy. Catching up takes just a few minutes.
0:12
Unidentified speaker (speaker_1): Three, it's available 24 hours a day whenever you need it.
0:15
Unidentified speaker (speaker_0): Two, it's convenient. Listen on your smartphone or smart speaker.
0:19
Unidentified speaker (speaker_1): And number one, it's easy to find. Subscribe on Apple Podcasts, Spotify, or anywhere you listen.
0:25
Unidentified speaker (speaker_0): Five reasons, one simple way. to stay on top of the news. Bloomberg News Now.
0:33
Unidentified speaker (speaker_1): Bloomberg Audio Studios,
0:35
Unidentified speaker (speaker_2): Podcasts, Radio News.
0:38
Unidentified speaker (UNKNOWN): If
0:38
Unidentified speaker (speaker_3): you're accessing the DeepSeek model, there's a kind of censorship embedded at the software layer. Even the open weight models, when locally hosted, have this very deep level of censorship and alignment to CCP topic. talking points. And this trend has been increasing over time. And what I think is especially worrying is some other research that's found
0:58
Unidentified speaker (speaker_3): how ideological bias intersects with, let's say, more functional uses of the models. So there's this study that found that when you asked a DeepSeek model about Tiananmen Square and then asked it to produce code, the code was is more likely to be malicious or broken.
1:15
Unidentified speaker (UNKNOWN): I'm
1:15
Unidentified speaker (speaker_0): Stephanie Flanders, head of Bloomberg Research, and this is Stefanomics, your weekly guide to the forces transforming
1:21
Unidentified speaker (speaker_1): the
1:22
Unidentified speaker (UNKNOWN): global economy. We
1:23
Unidentified speaker (speaker_1): have all been reeling the last few weeks at the thought that China might have been undermining America's technological leadership from
1:30
Unidentified speaker (UNKNOWN): the
1:30
Unidentified speaker (speaker_1): inside,
1:30
Unidentified speaker (UNKNOWN): via
1:31
Unidentified speaker (speaker_1): Jensen
1:31
Unidentified speaker (UNKNOWN): Wang, and
1:32
Unidentified speaker (speaker_1): by extension the NVIDIA chief's relationship
1:34
Unidentified speaker (UNKNOWN): with the
1:35
Unidentified speaker (speaker_1): US President.
1:36
Unidentified speaker (UNKNOWN): But what
1:36
Unidentified speaker (speaker_1): if China's most
1:37
Unidentified speaker (UNKNOWN): potent
1:38
Unidentified speaker (speaker_1): weapon in
1:38
Unidentified speaker (UNKNOWN): this race for AI supremacy was
1:41
Unidentified speaker (speaker_1): hiding in
1:42
Unidentified speaker (UNKNOWN): plain sight? The NVIDIA fair is rightly
1:45
Unidentified speaker (speaker_1): a major
1:45
Unidentified speaker (UNKNOWN): scandal,
1:46
Unidentified speaker (speaker_1): maybe
1:46
Unidentified speaker (UNKNOWN): the greatest national security scandal of recent times, potentially implicating not only Jensen
1:52
Unidentified speaker (speaker_3): Wang
1:52
Unidentified speaker (UNKNOWN): himself but senior members of the Trump
1:54
Unidentified speaker (speaker_0): administration,
1:55
Unidentified speaker (UNKNOWN): who may have turned a blind eye to the smuggling of America's most sophisticated chips to China and worked work to make US policy in this area. more China friendly. It'd be difficult not to think about all of that this
2:07
Unidentified speaker (speaker_0): week
2:07
Unidentified speaker (UNKNOWN): as we see Presidents Trump and Xi meet
2:11
Unidentified speaker (speaker_0): again,
2:11
Unidentified speaker (UNKNOWN): but there's another front to the AI battle, where China seems to be
2:15
Unidentified speaker (speaker_3): winning, and there's no secret about how it's doing it. The only mystery is why people aren't paying more attention. Daniel Remler is a senior fellow at the Centre for a New American Security. who focuses on the implications of AI and emerging technologies for US national security. I wanted to have him on the show after
2:35
Unidentified speaker (speaker_3): reading a provocative piece he wrote for Foreign Affairs recently on China's AI diplomacy, where he argues in effect that the
2:43
Unidentified speaker (speaker_0): US focus on AI security and technological controls
2:48
Unidentified speaker (UNKNOWN): has
2:49
Unidentified speaker (speaker_0): allowed China to gain the competitiveness. competitive edge when it comes to actually getting Chinese AI models embedded in companies and countries around the world. And that in turn has a lot of potentially very disturbing implications. Daniel, thanks so much for coming on Stefanomics.
3:04
Unidentified speaker (speaker_3): It's great to be with you, Stephanie.
3:08
Unidentified speaker (UNKNOWN): Now
3:08
Unidentified speaker (speaker_0): I should say Today, my colleague, the senior
3:10
Unidentified speaker (UNKNOWN): geoeconomics analyst in this area, Michael Deng, who's been on this show a few times, was at pay pains for me to understand quite how respected you were in this field, that you weren't just some policy won who'd written a piece for foreign affairs, and he wanted me particularly to understand that you were the first to identify distillation, which we've talked about a bit in the past, as a
3:30
Unidentified speaker (UNKNOWN): key issue in the US-China AI gap. So I'm even more grateful that you've been able to come on. Explain
3:37
Unidentified speaker (speaker_0): briefly what you think China has got right in its public AI diplomacy than perhaps America's been getting wrong.
3:46
Unidentified speaker (speaker_3): Right. So
3:47
Unidentified speaker (UNKNOWN): I
3:47
Unidentified speaker (speaker_3): think it's important to trace China's AI diplomacy from 2023. And it comes just after the launch of ChatGPT at the end of 2022 and the imposition of pretty severe export controls by the United States on exports of chips going to China. And that moment really crystallized for Beijing both that the
4:07
Unidentified speaker (speaker_3): United States had an early lead, if not a mastery, in this new paradigm of artificial intelligence. And then the chip export controls showed that the United States would take very serious action to maintain this lead over China. That caused a rethink of China's national strategy, including this diplomatic component that I also. outline
4:27
Unidentified speaker (speaker_3): in the piece. So it begins with the end of October 2023. They released this global AI governance initiative. It's one of these sort of turgid party speak documents, but I think it's important because it lays down a few markers that would become kind of the hallmarks of China's AI diplomacy over the next several years through to
4:47
Unidentified speaker (speaker_3): today. And you can really pinpoint these two goals of their. strategy. First is positioning China as the champion of the global south of developing countries. And second, overcoming US technology restrictions. Since then we've seen Beijing make several moves in this direction in its diplomacy. So it
5:06
Unidentified speaker (speaker_3): pushed for these diplomatic initiatives at the UN General Assembly that would promote capacity building for developing countries. It would tout its open weight AI models. And then it has engaged in this AI dialogue with the United States, both during the Biden administration, when I was there, and most recently under the Trump administration, where it has similarly
5:26
Unidentified speaker (speaker_3): pushed this message of opposing US export controls and positioning itself as a champion of the global south, even though the kind of ostensible purpose of the dialogue is risk and safety.
5:37
Unidentified speaker (UNKNOWN): So
5:37
Unidentified speaker (speaker_3): what I argue in the beast is that. this kind of very focused approach on these two goals has allowed China to consistently make the argument that the United States is excessively focused on risk or focused on risk in a way that is merely a cover or a smoke screen for inhibiting China's AI development and by extension, the development of AI in the
5:57
Unidentified speaker (speaker_3): rest of the developing world, which China sees itself as a member of. And China, by contrast, has a more inclusive approach to AI. It builds these open way AI models that can be picked up and used by countries and governments and enterprises around the world. And it's building kind of parallel institutions that favor that vision. So President Xi
6:16
Unidentified speaker (speaker_3): Jinping in July was at the launch of China's World AI Cooperation Organization, which is this new multilateral forum that has 37 members, essentially all local. low and middle income countries, to talk about AI capacity building, open models, et cetera, et cetera. No focus on risk or safety there. So again, that kind
6:36
Unidentified speaker (speaker_3): of pitch from China.
6:38
Unidentified speaker (speaker_0): What's the attraction of that pitch for developing countries and why has it been enabling China to get quite a foothold in these countries?
6:51
Unidentified speaker (UNKNOWN): from
6:51
Unidentified speaker (speaker_0): an AI standpoint?
6:52
Unidentified speaker (speaker_3): There's a few factors. So the first and and most uh essential really is the basic technological offer from China. So these open weight models that I mentioned from China are eight to ten, maybe twelve months behind the US frontier. They're still quite good models from a capabilities perspective, but users don't need to pay through the
7:12
Unidentified speaker (speaker_3): nose and API costs. Um they can fine-tune them for their own local purposes. And so uh there's a greater degree of data privacy sovereignty in using these open weight models for various use cases.
7:25
Unidentified speaker (UNKNOWN): Uh,
7:25
Unidentified speaker (speaker_3): and that's of course especially important for price-sensitive users, not just in low and middle countries, but even you know, enterprises in the West.
7:34
Unidentified speaker (UNKNOWN): So
7:34
Unidentified speaker (speaker_3): that basic technological offer remains pretty strong. Chinese companies have paired these. models with hardware, which of course they're very proficient at. So you've seen Huawei Alibaba go around the world pitching kind of packages of models with hardware solutions for low and middle income countries. So that technological offer is very real.
7:54
Unidentified speaker (speaker_3): Second, at the kind of political level, there is a sense from a lot of low and middle income countries that they don't have a seat at the table in terms of global. AI diplomacy and that decisions are being made over their head, not just by the United States unilaterally, but potentially by the United States and its allies and uh advanced economies.
8:14
Unidentified speaker (speaker_3): And so China is sort of again positioning itself as their champion says, we will give you a voice, right? We will give you the technology, we will kind of lift your voice up at the UN in our own organization. So there's that political element. And then Beijing just asks little of its own partners. And so there's a sense that when Beijing comes to kind
8:34
Unidentified speaker (speaker_3): of bring a technology offering or to ask you to participate in their organization or support their political initiative, there's less of an expectation that that would come with some other kind of concession. And so it's a relatively low ask for countries that may not be part of so many multilateral organizations or not have the kind of bevy of alliances and partnerships that the United
8:54
Unidentified speaker (speaker_3): States and its partners have to support Beijing.
9:02
Unidentified speaker (UNKNOWN): The
9:02
Unidentified speaker (speaker_0): lights are twinkling,
9:03
Unidentified speaker (speaker_3): the
9:04
Unidentified speaker (speaker_0): playlist
9:05
Unidentified speaker (speaker_3): sparkling.
9:09
Unidentified speaker (speaker_3): Is here.
9:11
Unidentified speaker (speaker_3): Canada's home for Christmas music with non-stop festive favorites. 24 7.
9:17
Unidentified speaker (speaker_3): Turn it up and unwrap the holidays everywhere you go.
9:25
Unidentified speaker (speaker_3): Just ask your smart speaker to play iHeart the Holidays or stream it anytime, anywhere
9:31
Unidentified speaker (speaker_0): on the free iHeartRadio app.
9:37
Unidentified speaker (UNKNOWN): I said at the start this was
9:39
Unidentified speaker (speaker_3): all a sort of AI diplomacy in plain sight. I guess that the somewhat illicit or the non-transparent aspect to this is the messaging embedded in the AI models themselves. By law, Chinese AA models can't go against
9:55
Unidentified speaker (UNKNOWN): the
9:55
Unidentified speaker (speaker_3): Chinese government. You can sit there on DeepSeek Sea and can
9:58
Unidentified speaker (speaker_0): ask them about
9:59
Unidentified speaker (UNKNOWN): the
9:59
Unidentified speaker (speaker_0): Chinese Communist Party or Chinnen. and you will get very regulated answers that are consistent with the importance of the CCCP and core socialist values. If one's thinking about the geopolitical advantage that China could get from this diplomacy, there's a real
10:19
Unidentified speaker (speaker_0): sense in which its messaging is being built into the AI that these countries are using.
10:25
Unidentified speaker (speaker_3): Absolutely. And there's a slight misconception here because the argument's made that, well, open weight models, because they can be locally hosted and fine-tuned, right? Any user or enterprise that is worried about CCP censorship or ideological bias can fine-tune that out, right? Can impose guardrails that would effectively unbrainwash
10:45
Unidentified speaker (speaker_3): the model, let's say. And it's true that, of course, if you're accessing the DeepSeek model via the app, right. which is hosted by DeepSeek itself, of course, there's a kind of censorship embedded at the software layer. But analysis, including by the US government, has found that even the open weight models, when locally hosted, have this very deep level of censorship and kind of
11:05
Unidentified speaker (speaker_3): alignment to CCP talking points. And it's not just uh the topics that that frankly everyone cites correctly, like Tiananmen Square or the status of Taiwan. It extends to these contentious historical topics that matter to the CCP. And this trend has been increasing over time.
11:22
Unidentified speaker (UNKNOWN): And
11:23
Unidentified speaker (speaker_3): what I think is especially worrying is some other research that's found how ideological bias intersects with, let's say, more more functional uses of the models. So there's this study from CrowdStrike that found that when you asked a DeepSeek model about Tiananmen Square and then asked it to produce code. the code was more likely to be malicious or broken.
11:42
Unidentified speaker (UNKNOWN): So
11:42
Unidentified speaker (speaker_0): if you're sort of a suspicious character, they then you get a sort of worse product to put into your system.
11:48
Unidentified speaker (speaker_3): Yeah. This Booz Allen Hamilton study um made that point even more sharply because they interacted with the model kind of as if they were a US government employee uh and then asked it to produce code, and the code was m more broken, more more malicious. So there's the these interactions between kind of censorship, ideological bias, and the actual use of the model
12:08
Unidentified speaker (speaker_3): that are are concerning based on what we've we've read. And this is really an area of research that's still nascent and growing, but there needs to be more done.
12:16
Unidentified speaker (UNKNOWN): I
12:16
Unidentified speaker (speaker_0): mean it is extraordinary. I mean to your point about people not paying more attention, it seems like a long time ago, but you know, the focus on TikTok. But if these kind of chatbots and LLMs are going to be embedded in companies, maybe American companies, who think that they've taken out the communist-friendly pieces of it. And in fact, there's this kind of insidious way that it's
12:36
Unidentified speaker (speaker_0): actually potentially delivering a worse product dependent on what stance they have towards China. That's extraordinary. I mean, that just feels like we should be talking about that a lot more.
12:46
Unidentified speaker (speaker_3): Yeah, absolutely. And i the problem extends, frankly, farther than just censorship and ideological bias. So, of course, in China, there's a very strong role for the security services to intervene in the development of products and kind of identifying bugs. They have a sort of right of first communication on any bugs
13:06
Unidentified speaker (speaker_3): that IT or software firms might identify. And so you could imagine that the security services would ask one of China's AI developers to build backdoors or other kinds of vulnerabilities in a model that could then be exploited later by the security services. And Cohere, the Canadian AI company, actually identified something that
13:26
Unidentified speaker (speaker_3): looked like a backdoor in an Alibaba model very recently. And this is something that had only been kind of speculated in academic research work, but this is the first sighting in the wild of of something like this from a Chinese AI model. And so as you have these open-way Chinese AI models potentially taken up by enterprises around the world, we're just creating more and more vectors
13:46
Unidentified speaker (speaker_3): of vulnerability. Trevor
13:47
Unidentified speaker (UNKNOWN): Burrus, Jr.
13:48
Unidentified speaker (speaker_0): Okay. Well, while we sort of try and absorb rather distressing news, obviously a natural kind of question around this is how America could respond if it understood the arguments and the evidence that you've been making more clearly. or more keenly. One point that my my colleague Michael said to me was
14:08
Unidentified speaker (speaker_0): that maybe we overestimate the role that openness plays in increasing the popularity of Chinese AI models, because he thinks a lot of users will not care that much about the open weight aspect, but they do care a lot about how easy it is to use and how cheap it is and how reliable it is. And if the US was
14:28
Unidentified speaker (speaker_0): just better on that, you wouldn't necessarily have to give up that basic distinction between open and closed.
14:34
Unidentified speaker (speaker_3): Yeah, I think that's essentially right. The United States needs to enhance the diversity of its offerings, not just at the frontier, but kind of right behind the frontier. Um And package this, I I believe, with a stronger diplomatic strategy, because it's not as if the United States is lacking for companies that have incredible technical prowess that could
14:54
Unidentified speaker (speaker_3): release open weight models that are as good or or better than the Chinese. And of course, we've seen recently Meta release Muse Spark, which is quite a strong open way model. But essentially w what I think we're really missing here is a kind of diplomacy, offense, and defense on this question. So we need to go on offense to push out America's
15:14
Unidentified speaker (speaker_3): vision. for AI governance that both is about managing the risks, but also about enhancing opportunity around the world. And this is something that we are just not really in the game on right now, in contrast to Beijing standing up this world AI cooperation organization. So we just need to be much more
15:34
Unidentified speaker (speaker_3): present there and pair it with our own technology offerings as. as China does. I also think we need to put pressure on the Chinese in these areas of kind of vulnerability that we've we've just been talking about. So the US has a government organization that can evaluate AI models. It has evaluated Chinese AI models
15:54
Unidentified speaker (speaker_3): in the past and produce some of this research that I mentioned around censorship, ideological bias. And we could be doing that on a much more regular, rigorous basis. pushing out that information globally, working with our allies and partners like the United Kingdom, like Japan, that have this capacity as well. So that every time Beijing asserts the inclusiveness
16:14
Unidentified speaker (speaker_3): of its, of its models, you know, openness, low price, we are ready with a counter message that brings facts to bear on, well, actually, how good are these models? How expensive or inexpensive are they? But also what are the potential risks and and vulnerabilities and dangers? Because we're just missing an action on on that question. It's not the kind of message that US companies by themselves are,
16:34
Unidentified speaker (speaker_3): of course, going to push. And so I think if you have that kind of positive offense and kind of uh defensive approach at the diplomatic level, I think you could really shift the playing field uh in favor of the United States.
16:47
Unidentified speaker (speaker_0): I guess at some level, China's right. We may think that the US have had a rather kind of patchy approach to safety and security over the last couple of years. I don't get the impression that President Trump's going to be raising these kind of concerns about open weight models in his, you know, when he has dinner or whatever with with President Xi. But do you have a sense that
17:07
Unidentified speaker (speaker_0): any of this is percolating up to some of the senior levels who are interacting with the Chinese? We talk in terms of kind of hawks and doves towards China within the administration, but this is quite a technical area where, as you said, there's quite a lot of evidence.
17:23
Unidentified speaker (speaker_3): Certainly. So Secretary of the Treasury, Scott Bessant, who leads for for the US side, not just the economic relationship and the trade talks, but also So this AI dialogue on on risks with China, is quite concerned about risks from not just open weight models but increasingly capable AI models in general, and has been very
17:43
Unidentified speaker (speaker_3): vocal in his view that we need to manage these risks by talking to Beijing in some way.
17:57
Unidentified speaker (speaker_1): Against shared risk, prevent misunderstandings between our two systems. We also expect to exchange best practices for evaluating and testing advanced
18:07
Unidentified speaker (speaker_0): models.
18:08
Unidentified speaker (speaker_3): He's also been clear, especially in recent days, about the importance of export controls on AI chips. And so he's signaling very clearly ahead of the visit that while managing AI risks in dialogue with China is a priority, it is not a priority at the expense of these necessary. national security actions,
18:28
Unidentified speaker (speaker_3): nor at the expense of getting the broader economic and and trade relationship right. So I do expect him to bring it up. But of course, even though he says truthfully that the United States is not going to negotiate export controls, the Chinese will, of course, bring them up because it is just part of the the playbook that that we've been
18:48
Unidentified speaker (speaker_3): discussing that I mentioned in in my piece where they will position US X war controls as the obstacle to cooperation to development and so they will try to kind of create linkage across these issues even as we try to silo them and separate them.
19:06
Unidentified speaker (UNKNOWN): The
19:06
Unidentified speaker (speaker_3): lights are twinkling the playlist sparkling
19:12
Unidentified speaker (speaker_3): iHeart the holidays is here.
19:15
Unidentified speaker (speaker_3): Canada's home for Christmas music with nonstop festive favorites 24-7.
19:21
Unidentified speaker (UNKNOWN): Turn it
19:22
Unidentified speaker (speaker_3): up and unwrap the holidays everywhere you go.
19:30
Unidentified speaker (speaker_3): Just ask your smart speaker to play iHeart the Holidays or stream it anytime, anywhere on the free iHeartRadio app.
19:42
Unidentified speaker (speaker_0): Rightly or wrongly, it's still always considered to be this kind of race for supremacy, that there's the technological supremacy, but also the kind of ins assumption is that brings with it a kind of geopolitical influence. I said at the start we've tended to focus on the kind of chips piece, and the Jensen Wang scandal has raised lots of
20:02
Unidentified speaker (speaker_0): questions and concerns. around that, and there does seem to have potentially been a really significant effect on a US national security of that. That's one kind of power that the US currently has and is trying to hold on to that supremacy without making the rest of the world feel like its technology is completely inaccessible. But at
20:22
Unidentified speaker (speaker_0): the same time, you have this more kind of amorphous power that comes from being the LLM sitting inside companies, sitting in countries. How do you think about the sort of respective weight of those two things? If everybody's using Chinese LLMs, in what sense is China winning if
20:42
Unidentified speaker (speaker_0): the US is still quite far ahead technologically?
20:45
Unidentified speaker (speaker_3): Well, in my view, it's a big world. And we often, I think, think of these issues through the very narrow. lens of the United States and China or the United States and a few other actors around the world. But the direction of history, as President Xi would say, is moving towards more and more countries having greater economic weight,
21:05
Unidentified speaker (speaker_3): greater say in the world. And so it matters which of their leading firms, their government industries adopt which kinds of technology. And as we we've discussed, right, Chinese AI models have this ideological inflection. They have these risks because of the tight linkage between China's AI ecosystem and the Chinese government. And of course, in my view,
21:25
Unidentified speaker (speaker_3): the United States and our allies, the AI technologies that we produce have a scope for kind of multiplicity, liberalism. They carry values that we care about. And so we should think hard about whether we are not seeding ceding the kind of context of ideas and soft power in much of the world as we focus on this kind of narrow
21:45
Unidentified speaker (speaker_3): competition, even though it is quite important, right? We have to get the hard power elements right. We do have to deny leading technologies to China because those matter for its military, its intelligence, its surveillance state. But it would be a kind of empiric victory if we succeeded there, but effectively ceded much of the world to influence by. China that effectively
22:04
Unidentified speaker (speaker_3): congeals into political alignment.
22:06
Unidentified speaker (UNKNOWN): When
22:06
Unidentified speaker (speaker_0): I'm listening to you, it sounds like it really does come down to how easy it is to swap out of these Chinese LLM models. I mean, if the US ends up with the quote unquote best AI that everybody wants to adopt, then it doesn't really matter that they use Chinese models on the way to that better.
22:27
Unidentified speaker (speaker_0): US tech. The US still gets the advantage. But if it's sticky, if those Chinese models are actually quite hard to swap out or they get really embedded in systems, then that seems much more problematic, even if the US remains technologically ahead. I know members of the intelligence service that I've talked to, you know, that's
22:46
Unidentified speaker (speaker_0): precisely the dilemma they have. They want to have a diversity of tech, they don't want to just be using US hardware. On the other hand, they don't want to do anything that's going to be irreversible, then they're going to find themselves years from now trying desperately to unpick the Chinese software. So have we reached that point? Aaron
23:04
Unidentified speaker (UNKNOWN): Powell
23:04
Unidentified speaker (speaker_3): At this point, it seems like at a technical level, it's fairly easy to swap between models. Now, I say at the technical level because it's important to distinguish that organizations are different. And just because a high-flying startup in London or Silicon Valley can easily swap between models whenever it likes for whatever tasks
23:24
Unidentified speaker (speaker_3): it likes doesn't mean that a Fortune 500 company or even let's say like a conglomerate or a large enterprise in the developing world.
23:31
Unidentified speaker (speaker_0): Or a hospital that can afford to change its IT once every
23:34
Unidentified speaker (UNKNOWN): 30
23:34
Unidentified speaker (speaker_0): years.
23:35
Unidentified speaker (speaker_3): Exactly. A hospital, some uh large institution in a kind of more rural area or a less economically advanced part of a country. It's just going to be more, more challenging, right? As time goes on, these systems will become increasingly important to the functioning of enterprises. And that means that employees will be trained around using them, right?
23:55
Unidentified speaker (speaker_3): Workflows will be developed and kind of codified. And so you are adding these institutional elements that make swapping between. models, I believe, more difficult. That's even before we get to the question of how do the Chinese change their relationship to their own AI products? So we've seen a
24:15
Unidentified speaker (speaker_3): recent trend of these open weight model developers adding a few kind of conditions to their licenses. So for example, Kimi K3 from Moonshot, which is a Chinese A developer. developer, that model's license says that you need a special agreement with Moonshot in order to host the model if you are
24:35
Unidentified speaker (speaker_3): an enterprise above a certain kind of revenue threshold. And that special agreement effectively is a way for Moonshot to claw back some revenue from companies, in particular US hyperscalers Cyperscalers hosting that model. And so we may see as Chinese companies come under increasing financial. pressure, you may see them add these conditions onto the licenses, which
24:55
Unidentified speaker (speaker_3): also means that they're creating leverage against companies or countries that are using these AI models. And that itself may be a hindrance to models being swapped here and there.
25:06
Unidentified speaker (speaker_0): Daniel Remnler, this is a I mean it's an aspect of the AI IRAS race that we have not talked about nearly enough, and I think um should be thinking about very hard. even as we um read all the headlines about Jensen Wang and look at this ASEAN summit this week. Thanks very much.
25:21
Unidentified speaker (speaker_3): Absolutely thank you Stephanie
25:35
Unidentified speaker (UNKNOWN): Thanks
25:35
Unidentified speaker (speaker_0): for listening to Stefanomics from Bloomberg. It was hosted by me, Stephanie Flan. Flanders, and I was joined by Daniel Remler from the Center for New American Security. Stefanomics was produced by Summer Sadi and Moses Andam with editorial guidance from Megan O'Neill and Beth Ponsett. Amy Keene is the executive producer for Talk Podcasts at Bloomberg, and Cheryl
25:55
Unidentified speaker (speaker_0): Brumley is our head of podcasts. Podcast. Sound design was by Blake Maples and Aaron Kasper. To help others find Stefanomics.
26:03
Unidentified speaker (speaker_0): please rate and review us highly wherever you listen.
This page was produced automatically: the audio was transcribed with nvidia/parakeet-tdt-0.6b-v3+sortformer-v2.1 and interpreted by Qwen3.6-35B-A3B-UD-IQ3_S.gguf. Transcription and interpretation can contain errors; always confirm against the original episode. Nothing here is investment advice.

More from Stephanomics

Stephanomics

What Happens When AI Replaces the Economists?

Episode discussion: Jason Furman discusses AI replacing economists, policy impacts of AI data centers, and his EconVitals platform

Speakers’ claims: Stephanie Flanders interviews Jason Furman about the economic and political implications of artificial intelligence. Furman argues that voter backlash against AI data centers stems from concerns over utility costs and job displacement. He outlines policy recommendations including healthcare affordability and permitting reform to lower energy costs. Furman also discusses his EconVitals platform, an AI-built macroeconomic tool he claims replaces the work of 20 economists. He expresses concern that AI may reduce on-the-job training opportunities for junior economists.

Artificial IntelligenceEconomic PolicyUS Elections
34 min0 assets10 claims to verify
Stephanomics

Reintroducing Stephanomics

Episode discussion: Stephanomics returns to cover global economic shifts beyond US policy

Speakers’ claims: The host introduces the podcast's renewed focus on the global economy's seismic shift driven by inflation, AI, energy shocks, and political threats. The show previously tracked Trump's economic policies under the name 'Trumponics' but is now widening its lens to cover broader global economic forces. Stephanie Flanders, Bloomberg's head of government and economics, outlines the new format. The show will use a global newsroom and interview economists and policymakers to explore big questions reshaping the global economy.

Global EconomyMonetary PolicyGeopolitics
2 min0 assets5 claims to verify
All insights