US Loses Ground to China in Race to Set Global AI Tech Standards

AI, National Security & Geopolitical Competition with AI For Good Impact and Tiedrich Global — Ai4

Ai4

Technical measurement standards and Track II scientific diplomacy offer the only realistic path to global AI safety consensus as geopolitical rivalry splits US proprietary stacks from China's open-source push.

AI-assisted summary of AI, National Security & Geopolitical Competition with AI For Good Impact and Tiedrich Global — original source. Check the original for full context.

Source published: 2026-08-21

Detailed notes

## Background Published on YouTube by Ai4 on August 21, 2026, this panel features Daniel Dobos of the AI For Good Impact Initiative and Lee Tiedrich of Tiedrich Global Strategies. The session examines the intersection of national security, export controls, and AI governance, highlighting how political attention and funding are increasingly concentrated in the geopolitical policy arena.

## TL;DR Panelists Lee Tiedrich and Daniel Dobos debate the dual dynamics of competition and collaboration in global artificial intelligence governance. Tiedrich emphasizes the strategic race between the US and China, advocating for voluntary technical standards, measurement science, and Track II scientific diplomacy to bridge evaluation gaps. Dobos contrasts regional innovation models, arguing that open-source transparency and civil aviation-style safety protocols are crucial for global trust, highlighting immediate common ground in areas like child protection.

## Atmosphere & style The discussion carries a measured, intellectually rigorous tone, characterized by the diplomatic poise and deep policy expertise of both speakers. Moving past sensationalist 'arms race' headlines, the dynamic between Tiedrich's institutional legal perspective and Dobos's precision-focused scientific outlook creates a balanced, constructive dialogue. It feels like observing an earnest, high-level multilateral working session where realpolitik strategic concerns are balanced with practical, engineering-minded solutions.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=4s) Global AI governance operates under a dual dynamic of fierce geopolitical rivalry at the state level and active multilateral collaboration among technical standard-setting bodies. - [02:29](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=149s) The core bottleneck in AI policy is an 'AI evaluation gap'—the lack of scientific consensus and measurement tools to benchmark AI capabilities before and after deployment. - [02:29](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=149s) Voluntary technical standards provide regulatory flexibility, allowing deregulatory regimes like the US to adopt them as best practices while binding frameworks like the EU AI Act enforce them legally without fracturing technical baselines. - [02:29](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=149s) Drawing an analogy to commercial aviation safety, global AI governance must build shared trust and standardized safety benchmarks separate from national defense rivalries. - [05:46](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=346s) International standard-setting functions as a strategic geopolitical tool, where the US exports its proprietary AI stack while China builds coalitions around open-source architectures to capture emerging markets. - [09:07](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=547s) Child protection represents an immediate, high-stakes domain where shared societal pressures enable international consensus and technical guardrails across geopolitical divides. - [10:17](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=617s) Open-source and open-weight models trigger intense debate: critics point to downstream safety vulnerabilities and IP theft, while proponents argue open architectures are essential for independent academic auditing and cybersecurity verification. - [14:44](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=884s) Global innovation models diverge significantly: the US relies on capital-intensive brute-force scaling, China leverages massive parallel manpower and state backing, and Europe/Switzerland focuses on high-precision vertical niches. - [14:44](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=884s) The European Union's comprehensive AI Act provides developers with a predictable regulatory framework, contrasting with the fragmented landscape of over 2,000 state-level AI bills across the United States. - [17:42](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=1062s) Track II scientific diplomacy focused on empirical measurement and ground-zero truth is far more viable for international AI consensus than high-stakes bilateral political treaties.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=4s) Competition vs. Collaboration in AI Governance

The discussion opens with the moderator challenging the popular narrative that global artificial intelligence policy is purely an "AI arms race." While geopolitical rivalry dominates headlines, substantial behind-the-scenes work on international safety standards tells a more nuanced story. The panelists are asked to introduce themselves and immediately take a stance on whether AI governance today is defined primarily by competition or collaboration.

Lee Tiedrich, President and CEO of Tiedrich Global Strategies, draws on over thirty years of experience in law, engineering, and policy—including work at the National Institute of Standards and Technology (NIST), advisory roles at the Organisation for Economic Co-operation and Development (OECD), and contributing alongside Yoshua Bengio and Geoffrey Hinton to the 2025–2026 International Scientific Report on Advanced AI Safety. She firmly states that the world is in an era of competition: "If you read the Trump administration policy, we are in a race with China. That is definitely the tone that's been set," pointing to events like China's World AI Conference echoing that exact theme.

Dr. Daniel Dobos, a particle physicist who worked at CERN on real-time AI inference, served as Strategic Foresight Director at the United Nations, and represents Switzerland in international AI standardization, counters that both dynamics exist simultaneously. While acknowledging deep geopolitical competition, Dobos emphasizes that technical experts routinely engage in multilateral collaboration on foundational safety standards.

```mermaid flowchart TD accTitle: AI Governance Dynamics accDescr: Diagram showing the dual forces of competition and collaboration in global AI governance.

subgraph Rivalry["Geopolitical Competition"] US["US AI Action Plan & Stack Export"] CN["China Multilateral Alliances & Open Weights"] end

subgraph CommonGround["Multilateral Collaboration"] Standards["Technical Standards (ISO, ITU, NIST)"] Science["Scientific Benchmarks & Safety Reports"] NicheSafeguards["Shared Guardrails (e.g., Child Protection)"] end

Rivalry <--> CommonGround ```

---

[02:29](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=149s) Technical Standards as Innovation Enablers and the Aviation Analogy

Lee Tiedrich clarifies the fundamental distinction between voluntary technical standards and mandatory statutory regulations. Citing a recent piece published by the OECD and findings from the International AI Safety Report, Tiedrich argues that the core obstacle in modern AI policy is an "AI evaluation gap"—a lack of consensus on how to measure and evaluate AI capabilities, both in pre-deployment benchmarks and post-deployment real-world settings. Referencing calls by AI pioneer Fei-Fei Li for deeper measurement science, Tiedrich asserts that establishing technical standards gives countries regulatory flexibility: jurisdictions favoring deregulation (such as the US) can treat standards as voluntary best practices, while more prescriptive regimes (like the European Union) can make compliance legally binding without fracturing the underlying technical baseline.

Daniel Dobos responds by warning that the United States is currently losing relative global influence and trust in international standard-setting bodies. To illustrate the power of standard setting, Dobos draws an analogy to the history of commercial aviation:

> "A hundred years ago aviation was highly risky and by global collaboration, by setting global standards, we reached all together that flying commercial flying is the safest way of transportation that we have today globally... At the moment the US is losing a little bit of this power because it's losing trust, and I think that is really sad for the entire world."

Dobos stresses that just as global civil aviation isolated commercial safety protocols from defense and national security rivalries, the international community must cultivate shared trust around commercial AI safety benchmarks.

---

[05:46](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=346s) Standard Setting as Geopolitical Strategy and the Global AI Stack

The moderator turns to the strategic reality of standards being leveraged as trade mechanisms and market capture tools, recalling historical telecommunications disputes over GSM, CDMA, and 5G infrastructure with Huawei.

Tiedrich acknowledges that the United States is actively investing in domestic evaluation ecosystems—such as NIST’s Center for AI Standards and Innovation (CAISI) and executive directives coordinating frontier model testing across agencies like CISA and the NSA. However, she notes that US foreign policy explicitly focuses on exporting the complete "US AI stack," organizing Department of Commerce industry forums to encourage foreign enterprise adoption. In parallel, China utilized its World AI Conference to establish a multilateral coalition spanning over two dozen nations, focused on proliferating and democratizing open-source and open-access models. If emerging markets standardize around Chinese architectures, those technical norms become locked in, creating durable market barriers for Western technology providers.

- **Key Standard-Setting Bodies:** The International Telecommunication Union (ITU) and the International Organization for Standardization (ISO) remain essential venues where multi-stakeholder participation attempts to reconcile these competing state strategies. - **Strategic Objectives:** Standard setting functions simultaneously as technical harmonization and an instrument of market dominance across emerging regional markets.

---

[09:07](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=547s) Practical Areas for Common Ground: Child Protection and Guardrails

When asked whether Chinese participation inside international standard bodies represents authentic collaboration or thinly veiled competition, Dobos maintains that technical experts can work constructively across political divides when addressing specific societal harms.

Dobos highlights **child protection** as an immediate, non-negotiable candidate for international consensus. Governments worldwide are currently grappling with the psychological and developmental impacts of generative systems on younger generations. Because every state faces identical domestic pressures regarding child welfare, scientific and technical working groups can establish shared guardrails regardless of broader geopolitical friction. Successfully demonstrating harmonized standards in tangible, high-stakes domains like child protection can establish the diplomatic and methodological blueprints needed to tackle more complex frontier AI risks later.

---

[10:17](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=617s) Open-Source Models, Cybersecurity, and Intellectual Property

The conversation shifts to the explosive proliferation of open-source and open-weight models, particularly those originating from Chinese research labs. Tiedrich cautions that open-weight models introduce severe governance vulnerabilities because downstream users can easily strip out embedded safety guardrails and fine-tune models maliciously. Furthermore, economic pressures have driven Western enterprises to download and run Chinese open models on local servers due to dramatically lower inference costs compared to proprietary US frontier APIs. This trend has sparked intense debates regarding knowledge distillation and intellectual property theft, with US frontier labs arguing that their proprietary training outputs are being ingested to bootstrap rival architectures.

Dobos offers a sharply contrasting perspective from the standpoint of academic research and cybersecurity:

```mermaid flowchart LR accTitle: Open Source Inspection Pipeline accDescr: Flowchart illustrating how open-source models allow independent auditability.

subgraph OpenArchitecture["Open Model Lifecycle"] Weights["Open Weights / Datasets"] --> Audit["University & Security Audits"] Audit --> Fixes["Vulnerability & Bias Detection"] Fixes --> Trust["Certified Safe Infrastructure"] end ```

Dobos argues that critical digital infrastructure has historically achieved maximum security only through open source, citing foundational web servers and web browsers: - Open architectures eliminate short-term return-on-investment pressures and give independent academic researchers full visibility into training data, hidden algorithmic biases, and emergent vulnerabilities. - He cites Switzerland’s fully open-source foundational model, **Apertus 1.5**, as an example of transparent modeling that empowers global researchers to inspect systemic risks before deployment. - Mirroring aviation safety, independent inspectors must be able to audit and certify every component across the value chain.

Tiedrich concludes the segment by noting that history shows proprietary and open architectures ultimately co-exist and find distinct commercial roles, just as proprietary enterprise software and Linux co-exist today.

---

[14:44](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=884s) Europe and Switzerland’s Niche Innovation Philosophy

Addressing Europe's perceived absence in the US-China AI duopoly, Dobos challenges the assumption that technological leadership requires brute-force scaling. He outlines three fundamentally contrasting regional philosophies:

Dobos explains that Switzerland—ranked as a global innovation champion for twelve consecutive years—leverages targeted capital to dominate vertical niches, creating "hidden world champions" in precision mechanics, high-end timepieces, and specialized software. While the EU AI Act is frequently criticized for strictness, it offers a predictable legal structure compared to the United States, where developers face a confusing federal void overlaid by more than 2,000 state-level legislative proposals.

---

[17:42](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=1062s) Scientific Track Diplomacy vs. Political Division

The panel concludes on the viability of bilateral agreements, examining US efforts around initiatives like Pax Sapiens / Pax Silica against China’s coalition of roughly twenty-nine partner nations. The moderator asks whether direct bilateral dialogues—such as the AI safety talks broached during presidential summits between Joe Biden, Donald Trump, and Xi Jinping—can yield substantive governance.

Lee Tiedrich argues that diplomatic breakthroughs are feasible primarily when negotiations remain grounded in "Track II" scientific diplomacy rather than high-stakes political bargaining:

> "If we can sort of keep the discussion at the scientific level... like when I worked on the International Safety Report for 2025–2026, the report was clear: we were not making policy recommendations. This was just trying to come up with ground-zero truth on the science. And there we actually had collaborators from all different countries—China as well as Western countries."

Tiedrich concludes that bringing researchers and measurement scientists into the room to agree on empirical benchmarks and evaluation tools is far more attainable than seeking high-level political accords between rival state leaders. The panel wraps up with an acknowledgment of the distinct American and European viewpoints, with the moderator expressing hope to incorporate direct Chinese perspectives in future summits.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=4s) If you read the Trump administration policy, we are in a race with China. That is definitely the tone that's been set — Lee Tiedrich

> [00:04](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=4s) AI arms race

> [02:29](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=149s) AI evaluation gap — Lee Tiedrich

> [02:29](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=149s) A hundred years ago aviation was highly risky and by global collaboration, by setting global standards, we reached all together that flying commercial flying is the safest way of transportation that we have today globally... At the moment the US is losing a little bit of this power because it's losing trust, and I think that is really sad for the entire world. — Daniel Dobos

> [05:46](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=346s) US AI stack — Lee Tiedrich

> [14:44](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=884s) Throws money at the scaling problem — Daniel Dobos

> [14:44](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=884s) hidden world champions — Daniel Dobos

> [17:42](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=1062s) Track II — Lee Tiedrich

> [17:42](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=1062s) If we can sort of keep the discussion at the scientific level... like when I worked on the International Safety Report for 2025–2026, the report was clear: we were not making policy recommendations. This was just trying to come up with ground-zero truth on the science. And there we actually had collaborators from all different countries—China as well as Western countries. — Lee Tiedrich

## Mentioned Daniel Dobos, Lee Tiedrich, AI For Good Impact Initiative, Tiedrich Global Strategies, RegulatingAI, NIST (National Institute of Standards and Technology), NIST Center for AI Standards and Innovation (CAISI), OECD (Organisation for Economic Co-operation and Development), Yoshua Bengio, Geoffrey Hinton, International Scientific Report on Advanced AI Safety, CERN, United Nations, World AI Conference, Fei-Fei Li, CISA, NSA, US Department of Commerce, ITU (International Telecommunication Union), ISO (International Organization for Standardization), Huawei, Apertus 1.5, EU AI Act, Pax Sapiens, Pax Silica, Joe Biden, Donald Trump, Xi Jinping

## Video description Daniel Dobos - Co-Chair, AI For Good Impact Initiative Lee Tiedrich - President & CEO, Tiedrich Global Strategies

National security is the fastest-growing dimension of AI policy, with bipartisan urgency around export controls, autonomous weapons, and the intelligence community's AI adoption. This panel ensures the Summit addresses the full spectrum of AI governance — not just enterprise compliance — and elevates RegulatingAI's positioning in the national security policy space, where AI governance funding and political attention is increasingly concentrated.

[Source](https://www.youtube.com/watch?v=hgAcbSQiVSU)

中文简报

美国在全球AI技术标准竞争中落后于中国在地缘政治竞争使美国专有技术栈与中国开源推进相分立的背景下，技术度量标准与二轨科学外交成为达成全球AI安全共识唯一现实的路径。与会专家Lee Tiedrich与Daniel Dobos探讨了全球人工智能治理中竞争与合作的双重动态。Tiedrich强调中美之间的战略竞争，主张通过自愿性技术标准、度量科学以及二轨科学外交来弥合评估差距。Dobos则对比了各区域的创新模式，认为开源透明度与民航式的安全协议对于建立全球信任至关重要，并突出了儿童保护等领域当下即可达成的共识基础。全球 AI 治理呈现地缘政治竞争与技术标准多边合作并存的格局；专家指出当前政策的核心症结在于“AI 评估鸿沟”，即各方对部署前后的科学测量工具与评测基准缺乏共识。 (2:29) 国际标准制定已演变为地缘战略工具：美国致力于向外输出专有的“美国 AI 技术栈”，而中国则通过多国联盟推广开源开放架构，力争在新兴市场锁定底层技术标准。 (5:46) 开放权重模型引发激烈博弈，批评者担忧其安全护栏易被剥除并存在知识产权蒸馏争议，而倡导者则强调透明开放架构对于学术独立审计和网络安全验证不可或缺。 (10:17) 全球 AI 创新模式分化显著：美国依赖密集资本进行暴力缩放（brute-force scaling），中国凭借高强度人力与国家支持进行并行试错，欧洲与瑞士则依托合规框架深耕高精度垂直细分领域。 (14:44)

Reading summary

Global artificial intelligence governance is defined by a simultaneous clash of geopolitical rivalry to capture global markets and multilateral cooperation to establish technical safety baselines. Panelists Lee Tiedrich and Dr. Daniel Dobos argue that while high-level political accords face gridlock, international standard setting, objective scientific benchmarks, and shared non-negotiable concerns offer practical pathways for international coordination. Competing AI Stacks and Strategic Standard Setting (5:46) Although international bodies like the International Telecommunication Union (ITU) and the International Organization for Standardization (ISO) aim to harmonize technology, technical standard setting increasingly functions as an instrument for market dominance. Lee Tiedrich notes that US policy focuses explicitly on exporting the integrated "US AI stack" through Department of Commerce industry forums alongside domestic evaluation efforts led by NIST's Center for AI Standards and Innovation (CAISI), CISA, and the NSA. In parallel, China has engaged a coalition of over two dozen partner nations to distribute and normalize open-access and open-weight architectures. If emerging markets standardize around Chinese technical norms, this creates durable technological lock-in that threatens to exclude Western technology providers from regional ecosystems. The Evaluation Gap and the Aviation Analogy (2:29) Tiedrich identifies an "AI evaluation gap"—the absence of international consensus on measuring AI capabilities both pre-deployment and in real-world environments—as the core challenge in contemporary policy. Establishing voluntary technical standards provides regulatory adaptability: jurisdictions favoring deregulation, such as the US, can treat standards as best practices, while prescriptive regimes like the European Union can legally mandate compliance without fracturing global technical baselines. Dr. Daniel Dobos draws an analogy to commercial aviation, which became the safest global transit system by isolating commercial safety protocols from national defense conflicts. Dobos warns, however, that the US is currently losing trust and influence in international standard-setting bodies, which risks undermining collective confidence in global AI safety baselines. Open-Weight Models: IP Risks Versus Auditability (10:17) The rapid spread of open-weight models, particularly from Chinese labs, has triggered debate over systemic security and intellectual property. Tiedrich cautions that open-weight architectures introduce major governance risks because users can strip embedded safety guardrails and fine-tune models maliciously. Furthermore, Western enterprises facing economic pressures increasingly deploy lower-cost Chinese models locally, fueling concerns among US frontier labs that proprietary outputs are being ingested through distillation. Dobos presents an opposing perspective from computer science and cybersecurity, arguing that critical digital infrastructure historically achieves true security only through open source. Open access allows independent researchers to audit data, algorithmic biases, and emergent vulnerabilities without commercial return-on-investment pressures. Citing Switzerland's Apertus 1.5 foundational model, Dobos advocates for independently verifiable infrastructure, while Tiedrich anticipates that open and proprietary models will ultimately coexist much like Linux and proprietary enterprise software. Diverging Regional Approaches to Innovation and Regulation (14:44) Dobos contrasts three competing regional philosophies governing artificial intelligence. The United States pursues capital-intensive brute-force scaling, while China deploys manpower and parallel execution by launching hundreds of state-directed startups to establish a national champion. In contrast, Switzerland and Europe emphasize targeted vertical precision, with Switzerland deploying specialized capital to establish "hidden world champions" in precision mechanics and specialized software. These strategies align with their regulatory environments: developers in the US must navigate an unpredictable patchwork comprising more than 2,000 state-level legislative proposals, whereas the European Union's AI Act, despite its strict reputation, provides developers with long-term legal predictability under a single unified framework. Scientific Track Diplomacy and Pragmatic Guardrails (17:42) While direct high-level bilateral summits between superpowers struggle to produce binding political treaties, the panelists argue that practical collaboration remains viable in targeted areas. Dobos highlights child protection as an immediate candidate for international consensus, noting that universal domestic pressures around youth mental welfare allow technical working groups to agree on guardrails irrespective of geopolitical rivalries. Tiedrich asserts that "Track II" scientific diplomacy offers the most effective forum for durable alignment. Pointing to the 2025–2026 International Scientific Report on Advanced AI Safety, Tiedrich observes that researchers from China and Western nations successfully established shared empirical truths because the mandate focused strictly on technical measurement science rather than policy prescriptions. An unresolved challenge is whether consensus around scientific measurement and child welfare can survive deepening economic friction over open-weight models and proprietary stack exports. Furthermore, the discussion reflects Western legal, scientific, and European regulatory perspectives, leaving open the degree to which direct Chinese stakeholders and developing nations will align with these proposed safety frameworks.

精炼阅读摘要

在这场关于人工智能地缘政治与安全治理的研讨中，Tiedrich Global Strategies总裁Lee Tiedrich与瑞士国际标准代表Daniel Dobos指出，全球AI生态正处于激烈的大国地缘博弈与紧迫的多边技术协作双重张力之中；尽管中美在技术栈出口与市场主导权上竞争加剧，但通过以实证科学为基石的“二轨外交”建立统一的评估基准与具体安全护栏，依然是避免全球技术脱钩与治理失序的关键路径。地缘竞争与全球技术栈输出博弈 (5:46) 研讨开篇指出，全球AI政策不能被简单视为无序的“军备竞赛”，但大国博弈的主调已经确立。Tiedrich指出，美国政府的政策定调明确将与中国竞争置于核心，中国在世界人工智能大会（WAIC）等场合也释放了针锋相对的竞争信号。目前，美国正通过国家标准与技术研究院（NIST）的人工智能标准与创新中心（CAISI），以及协调网络安全和基础设施安全局（CISA）、国家安全局（NSA）等跨机构测试构建前沿评估生态，并由商务部组织行业论坛积极向海外推广完整的“美国AI技术栈”（US AI stack）。与此平行，中国在WAIC上联合二十余个国家建立多边联盟，致力于推广和普惠开源及开放权重模型。这种技术输出竞争的本质在于抢占未来技术规范并形成市场锁定。Tiedrich分析认为，一旦新兴市场国家围绕中国的技术规范和开源架构完成基础设施标准化，将构筑起持久的路径依赖与准入壁垒，令西方商业技术提供商面临市场阻隔。在此背景下，国际电信联盟（ITU）和国际标准化组织（ISO）等多边平台已演化为多方博弈的前线，标准制定正同时承担着全球技术协调与争夺区域市场支配权的双重地缘战略功能。评估鸿沟与民航标准协同的启示 (2:29) Tiedrich界定了自愿性技术标准与强制性法律法规之间的功能差异，指出当前全球AI治理的核心障碍在于“AI评估鸿沟”（AI evaluation gap），即国际社会在如何衡量与评估AI能力（包括部署前的基准测试与部署后的现实运行表现）上缺乏科学共识。呼应AI学者李飞飞对发展度量科学的倡议，Tiedrich主张技术标准能为各法域提供灵活性：推崇去监管的环境（如美国）可将标准作为自愿性最佳实践，而采取严格规制模式的地区（如欧盟）则可将其作为法定合规依据，同时确保全球底层技术基线不至于彻底碎裂。 Dobos则引入商业民用航空的发展历史作为对照。他回顾指出，一个世纪前的航空业曾伴随极高风险，但国际社会通过跨越地缘政治分歧的多边合作与全球安全标准的制定，最终使商业飞行成为当今全球最安全的交通方式。Dobos警告称，当前美国由于国际信任度下降，在国际标准组织中的相对话语权与信任资产正在受损。他呼吁国际社会应当汲取民航将民用安全协议与国防安全竞争脱钩的历史智慧，在商业AI安全评测领域重新凝聚跨国互信。开源模型的安全审计与知识产权争议 (10:17) 源自中国研究机构的开源与开放权重模型迅速普及，引发了关于系统安全性与知识产权的激烈交锋。Tiedrich提出风险预警，认为开放权重模型具有显著的治理脆弱性，下游恶意行为者可以轻易剥除预设的安全护栏并实施定制化微调。此外，由于此类模型的本地推理成本远低于美国闭源商业前沿API，许多西方企业出于经济考量转而在本地运行中国模型，这加剧了关于“知识蒸馏”与知识产权侵权的指控，部分美国前沿实验室指责对手使用其闭源模型的输出数据训练竞争性开源架构。针对开源风险，Dobos从学术研究与网络安全的角度给出了相反观点。他认为，历史经验表明网页服务器和浏览器等关键数字基础设施恰恰是依靠开源架构才实现了最高级别的安全性。开源消除了短期商业投资回报的功利压力，赋予全球独立研究人员审查训练数据集、算法偏见与未暴露漏洞的充分能见度。Dobos以瑞士完全开源的基础模型Apertus 1.5为例，说明透明架构允许全球科学界在模型部署前进行独立审计。Tiedrich对此总结指出，闭源专有模型与开源生态并非绝对对立，二者未来更可能像专有商业软件与Linux系统一样在市场中共存分工。区域创新范式与科学二轨外交突破 (14:44) 面对外界关于欧美在中美双雄竞争中落后的论调，Dobos反驳了唯有依赖“暴力堆叠资源”（brute-force scaling）才能保持领先的技术迷思，并系统梳理了三大区域的创新哲学差异：美国依赖雄厚资本驱动规模扩展，但深陷各州碎片化立法的泥潭（全美累计提出2193项州级AI法案，例如内华达州提出6项立法并颁布了3项）；中国依托庞大的人口红利与并行执行力，针对单一概念启动数百家初创企业快速试错；而瑞士与欧洲则坚持垂直利基市场的精准创新，通过高效部署专业资本在精密制造等领域培育“隐形冠军”，且欧盟《人工智能法案》（EU AI Act）为产业提供了清晰可预期的法律边界。在双边与多边外交路径上，讨论认为领导人高层峰会往往受困于政治博弈，而基于“二轨外交”（Track II）的实证科学合作具有更高的可行性。Tiedrich回顾其参与撰写2025–2026年《国际先进AI安全科学报告》的经历指出，当跨国讨论严格限制在纯粹的事实求证与度量科学范畴、不掺杂政策游说时，中美及西方科学家能够保持高效协作并达成底线共识。Dobos亦指出，各国社会在儿童保护等心理和发展风险上面临同等挑战，此类高利害、强共识的具体领域正是建立全球统一安全护栏与技术标准的首要突破口。该研讨表明，尽管诸如Pax Sapiens或Pax Silica等排他性倡议难以在短期内弥合大国之间的地缘安全互疑，但将治理聚焦于度量科学、开源独立审计以及儿童保护等切实现实危害，是维系全球技术互通的最优解。不过，本次讨论主要呈现了美欧专家的学术与政策洞察，缺乏来自中国政策制定者与产业机构的直接参与，其倡导的科学协作共识在新兴地缘摩擦中的实际约束力仍有待观察。

Detailed notes — zh

## Background 本内容由 Ai4 于 2026 年 8 月 21 日发布在 YouTube 平台上。视频记录了 Tiedrich Global Strategies 总裁兼首席执行官 Lee Tiedrich 与 AI For Good Impact Initiative 联合主席 Daniel Dobos 的专题讨论，聚焦于国家安全、出口管制、自主武器与情报界采纳等 AI 政策最前沿议题，旨在从超越单纯企业合规的视角探讨全球 AI 治理格局。 ## TL;DR 本场研讨会深入探讨了当前全球 AI 治理中“地缘竞争”与“多边合作”并存的复杂态势。专家们区分了标准制定与法律监管的差异，强调了通过建立底层“测量科学”来弥补 AI 评估差距的紧迫性。对话进一步剖析了开源与专有模型的安全博弈，以及中美欧三方截然不同的技术发展路径（资本密集、人力密集与垂直精准创新），并最终指出“科学外交”是超越政治分歧、促成全球技术底线共识的最具可行性的突破口。 ## Atmosphere & style 整场对话呈现出理性、严谨且极具战略高度的学术与政策研讨氛围。两位专家分别依托美国政策智囊与欧洲科研中立的视角展开交锋与呼应，既有对地缘对抗现实的清醒剖析，又始终怀揣着如民航安全合作般的全球主义理想。叙事节奏层层递进，从微观的模型评估与开源透明性一路推演至宏观的地缘经济秩序，带给听众一种在技术浪潮与大国博弈交织下探寻理性破局点的深刻洞察感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=4s) 当前全球 AI 格局在宏观上由美中地缘政治竞争主导，但在技术专家与安全标准的制定层面，跨国多边合作依然在积极发生。 - [02:29](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=149s) 全球 AI 治理目前面临显著的“AI 评估差距”，业界在模型部署前后准确测量与评估系统性能方面仍缺乏有效工具与科学共识。 - [02:29](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=149s) “技术标准”与“监管政策”需作本质区分：坚实的测量科学标准是客观底座，各国可据此灵活选择自愿合规框架或强制性法律落地。 - [02:29](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=149s) 民用航空业将商业飞行安全与国防军用安全解耦治理的历史经验表明，建立全球互信的技术安全标准是管控前沿新兴技术风险的关键。 - [05:46](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=346s) 技术标准制定具备强烈的地缘经济属性与市场锁定效应：美国正致力于向盟友输出“美国 AI 技术栈”，而中国则通过多边联盟力推开源模型体系。 - [08:51](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=531s) 儿童保护等具备高度普世社会共识的议题，能够超越大国地缘博弈，成为各国率先建立 AI 安全协作与治理信任的突破口。 - [08:51](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=531s) 开源权重模型降低了企业部署成本并加速了普及，但也引发了安全护栏易被恶意剥离以及模型蒸馏等知识产权维权争议。 - [12:04](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=724s) 开源与透明性是保障关键数字基础设施安全的重要机制，使全球科研人员能够像航空适航认证一样对 AI 数据集与偏差进行白盒审查。 - [14:39](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=879s) 美、中、欧三方呈现截然不同的 AI 路径：美国依靠资本强推模型规模化扩张，中国依赖人力与初创公司密集赛马，欧洲/瑞士则聚焦精准创新与利基市场。 - [17:42](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=1062s) 科学外交在 AI 治理中扮演核心角色：让跨国科学家聚焦于测量科学的客观事实基准，比在政治层面直接达成大国战略共识更具可行性。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=4s) 引言与嘉宾介绍：竞争与合作的时代定调主持人开场指出，外界普遍倾向于将当下的全球 AI 格局定义为一场“AI 军备竞赛”，但这仅仅揭示了全貌的一半；在幕后，围绕安全标准的制定与推进其实正进行着大量工作。本场研讨会邀请到两位分别代表美国与欧洲视角的重磅嘉宾，并抛出首个核心议题：当下的 AI 治理究竟更多是由“竞争”还是“合作”所定义的？ Tiedrich Global Strategies 总裁兼首席执行官 Lee Tiedrich 率先作了自我介绍。她拥有工程学与法学双重背景，曾在 AmLaw 50 顶尖律所担任合伙人，在深耕新兴技术领域逾 30 年的过程中，曾供职于美国国家标准与技术研究院（NIST），担任经合组织（OECD）顾问，并与图灵奖得主 Yoshua Bengio、Geoffrey Hinton 共同参与撰写了 2025/2026 年《国际 AI 安全报告》（International AI Safety Report）。针对主持人的提问，Lee 明确表示当前无疑处于一个“竞争时代”。从特朗普政府的政策导向来看，美国正与中国展开直接竞逐；而中国在近期举办的世界人工智能大会（WAIC）上也同样延续并呼应了这种竞争基调。随后，AI For Good Impact Initiative 联合主席 Daniel Dobos 博士介绍了自己的背景。他接受过粒子物理学训练，职业生涯的大部分时间都在欧洲核子研究中心（CERN）从事实时 AI 推理研究，曾任联合国战略前瞻总监，并代表瑞士参与全球 AI 标准化工作。Daniel 认为当前局势既有竞争，也有合作的窗口。尤其在技术专家跨国协作制定安全标准的层面，多边合作不仅可行，而且正在发生，但不可否认在宏观层面上竞争依然占据主导。 --- ### [02:29](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=149s) 标准制定与监管的区别：弥补“AI 评估差距” 针对“技术标准是否会阻碍创新并催生官僚主义”的常见担忧，Lee Tiedrich 澄清了“标准（Standards）”与“监管（Regulation）”的本质差异。她援引近期为 OECD 撰写的一篇文章及国际 AI 安全报告指出，全球目前面临着巨大的“AI 评估差距”（AI evaluation gap）——无论是在模型部署前还是进入现实环境后，业界在如何准确测量与评估 AI 系统的表现上仍缺乏有效工具与共识。 ```mermaid flowchart TD accTitle: AI标准与监管分层治理架构 accDescr: 展示从底层测量科学共识到技术标准，再到各国不同监管路线的分流逻辑。 subgraph S1["底层基础：测量科学"] A["多学科科研协作 （跨国科学家共识）"] --> B["填补 AI 评估差距 （建立评估与度量技术）"] end subgraph S2["共识层：全球统一标准"] B --> C["统一技术标准体系 （ISO / ITU / NIST 等）"] end subgraph S3["应用层：多元化政策落地"] C --> D["去监管/自愿框架 （如美国模型自愿测试）"] C --> E["强制性合规法案 （如欧盟 EU AI Act）"] end ``` Lee 强调，正如李飞飞（Fei-Fei Li）等学者所呼吁的，当前亟需建立坚实的“测量科学”（Measurement Science），让多学科科学家就测试与评估技术达成底层共识。有了统一的技术标准作为底座，各国便可在监管策略上享有更大自由度： - **去监管化倾向的司法管辖区（如美国）**：可以将标准作为非强制性的行业自律基准或自愿测试指南。 - **重监管倾向的地区（如欧盟）**：可以直接将这些客观技术标准转化为强制性合规法律。 Daniel Dobos 对此补充并提出警示，他认为美国在主导全球标准制定方面正在“相对地”失去话语权，其根源在于全球信任的削弱。他以百年来全球民用航空安全治理作为类比：100 年前航空业充满风险，但正是通过全球协作与统一安全标准的建立，商业飞行才成为当今全球最安全的交通方式。Daniel 强调，航空领域成功地将商用飞行安全与国防军用安全解耦治理；若在 AI 领域缺乏类似美国领导的全球互信机制，全世界在风险管理上都将承受巨大损失。 --- ### [05:46](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=346s) 标准制定作为地缘经济武器与技术栈输出主持人将话题引申至标准制定的地缘政治属性：历史上 GSM 与 CDMA 的较量以及涉及华为的通信标准之争表明，制定标准往往也是掌控市场和推行贸易战略的重要手段。 Lee Tiedrich 赞同这一观察，并详细阐述了美中两国的战略布局： - **美国的技术栈输出战略**：美国正全力构建 AI 评估生态体系，包括 NIST 旗下的 AI 标准与创新中心（CAISI），以及总统行政令中引入国家安全局（NSA）和网络安全与基础设施安全局（CISA）对前沿模型开展自愿测试。在政策层面，美国商务部正组织企业联盟，致力于向全球市场输出“美国 AI 技术栈（US AI Stack）”，以确保全球盟友与新兴市场深度绑定美国技术底座。 - **中国的多边开源联盟**：中国在世界人工智能大会上发起成立了涵盖二十多个国家的多边组织，重点推进开源与开放获取（Open Access）模型，试图通过“技术民主化”加速其技术体系在发展中国家的渗透。 Lee 指出，技术标准的锁定效应（Lock-in effect）极其显著，一旦某种标准率先在新兴市场扎根，其他国家将极难切入竞争。因此，国际电信联盟（ITU）和国际标准化组织（ISO）等多边平台的多利益相关方参与显得尤为关键。 --- ### [08:51](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=531s) 全球标准共识的突破口与开源模型治理之辩面对中国深度参与多边标准组织的现状，Daniel Dobos 坦言这既是竞争也是合作。他举例指出，“儿童保护”（Child Protection）是超越地缘政治竞争、能够迅速达成全球协作的典型领域。各国政府目前在评估 AI 对下一代影响的科学证据上面临共同挑战，在这类具有普遍社会共识的基础议题上先行建立防护栏，能够为未来解决更复杂的 AI 标准化挑战奠定信任基石。随后话题转向开源模型的爆发式增长（尤其是中国开源模型的迅速崛起）对安全监管带来的冲击： Lee Tiedrich 分析了开源权重（Open-weight）模型带来的治理难题： 1. **安全护栏易被剥离**：开源权重使得恶意使用者能够轻易移除原有的安全限制，加剧了滥用风险。 2. **企业使用考量与成本驱动**：如《金融时报》等媒体报道，许多美国企业出于成本考量，选择将中国开源模型部署在本地服务器上运行，其使用成本远低于专有闭源模型。 3. **模型蒸馏与知识产权纠纷**：部分美国企业指责中国模型开发者通过蒸馏（Distillation）技术侵犯其专有模型的知识产权，引发了激烈的 IP 维权争议。 --- ### [12:04](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=724s) 开源透明性、网络安全与历史软件模式的呼应 Daniel Dobos 从网络安全与科研人员的角度对开源模型的风险论提出了反驳。他回顾了 Web 服务器和浏览器的发展史：如今几乎没有人在构建非开源的基础 Web 服务器或浏览器，因为历史证明全开源、透明审查才是构建关键数字基础设施最安全的方式。 Daniel 认为开源 AI 模型具有不可替代的安全优势： - **免受投资回报率（ROI）压力**：相较于部分商业闭源模型，开源模型不受短期商业利益驱使。 - **全链路白盒安全审查**：大学及全球研究人员可以深入剖析模型偏差（Biases）与训练数据集。例如瑞士近期发布的完全开源基础模型 **Apertus 1.5**，以及新加坡、阿联酋和中国等地的开源模型，使得全球科研界能够像航空监管机构拆解检查飞机的每一个零部件并进行适航认证一样，对 AI 价值生成链的每个环节进行风险审查与认证。 Lee Tiedrich 随后借用名言“历史不会重复，但总是押着相同的韵脚”（History doesn't repeat itself, but it rhymes）进行了总结。她指出，数十年前专有软件企业也曾对开源软件充满敌意与恐慌，但如今专有软件与开源生态已然并存繁荣。随着 AI 市场的成熟，专有模型与开源模型必将在不同的应用场景中各司其职。 --- ### [14:39](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=879s) 美、中、欧/瑞三方路径：资本扩张、人力密集与精准创新针对欧洲是否在美中竞争中沦为“旁观者”的疑问，Daniel Dobos 对比了三大区域截然不同的 AI 发展哲学： | 发展路径 | 核心特征与策略 | 典型优势与模式 | 面临的挑战 | |:--- |:--- |:--- |:--- | | **美国模式** | 资本密集型（Throwing money at the problem） | 依托雄厚资金强力推进模型规模化（Scaling）扩张 | 联邦缺乏统一法律，各州立法割裂（全美达 2,193 项州级法案，如内华达州仅生效少数法案） | | **中国模式** | 人力密集型（Throwing manpower at the problem） | 围绕特定商业模式快速涌现数百家初创公司赛马淘汰 | 内部同质化竞争激烈，依赖海量工程人力快速试错 | | **欧洲/瑞士模式** | 精准与高能效创新（Efficient & Targeted Innovation） | 连续 12 年蝉联全球创新冠军，聚焦垂直细分利基市场，培育“隐形冠军” | 资源总量有限，整体监管框架（如 EU AI Act）常被诟病过于繁琐 | Daniel 解释道，瑞士等欧洲国家的策略并非盲目跟风拼算力规模，而是如同制造高端机械手表、精细化工或高精度微机械一样，通过高度聚焦的精准投资培育垂直领域的全球市场领跑者，同时通过清晰的监管为产业提供确定性。 --- ### [17:42](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=1062s) 科学外交 vs. 政治博弈：Pax Silica 与多边对话前景在研讨会尾声，主持人提及美方的“硅和平”（Pax Silica）构想以及中国联合 29 国发布的多边倡议，询问在两国元首会晤探讨过建立安全对话的背景下，高层战略对话是否可行。 Lee Tiedrich 强调了“科学外交”的重要性。她回忆自己在参与由英国主导、Yoshua Bengio 等科学家领衔的《国际 AI 安全报告》编写过程中的经历，指出当时的报告严格界定为“不提供政策建议”，只专注于寻找测量科学的客观事实基准（Ground zero truth on the science）。在该机制下，包括中美在内的多国顶尖科研人员实现了高质量的紧密合作。 Lee 总结认为： - **科学与技术层面**：在度量科学、模型评估技术等纯科研议题上，中美及全球多边合作完全具备可行空间。 - **政治与地缘层面**：一旦议题上升到政治挂帅与综合国力竞逐，达成共识将极其艰难。让科学家和研究人员坐在同一间会议室里寻找事实，远比让政治领导人达成一致要切实可行得多。主持人最后总结致谢，指出本次对话呈现了深入的美国与欧洲/瑞士视角，并期待未来能进一步引入中国视角的直接交流。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=4s) 当下的 AI 格局定义为一场“AI 军备竞赛”，但这仅仅揭示了全貌的一半；在幕后，围绕安全标准的制定与推进其实正进行着大量工作。 — 主持人 > [02:29](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=149s) 全球目前面临着巨大的“AI 评估差距”——无论是在模型部署前还是进入现实环境后，业界在如何准确测量与评估 AI 系统的表现上仍缺乏有效工具与共识。 — Lee Tiedrich > [02:29](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=149s) 100 年前航空业充满风险，但正是通过全球协作与统一安全标准的建立，商业飞行才成为当今全球最安全的交通方式。 — Daniel Dobos > [05:46](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=346s) 美国商务部正组织企业联盟，致力于向全球市场输出“美国 AI 技术栈（US AI Stack）”，以确保全球盟友与新兴市场深度绑定美国技术底座。 — Lee Tiedrich > [05:46](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=346s) 技术标准的锁定效应极其显著，一旦某种标准率先在新兴市场扎根，其他国家将极难切入竞争。 — Lee Tiedrich > [08:51](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=531s) “儿童保护”是超越地缘政治竞争、能够迅速达成全球协作的典型领域。 — Daniel Dobos > [12:04](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=724s) 历史证明全开源、透明审查才是构建关键数字基础设施最安全的方式。 — Daniel Dobos > [12:04](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=724s) 历史不会重复，但总是押着相同的韵脚。 — Lee Tiedrich > [17:42](https://www.youtube.com/watch?v=hgAcbSQiVSU&t=1062s) 让科学家和研究人员坐在同一间会议室里寻找事实，远比让政治领导人达成一致要切实可行得多。 — Lee Tiedrich ## Mentioned Daniel Dobos, Lee Tiedrich, Yoshua Bengio, Geoffrey Hinton, 李飞飞 (Fei-Fei Li), Tiedrich Global Strategies, AI For Good Impact Initiative, RegulatingAI, NIST (美国国家标准与技术研究院), OECD (经合组织), CERN (欧洲核子研究中心), CAISI (AI 标准与创新中心), NSA (美国国家安全局), CISA (网络安全与基础设施安全局), 美国商务部, ITU (国际电信联盟), ISO (国际标准化组织), WAIC (世界人工智能大会), EU AI Act, Apertus 1.5, 国际 AI 安全报告 (International AI Safety Report), Pax Silica (硅和平) ## Video description Daniel Dobos - Co-Chair, AI For Good Impact Initiative Lee Tiedrich - President & CEO, Tiedrich Global Strategies National security is the fastest-growing dimension of AI policy, with bipartisan urgency around export controls, autonomous weapons, and the intelligence community's AI adoption. This panel ensures the Summit addresses the full spectrum of AI governance — not just enterprise compliance — and elevates RegulatingAI's positioning in the national security policy space, where AI governance funding and political attention is increasingly concentrated. [Source](https://www.youtube.com/watch?v=hgAcbSQiVSU)

Source quotes

[0:04] 主持人: 当下的 AI 格局定义为一场“AI 军备竞赛”，但这仅仅揭示了全貌的一半；在幕后，围绕安全标准的制定与推进其实正进行着大量工作。 [2:29] Lee Tiedrich: 全球目前面临着巨大的“AI 评估差距”——无论是在模型部署前还是进入现实环境后，业界在如何准确测量与评估 AI 系统的表现上仍缺乏有效工具与共识。 [2:29] Daniel Dobos: 100 年前航空业充满风险，但正是通过全球协作与统一安全标准的建立，商业飞行才成为当今全球最安全的交通方式。 [5:46] Lee Tiedrich: 美国商务部正组织企业联盟，致力于向全球市场输出“美国 AI 技术栈（US AI Stack）”，以确保全球盟友与新兴市场深度绑定美国技术底座。 [5:46] Lee Tiedrich: 技术标准的锁定效应极其显著，一旦某种标准率先在新兴市场扎根，其他国家将极难切入竞争。 [8:51] Daniel Dobos: “儿童保护”是超越地缘政治竞争、能够迅速达成全球协作的典型领域。 [12:04] Daniel Dobos: 历史证明全开源、透明审查才是构建关键数字基础设施最安全的方式。 [12:04] Lee Tiedrich: 历史不会重复，但总是押着相同的韵脚。 [17:42] Lee Tiedrich: 让科学家和研究人员坐在同一间会议室里寻找事实，远比让政治领导人达成一致要切实可行得多。

Source quotes

[0:04] Lee Tiedrich: If you read the Trump administration policy, we are in a race with China. That is definitely the tone that's been set [0:04] AI arms race [2:29] Lee Tiedrich: AI evaluation gap [2:29] Daniel Dobos: A hundred years ago aviation was highly risky and by global collaboration, by setting global standards, we reached all together that flying commercial flying is the safest way of transportation that we have today globally... At the moment the US is losing a little bit of this power because it's losing trust, and I think that is really sad for the entire world. [5:46] Lee Tiedrich: US AI stack [14:44] Daniel Dobos: Throws money at the scaling problem [14:44] Daniel Dobos: hidden world champions [17:42] Lee Tiedrich: Track II [17:42] Lee Tiedrich: If we can sort of keep the discussion at the scientific level... like when I worked on the International Safety Report for 2025–2026, the report was clear: we were not making policy recommendations. This was just trying to come up with ground-zero truth on the science. And there we actually had collaborators from all different countries—China as well as Western countries.

Original transcript

[0:04] We have great panelists. Everyone wants to call this an AI arms race, especially globally, [0:09] but I think that's only half the story. Behind the scenes, there's a lot of work on [0:14] safety standards that's going on. And we have two great panelists hopefully that'll talk about it, [0:21] disagree, and give some nonwaffling answers. So, we'll have them introduce themselves and also [0:30] answer one question. Is AI governance now defined more by competition or collaboration with you? I [0:41] say it's governed more by introduction. My name is Lee Tedri. I'm president CEO of Tri Global [0:46] Strategies. I've been working in emerging tech for over 30 years. I was an Amlaw 50 law partner [0:52] um with a background in engineering and law. um have worked at NIS, the US National Institute of [0:57] Standards and Technology, been a consultant for the OECD and I had the honor of working [1:01] with Joshua Benjio, Jeff Hinton on the 2025, 2025, 2026 International AI safety reports. Oh, [1:09] the question the question I think it is we're definitely in an era of competition. Picking [1:15] up on what our prior panel said, if you read the Trump administration policy, [1:19] we are in a race with China. That is definitely um the tone that's been set and China has recently [1:26] held their world AI conference and they are also picking up on on the race theme. Daniel, [1:33] let's hear it. My name is Dr. Daniel Doash. I'm a particle physicist by education. Have been worked [1:39] at CERN on real time AI inference for a lot of my career. uh worked as strategic foresight um [1:48] director at the United Nations and also work in AI standardization uh representing Switzerland there [1:54] and yes I totally agree we are in collaboration there is collaboration let me put it this way [2:02] there's also collaboration especially if we look into safety standards if we talk about technical [2:08] experts who work globally together there is collaboration there's multilateral collaborations [2:14] possible but in many aspects we are in competition in a competition well that's why we asked you to [2:21] come here from Switzerland so you could talk about competition with her but u u so Lee you [2:29] have argued that actual standards help innovation helps startups etc the general perception is that [2:38] sometimes they create more bureaucracy et yeah so just to take a step back When we talk about [2:44] standards, uh, standards are different than regulation. And I just wrote a piece that the [2:50] OECD published about a week ago on and this this comes from the AI safety report. You know, we have [2:56] an AI evaluation gap that we need to close. Um, we don't know how to measure and evaluate [3:02] um AI uh as well as we need to um both in the pre-eployment setting as well as when it goes out [3:09] into the real world. And standards are technical standards, right? Um, you know, we have standards [3:14] for all different types of technology and everyone seems to agree that we have an evaluation gap. [3:22] Part of the Trump administration AI action plan is we need to build an AI evaluation ecosystem. So, [3:28] what I'm trying to do is get people to focus on where do we have agreement. Um, evaluations are [3:33] really grounded. You heard uh Fafy Lee and and the panelists say, you know, we need more measurement [3:38] science and let's try to get the scientists to agree with multiddisciplinary input on what the [3:45] measurement techniques are, what the evaluation techniques are, fill the gap and then once you [3:50] have that established, you can then um figure out how to how to regulate and having common standards [3:56] creates more latitude. So if the US and certain countries prefer a more deregulatory approach [4:02] that works because standards don't necessarily need to be regulation but then if you look at the [4:07] EU and other jurisdictions where they prefer more regulation they can they can make them mandatory [4:12] but we at least have a common technical standards and then companies know what they need to do to [4:17] comply. So do you want to respond to that? Let me let me add one that's why we had you let me add [4:25] add one aspect to that that we had a little bit in the pre- discussion if the US is losing you [4:31] know some of the power in setting this global standards and my answer to that would be yes [4:38] not absolutely but relatively and I would like to compare it for example to setting safety standards [4:43] in global aviation 100 years ago aviation was highly risky and by global collaboration by [4:50] setting global standards we reached all together that flying commercial flying is the most safest [4:57] way of transportation that we have today globally and you can now say oh look this is you know maybe [5:04] national security concerns but we don't use the same kind of say AI safety standards or aviation [5:10] safety standards for defense purposes than for commercial flying so my argument would be that at [5:16] the moment the US is losing a little bit of this power because It's losing trust and I think that [5:23] is really sad for the entire world because a lot what we reached together in global risk management [5:29] of aviation came from strong trust that existed globally also led by the US and if we don't [5:38] have the same kind of trust in setting global AI standards we are losing all together. Yeah. So, [5:46] u Le is to both of you. Some people view standard setting as another way of, you know, capturing [5:56] competition. You know, the country that sets the standards kind of dominates the marketplace, [6:02] etc. How do you view this? Well, I I think a a couple things. I mean first of all um you [6:11] know I guess it's you know the US is investing a lot in building the AI evaluation ecosystem. Um, [6:19] you know, I worked at NIST for a year and NIST is very busy, including the center for AI standards [6:25] and innovation, um, Casey, which is part of NIST, the president's, um, recent executive [6:31] order establishing voluntary testing for the frontier models where he's bringing in NSA, [6:37] um, CISA. So you know that there I agree that you know on the international stage you know compared [6:44] to the Biden administration um the US is is not as involved um in some of the UN activities but [6:50] in terms of trying to develop the the safety mechanisms I think um you know we're we're [6:56] still there I'm sorry the question I'm saying is that standard setting sometime is uh you know [7:03] you use it as a trade mechanism too when you set standards GSM CD DMA and we saw that with cellular [7:12] um you know we had the competition over you know with Huawei that um and and this is where [7:18] you start to see the US China um competition come in you know one of the prongs in the president's [7:25] AI action plan is we want to you know global collaboration which h you know has translated [7:30] into we want to export the um US AI stack and the commerce department is pulling together [7:36] companies or creating forums for companies to come together to export the stack because um you know [7:42] we from the US perspective US policy we want c countries and companies in those countries to be [7:49] using US technology I mean similarly China they just had their world AI conference and they've [7:55] created a new multi-aterateral organization with uh over two dozen countries and they're very much [8:02] focused on open-source open access um models and it's very much you know they want to proliferate [8:08] it, democratize it. And there is a concern that if one technology gets into some of these emerging [8:14] markets um and it has a certain set of standards that those standards end up getting locked in and [8:19] it becomes harder for other countries to compete. So you know standards can be a competitive issue [8:26] and that's why I mean I really value the work of standards organizations. We have ITU where [8:31] I know Daniel's been very involved which is part of the UN. We have ISO, the International [8:35] Standards Organization. And really, you know, the hope of those organizations is they bring [8:40] together different stakeholders from around the globe and we can come up with something that can [8:45] work globally that's acceptable to to everyone and has that multistakeholder input. Daniel, [8:51] uh, China is also in a lot of these standard bodies. Is that collaboration or is there some [8:59] competition or you know things going on there? Uh I know you don't like waffling around but I have [9:07] to say it's both. Okay because it's I think the technical experts are able to work together and [9:14] identify you know what kind of guard rails do we need where it makes sense to create global [9:20] standards and where it makes sense to not do that. Let me give you one example uh child protection. [9:27] I think child protection is one topic where it makes sense. Every government is having trouble [9:35] at the moment to really understand to find the scientific evidence what it will do with the next [9:40] generation if we are not putting some guardrails into protecting our children. There I think [9:47] collaboration is possible despite all kind of geopolitical you know competition. I think their [9:53] collaboration is possible. And so this is where my hope is that you know with this few kind of things [9:58] like child protection and maybe some other safety related standards we can do some first steps which [10:04] allow for you know better exchange and seeing how this global standards on this maybe easier [10:10] task work out and see can we also then go to some more complex kind of AI standardization. [10:17] Lee, open-source models are now, you know, proliferating and they're growing and growing [10:23] faster, especially the Chinese models. Does that make it harder to do regulation around them [10:31] uh to create some safety standards? Does open source make it a little harder? I think it does. [10:37] And I you know to sort of take a step back um you know the open- source models that are attracting [10:45] a lot of attention are coming out of China but if you go to the um president's AI action plan [10:52] again it does talk about the importance from a US perspective of open innovation [10:58] um you know I think that the regulation of it is tricky for several reasons. I mean with the open [11:04] weight models it is easier to strip out um some of the safeguards which then raises you know the [11:11] stakes on some of the safety issues and raises additional questions. You know I think right now [11:16] we're also seeing um an interesting phenomena that a lot of it's been reported in financial times [11:23] and other newspapers but a lot of you know US companies are starting to use the Chinese models. [11:28] they bring them onto a local server and from a cost perspective um it is it's cheaper for them [11:35] to run them than the proprietary models and then there's the whole debate over the distillation [11:40] um there are US companies who believe that um their intellectual property is being violated [11:46] by the Chinese companies through this distillation um is that IP threat I think that you know issue [11:53] of whether or not there is um violation of IP rights is something that's going gets a lot of [11:59] attention. So I think for all those reasons it does make it more complicated. It makes it more [12:04] sorry please. from a researchers perspective and somebody looking also into cyber security I would [12:10] disagree here and looking into our history what we have seen with look at web servers look at [12:16] web browsers nobody tries to build a nonopensource web server or web browser today because everybody [12:22] we learn together that this is the safest way of building this kind of critical infrastructure and [12:29] this I would argue also for AI on the long term I don't know say that it will happen immediately [12:35] But this open source models they don't have the same kind of return on investment pressure as [12:40] some of the open weights model have. Also the open source allows when researchers, safety, [12:47] security researchers at universities to look into biases uh look into understand for example [12:53] if you some of the open-source models that you mentioned coming from China or from Singapore or [12:58] from UAE or from Switzerland like the opensource fully open-source Swiss model Apertus which just [13:05] came basically out in the second version 1.5 a few days ago. It allows researchers worldwide [13:12] to look what kind of biases exist, what data was it trained for and really understand what kind of [13:18] safety risks it it's uh it creates. In a similar way like aviation where in aviation you also have [13:26] uh aviation regulators who look at every single piece of an airplane and understand you know where [13:33] the risks are and if they find risk they basically regulate it. They certify each part to make sure [13:38] that the entire value generation chain is safe. Well, you keep talking about aviation. It was a [13:44] long flight from Switzerland, right? [laughter] I think it's a good we learned together there in [13:49] aviation very well how to manage risks together on a global scale and to do that for commercial [13:55] things but on the other side keep away all the you know global competition and the defense [14:01] competition out of that. And I think also and just you know on the open weight versus the proprietary [14:07] models I mean history doesn't repeat itself but it rhymes and you know you can go back a couple [14:12] decades and the proprietary software companies were very concerned about open source and you [14:18] know lo and behold look fast forward to today they coexist like there is a marketplace for [14:22] proprietary software and as Daniel said there's a marketplace for open software and I imagine that [14:28] you know as the market for AI continues to mature there's going to be roles for all these different [14:34] types of applications. Daniel, we've so far talked about US, China, so to speak, the you know thing [14:44] going on. You come from Europe. Where does Europe just briefly fit into this thing? We just we the [14:52] conversation is always about US and China and it seems like Europe is an afterthought. Apologies [14:58] intended, not by me. No, I I I think it's a very good question and this is something that we have [15:04] seen in technology development over the last decade to multiple times. There's nothing new [15:08] in AI. I think having this competition between us and China, I think everybody accepts that. [15:17] The European perspective is a little bit different on that because maybe innovation we understand [15:23] especially in Switzerland as uh world innovation champions uh for I think now 12 years in a row we [15:30] we basically have a perspective at the that the US is often throwing money at the problem the AI [15:36] problem basically you know how to scale AI there's a lot of money thrown China is a little bit [15:42] different it throws mostly manpower on it where there's a business idea and then where hundreds of [15:47] startups created following the same business idea and one team might survive. I think especially [15:52] Switzerland is trying to do some kind of you know with a very limited resources that we have as a [15:58] small country some efficient innovation trying to understand you know what is really necessary what [16:05] is really efficient and then with very targeted investments trying to invest into the right kind [16:12] of you know businesses right kind of solutions and this is where Switzerland has a lot of hidden [16:20] world champions small companies who are very very you know niche in the vertical and they are the [16:27] world marked leaders of uh certain solutions. think about watches or chocolate or high um you [16:35] know high precision fine mechanics and I think this is more of the Swiss European approach to [16:41] it to try to understand also not to rush too fast have a regulation that is that is you know giving [16:49] certainty to some players and you know everybody complains about the EUI act and on the other side [16:55] we heard basically you know that there's quite some you know confusion about what exists as [16:59] as federal laws in the US and on the other side having 2,193 state laws dealing with AI. I think [17:08] Nevada is one that has only very little I think six bills in total and three in in in in action [17:14] engaged um in in power. So this is something I think where this race is very interesting to [17:21] follow in the three regions and there are three different kind of philosophies how to do that. [17:26] they're not necessarily compatible which makes it for global companies very difficult but I [17:31] think that we will learn how together you know act in that and how to basically use the power [17:36] of each of his philosophies and also get rid of the weaknesses of all of the three philosophies. [17:42] U Lee we are a little short on time but let me ask you this. We have the PAC silica in the US and you [17:49] know you and president X she announced in China their grouping of I think about 29 countries. [17:57] It seems like you know hey who can I recruit better faster more? When President Biden had met [18:06] uh President Xi in California, there had been some conversations about creating some kind of a safety [18:13] dialogue and there was at least the thought that when President Trump and she met that there would [18:19] be some thoughts. What are your views? I mean should there be that level of dialogue? Is there [18:25] a possibility of that dialogue? And how do you see PAX silica and the the Chinese announcement [18:31] going on? I I think if we can sort of keep the discussion at the scientific level like when [18:37] I worked on the international safety report for 2025 2026 we had it was the report was clear we [18:44] were not making policy recommendations. This was just try to trying to come up with ground zero [18:49] truth on the science and there we actually had collaborators from all different countries China [18:56] as well as western countries. So I think if we can find a forum and keep it focused on you know what [19:03] is the science um what are measurement what are evaluation techniques there's an o potentially an [19:10] opportunity to collaborate I mean the UK um led the safety reports and that worked out well and [19:16] the UN is trying to pick up some some of that work um you know when you get into some of the [19:21] political issues and the competition issues that that may be a bit more challenging so you don't [19:26] see at a political level, it happening more at a safety level. Um, I think getting researchers [19:34] in the room is going to be a little easier than trying to get the political leaders in the room to [19:39] reach agreement. Okay. Well, you heard the Swiss perspective, we heard the US perspective. Uh, next [19:47] time we'll try to bring the Chinese perspective, too. So, thank you both for a fantastic gathering. [19:53] Really appreciate it. Thank you very much. Thank you very much. [applause] Thank you.