US AI Dominance Push Risks Global Alienation as China Taps Open Source

美力推AI主导权恐遭全球疏远，中国借力开源突围 美国对技术霸权的激进追求可能导致国际同盟破裂并加速欧洲脱钩；与此同时，中国正通过开源模型调动工程规模效应，绕开硅谷在资本和顶尖人才方面的优势。 在Sanjay Puri的主持下，讨论嘉宾Thorsten Jelinek、Ryan Wang与Justin Hendrix对比了推动美国、欧盟、中国和“全球南方”AI治理的独特法律传统、资本动态与战略哲学。研讨小组将美国以私人资本为主导的闭源范式与中国国家驱动的基础设施建设及开源模型优化进行了对比，同时指出了市场炒作和各州碎片化立法的风险。在评估自主AI智能体的兴起与日益加剧的地缘政治竞争时，发言专家呼吁建立技术与治理互操作性，并推行类似冷战时期“开放天空”的能力通报框架，以防范灾难性的升级风险，确保技术红利惠及全球大多数人。 全球 AI 监管分歧源于深层的法律传统差异：美国依托普通法系且缺乏全国统领性法规，欧洲遵循成文大陆法系，而中国采用国家集体主义模式；伴随地缘政治对抗加剧，各国对技术主权的追求正逐渐模糊与经济保护主义的界限。 (6:21) 资本与人才分布塑造了中美截然不同的技术路线：美国凭借充沛的全球流动性与极端估值溢价牢牢把控顶尖闭源人才，而中国在美元基金撤退背景下，正系统化调动庞大的本土 STEM 研发资源，在 DeepSeek 和阿里通义千问（Qwen）等开源基座模型上形成独特优势。 (9:02) 自主 AI Agent 的爆发正将核心算力需求由预训练转移至推理端，加剧了底层能源与算力基建压力；相较于中国自上而下、高速落地的基础设施建设，美国类似“星门计划”（Project Stargate）等大型项目仍深陷多方商业与土地利益博弈的延误中。 (22:44) 针对多 Agent 串联与协作引发的治理风险，专家强调法律责任与道德问责必须严格保留在人类开发者与部署企业身上，绝不能将失控归咎于 Agent 本身；同时需防范市场估值泡沫一旦退潮，政策制定者可能因过度反应而仓促放弃必要的长期安全治理。 (26:50) 为缓解中美间零和博弈与战略恐慌，专家建议参考冷战时期的“开放天空”条约建立前沿模型通报机制，将底层算力竞争与安全对齐、生物威胁等全球风险管控彻底解耦，在共同底线问题上维持跨国协作。 (32:57)

Divergent legal traditions and venture capital structures between the United States, China, and the European Union are driving fragmented approaches to AI regulation, leaving developing nations caught between competing geopolitical models. Legal Traditions and Domestic Regulatory Fragmentation (6:21) Thorsten Jelinek traces diverging global regulations to underlying legal traditions rather than partisan rhetoric alone: the case-based US common law system, Europe's codified civil law, and China's state-collectivist hybrid model. This foundational divergence explains the absence of an omnibus federal AI statute in the United States, where governance evolves incrementally through case-by-case adjudication alongside political gridlock. Both Jelinek and Justin Hendrix warn that the intensifying pursuit of digital sovereignty risks blurring legitimate national security concerns with economic protectionism. Hendrix emphasizes that US domestic governance faces further internal fragmentation. While conservative states lean toward deregulation and federal moratoriums on state-level oversight, states like Colorado, California, Connecticut, and Texas are establishing distinct local mandates. Hendrix argues this patchwork remains manageable for now, though friction between state-level consumer protections and federal deregulatory ambitions persists. Capital Asymmetry and the Open-Source Divide (9:02) Ryan Wang highlights a sharp divergence in capital and talent distribution between the US and China. US markets support extreme startup valuations—exemplified by robotics firm Figure AI raising near a $40 billion valuation pre-product—and offer liquid exit paths that draw the top tier of frontier researchers into proprietary, closed-source ecosystems via substantial compensation packages. Conversely, Chinese foundational model ventures pivoted to open-source architectures like DeepSeek and Alibaba's Qwen following the retreat of foreign US-dollar venture capital over recent years. Constrained by tighter capital and lower valuations, Chinese firms rely on a large domestic pool of STEM PhDs to systematically refine model architectures using public research papers, establishing an open-source engineering edge while US incumbents maintain closed frontier systems. Infrastructure Scaling and Autonomous Agent Accountability (22:44) The rapid development of autonomous agents using chain-of-thought reasoning shifts technical demand from model training to intensive inference, placing severe pressure on foundational power grids and data centers. Wang contrasts the slow, complex private-sector financing negotiations of US megaprojects like Project Stargate against China's centralized command structure, which continues building infrastructure rapidly despite US export restrictions on advanced semiconductors. On the governance of autonomous agent swarms, Hendrix cautions against granting agents quasi-legal or moral autonomy that shields creators from liability. Because multi-agent commercial transactions and daisy-chained models can trigger unexpected systemic failures, he insists legal liability must strictly remain with the developers and corporate entities that deploy catalytic software systems. Great Power Competition and the Global Majority (32:57) Hendrix critiques the US official posture for demanding unquestioned global technological dominance, warning that aggressive hegemony alienates European allies and pushes international partners away. Meanwhile, China leverages multilateral forums like BRICS and the UN to offer transactional infrastructure alternatives, leaving developing economies across Latin America, Africa, and India navigating polarized spheres of influence. To prevent destabilizing escalation, Jelinek proposes adapting historic arms-control concepts into an "Open Skies" framework for frontier AI. Under this model, competing powers would implement mutual notification protocols when deploying hyper-capable models, preserving competitive development on model capability while establishing mandatory bilateral cooperation on biosecurity, counterterrorism, and catastrophic risk alignment. The panel reaches no consensus on whether the American, European, or Chinese regulatory framework will ultimately dominate. Hendrix concludes that long-term legitimacy will belong to whichever governance model tangibly serves the interests of the global majority over the coming decades, leaving open whether international collaboration on safety can endure escalating geopolitical rivalry and market volatility.

在 Ai4 国际人工智能治理研讨会上，Thorsten Jelinek、Ryan Wang 与 Justin Hendrix 指出，美、欧、中在法律传统、资本逻辑与战略意图上的根本分歧正加剧全球监管碎片化与地缘零和对抗；专家们认为，面对自主智能体演进、算力基建压力以及全球南方的治理诉求，国际社会亟需超越技术霸权与保护主义，建立类似冷战“开放天空”条约的跨国安全通报机制与多边互操作治理框架。 法律传统差异与美欧监管路径分歧 (6:21) Thorsten Jelinek 指出，美、欧、中三方监管路径的割裂根植于底层法律传统，而非单纯的党派政治极化。欧洲依托大陆法系传统的成文法典，倾向通过自上而下的全面法规（如欧盟《人工智能法案》）进行规范；美国则立足普通法判例传统，依赖法庭判决个案迭代演进，这解释了为何美国联邦层面长期缺乏综合性全国 AI 法案。同时，各国对“技术主权”的追逐正在模糊正当的国家安全防御与经济保护主义之间的界限。 Justin Hendrix 进一步分析称，美国官方政策并非中立的“创新优先”，其行政部门明确追求不可挑战的技术霸权，但国内面临严重的监管碎片化。保守的“红州”倾向配合联邦的去监管诉求，尽管一项联邦全面暂停令在国会以微弱差距被否决，但随后联邦 AI 行动计划提出以审查各州监管环境来挂钩联邦资金的“温和暂停”方案；科罗拉多、加利福尼亚、康涅狄格、纽约和得克萨斯等州则各自推进本地法规，导致全美合规环境陷入持续博弈。 资本市场估值与中美开源闭源分流 (9:02) Ryan Wang 从风险投资与资本结构切入，对比了美中两国的产业生态。美国市场享有极高的流动性与估值溢价，例如人形机器人初创公司 Figure AI 在产品尚未商业化阶段估值即接近 400 亿美元，Scale AI 49% 股权估值达 140 亿美元，Meta 等科技巨头更为顶尖研究员开出高达数亿美元的薪酬包。由于全球前 0.0001% 的前沿人才高度集中于硅谷，促使美国头部机构普遍构筑专有闭源生态。 相比之下，中国在过去数年经历了西方美元风险投资的大幅撤离，初创企业面临估值压缩与退出渠道受限的困境。为此，中国企业（如 DeepSeek、阿里巴巴通义千问、腾讯等）战略性地转向开源基础模型。Wang 认为，这一转变并非技术劣势，而是依托国内庞大的 STEM 博士与工程研发人才储备，通过公开论文进行底层架构优化与极高效率的工程落地，使开源模型成为缩小技术代差与提供普惠基础设施的关键工具。 智能体演进、算力基建与问责机制 (22:44) 随着行业重心转向依赖思维链推理的自主 AI 智能体（如中国的 Manus），计算负荷的核心正从模型训练向推理端发生重大转移，对电力与数据中心等底层基础设施构成了空前压力。Wang 指出，美国的大型基建倡议（如星际之门 Project Stargate）面临土地、资金和跨企业谈判的重重阻力；而中国凭借集中的自上而下体制，即使面临英伟达先进制程芯片的出口管制，仍能以极高效率建设低成本、高可靠的算力基础设施。 Hendrix 则针对多智能体串联（daisy-chain）与自动化商业交易的潜在风险发出警告，坚决反对赋予智能体独立的法律或道德地位。他类比排放催化化学品的侵权责任，强调释放自主智能体所造成的不良后果必须严格追溯至部署该系统的个人或企业。Hendrix 预计短期内州级立法将首先聚焦聊天机器人的心理健康与儿童保护，随后逐步扩展至自主智能体商业行为的问责框架。 地缘霸权诉求与“开放天空”治理倡议 (32:57) 针对华盛顿两党普遍将 AI 视为赢者通吃的霸权竞赛、并促使微软等企业撤回在华研发团队的现状，Jelinek 警告当前 AI 资本支出已达上世纪电信投资的 10 至 100 倍，且西方算力主要掌控在私营科技巨头手中。为了避免误判导致灾难性升级，Jelinek 提议借鉴冷战时期的“开放天空”条约，建立前沿大模型能力的主动通报机制，呼应 Geoffrey Hinton 提出的在模型“善良（kindness）”与对齐层面保持合作、在原始智力层面展开竞争的构想，即使鹰派代表如 Eric Schmidt 也开始承认美中需就生物安全与防范恐怖主义等灾难性风险展开对话。 与此同时，印度、东盟、非盟及拉美等全球多数国家（全球南方）正面临被动卷入两极阵营的困境。Hendrix 指出，美国要求盟友顺从其霸权技术，中国则依托金砖国家（BRICS）和联合国推动务实交易。Jelinek 强调，治理不仅需要技术互操作性，更需要“治理互操作性”，国际社会必须吸取 20 世纪西方垄断国际电信联盟（ITU）的历史教训，防止单一霸权将排他性框架强加于发展中国家。 研讨最终表明，尽管技术理想主义呼吁设立保护人类的“AI 希波克拉底誓词”并将算力导向消除贫困等普遍福祉，但关于美、欧、中哪套治理框架将最终胜出仍未有定论。Hendrix 总结认为，最终的胜者将取决于未来数十年的实践检验中，哪一套框架能够真正向全球多数人口证明其能切实保障并促进全人类的共同利益；而在此之前，跨越地缘阵营的对话与防范系统性风险仍是各方无法回避的底线。

## Background 本视频由 Ai4 于 2025 年 8 月 28 日发布在 YouTube 平台。内容为一场以“国际 AI 治理：全球监管框架比较”为主题的跨国圆桌论坛，由 RegulatingAI 创始人 Sanjay Puri 主持，邀请了政策学者 Thorsten Jelinek、风险投资人 Ryan Wang 与媒体主编 Justin Hendrix 共同探讨美、欧、中及全球南方在人工智能监管路径上的分歧，以及这些政策对资本流动、人才竞争、主权安全与合规落地的深远影响。 ## TL;DR 圆桌深入剖析了美、欧、中三大 AI 监管模式的底层逻辑与法系差异：美国追求技术霸权但受困于各州碎片化与私营资本依赖，欧洲通过全面法典强化数字主权，中国则依托国家力量推进新基建并以开源生态实现突围。投资视角揭示了中美在估值泡沫、退出通道以及 STEM 研发人才基数上的悬殊反差，指出顶级闭源受阻反向加速了中国开源大模型的爆发。针对自主 AI Agent 带来的系统性风险与全球南方被迫站队的困境，嘉宾呼吁建立类似冷战时期“开放天空”的互通机制与治理互操作性。对谈最终总结指出，AI 治理模式的胜出不在于单边霸权，而在于哪套体系能最终向世界证明其真正符合全人类的根本福祉。 ## Atmosphere & style 整场圆桌讨论兼具宏观地缘政策的深邃思辨与一线创投市场的敏锐务实，现场气氛在犀利交锋与风趣调侃间自如流转。主持人 Sanjay Puri 节奏紧凑且妙语连珠，与学者 Thorsten 的制度剖析、投资人 Ryan Wang 的率真直言以及媒体人 Justin 对民主倒退的深切关切碰撞出强烈的思想火花。对谈的情感弧线从开场对地缘割裂与技术霸权的焦虑反思，逐步演进为抛弃非黑即白敌对叙事、呼吁回归人性最大公约数与跨国互信的理智共识，带给观众兼具警示感与建设性的听会体验。 ## Key takeaways - [06:21](https://www.youtube.com/watch?v=1bVutEOc58Q&t=381s) 三大主流经济体在法系传统与治理逻辑上存在本质差异：美国以判例法和逐案诉讼为主驱动敏捷创新，欧盟依托大陆法系出台全面法典《AI 法案》，中国则在成文法基础上强调与国家战略和产业发展目标协同。 - [06:21](https://www.youtube.com/watch?v=1bVutEOc58Q&t=381s) 各国对技术安全和数字主权的正当诉求，正面临演化为贸易保护主义壁垒的现实风险，加剧了全球治理与跨国监管的碎片化。 - [09:02](https://www.youtube.com/watch?v=1bVutEOc58Q&t=542s) 中美创投生态呈现巨大分化：美国初创企业估值高且具备顺畅的资本退出通道；亚洲及中国企业性价比高、工程师及 STEM 研发人才基数庞大，但面临美元基金撤出和退出渠道受限的挑战。 - [11:42](https://www.youtube.com/watch?v=1bVutEOc58Q&t=702s) 全球基础模型生态形成分化格局：美国在闭源顶级前沿模型（SOTA）上占据绝对领先地位，而中国企业（如 DeepSeek、阿里千问等）则因顶尖人才争夺成本等因素，选择在开源生态中实现突围与追赶。 - [15:33](https://www.youtube.com/watch?v=1bVutEOc58Q&t=933s) 美国 AI 政策核心转向追求无可动摇的技术霸权，但国内面临蓝州与红州监管极化、联邦试图冻结州立法的内部矛盾，同时霸权叙事也在反向促使欧洲等地区加快寻求数字主权。 - [20:04](https://www.youtube.com/watch?v=1bVutEOc58Q&t=1204s) AI 智能体（Agents）在大规模自主协同中有可能复刻并放大现存体制的不公；在治理归责上应采用“催化剂隐喻”，不可将 Agent 视为完全独立的实体而免除人类开发者与部署机构的法律责任。 - [22:48](https://www.youtube.com/watch?v=1bVutEOc58Q&t=1368s) 算力基建呈现体制差异：中国展现出自上而下的基建执行力与工程效率，而美国的大型基建计划则受制于电网、资金及跨机构博弈；投资人呼吁摒弃“非黑即白”的敌对叙事，将失控的 AI 风险视为人类共同威胁。 - [33:20](https://www.youtube.com/watch?v=1bVutEOc58Q&t=2000s) 国家能力架构截然不同：美国政府自身缺乏 AI 研发与算力能力并严重依赖私营科技巨头，中国政府自身拥有强大 AI 能力，欧洲优势则集中在工业垂直领域结合。 - [33:20](https://www.youtube.com/watch?v=1bVutEOc58Q&t=2000s) 国际协作可借鉴冷战时期《开放天空条约》的通报机制，在前沿超大模型突破时主动通报能力边界以避免战略误判，并在公共安全与善意 AI 等领域寻求最大公约数。 - [36:17](https://www.youtube.com/watch?v=1bVutEOc58Q&t=2177s) 广大“全球南方”国家正面临在中美两极标准与政治诉求夹缝中求生存的困境；国际社会不仅需要技术层面的互操作性，更亟需构建治理层面的互操作性以避免单极霸权倾轧。 - [43:53](https://www.youtube.com/watch?v=1bVutEOc58Q&t=2633s) 全球 AI 治理模式的最终胜出者并非单纯取决于地缘政治博弈，而在于哪种治理框架能长期证明自身真正能够保障并满足全人类的根本福祉与利益。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=1bVutEOc58Q&t=4s) 圆桌开场与嘉宾亮相：什么 AI 治理问题令你夜不能寐？ 主持人 Sanjay Puri（RegulatingAI 创始人兼主席）向现场观众致开场白。他指出，当前的国际 AI 治理与监管环境极其复杂，许多跨国企业都在苦苦摸索最基础的战略方向——究竟该去哪里开拓市场、在哪里投资、又在哪里销售产品。为了给与会者提供切实可行、能够落地的建议，本场圆桌汇集了来自政策、投资与媒体界的多维视角。在让每位嘉宾做自我介绍的同时，Sanjay 抛出了第一个破冰问题：“在 AI 治理领域，究竟是哪一个问题让你在夜里辗转难眠？” 来自德国柏林 Hertie School 电子政务中心的高级研究员 Thorsten Jelinek 首先发言。这是他第三次参加该大会，他谈到欧洲通常在 8 月处于休假状态，而今年 1 月欧洲迎来了一个关键的“觉醒时刻”。谈及最令自己担忧的议题，Thorsten 直言是“地缘政治背景”以及伴随而来的数字主权争端。他长期致力于中欧及美国之间的政策研究，并早在 2018 至 2019 年间就在中国组织了首批 AI 伦理研讨会，打破了当时西方普遍认为“中国不谈 AI 伦理”的偏见。作为公共政策学者和连续创业者，他当前还深入研究意识形态、AI 伦理以及机器意识等前沿交叉课题。 随后发言的 Outpost Capital 创始合伙人 Ryan Wang 戴着一顶惹眼的帽子登台，引发了主持人和 Thorsten 的趣味互动。Ryan 介绍称，自己在风险投资方面通过 Outpost Capital 布局前沿科技早期项目（涵盖 AI、机器人、空间计算和区块链），在私募股权方面则通过 Solaris Venture Partners 布局数据中心、数字基础设施及成长期 AI 机器人项目。曾在纽约从事投资银行业务、出生并成长于中国的 Ryan 表示，最让自己感到疲惫和担忧的是无休止的“中美对抗与竞争叙事”。他强调，双方存在大量误解，唯有推动更多跨国协作，才能加速迈向更好的未来。 最后，Tech Policy Press 首席执行官兼主编 Justin Hendrix 进行了介绍。在创立这家专注科技政策分析与报道的媒体之前，他曾在《经济学人》（The Economist）工作了约 12 年，并在纽约大学（NYU）领导了十年的新兴媒体技术实验室，目前仍在 NYU 和康奈尔科技学院（Cornell Tech）任教。作为一名 NYU 校友，主持人 Sanjay 听到这里会心地插话：“看到同校紫血人（Bleed Violet）总是格外亲切。”Justin 指出，真正令他夜不能寐的是全球范围内民主治理模式长达数十年的衰退，且这种衰退在包括美国在内的国家正呈现加速趋势。他每日核心思考的命题是：技术究竟是在为构建一个更加公正、公平、可持续的民主未来提供支撑，还是在暗中破坏它？ ```mermaid flowchart TD accTitle: AI治理夜不能寐的核心关切 subgraph Panelists["圆桌嘉宾与核心关切"] T["Thorsten Jelinek<br/>（Hertie School 学者）"] -->|核心担忧| T_Issue["地缘政治割裂与数字主权边界模糊<br/>（安全诉求演变为贸易保护主义）"] R["Ryan Wang<br/>（Outpost Capital 投资人）"] -->|核心担忧| R_Issue["中美科技竞争叙事与跨国误解<br/>（呼吁加强开源与底层研发协作）"] J["Justin Hendrix<br/>（Tech Policy Press 主编）"] -->|核心担忧| J_Issue["民主治理模式的全球性衰退<br/>（技术在赋能与破坏民主间的角力）"] end ``` --- ### [06:21](https://www.youtube.com/watch?v=1bVutEOc58Q&t=381s) 监管哲学与法系差异：欧美与中国的法律逻辑及主权博弈 Sanjay Puri 将话题引向欧洲与中国在监管环境上的深层差异，询问对于全球化企业而言，两地环境究竟有何本质区别。Thorsten Jelinek 指出，外界往往喜欢通过非黑即白的“法治（Rule of Law）”话语体系来进行地缘政治叙事，认为某些国家遵守法治而另一些国家不遵守，但这种二元对立的视角并不利于客观理解。如果从法学传统与法律文化的演进来看，三大主要经济体的监管路径差异有着深层次的制度根源： ```mermaid flowchart LR accTitle: 三大法系与AI监管模式对比 subgraph US["美国：普通法系（Common Law）"] US_1["判例法为主<br/>Case by Case 驱动"] US_2["两党政治竞争激烈"] US_1 & US_2 --> US_Out["缺乏统一综合性 AI 法典<br/>依赖行业敏捷诉讼"] end subgraph EU["欧洲：大陆法系（Civil Law）"] EU_1["成文法典化预先编码"] EU_2["法官严格解释条文"] EU_1 & EU_2 --> EU_Out["出台全面前瞻法案<br/>如欧盟《AI 法案》"] end subgraph CN["中国：混合大陆法系（Hybrid Civil Law）"] CN_1["成文法架构"] CN_2["融入国家与集体主义发展目标"] CN_1 & CN_2 --> CN_Out["自上而下敏捷标准与专项监管结合"] end ``` Thorsten 进一步分析称，当前笼罩在所有法律体系之上的是全球对“主权（Sovereignty）”的极力追求。在经历过去数十年主权被削弱的全球化阶段后，各国重塑经济韧性和治理能力、追求数字主权与技术安全本身无可厚非，但危险在于，眼下的主权诉求正在迅速推高壁垒，使得“正当技术安全”与“贸易保护主义”之间的界限变得愈发模糊，这是当前跨国监管中最严峻的现实挑战。 - **法系对比**：美国依靠判例法逐案应对新技术冲击，欧洲通过全面法典提前规范技术风险，中国则在成文法基础上强调与国家战略目标协同。 - **主权异化**：各国对技术自主权的正当诉求，正面临被异化为保护主义壁垒的风险。 --- ### [09:02](https://www.youtube.com/watch?v=1bVutEOc58Q&t=542s) 投资视角下的中美鸿沟：估值泡沫、人才红利与开源崛起 针对创业团队和投资者最关心的落地选择，Sanjay Puri 向 Ryan Wang 提问：站在一线风险投资人的角度，欧美与中国/亚洲究竟哪一个监管与市场环境更具投资吸引力？Ryan 坦言，自己主要聚焦于美国与泛亚洲市场（包括中国、东南亚、日韩及新加坡等），在对比两地生态时，呈现出两个极具冲击力的反差：估值逻辑的巨大分化与研发人才储备的悬殊。 首先在估值层面，美国 AI 初创企业的估值水平远远高于亚洲同类企业。Ryan 以人形机器人独角兽 Figure AI 为例，其最新一轮融资估值已逼近 400 亿美元，但该公司在当时甚至还处于“产品尚未正式商业化落地（Pre-product）”的阶段。Ryan 直言，这种在产品面世前就给出天价估值的现象，在当前的亚洲乃至全球其他市场是难以想象的。 其次在人才与开源创新层面，Ryan 回应了年初 DeepSeek 发布模型引发全球震动的事件。许多西方观察者对中国能够诞生如此顶级的开源大模型感到不可思议，但 Ryan 认为这完全在意料之中： - **开源机制透明**：开源世界的底层技术研发高度依赖学术界、研究员和高校博士群体，所有核心论文全网公开，并不存在所谓的信息黑盒或独占秘密。 - **理工人才断层优势**：如果横向对比北美与亚洲（尤其是中国）每年的 STEM 及计算机相关博士毕业人数，会发现亚洲在顶尖研发人才的绝对增量上存在着巨大的规模优势。只要有公开的知识体系与庞大的人才基数，出现突破性的技术成果只是时间问题。 ### [11:42](https://www.youtube.com/watch?v=1bVutEOc58Q&t=702s) 中美模型生态对比：闭源领先 vs 开源突围与资本退出困局 从全球 AI 竞争的格局来看，闭源与开源领域呈现出截然不同的图景。闭源领域无疑由 OpenAI 等美国企业领跑，占据着最顶尖（SOTA）的地位；而在开源领域，中国的 DeepSeek（深度求索）、阿里巴巴的 Qwen（千问）以及腾讯等开发的大模型则表现极为抢眼。中国团队通过前沿研究以及庞大的高水平博士与工程师团队，将系统性能和工程化优化推向了极致。 从投资回报与性价比（Bang for your buck）的角度衡量，亚洲及中国 AI 公司的估值相对较低、性价比较高，而美国公司则昂贵得多。然而，在资本退出路径（Exit Strategy）上，两地存在显著差异： ```mermaid flowchart TD accTitle: 中美 AI 企业资本与退出生态对比 subgraph US["🇺🇸 美国市场"] US_Cap["成熟金融市场与全球资本"] --> US_Val["高估值与顺畅的公开市场退出"] end subgraph CN["🇨🇳 中国市场"] USD_Pull["美元基金持续撤离 (近3-4年)"] --> CN_Gov["人民币资本 / 政府引导基金为主"] CN_Gov --> CN_Exit["退出渠道受限与退出预期承压"] end ``` - **资本成熟度与估值**：如果初创企业追求更高估值和顺畅的二级市场退出通道，成熟的美国金融市场仍然是首选，其汇聚了全球各地的资本与买家。 - **中国创投生态变化**：近三四年来大量美元基金（USD funds）逐步撤出中国，国际资本玩家减少。本土投资机构日益依赖各级政府背景资金的支持，导致退出机会面临较多不确定性。 针对“为何中国会在开源领域战略性走得如此靠前（如 DeepSeek、阿里千问等）”的疑问，Ryan Wang 坦言这虽然没有绝对标准答案，但核心原因在于顶尖人才成本与闭源策略受挫后的路径选择。诸如月之暗面（Moonshot AI）或李开复创立的零一万物（01.AI）等中国基础模型公司，起初都尝试过闭源路线，但全球排名前 0.001% 的顶级研发与工程人才绝大多数仍聚集在硅谷及美国本土。美国科技巨头（如 Meta）能够为超级智能团队开出单人上亿美元的天价薪酬，这种人才争夺成本是中国公司难以企及的。在顶级闭源竞争受限的环境下，中国企业通过开源生态迅速凝聚开发者并打开市场，开源因而成为中国模型突围的胜出路径。 --- ### [15:33](https://www.youtube.com/watch?v=1bVutEOc58Q&t=933s) 霸权焦虑、州级割裂与欧洲主权：美国 AI 治理与地缘政治困境 主持人将话题转向美国与欧中的监管体制对比。Justin Hendrix 强调，虽然外界常将美国描述为“创新优先”（Innovation-first），但现任美国总统及相关政策文件明文宣示的核心目标是“无可置疑、不受挑战的技术霸权”（Unquestioned and unchallenged technological dominance），以确保赢得所谓的“AI 竞赛”。 如果未来回望历史发现美国最终“输掉”了这场 AI 竞赛，其深层诱因与自我削弱的苗头如今已初见端倪： - **对技术霸权的恐惧与反弹**：美国追求绝对主导地位，引发了全球对“技术霸权干预他国政治”的深度担忧，反向加速了欧洲及其他地区寻求“数字主权”（Digital Sovereignty）的进程。 - **中国叙事的外交拉力**：中国采取开源技术路线，并在外交上积极倡导多边主义与国际合作，在全球争取盟友与技术伙伴。 - **资本过剩与思想贫乏**：美国市场充斥着巨额流动性资本，但也催生了极度浮躁的泡沫，例如马克·扎克伯格（Mark Zuckerberg）最新发布的宣言被批评为“毫无新思想，只剩庞大资本在支撑”，企业与决策层陷入盲目自满（“Drank its own Kool-Aid”）。 - **科学研究遭受压制**：若美国主动关停科研开放通道并陷入技术孤立，其创新优势将被严重削弱。 在欧洲方面，各国正通过大规模公共资金投入强化数字主权，政治层面上力图摆脱对美国技术依赖的声音正在成为主流。而在美国本土内部，监管格局高度碎片化且政治极化严重： ```mermaid flowchart LR accTitle: 美国联邦与州层面的 AI 监管博弈 subgraph States["各州分歧"] Blue["蓝州: 推进安全护栏与严格合规"] Red["红州: 倾向联邦创新放宽路线"] end States --> Fed["联邦行动计划与干预"] Fed --> Conflict["试图推行州级监管冻结期 (Moratorium Light)<br/>绑定联邦资金支持引发合宪性争议"] ``` 美国两党在 AI 规则上存在明显分野：部分蓝州积极推进设立安全护栏与监管法案；红州则更倾向于追随联邦政府的宽松创新导向。联邦层面曾试图通过“联邦暂停期”（Federal Moratorium）条款彻底架空各州的监管权，虽然该法案在参议院 7 月 4 日法案前被挫败，但类似“轻度暂停”（Moratorium Light）机制仍被写入最新的《AI 行动计划》（AI Action Plan），试图通过审查各州监管环境来决定是否拨发联邦资金，这种做法的合宪性至今仍存巨大争议。 --- ### [20:04](https://www.youtube.com/watch?v=1bVutEOc58Q&t=1204s) AI Agent 自主性与治理范式：技术重塑权力结构 随着技术焦点迅速转向自主运行的 AI Agent（AI 智能体），治理框架是否需要针对其自主性（Autonomy）做出根本性调整成为核心议题。 Thorsten Jelinek 指出，AI 作为一种颠覆性、转型性的技术，其治理规则必须保持动态调整。当海量 AI Agent 开始进行大规模、自动化的自主协作时，它们实际上是在复刻并放大现有的经济与社会系统。这就迫使人类必须反思现有经济系统本身的公平性与合理性： - **历史垄断与体系重演**：在上个世纪的电信行业，全球规则与核心技术长期被西方 5 个国家垄断，国际电联（ITU）等国际组织在技术转移与跨国创新上缺乏动力。 - **多极化与主权觉醒**：随着中国及全球新兴国家的崛起，世界进入多极化（Multipolarity），原本由少数国家主导的“开放”体系不再适用，曾经的“共同战略”演变为各自争夺的“单边国家战略”。 - **Agent 治理的核心**：技术本身既是打破旧等级、实现自由化的力量，但也极易复刻人类不希望看到的结构性不公与剥削。因此，Agent 的治理核心不仅在于制定孤立的技术参数或监管细则，更在于防范其在自动化协同中强化并重演现存体制的弊端。 ### [22:48](https://www.youtube.com/watch?v=1bVutEOc58Q&t=1368s) 算力基建竞赛与中美治理叙事：从“宿敌论”转向共同安全 在探讨 AI 智能体（Agents）的发展演进时，Ryan Wang 重点剖析了具备思维链（Chain of Thought）能力的智能体——例如中国初创团队开发的 Manus 等 Operator 式智能体。这类智能体能够自主拆解并处理复杂任务，但其顺畅运行对算力提出了极高要求，尤其是推理端（Inference）算力的爆炸式增长，直接对底层硬件与基础设施的承载能力形成了巨大考验。 对比中美两国在应对算力基建挑战时的表现，可以清晰看到治理体制带来的巨大差异。美国政界与体制高度碎片化，而中国则展现出极强的中央集权与自上而下的执行力。以美国雄心勃勃的“星际之门”（Stargate）超级计算计划为例，宣布半年多以来，在推进过程中仍面临繁复的融资难题、电网落地困境，以及 OpenAI、软银（SoftBank）与甲骨文（Oracle）之间复杂的权力博弈。反观中国，即便面临英伟达（Nvidia）先进制程芯片的出口管制与制裁，其新基建项目依然以“光速”推进，凭借举国体制在决策后“不争论、埋头干”的特质，实现了极高速度、极低成本且高可靠性的工程交付。 面对两套体制各自的优缺点，Ryan 呼吁重塑国际对话的叙事逻辑，特别是在美国语境下摒弃非黑即白的敌对思维： > “我们不应该一提到中国就指着鼻子说对方是坏人。我想改变这种叙事——我们或许是对立面、是竞争者，但绝非真正的敌人。人类真正的敌人是失控的 AI 技术对文明造成的生存威胁。我们应当聚焦于这一全人类的宏观命题，保持对话与合作的开放渠道。” --- ### [26:16](https://www.youtube.com/watch?v=1bVutEOc58Q&t=1576s) 美国各州立法拼图与智能体责任归属：催化剂隐喻与互操作标准 针对主持人和现场企业家关于美国各州碎片化监管的提问，Justin Hendrix 梳理了当前地方立法的活跃版图。尽管科罗拉多州率先推出了全美首个“全面性 AI 法案”（虽然目前在政治层面略有摇摆），加州、康涅狄格州、纽约州以及最新推出监管框架的德克萨斯州均在积极推进各自的法案。Justin 认为，现阶段虽然各州法案林立，但尚未形成让初创企业无法适从的“混乱迷宫”，大多数监管条文的规制目标依然相当明确且合乎常理。 ```mermaid flowchart TD accTitle: AI 智能体责任归属与治理逻辑 subgraph Agents ["智能体生态交互"] A["开发者 / 部署企业"] -->|"释放/串联"| B["半自主 AI 智能体"] B <-->|"多智能体协商与交互"| C["其他智能体 / 商业系统"] end subgraph Impact ["催化效应与责任认定"] B & C -->|"链式反应"| D["不可预测的负面产出"] D -.->|"法律责任溯源"| A end subgraph Governance ["未来规制路径"] E["底层技术标准 & 互操作性规范"] --> B F["各州法案演进: 心理健康/儿童安全 -> 智能体商业规制"] --> A end ``` 谈及智能体生态的深度治理，Justin 提出了一个深刻的责任隐喻： - **链式交互风险**：随着行业将更多 AI 系统串联（Daisy-chaining），成群结队的智能体将在商业与日常交互中进行自主协商与博弈。 - **催化剂隐喻与人类归责**：“如果我向自然界排放了一种化学物质，它与其他化学品催化反应并产生了有毒害的恶果，你必定会追究我的责任。我们绝不能过度将智能体视作‘独立自主实体’，从而免除最初按下释放按钮的个人、企业或组织的法律与道德责任。” - **标准与互操作性的漫长攻坚**：制定兼具技术细节与行业共识的标准往往耗时数年、极其磨人，但在智能体横行的时代，投资于互操作性标准已成为防范系统性风险的核心护栏。 - **立法演进趋势**：各州监管预计将经历两波浪潮——近期主要聚焦于针对聊天机器人、心理健康干预以及儿童数字安全的保护法案，随后将逐步延伸至涉及智能体自主交易与商业行为的法律问责框架。 --- ### [30:01](https://www.youtube.com/watch?v=1bVutEOc58Q&t=1801s) 泡沫叙事下的政策摇摆与大国主权 AI 竞争困境 针对资本市场与监管层高度敏感的“AI 泡沫论”，对话探讨了繁荣表象背后的狂热与潜在风险。从 Meta 开出 15 亿美元的“人才收购”（Acqui-hire）打包报价、动辄 2.5 亿美元的天价签约奖金，到 Scale AI 获得高达 140 亿美元的估值（对应 49% 股权），再到本次会议规模从 2,000 人暴增至 9,000 人的盛况，都印证了前所未有的资本集中度。 对于监管者而言，泡沫的存在与破裂带来深层的不确定性： 1. **监管过度风险**：欧盟出台《人工智能法案》（EU AI Act）的过程被一些人认为存在“操之过急”（Over its skis）之嫌，在某种程度上正是被业界狂热的情绪与末日般的恐慌所裹挟。 2. **潮退后的政策真空**：一旦市场泡沫破裂、投资降温，决策者可能会迅速转向撤退或忽视监管，导致在未来下一轮颠覆性技术飞跃真正降临之际，全球治理机制再度陷入应对乏力的空窗期。 ``` 【中美 AI 合作与地缘博弈的核心矛盾】 ┌─────────────────────────────────┐ ┌─────────────────────────────────┐ │ 大国竞争与主权叙事 │ │ 全球治理与合作需求 │ ├─────────────────────────────────┤ ├─────────────────────────────────┤ │• 两党共识："赢下 AI 竞赛即可主导世界"│ ◄-?─► │• 应对大模型与超级智能的系统性风险 │ │• 主权 AI (Sovereign AI) 战略崛起 │ 对立 │• 制定全球统一的安全与互操作标准 │ │• 美资企业撤退 (如微软缩减在华研究)│ 与撕裂 │• 跨国流动与算力/数据治理协调 │ └─────────────────────────────────┘ └─────────────────────────────────┘ ``` 面对这一现状，Sanjay Puri 提出了一个尖锐的矛盾：美国国会参众两院在跨党派层面已形成高度共识，坚信“赢得 AI 竞赛的国家将主导未来秩序”，导致主流美资科技巨头（包括微软等）纷纷撤回收紧在华研究中心；在此背景下，呼吁中美合作究竟该如何落地？ Thorsten Jelinek 对此给出了清醒的反思：当前围绕“AI 竞赛”的宏观叙事显得极其灰暗与单一。人们盲目追求所谓的领先，却未曾厘清这场竞赛的终点究竟是什么——是将 AGI（通用人工智能）或超级智能（Superintelligence）作为技术终局，还是真正利用它造福人类？在缺乏明确终极目标的情况下，各国盲目开展零和博弈并为此不惜牺牲一切伦理底线，正在将全球 AI 治理推向更加分裂的险境。 ### [33:20](https://www.youtube.com/watch?v=1bVutEOc58Q&t=2000s) 投资规模激增、国家能力差异与中美协作的可能性 Thorsten Jelinek 指出，当下的叙事已经深入人心，因为各方赌注极其高昂。与过去电信基础设施建设相比，如今 AI 领域的投资规模高出了 10 到 100 倍。过去电信建设因成本过高必须依赖西方政府的大规模国家干预与资金投入；但现在的根本转变在于，美国完全依赖私营部门投入，美国政府自身并没有本土 AI 研发与算力能力，必须借助于 Meta 等科技巨头。相反，中国是唯一一个政府自身就具备强大 AI 能力的国家。欧洲虽然在底层大模型上缺乏足够的自主能力，但其潜在机会和优势在于垂直工业领域的深度结合。 谈及跨国协作，Jelinek 提到了 Geoffrey Hinton 最近对中国的访问。Hinton 在中国的一场会议上提出，AI 分为“善良的 AI（kind AI）”和“聪明的 AI（smart AI）”，全球应该在“善良”上展开合作，而在“聪明/智力”上则属于竞争。Jelinek 借用过去冷战时期的《开放天空条约》（Open Skies Treaty）进行类比：虽然该防务条约因种种现实原因已不复存在，但其核心机制是当卫星飞越或采取重大行动时主动通报对方。未来的前沿大模型也可以采取类似机制——当一方训练出超大规模前沿模型时，主动知会对方其实力与边界，避免对方在突然目睹技术飞跃时陷入恐慌。 此外，不仅 Hinton 指出恐怖主义等跨国共同威胁需要全球协作，就连此前对华态度颇为强硬的 Eric Schmidt，如今也展现出更加开放的姿态，意识到美中合作的必要性。Jelinek 强调，AI 越来越像一种类似能源的公共品（Public Good），这自然会牵涉到开源机制。尽管中国在开源模型上有通过开源吸引更多用户、加速技术追赶的现实考量，且大模型的开源与传统软件开源存在很大差异，但国际社会绝不能放弃对话，必须持续寻找共识、建立机制以逐步缩小分歧。 ```mermaid flowchart TD accTitle: 全球主要经济体 AI 能力模式与协作设想 subgraph CAP["国家/地区能力架构差异"] US["美国模式<br/>（私营企业驱动，政府无自主能力，依赖科技巨头）"] CN["中国模式<br/>（国家政府自身拥有 AI 能力，积极推进开源追赶）"] EU["欧洲模式<br/>（底层通用算力较弱，强项在工业垂直应用）"] end subgraph COOP["国际协作设想：开放天空机制"] NOTIF["前沿模型突破信息透明互通<br/>（通报能力边界，避免误判与恐慌）"] COMMON["共同底线风险协同治理<br/>（反恐、公共安全、AI 善意/公共品属性）"] end CAP --> COOP ``` ### [36:17](https://www.youtube.com/watch?v=1bVutEOc58Q&t=2177s) 全球南方（Global South）的困境与两极分化站队 主持人 Sanjay Puri 提出了一个尖锐的国际地缘政治问题：目前的讨论大多围绕美国、欧盟和中国三大核心框架展开，但包括印度、印度尼西亚、菲律宾在内的广大亚洲地区，以及非洲和拉丁美洲等“全球南方”国家该何去何从？这些国家自身的治理架构是否最终会被吞并，被迫在美国模式、中国模式或欧盟模式之间做出三选一的抉择？ 针对这一疑问，Thorsten Jelinek 结合自己在巴西圣保罗和里约热内卢参加金砖国家（BRICS）会议的亲身经历表示，在技术和制度层面，多方寻求协作与对接的趋势已经发生。要避免彻底分裂，不仅需要技术上的互操作性（Interoperability），更需要“治理上的互操作性”。人类意识往往容易聚焦于差异，但仔细比对各国目前的监管文本，其实存在大量底层共性，必须主动构建对接机制。如果不加干预，全球治理很可能会重蹈上个世纪的覆辙，再次演变为由技术垄断者主导规则、并自上而下向外层层倾轧的单极或霸权格局。 Justin Hendrix 随后补充了全球南方面临的现实拉扯。印度即将主办下一届全球 AI 峰会（目前官方更倾向于称其为“机遇会议”），其核心议程预计将聚焦于技术公平性（Equity）以及如何将 AI 成果普及至全球绝大多数人口（Global Majority）。同时，非盟及拉美各国也在加速构思自己的方案。然而现实是，全球正深陷中美两极格局的张力之中： - **美国方面**：正在极力要求盟友签署其主导的愿景，但这一愿景如今已不再包装为“更民主、公平与可持续的未来”，而是赤裸裸地转变为“毫无争议、不可动摇的美国主导权（Unquestioned American dominance）”。 - **中国方面**：正通过金砖国家、联合国等多边机制推广另一套更具交易性质（Transactional）的合作方案。 - **盟友与第三国境况**：许多国家正处于进退维谷（Between a rock and a hard place）的艰难境地，在两大阵营的技术标准与政治诉求夹缝中求生存。 ### [40:36](https://www.youtube.com/watch?v=1bVutEOc58Q&t=2436s) 观众互动：跨越分歧的“AI 希波克拉底誓言”与人性的最大公约数 在观众提问环节，救助儿童会（Save the Children）首席技术官兼南加州大学（USC）AI 治理博士候选人 Ryan Gary 提问。他提到自己一个月前刚走访了中国上海等地，目前正在研究 AI 领域的“希波克拉底誓言”（即 AI 如何做到“安全为善/不作恶”）。他直言，学界和产业界已经写了太多关于“数字鸿沟”的陈词滥调，他迫切想了解：究竟是什么阻碍了中美等国放下分歧、共同利用 AI 去攻克癌症等全人类难题？如何才能提炼出行之有效的实践经验，促成真正的全球跨阵营合作？ Ryan Wang 针对这一问题做出了核心回应。他认为实现“AI 向善（AI for Good）”的关键在于回归最基础的底层逻辑： 1. **以人性为最根本的公约数**：不同文化与价值体系对“善”的定义各不相同，但全人类在基本人性上面是相通的。当世界上还有大量人口面临饥饿和贫困时，将 AI 用于解决这些基本生存危机就是无可争议的善。 2. **多维度的开源推动进步**：在开源世界中，权重（Weights）、源代码、训练数据可以分层次、多层级地开放。这种科研层面的共享是推动整个人类社会与全球经济向前迈进的核心动力。 3. **宏观经济视角**：很多现实世界的问题，往往需要依托经济跃升到更高阶段才能彻底化解。对于现场经常热议的“AI 是否存在泡沫”以及“泡沫破裂后会怎样”，Ryan 坦言，如果当前的 AI 泡沫破裂，全球经济在短期内几乎找不到下一个能承担引擎重任的驱动力，因此全行业必须努力维持技术与产业的发展势头，避免其过早破灭。 随后主持人准备在对谈正式结束前对嘉宾提出最后一轮挑战性问题，现场节奏紧凑进入收尾阶段。 ### [43:53](https://www.youtube.com/watch?v=1bVutEOc58Q&t=2633s) 最终胜出者：全球 AI 治理模式的博弈与终极诉求 在圆桌研讨的尾声，主持人 Sanjay Puri 向全场嘉宾提出了最后一个核心问题：在欧盟（EU）、中国（China）和美国（US）这三大主流治理路径中，究竟哪一种治理框架最终会在全球竞争中脱颖而出、赢得胜利？在提问的同时，Sanjay 还不忘幽默地提醒台上的嘉宾：“注意，你们说的可都被录制下来了，别忘了这一点。” 面对这道兼具地缘政治与科技伦理的难题，嘉宾们展现了极具智慧与幽默感的回答。一位嘉宾率先调侃道，能够胜出的治理模式必然是“具备慈母般关怀的 AI（motherly AI）或成为道德伙伴”的模型；另一位嘉宾紧接着表态：“最终是人类获胜（Humanity wins）。”面对如此充满情怀的回答，Sanjay 忍不住打趣道：“Ryan，你真应该去从政！我们可不是在竞逐奥斯卡最佳影片啊，各位。” 随后，嘉宾给出了更具深度的总结性阐述：能够最终胜出的治理模式，必然是那个能够向世界更多国家和地区证明自身确实“能够满足全人类根本利益”的体系。然而，要真正检验并证明这一点将耗费漫长的时间，在当下的技术与政策博弈阶段，尚无法给出一个非黑即白的明确答案。 随着论坛时间耗尽，Sanjay 笑着调侃大家听起来像是“选美大赛（Miss Universe）看多了”，随后正式向 Thorsten Jelinek、Justin Hendrix 和 Ryan Wang 表达了感谢，致谢他们带来的精彩洞见与深度思辨，全场在掌声与致意中圆满落幕。 - **核心议题**：美、欧、中三大 AI 监管体系谁将成为全球标准制定者与最终赢家。 - **嘉宾共识**：治理模式的最终胜利不取决于单纯的地缘竞争，而在于能否在全球范围内被证明真正符合人类福祉与利益（satisfy the interests of humanity），这一验证过程将是一个长期且复杂的演进过程。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=1bVutEOc58Q&t=4s) 在 AI 治理领域，究竟是哪一个问题让你在夜里辗转难眠？ — Sanjay Puri > [00:04](https://www.youtube.com/watch?v=1bVutEOc58Q&t=4s) 真正令我夜不能寐的是全球范围内民主治理模式长达数十年的衰退，且这种衰退在包括美国在内的国家正呈现加速趋势。技术究竟是在为构建一个更加公正、公平、可持续的民主未来提供支撑，还是在暗中破坏它？ — Justin Hendrix > [15:33](https://www.youtube.com/watch?v=1bVutEOc58Q&t=933s) 毫无新思想，只剩庞大资本在支撑 — Justin Hendrix > [22:48](https://www.youtube.com/watch?v=1bVutEOc58Q&t=1368s) 我们不应该一提到中国就指着鼻子说对方是坏人。我想改变这种叙事——我们或许是对立面、是竞争者，但绝非真正的敌人。人类真正的敌人是失控的 AI 技术对文明造成的生存威胁。我们应当聚焦于这一全人类的宏观命题，保持对话与合作的开放渠道。 — Ryan Wang > [26:16](https://www.youtube.com/watch?v=1bVutEOc58Q&t=1576s) 如果我向自然界排放了一种化学物质，它与其他化学品催化反应并产生了有毒害的恶果，你必定会追究我的责任。我们绝不能过度将智能体视作‘独立自主实体’，从而免除最初按下释放按钮的个人、企业或组织的法律与道德责任。 — Justin Hendrix > [30:01](https://www.youtube.com/watch?v=1bVutEOc58Q&t=1801s) 赢得 AI 竞赛的国家将主导未来秩序 — Sanjay Puri > [33:20](https://www.youtube.com/watch?v=1bVutEOc58Q&t=2000s) AI 越来越像一种类似能源的公共品 — Thorsten Jelinek > [36:17](https://www.youtube.com/watch?v=1bVutEOc58Q&t=2177s) 要避免彻底分裂，不仅需要技术上的互操作性，更需要‘治理上的互操作性’。 — Thorsten Jelinek > [40:36](https://www.youtube.com/watch?v=1bVutEOc58Q&t=2436s) 如果当前的 AI 泡沫破裂，全球经济在短期内几乎找不到下一个能承担引擎重任的驱动力，因此全行业必须努力维持技术与产业的发展势头，避免其过早破灭。 — Ryan Wang > [43:53](https://www.youtube.com/watch?v=1bVutEOc58Q&t=2633s) 最终是人类获胜 — Ryan Wang ## Mentioned Sanjay Puri, RegulatingAI, Thorsten Jelinek, Hertie School, Ryan Wang, Outpost Capital, Solaris Venture Partners, Justin Hendrix, Tech Policy Press, The Economist, New York University, Cornell Tech, Figure AI, DeepSeek, OpenAI, Alibaba Qwen, Tencent, Moonshot AI, 李开复, 01.AI, Meta, Mark Zuckerberg, International Telecommunication Union, Manus, Stargate, SoftBank, Oracle, Nvidia, Scale AI, EU AI Act, Geoffrey Hinton, Eric Schmidt, Open Skies Treaty, BRICS, Save the Children, USC, Ryan Gary ## Video description Thorsten Jelinek, Senior Fellow, Hertie School, Centre for Digital Governance Ryan Wang, Founding Partner, Outpost Capital Sanjay Puri, Chairman/Founder, RegulatingAI Justin Hendrix, CEO and Editor, Tech Policy Press Abstracts How are the world’s leading economies shaping the future of artificial intelligence—and what can innovators learn from their divergent paths? This cross-border conversation will examine the EU’s risk-based AI Act, the United States’ sector-driven, innovation-first approach, China’s rapid but tightly managed rollout of national standards, and emerging frameworks from regions such as the U.K., Canada, and Singapore. Panelists will unpack how each regime balances safety, privacy, and competitiveness, highlighting the real-world impact on funding flows, talent migration, and time-to-market for new AI products. We’ll explore where harmonization is possible, where regulatory fragmentation creates friction, and how multinational companies are adapting compliance strategies to stay ahead. Attendees will leave with a clearer picture of how today’s policy choices are reshaping tomorrow’s global AI landscape—and actionable insights for steering their own organizations through a patchwork of rules. [Source](https://www.youtube.com/watch?v=1bVutEOc58Q)

[0:04] Sanjay Puri: 在 AI 治理领域，究竟是哪一个问题让你在夜里辗转难眠？ [0:04] Justin Hendrix: 真正令我夜不能寐的是全球范围内民主治理模式长达数十年的衰退，且这种衰退在包括美国在内的国家正呈现加速趋势。技术究竟是在为构建一个更加公正、公平、可持续的民主未来提供支撑，还是在暗中破坏它？ [15:33] Justin Hendrix: 毫无新思想，只剩庞大资本在支撑 [22:48] Ryan Wang: 我们不应该一提到中国就指着鼻子说对方是坏人。我想改变这种叙事——我们或许是对立面、是竞争者，但绝非真正的敌人。人类真正的敌人是失控的 AI 技术对文明造成的生存威胁。我们应当聚焦于这一全人类的宏观命题，保持对话与合作的开放渠道。 [26:16] Justin Hendrix: 如果我向自然界排放了一种化学物质，它与其他化学品催化反应并产生了有毒害的恶果，你必定会追究我的责任。我们绝不能过度将智能体视作‘独立自主实体’，从而免除最初按下释放按钮的个人、企业或组织的法律与道德责任。 [30:01] Sanjay Puri: 赢得 AI 竞赛的国家将主导未来秩序 [33:20] Thorsten Jelinek: AI 越来越像一种类似能源的公共品 [36:17] Thorsten Jelinek: 要避免彻底分裂，不仅需要技术上的互操作性，更需要‘治理上的互操作性’。 [40:36] Ryan Wang: 如果当前的 AI 泡沫破裂，全球经济在短期内几乎找不到下一个能承担引擎重任的驱动力，因此全行业必须努力维持技术与产业的发展势头，避免其过早破灭。 [43:53] Ryan Wang: 最终是人类获胜

[26:50] Justin Hendrix: If I released a chemical into the world and it catalyzed with other chemicals and created a bad output, right, you'd want to hold me responsible. I hope we don't go too far in regarding agents as autonomous somehow that the blame for releasing that catalytic agent into the world doesn't rest with the person who took the decision to do so, or the company, or the organization. [15:33] Justin Hendrix: unquestioned and unchallenged technological dominance [32:57] Geoffrey Hinton: Collaborate on Kindness [36:46] Thorsten Jelinek: governance interoperability [38:29] Justin Hendrix: unquestioned, unchallenged American dominance [38:29] Justin Hendrix: between a rock and a hard place [43:45] Thorsten Jelinek: the one who has motherly AI or is a moral partner [43:45] Ryan Wang: Humanity wins. [43:45] Sanjay Puri: watching Miss Universe for a little while, too long [43:45] Sanjay Puri: we are not priming for the Oscars here

[0:04] Okay, [0:06] thank you all for coming. Uh [0:11] we are dealing with a very complex [0:14] international [0:16] governance regulatory environment and [0:19] some of you or a lot of you work for [0:21] companies that are global in nature and [0:24] are really trying to figure out you know [0:26] where should we go how where should we [0:28] invest where should we sell etc. Well, [0:31] the good news for you is that we've got [0:32] a amazing panel who bring really many [0:36] different perspectives and we're going [0:38] to try to not just talk about this [0:42] topics but hopefully leave you with or [0:44] they they are going to leave you with [0:46] something that'll be constructive and [0:49] tangible that maybe you can take away [0:51] from there. And if they don't then [0:53] please grab them at the end of the panel [0:56] and make sure they do. But uh the way we [1:00] are going to do it is I'll ask uh each [1:02] of them to introduce themselves but also [1:04] answer one uh specific question is what [1:09] uh issue in governance keeps each of you [1:12] up at night. [1:14] So we'll start with you Ten. [1:17] >> Yes. Thank you very much. Uh it's my [1:19] third time here and uh it's always a [1:22] great energy and I'm looking forward to [1:24] go to this uh showroom and uh I'm from [1:27] Berlin, Europe and last year it was like [1:29] that uh when Europe was on holiday on [1:31] August which is usually the time here [1:33] was so much energy but January was a [1:36] wakeup moment in Europe and I think uh [1:39] there's hope. Um anyway uh what wakes me [1:42] up of course that's the the geopolitical [1:45] context. I'm very much uh working in [1:48] that context on uh governance issues, [1:50] digital sovereignty and uh [1:55] and um I think the the way to find an [1:59] access point of collaboration I think [2:01] that's crucial. It would maybe not [2:04] >> about yourself too about myself. Yeah. [2:06] So I've been uh I have been in this AI [2:08] governance uh arena since a while. In [2:11] fact actually we had uh that was in [2:13] China. So I focused on China and Europe [2:14] a little bit on the United States. uh we [2:16] run the first uh AI ethics workshops in [2:21] 2018-19 where some outside China thought [2:24] well China is not talking about AI [2:26] ethics and so uh we had a good time uh [2:28] I'm uh try to publish so I'm a [2:32] researcher uh public policy researcher [2:34] but I'm also entrepreneur and uh focus [2:37] and uh I don't want to make my [2:39] advertisement for my session later I [2:41] look also into the issue of [2:43] consciousness AI and consciousness [2:46] Brian, introduction and uh what keeps [2:49] you up at night and where did you get [2:51] this cool hat? But anyway, let's uh get [2:53] into the introduction. [2:55] >> Thank you for noticing the cool hat. Um [2:58] and uh I I'm a big fan of uh you know [3:00] consciousness and AI, right? So we [3:02] should have a small discussion after. Um [3:05] so my name is Ryan Wong. I'm uh a VC and [3:08] PE investor. So on the venture capital [3:11] side, I manage a uh a firm called [3:13] Outpost Capital where we do early stage [3:16] frontier technology investments, right? [3:18] So things like AI, robotics, spatial [3:21] computing, blockchain, those are my [3:23] sectors. And on the private equity side, [3:25] I manage a firm called Solaris Venture [3:28] Partners where we do growth stage [3:30] opportunities in uh data center, digital [3:33] infra and uh also AI robotics. Actually [3:35] I have one of my colleague from uh [3:37] Solaris David. He's sitting right there. [3:40] Um and um yeah you know uh my uh before [3:45] that I was investing bank in New York [3:47] City and then before that um I was born [3:50] raised in China right so spent my [3:52] childhood in China. So um my perspective [3:55] and I think you know when I look into [3:58] the uh uh the governance and uh the [4:01] regulation um ecosystem out there what's [4:04] keeping me at night I think it's very [4:06] similar to what uh Torson has just [4:08] mentioned I am so tired of this whole US [4:12] China competition right rivalry type of [4:15] narrative and I think uh you know I [4:17] would love to see more collaborations [4:19] happening and I think uh there are [4:21] clearly a lot of misunderstandings [4:23] things on both sides of the table and [4:25] the collaboration is definitely help [4:27] very something very very helpful for us [4:29] to accelerate into a better version of [4:31] the future. So that's uh that's my take [4:34] on it. [4:34] >> Tired of this US China stuff? [4:38] We'll turn it over to you and uh [4:41] introductions and what's keeping you up [4:44] uh at night besides this whole US China [4:46] stuff. Something new. [4:47] >> Yes, sir. Uh so I'm Justin Hendrickx. [4:49] I'm the CEO and editor of Tech Policy [4:52] Press. Uh we are uh a website, a media [4:55] organization that focuses pretty much [4:57] what it says on the tin. We write about [4:58] tech policy matters uh both producing [5:01] our own and reporting and analysis but [5:04] also uh with expert contributors from [5:07] across the world. Um I started my career [5:09] at the economist. I was there for about [5:11] a dozen years. I ended up running a [5:13] emerging media technology lab out of NYU [5:15] for about a decade. I still teach at NYU [5:18] and at Cornell Tech. Um and in the [5:21] course of doing that I started teaching [5:23] around uh tech media democracy issues [5:26] and became you know uh very concerned [5:29] about the relationship between [5:30] technology uh and democracy uh in the [5:34] world. Democracy is a form of [5:36] governance. The thing that keeps me up [5:37] at night is the uh now multi-deade [5:40] decade decline of democratic uh form of [5:43] governance in the world. Um and that [5:46] decline appears to be hastening [5:47] including here in the United States. Um [5:50] and so the question I sort of try to [5:52] answer on uh most days is um how are we [5:56] going to build a more just equitable [5:58] sustainable democratic future? What role [6:01] does technology play in potentially you [6:05] know affording that future and what role [6:07] does it play in potentially sabotaging [6:09] that future? Um so when I think about [6:12] governance issues I'm looking at it [6:13] through that frame. I'm a NYU graduate [6:16] by the way. So always good to see alumni [6:19] bleed violet. [6:21] >> All right. So you've seen some different [6:24] perspectives and we'll try to kind of [6:26] dig into that. Thirsten you worked a lot [6:28] in the EU and the China space. So when [6:32] you compare those two from a you know [6:36] look at our audience if they've got [6:38] companies uh or they're looking at [6:40] companies etc. What are the few [6:43] differences when you look at the EU and [6:46] the China regulatory environment? [6:49] >> I we like to look at differences uh but [6:51] sometimes those differences are uh put [6:54] or narrative. So uh there's one when we [6:58] talk about AI governance is the legal [6:59] side. So there's this notion of uh [7:02] abiding to the rule of law and uh for [7:05] some some abide to that for some others [7:08] other countries many other countries or [7:10] they don't do right but I think it's uh [7:13] maybe the wrong way to put it uh I think [7:16] there are differences but uh to maybe [7:19] depoliticize those differences if you [7:21] look at the legal transitions in the [7:23] United States Europe and China you see [7:26] uh the US has a common law Europe has a [7:29] civil law and uh and China has a hybrid [7:32] civil law and let's say with state [7:34] collectivist uh goals and uh and they [7:38] have all lawyers and they have they [7:40] practicing law uh but if you look at [7:43] those those cultural differences then [7:44] you will also understand why the United [7:46] States has not uh let's say NI [7:49] comprehensive AI law uh is not only that [7:53] you have a very competing parties yeah [7:55] two parties but also that this kind of [7:58] case-based law kind of prevents that in [8:01] a way, right? And when you have a a [8:03] civil law, then everything is encoded in [8:06] the law and the judge has to interpret [8:08] the law and here is case by case. So new [8:11] legal changes happen by case. Um that's [8:14] one. So I think um um [8:19] so that's that's those differences and [8:21] what comes over it now is the push for [8:24] sovereignty. I think sovereignty is [8:27] needed to let's say rebuild the [8:30] democracy, rebuild the economy and and [8:33] we had decades where this sovereignty [8:36] was compromised and maybe for not all [8:39] the best ends. Uh the problem is with [8:41] this push for sovereignty, you uh you [8:44] you have the legitimate right to to [8:48] secure AI or secure technology uh to be [8:51] sovereign over your technology, but it [8:53] has pushed the boundaries uh between [8:56] security and protectionism and I think [8:58] that's a great a great issue we are [9:01] facing. [9:02] >> So Ryan um let's say we have some of our [9:05] folks in the audience who are trying to [9:07] build companies. when you look at the [9:09] investment you know as an investor which [9:14] environment regulatory environment is [9:16] from your standpoint the best [9:18] environment to invest in whether it's [9:20] Europe US or China [9:24] >> right I think this is a great question [9:26] right and um uh I have a limited [9:29] understanding about Europe so I'm going [9:30] to lean to you for those type of [9:33] expertise but uh between US and Asia [9:36] right and I think uh you China is a one [9:39] of the dominant force in Asia but uh the [9:41] the other markets like um you know [9:44] Southeast Asia, Japan, Korea, Singapore [9:47] right those are also very very [9:49] interesting market as well and I think [9:51] um you know in general what's happened [9:54] from an investment perspective when we [9:56] look into the uh landscape uh US versus [10:00] Asia right so some of the primary [10:03] differences right one I think um um the [10:07] valuation of the US companies is in [10:10] generally much much higher than um than [10:14] Asia- based companies right so you know [10:16] most recent round of figure AI is valued [10:19] close to 40 billion right when the [10:22] company is still um still a pre-product [10:24] company right they haven't really [10:27] officially launched their product um so [10:29] things like this is unimaginable in AI [10:33] in in in Asia right and also at the same [10:35] time I think the the talent pool is also [10:38] quite different, right? So, a lot of [10:41] people earlier this year, right, in [10:43] January when DeepS first dropped the uh [10:45] the R3, right, a lot of people are uh [10:48] are surprised that oh, you know, why [10:50] this open-source AI model is coming out [10:53] of China and I think, you know, to me [10:55] it's not a surprise to me, right? Why? [10:57] Because in the open-source word [11:00] development space um mo most of the new [11:05] R&Ds in that space are driven by [11:07] researchers, PhDs are driven by the [11:10] academia right and uh most of the papers [11:13] are public right so these are not secret [11:16] or you know uh so anyone in the world [11:19] have access to all this information and [11:22] when we look into the talent pool um [11:24] North American US to be specific [11:28] versus um versus uh Asia in particularly [11:31] China right we see we see how you know [11:34] how many new PhD graduates from China [11:37] every year and we can compare the [11:38] numbers right it's a huge huge gap right [11:42] and when you can you know think in this [11:44] kind of perspective naturally you can [11:46] understand oh you know why in in the [11:49] closed source world open AI is a leader [11:51] right you know they they are the they [11:53] are the state-of-the-art but why in the [11:56] open source board right why you know [11:58] deep sake um the chen by Alibaba [12:02] by tenscent right all these model are [12:04] performing so so well and how they can [12:07] you know use their you know advanced [12:09] research and uh like you know just army [12:11] of PhDs to um to optimize how the system [12:16] function to the extreme right so I think [12:19] you know there are opportunity all [12:21] across the board um more money for your [12:24] for your buck I think you know uh Asia [12:26] IA companies, Chinese companies are very [12:28] lower while US companies more expensive. [12:31] But in in term of exit strategies, [12:33] right? Like you know it's easier to [12:36] access um to exit as a US-based company, [12:38] right? Because you know you can uh US [12:41] stock market, US financial markets is [12:42] more mature. You have investors from all [12:45] over the world, right, for exit [12:47] opportunities. But in China the capital [12:49] situation uh a lot of USD fund has been [12:52] pulling away from China starting uh 3 [12:56] four years ago right? So you know [12:58] international players are leaving this [13:00] playground and also you know for a lot [13:02] of um uh domestic investors is mostly [13:05] backed by certain level backed by the [13:07] Chinese government for the domestic [13:10] investors. So for exit opportunities not [13:12] so promising. So it's different and I [13:15] think opportunities and um and uh and [13:18] the struggles happening in everywhere um [13:20] when we compare these two markets. But [13:23] the bottom line is if these guys want a [13:25] higher valuation it's good old USA. Um [13:28] but just very briefly Ryan why has China [13:32] and this is just a question maybe thirst [13:34] too gone so big into opensource is there [13:38] a strategic reason uh or is this a [13:41] company uh reason because when you look [13:43] at deepseek you look at Quen from [13:45] Alibaba and there's whole range of [13:47] others we only talk about them I mean [13:50] here we have you know Llama and there [13:52] are several others in France as Mistl [13:55] but why has China gone so far ahead [13:59] strategically in open source, [14:01] >> right? I I think that's a really good [14:03] question and uh I don't have a very [14:06] confident answer on this one. Uh my [14:08] interpretation is that um you know China [14:12] uh a lot of um you know foundation model [14:14] companies in China right like Moonshot [14:17] um like the one started by Kyu Lee right [14:19] the former uh CEO for Microsoft in China [14:22] as well um so they tried closed source [14:26] right um they've they tried to make it [14:29] work but I think uh maybe because of the [14:32] the best researchers the top 00 [14:37] top 001% [14:39] um type of researchers and uh you know [14:42] developers they are still very much um [14:45] based in the US and in particular based [14:47] in Silicon Valley right that's why you [14:50] know in this open-source to closed [14:53] source competition the open source [14:56] happened to be the winner in China right [14:59] versus you know um in the in the US the [15:02] playbook is different right um Meta just [15:05] spent like you know hundreds of million [15:07] per person to hire their super [15:09] intelligence team, right? And in in this [15:12] uh this kind of scenario is unimaginable [15:15] for companies in China, right? So if you [15:18] want to attract the top 00001% [15:22] uh expert, that's a cost and uh in China [15:25] most of the companies can afford this [15:27] kind of a cost and uh I think that's [15:29] kind of like my superficial assumption [15:31] on what's happening. [15:33] >> Uh Justin is feeling a little lonely [15:35] here. I'm going to come to you Tristan. [15:37] Uh Justin, you look at the US uh you [15:42] know regulatory environment, governance [15:44] uh environment in the US now we have a [15:48] innovation first pro- innovation kind of [15:50] an environment but we have a state level [15:52] kind of patches. How do you compare that [15:55] with a centralized regulatory system in [15:57] the EU and more so also in China? Yeah, [16:00] I think I'll come to the state, but I [16:02] also want to um just maybe comment on [16:04] the positionality of the US with regard [16:07] to Europe and China as well. Um and you [16:10] know, you said we have an innovation [16:11] first agenda. Uh but the stated agenda [16:15] of the president at the moment is [16:16] unquestioned and unchallenged [16:18] technological dominance. [16:20] >> Um that is the phrase right in the plan [16:23] to win the AI race. Um so you know uh [16:27] this question over whether we should [16:29] regard uh the the AI race uh you know [16:33] necessarily as as simply a kind of [16:34] binary competition uh perhaps between [16:37] the US and China. We could have a whole [16:39] of the panel on that. Um but if we just [16:42] take that phrase winning the AI race and [16:44] we think about um if we were to cast our [16:47] minds forward and in the eventuality [16:49] that we would look back and say we lost [16:52] the United States lost the AI race. What [16:54] would be the reasons? Um, I think we [16:57] could probably see the beginnings, the [17:00] germs of some of those. Now, you talk [17:02] about digital sovereignty and the push [17:04] for digital sovereignty. That's not [17:06] coming from nowhere, right? Um, there's [17:08] a real fear of unquestioned and [17:11] unchallenged technological dominance and [17:13] what that might mean for the politics of [17:15] other countries. uh there is you know a [17:19] real pull from China which is got this [17:22] sort of open approach and is talking in [17:25] the language of multilateralism and [17:27] international cooperation and trying to [17:29] make friends and allies uh etc. Um, and [17:33] then there's the extraordinary excess of [17:36] our market, right? It's great. There's a [17:38] lot of capital slloshing about, right? [17:40] But then you end up with these vapid [17:41] statements about super intelligence from [17:43] people like Mark Zuckerberg. Um, if if [17:46] you haven't read his most recent little [17:48] mini manifesto, I encourage you to do [17:50] it. It appears he's entirely out of [17:52] ideas and all he has left is capital. [17:54] Uh, so, you know, I think there's a real [17:57] possibility we'll look back and say US, [17:59] yeah, it was it was a leader. uh and it [18:02] self-sabotaged. It shut down its [18:04] science, [18:06] right? Uh it it made extraordinary [18:08] mistakes in this heady moment uh where [18:11] it believed its its uh own Kool-Aid. It [18:14] drank its own Kool-Aid, right? Um so [18:17] there's an extraordinary advantage [18:19] potentially for China and its [18:21] positionality as a as a challenger. Uh I [18:24] don't endorse that will be a better [18:26] world necessarily. uh if in fact the US [18:29] quote unquote loses the AI race. Um [18:32] we'll see what happens in Europe. [18:33] There's a lot of public spending there [18:35] right now at the moment. A lot of [18:37] concern about sovereignty. Uh and I [18:39] think that we'll begin to see whether [18:42] the political forces in Europe uh that [18:44] want to in essentially cleave themselves [18:47] off from reliance on the United States [18:50] uh whether that eventually kind of [18:52] becomes you know the majority view. Um [18:56] but you can understand the motivation. [18:58] Here in the US we have very fractured [19:01] politics uh across the US states. [19:03] There's a a you know a clear split [19:06] between uh red and blue states ones that [19:09] want to concern themselves with some of [19:11] the kind of regulatory guard rails um [19:14] that many believe are necessary. Uh red [19:16] states which may align more perhaps with [19:19] the current federal government's uh uh [19:21] point of view. Um there is a push to [19:24] essentially uh sideline all state [19:26] regulation in the US. We saw that [19:28] recently with the federal moratorum uh [19:30] which was defeated in the Senate before [19:32] the big beautiful bill was passed on [19:34] July the 4th. Um but that idea is still [19:38] germinating and may yet come back. It [19:40] was included in the AI action plan uh a [19:43] kind of moratorium light. you know, this [19:44] idea that the federal government would [19:46] review uh the regulatory scenario in [19:48] every state before offering uh federal [19:51] funds or you know, federal support. Not [19:53] exactly clear what it means or if any of [19:55] that would be constitutional, but um you [19:58] know, it still remains to be seen how [20:00] things will play out at the state level. [20:01] I said a lot. I'll stop there. [20:04] So uh one of the you know obviously we [20:07] have you know entrepreneurs as well as [20:09] policy leaders and you can't pick up at [20:12] least anything in the tech world without [20:15] talking about our favorite subject AI [20:18] agents. So I want to kind of get some [20:21] perspective that do AI agents need some [20:26] kind of a different governance [20:27] framework? Uh I want to open it up from [20:31] all of you to really understand because [20:34] it is autonomous. [20:36] Um and how is is the EU or China or the [20:41] US thinking about it uh person from your [20:43] standpoint? [20:44] >> Um uh I think we always need to adjust [20:47] the governance uh because it's a [20:50] transitionary transformative [20:52] transformative uh uh technology. So you [20:54] cannot just say we don't we need it. So [20:57] we need always to adapt and uh the [20:59] agents uh just assuming that they that [21:02] they really will start collaborating [21:04] like on a mass level. Um basically it [21:08] repeats a current economic system and [21:10] then you have to make up your own mind [21:12] is that economic system is it producing [21:15] the benefits for society? Is it [21:17] producing uh a more just more equal not [21:20] within the country also across [21:21] countries? uh because when we had this [21:24] openness in the past it also just worked [21:26] because there was a there was a [21:28] structural hierarchy in the world in the [21:30] past century let's say take [21:32] telecommunication uh it was always [21:35] monopolized in the west and there was [21:37] just five countries uh who had the [21:38] monopul uh and uh there was no there was [21:42] no almost no innovation going on there [21:44] was no technology transfer there was not [21:46] an interest even by the international [21:48] organization like the ITU the [21:49] international communications union etc [21:51] so Once uh uh now we we we go into the [21:55] sovereignty because now we become more [21:57] open equal right with China's rise but [22:00] also other countries rise this [22:02] multiparity etc suddenly this openness [22:04] doesn't work anymore. So what was um [22:07] what was a a a common strategy becomes [22:10] now only my strategy. So and the [22:13] technology just sits on top of it. It [22:15] transforms, it changes, it changes [22:17] hierarchies. Uh it also liberalizes, [22:20] right? Technology is a liberating force. [22:23] Um uh but it can also repeat something [22:26] which we don't want to repeat. So it's [22:28] not just about to regulate the [22:30] technology having a new governance [22:32] structure for for agents, but to see how [22:36] would it repeat what we don't want to [22:38] have repeated. Right? That's my view. [22:41] >> Anybody else on AI agents? [22:44] >> Yeah. So I think you know one of the [22:46] thing about the AI agent is that [22:48] especially for um those type of agent [22:50] that involve chain of thoughts right [22:52] like you know products like manis right [22:55] it's a Chinese company building a uh [22:57] agent operator style type of um uh type [23:01] of agent based solution that can handle [23:03] a lot of tasks um right so for for agent [23:07] application to run well it require a lot [23:09] of um computation power especially the [23:12] uh what we call as a inference side of [23:14] the computation, right? And um uh with a [23:18] with a demand for the inference side of [23:20] computation rise up, there's a huge huge [23:23] demand for the infrastructure side, [23:25] right? We need to have enough um uh [23:28] infrastructure to cover the potential um [23:32] increase on the on on the demand from [23:35] the application layer. So I think you [23:37] know um when I look into what's [23:39] happening in US uh with what's happening [23:42] in in Asia in particular in China right [23:44] as Justin you mentioned the US uh uh [23:47] politic is very fragmented whereas in [23:50] China it's very centralized right I [23:52] think that create quite a difference on [23:55] the on the build out of the next [23:58] generation infrastructure as well right [24:00] whereas we see you know uh Stargate [24:03] amazing very ambitious plan Right. And [24:06] uh you know right now we are like more [24:08] than half a year since the announcement [24:10] and uh where are they? What's a what's a [24:13] progress look like? You know there are a [24:14] lot of challenges handling the [24:16] financing, handling the buildout and the [24:19] certain power dynamic between OpenAI, [24:21] Soft Bank and Oracle right whereas you [24:24] know uh things are h things that happen [24:26] in China despite all the sanctions of [24:28] you know Nvidia chips exporting to China [24:31] and all that stuff a lot of new [24:32] infrastructure are being built in in [24:35] light speed right a lot of things are [24:37] done uh very fast very cheap and very [24:40] reliable right China is very famous for [24:43] it infrastructure building and I think [24:45] you know part of the reason is that it [24:47] has a centralized governing system and [24:50] everything is top down right there's [24:52] after the order is made there's no [24:54] debate and let's just heads on and get [24:56] things done right so that's kind of like [24:58] the China style in a nutshell so I think [25:01] um it's uh it it's it's tricky to see [25:05] how things will evolve from here but I [25:07] think you know both China and US had the [25:10] has certain level their leadership right [25:12] and I think uh you So uh uh at the end [25:15] of the day um there are pros and cons on [25:19] each of the structure and each of the um [25:22] uh each of the system right and I think [25:25] uh you know it's uh it's why I keep [25:29] saying that hey we should try to change [25:32] the narrative here right especially from [25:35] the US side I don't want all of us to [25:37] kind of when we think about China we [25:39] just point to China and say you guys are [25:41] the bad guy right I I want to change at [25:43] least that narrative to say hey you know [25:46] we are not enemies and uh is we might be [25:50] like opponents we might be you know [25:52] competitors but we are not true enemies [25:54] here right the true anime is a bigger [25:56] threat of the the AI right how AI could [25:59] be a bigger threat to humanity that's [26:01] that's where the big picture is can we [26:03] focus on that and try to collaborate try [26:06] to have some conversations open right so [26:08] I think um yeah you know that's uh I [26:11] think from agent It also involve like a [26:14] lot of things in that space as well. [26:16] >> Justin just Oh, sorry. Go ahead. [26:17] >> I was going to pick up on your AI agents [26:19] question if if you [26:20] >> No, it's actually, you know, some of our [26:22] uh folks who are working in the US, [26:26] there are some states who have come up [26:28] with uh AI legislation when you look at, [26:31] you know, Connecticut, Colorado [26:33] especially being high up there. How [26:36] should these folks be looking at these [26:39] kinds of different state regulations as [26:41] their their companies are working or [26:44] they're working as small entrepreneurs, [26:46] big entrepreneurs. What is your [26:48] perspective on that? [26:50] >> Uh well, I mean certainly um you know uh [26:53] states to watch include you know [26:54] Colorado which is now kind of I think [26:56] having a bit of a political waiver on uh [26:59] its first in the nation quote unquote [27:01] comprehensive AI bill. Um, but other [27:03] states, California, um, uh, as you [27:06] mentioned, Connecticut, uh, I just [27:08] talked last week at the National [27:10] Conference of State Legislators, which I [27:12] think I see another face that was there [27:13] in the room at that event in Boston, um, [27:16] with a legislator who's, you know, [27:18] working on AI regulation there. There [27:20] are things happening in my home state of [27:21] New York. There are many others. Um, I'd [27:23] also point to Texas, uh, which just [27:25] passed, you know, what could be [27:27] regarded, uh, as a another model of of [27:30] state AI regulation. Um so there's a lot [27:32] going on there. I do think at the moment [27:35] um you know that it it's complicated. Uh [27:38] I I don't uh buy the argument that we [27:41] are in such a sort of profound uh [27:44] confusing patchwork that uh you know [27:46] companies or or entrepreneurs can't make [27:48] sense of it. Most of the things uh that [27:50] are being regulated um are are you know [27:53] fairly straightforward. We'll see what [27:54] happens. Um just to pick up really quick [27:57] one thing on the AI agents question. Um [28:00] I I think it's going to be really [28:01] interesting to see what happens in the [28:03] next couple of years as more and more [28:06] including the people in this room uh we [28:08] we daisy chain more AI systems together [28:11] and we know we set these agents out to [28:14] kind of negotiate and uh their legions [28:16] amongst one another and various forms of [28:19] commerce or or you know what other types [28:20] of interactions happen. Um, but you [28:23] know, if I released a chemical into the [28:26] world and it catalyzed with other [28:28] chemicals and created a bad output, [28:30] right, you'd want to hold me [28:32] responsible. Um, so I hope we don't go [28:34] too far in regarding agents as [28:37] autonomous uh somehow that the the the [28:39] the blame uh for releasing that [28:42] catalytic agent into the world uh [28:44] doesn't rest with the person who took [28:46] the decision to do so uh or the company [28:49] or the organization. So that's one thing [28:50] that I would I would say about that. Um [28:52] but I think this is one area where um we [28:55] talked a little last night if you don't [28:57] mind me saying about standards uh at the [28:59] at the event that we were at um [29:02] standards interoperability extraordinary [29:05] excruciating a subject for many you know [29:07] the detailoriented process someone was [29:10] talking about how standards take years [29:11] to accomplish in many cases that type of [29:14] work is crucial especially in a world of [29:16] AI agents especially if we're afraid [29:18] that we're going to send these uh you [29:20] know catalytic semi-aututonomous [29:23] entities out to accomplish all manner of [29:25] things. Um we'll we need to deeply think [29:28] about those things and invest in uh in [29:31] standards and interoperability uh types [29:33] of concerns. And so um I would look for [29:36] uh legislation from the states uh [29:38] eventually. I think that there will be a [29:40] wave next year probably we're beginning [29:42] to see the beginnings of it now a wave [29:44] of AI chatbot legislation in the cha in [29:46] the states in including some of it [29:48] focused on applications around mental [29:50] health or concerns about child safety. [29:52] Um but I think you'll then eventually [29:54] begin to see uh legislation that tries [29:56] to get at some of these questions of of [29:59] commerce and agents. [30:01] So, um, you know, one of the things [30:05] that, uh, Justin, you mentioned is that [30:07] you're worried a little bit that AI is a [30:10] little bit of a bubble, so to speak. [30:13] Now, if let's say the bubble, well, in [30:17] this conference, we never wanted to [30:18] burst, but if that ever happened, what [30:21] would that mean for regulations, [30:23] governance, or collaboration between [30:26] countries? [30:28] So, this is a very good question and um [30:30] I'll leave it to smarter people to know [30:31] if uh it's a bubble, but right, but was [30:34] it a $ 1.5 billion offer for some person [30:37] to join Meta uh some aqua hire type uh [30:41] thing we you mentioned in your [30:43] presentation the $250 million were they [30:46] signing bonuses or whatever? I mean, [30:47] okay. [30:48] >> I mean, they paid 14 billion for scale [30:51] for 49% of it. So, [30:53] >> hey, maybe it's not a bubble, right? Um, [30:56] this conference went from 2,000 to 9,000 [30:58] attendees. Um, you know, there are [31:00] certain things that that look like this [31:02] and you can begin to ask yourself what's [31:04] going on here, right? Um, the amount of [31:06] capital deployed, the the amount of [31:08] enthusiasm. Um, we'll find out, you [31:10] know, eventually uh whether that bubble [31:12] bursts. Um, the question I think will be [31:15] when it bursts, what the policy makers [31:17] make of it. Um, some would argue that [31:19] the EU maybe uh got out a little bit [31:22] over its skis in terms of the uh AI act [31:25] partially in response to the degree of [31:28] enthusiasm um and the degree of concern [31:30] about artificial intelligence that maybe [31:32] some of the um breathless sweaty [31:35] meetings over AI governance and AI [31:37] safety over the last few years were [31:39] partly a result of uh a kind of response [31:42] to the the the bubble that has formed um [31:45] if it has formed. Um but on the other [31:48] hand, if it bursts, you know, will how [31:50] will policy makers react? Will they uh [31:53] you know, regard it as a moment to pull [31:54] back from concerns about AI governance? [31:56] Will that leave us in the lurch when [31:58] eventually uh the technology makes [32:01] another radical advance? [32:03] >> So Ten, you and Ryan have basically said [32:07] that we need to work closer. US and [32:10] China need to work closer. Um, I host a [32:14] podcast and I get a lot of members of [32:15] Congress and Senate that come in and [32:18] they basically consistently across the [32:21] board on bipartisan basis say the you [32:24] know the country that wins the AI race [32:26] is going to dominate whatever you want [32:29] to say it now and there is this whole [32:33] issue of sovereign AI. When we talk [32:35] about governance, we talk about [32:37] collaboration. How do you kind of uh [32:39] marry the two that whoever wins the AI [32:42] race is going to dominate? Well, we need [32:45] to collaborate. I mean, if you see most [32:47] US companies are kind of retreating, [32:50] they pull all the research when you look [32:52] at even Microsoft and others, [32:53] everybody's kind of pulled back from [32:55] China in many ways. [32:57] >> Yes. I mean, if you look at uh it's [32:59] quite quite gloomy. Uh the question is [33:02] what is the race? What's the goal? uh [33:05] and what are you willing to compromise [33:06] on you know to reach that goal and is [33:09] there an end goal in the AI because we [33:11] have this notion of AGI or super [33:13] intelligence is that the end goal or is [33:15] it then suddenly we use it for good etc. [33:17] So it's not clear uh it's it's a [33:20] narrative right so and everyone buys [33:22] into that and because also the stakes [33:24] are very high it's a lot of investment I [33:26] think is a 10 to 100 times more uh [33:29] compared when telecommunication was uh [33:31] put into place where you had also [33:32] massive state intervention because no [33:34] one can afford it and now we have a big [33:36] change that in the past it was [33:38] governments intervening in the west and [33:41] putting the money and now it's basically [33:43] the private sector in the United States [33:46] uh and China is uh China is the only [33:48] country where actually the government [33:49] has own AI capacity. [33:52] The US has no the US government has no [33:54] own AI capacity. It needs matter etc. [33:57] And in Europe I don't want to say has no [33:59] capacity at all but I think Europe's [34:01] strength here is vertically in the [34:02] industry. That might be the chance. So [34:05] on the collaborative side I would say uh [34:07] Jeffrey Hinton actually he visited China [34:10] just a few days ago right he went to the [34:12] the other conference and he said oh [34:14] there's kind AI there's smart AI and he [34:16] said we should collaborate on kindness [34:18] right and uh on the smart side I mean [34:20] that's a competition you don't want to [34:22] we don't want to I don't know whether [34:23] you can uh dissect a model in that way [34:26] but I I would make this analogy there [34:29] used to be the open skies treaty uh [34:32] which is is a defense treaty [34:34] Unfortunately, it doesn't exist anymore [34:36] for all the obvious reasons. But it [34:38] meant basically when I, you know, when [34:40] my satellites cross your country or when [34:42] I do something, I will inform you, [34:44] right? So what could happen is say if I [34:46] develop um you know this super duper [34:49] model whatever yeah uh uh I inform the [34:52] other side we have these great capacity [34:54] capabilities just be informed and don't [34:56] freak out when you see it. So that's one [34:58] way. uh Hinton is right that there's [35:02] certain issues uh like with some kind of [35:05] terrorism uh which is a common enemy but [35:08] I have to also say that the word [35:09] terrorism is also not so straightforward [35:12] these days uh but I think there are some [35:14] risks where there need to be [35:16] collaboration and I just realized that [35:18] even Eric Schmidt who is uh kind of [35:21] hawkish right towards China but even he [35:23] now is kind of uh more open and sees the [35:27] need to collaborate right And uh if AI [35:30] is something like energy, it's almost a [35:34] public good. So so you need to also open [35:37] up that comes maybe back on the open [35:38] source approach. But of course China uh [35:41] if I can just make one comment on that [35:43] which is usually done that China has no [35:45] choice than also to open source it uh to [35:48] catch up right because you get more [35:49] users. Uh but this open source is [35:51] different than in the old days software [35:53] opening up software. So uh anyway so on [35:56] the collaborative side um um I think [36:00] it's really this kind of what I said [36:03] like an open skies approach uh and uh [36:07] never stop trying to have this dialogue [36:09] always even if you say we have these [36:12] differences identify the differences and [36:14] maybe have a mechanism to reduce this uh [36:16] over time [36:17] >> we basically talked about three basic [36:20] frameworks you know the US EU and China [36:24] there's rest of the world that kind of [36:26] exists out there you know they call it [36:27] the global south or whatever it is [36:29] countries like India Indonesia [36:31] Philippines and that whole region are [36:34] their governance structures going to get [36:36] subsumed either you join the US [36:40] framework or you join the China [36:42] framework or you join the EU what's [36:44] going to happen I mean that's a huge I [36:46] mean Africa [36:48] South America Latin America where do [36:50] these uh frameworks go [36:52] >> I [36:52] If I can just uh uh so I had the [36:55] opportunity also to join one of those uh [36:58] bricks events which happened in Brazil. [37:00] So I was in Sa Paulo uh in uh Rio and uh [37:04] and um there's already on multiple [37:08] fronts from technology from institutions [37:11] the approach to collaborate so I think [37:14] it will happen uh in that way that you [37:17] try to you know to interface with [37:20] certain governance approaches you will [37:22] see that happening uh but in order to [37:25] keep it open you need to have not just a [37:28] technology interoperability but also [37:30] governance interpretability I mean and [37:32] that's the thing is uh you know our our [37:34] mind our consciousness works based on [37:36] significance on the difference but I [37:39] think we have so many commonalities if [37:41] you look at how AI is regulated for [37:43] example uh you find so many [37:46] commonalities uh despite the differences [37:49] uh and I think we need to find these [37:51] commonalities and find those interfaces [37:53] uh but otherwise I think we will see [37:55] unfortunately what we saw last century [37:58] uh that we will see institutions ions [37:59] which just support like say a monopoly [38:03] approach or whatever. So you it cascades [38:07] down unfortunately and the one who leads [38:09] it has the power to define the [38:11] institutions etc. So uh so it it it [38:15] permits all aspects. Uh so we have to be [38:19] um uh concerned or looking at all these [38:22] aspects and not just let it go but [38:25] always try to reach out again and try to [38:27] have this conversation. [38:29] Brian, anything on that Justin in terms [38:32] of the what's happening around the other [38:35] frameworks? [38:38] >> Well, I'll say uh you know India is [38:39] about to host the next AI summit. Um and [38:42] we'll see you know what uh the agenda is [38:44] uh for that event. I think it's in [38:46] February uh coming up. Um that will [38:49] they're not calling them summits now are [38:50] they? They're opportunity sessions or [38:53] whatever. Um but the uh the next one uh [38:56] you know whether the Indian government's [38:58] priorities which we know include more of [39:01] a focus on equity uh more of a focus on [39:04] um you know uh how to uh diffuse [39:08] technology across the global majority. [39:10] Um those types of concerns should rise [39:12] to the four uh in the agenda there I [39:15] assume and so that will be interesting [39:16] to to see uh you know what type of [39:18] thought comes out of that. Um and then [39:21] you know across the world whether it's [39:22] in Latin America or in Africa um you [39:25] know other governments other the African [39:26] Union they're making progress on [39:28] thinking about these questions. Um but [39:31] again it comes back to this you know [39:33] within this uh maybe two po polar domain [39:36] between China and the US. Um you know we [39:40] are asking uh c countries to sign on to [39:44] the US vision right um which is no [39:47] longer a stated vision of a more [39:50] democratic uh equitable and just [39:52] sustainable future uh but one simply of [39:55] unquestioned unchallenged American [39:57] dominance uh so you know come along with [40:00] us on that journey um where as opposed [40:03] to to China through bricks and through [40:05] the UN and through a various other uh [40:07] multilateral organizations is is uh [40:09] attempting to kind of sell a a different [40:12] kind of idea, more maybe you'd argue [40:15] more transactional transactional idea. [40:17] Um but uh our allies, I suspect our our [40:21] allies are between a rock and a hard [40:23] place. Um so we'll see what happens over [40:25] the next few years uh and whether [40:27] countries outside of uh those those [40:30] regions that you you described um what [40:33] they what they choose to do. We're [40:36] getting a little bit of a hook. Uh, but [40:39] we do want to open it up to some [40:41] audience questions. I've been bombarding [40:43] you. I have still a few ones left, but [40:45] let's see if we can have some [40:49] better questions than this guy's been [40:51] asking from the audience. And if they [40:53] don't, then I'll continue marching on [40:55] and we'll try to keep it short. Any [40:58] questions for them? But in the [40:59] meanwhile, I'll just Oh, please. [41:04] Sorry, my name is Ryan Gary. I'm the CTO [41:06] of Save the Children. I'm also a USC [41:09] doctorate candidate on governance and [41:11] AI. And I recently made a trip to [41:13] Shanghai, Shanghai and Chaning a month [41:16] ago. Um, met some friends strip ever. [41:19] Um, but my point though here is I'm [41:21] working on the hypocratic oath between [41:23] AI. Like how can we do it for good? And [41:26] I guess my question to all of you and [41:27] I'm going to send a LinkedIn to all of [41:28] you for a variety of research reasons. [41:30] like what do you think is the one thing [41:31] that's holding people back from really [41:33] seeing AI as this opportunity to kind of [41:37] come together with China solve cancer [41:39] find commonalities as opposed to the [41:41] differences and the reason I asked this [41:42] question is I don't want to do another [41:44] paper on the digital divide we heard [41:46] this a million times I want to find [41:48] practices that work so we could [41:49] replicate those and show that we can do [41:51] this right so I just want to get opinion [41:53] of anyone on the how can we get everyone [41:55] to find those commonalities so we can [41:57] actually move to a better place and I [41:59] say this as someone who was born in [42:01] Haiti, live in Chicago. I know there's [42:03] opportunity around and we could do this. [42:05] I just want to get your perspective. Um [42:07] and I love what um Mr. Hinton said, [42:09] Professor Hinton about you know the [42:11] mother AI I think is something that's [42:12] very important and your thoughts on [42:13] that. So I asked a lot of questions [42:14] there but how do we get this right in [42:17] your opinion? One quick answer because [42:21] running [42:22] >> quick answer I think is AI for good [42:26] right like you said we are driven by our [42:29] consciousness and our consciousness are [42:31] driven by stories right and then [42:34] different people different culture [42:36] different value system had different [42:38] definition on what is good but I think [42:41] the most fundamental unity of different [42:45] perspective is human nature right is [42:48] humanity right so we know that there are [42:51] still people living in hunger living in [42:53] poverty right if we can use AI to solve [42:56] those kind of problem is definitely for [42:58] good right and also you know um in the [43:01] open source world right you can open [43:03] weights you can open source code you can [43:05] open data right like different level on [43:08] how you can open different stuff and I [43:10] think you know from a research [43:12] perspective on further you can drive the [43:16] development and the advancement of the [43:18] AI sector. I think that's something good [43:21] for the for the humanity and for the [43:23] global economy as well. And um you know [43:25] a lot of real world problems will be [43:28] solved when the economy is getting to [43:30] the next stage. And uh you just talk [43:32] about whether we are in a bubble or not. [43:35] If the bubble burst what's going to [43:36] happen but I think you know if the AI [43:38] bubble burst I don't know what's the [43:40] next driver for the economy. So I [43:42] hopefully it's not bursting soon. U I'm [43:45] going to try to put them on the spot a [43:47] little bit before we end because that's [43:50] my job. So even though I think I'm way [43:53] way past uh final question to you [43:58] Tristan and Justin and Ryan in in the [44:02] end which governance framework is going [44:04] to win out of the EU, China, US [44:08] >> the one who has [44:09] >> you're being recorded here remember that [44:12] >> I repeat your question which [44:14] government's governance model will win [44:16] and it's very easy. the one who has [44:18] motherly AI or is a moral partner. [44:21] >> You got to be in politics, Ryan. Let's [44:23] see if you can hear it, man. [44:24] >> Humanity wins. [44:28] >> We are not priming for the Oscars here, [44:30] man. Just [44:32] >> I think the governance model will win. [44:35] um that has uh demonstrated [44:40] uh to more of the world um that it can [44:45] satisfy the interests of humanity. [44:48] >> Um and that will be uh something that [44:51] will take a long time to determine and I [44:53] think that unfortunately uh the answer [44:55] is not clear. Um [44:58] >> you guys have been watching Miss [44:59] Universe for a little while, too long, [45:01] but that's a whole different story. I [45:03] think I'm getting the hook. But uh thank [45:06] you guys for a fantastic panel. Really [45:09] really appreciate it. Great panelist. [45:11] >> Thank you all.