Chinese AI Startups Pivot to Domestic Silicon to Beat US Export Controls

应对美国出口管制，中国AI初创转向国产芯片 硬件整合与高性价比的后训练使中国开发者在克服严苛算力限制的同时，保持与西方前沿模型并跑。 主持人 Joe Weisenthal 和 Tracy Allaway 采访了研究员 Grace Shao，深入剖析了驱动中国人工智能生态系统的独特动力。讨论重点探讨了尽管面临美国芯片出口管制，中国 AI 实验室如何高度依赖开源策略、推理效率以及后训练优化来保持竞争力。Shao 阐述了自上而下的政府政策、密集的硬件供应链以及务实的企业整合，如何与硅谷对预训练规模和 AI 存亡论叙事的关注形成对比。此外，对话还涉及了务实的监管举措——例如杭州禁止因 AI 导致裁员——以及全球初创企业如何为了追求成本效益而日益采用中国的开源模型。 与西方AI领袖常从存在主义角度探讨AI不同，中国科技企业将AI视为务实工具，重点聚焦于商业应用与实体硬件（如电动汽车和机器人）的快速结合。 (2:40) 中国前沿AI实验室普遍采取开源路线以建立全球信誉并克服资本制约，DeepSeek-V4更是特意针对华为等国产芯片架构进行推理优化，为行业摆脱对英伟达等受限硬件的依赖提供了实践蓝图。 (9:21) 面对算力瓶颈与芯片出口限制，智谱AI、Minimax、月之暗面和DeepSeek等前沿实验室采取“做作业反推”策略，专注于后训练与推理效率，利用“智能蒸馏”等手段以极低成本保持在落后全球前沿6至9个月的窗口内。 (16:11) 得益于近几十年城市化积累的电网设施与“东数西算”战略，贵州、新疆等西部省份充沛的可再生能源能够承接算力负荷，有效缓解了中国AI基础设施的电力瓶颈。 (20:00) 杭州法院在一项里程碑式判决中明确禁止企业将AI替代作为解雇员工的合法理由，表明中国在推动“人工智能+”普及的同时，正通过法律与监管机制积极保护劳动者权益并防范数据安全风险。 (30:00)

Independent researcher Grace Hsiao argues that severe capital constraints, U.S. export controls, and domestic demographic pressures have steered China's AI ecosystem toward open-source architectures, post-training efficiency, and physical hardware integration rather than speculative frontier pre-training. Open-Source Models and Pragmatic Commercial Monetization (6:58) Chinese AI labs embraced open-source releases initially to build international credibility and overcome developer skepticism, a dynamic exemplified by DeepSeek founder Liang Wenfeng's push to share frontier research. While intense domestic competition remains, labs regularly integrate competitors' open-weight advances. Major independent frontier labs include Zhipu AI (GLM models for coding, listed in Hong Kong), Minimax (multimodal media, carrying an implied $20 billion valuation), Moonshot AI (the Kimi consumer interface), and DeepSeek. Compute Limits and Post-Training Homework Strategy (16:11) Because U.S. chip restrictions and data center buyouts cause acute compute shortages, Chinese labs focus resources on inference efficiency and post-training. Hsiao compares this approach to 'knowing what the answer to the homework is and working backwards,' letting Western firms shoulder speculative R&D costs before Chinese teams replicate frontier benchmarks within a 6-to-9-month lag at lower cost. Labs also acquire proprietary human execution datasets after exclusivity windows expire, dropping costs from $10 million to $1 million, while leveraging 'smart distillation' to use top models as evaluators and data labelers rather than brute-force copying. Infrastructure, Silicon Independence, and State Coordination (20:00) Chinese energy infrastructure supports AI compute via the national 'East Data, West Computing' framework, which routes processing workloads to western renewable energy hubs in Guizhou, Xinjiang, Inner Mongolia, and Sichuan to power eastern population centers. On the silicon side, DeepSeek delayed its V4 release by three to four months to engineer high-performance inference natively for Huawei's domestic silicon stack, demonstrating that labs could function without relying solely on Nvidia GPUs. Nationally, the Ministry of Industry and Information Technology and the Cyberspace Administration coordinate the 'AI Plus' campaign to embed generative AI across consumer and industrial applications, requiring public-facing models to register training sources. Hardware Agglomeration and Cultural Pragmatism Rather than indulging in existential 'AI psychosis,' Chinese discourse treats AI as a tool to offset an aging demographic and shrinking manual workforce, supported by legal boundaries such as a Hangzhou court ruling prohibiting AI replacement as grounds for worker terminations. China's dense industrial agglomeration, often containing entire supply chains within a 50-kilometer radius, allows EV and hardware developers to move concepts to production within 15 months at half the cost of global competitors. While physical AI faces short-term bottlenecks like two-hour battery limits and a lack of 3D spatial data for world models, cost-sensitive U.S. startups are already adopting Chinese open-source models for routine operations. While China's integrated hardware supply chains and manufacturing telemetry offer strong long-term advantages for grounded, physical AI, domestic labs still face unresolved hurdles, including enterprise data that remains young and unstructured, persistent hardware energy limits, and tight regulatory oversight over agentic applications.

独立人工智能研究者 Grace Hsiao 认为，中国的人工智能生态在严格的资本约束、芯片出口管制与深厚制造业基础的共同塑造下，走出了一条高度务实、开源协同并与实体硬件紧密绑定的差异化发展路径。 算力受限下的开源协同与后训练策略 (6:58) 中国人工智能实验室推行开源最初是为在国际市场上建立公信力并化解技术疑虑，随后在 DeepSeek 创始人梁文峰等人的推动下形成共享研究成果的行业文化，但在表面协作之下依然存在被称为“内卷”的激烈竞争。面对美国芯片出口限制与算力短缺，中国前沿实验室选择优化推理效率，并将计算资源集中于后训练（post-training）与数据微调。Hsiao 将其比作“已知作业答案后的反向推导”：在西方实验室投入数十亿美元通过试错摸索前沿方向后，中国团队能以极低成本将自身性能窗口维持在落后全球前沿仅 6 到 9 个月的区间。 在数据获取方面，中国团队通常等待专有数据集的独占期结束，以原本十分之一的成本购入顶级人类工作流数据。针对美方关于模型蒸馏的指责，研究员姚顺雨区分了简单复制答案的“愚钝蒸馏”与利用前沿模型充当评估及标注工具的“智慧蒸馏”，指出后者属于标准的企业微调流程。此外，DeepSeek-V4 特意推迟数月发布，旨在专为华为国产芯片架构重构推理系统，向行业证实了摆脱对英伟达 CUDA 依赖的可行性。 前沿实验室的商业化变现与分层调用 (11:06) 与美国初创企业在未见产品前即斩获巨额投机性估值的模式不同，中国大模型实验室面临更严格的资本环境，因而展现出清晰的商业变现纪律。Hsiao 梳理了四家代表性前沿初创机构：智谱AI主打代码能力并在香港上市；MiniMax 专注音视频多模态研发，香港隐含估值约 200 亿美元；月之暗面依托 Kimi 聚焦智能体与推理，正筹备 IPO；DeepSeek 则专注于基础研究并充当开源底层。这些公司通过提供托管 API 推理、企业级本地部署和安全护栏等服务，MiniMax 与智谱AI的预期年度经常性收入已达 1 亿至 10 亿美元区间。 在软件架构方面，Cursor 与 Harvey 等注重成本控制的美国初创企业已开始采用混合调用方案：在处理大通量常规任务时部署智谱 GLM 或月之暗面 Kimi 等中国高性价比开源模型，仅在最终评估与复杂仲裁环节调用 Claude 3.5 Opus 等西方顶级模型，以此大幅降低 Token 消耗成本。 “东数西算”基建与务实的社会治理取向 (20:00) 针对全球 AI 普遍面临的电力与电网瓶颈，讨论指出中国的能源供给优势并非源于对 AI 耗电的前瞻性预判，而是过去三四十年间快速城市化进程中对现代化电网、水电与太阳能系统大规模建设的自然沉淀。依托“东数西算”国家战略，算力中心被统一部署在贵州、新疆、内蒙古与四川等西部清洁能源富集区，高效承接东部沿海经济带（居住着约 90% 的人口并产生绝大部分数据）的算力负荷。 在社会观念与监管层面，中国 AI 社区表现出强烈的工程实用主义，普遍缺乏西方科技领袖对末日场景的形而上学推演。研究团队年轻化且以学者和学生为主，更关注实际业务落地。杭州法院在标杆判例中明确裁定，企业不得将“被 AI 替代”作为合法辞退员工或缩减合同的借口。尽管跨越式普及移动支付的中国公众对技术极为包容，但当“Open Claw”等智能体工具在民间引发数据安全风险时，中央监管部门随即介入警告，促使金融机构与国企迅速撤回了部署。 硬件制造集聚与物理AI的发展瓶颈 中国政府将人工智能定位为应对人口老龄化与劳动力缩减的宏观生产力工具，地方政府通过设立 11 至 12 个国家级 AI 先导区，以类似风险投资机构的角色为初创团队提供场地、宿舍与基础设施资金。深圳等珠三角工业集群具有极高的产业集聚密度，完整供应链常浓缩在 50 公里半径以内。这使得中国硬件与电动汽车厂商将新产品从立项推至量产的周期缩短至 15 个月以内（传统 OEM 需 3 至 5 年），生产成本不到全球同行的一半，且庞大的工厂网络为具身智能提供了天然的物理世界训练数据。 然而，Hsiao 强调消费级物理 AI 的成熟仍面临重大工程阻碍。人形与四足机器人无法仅凭大语言模型运转，极度匮乏专用的 3D 空间数据与“世界模型”，同时设备续航普遍被限制在 2 小时以内。尽管有企业尝试通过磁吸可拆卸电池仓等工程微创新延长工时，但物理硬件与底层算法的全面融合仍处于早期攻坚阶段。 总体而言，中国 AI 行业在算力封锁与资本收紧的双重约束下，凭借开源协同、逆向工程效率以及无与伦比的硬件制造供应链，在工程实用主义路线上建立了独特的竞争壁垒。然而，能否在受限硬件上持续跟进全球最前沿理论、突破物理 AI 的空间数据与续航瓶颈，并平衡强监管与自发创新之间的张力，仍是其长期演进中悬而未决的核心命题。

## Background 本期节目由 Bloomberg 旗下的播客《Odd Lots》于 2026 年 6 月 22 日在网络与 RSS 平台发布。节目由主持人 Tracy Alloway 和 Joe Weisenthal 在香港现场录制，特邀独立 AI 研究员、《AI Proem》Substack 专栏作者 Grace Shao（邵），深入探讨 DeepSeek 爆发后中国大模型产业的最新格局、开源生态演进、硬件制造优势、算力与资本约束，以及具身智能和数据市场等前沿话题。 ## TL;DR 节目全面解构了当前中国 AI 产业不同于硅谷的独特演进路径：在外部芯片出口管制与风投资本趋紧的双重制约下，中国前沿实验室（如 DeepSeek、智谱 AI、MiniMax 和月之暗面）放弃了高昂试错成本的预训练盲目竞赛，转而依托开源生态建立全球开发者信任，并将极其有限的算力聚焦于后训练与推理工程优化。与此同时，得益于过去数十年城镇化积累的崭新电网与“东数西算”调度体系，能源并未成为中国算力的主要瓶颈，而在长三角与珠三角方圆数十公里内高度集聚的供应链体系与仅为海外一半的硬件造价，使中国在具身智能硬件与智能制造落地层面具备强大韧性。此外，美国本土垂直软件正逐步转向“闭源做裁判、中国高性价比开源模型跑推理”的混合架构，展现出中美两地在商业变现与技术落地上的务实分工。 ## Atmosphere & style 整场对话充满现场感与机智幽默的思辨张力，既有雨季香港雷鸣下的故地重游叙事，也有主持人在拆解硅谷宏大玄学叙事（“AI 精神病”）时的辛辣吐槽。嘉宾 Grace Shao 的阐述冷静务实、洞见敏锐，以一线调研视角剥离了地缘政治的非黑即白，生动还原了中国工程师在机舱“引擎室”里的真实工程权衡与烟火气。节目整体节奏明快、信息密度极高，在宏观战略剖析与中老年人扫码用技术的微观细节之间自如切换，营造出兼具全球宏观视野与深厚地气现实的深度对谈体验。 ## Bulletin - `12:35` 面对资本紧缩与估值落差，中国大模型独角兽（如智谱 AI、MiniMax）通过提供托管 API 与推理服务实现商业闭环，单月收入已比肩去年全年总和，年底 ARR 预期达 10 亿至 15 亿元人民币。 - `16:32` 在芯片出口管制与研发预算受限下，中国前沿实验室采用“后训练工程优化”与排他期满后“错峰低价采购垂直数据集”的打法，造成中美模型之间存在约 6 至 9 个月的时间差。 - `20:00` 得益于近几十年城镇化新建的坚固电网及自上而下的“东数西算”西电东送与算力调度，能源目前未像美国那样成为制约中国 AI 数据中心扩张的核心瓶颈。 - `45:24` 中国硬件综合制造成本较海外低 50% 以上，但物理世界交互数据匮乏与电池续航仍是具身智能瓶颈；与此同时，Harvey、Cursor 等美国 AI 原生应用已反向采用中国开源模型以优化 Token 成本。 ## Key takeaways - `03:00` 中美 AI 产业呈现鲜明文化反差：美国由闭源前沿实验室主导，偏好宏大玄学的 AGI 宿命叙事；中国则由开源生态主导，将 AI 视作赋能实体经济与既有互联网业务的实用工具。 - `05:00` 中国大模型走向全面开源首先是打破海外“信任赤字”的务实品牌策略，并在算力、资本和人才受限的客观约束下倒逼出互借成果、平摊试错成本的开源协同生态。 - `10:00` DeepSeek 推迟发布以率先完成对华为昇腾国产硬件技术栈的适配迁移，不仅验证了全栈国产推理的可行性，也为其他前沿初创实验室提供了可参考的基础设施范式。 - `10:00` 面对研发资源瓶颈，中国“四小龙”创企走向差异化分工：智谱 AI 专注代码编程，MiniMax 聚焦多模态交互，月之暗面深耕 Agent 工作流，DeepSeek 持续追踪前沿基线。 - `12:35` 开源并不等于无法盈利，企业客户愿意为免去 GPU 采购、机房运维和合规护栏搭建的“云端推理 API 服务”付费，推理通道已成为创企稳定的高毛利来源。 - `16:32` 受限于无法承受“十投一中”的高昂探索成本，中国科研团队常跟踪海外明确技术路径，将有限算力重仓押注在后训练（Post-training）阶段进行极致工程实现。 - `22:15` 针对西方“搭便车”的指责，业内应区分死记硬背的“愚钝蒸馏”与利用前沿模型作为导师引导逻辑推理的“聪明蒸馏”，后者实为全球模型微调的标准工业工序。 - `25:00` 微信等超级应用坐拥超 14 亿月活却对接入原生 Agent 极为克制，核心在于保护极致的用户体验，并规避国民级应用中 AI 失控带来的交易与监管合规风险。 - `27:35` 海外顶尖华人学者回国更多受家庭生活方式、配偶职业连续性及国内极具竞争力的实际薪酬待遇驱动，而非单纯的地缘政治意识形态考量。 - `32:15` 杭州法院裁定企业以“被 AI 替代”为由解雇员工属违法，明确划定了保护劳动力稳定的司法底线，有效平息了社会对技术性失业的焦虑情绪。 - `34:52` 中国采取工信部推动“AI+”产业赋能、网信办严守安全合规的双轨协同监管机制，所有面向公众的大模型应用均须通过穿透式国家备案。 - `40:00` 地方政府以 KPI 为驱动扮演“VC 角色”，向初创团队提供算力、免租办公场地与人才宿舍，并依托大型开发者社区内测反馈形成闭环孵化网络。 - `42:47` 具身智能并非简单地将语言模型置入机械身体，真实世界 3D 物理交互数据的缺失与电池续航短板，是制约人形机器人落地的核心短板。 - `45:24` 中国不仅具备难以向海外迁移的高技能产业工人体系与极速量产能力，其高性价比开源模型也正在被美国头部软件反向采用以替代高价闭源 API。 - `48:36` 中国老龄群体对新硬件界面的接受门槛远低于西方，源于过去数十年移动互联网与超级应用普及所建立的高度数字化的日常生活范式。 - `52:45` 全球 B 端企业级软件正广泛采用“混合编排”架构，将核心复杂逻辑交给顶级闭源模型裁决，海量并发子任务交由低成本开源模型执行以追求投资回报率（ROI）。 ## Detailed walkthrough ### `00:00` 重返香港：从监管整顿到大模型狂潮 节目在香港开启。主持人 Tracy Allaway 曾在香港生活了近四年，于 2022 年离开，此番故地重游感触颇深。两人聊起香港独特的城市风貌——曾有人将其形容为“曼哈顿与毛伊岛（Maui）的结合体”，既有高密度的摩天大楼金融区，又坐拥宜人的海滩与自然风光；尽管录制正值台风多雨的雨季，电闪雷鸣，却丝毫没有削弱重返此地的兴奋感。 时隔数年，整个科技与宏观叙事已发生巨变。Tracy 回忆道，2020 年她在香港时，最轰动的头条还是中国对互联网科技行业的严监管整顿，当时外界普遍担忧此举是否会彻底扼杀中国的“企业家精神”。然而几年过去，创业与创新活力在各个角落依旧生机勃勃。特别是自 2022 年底 ChatGPT 横空出世以来，中国科技界敏锐地意识到必须在新一代软件与面向消费者的前沿技术上全面发力，各类大模型迎来了井喷式爆发。 --- ### `03:00` 中美 AI 景象对比：神圣叙事与实用主义 谈及对中国 AI 行业的直观印象，Joe 和 Tracy 总结了几个鲜明的反差。首先是生态形态：与美国由少数前沿实验室主导闭源模型的路线不同，中国的大模型呈现出极高程度的开源特征，甚至似乎连卖牙膏的传统公司都在训练发布自己的大语言模型（LLM），令人好奇其背后的商业化变现逻辑。其次是文化氛围与话语体系：美国前沿实验室的掌门人们往往用一种近乎“玄学与宿命论”（quasi-mystical）的宏大基调谈论通用人工智能（AGI）；而中国的 AI 氛围则接地气得多，人们普遍将其视作一项“工具性极强的新技术”，务实地服务于腾讯、阿里巴巴等巨头既有的业务基本盘。 ```mermaid flowchart TD accTitle: 中美 AI 发展路径对比 accDescr: 对比美国与中国在 AI 研发逻辑、话语体系和资本环境上的差异 subgraph US["美国模式 (US Mode)"] U1["闭源前沿模型为主<br/>(OpenAI, Anthropic 等)"] U2["叙事宏大玄学<br/>'AI 即终极业务'"] U3["资本充裕疯狂追捧<br/>资金直接砸向大模型"] end subgraph CN["中国模式 (China Mode)"] C1["全面开源生态为主<br/>(DeepSeek, 阿里开源等)"] C2["务实工具主义<br/>'赋能实体与既有业务'"] C3["资本受限与硬件结合<br/>精打细算应对算力制约"] end U1 -.->|生态对照| C1 U2 -.->|文化反差| C2 U3 -.->|资源逻辑| C3 ``` 此外，资本环境与产业基础也截然不同。在华尔街和硅谷，只要项目带上“AI”标签便能吸引天量资金；而在中国，企业融资面临更为苛刻的资本约束。但与此同时，中国在硬件科技、机器人、电动汽车（EV）以及动力电池等制造领域具备显著领先优势。这种将深厚的硬件制造能力与软件算法相交融的特殊环境，深刻塑造了中国 AI 技术演进的独特轨道。 --- ### `05:00` 开源之谜：信任赤字、开源战略与深度求索的哲学 独立 AI 研究员、《AI Proem》Substack 专栏作者 Grace Shao（邵 grace）应邀加入对话。针对“中国模型为何走上开源道路”这一核心问题，Grace 直言，外界习惯去寻找深奥神秘的哲学解释，但追根溯源，这首先是一个极其实际的商业与品牌决策。 由于地缘政治和认知隔阂，西方开发者与海外企业对中国模型天然存在信任门槛。通过将模型权重全面开源，中国技术团队得以有效打破黑盒、建立公信力，向全球证明其模型架构与技术实力是完全透明且扎实的。在此之上，开源理念也融入了领军人物的研发哲学——例如幻方量化与深度求索（DeepSeek）创始人梁文锋便多次公开表态，希望通过开源最前沿的研究成果，推动全行业共同进步。在严苛的算力禁运、顶尖人才瓶颈以及风险资本收紧的多重倒逼下，各家实验室更加珍惜每一分研发经费与算力资源，选择吸收彼此开源的突破成果，形成了互相借力、甚至在社交平台 X 上公开互道祝贺的研发共振层。 - **开源初衷**：以透明度换取海外开发者的信任，本质上是打破信任逆差的高效品牌破局策略。 - **产业协同**：开源代码与权重论文的公开分享，让不同实验室能快速集成友商的突破，间接平摊了高昂的试错成本。 - **要素制约**：算力短缺（Compute constraint）、人才限制（Talent constraint）与资本紧缩（Capital constraint）共同倒逼出极端追求效率的开源路线。 --- ### `07:00` 从狂热到“内卷”：港股 IPO 潮与 DeepSeek 的特殊定位 在谈及这种同行间其乐融融的分享氛围是否演变时，Grace 指出，这种“共享”最初并非出于某种纯粹的利他设计，而更像是一种竞争压力下的无心插柳。中国 AI 行业内部的竞争同样处于极度的“内卷”（involution）之中。 2025 年初伴随特朗普就职前后发布的 DeepSeek-V3 成了关键分水岭，它彻底扭转了全球科技界对中国 AI 水平的轻视态度，不仅带动了海内外资金重新回流此前低迷了三到五年的中国互联网板块，也让一众前沿独角兽加速驶入资本市场：智谱 AI（Zhipu AI / ZAI）与 MiniMax 已在香港联交所主板挂牌上市，月之暗面（Moonshot AI）亦在筹备上市进程。然而，DeepSeek 在这一拥挤的赛道中却占据着截然不同的超然生态位。当外界基于评测基准跑分对后续模型 DeepSeek-V4 议论纷纭、认为其跑分没有像前代那样拉开与美国顶尖实验室的差距时，北京一线核心研发圈内部反馈的信息却展现出了截然不同的另一面。 ### `10:00` 华为昇腾生态与中国模型前沿“四小龙”的分道扬镳 DeepSeek 此前曾将模型发布推迟了大约三到四个月，原因正是团队希望重新设计底层架构，将大量推理工作迁移到华为昇腾硬件技术栈上。尽管这绝不意味着华为已经完全替代了英伟达或成熟的 CUDA 软件生态——绝大多数开发者只要条件允许，首选依然会是 CUDA——但这次迁移是一次极其关键的“为团队挺身而出”的示范行动。它向整个行业释放了一个强有力的信号：完全基于国产全栈软硬件做大模型推理是切实可行的。 得益于 DeepSeek 全开源、开放权重的模式，其工程成果迅速成为了中国整个模型生态的共享基础设施。其他 AI 实验室得以参考其实践路径，逐步将自家的推理工作向华为全栈迁移。这种转向既有国家层面的战略宣示意味，更出于摆脱外部芯片卡脖子风险的务实考量。 如果暂且抛开国内科技巨头，聚焦于最具代表性的四家前沿初创实验室——DeepSeek（深度求索）、月之暗面（Moonshot AI，推出 Kimi）、智谱 AI（Zhipu AI，推出 GLM 系列）以及 MiniMax——可以发现它们各自面对现实约束作出了差异化选择。受限于算力、资金乃至顶级科研人才的瓶颈，这四家机构被迫走上不同的垂直赛道以争夺商业版图： ```mermaid flowchart TD accTitle: 中国大模型前沿实验室分工格局 accDescr: 展示四家核心初创实验室在算力与资本约束下选择的不同垂直发展方向 subgraph Labs["中国 AI 前沿实验室与差异化定位"] direction TB DP["DeepSeek<br/>(专注前沿突破/追赶基线)"] ZP["智谱 AI (GLM)<br/>(深耕代码编程/类似 Claude Code)"] MM["MiniMax<br/>(主打多模态交互能力)"] MS["月之暗面 (Kimi)<br/>(聚焦智能体 Agent 体验)"] end ``` - **智谱 AI（GLM）**：全面押注编程代码能力，其产品形态和定位更接近 Anthropic 的 Claude Code 或 OpenAI 的 Codex。 - **MiniMax**：深度聚焦多模态生成与交互能力。 - **月之暗面（Moonshot AI）**：将核心研发资源倾斜在智能体（Agent）工作流与长文本落地体验。 - **DeepSeek**：依然坚定致力于推进前沿基础模型演进，全力缩小与全球顶尖水平的差距，并带动国内生态加速跟进。 --- ### `12:35` 资本寒冬下的盈利现实与开源商业闭环 针对部分中国 AI 初创企业上市后的估值表现，现场讨论展现出鲜明的中美反差：MiniMax 以约 200 亿美元的估值进入二级市场，而对比之下，美国一些团队仅仅在 arXiv 上发了一篇学术论文、连实质性产品都尚未拿出的初创团队，在风投资本运作下就能轻易获得极高的估值。 中国风投资本环境在过去几年经历行业整顿后活跃度骤降，红杉等美元基金先后拆分或撤离。早在 2022 年，不少依托高校与科研机构孵化成立的大模型团队在融资层面步履维艰，上市初期的市值仅在 60 亿至 80 亿美元之间。然而，这种资金约束反而逼出了其商业变现能力。 智谱 AI 和 MiniMax 的公开财务数据显示，两家公司近期的单月收入已经比肩去年全年的营收总和，年底的年度经常性收入（ARR）预期已在 10 亿至 15 亿元人民币之间。所谓“开源即无法盈利”在产业界完全是个伪命题： - **托管服务与推理 API**：客户为模型 API 接口付费，本质上购买的是“免于自建”的服务。用户不必自行采购 GPU 芯片、架设算力机房，也不需要自行搭建安全护栏（Guardrails）、监控机制、部署框架与安全合规体系。 - **推理即利润中心**：虽然模型权重开放、支持完全本地化自建私有部署，但官方所提供的高可用性云端推理通道已经成为稳定且快速增长的盈利支撑点。 --- ### `16:32` 算力挤压下的“逆向推导”：后训练与错峰数据战略 当前制约中国 AI 模型发展的主动力来自两大因素：资金规模不及美国同行的挥金如土，以及日趋严苛的高端芯片出口管制。更严峻的是，部分国内互联网巨头甚至通过出资买断数据中心与初创实验室之间的算力租赁合同，直接截流算力资源。算力短缺导致初创实验室甚至无法满足汹涌而来的推理服务需求，只能优先保障高客单价、高质量的推理任务。 这种极度受限的环境直接重塑了中国模型团队的研发优化策略。与 OpenAI 将海量资金铺在探索性研发（R&D）、预训练（Pre-training）和后训练（Post-training）三大多元环节不同，中国团队无力承担“投十个方向只成一个”的高成本研发试错。 受访科研人员用了一个极为形象的比喻：“这就好比已经知道了家庭作业的最终标准答案，然后倒推解题过程。”中国实验室往往密切跟踪全球顶尖前沿实验室探索出的明确技术路径与架构方向，随后将极其有限的算力和研发资源压在**后训练（Post-training）**阶段进行极致工程优化。 在数据采购方面，中国团队普遍采取“错峰捡漏”策略。以 Mercor 等专业数据服务商提供的顶级细分垂直数据集为例，海外顶尖机构在排他期内可能需要支付 1000 万至 2000 万美元的高昂对价；而中国团队则会耐心等待三到六个月的独家排他协议期满，再以原价十分之一甚至更低的成本购入相同数据集进行定向精调。这一策略也恰好解释了外界普遍观察到的中美前沿模型之间存在的 6 至 9 个月的时间差。 ### `20:00` 能源并非中国 AI 的瓶颈：“东数西算”与电网红利 在美国，电力已成为制约 AI 发展的核心瓶颈——寻找具备稳定电力接入的数据中心变得愈发困难。但在中国，情况却截然不同。Grace Shao 指出，能源目前甚至不是中国 AI 面临的最大阻碍。外界有时会有一种误解，以为中国政府早有神机妙算，提前预判了 AI 热潮带来的巨大耗电量，但事实并非如此。中国真正的优势源于过去三四十年迅猛的城镇化进程。西方人如今到访的北京、上海、深圳等大都市，满街跑着电动车和机器人，这些现代都市几乎都是在过去二三十年间拔地而起的。正因如此，中国的电网基础设施极其崭新，政府此前为了应对城市扩张和工业用电激增，早已前瞻性地建设了大规模的水电、光伏等能源网络。这场意外与 AI 算力爆发在时间上撞了个正着，客观上为中国 AI 提供了充足的电力底座。 更深层的体制优势在于中央政府自上而下的超强执行力，各省必须协同跟进，而不会像美国那样各个州各行其是。最具代表性的就是“东数西算”工程：在贵州、新疆、内蒙古、四川等偏远山区和内陆省份，政府集中兴建了大量廉价的可再生能源基地与超算中心；而中国 90% 的人口集中在东部沿海（如京津冀、长三角、大湾区），那里才是数据生成与应用的核心市场。通过国家级管网将东部的数据传输到西部进行离线计算和模型训练，这种负载优化有效缓解了东部城市的能源压力。 ```mermaid flowchart TD accTitle: 中国东数西算战略架构 accDescr: 展现中国东部人口密集区产生的数据与西部低成本能源算力基地之间的协同流动关系 subgraph West["西部可再生能源与算力基地"] W1["偏远与内陆省份<br/>（贵州、新疆、内蒙古、四川）"] W2["廉价绿色能源<br/>（水电、光伏、风电）"] W3["大规模数据中心集群<br/>（承担模型训练与海量算力负荷）"] W1 --> W2 --> W3 end subgraph East["东部沿海核心经济带"] E1["人口聚集区<br/>（占全国90%人口）"] E2["一线超大城市<br/>（北京、天津、上海、深圳）"] E3["数据与商业需求来源<br/>（消费互联网与企业终端）"] E1 --> E2 --> E3 end E3 -- "产生海量数据与训练任务" --> W3 W3 -- "回传算力结果与模型权重" --> E3 ``` - 中国电网与西方老旧电网不同，整体设备普遍是在过去 20 至 30 年城镇化大潮中新建的，抗压与调度能力极强。 - “东数西算”让高能耗的计算集群直接建在风光水电资源极度富集的西部，避免了沿海核心城市电网过载。 --- ### `22:15` “搭便车”争议与聪明蒸馏：从数据采购到 DeepSeek 的算力困境 针对数据采购与模型训练的公平性，主持人提到一些专业数据服务商（例如 Mercor）会专门雇佣专业人员从事具体操作（例如现场制作复杂的 PowerPoint），借此采集高质量的人类交互数据并打包出售给 OpenAI。这些数据通常有一段“排他独家期”（Exclusivity Window），期满后供应商便可向市场公开发售。中国的大模型实验室往往会耐心等待独家期结束，以远低于首发的价格（比如以 100 万美元捡漏原价 1000 万美元的数据集）进行采购。这引发了西方同行的广泛非议——Anthropic 等美国实验室指责中国团队通过各种手段抓取美国前沿模型的输出，借由“模型蒸馏”（Distillation）搭便车；批评者认为，若没有美国顶尖闭源实验室不惜代价砸入巨额资本开拓技术边界，中国开源模型根本不可能保持如此接近前沿的水平。 Grace 坦言，中国团队在算力和资本上的双重约束是客观存在的事实，业内“没有人对此进行遮掩或装作不存在”，就连 DeepSeek 自己也公开承认面临算力不足的挣扎。但对于蒸馏的指责，需要理性区分技术细节。Grace 引用了 Google DeepMind 研究员 Yao Shunyu 的观点，指出蒸馏分为“愚钝蒸馏”（Dumb Distillation）与“聪明蒸馏”（Smart Distillation）： - **愚钝蒸馏**：属于纯技术层面的无脑硬搬，直接拿数千个问题去狂刷 Claude 的 API，强制将模型输出的答案照单全收、强行灌入自己的模型，逼迫自家模型强记出一模一样的答案。 - **聪明蒸馏**：更像是将前沿模型当作“陪练、伙伴或导师”。开发者借助前沿模型进行数据标注、逻辑判定与质量评估，利用其高超的推理能力来引导自家模型学习思考链路。Grace 强调，聪明蒸馏实际上正是全球所有企业在做模型微调时的标准工序，在技术伦理上并不应受到过分指责，整个边界目前依然处于一片灰色地带。 --- ### `25:00` 微信的 AI 困境与腾讯的 Agent 雄心 当谈及中文大模型的数据底座时，很多人直觉认为中国人口基数庞大、互联网极其发达，像腾讯拥有国民级应用微信，理应坐拥无与伦比的数据宝矿。然而 Grace 揭示了繁荣表象下的两个现实隐忧： 一方面，中国知识工作经济与现代企业软件生态的成熟度与欧美相比仍存在明显代差，大量沉淀的数据属于非结构化数据，缺乏像美国 Mercor 那样成熟细分的数据标注与工程化提供商，大模型真正所需的“高质量知识数据”实际上相当稀缺。 另一方面，坐拥海量数据的巨头在落地 AI 时往往瞻前顾后。以腾讯为例，其刚刚宣布正在秘密打造一款可以直接接入微信生态的原生 AI Agent，但这一举措在内部引发了剧烈反弹与争议： - **“产品之神”的领地意识**：微信缔造者兼核心负责人张小龙（Alan Zhang）在业界以极度保护用户体验而闻名，微信一直被他视作不可妥协的“亲生骨肉”，任何可能破坏简洁交互与用户体验的外部插件在微信内部都寸步难行。 - **合规与失控风险**：微信全球月活用户（MAU）已突破 14 亿（绝大多数在中国大陆本土，其余为海外华人及有对华业务往来的人群）。在如此庞大的国民级应用中，一旦 AI Agent 在聊天过程中“失控暴走”，或是在自动化执行交易、支付转账时发生误操作，将带来难以承受的声誉和合规风险。 此前腾讯在自研前沿模型上一直落后于阿里、百度等竞争对手，如今正在全力加速追赶。去年，腾讯从 OpenAI 挖来了同名研究员 Yao Shunyu 领衔这项攻坚。腾讯的核心目标并非一味在基准测试上刷分，而是打造“腾讯原生 Agent 模型”，彻底赋能自身既有业务，尽快将 AI 能力推送给最广泛的普通消费者。 --- ### `27:35` AI 人才的归国潮：家庭、生活方式与现实考量 主持人抛出一个尖锐观点：放眼全球顶尖 AI 实验室，华人科研人员的身影无处不在，“AI 必将由华人构建，唯一的分歧在于它究竟是由身处美国实验室的华人构建，还是由中国本土实验室的华人构建”。针对是否有大批在美科研人才正在回流中国的问题，Grace 拆解了隐藏在地缘政治叙事背后的烟火气。 媒体往往喜欢将归国浪潮简单归结为地缘政治审查或是学者心中的恐惧，但现实要务实、接地气得多——“大家终究只是想照顾好家庭、过上好日子，甚至就是看你给的薪酬待遇（Package）到底如何”。许多顶级研究员私下坦言，他们的配偶可能是国内的执业律师、护士或教师，这类专业资格与职业关系极难无缝迁移至美国职场。如果国内头部实验室能够开出极具竞争力的薪资，并提供可观的发展空间，他们往往更倾向于留在中国。 这种代际观念发生了翻天覆地的转变： - 三十多年前 Grace 的父母移民北美时，两地的生活质量有着清晰客观的差距，北美任何一个城市几乎都全方位胜出。 - 而今天，这已经演变成纯粹的个人生活方式选择。如果科研人员喜欢住在热闹的市中心、偏好快节奏、着迷于满街电动车与前卫科技感的未来都市生活，中国的超一线城市能完美满足这种期待。再加上照顾国内年迈父母的情感诉求，家庭与生活方式成了很多人归国的决定性因素。 谈及中外对顶尖技术人才的追捧风格，Grace 笑着指出，美国科技圈往往把顶级研究员的跳槽当作“职业体育明星转会”来炒作，动辄登上头条；而中国本土对待人才的挖角虽低调隐秘得多，从不公开讨论具体价码，但各大厂商给顶尖科研领军人开出的实际薪酬待遇同样极其丰厚。 ### `30:00` 顶尖人才薪酬对比与对美国“AI 精神病”的吐槽 在探讨中美 AI 行业的人酬差异时，Grace Shao 指出，中国顶尖科技大厂（如字节跳动）的高级产品经理，其薪酬水平完全可以与 Meta 等硅谷巨头的同行平起平坐；平均水平的 AI 研究员在中美两地的收入也相差无几。尽管硅谷可能存在个别天价签约——例如动辄数千万甚至上亿美元的天价签约奖金与期权——在中国可能罕见这种极端数字，但中国 AI 从业者在近期的 AI 热潮中，以及此前一波波的技术公司 IPO 中早已赚得盆满钵满，整体回报极为可观，只是绝对峰值规模略有差异。 随后，主持人 Joe Weisenthal 开启了一段半开玩笑的“播客大不敬”吐槽。他直言，自己作为主持人其实并不特别渴望邀请像 Sam Altman（OpenAI 首席执行官）或 Demis Hassabis（Google DeepMind 首席执行官）这类美国顶尖 AI 实验室的 CEO 来做客。Joe 坦言，自己实在厌倦了听他们大谈“人类工作的终结”、“生命的意义”或宏大的哲学叙事： - 这些高层往往患有某种程度的“AI 精神病”（AI psychosis），开口闭口全是宏大而空洞的概括； - 相比之下，他更渴望与身处真正“机舱/引擎室”（the engine room）的技术研发人员与基层从业者对话，聊聊具体而扎实的工程运转与现实操作； - 联合主持人 Tracy Alloway 打趣说，Sam Altman 现在应该主动申请上节目来“测试你的底线”；Joe 则笑称如果真能请到，他一定要逼着对方回答最枯燥、最细枝末节的运营问题，坚决不聊“如果我们全失业了，人生的意义是什么”。 --- ### `32:15` 中国 AI 社区的务实主义与杭州“AI 解雇案”判决 针对美国 AI 巨头们普遍散播的末日叙事——要么宣称白领工作将全盘消失，要么声称自己正在创造一个失控的“AI 怪物”——Joe 询问中国是否存在类似的生存危机讨论。Grace Shao 明确表示，中国 AI 社区的整体氛围要务实得多： 开源研究员 Nathan Lambert 近期走访中国各大 AI 实验室后的见闻极具代表性：他惊讶地发现，中国多数主流 AI 实验室的团队成员极为年轻，大量由学生和实习生挑大梁，核心研发团队往往仅由几位学术背景扎实的科学家统领。这种脱胎于学术界的研发文化使得他们远离资本包装与二级市场的狂热炒作，鲜有西方流行的末日妄想或宏大叙事（high level narrative）。 在公众反应与劳动力市场方面，尽管社会也存在一定焦虑，但监管机构展现了极为坚决的底线保护： - 杭州近期出现了一起里程碑式的劳动争议案件（杭州是阿里巴巴等多家科技巨头与 AI 机构的所在地）：一家公司试图以“你的岗位被 AI 替代”为由解雇员工，当地法院直接判决该解雇行为违法，裁定任何企业不得以 AI 为借口裁员、解雇或强行缩短劳动合同期限。 - 这一迅速而明确的司法裁决在很大程度上平息了公众对 AI 冲击就业的恐慌。 - 与美国企业哪怕毫无技术依据也喜欢在财报中将裁员归咎于 AI 形成鲜明反差，中国政策层面并不允许此类卸责行为。 - 与此同时，中国经济结构中纯知识工作者的比例相对较低，因此替代焦虑的迫切性并未泛滥。 - 整个东亚地区（涵盖韩国、新加坡等）均以高度务实的态度拥抱变革：疲惫的家长们热衷于让孩子掌握 AI 成为“AI 原生代”（AI native），学生们积极自我武装，整个社会聚焦于“积极适应未来”，而非“抵制未来”。 --- ### `34:52` 监管协同模式与“AI+”国家战略 谈及美国实验室标榜的对齐研究（Alignment）与 AI 安全投入，Grace Shao 分析了中国截然不同的治理架构与监管落地路径。中国并未陷入安全末日论的清谈，而是通过制度化的双轨协同与牌照注册体系进行管理： ```mermaid flowchart TD accTitle: 中国 AI 监管与发展双轨架构 accDescr: 展示工信部推动落地与网信办设置红线的协同管理模式，以及全量模型必须经由的国家备案流程。 subgraph GovFramework ["政府协同机制"] direction LR MIIT["工业和信息化部等<br/>推动经济发展与场景赋能"] CAC["国家网信办等<br/>设置合规红线与安全监管"] end subgraph Implementation ["国家备案与落地执行"] Registry["国家生成式 AI 备案库<br/>披露训练数据/潜在风险/算法安全"] Application["全产业应用：'AI+' 政策<br/>赋能智能制造/家电硬件/各行各业"] end GovFramework --> Registry Registry --> Application ``` - **政府部门双轨协同**：工业和信息化部（MIIT）等部门致力于推动产业落地，力主发展“AI+”（即将生成式人工智能赋能至千行百业）；国家互联网信息办公室（CAC）等部门则担当安全红线守门人，两者并行不悖。 - **强制备案与穿透式公开**：中国所有面向公众的生成式 AI 应用以及底层大语言模型（LLM）都必须通过国家备案机制，详细披露训练数据来源、安全机制及潜在风险。 - **“AI+”重塑实体产业**：正如 15 年前的“互联网+”政策，当前的“AI+”正成为国家级战略推手。特别是在 DeepSeek 引发业界震撼后，民间与实体制造业（包括传统家电制造厂商）爆发了将开源端侧模型嵌入硬件设备、快速实现商业化落地的浪潮。在务实的政策与产业共识下，贩卖技术恐慌在中国科技界始终无法成为主流。 ### `40:00` 自上而下的国家推力与地方政府“风投化” 谈及中国人工智能的发展驱动力，话题从市场上五花八门的消费电子产品切入——“AI 扫地机器人究竟能帮你多做什么？AI 电动牙刷又有什么意义？”这种看似不可思议的产品概念之所以大行其道，背后关键在于政府自上而下的政策推力。在宏观战略层面，中国政府将 AI 视作拉动经济增长的核心引擎，既为了提升全社会效率，也为了应对人口结构老龄化与劳动力总量下降所带来的用工缺口。如今许多年轻一代不再愿意走进工厂从事传统制造工作，工厂本身渴望自动化，而劳动力也倾向流向城市服务业。 在此背景下，各省市地方政府将推动 AI 发展化作核心 KPI，扮演起类似于风险投资机构（VC）的角色，积极在辖区内寻找并扶持“未来的 DeepSeek”。为了吸引初创团队，地方不仅提供专项资金，还直接配给写字楼、办公空间，甚至为青年创业者安排现成的员工宿舍。目前全国已落地约 11 至 12 个国家级人工智能创新应用先导区，供企业测试最新的 AI 产品。 在走访中国本土规模最大的 AI 开发者社区（拥有数十万名开发者）时，社区创始人透露，他们正充当着监管部门与民营科技巨头之间的桥梁。例如字节跳动研发出新产品后，会首先交给该开发者生态进行内测与除错；社区在测试并掌握数据后，再向提供场地与算力支持的地方政府汇报，协助评估是否值得跟进投资或提供产业配套。这种企业、开发者生态与地方政府之间高度协同的联动模式，构成了独特的产业落地网络。尽管企业对于接受国有资本仍存有顾虑与讨论，但随着 AI 与机器人被提升至涉及军事潜力与地缘政治博弈的战略高度，这一领域在中美两国都呈现出日益明显的国家化倾向。 ```mermaid flowchart TD accTitle: 地方政府与AI企业协同孵化机制 accDescr: 展示地方政府通过资金与基础设施扶持开发者生态，并与科技大厂产品形成闭环反馈 subgraph Gov ["地方政府 (VC化运作)"] G1["设定AI产业KPI"] --> G2["提供资金/办公楼/人才宿舍"] end subgraph Ecosystem ["产业开发闭环"] D1["全国先导试验区"] --> D2["数十万开发者社区"] T1["大厂新品研发 (如字节)"] --> D2 D2 -->|"内测排错 & 数据反馈"| T1 D2 -->|"项目成效报告"| G2 end ``` --- ### `42:47` 具身智能的现实瓶颈：不是插个大模型就能当机器人 许多人在童年时期幻想的人工智能，并非如今屏幕里的对话框，而是像电影《终结者2》中 T-1000 那样拥有实体的人形机器人。针对前沿模型实验室（如 DeepSeek、MiniMax、智谱 GLM 等）是否正与宇树科技（Unitree）这类先进机器人硬件企业紧密联合、共同打造未来的终极机器人，现实情况却比外界想象的更加复杂。 中国坐拥过去三十年积累的世界制造中心地位，掌控着无与伦比的供应链生态——在特定产业集群带，从基础原材料到代工厂（OEM）的终端成品，整条产业链往往被压缩在方圆 50 公里之内。这吸引了大量美国投资人与创业者亲临中国考察。然而，将纯软件算法融入实体硬件仍面临巨大的技术鸿沟： - **物理世界数据匮乏**：机器人企业根本无法直接将 MiniMax 这类语言模型当作通用“大脑”插入机械躯壳。物理 AI（Physical AI）与世界模型是截然不同的技术路线，缺乏真实世界的 3D 物理交互数据，导致语言模型（LLM）无法直接指挥双足或四足机器人进行平稳作业。 - **系统集成与续航短板**：硬件的核心瓶颈还在于机电集成能力与电池方案。虽然中国在动力电池产业链上处于领先地位，但大多数轻量化智能硬件的续航依然撑不过两小时。 - **工程补丁式创新**：在现有电池技术突破前，本土初创团队往往依赖极致的工程设计解决痛点。例如面对智能眼镜（如 Meta Ray-Ban）两小时断电的问题，科大讯飞（iFlytek）或 Rokid（若琪）等厂商推出了微型外挂“电池胶囊”，用户可直接轻量吸附在镜腿上，在不破坏佩戴体验的前提下延长数小时续航。 --- ### `45:24` 未来三年的比较优势：硬件造价减半与美国本土企业反向采用中国开源模型 对于中国未来几年的核心比较优势——究竟是高性价比的开源模型、深度嵌入传统产业的工业软件，还是机器人硬件制造，答案呈现出软硬件双轨并进的态势。 | 评估维度 | 中国市场的核心表现与比较优势 | 行业参照与实际案例 | |:--- |:--- |:--- | | **硬件制造成本** | 机器人、智能家电及硬件设备的综合制造成本，普遍比全球其他地区**低 50% 以上**。 | 海外竞品在成本控制与规模量产上难以抗衡。 | | **高技能产业工人体系** | 数十年积累的精细制造经验（Know-how）难以向海外直接复制迁移。 | 英国《金融时报》记者帕特里克·麦吉（Patrick McGee）在其关于苹果的专著中指出，苹果尝试将供应链移往印度时，面临的最大瓶颈正是数十年培育出的本土高技能技工短缺。 | | **敏捷研发与量产周期** | 硬件产品迭代极快，对市场反馈与供应链协同反应迅速。 | 本土电动汽车（EV）团队从产品立项、生产到下线展厅耗时**不到 15 个月**，而传统跨国 OEM 至少需要 3 至 5 年。 | | **开源模型的商业采纳** | 极具性价比且高度优化的开源模型正在渗透美国本土的高黏性 AI 原生应用。 | **Harvey**（法律 AI）与 **Cursor**（AI 代码编辑器）等美国头部软件，已开始采用混合模型（Hybrid Model）架构以削减昂贵的 Token 支出。 | 特别是在模型应用层面，一年前硅谷初创团队使用通义千问（Qwen）往往只是出于对成本的极度敏感，对合规和地缘政治并不在意。而如今，由于大模型的投资回报率（ROI）面临质疑，大量美国企业级原生软件为了支撑庞大的 Token 吞吐量，转向了混合模型架构：核心生成与上下文处理交由 GLM、Kimi 或 Qwen 完成，仅保留 Claude Opus 这类高成本旗舰模型作为终审裁判（Judge）或高难度逻辑引导。这些中国模型团队因承接了大量真实的海外商业请求，甚至一度面临推理算力供给不足的甜蜜烦恼。 --- ### `48:36` 物理具身对 AI 写作的假说与“中国大妈排队装机械爪”的真相 针对模型智力的演进，产业界私下流传着一种颇具哲学意味的假说：目前 AI 文本写作之所以依然带有某种违和的“机械味”，本质是因为模型被困在没有实体的冰冷数据中心里；一旦 AI 真正融入人形机器人、拥有在真实物理空间中受约束的肉身体验（Groundedness），其文字表达与思维深度将迎来飞跃。 对于海外社交平台上疯传的“中国八十岁老太太排队组装配置 OpenClaw 机械爪”等热门短视频，主持人表达了极大的怀疑，甚至怀疑纯属包装出来的假新闻。这一现象的背后具有双重逻辑： 一方面，视频内容确实存在博眼球与短视频炒作的成分；但另一方面，现实中老龄群体排队体验、甚至尝试配置 AI 硬件设备的场景却真实存在。中国社会的叔叔阿姨、大爷大妈对新技术的接受门槛显著低于许多西方国家。无论是主动拥抱还是生活环境所迫，现代中国的衣食住行已经完全建立在高度数字化的生活方式之上。长辈们日常离不开各类数字生态，早已习惯在日常生活中面对并接纳任何新技术界面。 ### `50:00` 中国对科技的独特乐观主义与“超级应用”生态中的 OpenClaw 热潮 在谈到中国社会对技术的接受度时，Grace Shao 分享了切身经历的生动细节：在北京，如果没有微信支付（WeChat Pay），你甚至连一杯星巴克都买不了。疫情结束后，她第一次回到上海，坐在一家餐厅里傻傻地坐着等人过来帮忙点餐，结果半天没有任何服务员上前。周围人的眼神仿佛在质问：“你怎么像个穴居原始人（caveman）一样？难道不知道直接扫桌上的二维码吗？”这种高度数字化的习惯背后，折射出中国与西方截然不同的科技乐观主义。在过去二三十年里，中国许多偏远农村地区原本根本无法触及一线城市的商品、资讯与公共资源，直到移动互联网与超级应用（Super Apps）彻底打破壁垒。许多村民年家庭收入或许只有一千美元左右，家里甚至没有电视机，但人手一部智能手机。正是这部手机，让他们能够获取小额信贷、在线网购农资生活用品，并让孩子连接外界的信息网络。因此，科技在中国社会普遍受到尊重与信赖，进入大型科技大厂（Big Tech）工作被视为光耀门楣的家族骄傲。 这种超级应用土壤也造就了此前独特的“OpenClaw 热潮”。由于 Anthropic 和 OpenAI 等西方前沿大模型公司并未在中国开展业务，普通人直接访问受限，OpenClaw 便成了许多中国用户最早能够切实接触到的西方前沿 Agent 产品。腾讯与阿里巴巴敏锐地抓住这一风口，迅速尝试将 OpenClaw 的功能整合嵌入到旗下的业务工具与产品矩阵中。在超级应用的架构下，这种交互具有极高的“功能邻近性”（functional adjacency）——用户在微信等应用的搜索框里一搜，就能调出 OpenClaw，进而直接操控微信小程序、自动下单购物，瞬间引发了大众级的狂热。 然而，监管层的反应极其迅速。最初，无锡等地方政府甚至还曾表态鼓励本地企业积极拥抱 OpenClaw，但中央监管部门迅速踩下刹车，明确警告必须高度重视数据隐私与系统安全，严令银行和国有企业（SOEs）不得随意接入，科技大厂随后迅速缩减了相关推广营销。令人捧腹的是，市场上甚至很快冒出了一批专门针对中老年群体的广告教程，手把手教各家“大叔大妈”如何把自己手机和智能设备上的 OpenClaw 彻底卸载掉。 ```mermaid flowchart TD accTitle: OpenClaw 在中国超级应用中的渗透与监管路径 accDescr: 展示 OpenClaw 借助超级应用快速落地中国并最终引发监管介入的过程。 subgraph 接入渠道 ["生态落地"] A["海外前沿模型受限<br/>(OpenAI / Anthropic 未入华)"] --> B["OpenClaw 成为首个<br/>大众可及的 Agent 工具"] B --> C["巨头生态集成<br/>(腾讯、阿里嵌入企业与消费端)"] C --> D["超级应用无缝交互<br/>(搜索入口 ➔ 调用小程序 ➔ 自动下单)"] end subgraph 监管转向 ["政策与市场降温"] D --> E["地方政府初期鼓励<br/>(如无锡尝试扶持)"] E --> F["中央监管迅速介入<br/>(严禁国企、银行使用，强调数据安全)"] F --> G["大厂收缩营销 & 出现反向需求<br/>(教程教长辈如何卸载 OpenClaw)"] end ``` - 微信与支付宝的无感普及重构了消费场景，线下服务全面转向“扫码即办”的自助范式。 - 农村与低收入群体跳过了 PC 与电视时代，直接通过移动智能设备实现了金融信贷、电商与教育资源的“代际跃迁”。 - 大模型落地不仅取决于技术能力本身，更依赖其与超级应用中既有工作流（如小程序生态、支付体系）的生态协同与功能邻近性。 --- ### `52:45` 结语与主持人复盘：分层模型架构与中国制造的数据复利 访谈结束后，主持人 Joe Weisenthal 与 Tracy Alloway 感谢了 Grace Shao 的分享，并相约未来三年内甚至更早再次邀请她返场，检验她对中国 AI 产业格局的预言。随后，两位主持人就本期对话的核心亮点展开了深度复盘。 首先引发热烈讨论的是企业级 AI 应用的技术架构策略。Joe 和 Tracy 指出，对像法律 AI 独角兽 Harvey 这样的垂直应用公司而言，“闭源与开源混合编排”正成为最具性价比的商业范式。企业并不需要在每一个计算环节都调用昂贵顶级的闭源前沿模型，而是可以将最核心、最复杂的逻辑交给一流闭源大模型，将大量并行、重复但对能力要求适中的子任务分配给几乎同样出色、但调用成本极其低廉的开源模型（如 DeepSeek 等）。这种在同一个应用层下“高低搭配”的组合拳，正为全球 B 端软件企业提供前所未有的套利空间。 其次，中国庞大的制造业优势是否会转化为不可逆的 AI 复利优势，成为本期的关键议题： - **物理世界数据反哺**：中国密集分布的无人化工厂、工业机器人以及各类家庭清洁设备，每天都在源源不断地生成海量物理世界具身数据（embodied physical data）。若能将这些高价值传感器与动作数据投喂给新一代物理世界基础模型，或将构建起其他国家无法复制的长期壁垒。 - **数据结构化的现实挑战**：针对“全中国都在用微信沟通、微信支付，必然拥有无穷无尽优质数据”的普遍预设，Grace 在节目中给出了反直觉的冷水：中国在工业制造和应用端的数据清洗、标准化与结构化程度，其实远未达到外界所想象的那般完善。 --- ### `55:00` 产业集聚效应、估值落差与数据资产交易市场 在探讨物理制造与软件创新时，两位主持人进一步引申到了区域产业生态与经济学上的“集聚效应”（agglomeration）。Tracy 联想起 Dan Wang（王丹）等科技观察家对深圳硬件生态的经典描述：在深圳，一个创业团队几乎只需在华强北和各元器件街区挨家挨户走上一圈，从一个门市部逛到下一个门市部，就能立刻完成一个全新硬件产品的全部零部件采购与打样。这种高密度的完备供应链网络，在当今美国任何地方都无法复制。 紧接着，Joe 对旧金山的 AI 繁荣与纽约的金融圈产生了一种颇为幽默的思辨： - 像机械制造这种需要螺栓、伺服电机和机器视觉硬件的产业高度集聚完全符合逻辑；然而，旧金山的软件 AI 研发本质上纯粹是“伏案脑力工作”（pure desk work），团队只要有网有电脑即可协同，却依然在极小的地理半径内表现出极端严密的产业集聚（“agglomeration”）。 - 在中美大模型赛道的资本定价上，估值差距堪称天壤之别。例如，像 MiniMax 这类在中国排名前列的明星大模型创企，即便估值冲上数十亿美元，相较于 OpenAI 和 Anthropic 等美国对手动辄上百亿乃至上千亿美元的估值水准，仍然显得微不足道。 - 此外，Grace 在访谈中提及的中国“数据要素市场”也令主持人大开眼界。例如，中国科技公司在大模型训练中，可以在排他期结束后出资竞购昂贵的专属专有数据集，这种明码标价的合法数据变现模式展示出了与西方完全不同的制度探索。节目最后，主持人提示听众关注 Grace Shao 的 Substack 专栏《AI Proem》，并欢迎大家加入 Discord 社区共同参与大模型与数据市场的后续研讨。 ## Notable quotes > `00:00` 曼哈顿与毛伊岛（Maui）的结合体 — Tracy Alloway > `16:32` 这就好比已经知道了家庭作业的最终标准答案，然后倒推解题过程。 — 科研人员 / Grace Shao > `22:15` 愚钝蒸馏 — Yao Shunyu / Grace Shao > `22:15` 聪明蒸馏 — Yao Shunyu / Grace Shao > `27:35` 大家终究只是想照顾好家庭、过上好日子，甚至就是看你给的薪酬待遇（Package）到底如何 — Grace Shao > `30:00` AI 精神病 — Joe Weisenthal > `30:00` 机舱/引擎室 — Joe Weisenthal > `50:00` 你怎么像个穴居原始人（caveman）一样？难道不知道直接扫桌上的二维码吗？ — Grace Shao ## Mentioned Grace Shao, Joe Weisenthal, Tracy Alloway, DeepSeek, 梁文锋, 智谱 AI (Zhipu AI / GLM), MiniMax, 月之暗面 (Moonshot AI / Kimi), 华为 (Huawei), 昇腾 (Ascend), 英伟达 (Nvidia), CUDA, OpenAI, Anthropic, Google DeepMind, Sam Altman, Demis Hassabis, Yao Shunyu, 张小龙 (Alan Zhang), 腾讯 (Tencent), 微信 (WeChat), 阿里巴巴 (Alibaba), 百度 (Baidu), 字节跳动 (ByteDance), Mercor, Harvey, Cursor, 通义千问 (Qwen), 宇树科技 (Unitree), 科大讯飞 (iFlytek), Rokid, OpenClaw, Dan Wang, Patrick McGee, Nathan Lambert ## Video description <p data-prosemirror-content-type="node" data-prosemirror-node-name="paragraph" data-prosemirror-node-block="true" data-pm-slice="1 1 []">China's AI industry has changed a lot since DeepSeek released its cheap frontier model last year, and briefly sent US tech stocks falling. After being locked out of the most advanced chips, Chinese companies are now allowed to buy some Nvidia H200s. In fact, many of the big Chinese tech companies &mdash; like Baidu &mdash; are making a push to become full-stack players, with their own chips, models, and cloud infrastructure. Today's guest is Grace Shao, an independent AI researcher and the author of the AI Proem Substack. She's a bit of an insider when it comes to China's AI industry, and when we were in Hong Kong we spoke with her about the latest in open-source models, the competition among Chinese frontier labs, DeepSeek's place in an increasingly crowded Chinese AI market, China's manufacturing edge, where bottlenecks exist right now (spoiler: it isn't data centers), if Chinese grandmas are actually using OpenClaw, and finally, of course, AI psychosis.</p> <p data-prosemirror-content-type="node" data-prosemirror-node-name="paragraph" data-prosemirror-node-block="true">Read More:<br data-prosemirror-content-type="node" data-prosemirror-node-name="hardBreak" data-prosemirror-node-inline="true"><a href="https://www.bloomberg.com/news/articles/2026-06-18/china-s-ai-race-spurs-emergence-of-winner-laggard-pair-trade?utm_medium=referral&utm_source=podcast&utm_campaign=odd_lots&utm_content=article" data-prosemirror-content-type="mark" data-prosemirror-mark-name="link">China AI Lab&rsquo;s 170% Stock Surge Cements Winner-Loser Pair Trade</a> <br data-prosemirror-content-type="node" data-prosemirror-node-name="hardBreak" data-prosemirror-node-inline="true"><a href="https://www.bloomberg.com/news/articles/2026-06-17/china-plans-mechanism-to-evaluate-ai-impacts-on-job-market?utm_medium=referral&utm_source=podcast&utm_campaign=odd_lots&utm_content=article" data-prosemirror-content-type="mark" data-prosemirror-mark-name="link">China Plans Mechanism to Evaluate AI Impacts on Job Market</a></p> <p data-prosemirror-content-type="node" data-prosemirror-node-name="paragraph" data-prosemirror-node-block="true">Only <a href="http://Bloomberg.com" data-inline-card="" data-card-data="">http://Bloomberg.com</a> subscribers can get the Odd Lots newsletter in their inbox each week, plus unlimited access to the site and app. Subscribe at&nbsp; <a href="https://www.bloomberg.com/subscriptions/oddlots?in_source=oddlotspodcast" data-prosemirror-content-type="mark" data-prosemirror-mark-name="link">bloomberg.com/subscriptions/oddlots</a></p> <p data-prosemirror-content-type="node" data-prosemirror-node-name="paragraph" data-prosemirror-node-block="true"><a href="http://bloomberg.com/subscriptions/oddlots" data-prosemirror-content-type="mark" data-prosemirror-mark-name="link">Subscribe to the Odd Lots Newsletter</a><br data-prosemirror-content-type="node" data-prosemirror-node-name="hardBreak" data-prosemirror-node-inline="true"><strong data-prosemirror-content-type="mark" data-prosemirror-mark-name="strong">Join the conversation:</strong> <a href="https://discord.gg/oddlots" data-prosemirror-content-type="mark" data-prosemirror-mark-name="link">discord.gg/oddlots</a></p><p>See <a href="https://omnystudio.com/listener">omnystudio.com/listener</a> for privacy information.</p> [Source](https://podtrac.com/pts/redirect.mp3/tracking.swap.fm/track/UVBrz8bN8aM2Xe47PEPu/traffic.omny.fm/d/clips/e73c998e-6e60-432f-8610-ae210140c5b1/8a94442e-5a74-4fa2-8b8d-ae27003a8d6b/ec8ef41e-9021-43ba-a6fb-b46d013a0fc5/audio.mp3?in_playlist=982f5071-765c-403d-969d-ae27003a8d83)

[0:00] Tracy Alloway: 曼哈顿与毛伊岛（Maui）的结合体 [16:32] 科研人员 / Grace Shao: 这就好比已经知道了家庭作业的最终标准答案，然后倒推解题过程。 [22:15] Yao Shunyu / Grace Shao: 愚钝蒸馏 [22:15] Yao Shunyu / Grace Shao: 聪明蒸馏 [27:35] Grace Shao: 大家终究只是想照顾好家庭、过上好日子，甚至就是看你给的薪酬待遇（Package）到底如何 [30:00] Joe Weisenthal: AI 精神病 [30:00] Joe Weisenthal: 机舱/引擎室 [50:00] Grace Shao: 你怎么像个穴居原始人（caveman）一样？难道不知道直接扫桌上的二维码吗？

[9:21] Grace Hsiao: They kind of like did one for the team... like look, guys, like you can actually do this. [22:30] Grace Hsiao: Dumb distillation is something I think most of us who are non-technical think about... you take a thousand queries, you take the answers of whatever Claude gives you, and then you force copy that into your model... Smart distillation is like using the frontier model almost as a partner to help you with the judgment for the evaluation and even the data labeling itself.

[0:00] Whatever your goal, trade show giveaways, client gifts, or team gear, 4imprint has the promo products to match. With thousands of options, from apparel and drinkware to tech and totes, it's easy to find the right fit for your brand and budget, with standout choices at every price point. And with their 360-degree guarantee, you can be 4imprint certain your order will show up just right, right on time. Explore more at 4imprint.com. 4imprint. 4certain. The thing about AI for business? It may not automatically fit the way your business works. At IBM, we've seen this firsthand. But by embedding AI across HR, IT, and procurement processes, we've reduced costs by millions, slash repetitive tasks, and freed thousands of hours for strategic work. Now we're helping companies get smarter by putting AI where it actually pays off, deep in the work that moves the business. Let's create smarter business, IBM. When you're running a business, the best days are the ones where priorities stay on track. For midsize and large companies, risk can affect multiple parts of the organization at once, from property and liability to cyber and regulatory challenges. At that level managing risk becomes an ongoing discipline At the Hartford the focus is on helping businesses manage risk before it turns into something more disruptive And when losses do happen, that work is paired with insurance coverage shaped by years of underwriting, risk engineering, and claims experience. Learn more at thehartford.com slash risk mitigation. Policies provided by Hartford Fire Insurance Company and its property and casualty affiliates, Hartford, Connecticut. Bloomberg Audio Studios. Podcasts. Radio. News. Hello and welcome to another episode of the Odd Lots podcast. I'm Joe Weisenthal. And I'm Tracy Allaway. Tracy, I love being in Hong Kong. I love it here so much. I love it here so much. I would come here a few times a year if we could. I'm sure you would. I lived here for like, I guess, almost four years. So it's kind of weird coming back. But Hong Kong has a lot of pluses, like great food, great weather for most of the year beaches I once heard someone describe it as Manhattan meets Maui which I think is like pretty accurate It so so nice The weather is actually not great It not great right now It literally this week We come during I guess it a monsoon season right Yeah, it's the rainy season. But oh, well, I like thunderstorms, so I'm enjoying it. Yeah, I'm enjoying it too. Anyway, one thing that has changed since the last time you were in Hong Kong, you left in 2022. We weren't doing as many AI episodes in those days, to say the least. We definitely weren't. In fact, so I remember one of the big stories when I was here in Hong Kong in 2020 was China's tech crackdown. Right. That's right. That's right. Right. And like there was all this concern about whether or not the crackdown was going to destroy China's entrepreneurial spirit. I'm doing air quotes on a podcast. I don't know why. But fast forward six years and there's entrepreneurism basically everywhere. And we talk a lot about how China is producing all these new AI models. OK, can I say like I know very little. I mean, I know very little about AI, but I know even less about Chinese AI. But here are some of my general impressions, which is, hey, it seems like there are so many open source. Okay so I know they largely open source It seems like every random company you see like some toothpaste company and they have produced an LLM So I very curious like how they making money on it I also get the impression like you know the heads of American AI labs speak in these like sort of quasi mystical terms, etc. It doesn't feel quite the same here, where it feels like a bit more of like yet another technology. But I'm glad you brought up the point about the tech crackdown, because at the time, the whole story was like, oh, there needs to be less focus on sort of digital tech and more focus on hard tech, which has been done extremely. That's been an extraordinarily successful endeavor. And then my last impression, though, is that since the release of ChatGPT in late 2022, that was the moment it's like, no, we really have to also compete on sort of this next era of software and sort of consumer facing tech breakthroughs. Yeah. Overall, the AI scene in China feels much more utilitarian to me. It's more about like, the big companies, the Tencents, the Alibaba sort of using AI for their existing business models rather than this existential thing, which it is in the US where like AI is the business. That's just it. Right? Yeah, that's exactly. AI is [5:00] It sort of sits in the middle of what you would call like software and hard tech because we consume it through the browser, right? Sort of the same way or in many cases through the browser, the same way that we would go to an Amazon or an online gaming or something like that. But it's clearly – it's a scientific endeavor. And so it sort of is this blend. And then you have to figure China is so far ahead of the US when it comes to things like robotics and EVs and batteries. And one thing I don't know anything about is the degree to which that melding of hardware capabilities with AI capabilities, how that influences the direction of the development of the AI tech. Yeah, I'm also very interested in like the capital stack for Chinese companies, because over in the US, we all know that people are flinging money at anything with the word AI in it. But in China, it's very different. I get the impression that it's like much harder to raise enormous sums of capital. And so I'm very curious how that limited capital actually influences the development of these models and the tech. I think it's safe to say that both of us have a lot of impressions. Yes. Right. I feel like how many times in this intro I was like I get the impression but I actually have no idea So that is a good reason to actually bring in our guest someone who has more than quote impressions unquote about the AI tech scene We going to be speaking to someone whose newsletter I a big fan of and everyone should read. We're going to be speaking to the perfect guest, Grace Hsiao. She's an independent AI researcher, and she has a great substat called AI Prom, and she joins us here in our Hong Kong office. So, Grace, thank you so much for coming on OddLots. Thank you so much for having me, Joe and Tracy. How did we do on our, quote, impressions, unquote? Those are pretty accurate impressions, I think. Okay, good. That was the episode. Like, let's start at a basic level. So the big impression, the one that everyone knows, is Chinese models are open source versus the closed frontier models of the U.S. Why did it develop that way? Yeah, I think people like to think of these mystical reasons. But really, it was a very pragmatic business reason to start with. To start with, a lot of the labs have cited that, you know, for Western companies or Western developers to trust them, they needed to open source their models to build that trust and credibility So in many ways it a branding decision Then on top of that I think you can see it as a philosophical drive You know the founder of DeepSeek Liang Wenfeng has openly said he wants open source his most frontier research to really help propel the whole industry as a whole And that kind of R sharing has now formed a layer for the whole ecosystem where each of the labs kind of integrate each other's kind of breakthroughs. You know, you see them congratulating each other, even on X, when they have new models announced. So you can say it's a bit more collegial. I wouldn't say they're not competing, though. However, because of the compute constraint they're faced with, talent constraint and the capital constraint you mentioned, they are a lot more conscious with where they want to put their money, where they want to put their time in R&D. And all of that forms the basis of a strong open source ecosystem. Is the culture as pro-sharing and pro-open source as it was even two years ago? Now, the Deep Seek moment was right around Trump's inauguration in early 2025. So about a year and a half ago. Since then, has the culture stayed the same or has that sort of competition bug, that intense competition bug that we know among American AI labs, has it spread to the Chinese labs at all I think the sharing is an unintentional result rather than an intentional effort in the beginning to even start with They are for sure extremely competitive And we all know the word involution. So China AI is dueting as well. That means there's involution in this ecosystem as well. However, I think bringing up DeepSeek, DeepSeek plays a very interesting role in the whole ecosystem. Like you mentioned, V3 propel the whole industry forward. everyone kind of start taking China AI more seriously. You know, it brought a lot of interest from investors globally back into the internet companies that Tracy mentioned, you know, prior to that, there was a bit of a slump for three to five years. However, you know, Zipu, ZAI is now publicly listed in Hong Kong. Minimax is publicly listed in Hong Kong. Moonshot is, you know, in preparation to go public next year. They are competing with each other to capture market share to capture developer mind share. But DeepSea plays an interesting role. I want to bring it back to DeepSea V4. So V4, you know, on the surface, you know, people said, OK, it wasn't as maybe impressive on evals and performance. It didn't catch up with the most frontier labs in the US, whatnot. Right. But it was a very interesting move because what I heard from researchers on the ground in Beijing was that the lab actually delivered. [10:00] laid their release for about three to four months because they wanted to re-engineer a lot of the inference onto Huawei. So I'm not saying this completely replaces NVIDIA or CUDA, not at all, because if you ask any developers, they still want to use CUDA if they can. However, it was the first effort to really, I kind of like did one for the team, like they kind of like put the resources. It's supposed to be like a signal, basically. Yeah. We're doing this all on like a Chinese stack. Yeah, they were like, look, guys, like you can actually do this. And they became a shared foundation layer for China's model ecosystem. So because again, everything is open source and open weight, other labs were able to study what they did to actually start inferencing on Huawei stack. And I think that was the first step, whether it's signaling or actually, you know, very pragmatic reason to start shifting some reliance on, you know, the China AI stack. Aside from DeepSeek, can you kind of describe the differences or what China is trying to do on the actual frontier side? Because there are some. I think if you really have to look at the ecosystem we can kind of put aside the big tech for now But looking at maybe the four most relevant startup labs DeepSeek Moonshot who has Kimmy Zia who has GLM and then Minimax they are still probably the most committed to frontier research research. However, because the constraint we mentioned that they face, whether it's compute, whether it's capital, or even frankly, talent, they have decided out of necessity to basically each focus on a different vertical in capturing a different kind of business share. So ZAI is very focused on coding capabilities. So if anything, their GLM plan is much more similar to maybe what you think of Claude, Claude Co, or Claude Code, etc. A codex, that kind of product. And then you look So Minimax, they're really focused on the multimodality capabilities. Moonshot, they're really focused on agents. And DeepSeek, again, really is just focused on pushing the frontier and kind of trying to play catch up and push the Chinese ecosystem as fast as possible. It's really crazy to look at some of the ones that have already gone public here. And just to put in, so Minimax is public, and in US dollar terms, it's a $20 billion company. I mean there are people in the U who have done nothing but publish a paper on archive who do not even have a product yet who have probably VC implied valuations over billion How do they make money You know again, open source, okay, like in these four models that you named, do they have different thoughts on how they plan to make money or different business models? Yeah. So China's VC space in general has not been that vibrant, frankly, since internet crackdown and a lot of USD funds did exit, you know, three to five years ago, with Sequoia being maybe the most like high profile, right? Like we all remember that. Now, people forget, even in 2022, a lot of these labs that we just talked about, they were struggling to even raise money, you know, raise capital. And a lot of them spun out of academic institutions, you know, you mentioned their rally anywhere roughly between 20 to 30 billion right now, but they went public between like six to 8 billion, that's like kind of tiny compared to American valuations right now. However, they are actually making money. You know, the publicly disclosed information, I think, from Minimax and Jipu indicates that they were making just as much like in their last month, they've made the same amount of money last month as they did last year essentially And their end of year AAR projection is anywhere between one to 1 billion right now So they are making money And how Well just because they open source doesn mean they don make money I think people forget you know, we had open source softwares before as well. People are paying for managed services. And when you're paying for an API through ZAI or Minimax, whatnot, you basically don't have to self host, you don't have to get your own GPU, you don't have to get you figure out your own compute, you don't have to figure out your own guardrails or deployment or security or monitoring whatnot, right? So just to be clear, you can self-host all of these models, but for the most part, they do offer that inference part of the stack, and that is a profit center for them. Yes, exactly. Whether you're planning a big tech event, launching a new campaign, or just stocking up on team gear, finding the right promotional products makes all the difference. 4imprint offers thousands of options, from on-trend apparel and premium drinkware to tech, totes, and giveaways, so you can find the right fit for any audience, purpose, or budget. You can customize it all. [15:00] your message, your look, and many items come with no setup charge to help you save. And if you're really watching the bottom line, you'll find standout choices at every price point so you can make a real impact while staying on budget. Plus, you'll get expert help, fast turnaround times, and their 360-degree guarantee so you can be 4imprint certain your order will arrive on time and look exactly right. Whatever your goal, 4imprint makes it easy to find your perfect promo match. Explore the possibilities today at 4imprint.com. 4imprint, 4certain. Support for the show comes from public.com. If you're actively involved in your portfolio, you probably catch yourself repeating the same actions. Buying the dip, manually sweeping idle cash, putting on a hedge. On public, you can now create AI agents that handle all these tasks on your behalf. Just describe what you want to do in plain English. Like, if the VIX hits 25, buy a put option on the S&P 500. or if my cash balance goes above $20,000, move the excess into my direct index. You approve the workflow and your agent handles the rest. Monitoring the market watching for your conditions and executing your strategies exactly as defined An investing platform driven by your intent not just your clicks You can also get full read and write access to your account via the public API Go to public slash market and fund your account in five minutes or less. That's public.com slash market. Paid for by Public Investing. Brokered services by Open to the Public Investing, Inc., member FINRA and SIPC. Advisory services by Public Advisors, LLC, SEC Registered Advisor. Complete disclosures available at public.com slash disclosures. Let's talk about health care for a second. It doesn't always work the way people expect it to. If you've ever waited on a prescription refill or had a hard time getting the care you needed, you know the feeling. The system should just work better for everyone. That's exactly what the people at Optum are trying to do every day. They're a health care company linking patient care and pharmacy services and using data and technology to drive the whole system so care is connected, not complicated. for patients and providers. Things like making it easier to get care that looks at the whole person, from primary care doctors to mental health support and even in-home care, and then using technology to make sure they all work together. Technology designed to help doctors spend less time on busy work and more time with their patients In those prescriptions Optum is working to bring costs down save patients money and make it easier to get refills Little by little Optum is helping make healthcare work as one for everyone. Head to business.optum.com to see how. All right, so there seem to be two major constraints on Chinese AI. Maybe energy is a constraint as well. We should talk about that. But there's the capital issue. So not as much capital available or people aren't flinging it at AI companies the way they are in the U.S. And then secondly, there are the export controls on chips. And we talked about that a little bit. But can you describe how those controls are actually, I guess, influencing the development of the models themselves and like, I guess, optimization? So some of these big tech are actually even buying out the contracts that data centers have with some of these labs. And, you know, they are taking over the compute. So these labs essentially are now optimizing for the highest quality inference demand, if that makes sense. They don't actually have enough even supply. They don't have enough compute power to even meet the demand that's coming through. So that's on how they're servicing clients. how they changing I guess or how they optimizing their training is that a lot of them are really focused on the post And this goes back to so you know how OpenAI has like three buckets where they chuck money at There's like the R&D, there's pre-training, there's post-training. R&D, a lot of times, a lot of money is spent, but say like one out of 10 things stick. But you need a lot of compute and resource and people to be figuring out where to go. For a lot of these labs in China, they frankly don't have that luxury. So they've even given me a metaphor and said, it's kind of like knowing what the answer to the homework is and working backwards. So they will wait till the frontier labs to come out with where the right direction is for the next frontier model, and they will work backwards and actually focus all their resources on post-training. So with post-training, they will optimize a lot of times the data they collect. For example, if a data provider like Mercore provides a very, very niche set of data set for like an open AI or whatnot, maybe they would charge them $10-20 million. The Chinese lab will wait out that exclusivity contract, three to six months time, let's say, and then pay a fraction, if not like a tenth of that price, the same data set. And that kind of plays into that like six to nine month leg that we hear about as well. That's really interesting. Let's talk about energy then, because the story in the US is that... [20:00] Electricity is really the big constraint on AI use. And, you know, you got to find a data center that has an electricity hookup and it has to be reliant and all of that. And it seems to be in short supply. Is it a similar story in China? honestly energy is probably not the biggest bottleneck right now in china and i think people like to say well some people like to say oh somehow the chinese government had foresight on the ai boom driving like the energy consumption but definitely not i think people forget that china's economic growth over the last three to four decades also meant a rise of urbanization and a lot of the cities that you know we are visiting these days like at least westerns are visiting like beijing Shanghai, Shenzhen with all these robots and EVs or whatnot. These are all really urbanized within the last two, three decades. And because of that, the grid is very new. And because of that, the government already foresaw that there was going to be a increase in energy demand. And, you know, so a lot of the energy plants, you know, the solar plants, hydro plants, whatnot, were actually built out in anticipation for that. Now, obviously, this has coincided with now the AI boom, and it's really helped out. Beyond that you know China has an advantage in the fact that they can actually drive top mandates and provincial governments will follow suit This is something quite unique to China because it not like decided by each state So when they pushed out the East State of West Compute, where it's basically a top-down initiative where they built a ton of renewable energy for cheap in rural mountainous areas in Guizhou province, like even Xinjiang, Inner Mongolia, Sichuan, you know. So those were like very easily executed, frankly. And then 90 percent of the population actually sit on the eastern coastal lines. Like we think about Beijing, Tianjin, Shanghai, Shenzhen. That's all on east. So that's where the data comes from. So that kind of optimization has also really helped them, you know, with the low that is the demand right now. I want to get back to something you said. So first of all, just to clarify, you mentioned companies like Mercore that sell proprietary data that they are able to collect and manufacture in various ways. Then they sell it to an open AI. So a company like Mercore will hire a bunch of people to, say, build PowerPoints, and then they'll collect the data on how they do that. And that is fresh data that they can sell. So those have exclusivity windows after which they can then sell them to anyone I not saying Merkur specifically but supposedly there are these data providers that do the sell and they have exclusivity windows And then the Chinese labs kind of weigh that out so they can pay like maybe a million dollars versus like 10 million for the same data set. So this gets to something generally speaking, which is that people are around the world correctly, like quite impressed by how high quality the Chinese models are, even if they're behind. But then you have things like that. And then you also have accusations from the likes of Anthropic that they're distilling models and that they're finding ways to collect the outputs of American models for training. So then you could say, well, yes, sure. That's like it's great. this open source model and it can stay close to the edge. But then the counter is that they can only be so advanced because there is this extremely capital intensive closed source model in the US that's really establishing the frontier and that these Chinese companies wouldn't be anywhere near where they were if they weren't sort of, I guess you would say, drafting off the American labs. Yeah, I think the compute constraint and the capital constraint is real. And frankly, like no one hiding that or pretending that that not an issue for them right now Like DeepSea has openly said they even were struggling right Like they needed more compute I think on the distillation allegations or accusation it is quite interesting Like recently I been thinking about this a lot and thinking about what it means for distillation and what it means for the models to catch up, right? So there was this one quote from Yao Shunyu, who is a Google DeepMind researcher. He said, there is smart distillation and dumb distillation. Dumb distillation is something I think most of us were frankly non-technical think about. It's like, okay, you take like a thousand queries, you take the answers of whatever Claude gives you, right? And then you kind of force copy that into your said model. And then you forcefully make them basically like get the exact same answer. Smart distillation is like you using the frontier model almost as a partner to help you with the judgment for the evaluation and even the data labeling itself. So you're using it as almost a teacher for your own model. It guides it a little bit versus really copy pasting the answer for that makes sense. And that part of it is frankly not that unethical or like, you know, that frown upon right now, because that is what enterprises do when they're fine tuning. So it's all kind of a bit of a murky area, to be honest. Okay. So you mentioned data just then. Talk to us about what the Chinese [25:00] data set actually looks like? Because I imagine if you're a Tencent, I mean, you've got WeChat, right? That must be a whole load of data on which to actually like build your AI. But on the other hand, I imagine like there are some restrictions around the internet, obviously. What does it actually look like here? So I'll actually split that into two parts. On the data itself, people often think China is so data intensive and you just have a vast amount of data to use for AI training. However, actually, people forget, again, China's enterprise build out or, you know, whatever the knowledge work economy is very new and not as sophisticated, frankly, as American ecosystem or the Western ecosystem, if you have to put it that way. So data is often unstructured and data, thus a lot of the specific needs for, you know, the kind of training we're seeing today is not as vibrant or the data ecosystem is not as sophisticated as what American data providers can provide, such as Mercore, like we just mentioned. Now, on the big tech side, it's a bit interesting. So I'm glad you brought up Tencent, because Tencent actually just announced last week that they are working in the works of creating an agent that can be plugged into WeChat This has been very controversial and it has actually had a lot of pushback even internally because WeChat product manager Alan Zhang has been famously or notoriously known to be kind of hard to work with if you want to push something within WeChat because he's so protective of that user experience. It's his baby, right? And Tencent, like you said, has WeChat, which is a super app, has more than 1.4 billion MAU globally, like mostly 1.3 billion people in China and the Chinese Aspera globally or people who work with China. That is immense value, but the risk and compliance risk of potentially an agent going rogue within that chatbot or that of agent going rogue in executing, you know, whether it's a purchase or whatnot, is very, that risk is very high. So they've been working on that. And on top of that, Tencent itself has been lagging behind compared to other big tech in their own proprietary models. And they've really been trying to really play catch up. They actually last year poached someone from OpenAI, who is a researcher called Yao Shun Yu as well, has the same name as the other researcher we just mentioned, to lead this whole initiative. And their whole goal is to basically build a Tencent agent native model And that is their biggest goal Because at the end of the day like you said in the very beginning their goal is to optimize existing businesses already and bring AI to the mass consumer as fast as they can. You know, you mentioned poaching a researcher from OpenAI, and it's like, the way I see it, AI will definitely be built by the Chinese. The question is whether it will be built by the Chinese working in the American labs or whether it will be built by Chinese working in Chinese labs. Has there been a gathering of steam of researchers from that had been working at American labs going to Chinese labs, or are they still sort of one off and somewhat rare? I think even during the Internet era, we saw a lot of Chinese nationals or Chinese ethnic people returning to China. Right. I think this it's easy to blanket statement as geopolitical headwinds. People are scared. But realistically, I think most people are just trying to take care their families and live a good life, right? I hate to sound so crass about it, but you know, sometimes it's what your package looks like. And to overgeneralize, I've heard from many researchers say, look, if my wife is a lawyer in China, my wife is a nurse in China, my wife is a teacher in China, that kind of employment opportunity is very, very hard to actually transfer to a new market If I can get a similar package and a growth opportunity in one of the big labs in China I will pick that over living in the US And on top of that I think that something lost in the nuance is my parents immigrated to North America 30 plus years ago. It was a very clean cut, like quality of life. It's just like objectively better in any city in North America compared to any city in China. Now that's kind of a personal debate, right? Because it depends on what you really value. If you want to be close to city center, you want that fast paced, like techno, EV, futuristic lifestyle, China actually gives that to you. And then on top of that, if you want to be close to your family, it's a very personal reason. So I met a lot of researchers actually decided to come back to China or this part of the world simply because they wanted to do it for personal family reasons. Are they paid as much as they are in the US? Because we get headlines all the time about, you know, so and so is joining whatever company and people treat that news like sports stars, right? Like teams trading their best players. Is it a similar thing here? I think you definitely get less of that sports star vibe or mentality here. They're still getting paid like hefty amounts, how much they don't usually disclose, but at least even in the [30:00] dense product manager can make just as much as a meta product manager. Similarly, if you're like an average AI researcher, you're probably making a similar amount. Although the star star players, like the ones that are signing 100 million bonuses, I don't know if we had anything like that big like in China. But look, they made their money with the IPOs. They made their money recently with all that this AI boom. It's just on a maybe slightly smaller scale. Doesn't mean that they're not making much more than the frankly average person. Tracy, can I say something that might be sort of a sacrilege for a podcast host to say? Okay. I'm only speaking for myself here. I'm not necessarily speaking for the team. But it occurred to me, like, I'm not really sure if I'd be that interested in, say, getting the CEO of like an American AI lab on the podcast as a guest. I don't know what I would ask them because like, like, do I really want to hear like Sam Altman or Demise Hissabes or whatever, like the future of work and all this stuff or like these, all the big, you know, I love doing AI episodes. I just feel like at that level I would rather talk to sort of like someone in actually the engine room rather than this sort of big picture person who may have some degree of AI psychosis and just like as on it speaks in like the biggest generalities OK, well, now Sam Altman needs to like invite himself on the show just to test your your commitment to not having AI CEOs. No, I would. I would do it. But I like let's just agree here that if we ever like get one of the really big like lab CEOs, I should ask the very like sort of mundane questions about operations and not like, what are we all going to do? And, you know, what is the meaning of life going to be when we don't have jobs? Because I'm so sick of those conversations. They may be important at some point. But, Grace, I'm sort of curious from your perception, like it does feel like the heads of the American AI labs have some degree of AI psychosis themselves. Either they talk about all white collar employment is going to disappear or that they're going to build a monster that if done wrong is going to be out of control and that they're not, you know, it'll escape the sandbox. Is there the same sort of existential discourse in the Chinese AI community? Yeah I think to start with in the AI community themselves I would say people are a lot more pragmatic And I think recently I was talking to Nathan Lambert who is open source researcher who just came to China and visited all the labs He said, look, I was shocked to see majority of labs are so young, as in like a lot of the researchers are still students. A lot of them are interns and the core research teams are maybe led by a handful of people. And then these people are academics by training. So maybe they're a bit less commercial. Maybe you can say they're less like sophisticated to manipulate the market, whatever you want to call them. So I definitely feel like there's less of that kind of psychosis or high level narrative going around. However, I would say that there is obviously anxiety from the public in some degree, not as much of a pushback. But recently there was a very interesting court case in Hangzhou, which is, you know, home to Alibaba and a lot of these AI labs. basically a company tried to like lay off a person saying you are being replaced with AI. And the court literally ruled say that is not allowed. And no company can use AI as an excuse to lay off or replace or even cut short their contract time So that was a really swift reaction from regulators And I think it really did serve as a calming factor for the public Obviously I also think I want to preface the fact that the knowledge worker economy makes up less of the overall economy in China as well. So that kind of fear maybe doesn't feel as imminent. But that conversation is being had. But I do think in Asia in general, not only in China, you look at South Korea, Singapore, all these countries are approaching AI in a very pragmatic way. And the tired moms are trying to train up the kids to be AI native. The students are trying to train themselves up to be AI native. People are preparing for the future versus pushing back on the future. That's interesting. You know, in the US, obviously, companies announced that they're laying people off and they cite AI even frequently when there's no evidence that AI had anything to do with it. So it's interesting they're not allowed to do it. In defense of American AI labs, Most of them lose a lot of money, and yet they actually do spend, at least the big ones, spend quite a decent amount on so-called alignment research, safety research, making sure the AIs don't go rogue, et cetera. How big of a part of the Chinese labs, how much do they spend on, I guess, what the American labs would put into the safety bucket? I do have to say I'm not a policy expert. [35:00] So I don't work with a lot of the safety people as much, but there are organizations in China that are definitely working like the regulators as well as the private sector working together. And for some context in China, there are various moving parts in the government. There's MIT, the CAC, etc. These agencies, basically, some are to propel economic development. So in this case, AI diffusion, the whole idea of AI plus, AI plus every single sector you can think of. Some act more like a guardrail as a protector. So they are working hand in hand. And on top of that, every single AI, Gen AI application, as well as LLM company have to go through the national registry in China. So they actually disclose what is being trained, what is the potential risk. That said, I think right now, no one really knows what the real impact of AI will be on economy. But, you know, definitely that fear mongering narrative is not mainstream in China. Thank you. Any items come with no setup charge to help you save. And if you're really watching the bottom line, you'll find standout choices at every price point so you can make a real impact while staying on budget. Plus, you'll get expert help, fast turnaround times, and their 360-degree guarantee. So you can be 4imprint certain your order will arrive on time and look exactly right. Whatever your goal, 4imprint makes it easy to find your perfect promo match. Explore the possibilities today at 4imprint.com. for imprint for certain support for the show comes from public.com if you're actively involved in your portfolio you probably catch yourself repeating the same actions buying the dip manually sweeping idle cash putting on a hedge on public you can now create ai agents that handle all these tasks on your behalf just describe what you want to do in plain english Like if the VIX hits 25 buy a put option on the S 500 Or if my cash balance goes above move the excess into my direct index You approve of the workflow, and your agent handles the rest. Monitoring the market, watching for your conditions, and executing your strategies exactly as defined. An investing platform driven by your intent, not just your clicks. You can also get full read and write access to your account via the public API. Go to public.com slash market and fund your account in five minutes or less. That's public.com slash market. Paid for by Public Investing. Brokered services by Open to the Public Investing, Inc., member FINRA and SIPC. Advisory services by Public Advisors, LLC, SEC Registered Advisor. Complete disclosures available at public.com slash disclosures. Let's talk about health care for a second. It doesn't always work the way people expect it to. If you've ever waited on a prescription refill or had a hard time getting the care you needed, you know the feeling. The system should just work better for everyone. That's exactly what the people at Optum are trying to do every day. They're a healthcare company linking patient care and pharmacy services and using data and technology to drive the whole system so care is connected not complicated for patients and providers Things like making it easier to get care that looks at the whole person from primary care doctors to mental health support and even in care and then using technology to make sure they all work together. Technology designed to help doctors spend less time on busy work and more time with their patients. And those prescriptions, Optum is working to bring costs down, save patients money and make it easier to get refills. Little by little, Optum is helping make health care work as one for everyone. Head to business.optum.com to see how. How would you describe, I guess, the approach of the Chinese government to AI in general? Because it feels like the tradeoff for maybe not being on the cutting edge with frontier models is, well, you're further along with sovereign AI and the government maybe has like a better handle on what the labs are actually doing. The Chinese government probably sees this in a much more pragmatic way. You know, just like how there was Internet Plus policy 15 years ago, there's now an AI Plus policy. When DeepSeek moment took off, there was a frenzy of private companies, even in home appliances, trying to embed DeepSeek. [40:00] You see, I'm just like, what is a AI vacuum going to do for you or AI like electric toothbrush going to do for you? It was wild, right? But the government picked that up. And I think what the Chinese government, going back to what we talked about earlier, is that they have the advantage of having the ability to push things down from top down. And at the very high level, they're seeing AI as an economic driver to propel maybe efficiency to address some of the labor shortage that is coming as the population continues to age and decline. It also addresses a lot of issues where a lot of the young people don't want to work in manufacturing roles. They want to be automated, actually, and they want to work in urban areas. So they have that now. And then each of the provincial governments will take that as kind of like a KPI. They're like, all right, let's go act like VCs, essentially, and go find, you know, future deep seeks and fund them. However, how much these companies want to take government money is a different discussion. A lot of them would then even support them by providing, you know, infrastructure like buildings, offices, even like some of them are heard like dormitory for these young entrepreneurs and then give them money and capital try things out There also these AI pilot zones being rolled out across the country I think now about 11 or 12 of them where you know people can try out new AI products I met with the largest AI developer community founder in China a couple of weeks ago, and she was saying there's more than a couple hundred thousand developers in this ecosystem, and they are working with regulators and the private sector. So if, say, ByteDance have a new product, they might go to them first and say, can you try this out? and then they will report and debug and see what's happening and then tell the whoever local government that's funding them with providing with infrastructure say hey this product might come out do you want to be part of it do you want to give it money do you want to provide it with whatever resource you want to so there is kind of this like cohesive ecosystem where they kind of all dance together how much these companies actually want to take state money i think that's debatable However, as AI and robotics become more and more sensitive and being recognized as not only an economic driver, but potentially a military use or geopolitical, I guess, talking point, at this point, it is becoming more and more nationalized, not only in China, but globally in the U.S. and so on. Robotics is obviously an area where China is just straight up ahead of the United States or at least according to all the videos on my Twitter and Instagram feed of humanoid robots and so forth How much does you know when we were all kids when we thought of like AI I think we thought of robots, right? We thought about the T-1000 or some version of it. And now when we think of AI, most people think of chatbots, but that's just one aspect of AI. For the advanced labs, whether we're talking about the DeepSeek, so the Minimaxes or GLM and so forth, are they actively working hand in hand with some of the unitaries and advanced robotics companies to figure out how you can actually have the true AI robot of the future? Yeah, I think China, having been the manufacturing hub of like literally everything under the sun over the last like three decades, has definitely an advantage of owning all the supply chain. And it's like not only just owning the supply chain, but, you know, there are literally regions where that whole supply from raw material to like the end product from OEM is all within like, say, 50 kilometers of each other. So what we seeing is a lot of American investors and entrepreneurs coming into China to kind of get a sense of that And like you mentioned because we been so fixed in software I think China having a very strong hardware background is now thinking about how can we actually integrate the software into the hardware? How ready that is to the mass market? I frankly don't think it's really there yet. So recently, I just met with some robotic companies. They actually can't just plug in a mini Macs, you know, that's like for them, they need to actually get physical data. This is where like, you know, now all the hype is on world models, physical AI, you know, that is a complete different set of kind of technology, essentially, where without the 3D data that these models need right now, the bottleneck right now is that, you know, these hardwares, these humanoids, quadrupeds, dogs, whatever you want to call them, they cannot be powered by LLMs. That's number one. Number two is despite that China being very strong on hardware, the bottleneck is actually a lot of times in the integration as well as the battery solutions. You know, you think of China having very strong battery solutions, but most of these gadgets can't last more than like, say, two hours. And there's no one that's really come out with a better solution so far. What I've seen the most creative thing so far is like, you know, those glasses you wear, like the meta glasses, they kind of die within two hours. But China, like iFly Tech or Rockhead, that's [45:00] kind of a newer player startup, they created these battery capsules where you can just like stick onto your glasses. It's very lightweight, doesn't really affect your user experience. And that's actually able to kind of extend it by a few hours. So to go back to your question, is China trying to do physical AI? Definitely. What is their edge? I think it's still in manufacturing. Is their software good enough? I don't think anyone really has good enough software right now as of now. Wait, I'm just going to press you on this. So if we fast forward 10 years, what would you say is most likely to be China's comparative advantage? Is it the cheap, open source, super optimized models? Is it software, AI software that's integrated with industry and existing business? Or is it robotics and the hardware side of AI? 10 years is a long time. I know. Sorry. We want to challenge you. A lot of these companies didn't exist 10 years ago. We're not even five years ago. Yeah, that's fair. Okay, I can shorten the time frame in three years. All right. So don't chase me down if I'm wrong in three years. But I think, you know, there's two parts. One is I think I agree with the hardware side China is definitely going to have I think more breakthroughs and have a lot of edge Not only is the supply chain all domestically there I think something overlooked by people is the fact that a lot of the know is also there And that's not easy to transfer overnight. You know, Patrick McGee's book recently in his Apple book saying how Apple tried to move this whole supply chain to India. The biggest bottleneck is actually these highly skilled laborist jobs that actually are so technical that cannot be even trained in one generation. It took decades to really train up the local community, labor force, whatnot. So that's still there. Now, because of that ecosystem, a lot of these robots, home appliances, whatnot, these tech gadgets are produced at less than 50% of the cost of where you could produce at anywhere else in the world. They are also extremely innovative. I've talked to people at EV companies, just for example, to ship out a new model from ideation to production to hitting the floors. That takes maybe less than 15 months. Wow. But for a traditional OEM, wait for at least three to five years, right? So there's the hardware side. I think another very underappreciated fact on the Chinese open source model is that people don realize So a year ago when I spoke to startups in silicon valley they were the most cost conscious frankly less compliance conscious and gives very little care about geopolitics they were building on top of quit now it's actually a lot of ai native american enterprises building on chinese open source because we're seeing headlines on the roi is not like showing it's extremely expensive for these token maxing projects whatnot So Harvey, Cursor, they've talked about using a hybrid model where they will build majority on GLM or Kimi. But kind of like what we talked about earlier, where they use like Opus to act as a judge or a guidance. So I think that's something where we continue to see. And these companies are generating a lot of revenue. And going back to the fact that like a lot of them don't even have enough capability to support the demand that's coming through. That's such a fascinating idea, and it makes a lot of sense that at the application layer, that probably a lot of different models can go into it. By the way, I was talking to someone at a dinner recently, and he said he thinks that AI writing will get a lot better when AI is embedded in humanoid robots because then we have this sort of groundedness in the real world where I have no idea if this is true But he said the reason his theory was that the reason why AI writing is still so weird is because it in this disembodied data centers And that as soon as they're really in robots, then they'll have a sort of real-world groundedness. All right, I have one last question. And, you know, after Open Claw came out, I started seeing again, I my entire Twitter feed and Instagram feed. I have done this to myself, but all I all I do all day is consume Chinese propaganda. But I started seeing all these videos, like all these grandmothers and stuff like setting up their claws and stuff like that. And I saw these videos and I said to myself, I just don't believe this. I think this is fake news. I do not actually believe that there's all these 80 year old grandmas or whatever, really excited about setting up their open claw or whatever. Are those real? Like, what's the deal with that? I think that was definitely a bit of a hype. I knew it. No, no. But I will say there were grandmas lining up to get it done. Okay. I'll answer this twofold. On the kind of the surface is I think Chinese aunties, uncles, whatnot, they are just much more open to technology because, you know, you go around, whether it's by force or by nature, you know, you can't really navigate modern Chinese life without being on Ali [50:00] You really can't, you can't like buy Starbucks in Beijing without like having WePay. After COVID, I went back to, I think, Shanghai for the first time. And I was sitting at a restaurant just like waiting for someone to like help me order food. No one came because they're just like, why are you so like, why are you a caveman? Don't you know how to like scan the QR code on your table? Okay, so that tangent aside, I think the overall optimism around technology is very different from the West. because in the last 20, 30 years, a lot of rural areas in China literally could not access resources, information, goods, whatever that, you know, like big cities could not until these super apps came about. So a lot of people don't have TVs in their homes and they live in a village and their maybe annual household income is like a thousand dollars, but they will have a smartphone and that smartphone will be able to actually enable them to get micro loans, to purchase goods, you know, to help their kids access information online, whatever that is. So technology is very much kind of accepted and respected And actually like big tech is love Like if you work for one of big tech you are like a pride of the family So there that very cultural aspect of it Then is like going back to super apps So I think the open claw frenzy was interesting because some people say it was the first agent that, you know, Chinese people could get their hands on, like a Western agent that they can't get their hands on because, you know, Anthropik and OpenAid doesn't actually operate in China. You can't access that. So when Tencent and Alibaba tried to embed open claw products into their own like series of products or business products, whatever offerings, it got people really excited. And because of these super app models that they have, it was a very natural way for people to access them. There's a functional adjacency to, you know, the search bar and then opening up an open claw and then trying to like run, like, you know, manage your mini programs within Tencent, WeChat, and then trying to order something. So all of that kind of took off. But that said, actually, the Chinese government, again, regulators acted very swiftly. In the beginning, I think local government like Wuxi tried to even encourage local businesses to embrace open claw. But the Beijing government immediately said, guys, actually be very, very careful of your data privacy your security Banks shouldn do this SOEs shouldn do this Be mindful of what you doing with this technology And then the big tech kind of rolled back a bit of their marketing And you can see actually hilariously there were advertisements for helping these aunties and uncles how to uninstall these open cloth on their gadgets. So anyway, that's a bit of a kind of background on that. All right, Graysha, thank you so much for coming on OddLots. Great to connect with you here in Hong Kong. And we'll have you back in three years. No, hopefully before then, We'll have you back at least certainly in three years to see how your predictions held up. Thanks so much. Tracy, that was fun. There was a lot of interesting ideas in a fairly short conversation. But one thing specifically, and then that like sort of stands out to me, is thinking about some of these application companies, how much it makes sense for them to sort of, yeah, they'll use like a state of the art American closed model for like some of the work. But then other just, you know, almost as capable models underneath. So you have like a legal AI app like a Harvey or something I sure some of these open source models work well for some tasks within that context and how much it makes sense to sort of combine them under one app layer Yeah. Well, you don't need the cutting edge model for everything. Right. But like if you can get some of the cutting edge model combined with like the cheapness of the open source thing, like that seems like a pretty good deal for a lot of companies. Yeah, definitely. I also think like, how is the massive manufacturing edge that China has not going to just keep compounding itself? I mean, this is like the multi trillion dollar question of the entire world. But it really does. I mean, when you think about, okay, there's a sort of presumably natural synergy, there's all this real world physical data that Chinese manufacturing companies can theoretically get from their various robotic factories. cleaners and so forth and then feed those into certain models that then we call AI. That really does seem like a potential leg up that, you know, just on the data collection alone from all those physical things, like a huge potential edge over the next several years. Yeah, although it was very interesting. Grace was saying that it's not as developed or as structured as it is in the [55:00] Because I thought the same thing. Like if everyone's talking over WeChat, if everyone's paying over WePay, you must have oodles and oodles of data. But yeah, that was interesting. The ecosystem thing, I think, is really important. And it just seems like really hard to get an ecosystem kind of going from scratch. And I remember maybe it was with Dan Wong, but someone describing how like if you go to Shenzhen, you can basically start like an entire company manufacturing a physical thing because every single supplier is there. You just go from like one storefront to another storefront to another storefront. I don't think that can be replicated anywhere in the US. I mean, this is a bit of a tangent, but, you know, this is economists talk about, quote, agglomeration all the time. The advantage is exactly that of having everything there and then you build these deep networks. One thing I find to be a little strange, and again, this is a tangent, is how San Francisco concentrated AI is, even though at the American level, it is sort of pure like desk work, right? Like it not the sort of we need the bolts manufacturer we need the servos manufacturer we need the robovision manufacturer and yet it still so agglomerated or the fact that finance is so agglomerated in new york city i find that or uh conglomerated i find that to be a little odd so um but yeah i thought glomerated is a good word glomerated i don't know whether it's conglomerated or agglomerated i'm just gonna say glomerated we know that there are certain industries in the u.s that are quote glomerated one way or another but But, yeah, I did find that to be – and, man, these – like Minimax, $20 billion company, that's like nothing compared to the valuations that we see with American companies. Pretty wild stuff. Also, the point we should do more – I mean there's a million – we got to do – we have – and there are plenty of non-AI things to do. So you can't just keep saying we should do an episode on X or Y, but data markets and like the idea of like, OK, these Chinese companies can buy very pricey proprietary data after some exclusivity window. There's some interesting stuff down there. Yeah, the monetization was really interesting to hear. OK, shall we leave it there for now? Let's leave it there. This has been another episode of the All Thoughts podcast. I'm Tracy Alloway. You can follow me at Tracy Alloway. And I'm Jill Waisendahl. You can follow me at TheStalwart Follow our guest Grace Chow She at GraceMZChow and check out her sub stack AI Prom Follow our producers Carmen Rodriguez at Carmen Armand Dashiell Bennett at Dashbot, Kale Brooks at Kale Brooks, and Kevin Lozano at Kevin Lloyd Lozano. And for more Odd Lots content, go to Bloomberg.com slash Odd Lots. We have a daily newsletter in all of our episodes, and you can chat about all these topics 24-7 in our Discord, discord.gg slash Odd Lots. And if you like Odd Lots, if you want us to do an episode on data markets, then please leave us a positive review on your favorite podcast platform. And remember, if you are a Bloomberg subscriber, you can listen to all of our episodes absolutely ad free. All you need to do is find the Bloomberg channel on Apple Podcasts and follow the instructions there. Thanks for listening. Thank you for imprint for certain. Healthcare doesn't always work great. If you've ever waited on a refill or couldn't schedule an appointment, you get it. That's the kind of stuff Optum is changing. They're using data and technology to integrate patient care, pharmacy, and everything else. So healthcare is connected, not complicated. What's that look like? Cheaper prescriptions that are easier to get and care that looks at the whole person, how you need it. Optum is helping make healthcare work as one for everyone. Learn more at business.optum.com. [60:00] So whether an AI agent supports a single user or your entire enterprise, with Okta, you'll turn risk into opportunity. Secure every agent. Secure any agent. Okta secures AI.