US Labor Dept, Anthropic See AI Lifting Output Over Job Cuts

美劳工部与Anthropic：AI重在提升产出而非削减岗位 雇主正在部署生成式AI以扩大整体运营能力并重组离散任务，而非缩减人员编制；与此同时，上下游的组织瓶颈暂时将宏观经济生产率的提升掩盖在跨度数年的J曲线之后。 专家小组成员通过将工作视为各项具体任务的复杂组合、而非面临被整体淘汰的单一角色，化解了关于技术性失业的普遍焦虑。Gregor Schubert 阐述了宏观经济生产率的 J 曲线及组织瓶颈；Pavel Khodalev 对比了全新开发与企业遗留代码库环境下开发人员的真实生产率；Sarah Heck 详细介绍了 Anthropic 经济指数（Anthropic's Economic Index）的实证使用数据及高薪任务暴露情况；Taylor Stockton 则分享了美国劳工部聚焦于基础设施岗位、劳动力灵活性与实时劳动力追踪的相关举措。最终，专家们主张将公众叙事从对被取代的恐惧和职场“作弊者”的污名化，转向普及基础 AI 素养、强化以人为本的技能以及制定基于实证的政策。 当前宏观经济正处于“生产力J曲线”初期，前期在实验、流程重构与基础设施上的巨大投入暂时掩盖了潜在效率收益，未来年均生产力预计将呈现1%至2%的温和渐进增长。 (1:56) AI物理基础设施的扩建正在成为短期劳动力需求的核心引擎，数据中心技术员、暖通专家和电网架线工等高薪蓝领及技术岗位迎来快速增长。 (9:13) Anthropic经济指数显示，自动化特定细分任务并不等于消除整个岗位；与以往技术革命不同，当前生成式AI的暴露度高度集中在软件工程等高薪专业领域。 (18:09) 目前生成式AI主要以“中间对中间”而非端到端的方式运作，上下游仍需大量人工进行数据准备与质量把控，企业普遍倾向于保留人员并借助AI大幅扩展整体产能。 (24:55)

Panelists at the Ai4 conference argue that artificial intelligence is fundamentally transforming the economy through the rebundling of individual tasks rather than wholesale job elimination, though realizing broad macroeconomic gains will require navigating complex organizational bottlenecks and cultural friction. The Productivity J-Curve and Bottlenecks (1:56) UCLA Anderson Assistant Professor Gregor Schubert explains that the economy is navigating an AI productivity J-curve, where heavy upfront capital spending, experimentation, and workflow restructuring temporarily mask efficiency gains or cause measured productivity to stagnate. While isolated laboratory experiments demonstrate discrete task speedups of 10% to 50%, real-world enterprise integration currently functions 'middle to middle rather than end to end.' Companies must supply significant human labor upstream to clean data and structure prompts, as well as downstream to conduct quality assurance and package outputs. Because organizational systems cannot instantly scale around accelerated intermediate steps, Schubert pushes back against sensational forecasts of 20% to 30% annual GDP surges, projecting instead a gradual ramp of roughly 1% to 2% in additional annual productivity growth over the coming decade as firms build out organizational capital. Software Engineering as an Early Indicator (3:30) Pavel Khodalev, CTO at First Line Software, highlights software engineering as a leading testing ground for AI adoption because developers can deploy tools autonomously without corporate friction. Breaking down the engineering lifecycle, Khodalev allocates roughly 20% of developer time to stakeholder analysis and architectural synthesis, 60% to coding and translation, and 20% to delivery and operations. AI primarily automates the coding phase, while human-centric competencies like empathy and stakeholder alignment remain unaddressed by models. Citing a Stanford study of enterprise teams, Khodalev contrasts measured productivity returns across codebase types: brand-new 'greenfield' projects can achieve up to a 60% efficiency gain, whereas legacy 'brownfield' systems yield only 15% to 20% improvements or even negative returns due to the overhead of debugging and integrating generated code. Overall, mature AI tooling is more realistically projected to deliver up to a 2x efficiency multiplier over time rather than extreme 5x to 6x industry claims. Task Rebundling and Labor Market Dynamics (11:05) Economists and policymakers on the panel describe jobs as complex bundles of distinct tasks rather than monolithic blocks vulnerable to instant elimination. Schubert notes that while early predictions suggested AI would replace radiologists, current systems address only about 2 of roughly 30 daily radiological tasks. Similarly, Taylor Stockton, Chief Innovation Officer at the U.S. Department of Labor, characterizes the mass displacement narrative as significantly overstated, pointing out that employers are largely using generative tools to expand total output and capacity rather than cut headcount. Stockton notes that AI is driving net-positive job demand across physical infrastructure roles—including data center technicians, electrical lineworkers, and HVAC specialists—as well as specialized engineering and AI-enabled cross-industry roles. At Anthropic, Head of Policy Sarah Heck observes via company telemetry that early usage was primarily augmentative, but autonomous task completions flipped ahead after the release of Claude Code. She emphasizes that AI exposure is uniquely concentrated among high-wage earners rather than lower-to-middle brackets. Cultural Stigma and Policy Adaptation A significant delta persists between corporate leadership's belief that adoption is lagging and the reality of covert, widespread usage on the ground. Khodalev explains that employees frequently hide their use of generative tools from managers out of a 'cheater' stigma, fearing their work will be perceived as illegitimate shortcuts. Accelerating enterprise benefits requires leadership to dismantle this psychological barrier by actively encouraging augmented workflows. At the policy level, Stockton cites Mary Meeker's data showing a marked optimism gap—83% AI optimism in China compared to 39% in the United States—and outlines federal efforts to weave foundational AI literacy into apprenticeships and education. In parallel, Heck and Stockton advocate for data-driven governance using real-time labor indicators, such as the DOL AI Workforce Research Hub and Anthropic's Economic Index, to guide public transitions and expand institutional agility. The panel views artificial intelligence as a general purpose technology whose long-term economic footprint hinges on organizational learning, updated labor infrastructure, and empirical tracking. Key unresolved issues include whether AI adoption will widen economic inequality across regions and how standard economic indicators will capture unmeasured consumer surplus generated outside formal workplaces.

在Ai4研讨会上，来自学术界、工业界与美国政府的专家指出，人工智能对经济的冲击正处于以高昂沉没成本与流程重组为特征的“生产力J曲线”初期；尽管全盘替代劳动力的悲观论调被严重夸大，但其实际效益的释放正深度受制于企业工作流瓶颈、遗留系统技术债务及职场文化心理阻碍。 生产力J曲线与企业中端执行瓶颈 (1:56) UCLA安德森管理学院助理教授Gregor Schubert指出，当前宏观经济统计中未能观测到生产力激增，主要是因为企业正处于技术初期的“生产力J曲线”阶段。在此阶段，组织必须在技术探索、软件采购和人员调整上投入大量前期资本与时间，导致短期内可衡量的生产力可能停滞甚至暂时下滑。Schubert反驳了短期内可实现20%至30%爆炸性GDP增长的激进预测，认为受限于传统企业僵化的架构，未来十年更为现实的发展节奏是每年带来约1%至2%的渐进式宏观生产力增益。 企业内部的流程阻碍进一步解释了实验室与实际落地间的巨大落差。Schubert强调，尽管受控实验环境下离散任务可测得10%至50%的效率提升，但生成式AI在真实业务中普遍表现为“中端执行”（middle-to-middle）而非端到端交付。上游环节依然极度依赖人工准备提示词与清洗数据，下游环节则必须耗费大量人力进行输出验证与交付包装。上下游固有的人工瓶颈削弱了中间生成步骤的加速红利，使得企业整体组织资本的构建需要5至10年方能显现。 任务捆绑理论与劳动力替代误读 (11:05) 针对技术性失业的公众焦虑，Schubert与美国劳工部首席创新官Taylor Stockton主张将职业视为“细分任务的捆绑包”，而非非黑即白的整体。Schubert引用过往对放射科医生的误判为例，指出医生日常承担约30项独立职责，而现有AI仅能覆盖其中约2项，因此技术引发的是任务的重新解构与捆绑，而非整个工种的凭空抹除。Stockton进一步表示，劳动力被成规模替代的论调被“显著夸大”；由于企业往往选择保留现有雇员并借助AI大幅提高总产能与服务质量，企业实际用工需求甚至伴随业务扩张而增加。 美国劳工部将AI在宏观上定调为美国经济的“净增益”，并指出了新的高薪就业扩张路径。除前沿系统研发岗位外，大规模数据中心与电网建设等物理基础设施直接催生了对HVAC暖通技术员、电网线路工和数据中心运维人员的大量需求。Stockton强调，劳工部已成立AI劳动力研究中心（AI Workforce Research Hub），通过动态追踪真实市场数据并注资联邦学徒制与AI素养项目，协助多元教育背景与地理区域的劳动者平稳过渡到新型岗位。 软件工程试验田与代码环境分化 (14:11) First Line Software首席技术官Pavel Khodalev提出，软件开发具备极高的自主试验空间，已成为观察认知类工作演变的关键风向标。Khodalev将工程周期划分为三个阶段：约20%时间用于需求同理、非言语信息捕捉与架构设计；60%时间用于代码转译；20%时间用于交付测试与安全合规。由于人类共情力与经验判断目前完全无法被AI取代，且代码校验带来大量认知开销，成熟AI工具对整体研发效率的现实提升上限约为2倍，远非行业宣称的5至6倍完全替代。 软件所处的具体架构环境直接决定了技术回报的成败。Khodalev援引斯坦福大学针对数千名工程师的研究指出，在从零搭建的“绿地项目”（Greenfield）中，AI代码生成最高可提升60%的开发效率；但在充斥技术债务的企业“棕地/遗留系统”（Brownfield）中，效率提升收窄至15%至20%，甚至因排查AI错误代码反向导致负收益。同时，Anthropic政策负责人Sarah Heck指出，其经济指数研究显示，随着智能体编程工具Claude Code上线，自动化任务占比已首次反超辅助性交互，且AI冲击呈现出高度向软件工程等高薪群体集中的异质性特征。 职场隐秘采纳、污名与家庭福利 (38:26) 在企业文化与心理层面，Khodalev揭示了管理层认知与一线员工行为间的严重脱节：企业高层多抱怨AI采纳滞后，但员工私下却普遍存在未经报备的自发应用。造成这种“地下采纳”的核心症结在于职场上的“作弊者污名”（cheater stigma）——员工担忧公开使用生成式工具会被视作投机取巧或能力缺陷。消除这一心理防御机制并将其转变为正向授权，是企业突破采纳僵局的前提。 在政策与宏观认知上，Stockton援引Mary Meeker的研究指出，中国公众对AI的技术乐观度达83%，而美国仅为39%，极易导致采纳意愿落后，因此迫切需要强化跨阶层的AI素养以重构公众信任。Heck则主张依靠近实时的使用数据而非主观臆测来指导监管与公共治理。此外，Schubert强调当前研究存在严重盲区：个人在工作之外利用AI进行自主学习与生活优化的非正式使用，正在创造极其庞大但未被GDP计入的“家庭福利盈余”，这一正面外部性在聚焦岗位更替的辩论中往往被严重低估。 本场研讨会的核心共识在于，AI作为堪比电力与蒸汽机的通用目的技术，其长周期转型注定伴随中端集成瓶颈与高昂重组成本，短期内既不会造成毁灭性的结构性失业，亦无法带来奇迹般的爆发式增长；悬而未决的关键命题在于，随着高薪技术暴露度的上升以及企业内部隐秘采纳带来的治理断层，公共政策与企业管理层能否依靠动态客观数据打破制度与文化阻力，确保技术转型的红利超越单纯的软件中端提速，向跨区域、多层次的劳动力群体平稳扩散。

## Background 该视频由行业会议频道 Ai4 于 2025 年 8 月 29 日发布在 YouTube 平台。视频记录了一场关于 AI 经济影响的圆桌讨论，汇聚了来自 Kearney、美国劳工部、UCLA、First Line Software 以及 Anthropic 的多位专家，重点探讨 AI 如何重塑生产力、变革劳动力市场并影响未来的宏观经济结构。 ## TL;DR 圆桌嘉宾围绕技术采纳的“J 曲线”效应展开剖析，指出 AI 当前仍处于早期的组织适配与探索阶段，短期内难现爆发式跃升，但中长期将重构岗位与任务边界。讨论破除了“AI 将引发全盘性失业”的过度恐惧叙事，强调工作是任务的集合体，AI 更多是在自动化特定中间环节，而整体经济将迎来伴随算力基建和技能升级的净正面效应。软件工程前线数据显示效率提升受限于遗留系统与端到端协同，学者与政策专家呼吁全社会必须从恐惧转向乐观叙事，重视人际软实力培养与广泛的家庭端福利增量。 ## Atmosphere & style 整场圆桌基调理性务实且充满前瞻洞察，嘉宾们既没有陷入科技乌托邦式的盲目吹捧，也坚决反对耸人听闻的失业恐慌。对话在学术理论、政策大局与一线工程实战之间自如切换，主持人穿插的育儿趣事与嘉宾对官僚流程“加州车管所难题”的自嘲让专业讨论显得格外接地气。现场氛围融洽且观点交融，观众能深切感受到决策者与从业者在面对未知变革时那种努力保持清醒、依赖实证数据破除迷思的严谨探索精神。 ## Key takeaways - [01:56](https://www.youtube.com/watch?v=JuPprHA3AMc&t=116s) 技术采纳遵循经济学上的“J 曲线”规律，初期因试错成本和流程探索可能导致短期生产力停滞，真正的生产力爆发将在中长期显现。 - [03:30](https://www.youtube.com/watch?v=JuPprHA3AMc&t=210s) 软件工程是 AI 变革的最前沿试验田，程序员群体率先经历人机协作模式重塑，其工作流演变为各行业预判劳动力市场转型的风向标。 - [09:13](https://www.youtube.com/watch?v=JuPprHA3AMc&t=553s) 算力数据中心建设带动了暖通空调技术员、高压电力线路工和机房运维等关键实体岗位的巨大需求，推动了新一轮高薪基础设施就业。 - [11:05](https://www.youtube.com/watch?v=JuPprHA3AMc&t=665s) 岗位本质上是多项任务的集合体而非单一黑盒（如放射科医生约30项任务中仅极少数能被自动化），AI 的核心影响是任务边界的重组与岗位重塑而非瞬间消灭工种。 - [13:17](https://www.youtube.com/watch?v=JuPprHA3AMc&t=797s) 软件工程中前端需求理解（依赖同理心与直觉）和后端系统交付依然高度依赖人类，破除夸大的神话后，AI 带来的实际综合效率提升上限约为 2 倍左右。 - [18:09](https://www.youtube.com/watch?v=JuPprHA3AMc&t=1089s) 随着自主 Agent 工具（如 Claude Code）的推出，自动化任务占比首次超越人机协同增强任务，且 AI 冲击力历史上首次高度集中于高薪高技能技术群体。 - [22:25](https://www.youtube.com/watch?v=JuPprHA3AMc&t=1345s) 现实中雇主更倾向于利用 AI 红利拓展新增业务做大蛋糕，而非机械化削减员工，政策重点应转向为劳动力构建敏捷重训体系而非盲目恐慌大规模失业。 - [24:55](https://www.youtube.com/watch?v=JuPprHA3AMc&t=1495s) 当前 AI 提速主要集中在“中段到中段”环节，受制于前端数据准备与后端质检交付两大人工作业瓶颈，企业需耗费 5 到 10 年积累“组织资本”才能释放宏观生产力。 - [27:48](https://www.youtube.com/watch?v=JuPprHA3AMc&t=1668s) AI 带来的宏观生产力更可能呈现每年 1% 到 2% 的渐进式增长，医疗生物科学、定制化教育及精简官僚流程的政府公共服务是三大高潜力落地领域。 - [30:40](https://www.youtube.com/watch?v=JuPprHA3AMc&t=1840s) 基础 AI 素养已成为未来职场的“入场券”，而建立信任、同理心与情感沟通等人际软实力将构成人类不可替代的核心互补优势。 - [33:02](https://www.youtube.com/watch?v=JuPprHA3AMc&t=1982s) 实证数据显示 AI 提效因系统架构而分化：从零搭建的新项目（Greenfield）最高可提效 60%，而维护企业遗留老旧系统（Brownfield）提效仅 15% 到 20% 甚至为负。 - [38:26](https://www.youtube.com/watch?v=JuPprHA3AMc&t=2306s) 企业与公众亟需打破将使用 AI 视作“作弊”的负面心理与恐惧叙事，将其正向重构为利用科技武装进化的新协同形态。 - [39:42](https://www.youtube.com/watch?v=JuPprHA3AMc&t=2382s) 人们在家庭生活与个人日常学习中使用生成式 AI 的频率甚至高于职场，这一维度创造了海量却被传统经济统计指标忽视的消费者福利。 - [40:49](https://www.youtube.com/watch?v=JuPprHA3AMc&t=2449s) AI 是比肩蒸汽机、电力和互联网的通用目的技术（GPT），在尚无成熟“标准操作手册”的早期历史窗口中，敏捷转型者将确立长期竞争优势。 ## Detailed walkthrough ### [00:05](https://www.youtube.com/watch?v=JuPprHA3AMc&t=5s) 圆桌开场与嘉宾介绍 主持人 Miguel Paredes（Kearney 合伙人兼吴恩达 AI Fund 顾问）首先向现场观众致意，并感谢 Ai4 盛会逐年扩大的规模与影响力。他指出，人工智能对日常生活、整体经济乃至人类社会结构的深远影响，是当前最受瞩目且充满未知的重要议题。尽管业内对未来的走向众说纷纭、尚无定论，但许多先锋机构和企业已经在第一线目睹了这项技术对其自身、客户以及政府运作方式的重塑。 本次圆桌汇聚了来自学术界、产业前沿与政策制定核心的四位重磅嘉宾： - **Gregor Schubert**：加州大学洛杉矶分校（UCLA）安德森管理学院助理教授，专注于 AI 对宏观经济与市场影响的研究。 - **Pavel Khodalev**：First Line Software 首席技术官（CTO），长期带领团队协助企业客户进行智能化转型及软件工程流程重塑。 - **Sarah Heck**：大模型前沿企业 Anthropic 的政策主管（Head of Policy）。 - **Taylor Stockton**：美国劳工部（U.S. Department of Labor）首席创新官（Chief Innovation Officer），并在近期刚刚代表官方发布了劳动力市场相关的最新研究报告。 --- ### [01:56](https://www.youtube.com/watch?v=JuPprHA3AMc&t=116s) J 曲线效应与中长期生产力增长预期 作为首位发言嘉宾，Gregor Schubert 教授开门见山地指出，当前大众所目睹的 AI 影响不过是“冰山一角”，未来十年内才会真正迎来颠覆性的经济重塑。他在 UCLA 教授一门面向管理者的 AI Agent 课程，从实际教学与产业互动中他观察到，绝大多数企业目前仍处于非常早期的探索阶段。在经济学上，这种现象被称为技术采纳的“J 曲线（J-curve）”。 ```mermaid flowchart LR accTitle: 技术采纳的 J 曲线效应 subgraph Phase1["当前阶段：初期探索与投资"] A["引入 AI 新技术"] --> B["时间与算力资源投入"] B --> C["短期生产力停滞或微跌"] end subgraph Phase2["中长期阶段：价值释放"] C --> D["跑通成熟商业用例"] D --> E["生产力爆发式跃升"] end ``` 在 J 曲线的初期低谷阶段，企业需要投入大量时间、算力与人力去摸索 AI 的落地路径，因此短期内不仅难以看到显著的生产力提升，甚至可能因试错成本而出现暂时的生产力负增长。因此，Gregor 提醒大家保持耐心，切勿根据按月发布的宏观经济短期波动去“占卜算卦”（"reading the tea leaves"）并草率断定 AI 是否有效。真正的生产力爆发和颠覆性应用场景，将在中长期的后台积淀之后集中显现。 --- ### [03:30](https://www.youtube.com/watch?v=JuPprHA3AMc&t=210s) 软件工程：AI 变革的先行试验场 First Line Software CTO Pavel Khodalev 将聚光灯拉回到自身所在的软件工程领域。他用“既令人着迷又令人畏惧”（"surprisingly fascinating and scary"）来形容程序员们当前的生存状态——软件工程师们亲手构筑了强大的 AI 系统，却也在第一时间成为了被 AI 助手与自动化能力深度冲击的对象。 与其他受到重重合规与预算限制的传统行业不同，软件工程师拥有得天独厚的自由度：他们无需向管理层申请专项资金或许可，就能在日常工作中直接上手实验各种 AI 工具。经过前期的迷茫与技术摸索，如今软件开发的未来形态正日趋清晰。Pavel 向各行各业的观察者建议：如果想预判 AI 将如何颠覆自己的行业，最直接有效的方法就是密切关注软件工程领域的演变范式，因为这里正是未来整个劳动力市场转型的先锋试验田。随后，主持人 Miguel 也补充强调，业界当前正热议纯粹的“编写代码（Coding）”与更高维度的“软件工程（Software Engineering）”在 AI 赋能下的本质差异。 --- ### [05:09](https://www.youtube.com/watch?v=JuPprHA3AMc&t=309s) Anthropic 经济指数：基于真实交互数据的洞察 Anthropic 政策主管 Sarah Heck 风趣地调侃道，如果自己真能准确看透未来，现在可能早就躺在“通用人工智能的度假海滩上”（"on my AGI beach"）享清福了。她严肃指出，过去业内缺乏大量基于现实交互的实证数据来评估 AI 的真实影响，而 Anthropic 作为底层模型提供方，正好处在数据的源头。为此，Anthropic 在 2 月推出了开源的“Anthropic 经济指数（Anthropic Economic Index）”，全面向公众开放用户与 Claude 交互的数据集。 结合美国劳工部的官方职业分类体系，Anthropic 追踪了用户与 AI 交互的具体模式，并梳理出几个关键趋势： - **行业分布**：除了最为集中的软件与信息技术领域外，教育学习、创意产业以及市场营销等领域的采纳率呈现出超预期的爆发态势，跨越了硬核科技与创意文科的广泛光谱。 - **任务性质划分**：研究重点区分了“增强型任务（Augmentative Tasks）”与“自动化任务（Automative Tasks）”。Sarah 特别澄清，自动化任务并不等同于将某个工作岗位连根拔起，而是自动化其中的具体环节。核心在于观察劳动者被解放出来的时间是被用于更高价值的生产活动，还是仅仅被闲置。 --- ### [08:15](https://www.youtube.com/watch?v=JuPprHA3AMc&t=495s) 美国劳工部视角：AI 带来的高薪新岗位与净积极效应 美国劳工部首席创新官 Taylor Stockton 指出，在很多社会舆论中，大家往往过度放大了 AI 对就业市场的替代冲击与潜在风险。美国劳工部虽然高度重视劳动者在转型期间所面临的结构性挑战，并全力提供再培训与兜底支持，但在整体宏观定调上，官方仍将 AI 视作一个带来“净正面效应”（Net Positive）的历史性机遇。 Taylor 强调，AI 正在为美国劳动力市场创造海量全新的高薪就业岗位。这种增量不仅限于底层大模型研发或技术开发岗（如 AI 工程师），更广泛地渗透到了医疗健康、市场营销等各行各业——在这些领域中，“熟练运用 AI”正演变为岗位本身的核心技能要求，从而催生出一大批此前从未存在过的全新职业形态。 ### [09:13](https://www.youtube.com/watch?v=JuPprHA3AMc&t=553s) AI 基础设施岗位机遇与劳工部 AI 劳动力研究中心的建立 泰勒·斯托克顿（Taylor Stockton）继续深入探讨了新岗位体系中的第三类——**AI 基础设施岗位**。他指出，在白宫上月发布的“总统 AI 行动计划”（President's AI Action Plan）中，政策重点聚焦于伴随大规模数据中心与算力基础设施建设而激增的人才缺口。这类建设催生了对暖通空调（HVAC）技术员、电力线路工（Linesmen）以及数据中心运维技术人员等实体关键岗位的巨大需求。美国劳工部的核心战略目标之一，就是通过对 AI 基础素养培育以及 AI 相关学徒制（Apprenticeships）项目的投资，打破地域、受教育程度和收入水平的门槛限制，为全美各阶层劳动者开辟通往这些高薪新岗位的成长路径。 斯托克顿进一步强调，当前关于 AI 影响的讨论必须摆脱浮躁的传闻与猜测，真正扎根于“实时客观数据”。与其听信耸人听闻的裁员头条新闻或盲目跟风“我表哥说他们公司发生了什么”，决策者与公众更需要建立主动的监测机制来摸清底层实况。为此，作为行动计划的一部分，美国劳工部正式宣布启动**“AI 劳动力研究中心”**（AI Workforce Research Hub），联合内部及多元化的私营机构合作伙伴展开持续性分析，以便更敏捷地支持劳动者应对这一经济新现实。 - **基础设施关键岗位**：大规模数据中心建设直接拉动了暖通技术人员（HVAC technicians）、高压线路工人（linesmen）以及机房运维专员的劳动力需求。 - **劳工部核心行动**：建立 AI 劳动力研究中心，推进 AI 普及教育与新型学徒制，强调以真实数据而非道听途说的个案作为政策制定依据。 --- ### [11:05](https://www.youtube.com/watch?v=JuPprHA3AMc&t=665s) 岗位是任务的集合体：AI 转型中的组织重构与“放射科医生”启示 主持人米格尔·帕雷德斯（Miguel Paredes）以轻松幽默的家庭趣事接过话题。他谈到自己有两名分别只有 13 岁和 10 岁的女儿，每当他竭力鼓励她们使用 AI 工具并为未来职业做准备时，孩子们总会无奈抗议：“老爸，快停下吧！我们知道你超爱 AI，但求求你放过我们。”随后，米格尔向加州大学洛杉矶分校（UCLA）的格雷戈尔·舒伯特（Gregor Schubert）教授发问：面对关于“工作全盘消失”还是“创造全新岗位”的传统技术演进争论，在短中期内，工作往往并非直接消失，而是被重构——许多过往由人类承担的任务将被 AI 接管，劳动者是否具备胜任那些“仅有人类才能完成的任务”的能力？ 舒伯特教授高度认可了将工作拆解为“任务”（Tasks）这一核心分析维度。他指出，Anthropic 的政策报告与斯托克顿的观点都印证了这一点：**岗位绝不是非黑即白的单一整体（Singular blobs）**，不能简单理解为“要么整份工作被自动化，要么完全不受影响”。舒伯特在面向企业高管的课程中常引用深度学习先驱杰弗里·辛顿（Geoffrey Hinton）当年关于“AI 将彻底取代放射科医生”的著名预言作为反思案例——实际将放射科医生的工作细分后会发现，在其多达 30 项具体任务中，AI 真正能够接管的只有约 2 项。因此，AI 带来的实际上是任务的重新打包（Rebundling）与岗位重塑。 ```mermaid flowchart TD accTitle: 岗位重构流程示意图 subgraph S1["传统岗位认知误区"] A["传统工作概念<br/>（单一黑盒 Blob）"] -->|过度焦虑/简化| B["全盘被替代 VS 完全无影响"] end subgraph S2["任务级别重组现实"] C["复合岗位<br/>（如放射科医生约30项任务）"] C --> D["自动化任务<br/>（约2项：图像初筛等）"] C --> E["人类核心任务<br/>（临床沟通、综合诊断等）"] D & E --> F["新型岗位打包与工作流重构"] end ``` 舒伯特强调，企业在管理这场技术转型时必须具备战略高度。如果企业只是草率地给每位员工分发一个 OpenAI 或 Claude 账号，指望员工自行摸索“AI 玩法”，或者在自动化部分任务后简单地解雇员工，最终必将陷入阵痛并付出重新招聘的高昂代价。企业必须将 AI 工作流视为内部产品进行战略设计，引入严格的用户体验（UX）测试与质量保障（QA），主动引导员工向高价值业务环节重塑与迁移。 - **任务集合体理论**：岗位是多项任务的组合包，技术变革首先改变的是任务边界，而非瞬间消灭整个岗位工种。 - **实施误区警示**：AI 落地是涉及流程再造与职位重构的企业战略议题，绝非纯粹的技术或 IT 工具发放问题。 --- ### [13:17](https://www.youtube.com/watch?v=JuPprHA3AMc&t=797s) 软件工程的任务拆解：理解人类需求与破除“5倍生产力”的神话 作为来自 First Line Software 的首席技术官，帕维尔·霍达列夫（Pavel Khodalev）从软件工程一线实践出发，详细剖析了技术研发流程中的任务切片。他指出，软件开发并非单纯的打字敲代码，而是由三大核心环节组成的复杂工程： 1. **需求理解与方案架构（约占 20% 时间）**：这是最关键的人类活动。软件工程师需要调动多重感官、同理心、非语言沟通能力、过往经验积淀的“智慧”，乃至开启“胡话过滤器”（BS radar）来去伪存真，精准捕捉客户与用户的真实意图，进而综合构思出契合目标的解决方案。目前的 AI 缺乏多维度输入通道与人类直觉，完全无法承担这一阶段的任务。 2. **实际编程与机器语言转换（约占 60% 时间）**：将构想好的逻辑方案转译为计算机可执行的代码指令。这正是当前各类 AI 工具主力攻坚的环节。 3. **软件交付与运维上线（约占 20% 时间）**：涵盖安全审查、高可用与可扩展性验证等严格流程，确保系统能够稳定交付给终端用户使用。 ```mermaid flowchart LR accTitle: 软件工程生命周期时间分配与 AI 介入现状 subgraph Phase1["需求与架构 (20%)"] P1["人类核心能力<br/>同理心/直觉/经验/去伪存真"] end subgraph Phase2["代码编写 (60%)"] P2["AI 主攻自动化区间<br/>（需人工审查校验/人在回路）"] end subgraph Phase3["交付运维 (20%)"] P3["系统工程落地<br/>安全合规/扩展性/上线验证"] end Phase1 --> Phase2 --> Phase3 ``` 霍达列夫针对业界宣称“AI 将取代所有程序员”或带来“5倍、6倍生产力飞跃”的狂热言论进行了理性降温。他解释道，尽管 AI 在代码编写阶段展现了优势，但整体仍处于早期阶段，无法独立闭环全流程。工程师在向 AI 助手下发指令并委托任务后，必须严密检查验证其产出的准确性，这意味着人类程序员在未来数年内依旧是“闭环控制者”（In the loop）。基于上述严谨的工时数学模型，即使在全流程配合最为顺畅的理想状态下，AI 实际能为工程师带来的综合生产力提升上限也就是 **2 倍左右（Double the productivity）**。因此，当下的核心诉求是对软件工程师进行技能升级与重塑（Upskilling / Reskilling），促使其从纯粹的基础代码编写者进化为能够熟练驾驭 AI 工具的系统级工程师。 ### [18:09](https://www.youtube.com/watch?v=JuPprHA3AMc&t=1089s) Anthropic 的实证观察：从“人机协同”到 Agent 自动化的逆转与高薪群体冲击 围绕 Anthropic 内部掌握的独家数据与宏观研究，Sarah Heck 深入拆解了 AI 在任务级别（task-based）对就业市场的实际重塑路径。她首先坦承，目前所有数据研究仍处于“极其早期的阶段”（early innings），但 Anthropic 致力于保持数据透明并与学术界深度联动。她开玩笑说：“我哪怕立刻招来 10 位经济学家，我们面临的未解之题依然数以千计。”正因如此，Anthropic 在“经济未来计划”（Economic Futures Program）下设立了专项资助，向全球外部学者开放内部索引与第三方数据集，以合力拆解劳动经济学谜题。 在具体数据走向方面，最引人注目的拐点发生在“任务增强”（Augmentation）与“任务自动化”（Automation）的比例变化上。在前两期报告（分别于 2 月和 4 月发布）中，数据清晰显示“增强型任务”远超“自动化任务”，人类用户主要将 Claude 作为迭代协作的副驾驶；然而在 5 月发布的第三期报告中，这一平衡被彻底打破——自动化任务占比首次实现了对增强任务的反超。这一转折点恰逢 Anthropic 推出了其首款 Agent 工具 Claude Code。虽然这主要反映在软件工程领域利用 Agent 独立执行多步骤编码，但这种由代理软件接管完整工作流的趋势正在全行业快速蔓延。 此外，研究揭示了一个反常的经济学现象：AI 暴露度极度集中于“高收入阶层”（Upper echelon of wages），尤其是高薪软件工程师。Sarah 强调，这与以往历次往往冲击中低技能蓝领的技术革命完全不同，是历史上首次出现高端白领阶层成为技术震中。企业界目前正为此深感焦虑：这类高薪群体究竟会被转岗重训（reskill），还是其工作板块将被直接剔除？在地域维度上，早期研究还在观测 AI 会成为“水涨船高”（the rising tide lifts all boats）普惠各地的技术，还是会让原本就具有高 GDP 优势的区域获得脱离重力的“逃逸速度”（escape velocity），进一步拉大经济鸿沟。 - **资助合作机制**：Anthropic 设立“经济未来计划”（Economic Futures Program），为使用其数据研究劳动市场的学者提供科研资金。 - **任务拐点轨迹**：2024 年 2 月与 4 月报告显示协同增强占主导，5 月伴随 Claude Code 发布，自主代理执行的“自动化任务”首次超越“增强型任务”。 - **高薪暴露异象**：AI 冲击重心异乎寻常地集中在高收入技术研发人群，颠覆了以往机械化/信息化率先淘汰低技能岗位的历史规律。 --- ### [22:25](https://www.youtube.com/watch?v=JuPprHA3AMc&t=1345s) 美国劳工部的视角：被夸大的替代叙事与劳动力的“敏捷肌肉” 针对外界对“大规模失业”的普遍恐慌，美国劳工部（U.S. Department of Labor）首席创新官 Taylor Stockton 给出了针锋相对的政策观察，明确指出当下流行的“技术置换论”（displacement narrative）被严重夸大了。他尖锐地指出，许多唱衰就业的悲观模型仅仅建立在“这项技术理论上能做什么”的技术能力视角，完全脱离了“雇主在真实商业环境中如何重新配置任务与岗位”的决策逻辑。 Taylor 解释称，很多人落入了思维陷阱，看到 AI 能够“用更少资源做更多事”（do more with less），便想当然地推导出雇主会“用更少的人力做同样的事”（do the same with less employees）。但现实中政府观察到的企业行为恰恰相反：多数雇主是在“保持相同甚至更多员工规模的前提下，借助 AI 达成前所未有的超额产出”，即利用效率红利开拓新增业务，而非机械式裁员。 尽管大规模失业论立不住脚，但岗位内部的任务转型与技能迭代却是实打实的。当天，美国劳工部联合教育部（Department of Education）及商务部（Department of Commerce）共同发布了最新的人才报告，其核心诉求是为全美劳动力市场注入“敏捷性”（agility）与灵活性。Taylor 强调，技能需求正在动态演变，政策层面不可能预判所有具体岗位，因此唯一的解法是为各州劳工部门和培训项目锻造“灵活性的肌肉”——建立敏锐捕获本地雇主新型技能需求的雷达系统，并快速拉起人才再培训管道，支撑人口随产业演化实现无缝升级。 ```mermaid flowchart TD accTitle: AI技术落地与劳动力决策分流图 subgraph Misconception["外界悲观假设（过度夸大替代）"] A["AI 实现降本增效<br>（Do more with less）"] --> B["削减员工规模<br>（Do same with fewer staff）"] B --> C["大规模劳动力置换与失业"] end subgraph Reality["真实企业决策与政策导向"] D["引入 AI 生产力工具"] --> E["保留/扩充员工规模<br>（做大业务增量）"] E --> F["重塑任务结构与技能栈"] F --> G["跨部门联动（劳工部/教育部/商务部）<br>构建敏捷劳动力培训管道"] end ``` --- ### [24:55](https://www.youtube.com/watch?v=JuPprHA3AMc&t=1495s) 经济学家的现实检验：“中段到中段”瓶颈与组织资本积累 针对企业究竟在何处看到了真实的生产力跃升这一核心难题，UCLA 经济学助理教授 Gregor Schubert 剖析了“实验室神话”与“企业现实”之间的巨大脱节。他指出，目前绝大多数惊人的效率增幅都出自高度受控的单任务实验环境——给被试者一个 Chatbot 去完成写作或特定代码编写，很容易测出 10% 到 50% 的生产力飙升。但一旦把工具扔进现实企业，生产力神话便撞上了由复杂工作流构成的冰冷现实。 Gregor 用了一个极为精辟的技术诊断来解释这一困境：“AI 解决的是‘中段到中段’（middle to middle）的环节，而非‘端到端’（end to end）的工作流。”即便 AI 能把中间的处理过程加速 30% 甚至翻倍，企业整体产出依然受制于首尾两端的严重瓶颈：在前端，必须有专门团队进行任务结构化、清洗与准备高质量输入数据；在后端，需要资深员工进行质量把控、合规校验，并将结果包装成终端客户能用的成熟形态。首尾两端如果无法自动化扩容，中间段的提速在全局层面上就毫无意义。 这种工作流的重构与岗位角色的平移，注定使真正的宏观生产力释放需要长达 5 到 10 年的时间周期。目前宏观经济指标中看到的微弱增长，很大部分只是数据中心和电力基础设施建设的资本开支拉动。Gregor 认为，当前企业界表面上看起来“砸了巨额资金却没换回多少即时产出”，但这并不代表泡沫破裂，而是企业正在将大量成本转化为“组织资本”（organizational capital）。正如 Taylor 所言，各家企业正在重构内部流程以练就这种组织肌肉，只有熬过前期的结构适配期，数年后才能真正迎来全要素生产率的大规模喷薄。 ```mermaid flowchart LR accTitle: AI中段到中段的工作流瓶颈示意 subgraph BottleneckFront["前端瓶颈（依赖高强度人工）"] In1["任务结构化"] --> In2["数据清洗与准备"] end subgraph AIProcess["AI 提速区（效率跃升 10%~50%）"] In2 --> Mid["中间层内容/代码生成<br>（Middle to Middle）"] end subgraph BottleneckBack["后端瓶颈（依赖高强度人工）"] Mid --> Out1["人工质检与风控"] --> Out2["封装交付给终端用户"] end ``` ### [27:19](https://www.youtube.com/watch?v=JuPprHA3AMc&t=1639s) 宏观生产力增长的现实节奏与 AGI 预期 在探讨宏观经济走势时，学者们认为 AI 带来的增长更可能呈现为一种平稳渐进的曲线，即在未来十年左右的时间里，每年为生产力带来约 1% 到 2% 的额外增长。由于现代企业组织架构的固化以及整个经济体系的高度复杂性，现实中很难出现某些人在畅谈 AGI（通用人工智能）或失控级超级智能（Runaway Superintelligence）时所宣称的“单年 GDP 暴增 20% 至 30%”的剧烈飞跃。 ### [27:48](https://www.youtube.com/watch?v=JuPprHA3AMc&t=1668s) Anthropic 视角：医疗、教育与政府公共服务中的 AI 潜力 主持人 Miguel Paredes 向 Anthropic 政策负责人 Sarah Heck 提问：在公司拓展市场并观察世界变化的过程中，哪些领域在采纳 AI 方面令人意外地滞后，或者说在哪些领域最有望看到 AI 带来的巨大变革？Sarah Heck 直言，目前全行业仍处于早期阶段，Anthropic 希望在各个领域看到更广泛的应用，但他们正重点与外部机构合作推进三大具有“超额影响力”（Outsized Impact）的领域： ```mermaid flowchart TD accTitle: Anthropic 重点布局的三大高影响力领域 subgraph ImpactAreas["Anthropic 重点推动的高影响力落地场景"] A["医疗与生物科学<br/>(Healthcare & Bioscience)"] B["教育体系重塑<br/>(Education & Curriculum)"] C["政府与公共服务<br/>(Government & Public Services)"] end A --> A1["减少文书负担，让医生专注于临床诊疗"] A --> A2["加速生物医药研发与数据处理"] B --> B1["支持教师备课与教学增效（非作弊工具）"] B --> B2["针对学生差异化需求实现因材施教"] C --> C1["精简官僚流程（解决'加州车管所 DMV 难题'）"] C --> C2["重建公众对政府及非营利机构的信任"] ``` 第一是医疗与生物科学。这一方面体现在将医生从繁重的文书工作中解放出来，让他们有更多时间陪伴患者、深入思考病情诊断；另一方面则体现在生物医药研发领域，AI 模型执行和加速科研任务的效率远超单个科研人员乃至研究团队。Anthropic 正在积极投资推进此类“有益部署”（Beneficial Deployments）。第二是教育领域。Anthropic 正与美国白宫及多家机构合作，探索教师如何利用 AI 提高教学效率并为不同需求的学生量身定制个性化课程，引导技术走向赋能教学而非沦为作弊工具。 第三是政府公共服务。Sarah 幽默地将其比作自己的“加州车管所（DMV）难题”——没有人想再去 DMV 排长队换驾照，如果 AI 能打破繁文缛节、大幅提速办事流程将意义重大。作为美国国务院的前雇员，她回忆起当年鲍威尔（Colin Powell）推动在每张办公桌上配备电脑的历程：从打字机线缆过渡到能向驻外使馆发送电子邮件耗费了漫长时间。AI 在公共部门的普及同样不会一蹴而就，但联邦、州及地方政府对 AI 的投入，将显著提升公共服务效率，并重新建立公众对政府机构及非营利组织的信任。 - **三大高潜力聚焦领域**：医疗与生物医药研究、定制化教育体系、政府与公共行政服务。 - **历史镜像**：回顾美国国务院当年普及电脑办公的渐进过程，说明政府机构技术采纳需要周期，但长期红利巨大。 ### [30:40](https://www.youtube.com/watch?v=JuPprHA3AMc&t=1840s) 人机协作新范式：基础 AI 素养与不可替代的人际软实力 针对人类工作任务逐步外包给 AI 的趋势，美国劳工部首席创新官 Taylor Stockton 深入解析了未来劳动者急需培养的核心素养。他强调，政府的第一要务是普及“基础 AI 素养”（Foundational AI Literacy）。虽然这并非万能解药，但却是进入未来职场的“入场券”（Table Stakes）。当前政府正将 AI 素养融入所有教育和劳动力资金支持项目中，覆盖 K-12 基础教育、职业与技术教育（CTE）、高等教育以及针对失业群体的公共就业服务机构，同时推动企业投资在职员工的技能升级。 随着 AI 以更快、更高效的方式接管常规任务，工作内部的技能构成正在发生深刻转移。人类真正不可替代的核心优势转向了人际交往能力（Interpersonal Skills）——例如建立信任、同理心与情感连接。以医疗行业为例，AI 可以胜任极其复杂的疑难杂症诊断并制定个性化治疗方案，但与患者建立情感纽带、传递关怀依然完全依赖医生的人性温度。未来的劳动力投资既要聚焦于驾驭、管理甚至构建 AI 的硬技能，更要大力押注这些具有互补性的人性化软实力。 ### [33:02](https://www.youtube.com/watch?v=JuPprHA3AMc&t=1982s) 来自一线工程的实战数据：Vibe Coding 局限与传统系统的效能落差 First Line Software 首席技术官 Pavel Khodalev 从软件工程前线的实战经验出发，打破了外界对“氛围编码”（Vibe Coding，即仅凭自然语言 Prompt 就生成现成软件）的神话。他指出，Vibe Coding 确实可以快速构建原型来验证想法，但真实软件的生命周期在于持续迭代和修改以契合人类的动态需求，这恰恰是当前纯 AI 编码无法独立支撑的，深度的系统实现依然离不开专业工程师的介入。 Pavel 引用了一项来自斯坦福大学针对数百个团队、数千名工程师的实证研究项目数据，揭示了 AI 在不同研发场景下的巨大效能分化： | 开发环境类型 | 业务特征与系统现状 | AI 带来的实际效率提升幅度 | 核心瓶颈与工程现实 | |:--- |:--- |:--- |:--- | | **全新项目 (Greenfield)** | 从零起步搭建系统，无历史包袱 | **最高可达 60%** (取决于所采用的技术栈) | 架构灵活，能最大化发挥大语言模型的代码生成优势 | | **既有系统 (Brownfield / Legacy)** | 维护与重构企业老旧遗留系统 | **最佳情况下仅 15% - 20%**，部分场景出现**负增长 (<0%)** | AI 对过时陈旧技术栈支持不足，工程师花费在集成与修复上的时间超过了收益 | Pavel 强调，从整体行业现状来看，当前工具链能为软件工程师带来的综合效率提升大约在 15% 到 20% 之间，且前提是工程师自身完成技能重塑。AI 正在深度介入人类最私密的工作体验与思维过程，要求工程师从心智、大脑到日常工作习惯全面调适。这种人机融合需要漫长的适应期，但也正因如此，工作岗位不会在一夜之间被彻底颠覆，反而为人类利用技术更好地赋能彼此提供了宝贵的缓冲与进化空间。 ### [35:53](https://www.youtube.com/watch?v=JuPprHA3AMc&t=2153s) 告别恐惧叙事：重塑公众对 AI 经济的乐观预期 在圆桌会议进入尾声之际，主持人邀请每位嘉宾分享一句最具启发性的前瞻建议。劳工部的 Taylor Stockton 提出，当前全社会最紧迫的任务是将围绕“AI 与就业”的讨论焦点，从“基于恐惧的叙事”（Fear-based Narrative）彻底转向“基于乐观的叙事”（Optimism-based Narrative）。 Taylor 引用了知名风险投资人玛丽·米克（Mary Meeker）在几个月前发布的重磅 AI 趋势报告中的调查数据：中美两国公众在看待 AI 对未来的影响时存在鲜明反差，中国民众将 AI 视为净利好（Net Positive）的比例显著不同，显示出不同文化与社会语境下对新技术乐观程度的差异。如何引导公众看到技术带来的生产力跃升与全新机遇，成为决定未来经济动能的关键课题。 ### [36:39](https://www.youtube.com/watch?v=JuPprHA3AMc&t=2199s) 缩小中美 AI 素养差距：从恐惧走向普遍机遇 Taylor Stockton 结合中美两国的 AI 采纳率数据指出（此前数据提到中国为 83%，而美国仅为 39%），这种差距背后潜藏着巨大的经济风险。随着 AI 在美国经济中创造出海量的新机遇，关键问题在于有多少普通美国人真正有能力抓住这些机会。他强调，公众需要提升 AI 素养（AI Literacy）、培养对新工具的兴趣与创造力，才能充分享受技术发展红利。这一波由 AI 带来的发展浪潮，不应受到地理位置、受教育程度或收入水平的限制，而应尽可能广泛地普惠到全美所有群体。因此，核心任务在于如何转变整体社会叙事，将公众普遍的恐惧感转化为乐观与行动力。 ```mermaid flowchart TD accTitle: AI 普及与社会叙事转型框架 accDescription: 描述从传统恐惧与采纳鸿沟到赋能型叙事与全民福利的转变路径。 subgraph Current["现状与痛点"] A["采纳差距与焦虑<br/>（美 39% vs 中 83%）"] B["公众叙事偏差<br/>（恐惧被替代 / 自认'作弊'）"] end subgraph Strategy["核心应对策略"] C["数据驱动决策<br/>（实时调整与敏捷试错）"] D["重塑赋能叙事<br/>（AI 赋能的新人类）"] E["普及 AI 素养<br/>（跨地区/教育/收入层级）"] end subgraph Impact["综合社会效益"] F["职场生产力释放"] G["家庭与个人福祉提升<br/>（生活优化与自主学习）"] end A --> E B --> D E --> C D --> C C --> F C --> G ``` ### [37:18](https://www.youtube.com/watch?v=JuPprHA3AMc&t=2238s) 拥抱“数据驱动”的敏捷转型：快速试错与实时调整 Sarah Heck 对此表示高度赞同，并指出现在正是美国赢得领先优势的关键时刻，核心在于如何谈论 AI、使用 AI、保护技术并将其融入日常生活乃至国家安全体系中。她引用了一句名言幽默地表达自己的观点：“除了上帝，其余所有人都必须拿数据说话（In God we trust, and everybody else bring data）。”她呼吁用扎实、真实的数据来跟踪和观察技术演进，因为只有客观数据才能帮助全社会在各个行业、社区乃至课堂中，尽可能平稳地完成这场深刻的转型。 她进一步阐述道，AI 技术以光速在向前演进，而它的巨大优势在于能够提供接近实时的数据反馈。有了实时数据，企业与政策制定者便能迅速、灵活地做出调整，即使遭遇失败也能立即改变策略。尽管现代社会通常并不擅长快速迭代和接受失败，但这恰恰是当前所有人面临的挑战与历史机遇。 ### [38:26](https://www.youtube.com/watch?v=JuPprHA3AMc&t=2306s) 破解“作弊者”心理：重塑职场 AI 使用的叙事模式 Pavel Khodalev 认为当前最大的阻碍其实出在公共叙事（Public Narrative）上。主流舆论充斥着“AI 将夺走工作、消灭岗位、取代人类”的论调，这种恐慌叙事导致企业内部产生了显著的认知断层。企业高管在做问卷调查时，往往认为员工在工作中对 AI 的使用远远不够；然而在私下的一对一交流中，却会发现许多员工实际上早已在偷偷使用 AI。 这种企业预期与实际应用之间的脱节，很大程度上源于公众舆论给使用 AI 贴上的负面标签。Pavel 分享了他在调研中得到的最令人震惊的真实反馈：许多人坦言“我很不好意思承认自己在用 AI，因为这让我觉得自己像个作弊者（I feel like a cheater）”。Pavel 强调，必须彻底扭转这种陈旧叙事，明确告知员工使用 AI 并不丢人，使用 AI 并不是作弊，而是进化为被强大科技武装的“新版人类”。只有从根本上改变这种心态，才能大幅加速技术的落地与采纳。 ### [39:42](https://www.youtube.com/watch?v=JuPprHA3AMc&t=2382s) 政策与研究盲区：家庭与个人层面的巨大 AI 福利 Gregor Schubert 赞同全社会需要转变态度以推进技能重塑，同时他指出了当前政策讨论与学术研究中的一个巨大“盲区”——生成式 AI 在家庭与个人生活层面所带来的巨大福利。根据他即将发表的最新研究，使用统计数据显示，人们在个人日常生活中使用聊天机器人和生成式 AI 的频率，实际上远高于在职场中的使用频率。 现有的多数研究都狭隘地聚焦于 AI 对企业工作岗位和生产力的冲击，从而忽视了家庭维度的巨大价值。普通人在工作场景之外，利用 AI 来优化日常生活、搜索解答疑惑、自学全新知识，这些都创造了海量却难以被传统经济指标衡量的消费者福利（Consumer Welfare Benefit）。他指出，在权衡职场转型期的短期摩擦与成本时，必须将这种广泛的个人用户端福祉一并纳入考量。 - **家庭端采纳率高于职场**：大量用户在个人生活与日常学习中使用生成式 AI，形成未被传统经济指标充分衡量的福利增量。 - **走出单一就业视角**：评估 AI 经济影响时，需平衡职场短期重组成本与家庭端广泛的效用提升。 ### [40:49](https://www.youtube.com/watch?v=JuPprHA3AMc&t=2449s) 总结与展望：通用目的技术的历史契机与前行之路 Miguel Paredes 在主持结语中对讨论进行了系统梳理。他指出，当前社会处于 AI 革命的最早期阶段，普遍存在的“AI 焦虑”是极其正常的自然反应。但正如 Taylor 所言，大家应当将焦点迅速转移到海量的潜在机遇上，积极为自身、家庭、朋友乃至整个企业组织进行 AI 技能重塑（AI Skilling）。 Miguel 强调，正如经济学家所定义的，AI 属于典型的“通用目的技术”（General Purpose Technology, GPT），与历史上深刻重塑人类社会与全球经济格局的蒸汽机、电力以及互联网处于同等量级。在这场前所未有的变革中，目前全球尚未形成既定的“标准操作手册”（Playbook），这正是企业、组织与各国政府脱颖而出、确立领导地位的绝佳历史窗口。最后，在主持人的提议下，全场观众以热烈的掌声向与会嘉宾致谢。 ## Notable quotes > [01:56](https://www.youtube.com/watch?v=JuPprHA3AMc&t=116s) reading the tea leaves — Gregor Schubert > [03:30](https://www.youtube.com/watch?v=JuPprHA3AMc&t=210s) surprisingly fascinating and scary — Pavel Khodalev > [05:09](https://www.youtube.com/watch?v=JuPprHA3AMc&t=309s) on my AGI beach — Sarah Heck > [11:05](https://www.youtube.com/watch?v=JuPprHA3AMc&t=665s) 岗位绝不是非黑即白的单一整体 — Gregor Schubert > [13:17](https://www.youtube.com/watch?v=JuPprHA3AMc&t=797s) Double the productivity — Pavel Khodalev > [18:09](https://www.youtube.com/watch?v=JuPprHA3AMc&t=1089s) 我哪怕立刻招来 10 位经济学家，我们面临的未解之题依然数以千计。 — Sarah Heck > [18:09](https://www.youtube.com/watch?v=JuPprHA3AMc&t=1089s) the rising tide lifts all boats — Sarah Heck > [24:55](https://www.youtube.com/watch?v=JuPprHA3AMc&t=1495s) AI 解决的是‘中段到中段’（middle to middle）的环节，而非‘端到端’（end to end）的工作流。 — Gregor Schubert > [37:18](https://www.youtube.com/watch?v=JuPprHA3AMc&t=2238s) In God we trust, and everybody else bring data — Sarah Heck > [38:26](https://www.youtube.com/watch?v=JuPprHA3AMc&t=2306s) 我很不好意思承认自己在用 AI，因为这让我觉得自己像个作弊者 — Pavel Khodalev ## Mentioned Miguel Paredes, Gregor Schubert, Pavel Khodalev, Sarah Heck, Taylor Stockton, Kearney, AI Fund, First Line Software, Anthropic, Anthropic Economic Index, Claude, Claude Code, Economic Futures Program, U.S. Department of Labor, U.S. Department of Education, U.S. Department of Commerce, AI Workforce Research Hub, President's AI Action Plan, UCLA Anderson School of Management, Geoffrey Hinton, OpenAI, Mary Meeker, Colin Powell, Stanford University ## Video description Miguel Paredes, Partner / Fellow, Kearney, AI Fund Taylor Stockton, Chief Innovation Officer, U.S. Department of Labor Gregor Schubert, Assistant Professor of Finance, UCLA Anderson School of Management Pavel Khodalev, CTO, First Line Software Sarah Heck, Head of Policy, Anthropic Abstracts This session will examine the far-reaching economic impact of AI on individuals, companies, and global economies. Panelists will discuss how AI is reshaping productivity, changing job markets, and influencing the future of work. Explore the potential for AI to drive economic growth and innovation, while also addressing concerns around job displacement, workforce reskilling, and income inequality. Learn how businesses and governments can harness AI to improve economic output and prepare for the inevitable shifts in labor markets and economic structures. [Source](https://www.youtube.com/watch?v=JuPprHA3AMc)

[1:56] Gregor Schubert: reading the tea leaves [3:30] Pavel Khodalev: surprisingly fascinating and scary [5:09] Sarah Heck: on my AGI beach [11:05] Gregor Schubert: 岗位绝不是非黑即白的单一整体 [13:17] Pavel Khodalev: Double the productivity [18:09] Sarah Heck: 我哪怕立刻招来 10 位经济学家，我们面临的未解之题依然数以千计。 [18:09] Sarah Heck: the rising tide lifts all boats [24:55] Gregor Schubert: AI 解决的是‘中段到中段’（middle to middle）的环节，而非‘端到端’（end to end）的工作流。 [37:18] Sarah Heck: In God we trust, and everybody else bring data [38:26] Pavel Khodalev: 我很不好意思承认自己在用 AI，因为这让我觉得自己像个作弊者

[1:56] Gregor Schubert: only the tip of the iceberg of the impacts that we're going to be seeing in the next 10 years or so. [3:30] Pavel Khodalev: surprisingly fascinating and scary existence. [5:09] Sarah Heck: AGI beach [8:01] Taylor Stockton: net positive [11:05] Miguel Paredes: Dad, stop. We understand you like AI, but leave us alone. [18:09] Sarah Heck: hire 10 economists and we would still have thousands of questions that we'd all want to answer [18:09] Sarah Heck: rising tide [that] lifts all boats [22:25] Taylor Stockton: significantly overstated [24:55] Gregor Schubert: middle to middle rather than end to end. [27:48] Sarah Heck: California DMV problem [36:39] Sarah Heck: In God we trust and everybody else bring data.

[0:05] Thank you. Uh good morning everyone. Uh [0:07] it's great being here with you. I'm [0:08] going to like this so I can see people [0:10] here. Um we're very excited about this [0:12] panel. Thanks to AI4 for always putting [0:15] up a great uh event growing each year. [0:17] Um thanks so much for you guys for [0:19] coming. It's such an important topic and [0:20] we got a amazing lineup of panelists. Uh [0:23] my name is Miguel Paredes. I'll be [0:24] moderating this panel. um a partner at [0:27] Carney and a fellow and adviser to AI [0:29] fund, Andrew Ing's AI fund. Um and we [0:32] just are going to go deep into what's [0:33] going on with AI and what's it how's it [0:35] going to affect our lives, our [0:36] economies, our society. And I'm going to [0:38] do quick intros. Uh please feel free to [0:40] later compliment things that are [0:42] relevant. Um I don't want to take away [0:43] from the content, but today we have [0:46] Gregor Schubert, uh who's a professor at [0:48] UCLA, an expert on on topics of the [0:51] impact of AI on the economy and markets. [0:54] Uh we have Pavl Kolalev, sorry if I [0:57] butchered that. Uh CTO uh at um a [1:01] startup that's doing a lot of work [1:02] around uh this transformation with [1:05] clients and working with software [1:06] engineers first line uh and we will hear [1:09] about all the stuff they're doing. Uh we [1:11] have Sarah Hec who's head of policy at [1:13] anthropic. Sarah, thank you so much for [1:14] joining us. And uh Taylor Stockton, who [1:17] is the chief innovation officer at the [1:19] US Department of Labor, uh and who just [1:21] released a report, uh today or [1:24] yesterday, so we're going to hear all [1:25] about it today. So, uh exciting news. Um [1:28] we're going to start off by asking you [1:30] guys uh and each one will have a time to [1:32] respond. What what is the impact of AI [1:35] on the economy, on our societies? What [1:37] are you your personal views? We're [1:38] hearing a lot out there um different [1:41] points of view. Uh I don't think anyone [1:43] knows exactly where this is going but a [1:45] lot of people are and a lot of [1:46] organizations are firsthand seeing how [1:48] this is transforming their work or their [1:50] clients work or their government's work. [1:52] So I'd love to dig into that. Maybe [1:53] Gregor start with you. [1:56] >> Okay. So I think the I think the impacts [2:00] or the way I'd like to talk about this [2:02] is that the the I think the impacts that [2:04] we're seeing right now are in some sense [2:05] only the tip of the iceberg of the [2:07] impacts that we're going to be seeing in [2:08] the next 10 years or so. So I teach a [2:11] course at UCLA on AI agents for managers [2:14] and one of the takeaways from the course [2:15] is that a lot of organizations are still [2:17] at a very early stage in terms of their [2:19] implementation. Um there's lots of [2:21] things that uh uh companies are trying [2:24] out but they're probably still in the [2:26] part of the experimentation spectrum [2:28] that's sometimes called the J curve uh [2:30] where you actually just investing in the [2:32] technology and you're trying to figure [2:33] out what to do with it. And in that part [2:35] you're probably not seeing huge [2:36] productivity gains. you may be even [2:37] seeing negative productivity impacts [2:39] because you're investing a lot in trying [2:40] to figure out what to do with this [2:42] technology and investing time and uh [2:44] resources into that. And so I think [2:46] we're much more likely to see big [2:48] productivity gains and sort of new [2:50] applications and new use cases emerge [2:52] over the coming years uh relative to [2:55] what we've seen so far, which is [2:57] probably only the early experimentation [2:58] stage. And so I'm both pretty optimistic [3:00] with regard to uh the transformations [3:02] we're going to see in the medium term. [3:04] Um, and also I guess would like to [3:06] caution people to not over interpret uh [3:09] sort of uh month-by-month economic [3:11] statistics and trying to read the tea [3:13] leaves there uh with regard to whether [3:15] or not we're seeing productivity impacts [3:17] of AI or not. Uh because oftentimes the [3:19] early investment phase doesn't look like [3:21] the big productivity gains just sort of [3:23] cooking on the back end. [3:25] >> Thanks Gregor Pav. [3:30] >> For me it's a very interesting topic. I [3:32] I'm not probably uh ready to disclose [3:34] like about the whole economy but uh [3:36] being a part of the software engineering [3:37] and software development industry that's [3:40] like a surprisingly fascinating and [3:42] scary existence right now because what [3:44] is happening that probably the job and [3:47] activities of software engineers are [3:48] impacted mainly by AI assistance and AI [3:52] capabilities and that's kind of strange [3:54] thing we kind of inventing that and at [3:56] the same time impacting ourselves at the [3:58] same time and I believe the response [4:00] here is not just that they area is great [4:02] for AI application is because uh [4:05] engineers have no limitations in doing [4:07] that. They don't need to ask for any [4:08] money. They don't need to ask anybody [4:10] for permission of doing that. They just [4:11] doing that. So just playing with the [4:13] technology and being part of that [4:15] evolution, it's actually was for quite a [4:18] long time very foggy and unclear what's [4:20] is actually happening and what what what [4:22] does make sense, what does make sense. [4:24] But from the current moment of the time, [4:26] it seems like it's shaping up into the [4:29] future uh setup. Obviously AI can [4:32] progress more and it will be progressing [4:34] more and cognitive jumps is going to [4:35] happen and that will going to change the [4:37] whole landscape but for now it's kind of [4:39] emerging into software development and [4:42] if you are interested in what's going to [4:43] happen to your area of industry and jobs [4:46] observe closely what's happening in [4:48] software engineering that's probably [4:49] like the the example of future [4:52] evolution. [4:53] >> Thanks Po. One thing uh Po and I were [4:55] talking about earlier was the difference [4:57] between these tools uh really improving [5:00] and augmenting and sometimes replacing [5:02] coding versus software engineering for [5:05] example and engineering in general but [5:07] we'll dig into that in a bit. Sarah, [5:09] you're at Anthropic uh one of the uh [5:11] most important companies in the space [5:13] and and allowing and enabling a lot of [5:14] this transformation. Um tell us a little [5:16] bit your your take on this. Yeah. Um, if [5:20] I knew what the future was going to look [5:21] like, uh, I'd maybe be on my AGI beach [5:24] or something already. I don't know. Um, [5:26] but, you know, I think at Anthropic, uh, [5:28] we think about this deeply every day. [5:31] Um, and one of the things that we [5:34] thought about as we were sort of looking [5:37] trying to look into the future is there [5:39] was not a lot of data out there on what [5:42] um, how people and companies are using [5:44] it. and we sit at a company that has all [5:47] the data on it. So back in February for [5:49] the first time we released real data on [5:52] how people were interacting with Claude [5:55] um through our uh anthropic economic [5:58] index um which is available to everybody [6:00] and you can actually go and download the [6:01] data and I encourage you to do so and [6:03] and please like go do your own research [6:06] with it. um basically with the idea that [6:08] we want to be able to track these things [6:09] over time because I think Gregor's [6:11] exactly right like we are we are in such [6:13] early innings we might we might just be [6:15] in like the warm-up here um and so we [6:17] don't know but in order for us to make [6:19] good choices for our companies or for [6:21] our our government or for our society we [6:24] need to root this in data um so you know [6:26] I think probably unsurprisingly our data [6:29] has shown that the you know the the vast [6:31] majority of users are using it in sort [6:33] of the software and technology space we [6:35] s see outsized use in education um and [6:39] learning as well. Also probably [6:40] unsurprising. We also see outsized use [6:42] cases in creative industries and [6:44] marketing. Um again maybe not so um [6:48] surprising but I think uh instructive in [6:50] terms of you know you see use cases that [6:52] range from like sort of the hard [6:54] sciences all the way to to more creative [6:56] industries. We've also been tracking [6:58] this um in what kind of conversations [7:02] people are having with Claude. So, not [7:04] just sort of where the the the tasks map [7:07] to occupations, which we're grateful to [7:09] the Department of Labor for their um [7:11] their uh data that we can cross [7:13] cross-pollinate with ours to to tell you [7:15] this, but um also whether people are [7:18] using it for augmentative tasks or [7:20] automative tasks. Um and when I say [7:22] automative, I want to be very clear. [7:24] It's not about necessarily automating a [7:26] job away, but what tasks can be [7:28] automated? And then we'll have to study, [7:30] okay, so how does that person use the [7:32] time? Are they doing something that's [7:33] more productive or are they actually [7:35] just, you know, going and sitting on [7:36] their AGI beach? I don't think so. Um [7:39] but um we're studying this over time. [7:41] We've had three reports right now um out [7:44] and and we'll um likely have another one [7:46] soon. So I think where we approach this [7:49] is how do we make sure that we are [7:52] putting out the data that can help the [7:54] people in this room and people outside [7:55] this room make really informed decisions [7:57] about where the economy is going. [7:59] >> Thanks Sarah. it uh super interesting to [8:02] look at that data and see the findings [8:04] you guys have had. If you guys have not, [8:05] I recommend you do and I also recommend [8:06] you guys do what Sarah said, which is [8:08] download the data and and play around [8:09] with it yourself. Taylor, tell us a [8:11] little about your perspective and how [8:13] you guys are doing this at the [8:14] government level. [8:15] >> Yeah, absolutely. I mean, I think when [8:16] when I'm involved in these [8:17] conversations, a lot of times there's a [8:19] tremendous focus on the challenges of of [8:22] what the impact of AI will be on the [8:23] labor market. But at the Department of [8:25] Labor, I think fundamentally though we [8:27] do see the challenges and we want to [8:29] make sure that we're able to help [8:30] support workers in navigating those, we [8:33] overall see AI as really net positive in [8:36] terms of the new opportunities that that [8:38] AI can bring to the workforce and and to [8:40] American workers in terms of the [8:42] tremendous number of new jobs, new [8:43] highwage jobs that are created. Uh you [8:46] know, we think about a lot about the [8:47] different categories of those. Um, yes, [8:49] there are some segments of jobs that are [8:51] fairly technical that AI is creating in [8:53] terms of actually building new forms of [8:55] AI, AI engineers, etc. But we're also [8:58] seeing a lot of different data around [9:00] new high-paying AI jobs across [9:02] industries in marketing and healthcare [9:05] where fundamentally using AI is a part [9:08] of the job and this type of job hasn't [9:10] existed before until now. Uh, and then a [9:13] third category example of new jobs is AI [9:15] infrastructure jobs. And so if you look [9:17] at the president's action plan, AI [9:19] action plan that came out last month, [9:21] there was a tremendous focus on as we [9:24] think about these massive data, you [9:26] know, data infrastructure, data center [9:28] investments, the talent pipelines that [9:30] are necessary to build those in terms of [9:32] HVAC technicians, linesmen, data center [9:36] technicians. And so all of these types [9:38] of roles are really new opportunities [9:40] that are being created in the economy. [9:42] And so fundamentally at the Department [9:44] of Labor, one of our biggest priorities [9:46] is how do we create more pathways for [9:49] individuals across the country, [9:51] regardless of geography, regardless of [9:53] education level or income level to [9:56] access those new types of high-paying [9:58] opportunities through investments in AI [10:00] literacy, through investments in in AI [10:02] related apprenticeships. And so I think [10:04] that's point number one is yes, there [10:06] will be challenges, but how do we also [10:08] focus on the tremendous number of [10:10] opportunities there? And then the second [10:12] thing that I'll just quickly say and [10:13] similar to Sarah's point is I think a [10:15] lot of these conversations we believe [10:17] can be more grounded uh in real- time [10:20] data of what we're actually seeing [10:22] because so often I think this can get [10:24] into a game of well what do I think and [10:26] what did my cousin say about what's [10:27] happening at his company and you know [10:29] how many people have been laid off uh [10:31] and and so I think rather than you know [10:33] think about those headlines and [10:35] predictions let's focus on data let's [10:37] focus on how do we have more active [10:39] measures to understand what's happening [10:41] on the ground and and better respond to [10:43] it. And the one that I'll just quickly [10:44] name on our side is that uh as a part of [10:47] the president's AI action plan last [10:48] month, the Department of Labor announced [10:50] that we are launching a AI workforce [10:53] research hub to have an ongoing analysis [10:56] both internally and with with various [10:58] private partners to really think about [11:00] what that impact is and how we can [11:02] better support workers in in navigating [11:04] that new reality. [11:05] >> Thanks, Taylor. Super exciting stuff you [11:07] guys are doing. Um, a lot of people are [11:09] worried about losing their jobs. Uh, I [11:11] have two daughters, 13 and 10. Um, I'm [11:13] wondering, hey, what type of jobs will [11:14] they have in the future? I'm I'm trying [11:16] to help them use these tools as much as [11:18] they can. They're like, "Dad, stop. We [11:20] understand you like AI, but let us leave [11:22] us alone." Um, Gregor, I'll ask you a [11:25] question. Uh, a lot a lot of people say, [11:26] "Oh, yeah, the jobs are going to [11:27] disappear." Some other people say, [11:29] "Well, not really. There's going to be [11:30] additional jobs." This is the typical [11:31] argument whenever there's technological [11:33] advances. There's there's an argument to [11:35] be made uh that jobs will be transfor [11:37] trans transformed. So before losing a [11:39] job your more what's going to happen in [11:40] the short midterm is your job's going to [11:42] be transformed. Tasks that you used to [11:43] do will no longer be done by you. It'll [11:45] be done by AI. That leads to also [11:47] questions of like does do the people [11:49] have the optimal skill set to be able to [11:51] do those tasks that only humans can do. [11:54] What are your takes on this as a someone [11:55] who studies these things? [11:57] >> Um so first off I want to highlight the [11:59] word task that you used. um which I [12:01] think is really important in thinking [12:02] about this and it's also something that [12:03] comes through in the report from [12:05] anthropic that Sarah talked about and in [12:08] what Taylor was saying about sort of [12:09] restructuring of jobs which is that um [12:12] we shouldn't think about jobs as these [12:14] like singular blobs where either your [12:16] your entire job is going to be automated [12:18] or you're going to be fine right it's [12:19] that jobs are bundles of tasks uh [12:22] there's a famous quote from Jeff Hinton [12:23] about sort of radiologists going away um [12:26] and I I show this to to my to in my [12:28] executive courses and I show a list of [12:30] all the tasks that a radiologist does. [12:32] And it turns out that maybe out of 30 [12:34] tasks, two of them AI can do anything [12:36] on, right? And so you should think about [12:38] jobs as bundles of tasks, some of which [12:40] can be automated, some of which cannot [12:42] be automated. And so as a result of AI, [12:44] we're going to have a restructuring of [12:45] what jobs look like. Like we might [12:46] rebundle these tasks into new kinds of [12:48] roles. New kinds of tasks might pop up [12:50] as Taylor's saying about that suddenly [12:53] need to serve this this AI economy. Uh [12:55] but we should not think of sort of jobs [12:57] as wholesale going away or staying in [12:59] some way. It's like it's much more [13:00] subtle about restructuring roles. And I [13:02] think what that implies for how to [13:04] manage this transition because in the [13:05] short run, some people might have jobs [13:07] where lots of tasks can be automated and [13:09] their employer might not be particularly [13:10] good at redeploying them. It's what's [13:12] really important is trying to manage [13:13] that transition well uh to avoid that [13:16] short-term pain of reshuffleling. I [13:17] think what that means is that for [13:18] companies uh that they need to think [13:20] about AI implementation not as oh we're [13:22] going to give everyone an OpenAI account [13:24] or an cloud account and then they're [13:27] just going to do the AI thing on their [13:28] own. It needs to be a strategic issue, [13:30] right? It needs to involve restructuring [13:32] positions, restructuring workflows. It's [13:34] not an IT issue. It is an at the end of [13:37] the day AI based workflows are internal [13:40] products and so they need to involve [13:41] user experience testing, Q&A, right? [13:44] Like there needs to be lots of internal [13:45] strategic discussions about where this [13:47] fits in our workflow and how to [13:48] restructure organizations in order to [13:49] make sure that sort of people redeploy [13:51] to highv value users quickly rather than [13:54] oh we replace some tasks and fire at the [13:56] people and then later on like have to [13:57] hire them back to figure out how to do [13:59] this differently. Yeah. So I think [14:00] strategic discussions are important to [14:01] sort of mitigate this potential [14:03] short-term pain as people are trying to [14:06] figure out what's new and what to do. [14:07] But I think overall there's going to be [14:09] lots of interesting new opportunities. [14:11] >> Thanks Gregor. Um Pav uh in in your [14:14] world working with software engineers [14:16] and with companies and and and [14:18] automating and augmenting a lot of these [14:19] things what are you seeing? Actually [14:21] it's a good topic about tasks because [14:23] that's exactly how it it is segment [14:25] segmentation is happening in software [14:27] engineering because uh for those of you [14:29] who are not familiar with software [14:31] engineering it's actually combination of [14:33] multiple activities and the first and [14:35] the most important activity there that [14:37] you as a human me as a software engineer [14:39] need to understand other humans needs [14:42] and for that I'm using multiple sources [14:44] and senses and AI right now have only [14:48] limited uh input information that it can [14:50] consume. I have also things like empathy [14:53] to you. I can read you non-verbal [14:55] information and I can turn on my BS [14:58] raider and kind of eliminate information [15:00] that is not important. I can also like [15:02] based it on my past experiences and [15:05] immediately pick up important things for [15:07] this specific activity from my past [15:08] experiences that people used to call [15:10] wisdom. Right? So all this kind of [15:12] intuition component of our uh [15:15] conversation that's all important aspect [15:17] for me to grasp all of the details of [15:20] the of the needs that you as a human [15:22] have and then I also need to have like [15:24] skills of engineering to take these [15:26] inputs and process them internally and [15:29] synthesize a solution that will fit your [15:32] purpose and that's an important task [15:35] that I don't see in the current version [15:37] of AI might be done by AI because so [15:40] many inputs and So many skills are not [15:42] yet there and I don't hear yet from the [15:45] PKI producers that they kind of working [15:48] on that like diversification of inputs [15:50] and information and cannot even imagine [15:52] how they can develop this to the moment. [15:55] But after that when you did this [15:56] synthesizing of the solution then you [15:59] move to actual kind of paraphrasing this [16:02] to a computer you put programming code [16:04] you kind of use a different language to [16:06] explain to the computer what need to be [16:08] done how software need to operate and [16:10] that's like second activity and then the [16:13] last activity is actually bring software [16:15] to the user. So you need to like do go [16:17] through the rigorous steps to make sure [16:19] that it's safe and uh uh scalable and [16:22] all of the great aspects and then kind [16:24] of users can use it. So if you split [16:26] these activities like roughly roughly [16:27] because it's very much dependent on the [16:29] software engineer let's take like [16:30] engineering itself analyzing and [16:32] engineering like 20% of the the time and [16:35] like actual coding programming coding [16:37] it's 60% of the time and another 20% to [16:39] deliver this to the end users. So AI [16:42] right now attacks actively this coding [16:45] activity 60% of your time but still that [16:48] is in the early stages. So it very good [16:51] in some activities in the in the coding [16:53] it does it's not yet good and doing the [16:55] whole thing completely and that requires [16:57] like an upscaling [16:59] a reskilling of the software engineers [17:01] that is happening right now because [17:03] software engineers are more expected to [17:04] be engineers and more expected to kind [17:07] of handle this supporting AI tools that [17:11] are augmenting them with the small tasks [17:13] but then when you delegate something and [17:15] when you formulate something to these AI [17:18] uh assistants you also need to check [17:20] back that they produced what you [17:21] expected. So you're still in the loop of [17:23] programming anyway uh even in the [17:25] upcoming future. I don't know like next [17:28] five years. Who knows like what's going [17:29] to happen with AI? Maybe it's going to [17:30] be AGI soon. uh I I cannot understand [17:34] how but it is still uh so and all that [17:37] kind of gives a true picture of [17:39] multiples because you probably hear like [17:41] we're going to replace all the engineers [17:43] soon and it's like AI can be replacement [17:46] for 5x 6x some some excess in terms of [17:51] productivity with all this math that I [17:53] gave it to you in the best case scenario [17:56] maybe double the productivity of [17:59] engineers maybe if like the whole thing [18:01] is taken by properly. It's not yet [18:03] there, but that's like a trajectory [18:04] where software engineering is going to. [18:06] >> Thanks, Pablo. Um, Sarah, you you we [18:09] talked about this data you have, the [18:10] studies you have. You mentioned it [18:12] briefly. Uh, would love for you to like [18:13] maybe do a deep dive in some areas that [18:15] you think is worth it for us to hear, [18:16] especially as it relates to just jobs [18:19] and and the task based approach and how [18:21] things will change. You guys are seeing [18:22] this at different companies. What would [18:24] you say about that? [18:25] >> Sure. Um, so maybe I'll just highlight a [18:29] few things that stand out to me and [18:31] again I want to like emphasize we are in [18:33] the early innings of this data. Um, and [18:35] we'll we are committed to continuing to [18:37] put it out and also um committed to [18:40] working with external researchers. Um, I [18:42] I joke that I could hire 10 economists [18:44] and we would still have thousands of [18:46] questions that we'd all want to answer. [18:47] And so we actually um put the index [18:50] under a economic futures program where [18:52] we're actually giving out grants for [18:53] people to use data and other external [18:55] data sources. So I just put that out [18:56] there for anybody that wants to come do [18:58] research with us. We're we're really [19:00] excited about it. Um you know I think [19:03] there a couple of other things that I [19:04] would highlight beyond what I sort of [19:06] said about some industry concentration [19:08] areas. Um, first going back to this sort [19:11] of augmentation and automation uh [19:14] statistic that we we put our data set [19:16] that we put out. Um, we in our first two [19:20] reports, one was in February and then I [19:22] think we did the next one in April and [19:25] then the the third one was in in May. [19:27] Um, in the first two we saw augmentation [19:31] outpace automated automated tasks and [19:34] that was sort of like yeah that seems [19:36] reasonable like people iterate with [19:39] claude and like it's not you don't [19:40] Claude doesn't do the exact same thing. [19:42] In our third report um we put it out [19:45] right after we released Claude code [19:47] which was our first agentic model. Um [19:50] and we saw that flip between automated [19:53] tasks and augmentative task. Now [19:54] obviously again super early innings [19:58] obvious that we are seeing software [19:59] engineers use um use agents and and use [20:03] AI uh tools in in sort of distinct ways. [20:07] But I did think that that was quite [20:09] noticeable in that like we're seeing [20:11] agents sort of proliferate around um not [20:14] just software engineering tasks uh not [20:16] only from anthropic but from from other [20:18] companies and well as well. And so I [20:20] think that's going to be something that [20:21] we're going to continue to sort of look [20:22] at, you know, how what is this ratio of [20:24] automated versus augmentative tasks. Um [20:27] the second one that I think is just [20:30] interesting and notable again um maybe [20:33] not uh what you wouldn't think but [20:35] something I think we should all talk [20:37] about is um the we studied where the [20:41] concentration was in wages. um and the [20:45] the concentration in wages is um sort of [20:47] in this upper echelon of wages that is [20:50] sort of software engineers. And so I [20:53] think for the first time you're going to [20:54] see a technological shift that um [20:58] affects a wage class that perhaps has [20:59] not been affected in other technological [21:01] revolutions. I'm not saying how that's [21:04] going to be affected, but I do think [21:06] it's something that's quite distinct and [21:07] is something that we are looking at. And [21:09] so I mean the question is like are is [21:12] that sort of wage group going to be [21:15] reskilled? Are certain parts of their [21:17] task or their jobs going to go away? How [21:20] is that going to work? And I think we're [21:21] all sort of watching industry grapple [21:24] with this um in the current moment. [21:26] We're also doing some internal research [21:28] specifically with our our software [21:30] engineers because we think of course um [21:33] like you know the software engineers we [21:35] have are certainly not uh you know [21:37] indicative of what's happening across [21:39] the industry at large but we sort of [21:41] want to know internally and so we'll [21:43] we'll be putting out some stuff around [21:44] around that too. Um but we're we're [21:47] really sort of interested in this and [21:49] then um I think some we've done some [21:52] early research on sort of geographic [21:55] um sort of takeaways and also how it [22:00] affects um geographies with high GDPs. [22:03] And so we're looking to see if like is [22:06] this going to actually be a technology [22:07] that is um uplifting sort of every [22:10] geography or is this something where the [22:12] geographies that are already have high [22:14] GDPs are just going to sort of have [22:16] escape velocity um or or [22:20] does the rising tide lift all boats and [22:22] we don't know yet but that's one of the [22:23] things that we're really interested in. [22:25] >> Super interesting. Thank you Sarah and [22:26] thanks for all the research you guys are [22:27] doing and sharing those insights. [22:29] Taylor, as we're talking about this [22:30] changing jobs, changing skill sets and [22:32] tasks, this new report on talent, please [22:35] tell us a lot about it and give us your [22:37] thoughts on this whole [22:38] >> Yeah, I'd love to talk about the new [22:39] report. Um, we'll briefly just uh add [22:41] one other comment on this this [22:43] displacement notion though because I [22:44] think part of our belief is that a lot [22:46] of this displacement narrative is [22:48] significantly overstated. A couple of [22:50] the reasons have already been mentioned, [22:51] but the one piece I would just quickly [22:53] add is that I think a lot of it is based [22:55] on studies of what can the technology do [22:59] and not necessarily on how might that [23:01] translate into employers decisions about [23:04] to shift tasks and shift roles. I think [23:07] a lot of times when people see this [23:08] notion that you can do more with less [23:11] with AI, they assume that therefore [23:14] employers will do the same with less [23:16] employees when in reality a lot of the [23:19] decisions that we're seeing employers [23:20] make are keeping the same number of [23:22] employees but doing more with it, right? [23:25] And so the AI is seeing that [23:27] productivity, that efficiency gain, but [23:29] not necessarily with a drastically lower [23:31] number of employees. in fact sometimes [23:33] with increased employees. And so just [23:34] one of the many reasons why why we [23:36] believe that uh a lot of the [23:38] displacement narratives are are quite [23:39] overstated. But to get back to your core [23:41] question, we do believe that there will [23:43] be new types of jobs out there. We do [23:45] believe that there will be some type of [23:47] impact in terms of roles that decrease [23:49] in in terms of number that are in the [23:51] economy. And we do believe that across [23:53] many of these roles, there will be new [23:55] sets of tasks and new sets of skills [23:56] that people have to take on. And so part [23:59] of what we put out in our report today [24:01] that was between the Department of [24:03] Labor, the Department of Education, the [24:04] Department of Commerce was really [24:06] showing how we can bring more agility to [24:09] the labor market and to the workforce [24:11] system. Because the reality is we know a [24:13] little bit about how skills and tasks [24:15] will change, but it's going to keep [24:17] changing. And so really what we need to [24:19] build in our into our system is these [24:21] muscles of flexibility that allow state [24:24] workforce agencies and and other [24:26] programs that we support to move quickly [24:28] at saying what are the types of demands [24:31] uh from employers in our local economy? [24:33] What are the types of of skills that [24:35] they're looking for and how do we [24:36] quickly build talent pipelines to to [24:39] better support those needs, better [24:41] address those needs? And so a lot of the [24:43] specific actions that we released today [24:45] have to do with how do we create more [24:47] flexibility, more agility, and more [24:48] innovation uh to continue to reskill the [24:51] population as the economy evolves. [24:55] >> Gregor, um as an economist, an expert in [24:57] this space, where are the productivity [24:59] gains? Where where are the evidence [25:01] studies? There's there's some and [25:02] there's some areas that have shown some [25:04] promise. What are you seeing? What are [25:06] where where's the evidence? Um so I [25:08] think one one import one important thing [25:09] to distinguish there uh is and I think [25:11] Taylor was hinting at this just now is [25:13] that we have lots of studies of large [25:16] productivity gains in the lab right you [25:18] give people a single task you you hand [25:20] them a chatbot and you say uh do this [25:22] task and they tend to be much more [25:24] productive uh when they when they do [25:26] that and you can see for specialized [25:28] tasks like sort of white collar writing [25:30] tasks and software engineering people [25:32] throw out numbers somewhere between like [25:33] 10 and 50% productivity gains um but I [25:36] think there there's an important [25:37] difference from that to the [25:39] implementation in actually existing [25:41] firms, right? Where people don't do [25:43] isolated tasks. People do uh tasks of uh [25:47] really complex workflows that have lots [25:49] of other bottlenecks. So even if you're [25:50] twice as productive, all the other [25:52] bottlenecks are not going away anytime [25:53] soon. Um and so I think the the [25:55] productivity gains there are going to be [25:57] much slower and going to be a roll out [25:59] more over a horizon of like 5 to 10 [26:01] years or something like that. Um the I [26:03] like the expression that uh what AI does [26:06] it goes middle to middle rather than end [26:08] to end. And so in a lot of companies um [26:10] what that means is that they need to [26:12] invest much more in staff that does the [26:14] input part for the AI like structuring [26:16] tasks uh preparing data and so on and [26:18] then deals with the output uh doing [26:20] quality control on that uh putting it [26:22] into useful products for other people um [26:25] or otherwise sort of delivering it to to [26:27] end users. And so shifting those roles [26:29] around um uh takes some time. And so [26:32] just the fact that the middle to middle [26:33] piece suddenly becomes 30% more [26:35] productive doesn't help us if the ends [26:37] are not built out. And so I think as a [26:38] result um in the overall statistics [26:41] maybe we're seeing some small tickups in [26:43] overall um productivity growth. This is [26:44] a bit hard to distinguish from just the [26:45] like uptick in investment in um in the [26:49] uh data centers and um and capacity that [26:52] that's needed for that. But what we're [26:53] mostly seeing is firms investing and [26:55] firms putting lots of money into trying [26:57] to build out these processes, [26:58] experimenting with stuff. Um, and in the [27:00] short run, that looks like we're [27:01] spending lots of money, but we're not [27:02] getting a lot of things back. But it's [27:04] really that you're building this [27:05] organizational uh muscle, as Taylor [27:07] said, um, and this organizational [27:09] capital uh, as you're restructuring your [27:11] your firm to then be able to quickly [27:14] ramp up and have those productivity [27:15] gains uh, a couple years from now. Um, [27:17] so in terms of sort of overall economic [27:19] growth stories, I'm thinking this is [27:21] going to be much more like a slow ramp [27:22] of 1 to 2% maybe additional productivity [27:25] growth per year over the next 10 years [27:26] or so. Um, but because of the way the [27:28] the way that firms are structured and [27:30] how complex the economy is, I don't see [27:32] the like 20 30% like jumps in GDP uh in [27:37] in a in one year to be likely that [27:40] people sometimes talk about when they're [27:41] talking about AGI and sort of runaway [27:43] super intelligence. [27:48] Sarah, um, what are areas you guys are [27:51] surprised where organizations or people [27:54] are not using your tools as much and as [27:56] you guys are building the company and [27:58] thinking about go to market and thinking [27:59] of how the world is changing, where [28:01] would you guys expect more use and more [28:04] transformation in either people's lives [28:06] or organizations or governments? [28:08] >> Um, [28:10] again, I think going back like [28:12] everybody's in the early stages, so we'd [28:14] like to see more more use everywhere. I [28:16] think there's a couple of areas that we [28:17] are specifically working with [28:19] organizations be because we think [28:21] they're going to have outsized impact. [28:23] Um, and so I'll I'll sort of mention [28:25] three. I think healthc care and [28:27] bioscience is a huge one. Um, I think we [28:30] all know this. Uh, you know, and that is [28:33] everything from making sure that your [28:35] doctor has more time to be with you and [28:37] think critically about your illness and [28:39] not doing paperwork because AI can do [28:41] that. And that's um and then on the [28:43] other side of it um you know the the [28:46] gains that we'll see in biomedical [28:48] research because you know a a AI model [28:51] can actually run research so much faster [28:53] and and and than than a than a human or [28:55] a group of humans can. Um and so I think [28:58] we're really investing in both um what [29:01] you know what we call beneficial [29:02] deployments in terms of invest anthropic [29:05] actually investing with with [29:06] organizations in that but also seeing [29:08] our our our users doing that as well. [29:10] Education is another one. You know, we [29:12] we working with the White House on [29:14] education um and many other [29:16] organizations. But I think there's going [29:17] to be this transformation that um you [29:20] know, we see teachers wanting to use [29:21] this and figuring out how to use it um [29:23] and not again not using for cheating, [29:25] but actually seeing productivity gains [29:27] and being able to tailor um curriculum [29:30] to students that have, you know, every [29:33] student has a different need and AI can [29:34] do that. Um and then I think another one [29:37] is is frankly government services. I [29:39] mean, I think we all I I call it my like [29:41] California DMV problem. I never want to [29:43] go to the DMV and get a new license [29:44] again, but like imagine if AI could make [29:46] it like really fast and cut all that [29:48] bureaucracy. Um, but you know, I think [29:51] that that's an area where, you know, you [29:52] can speak to this, but you know, we're [29:54] really excited to see both the federal [29:56] government, but also state governments [29:57] and local governments really getting [29:59] into into the business of AI and um [30:02] making that investment. You know, I'm a [30:03] I'm a alumni of the the State Department [30:06] and when Coen Pal invested in putting [30:08] computers on every desk, it took a [30:10] really long time to move from like, you [30:12] know, typewriter cables to like, you [30:13] know, we could send an email message to [30:15] an embassy. You know, it's not [30:17] overnight, but I think that there's [30:18] going to be huge gains in um government [30:21] services. And I also think that that [30:22] puts trust back in institutions, both on [30:24] government institutions, but also um [30:26] sometimes nonprofit institutions, which [30:28] could be using these tools um for really [30:31] exciting things to deliver services to [30:33] people that really need them um when [30:35] they need them, where they need them. [30:36] And so I think that those would be some [30:37] areas we're we're particularly excited [30:39] about. [30:40] >> Thanks, Sarah. So as uh we're seeing and [30:42] we'll see more and more outsourcing of [30:45] tasks and and things that humans do to [30:47] AI and humans doing other things. Uh [30:49] Taylor, what how should we think about [30:52] the skills and competencies that humans [30:54] should develop more and more in an age [30:56] of AI? [30:57] >> Yeah, I think our our belief is that the [31:00] first priority is really a foundational [31:02] AI literacy, which is which is not the [31:04] entire answer, but we do believe it's [31:06] the first step. And so one of the things [31:08] that you've seen uh from this [31:09] administration is a tremendous [31:11] prioritization on AI literacy across all [31:15] education and workforce funding streams [31:17] because we really believe as these new [31:19] jobs and new opportunities, new [31:21] opportunities for productivity arise, [31:24] that will be the table stakes element [31:27] that people will need for that. And so [31:28] we're really working to say how do we [31:30] work this into our K- through2 systems? [31:32] is how do we work this into our career [31:34] and technical education programs, our [31:35] post-secondary institutions, um our our [31:38] public workforce uh agencies that [31:41] support unemployed workers, but also [31:43] working with companies to actually say [31:45] how do you invest uh in your own current [31:48] workforce to make sure that they are [31:50] able to see these different gains uh and [31:52] new forms of opportunity. And so I think [31:54] point number one is just those [31:55] foundational AI literacy skills that [31:58] that allow them to to be a part of the [32:00] the conversation. The the second thing I [32:02] would point to is that back to this idea [32:04] of the shifting of skills within jobs. I [32:07] think one of the things that we're [32:08] seeing is that as AI can take over uh [32:12] some of the different tasks and do it in [32:14] a faster and potentially more effective [32:16] way than humans can, the different tasks [32:18] that are left over that humans are [32:21] really able to spend more time on are [32:22] are the interpersonal skills are are [32:25] these elements of and a great example is [32:27] is in the medical fields. uh you know [32:29] doctors are are seeing tremendous [32:31] ability to use AI to figure out tricky [32:35] diagnosis and uh better personalized [32:38] treatment plans but that human [32:40] relationship that trust building that [32:42] empathy um and all of those [32:44] interpersonal skills are so so unique to [32:46] humans and and so I think yes it's the [32:49] AI skills being able to manage interact [32:51] uh and in some cases even build but also [32:54] saying what are those complimentary [32:55] human skills that will even be more [32:58] important in in this new economy and so [33:00] that's something that we're also going [33:00] to be making a lot of investments in. [33:02] >> Great. Thanks, Taylor. Pavville, um [33:04] you're in the trenches with your company [33:06] working with customers, corporations, [33:08] like what how are you seeing things [33:09] change and how are you guys creating [33:11] value for them? What what are like what [33:13] AI tools you guys are building? What are [33:15] they doing? What are humans still doing? [33:17] Tell us a little about your perspect [33:18] perspectives. [33:20] we uh in the in the continuous process [33:23] of upskilling and reskilling our [33:25] workforce for sure and that's that's [33:27] what we observe right now and actually [33:28] about the data there is some data [33:29] already from the trenches it's not about [33:31] just wipe coding because everybody [33:33] probably heard about coding that you you [33:36] ask AI to do something through the [33:37] prompt and then it produces you almost [33:39] ready to go software the challenge with [33:41] this software that it looks like it's [33:43] ready to go but then the issue with the [33:45] software that's always evolving so you [33:47] need to do adjustments you need to make [33:49] amendments to the software to keep it on [33:51] track with the real human needs and [33:53] that's what V coding fails on. So you [33:56] can create some good prototypes. So you [33:57] can cons confirm the idea. But then uh [34:01] in deep implementation you really still [34:03] need like engineers and uh deeper [34:05] evolution of that and in enterprise [34:07] world there is a great project from [34:09] Stanford that actually observing right [34:12] now hundreds of teams and like thousands [34:15] of engineers and what they figured that [34:17] for the green field area so like when [34:20] you're creating something from ground [34:22] zero then you can see efficiency gains [34:25] like up to 60% depending on technology [34:27] that you use. But if you go into the [34:29] brown field that is called legacy in our [34:32] world when you work with a old systems [34:34] and so on that might go to maybe in the [34:37] best case scenario like 15 20% of the [34:39] current tool set but in some scenarios [34:41] it might go actually into negative [34:43] percentages because the technology is [34:46] not yet there to give you like outdated [34:49] legacy stack uh support and as a result [34:52] engineers spend more time on integration [34:54] of what AI produced rather than just [34:56] leveraging that. So these are kind of [34:58] numbers. So like in our industry for now [35:00] with the current tool set with the [35:01] current uh level of evolution it's maybe [35:04] 15 20% gain that you can get if your [35:08] software engineers actually reskill [35:10] themsel because that's another problem [35:12] of AI uh real problem that you're [35:15] actually getting into the very intimate [35:17] zone of human doings. You you actually [35:21] not giving them process, you're giving [35:23] them other tools that they need to [35:24] integrate into everyday job that [35:26] actually happening within their minds, [35:29] brains, body, uh uh tactical feelings, [35:33] everything and they need to integrate [35:35] the skills and that's going to take some [35:36] time. Humans need to have some time to [35:38] adjust and reevaluate and that's gives [35:41] us a hope for not in like an immediate [35:44] jump on the invenation of the jobs and [35:46] it it actually empowering of us to be a [35:49] greater humans to other humans. That's [35:51] that's like my view from the trenches. [35:53] you pav as we're ending or getting close [35:55] to the end of the panel I want to ask [35:56] each of you to share uh something brief [35:59] with the audience something that we may [36:01] not have touched upon like what is the [36:03] best advice you can give them in this AI [36:04] age of AI and how it's going to impact [36:06] the economy the world how it's going to [36:07] potentially transform our everyday lives [36:09] our relationships Taylor what what [36:12] nugget of wisdom can you give on this [36:13] topic [36:15] >> I I would say that the biggest thing [36:16] that I'm thinking about is how do we [36:18] shift this conversation around AI and [36:20] jobs from more of a fearbased narrative [36:23] to more of an optimism based narrative. [36:25] Uh there was a stat that came out in in [36:27] Mary Mer's report a couple months ago on [36:29] AI around how in China AI is viewed [36:33] versus how in in America AI is viewed [36:35] and what percentage of the population is [36:37] viewing this as a net positive. And it [36:39] was 83% in China and 39% in the US. And [36:42] I think the the risk that I see in that [36:45] is that again as all these opportunities [36:47] that we're already seeing are created by [36:49] AI in the economy, how many Americans [36:52] will be able to to pursue those? How [36:53] many Americans will are, you know, be [36:56] developing the AI literacy skills and uh [36:58] and interest and creativity to be able [37:01] to to take advantage of the the [37:03] opportunities that were created. We want [37:05] this wave of opportunity to be as widely [37:07] as accessible as possible across the [37:09] country regardless of geography, [37:11] regardless of education level, income [37:12] level. And so how do we shift this [37:14] overall narrative from fear to to op [37:17] optimism? [37:18] >> Thank you, Sarah. [37:19] >> Couldn't agree with you more. I think we [37:21] have I think it's not this is the moment [37:24] for America to win. It is about how we [37:27] talk about it, how we use it, how we [37:29] protect it, how we integrate it into our [37:32] lives and into our national security [37:34] apparatus. Anyway, we we we can talk [37:37] about that after. I guess my my one [37:38] parting thought is like in God we trust [37:41] and everybody else bring data. Let's [37:43] talk about these things in like real [37:45] terms and watch them over time because I [37:47] think the data will help us all do uh [37:51] make this transformation smoother for [37:53] the most amount of people. whether they [37:55] are in an industry, in a sector, in a [37:57] community, um in a classroom. Uh and the [38:01] good hopefully the good news about AI is [38:03] that we it it moves at the speed of [38:06] light. And so if you have real time data [38:09] or close to real-time data, you can make [38:11] those adjustments quickly and flexibly [38:13] and fail and and make a different [38:16] choice. Um and we're not particularly [38:18] good at that in general in our society, [38:20] but I think that this is sort of the [38:22] challenge and opportunity for all of us. [38:26] Yeah, I believe the the biggest problem [38:28] here is our narrative, public narrative [38:30] because uh the public narrative is [38:31] actually the AI will take that and AI [38:34] will remove that and it will replace [38:36] these people. But in reality like there [38:39] is a there is a gap in in the in the [38:41] surveys when you when you survey top [38:44] managers of the companies they believe [38:45] that AI is not used enough in their [38:47] companies for the workforce. But if you [38:50] go like into oneonone conversations, you [38:52] figure that people actually using AI [38:54] even they saying that it's not going to [38:56] happen. And that's what's happening like [38:58] it's it's a delta between real adoption [39:00] and what is expected from the corporate [39:03] perspective. And I believe it's a it's a [39:04] derivative from the narrative because we [39:06] instead of saying that you guys [39:09] welcome to use AI in everything that you [39:12] need to do and it's not an embarrassing [39:15] thing because the most kind of striking [39:17] feedback that I've got that people are [39:19] saying I'm embarrassed to share that I'm [39:21] using AI because I feel like a cheater [39:24] and that like flip over that [39:26] conversation and narrative that you are [39:28] not a cheater you are new version of the [39:31] human being that actually empower [39:33] powered by such a great technology [39:35] that's like a narrative that can speed [39:37] up adoption implementation of [39:40] >> thanks Pablo Gregor [39:42] >> I agree with everyone else that I we [39:44] need a broad change in attitude towards [39:46] sort of reskilling and sort of grasping [39:47] this opportunity what I want to [39:49] highlight I think is a it's a bit of a [39:50] blind spot in terms of the overall um [39:53] policy discourse and research on this um [39:56] which I think are the large household [39:59] benefits of generative AI right and so [40:01] this is related to some research that [40:02] I'm currently working on that's [40:03] hopefully coming out soon which is this [40:04] idea that if you look at usage [40:06] statistics more people are using [40:08] chatbots genative AI in their personal [40:10] lives than do at work. Um and so and [40:13] that's a very large dimension of welfare [40:16] benefit uh resulting from this. You know [40:18] people optimizing their lives asking for [40:20] information teaching themselves things [40:22] that all happen outside of a corporate [40:23] context. Um, but that's a benefit that's [40:26] largely unmeasured in a lot of the the [40:28] studies related to this where we're just [40:30] focusing narrowly on the impact on jobs [40:32] and and and so on. But there's a broad [40:34] consumer user benefit to these tools [40:37] that that always needs to be uh taken [40:39] into account, right? There's lots of [40:41] people benefiting from them. Um, and so [40:43] that always needs to be weighed against [40:44] sort of uh short-term cost in terms of [40:47] trying to implement this correctly in [40:48] the workplace. [40:49] >> Great. Thank you so much, Gregor, and [40:51] all of you. I just want to highlight a [40:53] couple things that were said which I [40:54] think are very important. We're just in [40:55] the beginning of this. Um there's a lot [40:57] of AI anxiety. I think that's normal. I [41:00] think we should switch to a uh as Taylor [41:02] said, there's so many opportunities. So [41:04] focusing on the opportunities and really [41:06] AI skilling ourselves, our our families, [41:09] our friends, our companies um I think is [41:11] a big opportunity right now. Uh AI is as [41:14] economists call it is a general purpose [41:15] technology. Technologies that transform [41:17] the world, societies, economies. So, [41:19] electricity, the internet, steam engine. [41:22] Um, so it's going to be interesting to [41:23] see how this happens. There's no [41:24] playbooks yet, and I think there's an [41:26] opportunity for us and our [41:27] organizations, our governments to lead. [41:30] So, with that, uh, I ask you guys to [41:31] thank the panelists with me, and thank [41:33] you so much for being here.