The Economic Impact of AI: Jobs, Productivity, and Future Economies

人工智能的经济影响：就业、生产力与未来经济 专家小组探讨了人工智能如何重塑现代工作模式，将核心经济价值从机械的技术执行转向审慎的验证与综合统筹。尽管企业数据表明深度应用 AI 能够推动商业扩张与招聘，与会嘉宾仍深入剖析了企业裁员的现状，将表演性质的“AI 洗白”（AI washing）与科技行业中已有记录的初级岗位实质性缩减区分开来。专家们探讨了被迫摒弃既有习惯的资深员工与适应力强的初级人才所面临的摩擦，以及初级研究学徒岗位受到的严重挤压。讨论最终得出结论：要缓解财富过度集中并妥善应对这一指数级转型，亟需前瞻性的政策设计、持续的教育支持以及独立的国际治理。 推动当前 AI 资本支出热潮的核心在于模型技术能力与实际经济效用的稳步提升，这种中长期价值增长将持续吸引对底层基础设施的巨额投资。 (1:34) 职场核心价值正从底层机械编码转向高阶批判性思维与输出验证；企业软件开发也逐步向 AI 编排演进，推动交付速度提升 5 至 20 倍。 (3:16) CalMatters 的“数字民主”项目利用符号 AI 与规则机器学习弥补地方新闻业的人力空缺，将议会发言直接关联至竞选资金数据，并支持带时间戳的视频溯源。 (13:31) 深度应用 AI 的企业在追踪数据中呈现出整体员工规模扩大的趋势，多位专家指出近期大型科技企业的裁员多属于消化疫情期过度招聘的“AI 洗白”借口。 (23:42) AI 的普及正在剧烈扁平化企业层级并冲击中层管理；尽管无包袱的初级技术人员适应迅速，但在研究等领域，初级岗位的萎缩也可能切断培养高阶专业直觉的传统路径。 (40:33)

Artificial intelligence is restructuring enterprise workflows from manual execution toward high-level orchestration, but whether this transition drives broad prosperity or labor displacement depends on technological design, institutional policy, and the speed of adoption. While proponents highlight productivity gains that can expand commercial capacity, economists and researchers warn of acute friction in entry-level hiring pipelines, middle-management cuts, and compounding wealth concentration. Transition Speed and the Automation Trajectory (18:32) Anton Korinek observes that continuous gains in technical capability and utility justify rising capital expenditures in AI infrastructure, but the ultimate macroeconomic outcome depends on whether models augment human workers or replace them entirely. While AI has largely served as a complementary tool, Korinek emphasizes this was a design choice rather than an inherent economic law; shifting toward autonomous end-to-end execution could depress aggregate labor demand and wages. Drawing comparisons to historical structural shifts, such as the two-century transition away from agricultural labor, Korinek stresses that the central risk is transition speed. Society absorbed recent incremental tools, but an abrupt, exponential acceleration could exceed the structural capacity of institutions and labor markets to adapt. Enterprise Restructuring and the Shift to Orchestration (8:14) Within enterprises, panel evidence indicates that AI transforms the composition of work rather than simply eliminating headcounts. Bob Wollheim describes how software development at CI&T transitioned from legacy methods to AI-native orchestration under the philosophy 'AI codes, we build,' enabling teams to deliver solutions at 5 to 20 times their prior speed. In commercial applications like Domino's delivery operations, AI was deployed to optimize back-office demand patterns and dynamic promotions rather than consumer-facing novelty bots. Foaad Khosmood argues that economic value in this environment moves away from rote mechanics and syntax toward critical thinking and verification. To mitigate errors and hallucinations, Khosmood highlights CalMatters' Digital Democracy platform, which pairs traditional rule-based algorithms with timestamped primary video records to assist investigative journalists rather than generating fully automated narrative copy. The Junior Hiring Paradox and Flattened Hierarchies (35:34) Corporate structures are rapidly flattening as automated systems take over repetitive aggregation and reporting tasks, disproportionately exposing middle management to layoffs. Regarding entry-level talent, the panelists present competing empirical realities. Wollheim cites vendor tracking data showing that AI-intensive companies hired 10% more workers overall—growing hiring 7% in the first six months and 32% over 18 months—and notes CI&T successfully onboarded 400 junior trainees who adopted AI tools rapidly because they had no legacy habits to unlearn. He characterizes tech industry layoffs as corporate 'AI washing' to disguise pandemic-era overhiring. Conversely, Korinek cites the economic study 'Canaries in the Coal Mine,' which documents measurable hiring contractions among entry-level programmers. Elizabeth Rhodes reports that at OpenResearch, junior research assistant positions are increasingly difficult to justify when automated systems produce baseline research drafts in hours rather than weeks. Rhodes warns this contraction risks severing the foundational apprenticeship pipeline where junior professionals traditionally cultivate the domain judgment and intuition required to guide and prompt AI systems effectively. Distributional Disparities, Social Safety Nets, and Governance (4:57) Addressing wealth concentration from frontier model developers, Rhodes rejects technological determinism, arguing that inequality is shaped by deliberate institutional choices and policy frameworks rather than technology alone. Her field research with low-income families and rural workers in Texas, Illinois, and West Virginia shows that immediate survival challenges eclipse technological adaptation. Because foundation models evolve faster than traditional decade-long retraining cycles, Rhodes and Khosmood argue that skill acquisition cannot be treated solely as an individual burden without robust social safety nets and childcare support. On systemic oversight, Khosmood compares unchecked AI development to the regulatory inaction that previously enabled corporate data harvesting, advocating for baseline state regulation paired with an independent international consortium to enforce binding compliance rules on frontier developers. The discussion leaves unresolved whether market economies can preserve entry-level apprenticeship pathways as baseline cognitive tasks are automated, and whether regulators can construct effective national and international governance mechanisms before rapid capability advances outpace existing institutional guardrails.

在Ai4研讨会上，与会专家指出，人工智能的技术实用性提升正推动庞大资本支出与开发范式重塑，将核心价值从机械编码转向系统编排与批判性思维；然而，关于AI究竟会扩大就业还是加速替代，以及它对阶级不平等与组织层级的冲击，学界与业界仍充满高度不确定性与分歧。 资本支出、生产力提速与招聘反差 (1:34) 弗吉尼亚大学教授、Anthropic经济学家Anton Korinek指出，尽管短期市场存在波动，但AI技术能力与实用价值的持续演进，构成了支撑基础设施资本支出（Capex）的核心经济基本面。CI&T执行副总裁Bob Wollheim以企业软件交付为例指出，行业范式已跃升至“AI编写代码，人类负责构建”，使交付速度提升5至20倍；例如Domino's利用AI动态配置促销策略以替代繁琐的月度人工规划。针对科技业裁员潮，Wollheim反驳了“AI导致失业”的叙事，认为裁员主要是科技巨头纠正疫情期间过度招聘的公关掩护（即“AI洗白”，AI washing）。 Wollheim援引企业采购数据指出，深度采用AI的企业在整体招聘上比同行高出10%，在采用前6个月招聘增长7%，18个月内增幅达到32%，表明生产力飞跃拓展了业务规模而非单纯缩减人手。然而Korinek持谨慎态度，并援引经济学论文《煤矿中的金丝雀》指出，尽管整体净就业数据仍存争议，但在计算机编程等高度暴露的岗位上，初级职位的招聘萎缩已呈现可量化的客观事实。 互补与替代博弈：转型速度决定适应成败 (18:32) 针对AI究竟创造还是摧毁就业，Korinek坦言自身“高度不确定”。他警示称，近年来AI对知识工作者的赋能主要体现为“互补性”，但这更多是人为的技术设计取向，而非不可逆的自然法则。一旦软件开发走向自主端到端执行，互补关系可能迅速转变为纯粹的人力替代，导致整体劳动需求下滑并压低工资水平。 加州理工州立大学教授Foaad Khosmood列举历史先例指出，自动化曾彻底淘汰电话接线员，而1990年代科技公司内部普遍设立的印刷电路板（PCB）与芯片设计团队也在十年内被外部代工基地取代。Korinek认为，尽管两个世纪前超过90%的人口从事农业并成功实现职业转型，但当前变革的核心变量在于“转型速度”。社会尚能平稳吸收过去两年的渐进式演变，但若技术发生指数级跃迁，其冲击速度将可能超出社会制度与就业市场的承载极限。 组织架构扁平化与初级人才的断层悖论 (35:34) 在企业日常运作中，Wollheim发现日常报表与进度追踪等中层管理职能正被AI智能体快速取代，企业组织正急剧扁平化，打破了原有的晋升阶梯。而在代码开发层面，资深工程师面临痛苦的“习惯遗忘”（unlearning）阻力；相反，CI&T招聘的400名初级培训生由于没有旧开发模式的思维包袱，能够直接基于AI原生范式构建软件，展现出极高的生产力。 然而，OpenResearch研究总监Elizabeth Rhodes从非营利研究机构的视角提出了不同考量。在经费受限的组织中，AI系统只需数分钟至数小时即可生成初级研究助理（RA）以往耗时一个月才能完成的草稿，迫使机构难以保留传统RA职位。Rhodes警告称，这种收缩切断了初级员工积累专业直觉和提示词辨析能力的“学徒期”阶梯。针对“AI仅局限于低端任务”的假设，Korinek补充强调，AI已在数学定理证明与系统漏洞挖掘中展现出“批判性思维的微光”，未来五年其能力边界绝不可被轻易设限。 结构性不平等、底层防线与全球监管构想 (4:57) 针对前沿实验室高估值与潜在IPO对财富分配的冲击，Rhodes强调不平等并非技术决定的宿命，而是制度与选择的结果。结合对西弗吉尼亚州农村社工及得州、伊州低收入母亲群体的田野研究，她指出许多底层劳工在生存线上挣扎，无法独立承担技能转型的沉重成本。单纯的现金转移并不能解决问题，若AI研发仅服务于精英群体，技术将进一步拉大阶级差距，决策者必须在工具设计上覆盖更广泛的社会现实痛点与多语言需求。 在技能储备与宏观风险管控上，Khosmood认为未来最重要的素养是高阶批判性思维，用于甄别和校准那些“看似合理但潜藏错误”的自动化结果；同时，计算范式应逐步摆脱类似“看百万小时下落视频去学重力公式”的高能耗暴力训练，转向端侧轻量化模型。针对奇点与安全失控风险，Khosmood回顾了过去二十年对社交媒体数据泛滥监管滞后的教训，呼吁在国家基线监管之外，组建类似国际防核扩散机制的独立国际联盟，为前沿AI模型设定具备强制约束力的安全规范。 整场研讨表明，AI对经济的冲击并非简单的劳动力总量置换，而是一场由转型速度驱动的深远结构重塑。然而，在初级工作被压缩的背景下如何维系专业人才梯队、如何化解中层管理空心化带来的晋升体系失效，以及如何在兼顾底层弱势群体安全网的同时建立全球统一监管框架，仍是学界、政界与产业界尚未攻克的核心议题。

## Background 该视频由频道 Ai4 于 2026 年 8 月 22 日发布在 YouTube 平台上。视频记录了一场关于人工智能对经济影响的高端研讨会，汇集了来自 Anthropic、OpenResearch、CI&T、加州理工州立大学及《华盛顿邮报》的专家学者与行业高管，共同探讨 AI 在重塑生产力、重构劳动力市场、加剧或缓解收入不平等以及影响未来经济结构方面的深远变革。 ## TL;DR 研讨会深入剖析了当前 AI 投资热潮与劳动力市场的实质变化，指出劳动力的核心价值正从基础编码转向批判性思维、结果验证与全局编排能力。嘉宾们揭示了许多科技大厂将疫情期间过度招聘的裁员归咎于“AI 洗白”的公关策略，并探讨了企业组织扁平化过程中中层管理者受到的巨大冲击与初级员工面临的机遇与断层。最后，与会者强调不平等是制度与政策选择的结果而非技术宿命，呼吁建立健全的社会安全网兜底机制以及独立的全球 AI 治理监管体系。 ## Atmosphere & style 整场研讨会节奏紧凑、气氛严谨且充满思想交锋，既有宏观经济学的理性推演，又有非营利科研视角的人文关怀与企业实战派的坦率直言。企业高管的务实乐观与学者专家的审慎警惕形成了极具深度的张力，发言者毫不避讳地拆穿行业公关话术与盲目恐慌论。聆听过程既能感受到前沿技术带来的颠覆性兴奋，也能体会到关于底层劳工生存困境与组织转轨阵痛的现实沉重感。 ## Key takeaways - [01:34](https://www.youtube.com/watch?v=CvHKbdFiUns&t=94s) AI 投资与资本支出的长期支撑：中长期来看，全球资本市场向 AI 领域的持续注资，核心驱动力在于模型与技术本身的实用性（Usefulness）以及其在全经济范围内创造的实际经济价值。 - [03:16](https://www.youtube.com/watch?v=CvHKbdFiUns&t=196s) 核心技能重心转移：随着代码与基础内容生成的自动化，劳动力的核心竞争力正从单纯的基础编码或执行，转向批判性思维（Critical Thinking）、复核复杂产出及工具编排验证能力。 - [05:05](https://www.youtube.com/watch?v=CvHKbdFiUns&t=305s) 不平等是人类的制度选择而非技术宿命：财富向少数科技巨头集中已成现实，但技术是否会拉大贫富差距取决于人类的制度设计，技术研发必须从一开始就兼顾多元群体的普惠性。 - [08:14](https://www.youtube.com/watch?v=CvHKbdFiUns&t=494s) “AI 编码，人类构建”的范式转变：企业正在将基础开发工作移交 AI，开发速度提升数倍至数十倍；传统初级程序员需求减少，而擅长顶层设计与系统编排的 AI 架构师和 AI 工程师急剧紧缺。 - [13:31](https://www.youtube.com/watch?v=CvHKbdFiUns&t=811s) AI 填补公共监督真空而非取代人类：加州“数字民主”项目通过 AI 记录和分析立法机关言行，有效填补了地方媒体因长期萎缩而留下的监督空白，以人机协作模式赋能调查记者。 - [18:32](https://www.youtube.com/watch?v=CvHKbdFiUns&t=1112s) 互补性与完全替代的临界翻转风险：目前 AI 主要充当知识工作者的强互补工具，但这源于开发者主动设置的权限限制；未来技术能力一旦突破特定临界点，存在从辅助转向全面替代关键职能的风险。 - [21:02](https://www.youtube.com/watch?v=CvHKbdFiUns&t=1262s) 技术转型的核心矛盾在于速度：历史证明社会能够消化技术更迭（如电话接线员淘汰或农业人口转移），但如果当前 AI 的能力跃迁速度过快，整个就业与社会保障体系将面临剧烈震荡。 - [23:42](https://www.youtube.com/watch?v=CvHKbdFiUns&t=1422s) AI 赋能招聘与科技大厂“AI 洗白”：数据显示重度使用 AI 的企业招聘量反而显著增长；大厂近期的裁员潮本质上是在修正疫情期间盲目扩张的人员冗余，AI 仅被高管当作体面的公关借口。 - [26:21](https://www.youtube.com/watch?v=CvHKbdFiUns&t=1581s) 底层阶层忽视与社会安全网缺位：AI 讨论过度聚焦于高知软件工程师，忽略了抗风险能力极低的底层劳工与弱势家庭；若自动化降本的红利无法回馈并建立普惠兜底的安全网，将加剧全社会的生存焦虑。 - [28:53](https://www.youtube.com/watch?v=CvHKbdFiUns&t=1733s) 初级岗位受冲击与“煤矿里的金丝雀”：学术研究警示初级岗位入职机会已在易受 AI 影响的领域出现下滑，结构性转型使得职业培训无法再作为一次性任务，必须演变为贯穿一生的持续重构。 - [35:34](https://www.youtube.com/watch?v=CvHKbdFiUns&t=2134s) 组织架构扁平化与中层管理危机：例行流程和报表型中层管理岗正被智能 Agent 快速替代；资深员工面临痛苦的“反学习”（Unlearning）过程，而没有传统思维包袱的初级员工在原生 AI 协作中展现出更高适应力。 - [40:33](https://www.youtube.com/watch?v=CvHKbdFiUns&t=2433s) 初级人才培育阶梯的断裂隐忧：虽然初级开发者无思维定势，但在预算受限的非营利机构或科研团队中，基础任务被 AI 秒级完成后，资助初级研究助理的必要性降低，可能导致新人积累专业直觉的阶梯被切断。 - [44:32](https://www.youtube.com/watch?v=CvHKbdFiUns&t=2672s) 前瞻性监管与独立跨国治理机制：汲取过去二十年社交媒体缺乏监管导致严重恶果的教训，面对逼近人类智力奇点的 AI 风险，人类亟需建立中立、独立的全球联合治理组织以制定底层约束规则。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=CvHKbdFiUns&t=4s) 开场介绍：AI 资本支出与宏观经济问题 主持人 Yun-Hee Kim（《华盛顿邮报》情报部副主编）开启了本场专题研讨。当前全球科技巨头正在展开规模空前的 AI 资本支出（Capex），资本市场上 AI 概念股估值居高不下，而劳动力市场中初级求职者已开始感受到摩擦与阵痛。针对这一系列现象，本场圆桌会议旨在深入探讨 AI 繁荣对宏观经济、企业运营以及普通职场人士的深远影响。 Yun-Hee Kim 依次介绍了现场的四位专家： - **Foaad Khosmood**：加利福尼亚州立理工大学（Cal Poly）教授，非营利新闻机构 CalMatters 数字民主首席架构师； - **Anton Korinek**：Anthropic 变革性 AI 经济研究主管，弗吉尼亚大学经济学教授； - **Elizabeth Rhodes**：旧金山非营利研究机构 OpenResearch 研究总监； - **Bob Wollheim**：IT 与软件开发服务公司 CI&T 合伙人兼执行副总裁。 --- ### [01:34](https://www.youtube.com/watch?v=CvHKbdFiUns&t=94s) 宏观经济视角：AI 实用性与资本投资的长期驱动 面对日益攀升的 AI 投资规模，Yun-Hee Kim 向 Anton Korinek 提问，探讨未来几年 AI 将带来的经济影响规模与根本性变革。Anton Korinek 首先澄清，自己刚入职 Anthropic 不久，本次发言仅代表个人及弗吉尼亚大学教授的学术立场，不代表公司观点。他主张从经济学基本面来审视这一轮大潮：决定长期走势的核心基本面在于 AI 技术的持续进化以及其实用性（usefulness）的不断跃升，这必然会带动全经济范围内创造的经济价值持续增长。 Anton 指出，尽管经济学家通常不愿对短期的日常市场波动进行投机性预测，但从中长期来看，正是由于 AI 展现出的巨大实用价值，全球资本市场将继续向该领域注入海量资金，以支撑模型开发与基础设施建设，推动投资与资本支出周期呈现指数级扩张。 --- ### [03:16](https://www.youtube.com/watch?v=CvHKbdFiUns&t=196s) 职场冲击：从基础编码到批判性思维与验证能力 针对 AI 对工作与职场带来的最核心变化，Foaad Khosmood 认为最深刻的变革在于“人们对工作本质的认知方式发生根本性转变”。在脑力经济（brain economy）的演进过程中，劳动力的附加值重心正在经历历史性转移：过去对程序员核心技能的定义是“能否写出优质的代码”，而未来的核心竞争力则转向了“批判性思维”（critical thinking）、复核复杂自动化产出的能力，以及掌握各类工具与验证机制来甄别“看似正确但实则存在隐患”的内容。 Foaad 结合自身在计算机科学领域的教学经验指出，以往高校和企业在培养入门级劳动力时，往往过度侧重黑客攻坚与单纯的编程编码技能，忽视了批判性审视与验证能力的系统训练。随着 AI 在代码生成等任务上的普及，这种在计算与工程中保持怀疑、审慎验证的批判性思维将变得极具价值，且这一规律同样适用于计算机之外的各行各业。 ```mermaid flowchart LR accTitle: 劳动力技能价值重心转移示意图 subgraph Past ["过去的核心价值"] A["基础编程与实现能力"] --> B["黑客攻坚技巧与语法产出"] end subgraph Future ["未来的核心价值"] C["批判性思维与审慎复核"] --> D["AI 输出的有效性验证"] D --> E["跨系统工具调度与编排"] end Past -.->|"AI 浪潮推动范式转变"| Future ``` --- ### [05:05](https://www.youtube.com/watch?v=CvHKbdFiUns&t=305s) 财富分配与技术普惠：不平等并非 AI 的宿命，而是人类的选择 围绕 Anthropic、OpenAI 等明星企业潜在 IPO 所带来的硅谷巨额财富聚集，Yun-Hee Kim 询问收入差距是否会进一步恶化。Elizabeth Rhodes 提出了极具洞察力的判断："Inequality isn't something that AI does to us. Inequality is something that we create or prevent by the choices that we make."（不平等并非 AI 强加于我们的客观结果，而是由我们自身的制度与战略选择所促成或避免的）。虽然财富确实在加速向少数企业与投资机构集中，IPO 也会强化这一趋势，但关键在于人类如何设计利益分享与再分配机制。 Elizabeth 强调，根据她在 OpenResearch 的实证研究，单纯依靠“直接发放资金”的现金再分配远远不够，必须建立基于科学证据的完整政策框架。更为本质的议题是技术本身的普惠性与设计初衷： - **服务对象与痛点**：该技术究竟是为谁开发的？解决的是谁的实际问题？ - **文化与语言包容**：模型能够理解谁的语言和表达方式？ - **技能门槛与参与度**：哪些群体拥有有效使用该技术的技能？在技术设计和评估过程中，边缘群体与普通大众是否被纳入考量？ Elizabeth 警告，如果 AI 仅仅用来放大和赋能那些本就拥有高等教育背景与优势资源的精英群体，社会的鸿沟只会越拉越大。她建议企业管理者不要只盯紧短期内利润最高、最显眼的商业用例，而应倾听更广泛的社会声音，使模型本身适应更多元人群的需求，从而在提高生产力的同时，真正实现“扩大机遇”（expand opportunity）。 --- ### [08:14](https://www.youtube.com/watch?v=CvHKbdFiUns&t=494s) 企业落地与岗位演进：从“程序员”到“AI 架构师”的范式转变 作为帮助强生（Johnson & Johnson）、达美乐披萨（Domino's）等大型客户落地 AI 的技术服务商，CI&T 的 Bob Wollheim 幽默地回应了 Anton 的观点：“Anton 说 AI 很有用，我想说它何止有用，简直是‘极其有用’（tremendously useful）。”他指出，软件开发模式经历了从传统的慢速瀑布流（Waterfall），到后来的精益数字化（Lean Digital），而如今迎来的不仅是效率提升，更是一场彻底的颠覆（disruption）。CI&T 内部流行一句话——**“AI codes, we build”**（AI 负责编码，我们负责构建），工程师的角色正由底层的 Coder 蜕变为高阶的 Builder，开发速度实现了 5 倍、10 倍乃至 20 倍的跃升。 以达美乐披萨为例，Bob 说明现代传统企业本质上都已演变为技术和软件驱动型公司。达美乐业务增长的核心支撑在于其配送与订单系统。CI&T 在几个月前为其推出了全新的交付系统，但团队并未跟风开发花哨却低效的聊天机器人，而是将 AI 深度植入到后台（back office）运营中： - 实时洞察与预测局部区域的市场需求； - 动态调整、快速生成并投放针对性的促销方案； - 免除以往人工需要提前一个月规划、结构化思考的繁重周期，实现全业务流程的即时响应与自动化运转。 针对“AI 是否会导致大规模裁员”的担忧，Bob 回顾了去年的恐慌情绪，并直言所谓的“AI 毁灭就业浩劫论”并没有成为现实。虽然具体的重复性任务被 AI 大量替代，但现代业务与技术系统的复杂性（complexity）正在呈爆发式增长。企业对人才的需求总量不降反升，只是岗位形态发生了根本演变——行业不再需要大量基础编码人员，转而急缺**AI 编排师（AI orchestrators）**与**AI 工程师（AI engineers）**等新型复合角色。 ### [13:31](https://www.youtube.com/watch?v=CvHKbdFiUns&t=811s) CalMatters 的“数字民主”项目：利用 AI 重建地方新闻监督 在讨论 AI 是否正在取代人类工作时，Foaad Khosmood 针对主持人的提问回应道，答案是“既是也不是，但主要是‘否’”，这与他们所处的具体领域密切相关。他介绍了由 CalMatters 运营的“数字民主”（Digital Democracy，平台网址为 digitaldemocracy.org）项目。该系统是一个由 AI 驱动的政府透明度工具，旨在记录并提供加利福尼亚州立法机关（包括州参议院和州众议院）中每个人说过的每一句话，并深度挖掘议员与资助者、说客之间的复杂关联网络。系统分为两个版本：一个是面向公众的开放版本；另一个是名为“AI Tip Sheets”的记者专用版，专门为新闻记者发掘报道线索。 Foaad Khosmood 解释了为什么说该项目并没有真正“抢走”记者的饭碗：早在 15 到 20 年前，传统新闻业就遭受了毁灭性打击，大量地方报纸和媒体机构被迫裁撤驻州议会大厦的常驻记者，地方议员的立法活动彻底脱离了公众监督视野，留下的采编团队只剩维持运转的“骨架规模”。数字民主项目的诞生并非取代现有岗位，而是填补这一监督真空，帮助资源匮乏的媒体监控民选官员，极大地节省了调研时间与精力。 这项技术不仅实现了基础监控，还促成了过去完全无法开展的深度数据分析。Foaad Khosmood 分享了一个成功案例：他们去年发布了一篇题为《在加州，权力意味着无需说“不”》（Power is not having to say no in California）的调查报道。该报道揭露了加州立法程序中一项特殊的“未投票”（not voting）机制——它看似弃权，但实际效果等同于投反对票，政客们常以此在不留下公开反对记录的情况下暗中扼杀法案。该调查报道最终斩获了艾美奖（Emmy Award）。他指出，虽然每个行业的情况不同，但在非营利新闻业赋能与政府透明度建设方面，AI 展现了巨大的社会价值。 ```mermaid flowchart TD accTitle: 数字民主系统工作流 subgraph S1["数据摄取与结构化"] A["加州立法机关音视频及发言记录"] --> B["符号/规则与传统机器学习处理"] end subgraph S2["双轨输出应用"] B --> C["公众透明度平台<br/>digitaldemocracy.org"] B --> D["记者专用线索生成<br/>AI Tip Sheets"] end subgraph S3["新闻赋能与监督"] D --> E["地方媒体与调查记者"] E --> F["深度报道与事实问责<br/>如获艾美奖的立法投票调查"] end ``` --- ### [16:30](https://www.youtube.com/watch?v=CvHKbdFiUns&t=990s) 幻觉与准确性控制：基于一手信源与记者把关的防御机制 针对主持人关于 AI 错误和幻觉（Hallucination）风险的提问，Foaad Khosmood 详细阐述了系统的底层架构与质量控制机制。该项目的研发始于 2023 年底大语言模型（LLM）爆发之前，因此其核心主要构建在符号 AI（Symbolic AI）、基于规则的系统（Rule-based AI）以及传统机器学习技术之上，而非不可控的黑盒生成模型。 为了防止虚假或错误信息流入公众报道，系统设置了两道坚固的防线： 1. **一手信源可追溯性**：呈现给记者的每一个事实点都严格绑定并引用原始来源。记者只需点击链接，就能直接跳转到对应的发言视频片段，亲眼验证政客的准确措辞与投票行为。记者在发表报道时也会直接嵌入这些视频链接，从而彻底杜绝了被误读或曲解的指控。 2. **人类处于决策核心**：系统定位是生成新闻提示（Tips）而非自主撰写稿件。最终的故事整合、背景事实核查以及结合现实经验的深度报道，完全由人类记者把关与执笔。 Bob Wollheim 随后补充指出，在实际商业运营中，AI 替代的本质上是重复性劳动（“当我们像机器一样工作时的那些重复任务”），而真正体现人类价值的核心工作并未被取代。未来的用工需求正在发生结构性转移——从事低级重复性工作的人员减少，而负责全局编排（Orchestrating）、协调决策和承担风险的管理与核心角色需求增加，这代表着工作形态的质变而非总工作量的萎缩。 --- ### [18:32](https://www.youtube.com/watch?v=CvHKbdFiUns&t=1112s) 技术替代还是劳动力互补：人类能掌控技术发展方向吗？ 当被问及 AI 是否会消灭某些工种但创造新工种时，Anton Korinek 直言自己对长远走向“高度不确定”（highly uncertain）。他指出，过去几年 AI 在很大程度上充当了知识工作者（包括记者）的“强互补工具”（Complementary Tool）。但他也提醒，数字民主系统之所以呈现出互补性，是因为开发团队做出了明确的设计选择（Design Choice），主动限制其仅提供素材而不让其自动写稿。 Foaad Khosmood 赞同这一观点，并指出若没有大量记者的积极参与反馈，该系统绝不可能成熟；而记者之所以愿意深度合作，正是因为他们确信该技术是在为他们赋能而非取代他们。Anton Korinek 进一步提炼出两点核心启示： - 过去几年 AI 带来的实际冲击比许多人预期的要温和良性。 - 人类社会和技术开发者在技术应用路径上拥有能动性与选择权（Agency and Choice）。 然而，Anton Korinek 强调，AI 必然作为“互补工具”存在并不是一条不可动摇的自然法则。未来存在一种重大风险：技术特性可能在某一临界点发生逆转（Complementarity Flips），从辅助走向全面替代，直接接管更大比例的工作职能。 ```mermaid flowchart LR accTitle: AI对劳动力市场的双重影响路径 subgraph PathA["互补路径（增强模式）"] A1["作为辅助工具应用"] --> A2["知识工作者效率跃升"] A2 --> A3["劳动力需求与薪资获支撑"] end subgraph PathB["替代路径（自动化模式）"] B1["完全自动化关键职能"] --> B2["削减对人类雇员需求"] B2 --> B3["导致薪资承压与结构性失业"] end ``` --- ### [21:02](https://www.youtube.com/watch?v=CvHKbdFiUns&t=1262s) 历史转型速度与当前技术周期的冲击 围绕失业风险的具体情景，Anton Korinek 阐明了纯经济学维度的逻辑：技术越是倾向于赋能互补，劳动生产率、工作岗位和工资水平就会得到提升；一旦企业转向彻底的自动化替代，劳动力总需求将直接下挫，带来工资停滞与失业潮。 Foaad Khosmood 对替代的不可避免性表示赞同，但他将这种演变置于技术历史长河中进行类比： - 20 世纪 40 至 50 年代，自动化技术让传统电话接线员彻底退出历史舞台。 - 20 世纪 90 年代他刚步入职场时，每家科技公司都有自己的硬件设计部门，在内部完成 PCB 电路板和芯片设计；而在随后的 5 到 10 年内，这些设计制造能力迅速集中到少数专业代工厂手中，企业不再需要在内部保留庞大的硬件设计团队。 Anton Korinek 补充道，社会本身需要技术革新和产业更替。如果倒退 200 年，超过 90% 的人类祖先都在从事农业，而今天农业人口的大幅缩减显然是社会进步的体现。他强调，**转型的核心矛盾在于演变速度**：如果保持过去两年的温和推进节奏，社会完全有能力消化吸收这种结构性转移；但如果技术能力发生急剧加速的跃升，整个社会与就业体系将面临极大的震荡。 --- ### [23:42](https://www.youtube.com/watch?v=CvHKbdFiUns&t=1422s) 数据揭示 AI 与招聘正相关：科技大厂裁员真相与借口 Bob Wollheim 提出了一个反直觉的最新统计数据：近期有两家美国初创公司（一家专注支付数据监测，追踪向 OpenAI、Anthropic 等持续采购 AI 服务的企业开销；另一家追踪招聘动态）联合分析了 2021 年至 2026 年（预期）的数据，结果显示： - 重度使用 AI 技术并产生密集资本开销的企业，其招聘人数比其他企业**多出 10%**。 - 企业在深度接入 AI 的前 6 个月，招聘规模平均增长 **7%**；而在持续重度使用 AI 18 个月后，招聘规模增长达到了 **32%**。 针对“既然 AI 提升了生产力，企业为何还要雇佣更多人”的疑问，Bob Wollheim 解释这背后的核心驱动力是**业务的高速增长**。生产力的释放打开了全新的商业机遇和扩张空间，促使企业吸纳更多劳动力。 对于主持人提出的“如果 AI 如此高效，为何科技行业近期出现大规模裁员”这一尖锐问题，Bob Wollheim 直言不讳地指出，这与 AI 毫无关系，AI 只是被科技大厂高管拿来当裁员的挡箭牌。大裁员的真正根源是疫情期间的盲目过度招聘导致组织机构严重臃肿（“Too Fat”）。当企业需要去库存、挤泡沫、削减冗余人力时，声称“因 AI 提升效率而优化人员”成为了应对市场最体面的说辞。他援引知名投资人 Marc Andreessen 的观点印证：在大厂高管眼里这几乎成了一个公开的笑话，各大科技巨头（如亚马逊、Meta、甲骨文、Block 等）将业绩回调期的人员精简归咎于 AI，而资本市场现在已经不再轻信这一套说辞。 ### [26:21](https://www.youtube.com/watch?v=CvHKbdFiUns&t=1581s) 裁员叙事、阶层差异与社会安全网的缺位 关于科技行业裁员以及企业 CEO 们将裁员归咎于 AI 的现象，伊丽莎白·罗兹（Elizabeth Rhodes）指出，公众舆论往往存在盲区。虽然她自己来自旧金山湾区，但她日常接触的研究对象完全不是湾区精英——他们是西弗吉尼亚州偏远乡村的社工，或者是得克萨斯州和伊利诺伊州的低收入母亲。当前关于 AI 的讨论过度聚焦于软件工程师或知识工作者，却忽略了大量本就勉强维持生计的底层群体。例如亚马逊仓库的基层工人，虽然机器人可能随时取代他们的岗位，但他们当下正为下个月的房租发愁，根本无暇去思考 AI 长期会对自己产生何种影响。在制定应对岗位替代与经济转型的政策时，绝不能忽视人们起点上的巨大差距，这背后是数十年投资或缺乏投资所造成的现实。社会必须做出选择：是把适应新经济的沉重负担全部推给个体，还是将其作为全社会共同投资的责任？ 福阿德·霍斯穆德（Foaad Khosmood）对此表示“百分之百赞同”，甚至要“画三条下划线来强调”。他指出，大量枯燥、重复性的工作确实极易被技术和 AI 取代，但公众产生焦虑的根源并不是“枯燥的工作消失了”——如果没有生存威胁，没有人会留恋重复性劳动。问题的核心在于社会安全网究竟在哪里。AI 带来的巨大成本削减和效率收益，目前看来并没有回馈并下沉给整个社会，这才是普遍焦虑的经济根源。 ```mermaid flowchart TD accTitle: AI冲击下的阶层分化与应对机制 subgraph S1["社会现实分歧"] A["知识工作者与湾区精英"] B["底层劳工 / 乡村社工 / 低收入家庭"] end subgraph S2["核心困境"] C["AI 替代与岗位演变焦虑"] D["眼前生计危机：房租、育儿与零抗风险能力"] end subgraph S3["政策抉择路径"] E["个体自行承担转岗适应成本"] F["社会共同投资与安全网兜底"] end A --> C B --> D C & D --> E C & D --> F ``` - **底层生存困境**：对于许多低收入家庭而言，缺乏缓冲资金使得他们无法在面临技术颠覆时从容参与长期职业培训。 - **利益分配失衡**：企业因自动化削减开支所获取的超额利润，尚未形成有效的普惠分配机制。 --- ### [28:53](https://www.youtube.com/watch?v=CvHKbdFiUns&t=1733s) “AI 洗白”与经济学争论：“煤矿里的金丝雀” 安东·科里内克（Anton Korinek）为裁员潮的成因补充了经济学维度的细致视角。他认为，现实中确实普遍存在所谓的“AI 洗白”（AI washing）现象——许多企业仅仅是将出于其他财务或经营原因的裁员，伪装成由 AI 驱动的技术升级。然而，经济学界对这一现象的实质冲击存在激烈争论。其中一篇代表性论文题为《煤矿里的金丝雀》（*Canaries in the Coal Mine*），该研究表明，在极易受 AI 影响的职业领域（尤其是编程与软件开发）中，初级入职员工（entry-level workers）的就业机会相比几年前确实出现了明显下滑。尽管这篇论文本身的方法论在学术界仍有争议，但它确实揭示了初级岗位正在承受实际冲击的苗头。 鲍勃·沃尔海姆（Bob Wollheim）随后补充道，虽然相关研究也观察到初级岗位招聘在某些时段存在短期反弹（例如有数据指出初级岗位招聘增长了 12%），但这仅仅是眼下的短期波动。AI 带来的颠覆性力量极其强大，商业组织、政府部门与监管机构之间必须建立全新的利益协调机制。他澄清自己并非盲目乐观，但他坚信生产力的跃升将催生创业浪潮的爆发：“工作不会凭空消失，而是会被重塑与替代。”然而，如果整个经济与教育体系不进行全方位的重构，传统教育培养出的人才将无法匹配未来的岗位需求，导致更多人被时代抛下。 - **AI 洗白现象**：企业将宏观经济下行或战略失误导致的裁员包装为“拥抱 AI”的公关策略。 - **初级岗位风险**：学术研究《煤矿里的金丝雀》警示，高度暴露于 AI 自动化的工种正率先压缩应届生与新人的准入门槛。 - **系统性重构**：从科研教育到商业运作的全链条必须重新设计，单靠初创企业或投资机构无法独立完成转型。 --- ### [31:54](https://www.youtube.com/watch?v=CvHKbdFiUns&t=1914s) 谁来承担重构责任：终身学习与批判性思维的必要性 针对主持人关于“谁应该为系统重构负责”的提问，伊丽莎白·罗兹强调这必须是一项“共同责任”（shared responsibility）。将职业再培训简单包装成应对技术转型的万灵药，往往成了把结构性转型压力转嫁给普通工人的借口。当前的 AI 革命与以往的技术革命有着本质不同：过去人们或许只需在十年间完成一次技能重塑，但如今底层模型与工具每个月都在进化，成功所需的技能也在不断变动。这意味着再培训不是一次性任务，而是贯穿一生的“持续学习与自我重构”。如果政府和企业不能在培训期间提供基本的生活支持，一个连房租和托儿费都付不起的单亲妈妈根本无从谈起技能升级。 在谈到未来人类最不可或缺的能力时，福阿德·霍斯穆德着重强调了“批判性思维”（critical thinking）与探究精神。当 AI 能够直接输出一个完整的任务结果时，人类往往无法直观判断该结果的效率、质量或长期可维护性。从业者必须具备敏锐的审视能力，时刻追问“这里可能会出什么错？”，并通过多模型交叉验证与压力测试来把关。 此外，福阿德对当前基于大规模预测的“Token 革命”提出了深刻的能效反思。他在课堂上打过一个生动形象的比方： > “这就像是你想向学生传授重力理论，却让他们看了一百万个小时物体落地的视频；但事实上，整套理论本来就只有一个简单的物理公式。如果你一开始就掌握了这个公式，根本不需要看视频并耗费那么庞大的能量。” 他指出，行业最终会意识到粗暴堆叠算力的能效极限，模型端侧化与轻量化正是在笔记本电脑和手机上蓬勃发展的趋势，但当 AI 产出与现实商业决策碰撞时，依然需要具备批判性思维的人类站在舵手的位置上做最终抉择。 - **技能重塑的频次升级**：从“十年一次转行培训”演变为基于模型快速迭代的“终身持续学习”。 - **探究式思维框架**：面对生成式结果，需要主动构建“多模型交叉验证”与“潜在故障预判”的工作流。 - **算法能耗反思**：Token 预测机制存在巨大的能效冗余，未来技术演进将倒逼模型向轻量化与高能效架构转移。 --- ### [35:34](https://www.youtube.com/watch?v=CvHKbdFiUns&t=2134s) 组织扁平化与中层管理危机：“学会遗忘”与初级员工的崛起 当主持人问及 Meta 等大型科技公司调整管理层级、中层管理岗位随着 AI Agent 普及而逐渐消失的现象时，鲍勃·沃尔海姆给出了极为直接的判断：企业的层级结构必然会大幅“扁平化”，大量传统中层管理者正在不可逆转地消失。过去的中层工作本质上充斥着极度重复的例行事务，例如汇总数据、编写周报月报与推进常规流程，因此当自动化来临时，这部分岗位受到的冲击最为剧烈。 鲍勃结合自身企业在全公司推行“AI 增强型软件开发生命周期”（AI-centric Software Development Life Cycle）的实战经历，分享了组织变革中的摩擦与意外收获： ```mermaid flowchart LR accTitle: AI驱动下的企业组织结构演变 subgraph OLD["传统金字塔架构"] direction TB T1["高层管理"] --- M1["庞大中层协调层（报表/流程）"] --- J1["初级执行人员"] end subgraph NEW["AI时代的扁平化架构"] direction TB T2["核心决策与架构层"] --- J2["AI 原生初级员工 + 智能 Agent"] end OLD ==>|"淘汰冗余中层 / 流程自动化"| NEW ``` 在实际推进过程中，组织最先显现的变化就是中层被严重压缩、整体团队精简。此前业内有一种普遍的悲观论调，认为 AI 时代企业将不再需要初级员工，但鲍勃的实践彻底打破了这种迷思：公司当年直接招募了 400 名初级管培生，而这批新人的表现极其出色。原因在于，当前职场面临的最大挑战不是“学习”，而是“学会遗忘旧经验”（unlearn）。那些拥有十几年经验的资深工程师往往陷在传统开发思维中痛苦挣扎，而初级员工如同一张白纸，能够直接以原生的方式利用 AI 工具进行协作与编程。 这种剧烈的结构重塑必然伴随着阵痛与利益冲突。过去企业有着极其清晰的晋升阶梯、职称体系与薪酬梯度，当层级被强行拉平后，大量人员堆叠在同一维度，引发了激烈的内部竞争与薪资评定摩擦。然而，在管理阵痛的背面，积极主动“遗忘旧范式并拥抱新技能”的人员正在为企业开拓出全新的业务空间与营收增长点。 - **中层职能瓦解**：高度依赖例行报表、流程传递的传统管理岗正被 AI Agent 迅速替代。 - **400 名管培生实践**：初级开发者因无历史技能包袱（No learning to unlearn），在 AI 原生编程中的适应速度显著超越部分资深技术人员。 - **晋升体系断裂**：组织扁平化彻底打破了传统的职级阶梯，引发了关于职衔分配、薪酬衡量与内部竞争的深层摩擦。 ### [38:55](https://www.youtube.com/watch?v=CvHKbdFiUns&t=2335s) 职场“反学习”与中高层危机：AI 正在超越低阶任务 在团队协作中，初级培训生在使用 AI 后展现出了惊人的工作热情与生产力爆发。这种现象直接倒逼了资深员工的心态转变——许多高级员工意识到原有的经验模式已不再适用，纷纷感叹：“我必须做出改变了，我愿意彻底推翻以往的做法。”这就是所谓的“反学习”（unlearning）过程。然而，这种转型绝非温和无痛，伴随而来的将是残酷的职场淘汰与阵痛（"there's going to be blood"）。 针对“AI 目前缺乏批判性思维、只能处理低阶琐事”的传统观点，嘉宾提出了强烈预警：AI 系统正在以前所未有的速度迭代进化，其临界点已初现端倪。目前，AI 已经能够证明全球顶尖数学家数十年未能攻克的高难度数学定理，在网络安全领域也能够发现顶尖专家多年未曾察觉的安全漏洞。面对这种技术跃迁，人类必须保持谦逊，绝不能断言五年后 AI 仍旧停留在低端辅助层面。事实上，这种颠覆已经波及组织架构——有高校导师证实，其实习生在经历企业裁员风暴时发现，被裁掉的往往不是底层实习生，而是他们的直属中层管理者。 ```mermaid flowchart TD accTitle: AI对企业层级的冲击机制 subgraph Impact["AI 冲击下的组织变迁"] A["初级员工 / 实习生"] -->|"无历史包袱，快速拥抱 AI"| B["生产力大幅爆发"] C["中高层管理者"] -->|"受限于传统经验，存在抵触"| D["面临反学习与裁员压力"] E["AI 能力跨越"] -->|"数学定理证明 / 攻防漏洞挖掘"| F["向高级批判性思维渗透"] end ``` --- ### [40:33](https://www.youtube.com/watch?v=CvHKbdFiUns&t=2433s) 观众问答：初级岗位的消失与重构争议 来自加州的教育工作者凯莉·麦考伊（Carrie McCoy）向鲍勃·沃尔海姆（Bob Wollheim）提出质疑：面对大量关于初级编码和商业入门岗位缩减的数据报告，如何解释当前初级职位的生存现状？ ```mermaid flowchart LR accTitle: 初级人才在AI时代的发展困境与机遇对照 subgraph Enterprise["企业视角 (Bob Wollheim)"] direction TB E1["传统层级晋升失效"] --> E2["需要无偏见的新视角"] E2 --> E3["资深员工抵触自动化"] E3 --> E4["初级员工无包袱更具适应力"] end subgraph NonProfit["非营利科研视角 (Elizabeth Rhodes)"] direction TB N1["AI 数分钟生成基础产出"] --> N2["初级 RA 缺乏专业直觉与 Prompt 经验"] N2 --> N3["资源受限机构减少初级招聘"] N3 --> N4["传统人才直觉培养路径断裂"] end ``` 鲍勃回应称，自己一年前也曾深感担忧，但实际趋势颠覆了传统认知。过去，初级岗位的逻辑是“从无知起步、逐步跟随前辈学习爬梯子”；但在当前的颠覆性变革中，企业反而最需要这种未被传统经验固化的“白纸”状态。中高层员工往往自视甚高，对自动化和流程变革充满抵触与质疑，甚至为了保住个人职位而阻碍自动化推进。像 IBM 等大型企业在短暂取消初级培训生计划后迅速纠偏，重新招募初级人员，因为他们没有固有思维阻碍。 然而，伊丽莎白·罗兹（Elizabeth Rhodes）结合非营利科研机构的实际运营提出了完全相反的现实困境： - **产出效率的断崖式差距**：以往需要初级研究助理（RA）耗时一个月整理的基础产出，现在利用 AI 几分钟或几小时就能完成同等水平的初稿。 - **直觉与 Prompt 能力缺失**：初级科研人员缺乏足够的专业判断力与直觉，往往不知道如何准确地向 AI 提问和调试结果。 - **人才培育阶梯的断裂**：在预算极其有限的非营利组织中，继续雇佣大量初级 RA 在经济上变得难以支撑；但所有资深研究人员最初都是从 RA 做起并在此过程中积累专业直觉的，这一入门通道的萎缩令人深感忧虑。 --- ### [44:32](https://www.youtube.com/watch?v=CvHKbdFiUns&t=2672s) 奇点、核级风险监管与全球独立治理机制 现场观众提出了关于技术失控与全球治理的深层议题：鉴于大模型越狱攻防、逼近人类智力奇点的狂热军备竞赛等现象，AI 的深远影响已远超工业革命和互联网浪潮。人类是否需要“为了加速而先慢下来”？是否应当将 AI 提升至堪比核武器的高度，通过建立世界级的联合监管机构来实现跨国对齐与管控？ 福阿德·霍斯穆德（Foaad Khosmood）对此表示认同，并指出人类在 20 至 25 年前就犯过严重错误——当时对社交媒体的信息抓取、隐私转售和数据监控缺乏前瞻性监管，最终导致了严重的社会恶果与系统性危机。 针对未来的治理架构，福阿德指出： - **监管的必要性与局限**：鉴于当前的政治气候，完全依赖各国政府可能并非最佳方案，但基础的政府监管依然不可或缺。 - **独立国际联合体构想**：尽管在现实中可能属于一种理想主义设想，但建立一个真正中立、独立的全球财团或国际组织，为各大科技巨头制定具有约束力的底层规则与合规义务，是应对 AI 潜在失控风险的关键路径。 主持人尹熙·金（Yun-Hee Kim）随后宣布本场研讨圆满结束，并向全体与会嘉宾及现场观众致谢。 ## Notable quotes > [05:05](https://www.youtube.com/watch?v=CvHKbdFiUns&t=305s) Inequality isn't something that AI does to us. Inequality is something that we create or prevent by the choices that we make. — Elizabeth Rhodes > [08:14](https://www.youtube.com/watch?v=CvHKbdFiUns&t=494s) AI codes, we build — Bob Wollheim > [13:31](https://www.youtube.com/watch?v=CvHKbdFiUns&t=811s) Power is not having to say no in California — Foaad Khosmood > [31:54](https://www.youtube.com/watch?v=CvHKbdFiUns&t=1914s) 这就像是你想向学生传授重力理论，却让他们看了一百万个小时物体落地的视频；但事实上，整套理论本来就只有一个简单的物理公式。如果你一开始就掌握了这个公式，根本不需要看视频并耗费那么庞大的能量。 — Foaad Khosmood > [28:53](https://www.youtube.com/watch?v=CvHKbdFiUns&t=1733s) 工作不会凭空消失，而是会被重塑与替代。 — Bob Wollheim > [38:55](https://www.youtube.com/watch?v=CvHKbdFiUns&t=2335s) 我必须做出改变了，我愿意彻底推翻以往的做法。 — Bob Wollheim > [38:55](https://www.youtube.com/watch?v=CvHKbdFiUns&t=2335s) there's going to be blood — Bob Wollheim > [08:14](https://www.youtube.com/watch?v=CvHKbdFiUns&t=494s) Anton 说 AI 很有用，我想说它何止有用，简直是‘极其有用’ — Bob Wollheim ## Mentioned Anton Korinek, Elizabeth Rhodes, Yun-Hee Kim, Bob Wollheim, Foaad Khosmood, Carrie McCoy, Marc Andreessen, Anthropic, OpenAI, OpenResearch, The Washington Post, CI&T, Cal Poly, CalMatters, Johnson & Johnson, Domino's, Meta, Amazon, Oracle, Block, IBM, Digital Democracy, AI Tip Sheets, Canaries in the Coal Mine ## Video description Anton Korinek - Head, Transformative AI Economic Studies, Anthropic; Professor of Economics, University of Virginia Elizabeth Rhodes - Research Director, OpenResearch Yun-Hee Kim - Deputy Editor, Washington Post Intelligence, The Washington Post Bob Wollheim - Partner, EVPC I&T Foaad Khosmood - Professor and Chief Architect of Digital Democracy, Cal Poly / CalMatters 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=CvHKbdFiUns)

[5:05] Elizabeth Rhodes: Inequality isn't something that AI does to us. Inequality is something that we create or prevent by the choices that we make. [8:14] Bob Wollheim: AI codes, we build [13:31] Foaad Khosmood: Power is not having to say no in California [31:54] Foaad Khosmood: 这就像是你想向学生传授重力理论，却让他们看了一百万个小时物体落地的视频；但事实上，整套理论本来就只有一个简单的物理公式。如果你一开始就掌握了这个公式，根本不需要看视频并耗费那么庞大的能量。 [28:53] Bob Wollheim: 工作不会凭空消失，而是会被重塑与替代。 [38:55] Bob Wollheim: 我必须做出改变了，我愿意彻底推翻以往的做法。 [38:55] Bob Wollheim: there's going to be blood [8:14] Bob Wollheim: Anton 说 AI 很有用，我想说它何止有用，简直是‘极其有用’

[4:57] Elizabeth Rhodes: inequality isn't something that AI does to us. Inequality [is] something that we create or prevent by the choices that we make. [8:14] Bob Wollheim: AI codes, we build. [13:31] Foaad Khosmood: yes and no, and mostly no. [18:32] Anton Korinek: In two words, I feel highly uncertain about it. [23:42] Bob Wollheim: It's a good excuse because you're too fat. [31:54] Foaad Khosmood: It's like saying that you want to teach about the theory of gravity by watching a million hours of videos of things falling down, where the whole thing is basically just one formula. If you had that to begin with, you wouldn't need to watch the videos and expend all that energy. [38:55] Bob Wollheim: That's the unlearning process. But there's going to be blood. [38:55] Anton Korinek: glimmers of critical thinking

[0:04] Thank you everyone for joining this session  this afternoon. My name is Yunhi Kim and [0:09] um we will talk about the big AI spending that's  happening right now and the impact on the economy. [0:16] So we're seeing massive capital expenditures by  the big tech companies and AI sky-high market [0:21] valuations in AI stocks and some early signs of  labor market friction for new workers. What does [0:28] that mean for the economy, businesses, and all  of us? That's what we will explore in this panel [0:33] discussion today. So, I'm thrilled to introduce  to you our expert panelists for this discussion. [0:40] To my very left is Foyad Cosmood. He's professor  of um California Polytenic State University and [0:49] the chief architect of digital democracy at Cal  Matters, which is a nonprofit news organization. [0:56] Next to him is Anton Corinick who is head of  transformative AI economic studies at Anthropic [1:04] and also a professor of economics at University of  Virginia. Next to him is Elizabeth Rhodess. She's [1:10] a research director at the nonprofit research  organization Open Research based in San Francisco. [1:18] And last but not least, Bob Wolhim, partner,  executive vice president at CIT, an IT and [1:26] software company. So let's start with the the big  picture. Um Anton, the spending on AI keeps on [1:34] increasing. Give us a sense of the scale of the  economic impact of AI um that AI will bring. Um [1:42] what kinds of fundamental changes can we expect  in the next year? Um Thank you very much uh for [1:49] organizing this and I should say I have joined  Enthropic very recently. Uh so I'm still speaking [1:55] under my hat as a professor at the university for  the purposes of today's panel and everything I'm [2:01] saying is just my personal uh perspective on this.  So I think um I like to look at the fundamentals [2:09] uh when it comes to like bigger picture economic  questions like the one that you mentioned and the [2:16] fundamental is uh that AI continues to improve  I think um is very likely to continue to improve [2:24] over the next few years and its usefulness uh  frankly continues to improve and uh for that [2:33] reason I anticipate state that the value that it  can generate across the economy will continue to [2:39] go up and now I don't as an economist I don't like  to speculate about dayto-day market movements and [2:46] so on uh but I think in the medium term uh it is  very likely that we will continue to invest more [2:53] and more uh into uh this technology just because  it's useful so the investment capex inment cycle [3:01] will continue to rise expon exponentially. I I  think the usefulness will continue to rise and [3:08] I expect that in the medium term that's going  to also pull in investment to kind of underpin [3:16] what these systems are producing. Forad um what  is the single biggest change you think AI will [3:22] bring when it comes to work in the workplace?  Uh thank you for inviting me and uh I also feel [3:30] compelled to say that um although Cal Matters is  a a great partner research partner with us uh I [3:37] don't represent Cal Matters uh here. Um so uh in  my opinion the um biggest uh AI um impact will be [3:50] in the way people think about work. And so um you  know we've had transformations like this in the [3:56] past and I believe the sort of the shift in the  value ad in the sort of brain economy will move [4:04] from you know being able to code and code well and  and so on to being able to do critical thinking [4:12] and doublech checkck the work of you know some  complex work that's been done or at least know [4:16] the tools and the buttons to push to be able to  verify things that sort of seem seem seem okay [4:23] but may not be okay. Um that critical thinking  has not always been there at least in entry level [4:29] workforce hasn't been really cultivated. People's  kind of like coding skills and hacking skills [4:33] have been valued you know from what I see from my  students and others. Um but I think very soon uh [4:39] we're going to get to a point where um that that  critical thinking within computing is going to be [4:45] valued. And so this is kind of what I see. I mean  I have a little bit of a narrow focus on because [4:51] I'm in computer science but um yeah critical  thinking I honestly don't think it's going to be [4:57] much different in any other sector and Elizabeth  AI is already creating enormous wealth in Silicon [5:05] Valley with the coming IPOs of anthropic and open  AI. Um what do you think do you think that income [5:12] disparity will likely increase as a result?  Well, I think it's important to recognize that [5:20] inequality isn't something that AI does to us.  Inequality um something that we create or prevent [5:29] um by the choices that we make. So obviously you  know there is enormous wealth being generated. [5:36] It is being concentrated and I think the IPOs  will continue that trend. But at the same time [5:44] um you know it does raise questions about  redistribution about how do we ensure that the [5:49] gains the economic gains are uh distributed and  I I think redistribution is really hard uh to do [5:58] it well to get the incentives right I think it's  going to take more than just in my own work I've [6:04] learned that just giving um money isn't going to  be enough and so I I think thinking through that [6:09] um and creating a policy framework there based  on evidence is important. But I also think the [6:15] question is actually a little bit broader. It's  it's also about, you know, who the technology is [6:21] built for and who benefits from it. Like whose  problems does it does it solve? Like whose [6:28] language does it understand? Who has the skills  and expertise to you know use it effectively? [6:35] and who what communities and individuals um you  know are included when they think about how to [6:40] design and evaluate it. And so I think you know  in addition to how do we redistribute the gains [6:47] or how do we ensure that economic money received  by AI is more equally distributed I think we also [6:53] have a chance to make sure that technology itself  um is also working for more people and and in [7:00] some ways I'm a little bit of an optimist here. I  think if it just amplifies and, you know, augments [7:05] people who already have the education and skills,  then it's going to, you know, that disadvantage or [7:10] that advantage is going to just continue to grow.  But I do think there's choices that we can make [7:15] that will, you know, help expand opportunity,  not just um productivity. If you were to give [7:23] the business leaders in this room some advice  about about how to avoid that, what would it be? [7:32] That's a great question. I think there's a  couple places. I think one is the use cases [7:37] that are going to benefit more people may not  be the use cases that are going to, you know, [7:41] generate the most profit certainly in the  short term. And so I think it's, you know, [7:45] recognizing we do need to generate profit. Let's  focus there. But it's also really looking out at [7:51] much broader voices in society about what are  the problems that they need solving and and in [7:58] what ways. I I think you know we we've talked to  a lot of people in our work and some people are [8:04] like oh yeah this is great some people try it and  it's like you know are turned off by it how do we [8:09] make it um more it's not just access I think it's  also making the models themselves uh work better [8:14] for more more people and Bob you your company  works with clients such as J&J and Dominoes to [8:23] help them roll out AI within their businesses can  you talk a little bit about how those companies [8:31] have kind of transformed their operations as a  result of AI. What's been beneficial but what [8:37] have been the challenges? All right, good.  Um, I think Anton said that uh it's useful, [8:47] right? So, I would say it's a little bit more than  useful. [laughter] Uh, it's tremendously useful, [8:54] right? So, in what way? And I'll I'll explain you.  Um I I think Canti does you know development for [9:02] for plants like like the ones you mentioned  and develop development software development [9:09] has been evolving obviously uh they used to  do like the waterfall ways you know all very [9:15] slow then the lean digital technology principles  came and and increased a lot but now it's it's a [9:25] disruption is not a a a productivity change. We  are you know we we have a phrase in the company [9:35] that's says you know uh AI codes codes we build.  So a coder, someone that does software as we do, [9:47] it's becoming no longer a coder, it's becoming a  builder and with, you know, five times, 10 times, [9:56] 20 times more speed, which means that you can, you  know, if you look at the other way, we're bu we're [10:04] building the same thing in in a small portion of  the time and effort and money we needed to do at [10:12] At the same time, companies are understanding that  they can build things that they never imagined. [10:20] They can open new revenue opportunities. So, it's  we are in the piece of software development. But [10:28] businesses which are more and more and more  software companies, every every company is [10:33] a software company. Even like Domino's as you  mentioned, you know, sells pizza. you know the [10:39] whole uh boost of the business is is technology  right through delivery and it's AI being deployed [10:46] within Dominoes. Give us an example. We just  launched uh recently a couple months ago [10:52] uh their their new uh delivery system and AI is is  not we we didn't fall the trick of like oh let's [11:01] do fancy little bot style that no one uses or make  jokes about it because it's not that efficient. [11:08] We're using nowadays we're using technology  in in the back off backstage uh back office [11:15] helping you know they understanding demand they're  understanding what you know interesting promotions [11:21] could be held differently and and all life right  so you don't need to you you take out the people [11:28] from planning and structuring thinking about  things that will happen next month and everything [11:34] becomes immediately right so I think it's just To  summarize, it's it's a it's a it's a disruption [11:41] moment. We are reinventing our company in the  way we deliver what we do deliver. Uh we're you [11:48] know the short term is huge amount of productivity  gains but that's not our game. The way we see it [11:56] is like the increase of enormous possibility  for the for our clients to develop things [12:02] they never were able to. In those two clients as  examples, as AI agents get deployed even further, [12:10] did you notice that some of the tasks that humans  need to do had to go away as a result of AI? Yes, [12:17] a lot. But and that's the interesting but you know  last year I was here at this event and that's you [12:26] know I was on the other year too but last year  we were afraid that that that answer yes that [12:32] I just gave would mean less jobs right what we're  seeing now that that was the biggest AI slop was [12:42] thinking that you know the AI uh catastrophe was  real and it was is not right. What we are seeing [12:50] now and we are a live living proof of that is  that complexity is so huge. complexity is getting [12:58] more more more you know everything is getting more  complex that we need more people more jobs not the [13:07] same jobs that's that's the important thing right  so they're not you know we don't need more coders [13:14] we need AI orchestrators or AI engineers for in  for instance so um AI could kill certain tasks and [13:23] job functions but you think it will create new new  roles. I want to go to Foyad because you have a [13:31] very interesting project um at Cal Matters and the  tasks that humans used to do are now being done by [13:39] AI. So tell us a little bit about that project.  Um well um I would say yes and no and mostly no. [13:47] Um and this might be just unique to what we  do. So our project digital democracy digital [13:53] democracy.org or if you want to take a look at it  um is a um AIdriven tool that provides government [14:00] transparency by literally delivering everything  anyone ever said in the state legislature, [14:07] state senate and state um assembly in California  and um all kinds of interesting connections with [14:14] like funders and uh lobbyists and things like  that. We have a um public version of it which [14:21] is in digital democracy.org And then there is a  reporterssonly version which is uh it's called [14:27] AI tip sheets and it provides tips to reporters  about what to cover. Now the reason I said yes and [14:32] no is because the reason we even entered the space  was because um maybe 20 15 to 20 years ago uh the [14:42] news industry was devastated. And I mean you know  this there was um all kinds of hometown papers and [14:51] and media organizations that used to each have  reporters that they would send to uh state the [14:57] state capital to do reporting. and uh they would  cover you know their own local legislator and all [15:03] the other things but those jobs were already  gone. They were cut. They were first to go and [15:08] as a result these news operations were just down  to skeleton groups very very small operations and [15:14] uh they were not covering the state legislature.  So with this tool they can cover the state [15:20] legislature. They can keep tab on what the what  their own elected representatives are doing. It's [15:26] going to save them enormous research uh time and  effort and frankly there's going to be some things [15:32] in fact we've already proven it. There are some  analysis that could never have been done before [15:37] that is now being done for the first time and  it's being written about. Uh a story that we wrote [15:42] last year called power is not having to say no in  California is about this very specific California [15:50] um process in the legislature where you can you  can be registered as not voting. Sounds kind of [15:57] like abstain but it's not an abstain. It counts  as a no. And uh politicians take this position [16:03] when they don't want to be on the record saying  no but then they could still kill a bill using [16:07] this. and uh we wrote a story about it and it won  an Emmy. It did really well and so so that's the [16:14] that's the kind of thing that I see here. Now I  don't I fully fully admit it may not be the same [16:19] in every industry but where we want to make the  impact which is government transparency through it [16:24] uh empowering the news industry especially  nonprofit news industry. Um this is this is [16:30] how it start. Could you address the the act the  potential for errors with AI or hallucination? I [16:38] mean how accurate is it? Very very good question.  We get this all the time. So um first of all our [16:45] work started before the sort of LLM revolution  in late 2023. So almost everything we have is [16:52] in terms of uh symbolic AI and rule-based AI and  machine learning traditional machine learning. [16:59] And uh there is I would say two big um uh gates  that would prevent from errors being published [17:06] in the in uh news that goes out. The first one  is that everything that we do comes from primary [17:14] sources and those primary sources are cited  for every fact that is presented to the to the [17:20] uh reporter and the reporter can you know if  reporter wants to say like so and so said this [17:25] or voted this way they can just click on something  they go and see the video of them literally saying [17:30] it and so and they started actually they started  um linking to this like in the stories you [17:36] know and that that cut down on the number of  claims that I was miscoded and so on because [17:41] it's literally the video uh on there. Uh and the  second thing is that it is not writing stories, it [17:47] is providing tips and the reporter is ultimately  responsible for pulling the information together, [17:53] adding their own human knowledge and lived  experience and making a story out of it. Can [17:58] I just add uh in in in at CINT what we're seeing  is that AI is replacing when we are machines the [18:05] repetitive work the real human work is not being  replaced and what we are seeing the shift is that [18:11] we have we need less repetitive people doing  repetitive work and more people orchestrating [18:19] coordinating deciding taking the risk right so  the repetitive work is being replaced but again [18:26] that doesn't mean that it's less work. It's  different work. Anton, where do you fall in [18:32] this debate um about whether AI will kind of kill  certain job functions but create new ones? Yeah, [18:41] in in two words, I feel highly uncertain  about it. [laughter] And um I think it it [18:49] is uh absolutely true what you just described that  over the past uh couple of years uh we have seen [18:57] uh AI being by and large a highly complimentary  tool for knowledge workers and I think this is [19:07] also kind of for us example for journalists but  in some ways what you did was also a design choice [19:13] right you intentionally chose supposed to build  this system that provides material for journalists [19:20] and does not write this story by itself, which you  could with today's language model as well. Yeah. [19:28] Um juries out whether or not it would be as high  quality, but you're absolutely right. And in fact, [19:33] the system couldn't have couldn't have been where  it's at without massive input from reporters. And [19:38] those reporters participated because they thought  that this was a fruitful way to go forward, not [19:43] not that they're being replaced. Mhm. Yeah. And I  think this illustrates two things. First of all, [19:50] uh that uh in the past uh the impact was actually  more benign than a lot of people feared. Secondly, [19:59] that we do have agency and choice about it. But  at the same time I also want to emphasize it's [20:06] not a law of nature that AI systems always have to  be complimentary and I think there is a risk that [20:15] going forward at some point the complimentarity  flips and at some point these systems may take [20:23] over more significant parts of jobs functions  and frankly part of that is a societal choice [20:31] and part of it is also we don't know as well how  the technology will evolve going forward. Can we [20:39] talk a little bit about some of the risks that  you're talking about? So if AI becomes even more [20:47] powerful and we've seen with the latest hacking  capabilities at OpenAI and Enthropic that at a [20:55] certain point AI is beyond human control. So what  are some of the what are some of the scenarios [21:02] that you're looking at that could potentially  lead to you know huge job losses in in white [21:10] color work or entry- level jobs. Yeah. So, so the  the control scenarios that's a little bit out of [21:17] my wheelhouse. But on a pure economic front,  uh the more these systems complement workers, [21:24] uh the more it basically increases worker  productivity and supports jobs, supports wages. [21:34] the more we use them to just flat out automate  away jobs, the more it reduces labor demand, [21:41] uh, which leads to declining wages and declining  employment. And, um, I think both are distinct [21:51] possibilities. I completely agree if I don't mind  if you don't mind me timing in a little bit. Um, [21:58] um, I think yes, there is these two possibilities  and unfortunately I don't think we can control [22:02] for them. So there will be ones that are replacing  jobs like I feel like that is inevitable. However, [22:09] you can compare that to technological innovations  that have done this in the past. You know in the [22:14] what in the 40s and 50s all the phone operators  were basically obsoleted. Um in fact the better [22:21] example in the 90s when I entered the workforce  every tech company used to have its own little [22:26] hardware design group and they used to do design  PCB design chip design all inhouse and all of [22:33] that basically in the next you know 5 to 10 years  after that uh was um consolidated into just a few [22:40] manufacturers and nobody needed to know those  design skills inhouse anymore and people just [22:46] it was much more efficient to have it be on and  then sent back to you. So similar similar thing [22:52] idea. So so yeah, one of the interesting things is  um we want technological change and technological [22:59] turnover, right? So if we went back 200 years in  history, it's probably true that the ancestors [23:06] of more than 90% of us were all engaged in  agriculture and probably we don't mind that that's [23:14] no longer the case today. But I think what matters  is kind of the speed of that change. If it is at [23:22] the pace that we have experienced for example for  the past two years then it is no big problem for [23:29] our society to absorb that change. If it were to  rapidly accelerate then I think it could create [23:36] significant problems. Yeah, it's hard to agree  or disagree with that, but I was just looking [23:42] at some data more recently and the two startups uh  uh one that uh processes payments they they could [23:51] you know on the US they could check everyone that  is using AI buying you know from open AI entropic [23:58] etc and the other one uh checking hiring and  what they discovered that uh from 21 to 26 [24:06] the companies that are really using AI intensive  expenses on AI not just the chatbot they're hiring [24:15] 10% more people than all the others right so and  they have more statistics and and in the s first [24:22] 6 months is 7% more and 18 months of intense AI  uses 32% more uh hiring my understanding is that [24:32] that comes from business growth right it's not  the whole sometimes we're looking too much into [24:38] productivity or job cancellations and we're not  looking at the opportunity. This people they're [24:45] not, you know, having more productivity. Why would  they hire more? They are having more productivity [24:51] and growing more and therefore hiring more. So why  is the tech industry laying off so many workers [24:58] if AI is so productive and efficient? Oh, that's  a great question. And it's been responded in the [25:03] Silicon Valley by a bunch of people more famous  than me. But it it has nothing to do with AI. [25:10] It's a good excuse because you're too fat. You  hired so too much during pandemic. Has nothing [25:16] to do with AI or at the most because you want to  invest more AI. But even though it has nothing to [25:23] do you are too fat, you have productivity gains  that you don't need all those people. It's hard [25:29] to say this but it's real right. So it has nothing  to do because the trick is that if that was real [25:37] the no one will be questioning AI productivity and  everyone is still questioning. So those companies [25:45] they are firing and they are using just AI  as the best excuse or they used to because [25:50] now the market is not buying anymore. So the CEO  the CEOs of Amazon, Meta, Oracle, and Block are [25:56] all kind of using AI as an excuse to cut their  people. They are. Okay. Mark Andre, very famous, [26:03] more trustable than me, eventually said that. If  you look on online, you're going to find a podcast [26:09] video. It's funny because he he laughs at the  question and says, "This is the biggest joke that [26:16] these people are doing. It's nothing to do with  AI." Elizabeth, you're from the Bay Area. What [26:21] are your thoughts about just the industry layoffs  and the messaging that CEOs are giving about why [26:26] they're cutting workers? So, I am from I am from  the Bay Area, but the people that I work with are [26:34] are decidedly not u from the Bay Area. There are  uh rural social workers in rural West Virginia. [26:41] They are low-income moms in Texas or Illinois. And  I think often in some of these conversations like [26:50] we're focused a lot on you know or the software  engineers or or the knowledge workers. Um and [26:56] there is there is a whole group of people who  are um you know barely making ends meet. Now [27:03] they might be working at an Amazon warehouse and  you know robots may be coming to take their jobs. [27:08] Um but even still, you know, right today, like  they are not able to um you know, they're they're [27:14] so worried about how they're going to pay next  month's rent that they're not thinking about, [27:18] okay, how is this going to to affect me? And um I  think thinking you know as we think about policies [27:25] to address any sort of job displacement or you  know job changes or how do we prepare people for [27:31] the AI economy I think we really can't forget  that people are starting from very different [27:37] places and that's the result of you know decades  of investment or disinvestment and we have you [27:45] know we have decisions to make like are we going  to make this sort of an individual responsibility [27:50] um you know people bear the burden of figuring  out how to adapt or are we going to um create sort [27:56] of shared investment in in this. It's a really  good point. Yeah. Um I was going to say uh first [28:02] of all 100% agree with this. Underline it. Triple  underline it. Um it's absolutely true that lots of [28:10] um mundane repetitive work can be those are the  ones that can be easily replaced uh with with [28:16] technology and AI. But people have anxiety not  because mundane work is going away. You know, [28:24] no one no one would have a problem if they  just said uh you know work doesn't have [28:28] to be repetitive anymore. The the issue is  what is the social safety net? What is the [28:34] uh what what does it mean economically? And you  know unfortunately the benefit at least at this [28:39] point it looks like the benefits of all those  enormous savings are not being passed down to [28:45] uh society in general. And because of that there's  that anxiety. May I perhaps also add uh some more [28:53] thoughts on that? So I think by and large the  perspective that you presented is probably true [29:01] but I would express it with less certainty. So I  think there has definitely been some of what has [29:07] been called AI washing, right? That layoffs that  were actually for other reasons are disguised as [29:14] layoffs due to AI. But there's also an intense  debate on this in the economic literature and [29:21] kind of the leading paper on this topic is called  canaries in the coal mine aptly uh does uh show [29:30] that especially among entrylevel workers uh  you can see that the job market opportunities [29:38] uh in highly affected uh occupations and that  means for example uh coding think are less than [29:47] they used to be a couple years ago. Now this paper  in itself is being contested but um my best read [29:56] of it is that there is something there although  it's probably not quite as significant as what [30:04] some of the layoff announcements that you referred  to is capturing. Uh yeah I I do agree this study [30:11] also says that you know there were more hiring  hiring 12% more in in junior uh positions right [30:17] but that I I don't disagree with you that this is  short term this is looking now right so I I I'm [30:23] not saying this is the future right it's hard  to to say but I also just want to comment that [30:28] uh the the the disruption is so strong that we  will need to have new arrangements among you know [30:37] uh business owners, governments, regulation.  There's there's no doubt about that, right? [30:44] So I I don't have a a naive optimistic point of  view. I think what I'm what I'm saying is that [30:50] uh there are gains of productivity and the  entrepreneurship will explode. In my point of [30:56] view, jobs will not be cut but will be replaced.  But but for sure the the whole economic needs to [31:04] be rearranged otherwise there was going to be  more people left behind. I I that with that I [31:10] agree right because but but it's it's a bigger  job I don't think it's only businesses startups [31:17] investors you know more players you know the all  all the scientific uh education all those things [31:25] will need to change right so because if you train  someone in a traditional way school university [31:33] etc to get a traditional job that won't exist  you got to you got to refigure the whole system, [31:40] right? So, and it's I think it's it's it's about  social safety in in the short run, but it's also [31:45] about reconfiguring the whole thing, right? So,  who should be responsible for for reconfiguring? [31:54] Is it governments? Is it business leaders? Is it  is it the workers themselves? That's an amazing [32:01] question. Uh I would say I I would say a lot  of different people have responsibility here. [32:08] I think it's a shared responsibility. I think  I do think that there is a lot of uncertainty [32:13] and that individual workers should not bear  that like there's a story in which we just [32:17] make rescaling sort of a story that comforts  us and trans you know a structural transition [32:23] but individual burden. Um, but I also think it's  it's different than like you you were saying it's [32:29] different than past transitions where you know  if you just reskill once or maybe once a decade [32:35] um you know the models are changing constantly  and the skills that need you know that help people [32:42] succeed change and so it's about reinventing it's  about like lifelong learning and building in these [32:47] systems both in within businesses in terms of  training their own workforces in government [32:53] um And so it's about both preparing people and  then it's also about supporting people through the [32:58] transition because again, you know, the single mom  who can't afford to pay rent is not going to be, [33:04] you know, well positioned to, you know, take  part in some of even if even if it's not paid [33:09] for by someone else. It's like, you know, how  are she going to pay rent and childare and [33:12] all those things during it. So I think it is  um, you know, it's it's not just reskilling. [33:20] What should people focus on when they're looking  to reskill for? You talked a little bit about [33:28] kind of creativity is important. Yeah. Yeah.  I um I mean I have to go back to what I said [33:32] the first time. Um critical thinking. Um, I  feel like critical thinking is becoming super [33:38] super important because now you're at a situation  where you have the result of a of a task that has [33:48] been done for you yet you have no idea how good,  how efficient, how productive, uh how will it do [33:55] like 2 years from now, how well-maintained  it will be. And uh these are skills that um [34:01] you don't you know I think it helps at least  in the tech sector if you have that sort of [34:05] uh computer computing knowledge but really in any  sector you just need that inquisitiveness you need [34:11] that you need that um you know um um that thought  process that leads you to say like what could go [34:20] wrong and let me double check this and let me run  a few different models and and work them against [34:25] each other and so on. Um at the same time I think  what's happening in you know in when this sort of [34:31] the the token prediction revolution that we're  in the middle of is that I think at some point [34:37] um companies are going to realize that this is uh  just enormously energy inefficient. Um essentially [34:43] what you're doing is like you know this this  is an analogy I made in my class um that um [34:49] it's like saying that you um want to teach about  theory of gravity by watching like a million hours [34:56] of videos of things falling down right where the  whole thing is basically just one formula if you [35:02] had that to begin with you don't you wouldn't need  to watch the videos and expend all that expend all [35:07] that energy so I do see like uh and we're already  seeing it because all these models are becoming [35:12] more more efficient and they're running on your  laptops now and on phones and everything. So, [35:16] it's already, you know, going in that direction.  It's going to be more efficient. At the same time, [35:22] um that critical thinking is super necessary  just so when when it meets reality and it has [35:28] to meet business decisions, uh someone is sort  of there at the helm making the making those [35:34] um tough decisions. Bob, how will managerial jobs  change? Um, I've been reading a report about how [35:41] Meta is kind of changing its rankings of their  managers and middle manager jobs are disappearing [35:48] with agents starting to do some of the repetitive  work that humans are doing. What does that mean [35:54] for the hierarchy of an organization? It it means  it's going to get much flatter. Uh they are really [36:01] disappearing. um they they were mostly very  repetitive, you know, daily, monthly, creating [36:11] reports, doing always the same thing. That's why  people are so freaking out because they're afraid [36:16] of their own machines, not afraid of their own  humanity. uh at the same time uh we started a [36:24] we call aic stillcantic software development life  cycle inside our company to our clients and and [36:31] so it's been a couple months that we are you know  disrupting every little client every little or big [36:38] client all clients at CNT and what we learned then  we do that for real first effect is we flatten we [36:47] need less people and the middle is suffering  more the there was a fallacy that we would [36:54] never be able to invite junior people to start  that's not happening they are coming we hired [37:01] uh this year 400 uh junior trainees and they're  the best and one thing that you know about your [37:09] questions your your answer about like what we  need to learn and I think we need to unlearn [37:14] and those junior peoples they they have no  learning so they came fresh. It was amazing. [37:21] They were starting to develop code using AI as it  is supposed to do, not as the way they used to do [37:29] 10 years ago. Right? So, but yes, and and then  what we're seeing is that and and there's a lot [37:36] of friction. This is not pure happiness at all.  But we are reshaping you know a lot of the the [37:43] people and some will reskill as we discussed some  will not right so but obviously this is a smaller [37:52] uh perspective is just a company a business uh I  completely understand there's a bigger perspective [37:58] that we need to consider as civilization but in  the business we are already seeing people that are [38:05] willing to unlearn and relearn and they're will  be opportunities. There will be a lot of friction [38:12] about you know titles and promotions and salaries.  There will because the leather was very defined. [38:19] If you flatten what happens a lot of people  here and a lot of people a lot of competition [38:25] yes for sure and and I think what we don't see and  we're we are experimenting this you know what we [38:32] don't see is is not easy to see I would say is the  opportunities that those things bring for both the [38:41] company right so more revenue more business and  for the people right so it's interesting to see [38:47] that very experienced people coders, long time,  you know, top quality. They were suffering. And [38:55] then they were seeing the junior trainee just just  exploding in happiness and productivity. And by [39:03] having those people work together, a lot of senior  people said, "Ooh, I got to change. I'm going to [39:10] move. I I'm happy to do it all differently than  I used to do." That's the unlearning process. [39:17] But there's going to be blood. So we are talking  about AI not being quite so good at critical [39:24] thinking yet and doing lower level work and so  on. But I do want to emphasize that these systems [39:31] are getting better and are likely to continue  to get better. And we can already see glimmers [39:36] of critical thinking. We can see AI proving  mathemat mathematical theorems that the world's [39:44] top mathematicians haven't been able to prove.  uh to do cyber security tasks uh that the leading [39:52] minds in their fields have not uh been able to  find uh in the past decades and I think honestly [40:01] we need to be humble about where this is going and  I can't predict with any certainty that let's say [40:10] in five years from now the AI will still be doing  only the lower level tasks uh because they are [40:16] getting smarter and smarter. I was going to say  I can I can confirm the first part of what you [40:21] said because I've had student interns reporting  that you know when they went through layoffs [40:26] it wasn't them and their peers but it was their  managers they got laid off. Yeah, that's scary. [40:33] Yeah. Okay. I want to give our audience a chance  to ask their questions. So if you have a question, [40:39] please raise your hand. And we've got mics coming  around. Please identify yourself. Sure. My name [40:44] is Carrie McCoy and I'm from California and I'm  an educator and I was just wondering for Bob and [40:48] you're saying that you're not concerned about  um job displacement. What do you think about [40:53] the data and the reporting that there is about  less entry levelvel positions? How do you justify [40:58] what's happening with like coding or entry-  level business people? Yeah. Uh last year I I [41:05] I was talking to some friends before this this  panel that last year I was much more concerned [41:12] um and and and I think the what we couldn't see is  that the let's say ignorance of entry level that [41:23] was always something that we suppose that they  will entry and then start learning and being less [41:34] ignorant sort of saying, right? That was the  traditional point of view of an entry level, [41:41] right? You learn with your peers, you get better  and better and you you you grow into the company, [41:48] right? The revolution is so big that we are and  we're seeing that live at CIT that the ignorance [41:57] sort of saying of entry level is what companies  need more And we are seeing IBM announced a lot of [42:06] other big corporations they already seen it that  first you know last year beginning of this year [42:11] they cancelled their junior programs trainee  programs but they've seen how wrong that was [42:20] because those people come fresh and if you're if  you are thinking about the old perspective of a of [42:26] a leather yes that won't happen this is the way it  used to happen. If you look at a new perspective [42:34] that you need fresh completely fresh look and  even the education that those junior people had [42:44] is not very useful. It's it's good that they are  junior because if you're experienced if you're a [42:51] middle level or a high senior person, you believe  you know and you resist to change and you're not [42:59] productive. you're always questioning you fight  for your own automation and those people don't [43:06] do that. So a year ago I was here in this stage I  was saying something different but data is showing [43:14] that I was wrong and many people were wrong. That  that doesn't sell an optimistic view of the world. [43:20] Let's not mix those things. Yeah. I would just  sort of uh offer a somewhat contrasting example. [43:26] um you know we're a nonprofit research  organization and you know we used to have [43:30] a number of RAS and we have found that it's much  you know they didn't have necessarily the judgment [43:40] um so you know they would take maybe a month to  produce an output that then you know we would [43:45] take and and improve on or or give feedback and  AI could produce something somewhat similar in [43:51] maybe five minutes or you know a couple of hours  and you know the more junior RAS didn't have sort [43:59] of the the intuition to know like exactly how  to prompt the AI and so you know and as being [44:06] a very resource constrained nonprofit you know  it's hard you know it's hard to justify okay do [44:13] we keep these RAS or do we go here and so I I do  worry that you know also we all started as RAS and [44:19] that's how we developed the intuition so I I do  worry that in some environments that it might not [44:24] be uh quite as optimistic. We had a question. Yes,  we have a question. We can make it quick. Here [44:32] we go. First of all, thank you. This was really  insightful. Thank you so much. So, I'm going to [44:39] try to keep it very small. Um, so new technologies  lead to new prospects, new job descriptions. We [44:45] talked about the industrial revolution. We talked  about the newspapers after the internet boom. [44:49] My question is, but this seems to be much more  profound. AI is much more profound. For example, [44:56] the model hacking example on open AI hacking  tricking uh the singularity in terms of beating [45:02] uh human intelligence. There is also a mad race  in terms of uh uh you know um really going fast, [45:11] right? My question is do we need to slow down to  speed up? number one and second is should we start [45:18] looking at this especially from the singularity  perspective as something as profound as nuclear [45:24] uh thing and start looking at world  level uh consortium to basically have [45:30] alignment across all the countries to  move forward what's your thought on that [45:38] okay um I can take it thank you um I don't know  if I can u it was a it was a very very ambitious [45:45] question involving you know the international  community and so on. Uh but in general I agree [45:50] with uh with more oversight and and more sort  of control. Uh I think we already made a mistake [45:57] like maybe 20 years ago or 25 years ago by not  having oversight over things like social media and [46:06] um you know um harvesting of information and  reselling of information and personal information [46:11] and so on. And we're, you know, we kind of like  entered an era where uh lots of terrible things [46:17] happened because of that because there was  not enough, you know, oversight. Now I don't, [46:22] you know, political climate being the  way it is, I don't know if government [46:27] uh are the best uh way to do that. I think some  government oversight is definitely necessary. [46:32] But if and you know this might be a pie in the  sky, but if some independent truly independent [46:37] consortium could lay down basic rules and basic  uh obligations uh for uh the major players that [46:45] would be great. Well, I'm afraid we're out of  time. Thank you so much for joining us. Um Fyad, [46:52] Anton, Elizabeth, and Bob. Thank you.  Thank you you all. Thank you so much.