OpenAI Targets Automated Researchers by 2028 as Capabilities Shift Safety Strategy

OpenAI拟于2028年前实现自动化研究员，技术跃升驱动安全策略调整 OpenAI预计将在2028年3月前构建出自动化AI研究员，推动加速反馈回路，这将需要去中心化的安全防护以及普及性基础算力等结构性经济调整。 在本次小组讨论中，OpenAI领导层与研究人员探讨了公司最新发布的政策蓝图以及迈向超级智能的迅猛进程。他们深入探讨了加速科学发现、廉价能源和普及软件开发等主要益处，同时强调了充沛算力基础设施的关键必要性。为管控风险，他们提出了防御性AI部署、航空业级别的安全报告机制，以及便携式福利和普及性基础算力等新型经济安全网。最后，小组反思了人际联结的独特价值，并将实现自动化AI研究员的内部目标设定在2028年3月。 Sam Altman将当前AI的指数级演进比作2020年初COVID-19爆发前的隐秘阶段，指出技术层面的剧变已然发生，亟需社会各界尽早启动公共治理辩论。 (0:43) AI将大幅削减创业壁垒，未来仅需两到三名成员配合充足的GPU算力即可打造出功能完备的大型企业，同时也将推动“通用基础算力”等新型保障模式的发展。 (18:00) OpenAI强调普及AI的关键在于建设充沛的物理数据中心基础设施，否则稀缺的算力将被高价垄断，导致权力与资源向极少数富裕群体集中。 (20:35) OpenAI设定了在2028年3月实现自主“AI研究员”的内部目标，该里程碑不仅代表高阶认知劳动的突破，还将引发AI自我递归迭代的极速发展。 (32:27)

OpenAI leadership and researchers argue that the rapid acceleration toward highly capable artificial intelligence demands proactive institutional and economic redesign, warning that society must prepare for deep structural shifts before crisis-level decisions are forced upon it. Exponential Trajectories and Automated Research Milestones (0:05) Sam Altman compares the current period of AI development to late January 2020, recalling how OpenAI began tracking COVID-19 data, planning work-from-home contingencies, and coating door handles in copper while facing public ridicule as society remained oblivious to an impending exponential curve. Adrien Ecoffet observes a corresponding internal shift: researchers recently moved from manually writing code to relying predominantly on AI generation, heightening their sense of urgency regarding the arrival of powerful systems. OpenAI has set an explicit internal target date of March 2028 to achieve an 'automated researcher' capable of independently conducting advanced AI research. Ecoffet characterizes this milestone as a disruptive inflection point that simultaneously proves full autonomy in complex cognitive labor and recursively accelerates future AI capabilities. While Altman acknowledges the team could encounter technical barriers, he maintains that the plausible prospect of compressing a decade of scientific progress into a single year requires immediate public and political deliberation. Transitioning from Lab Restrictions to Multi-Layered Defense (9:56) Ecoffet and Altman argue that 'classical AI safety'—which presumed a small number of centralized, tightly restricted models—is insufficient for a future populated by numerous open-source and proprietary models. Because frontier coding capabilities will inevitably allow models to uncover cyber vulnerabilities, Altman asserts that trusted defenders must leverage AI first to patch fragile infrastructure, such as power plant systems that have gone decades without updates. To manage deployment risks, the panelists propose post-deployment safety layers, including an incident-reporting database modeled after commercial aviation protocols to track near-misses. In biosecurity, Josh Achiam notes that model-level restrictions must be matched with external detection networks and medical countermeasures, pointing to global food supply chain biosecurity as an under-attended, vulnerable sector where AI defense can raise the financial and logistical costs for attackers. Compute Abundance and Restructured Startup Economics (20:35) Altman argues that true democratization requires both public input on AI development and universal access to raw compute infrastructure. Rejecting the idea that OpenAI will soon escape its 'compute crunch,' he explains that falling unit costs paired with rising capabilities keep aggregate demand uncapped. Without aggressive data center expansion, wealthy entities could bid up scarce compute and monopolize power; Altman compares abundant compute infrastructure to historical investments that drastically lowered the cost of electricity. Lowered technical friction is already altering business development. Drawing a historical parallel to Amazon Web Services eliminating the need for early founders to manually rack physical servers, Altman anticipates that two- or three-person teams equipped with large GPU allocations will soon build functional enterprise-scale companies, utilizing AI to execute specialized business functions where the founders lack internal expertise. Labor Protections, Taxation Shifts, and Human Agency (24:05) As AI performs a larger share of intellectual tasks, the panel forecasts that traditional tax regimes tethered to human labor income will need to evolve. Altman discusses structural proposals such as citizen equity stakes in AI upside and 'universal basic compute'—giving individuals raw productive capability rather than passive cash stipends—while Ecoffet notes blueprint discussions around a 32-hour workweek and counter-cyclical social safety nets. Immediate worker safeguards highlighted include portable benefits. Altman criticizes employer-tied healthcare in the United States, arguing that individuals should never lose health insurance due to job transitions. Achiam emphasizes that labor unions and workers must be empowered to co-author workplace deployment, mandate AI literacy, and restrict invasive surveillance. Despite widespread automation, the panelists predict that direct human connection will retain unique value, shifting long-term societal emphasis toward care-focused roles such as teaching, elder care, and nursing. OpenAI presents its industrial policy proposals as an opening framework for public debate rather than a definitive blueprint, inviting policy engagement through grant pilots and a planned Washington, D.C. workshop in May 2026. However, whether institutional governance and social safety nets can adapt quickly enough to keep pace with accelerating technological progress remains an unresolved challenge.

OpenAI 决策层与研究人员在本次政策讨论中指出，人工智能技术正以指数级速度向超级智能演进，社会必须在潜在危机爆发前全面启动前瞻性的制度应对。他们主张，唯有通过大规模扩建算力基础设施、构建多层级的主动防御生态，并以便携式福利与新型分配机制重塑社会安全网，才能在技术奇点临近时确保普惠与社会稳定。 指数级跃迁与2028年自动化研究员目标 (0:05) Sam Altman 将当前公众对 AI 发展速度的迟钝感知，比作 2020 年初 OpenAI 内部预判 COVID-19 疫情时的情形：当时团队已开始制定远程办公计划并在门把手上镀铜以防病毒传播，甚至因此被媒体讥讽，而旧金山街头的公众当时仍对即将到来的封锁浑然不觉。Altman 强调，从事指数级技术研发的人往往会提前感知到剧烈拐点；当前底层技术的质变已经发生，但社会系统尚未消化其全部冲击。在研发一线，研究员 Adrien Ecoffet 透露，团队自身的工作模式近几个月发生了根本性转变，研究员已从手工编写大部分代码转向依赖 AI 生成绝大多数代码，这种切身体验赋予了政策筹备极强的紧迫感。 关于技术突破的时间表，Altman 明确透露 OpenAI 内部设定的官方攻坚目标是在 2028 年 3 月前开发出能够独立进行前沿 AI 研发的“自动化研究员”（automated researcher）。Ecoffet 指出，该里程碑具有“双重颠覆性”：它不仅直接证实模型已具备承担高难度自主认知劳动的能力，更意味着由 AI 驱动的 AI 研发将触发前所未有的递归式自我加速。由于这种飞跃并非单点事件而是持续的爬坡过程，OpenAI 认为政治决策机构必须尽早展开公开辩论，以避免在危机临头时仓促做出低质量决策。 从实验室对齐转向多层级主动防御体系 (9:56) Adrien Ecoffet 与 Sam Altman 指出，传统的“古典 AI 安全”思路假设未来仅存在极少数高度受控且完全对齐的集中式模型，但未来的真实世界将充斥着大量不同来源的模型。单纯依靠发布前的红队测试、模型级安全限制或实验室内部约束，已不足以应对开源高能力模型扩散带来的滥用风险。为此，OpenAI 在蓝图中提议借鉴商业航空安全体系，建立跨行业共享的“未遂事故”（near-misses）与安全事件报告数据库，以便所有开发者共同追踪新型隐患并协同调整防御措施。 面对网络安全与生物安全领域的系统性漏洞，讨论强调防御方必须率先部署高级别 AI。随着模型编程与代码推理能力的剧烈提升，攻击者利用自动化技术挖掘软件漏洞的成本大幅降低，因此可信防御者必须使用 AI 抢先对电网、运行数十年未曾维护的老旧关键基础设施进行系统性修补。首席未来学家 Josh Achiam 进一步指出，COVID-19 揭示了全球物资与食品供应链的极度脆弱，AI 的核心价值在于逆转攻防不对称性，通过大幅提高攻击者的经济与操作门槛，为长期缺乏防护的食品供应链建立规模化生物防御盾牌与早期预警网络。 算力丰裕、精益创业与经济分配范式重构 (18:00) Altman 将 AI 带来的创业门槛降低比作当年亚马逊 AWS 云计算的普及，但预测其颠覆力将远超以往。以往依赖庞大跨职能团队的传统创业模式，正被“两至三名人类搭配海量 GPU”的极简精益团队取代，AI 能够无缝填补团队在软件工程或商业职能上的所有专业技能空缺。然而，这种创新活力的普惠完全取决于底层物理算力的供给。Altman 驳斥了算力瓶颈能够自发消失的论调，指出随着单位智能成本下降，全社会的有效算力需求几乎无上限；如果数据中心基础设施建设无法跟上节奏，稀缺的算力将被极少数财富寡头竞价垄断。他主张大规模扩建数据中心本质上是一项平等主义工程，其社会效应堪比历史上压低电力成本对全民生活水平的革命性提升。 在收益分配机制上，Altman 提出不应局限于传统的现金式全民基本收入，而应探索“全民基本算力”（Universal Basic Compute），直接赋予个体算力资本以激活创造力。随着 AI 逐步接管大部分智力劳动，传统的劳动所得税基础将被大幅侵蚀，资本与劳动的历史平衡可能被彻底打破。为此，政策讨论提议对国家税基进行根本性现代化重构，探索让公民直接持有 AI 生产力红利的股权所有制方案，甚至前瞻性地将法定工作时长缩减至每周 32 小时纳入制度考量。 劳工转型保障、人际连接与关怀经济价值 (30:54) 针对普通劳动者面临的技术替代焦虑，讨论强调必须率先建立逆周期的动态制度安全网。OpenAI 蓝图将“便携式福利”（portable benefits）列为核心举措，Altman 明确批评美国将医疗保险与传统雇佣关系强行锁定的现状“极其荒谬”，主张任何人都不应因失去工作而丧失基础医疗保障。Achiam 提出，制度制定者必须保障工会与劳动者深度参与 AI 引入工作场所的规则制定，赋予员工对 AI 职场监控等侵入性行为的防范权，并通过全方位的 AI 素养培训促进平稳过渡；同时应在政府与私企之间建立新型中间机构，对宏观经济结构的剧变展开高精度实时监测。 在技术渗透加速的背景下，讨论特别强调了人际连接与关怀经济的不可替代性。Altman 分享了个人的“机器人咖啡馆”体验：尽管机械臂制作咖啡效率极高，但完全缺失了星巴克咖啡师日常问候与微笑的交互过程令人感到“极度乏味”，证明纯粹的人际情感连接依然具有不可替代的价值。Ecoffet 与 Achiam 一致认为，随着 AI 接管繁重琐碎的行政事务并协助完成复杂的医疗诊断与保险核验，人类劳动将进一步回流至护士、养老照护、基础教育等以同理心、关怀和人际承诺为核心的岗位，实现公共服务的高质量普惠化。 OpenAI 强调本场论坛涉及的政策构想属于开放性对话的起点，而非定型的强制性蓝图。为推动政策界与公众深化协作，OpenAI 宣布设立专项试点资助计划，提供最高 10 万美元的直接研发资金及 100 万美元的 API 算力额度，并计划于 2026 年 5 月在华盛顿特区设立新的实体工坊以召集各界深入协商。但在全球地缘竞争与资本加速布局下，如何确保算力扩张不造成极端能源挤压、各主权国家如何同步推进便携式福利及税基改革，仍是悬而未决的现实挑战。

## Background 本期视频由 OpenAI 官方 YouTube 频道于 2026 年 4 月 7 日发布，时长 46 分 11 秒。视频记录了 OpenAI 论坛的一场深度对谈，首席执行官 Sam Altman 联同首席未来学家 Josh Achiam 与资深研究员 Adrien Ecoffet，围绕人工智能以超乎外界想象的速度演进、前沿技术在科研与创业模式上的赋能潜力，以及全社会制度体系如何应对伴生的结构性挑战展开探讨。 ## TL;DR 对话围绕 OpenAI 当日发布的超级智能行动蓝图展开，透露团队内部已全面转向由 AI 编写代码，并设定了 2028 年 3 月攻克“自动化研究员”的关键技术节点。嘉宾们深入展望了超级智能在一年内推进数十年科研突破、赋能“极简人员加海量算力”微型创业，以及大幅消除医疗官僚壁垒的普惠前景。针对网络与生物安全风险，他们主张告别单体模型的精准对齐幻想，借助充沛的 AI 能力构建航空级事件上报与社会自发防御体系。面对普通劳动者的替代焦虑与劳资失衡，与会者提出建设如电力般充沛的数据中心以避免算力垄断，并探索推行“全民基础算力”（Universal Basic Compute）、向 AI 自动化智力产出征税以及与雇主解绑的便携式福利保障。 ## Atmosphere & style 整场对谈弥漫着一种洞悉技术质变却发现外界浑然不觉的清醒与紧迫感，Sam Altman 用 2020 年疫情前夜独自行走在旧金山街头的冰冷回忆，精准勾勒出技术前沿与社会认知之间的巨大脱节。研究人员与未来学家的对话展现出罕见的谦逊与务实，既有看到孩子用自然语言借助 Codex 当场写出电子游戏时的惊叹，也有面对政策落地与制度赤字时的审慎推敲。在探讨宏大技术蓝图的同时，Altman 坦率讲述自己在无人化机器人咖啡馆中感受到的空虚与索然无味，让整场讨论在冰冷的算法加速之外，沉淀出一份对真实人际联结、照护经济与人类劳动尊严的深切关怀。 ## Bulletin - [00:05](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=5s) OpenAI 发布应对超级智能（Superintelligence）的政策行动蓝图，旨在赶在具备强大能力的前沿模型密集走出实验室之前，提前引发公众与政策层面的充分辩论。 - [32:27](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=1947s) OpenAI 官方设定的全自主“自动化研究员”（Automated Researcher）研发目标时间为 2028 年 3 月，这将推动高阶科研认知工作的实质性自动化并触发递归加速研发循环。 - [28:45](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=1725s) Sam Altman 主张将未来福利政策从传统的发放无条件现金补贴（UBI）转向提供“全民基础算力”（Universal Basic Compute），通过基础设施建设平权释放普通大众的创造力与能动性。 - [44:46](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=2686s) 针对 AI 时代的产业政策研究，OpenAI 设立了专项扶持基金，单个入选项目最高可获 10 万美元现金资助及 100 万美元的 API 算力额度，并计划于 5 月在华盛顿特区设立研讨实体空间。 ## Key takeaways - [02:26](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=146s) OpenAI 内部研发人员工作流已在数月内全面过渡到由 AI 代写大部分代码，技术人员从一线深刻感受到了演进烈度，推动其直接参与政策条款的起草。 - [04:36](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=276s) 萨姆·奥特曼将当前 AI 的临界质变与社会反应，比作 2020 年初新冠疫情暴发前夜少数人洞悉指数曲线却难以让大众相信的历史重现。 - [06:33](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=393s) 超级智能有望大幅摧毁高昂的边际成本壁垒，在医疗、住房和清洁能源等公共基础设施领域带来前所未有的超低成本供给与繁荣。 - [10:23](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=623s) 社会系统级韧性需要超越传统的实验室安全红队测试，借鉴航空工业建立跨行业、统一标准的“事件报告机制”（Incident Reporting）以快速共享险情并加固防御。 - [12:34](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=754s) 在充斥数百万个 AI 的多智能体世界中，单靠个别厂商的模型对齐已无法保障安全，必须通过赋能可信防御者构建涌现式的网络与生物防御盾牌。 - [18:44](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=1124s) 初创企业形态正在经历深层范式转移，从早期的自建机房机架，到云计算时代，演变为未来仅凭两三人团队搭配海量 GPU 即可支撑的全功能商业体。 - [20:17](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=1217s) 全球对算力的需求在现实中没有天花板，若基础设施扩张滞后将导致算力被极富阶层垄断，大规模扩建数据中心在本质上具备极强的平权主义属性。 - [23:45](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=1425s) 随着 AI 替代人类劳动并推动财富流向资本与自动化生产，财政体系需要探索向 AI 创造的智力价值征税，重构现代化税基。 - [25:32](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=1532s) 保障普通劳动者权益需推行让工会参与 AI 部署准则的制定、坚决抵制职场全天候 AI 监控，并为普通员工提供系统的 AI 素养普及。 - [30:39](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=1839s) 为缓解转型期的劳动力冲击，政策蓝图建议设立“便携式福利”（Portable Benefits），将医疗保险等生存底线权益与特定雇主解绑。 - [33:58](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=2038s) 当冷冰冰的认知与操作技能被 AI 大规模承接后，人类情感联结的价值反而大幅跃升，护理、育儿与教学等“照护经济”（Care Economy）将被重新重估。 - [37:12](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=2232s) AI 能够穿透错综复杂的医疗官僚报销与转诊迷宫，不仅大幅削减医护繁重的文书行政负担，还赋予患者随时解读指标的自主权。 - [39:45](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=2385s) 当前社会存在严重的“能力悬置”与公众认知滞后，许多用户对带有思考链路的推理模型一无所知，老一辈程序员也常因习惯困难而陷入“苦工盲区”。 - [42:23](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=2543s) 顶尖 AI 正在演进为终极“思维伙伴”，未来将通过系统化检视个人历史邮件、备忘录与灵感碎片，直接辅助创作者筛选并落地项目构想。 ## Detailed walkthrough ### [00:05](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=5s) OpenAI 论坛开场与超级智能蓝图的发布背景 OpenAI 论坛主持人克里斯·尼科尔森（Chris Nicholson）宣布活动开始，并介绍论坛的核心主旨：深入探讨人工智能在真实世界中的落地应用、从中汲取的教训，以及公众如何参与塑造其发展轨迹。当天的讨论聚焦于技术领域最根本的问题之一——随着 AI 系统能力的飞跃式提升，它将对科学探索、工作模式、社会组织形态以及治理结构带来怎样的深远影响。本次对谈邀请到 OpenAI 联合创始人兼首席执行官萨姆·奥特曼（Sam Altman）、首席未来学家乔希·阿奇亚姆（Josh Achiam）以及资深研究员阿德里安·埃科菲（Adrien Ecoffet）。 话题首先切入 OpenAI 当天清晨刚刚发布的一份重磅蓝图（Blueprint），该文件重点探讨了“超级智能”（Superintelligence）的演进与应对。面对“为何选择在此时发布该蓝图”的提问，萨姆·奥特曼直言，核心原因在于技术的进步速度仍在不断加速，内部团队认为“超级智能已经离我们非常接近了”。这不会是一个孤立发生的单点事件，而是在未来几年内，一系列能力极其强大的前沿模型将陆续走出实验室并重塑现实世界。 萨姆指出，研究人员在梳理这套早期构想时展现了非凡的洞察力，发布这份蓝图的初衷是尽早抛砖引玉、引发公开辩论。他坦言团队并非全知全能，“我们可能会犯错，也可能会撞上一堵未知的技术高墙”，但根据目前所看到的模型技术储备与研发管线（pipeline），能力极其强悍的模型在短期内降临已是高度确定的趋势。回顾历次重大技术转型，如果公众、政策制定者和政治体系能在必须做出关键抉择之前拥有充分的时间来公开探讨各种构想，最终形成良性决策的概率就会大幅提升。 ```mermaid flowchart TD accTitle: 技术演进与公共政策辩论飞轮 accDescr: 展现从实验室内部感知加速到前置推动公共辩论的逻辑闭环 subgraph Lab["前沿技术现状"] A["研发管线快速演进"] --> B["前沿模型能力激增"] B --> C["内部形成紧迫感"] end subgraph Public["社会与制度响应"] D["发布超级智能行动蓝图"] --> E["政治与公众提前充分辩论"] E --> F["在危机前建立良性治理与制度缓冲"] end C --> D ``` --- ### [02:26](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=146s) 研究人员跨界政策制定与“让 AI 编写代码”的紧迫感 主持人谈到，在这份蓝图的成文过程中，技术研发团队与政策团队展开了极为罕见的深度跨界合作。资深研究员阿德里安·埃科菲分享了参与政策文件起草的切身体会。他坦言这是一次令人感到“谦逊”的经历：研究人员往往习惯于漂浮在抽象概念中——比如随口讨论 AI 的经济影响或安全监管政策——但真正伏案下笔，把这些想法凝练成具体、严谨、经得起同行严格质询推敲的政策条款，完全是另一回事。 阿德里安认为，让处于一线的技术人员尽早且深度地参与这类前瞻性政策至关重要。政策制定必须依赖那些每天亲手构建系统、深刻理解安全技术栈（safety stack）运作原理、并且对底层迭代节奏具有第一手感知的研发者。 更为关键的是，这种紧迫感来自技术人员日常工作方式的真实质变。阿德里安透露，在编写蓝图的近几个月里，OpenAI 内部大量研究人员亲身经历了一场戏剧性的转折——从“主要由自己亲手编写代码”全面转向“让 AI 编写绝大部分代码”。这种在生产力一线的切实冲击，让研发团队深刻体会到技术绝非虚妄的未来概念，它的演进烈度与真实性远超外界普通观察者的想象，这也是蓝图必须争分夺秒向公众发布的核心动因。 - **跨界协作范式**：一线模型研究员与政策团队共同撰写行动纲领，将技术实现细节直接映射为制度设计。 - **内部生产力拐点**：研究员在几个月内完成了工作流从手写代码到 AI 代写的跃迁，直观验证了自动化能力的突变。 --- ### [04:36](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=276s) 2020 年初疫情前夜的回忆：指数级变化的非对称感知 面对这种“内部深知剧变将至，外界却浑然不觉”的错位感，萨姆·奥特曼在现场分享了一段私密而生动的故事。他回忆起 2020 年 1 月底至 2 月初新冠疫情刚刚露头的日子。当时 OpenAI 的研究团队由于日常就在与指数增长（exponential）打交道，对数字曲线天然极为敏锐，在外界尚未警觉时，整个团队就已经近乎痴迷地紧盯疫情数据，并笃信全球大停摆必然发生。 OpenAI 团队随即开始筹划居家办公，甚至在办公室门把手上安装了具有抗菌特性的铜制把手。这一超前举动在当时被媒体作为奇闻报道，记者发文公开嘲讽这群“疯狂的 OpenAI 员工”。萨姆回忆，在旧金山正式发布封锁令之前的一个寒冷冬夜，当时住在教会区（Mission District）的他心想自己即将被困在家中数月，于是决定独自在夜色中做最后一次漫长的城市散步。 在数小时的寒夜步行中，萨姆戴着口罩，透过沿街餐厅与酒吧明亮的玻璃窗，看着里面的顾客在毫无防备地贴面交谈、推杯换盏。整条街上只有另一个路人戴着口罩，两人擦肩而过时彼此默默点头致意；其余的生活场景看上去完全风平浪静。萨姆感慨，自己这一生中从未像现在这样，如此强烈地重温那个寒夜的冰冷触感： > “当下的感觉和当年完全一样：剧烈的变革已经发生，模型已经达到了某种临界水平，但整个社会尚未消化这一事实。我们觉得自己看得很清楚，正在拼尽全力告诉世界它即将到来，而这种信息极难传递。这感觉就像是回到了疫情初期的那个夜晚，独自穿行在街头。” --- ### [06:33](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=393s) 科学突破与物质丰裕：超级智能带来的潜在前景 当主持人将话题从疫情初期的沉重转回到 AI 蕴含的巨大红利时，萨姆强调，尽管蓝图花了大量篇幅探讨如何防范风险，但背后的正向潜力同样不可估量。在科学探索领域，最直接的前景是“在一年内完成相当于过去数十年的科学突破”。这涵盖了治愈大量疑难杂症、推动真正个性化的精准医疗，以及发现全新材料以制造廉价、清洁且安全的能源。 在个体创造力方面，超级智能将彻底颠覆门槛。未来任何怀揣创业点子的人，都可以依靠 AI 完成全套软件的落地开发与业务执行；人们甚至可以指示系统生成专属的、符合个人口味的最佳视频游戏。萨姆透露，自己在通读完研究人员撰写的蓝图后，第一反应是既震撼又感到有些不可思议：“我们居然已经能够极其严肃、可信地探讨这些以往只存在于科幻中的事物了。”更重要的是，如果系统真能达到预期的高度，AI 本身也将成为人类化解次生风险的最强工具。 随后，首席未来学家乔希·阿奇亚姆从社会基础设施（social infrastructure）的角度拓展了这一愿景。人类社会长久以来一直渴望为每个人提供基本的食物、居所、电力和医疗保障，让每个人都能从生存焦虑中解脱，专注于生命中更本质的事物。但历史上社会一次次遭遇的阻碍往往是“成本过高、无力负担”。而超级智能有望从根本上摧毁这种由高昂边际成本构筑的壁垒，以人类历史上前所未有的超低成本供给优质资源与公共服务。乔希强调，正因如此，作为这一变革的核心守护者，技术构建者背负着不可推卸的时代责任。 ### [09:26](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=566s) 广泛繁荣的责任与系统性冲击的防范 Sam Altman 强调，确保 AI 带来的红利被全人类共享是一项核心责任。这项使命不仅关乎技术的突破，更关乎能否建立起切实保护工薪阶层、中产阶层以及低收入国家群体的制度与政策体系，避免让 AI 仅仅沦为极少数富裕阶层的特权工具。在为这种普惠潜力感到振奋的同时，Sam 也明确指出潜在的负面风险极具毁灭性，甚至可能引发波及全球的系统性冲击，这要求所有人必须保持清醒、诚实面对并提前做好准备。 主持人由此引出“系统韧性”（Resilience）的话题。韧性绝非仅仅体现在模型发布前的实验室阶段，更在于模型进入真实社会后，整个人类社会如何构建多层次的防御与响应体系。这一问题被抛给了现场的几位对谈者，探讨他们各自对 AI 时代“韧性”的深度思考。 ### [10:23](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=623s) 从经典安全演进到分层防御与航空级事件报告 Adrien Ecoffet 对传统的安全防护与社会级韧性做出了清晰的区隔。传统的做法主要集中在内部安全评估（Safety Evals）、红队对抗测试（Red Teaming）以及部署对齐缓解措施，这些做法至关重要且必须持续加大投入。然而，仅靠这些远远不够——社会必须做好准备，应对那些完全不做安全测试的恶意或疏忽行动者所造成的冲击，以及即使经过严密测试依然可能发生的系统意外或险情（Near Misses）。 ```mermaid flowchart TD accTitle: 航空启发的事件报告机制 accDescr: 展示从AI事故或险情上报到全行业数据库并协同修复的韧性防御闭环 subgraph S1["前线监测与响应"] A["AI 安全事故 / 险情 (Near Miss)"] --> B["标准化事件上报机制"] end subgraph S2["全行业共享与防御"] B --> C["公共事件数据库 (借鉴航空业经验)"] C --> D["各机构同步风险动态"] D --> E["全社会共同部署针对性缓解措施"] end ``` 为了应对潜在的溢出风险，Adrien 重点提到了借鉴航空工业经验的“事件报告机制”（Incident Reporting）。在民航体系中，任何细微的安全险情或异常都会被汇总至统一数据库，让所有航空公司共同吸取教训并快速部署防御措施；AI 领域同样需要在全社会层面推行此类机制。此外，随着大模型在代码生成能力上的飞跃，网络安全战攻防格局正被剧烈重构，社会韧性建设必须确保将 AI 工具转化为巩固软件安全、全面修补系统漏洞的防御性武器。 - **多层安全架构**：传统安全评测与红队测试是不可或缺的基底，但必须向社会系统级防护扩展。 - **民航模式启示**：建立轻微险情即时上报的集中数据库，帮助全行业及政府机构提前获悉新兴威胁。 - **代码双刃剑**：强大的代码生成能力直接等同于潜在的网络攻击能力，必须主动使用 AI 进行防御加固。 ### [12:34](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=754s) 告别单体幻想：多智能体世界中的涌现式安全与防御优先 Sam Altman 深入剖析了“经典 AI 安全”理念的局限性。早期研究普遍设想未来世界上只会存在极少数几个尖端超强模型，只要确保这几个 AI 不做坏事、完成精准对齐，世界就能高枕无忧。但在 Sam 看来，现实图景虽然更加复杂，却也更加稳定：未来世界上将充斥着数以百万计的各式各样的 AI，单凭一两家前沿厂商“承诺管好自家模型”已不可能保障绝对安全，唯有依靠全社会自发形成的“涌现式响应”（Emergent Response）与多层防护盾牌。 在网络安全领域，AI 挖掘系统漏洞的能力将呈现指数级爆发，人类将惊愕地发现现有软件设施远比想象中更加脆弱。面对开源模型和代码能力的迅速普及，封锁技术不仅不切实际，更无法阻挡漏洞挖掘技术的扩散。真正的解法在于建立防御者倾斜的机制（如 OpenAI 面向可信防御者的专享扶持计划），赋能那些保卫核心基础设施（如二十年未打补丁的发电厂工业软件）的安全团队。而在生物安全领域，Sam 同样断言，仅仅限制自身模型无法合成病原体是不够的，未来终将有人利用其他未受控的模型去研发病毒；唯一的安全解法是社会必须借助前沿 AI 建立起由检测雷达、快速响应治疗和生物防护网构成的坚固护盾。前沿对齐测试为社会争取到了关键的“时间窗口”，但最终救赎在于社会自发涌现的综合防御体系。 ### [15:59](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=959s) 历史经验的映射：新制度构建与食品供应链的生物韧性 Josh Achiam 拓展了讨论的历史厚度，指出历史上每一次重大的技术范式转移，都会催生出全新的制度体系与民主治理机制。AI 所放大的诸多脆弱性——如民主选举遭受的恶意操纵、以及新冠疫情（COVID-19）期间暴露无遗的全球供应链脆弱性——本质上是长期存在的内生缺陷，AI 的到来只是极大加速了采取行动的紧迫性。 Josh 感到最振奋的是，社会可以利用 AI 构建全新的国家能力（State Capacity）与治理机构。AI 让人类历史上第一次有能力对深层系统漏洞展开系统化扫描、大规模自动化修复，从而在网络与生物领域全面拉高攻击者的经济成本，使袭击变得“无利可图”。特别是在生物安全维度，Josh 极力呼吁关注常被大众严重低估的“食品供应链生物风险”（Food Supply Chain Bio Risks）。一旦农业或牲畜供应链遭到病原体精准打击，现代文明的基础将面临严峻考验；而利用 AI 构建低成本、可规模化落地的农业防疫与监测韧性，正是当务之急。 ### [18:00](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=1080s) 个体能动性的解放与组织形态的根本重构 谈到技术冲击下的个体转型与劳动力经济价值的重组，Josh 认为 AI 经济将带来普通人“能动性”（Agency）的空前爆发。过去的创业与价值创造受限于高度专业的门槛与庞大的人力分工，而未来个体只需凭借一个创意，就能调动一整支专属的 AI 团队，全面包揽自己毫不擅长的商业运作、财务、法律以及技术实现等全链条流程。这一转变不仅极大降低了全新商业体的启动成本，更在根本上重构了经济生态的组织与分配逻辑。 ### [18:44](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=1124s) 极简团队与算力迸发：AI 时代的初创企业重构 谈及初创企业的发展格局，Sam Altman 结合自身创立公司以及长期掌舵 Y Combinator 的深厚经验指出，每一次创业门槛与摩擦成本的大幅下降，都会催生行业的巨变。他回忆起自己早年创业时经历的技术浪潮——当时正值亚马逊 AWS 等云计算基础设施兴起，创业团队再也不必在壁橱里手动装配和维护实体服务器机架（"manage racks in the closet"），只需几个人就能迅速构建起全球级产品。 如今由 AI 引领的变革规模远超当年的云转型。Altman 表示自己对探索这一新空间抱有极大执念，他敏锐地察觉到某种更深层的范式转移：未来的初创企业可能只需两到三个人、甚至单枪匹马，搭配海量 GPU 算力，就能在极短时间内完整建立起一家公司。这种极低摩擦的构建方式将从根本上重塑创业生态，让任何有创意的人都能平权参与商业创新。 - **创业摩擦递减的历史演进**：从自建机房机架，到 AWS 云计算普及，再到如今由大模型与 GPU 支撑的微型高能团队。 - **初创企业形态重塑**：从劳动密集型团队转型为“极简人员 + 庞大算力”的高效协同模式。 ```mermaid flowchart LR accTitle: 创业门槛演进与 AI 范式转移 accDescr: 展示从早期自建机房到云计算再到 AI 算力驱动下初创团队形态的变迁 subgraph Phase1["传统时代"] A["实体机架与机房维护"] --> B["高资本与人力门槛"] end subgraph Phase2["云计算时代"] C["AWS 等公有云普及"] --> D["小团队免除硬件运维"] end subgraph Phase3["AI 时代"] E["海量 GPU + 先进模型"] --> F["两三人支撑全功能初创公司"] end Phase1 --> Phase2 --> Phase3 ``` --- ### [20:17](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=1217s) 民主化的双重维度与算力短缺的“无底洞” 在讨论如何让更多人获得 AI 赋能时，Altman 明确区分了“AI 民主化”的两个核心层面：一是使用权的普及（Share the Access），确保每个人都能获得充足的 AI 工具来改善生活和开创事业；二是话语权的普及（Voice in where it goes），即让公众能够参与决定技术的发展走向。OpenAI 之所以公开发布政策蓝图并尽早将产品推向市场，正是为了让全社会提前感知技术并展开公开辩论，随后将公众的反馈重新吸纳进系统的迭代中。更重要的是，普及绝不能仅停留在 ChatGPT 这类前端交互产品，而必须将真正高价值、高算力消耗的底层能力开放给大众，用于科研突破或商业创业。 针对外界关于 OpenAI 何时能彻底走出算力短缺（Compute Crunch）的疑问，Altman 给出了一个出人意料的清醒判断：只要团队做好了本职工作——持续降低智能成本并提高智能上限，全球对算力的需求在现实中就是没有天花板的，算力短缺可能永远无法终结。如果 AI 基础设施建设滞后，算力就会沦为极度稀缺的资源，全球最富有的阶层和跨国巨头将通过无休止的竞价垄断使用权，导致权力与资本走向极度集中。因此，真正的民主化路径只有一条：竭尽全力建设铺天盖地的算力基础设施，让算力变得如空气般充沛，甚至充沛到人们需要绞尽脑汁去构思“如何把这些算力用完”。 - **民主化的两大支柱**：平等的技术使用权（包括核心高算力服务）与双向反馈的社会决策话语权。 - **需求无上限假说**：智能成本每下降一个数量级、能力每提升一档，衍生出的新应用需求便会呈指数级爆发。 - **算力匮乏的分配风险**：当供给严重受限时，市场机制会导致价格被高资产竞购者单向推高，加剧阶层壁垒。 --- ### [22:50](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=1370s) 能源历史映射 AI：数据中心建设的平权属性 为了更直观地阐释基础设施与社会公平的关系，Altman 回溯了人类现代化的能源演进史。过去 50 年乃至 200 年间，人类生活质量的质的飞跃，最根本的动力之一就在于将电力和能源的成本打到了极低水平。长久以来，人均能源消耗与人均生活质量高度正相关；而未来，AI 智能很可能会接过能源的接力棒，成为决定人类生活水准的核心驱动力。 对话指出，大规模扩建数据中心表面上看是重资产工业扩张，但其内核实际上是一项极具平等主义色彩的举措。正如低廉电价消除了能源特权一样，只有在技术层面不断创新以降低推理成本，并在物理层面大规模扩张数据中心网络，才能使高阶 AI 摆脱奢侈品属性，像电力一样普惠到全球每一个普通人的日常生活中。 - **历史类比**：廉价且充沛的电力极大提升了全人类的生活水准；AI 将成为新时代的通用动力。 - **基建平权逻辑**：通过规模化建设数据中心扩大供给，抑制特权阶层的价格垄断，确保普惠获取。 --- ### [23:45](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=1425s) 应对资本与劳动格局变化：重构税基与普惠红利 Adrien Ecoffet 将话题引入更广泛的普通劳动者群体：那些无意创立公司的大众如何不被浪潮抛下？AI 的爆发正在深刻改变宏观经济结构，推动经济分配的天平向资本或“由 AI 执行的劳动”加速倾斜。为了避免繁荣成果仅仅演变为少数技术巨头的极端财富，OpenAI 在政策蓝图中提出了现代化的税基改革设想，探讨如何在 AI 深度重构生产力的时代实现全民普惠分配。 主持人随即抛出探讨重点：如何让全社会捕获 AI 带来的红利？劳动者如何与技术协同？Josh Achiam 首先补充了算力分配的紧迫性——在算力完全充沛之前的这几年过渡期里，如何公平分配稀缺算力将是社会治理面临的最艰巨课题。如果算力过于昂贵，当面对“是否应该向全民提供某种前所未有的公共利益服务”时，整个社会就会陷入痛苦的财政权衡与成本拷问中。因此，迅速扩大全球总算力储备依然是化解分配矛盾的前提。 - **宏观经济重构**：从传统的人力劳动创造价值，逐渐过渡到资本与 AI 自动劳动主导。 - **财税体系现代化**：探索适应 AI 自动化生产的新型税基机制，防范财富单向聚集，建立普遍的繁荣共享体系。 - **过渡期权衡风险**：算力供给不足时，极具社会效益的公益性 AI 项目往往会因高昂的推理成本而受阻。 --- ### [25:32](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=1532s) 直面劳动者焦虑：工会赋权、抵制监控与国家治理能力 Josh Achiam 直言不讳地指出了对话中不可忽视的“房间里的大象”（"the elephant in the room"）：大量普通劳动者正在对 AI 产生强烈的焦虑与抵触情绪。工人们的普遍直觉并非兴奋地拥抱新技术，而是恐惧自身岗位被系统直接替代。他认为，作为 AI 技术的缔造者和守护者，OpenAI 必须率先通过蓝图等纲领性文件向社会作出承诺，积极倡导建立健全的托底安全网，确保无论经济结构如何剧变，劳动者都能获得充分的生计保障。唯有在确立了这种信任与安全感的前提下，劳资双方才能展开坦诚对话。 在具体落地方向与新型机构形态上，Achiam 强调了以下几个关键维度： - **工会与劳动者赋权**：推动工会深入参与企业引入 AI 的决策过程，让一线员工“共同制定（co-author）”AI 在工作场所的具体部署准则与可接受范围。 - **坚决抵制技术滥用**：明确回应员工对“AI 职场全天候监控（AI surveillance）”的担忧，确保劳动者对隐私和尊严拥有实质防卫权。 - **全民 AI 素养普及**：提供系统化的工具培训，帮助普通人掌握运用 AI 改善生活、经营微型业务或释放潜能的技能。 - **国家治理感知力升级**：在制度创新层面，借助高性价比且具备规模化扩展能力的 AI 工具，强化国家在微观与中观经济层面的高精度度量能力（granular depth），从而在特定行业或区域发生剧烈经济阵痛时迅速做出精准的政策响应。 ### [27:52](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=1672s) 构建介于企业与政府之间的新型治理机构 随着技术演进速度超越既有制度的响应能力，现场讨论聚焦于制度层面的深层空白。传统格局中，私人企业治理维度偏向股东利益且问责机制有限，而政府行政体系虽然具备合法性，反应速度却相对迟缓。针对这种两极分化，与会者提出需要探索一种介于企业与监管部门之间的新型组织形态——一种“介于公司董事会与监管机构之间”的缓冲载体，承担起快速原型测试、政策沙盒试验以及提供社会兜底安全网的职能。 这种设想虽然带有头脑风暴与即兴探讨的色彩，但直指核心困境：应对指数级跃迁的技术挑战，既不能全盘寄托于缺乏外部问责机制的商业公司，也无法指望按部就班的立法官僚机构能够神速应对，催生出敏捷且具备公共利益属性的“中间态机构”已成为紧迫的制度诉求。 ### [28:45](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=1725s) 全民基础算力与重构资本主义税收体系 面对未来社会如何广泛分享 AI 红利的问题，山姆·奥特曼（Sam Altman）梳理了从传统福利思维向技术赋能转型的逻辑演变。他指出，与其沿袭传统无条件基本收入（UBI）那种在 AI 取代人类工作后按月发放现金补贴的思维，不如推行“全民基础算力”（Universal Basic Compute）。奥特曼强调，普通人极具自主性与创造力，非常清楚自己需要什么，只要不被剥夺获取核心算力资源的权限，就能催生出意想不到的生产力与价值。 在宏观经济制度上，奥特曼指出传统资本主义根植于劳动力与资本的平衡机制，一旦 AI 接管了绝大多数当代人类的智力劳动，原有的天平将被彻底打破。为此，分配与财政体系必须经历结构性变革： - **税制重构**：从传统向人类劳动所得征税，转向对 AI 所创造的智力劳动价值与自动化产出征税。 - **动态兜底**：探索新型过渡期扶助计划与适应 AI 冲击的失业保险体系。 - **所有权让渡**：打破少数资本巨头垄断，设计能让公众广泛享有技术增长红利的新型所有权架构。 奥特曼坦言自己始终保持开放态度，“也许我们预判有误，也许未来根本不需要做出任何改变”，但提前在窗口期展开探讨与准备至关重要。 ```mermaid flowchart TD accTitle: 应对 AI 冲击的宏观经济与政策防御框架 accDescr: 描绘从 AI 智力替代到劳资失衡，进而触发税制重构、基础算力分发及反周期防御机制的流程图。 subgraph Disruption["AI 带来的结构性冲击"] A["AI 接管核心智力劳动"] --> B["传统劳资平衡被打破"] B --> C["传统所得税税基萎缩"] end subgraph InstitutionalResponse["制度层面的重塑举措"] D["全民基础算力 (UBC)"] E["向自动化智力产出征税"] F["反周期失业保障机制"] G["32 小时工作制探索"] end subgraph SafetyNet["社会稳定与共享"] H["便携式福利 (福利与雇主解绑)"] I["医疗与基础生计绝对兜底"] end C --> E B --> D B --> F F --> G D & E & G --> H H --> I ``` ### [30:39](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=1839s) 反周期防御机制与便携式福利制度 针对转型阵痛，团队探讨了政策蓝图中的预警与反周期调节工具。由于激进政策在常规时期往往伴随巨大的经济扭曲，蓝图主张将保障措施设计为“条件触发式”的反周期机制——仅当 AI 对特定行业就业产生实质性破坏时，才激活更高水平的失业保障金及缩短法定工时（例如推行 32 小时工作制）。 政策讨论中获得最高共识的提议是“便携式福利”（Portable Benefits）。在美国现有体制下，医疗保险等关键福利被深度绑定在雇主身上，一旦离职便面临保障断崖。奥特曼对现行体系的不合理表达了强烈态度，直言现有模式早已畸形： > “没有任何人应该因为失去工作而失去医疗保险，这种事情根本就不应该发生。” - **便携式福利**：将医疗与核心福利与特定雇主解耦，转变为跟随个人身份的跨行业保障账户。 - **动态缓冲阀**：将缩短工时与提高补助作为应对结构性失业的自动化经济杠杆。 ### [32:27](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=1947s) 2028年3月里程碑：自动化研究员与双重技术冲击 当主持人询问社会与制度跟上技术迭代的适应窗口期还剩多久时，研究员阿德里安·埃科菲（Adrien Ecoffet）坦承时间极其紧张。他提到开发全自主“自动化研究员”（Automated Researcher）的规划时，乔什·阿奇亚姆（Josh Achiam）当场补充纠正了明确的路线图时间节点：“官方设定的目标是 2028 年 3 月”。 埃科菲深入剖析了自动化研究员问世所带来的“双重冲击”（Double Whammy）： 1. **高阶认知替代**：模型具备自主开展前沿 AI 算法与科研工作的能力，标志着高复杂度脑力工作被实质性自动化。 2. **递归加速循环**：当能够自主从事 AI 研发的智能系统加入研发循环后，整个技术的自我演进速度将发生质的跃升，迭代周期从那时起将远远快于以往任何阶段。 这种级联加速效应意味着留给社会制度适配、法规出台及就业转型的窗口期正在迅速收紧。 ### [33:58](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=2038s) 观众提问：未来最关键的人性特质与“机器人咖啡馆”的顿悟 在观众互动环节，观众斯维特兰娜·罗曼诺娃（Svetlana Romanova）提问：“随着 AI 能力日益强大，未来哪些人类特质将变得最为核心？” 阿奇亚姆认为品格、奉献、努力以及同理心这类永恒的人类美德依然不可替代。奥特曼则认同创造力与理解他人需求的核心地位，并分享了一段令他触动深刻的切身经历。奥特曼原本对打卡全自动化的“机器人咖啡馆”充满期待，自认是个实用主义者，不需要星巴克咖啡师的问候与微笑；然而当他真正站在屏幕前点单、看着机械臂端出一杯完美的咖啡时，他感到前所未有的空虚与索然无味： > “走进去在屏幕上按几下，让机器人完成动作端出一杯美味的咖啡，这完全是一次极度缺乏满足感、我再也不想经历的过程。我原以为自己根本不会在乎那些人际互动。” 埃科菲进一步阐述，社交媒体与科技发展在过去数年间其实削弱了现实中的人际联结。随着冷冰冰的认知与操作任务被 AI 大量剥离，人类社会对“情感联结”的渴求反而会被急剧放大。这不仅将全面重估“照护经济”（Care Economy）——如护理、养老、幼儿教育与教学等岗位的社会价值，也将把“你作为一个真实的人，对周围同类究竟有多友善”确立为未来最核心的品质标尺。 ### [36:36](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=2196s) 医疗与照护体系：打破医疗迷宫与普及优质服务 针对美国社会中医疗、托育与养老成本畸高的痼疾，阿奇亚姆展望了 AI 技术重塑医疗服务公平性的前景。他指出，当前医疗体系最大的痛点不仅在于治疗本身的昂贵，更在于复杂官僚系统筑起的“迷宫”——普通患者在错综复杂的保险报销条款、专科转诊流程中晕头转向，经常耗费数年时间在不同诊所间周旋却得不到确切诊断。AI 助手能够充当全天候顶尖全科顾问与医疗导航员，穿透信息壁垒，将全球顶级水平的医疗智慧以边际成本近乎为零的方式普及给每一个普通家庭。 ### [37:12](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=2232s) 医疗体系重塑与患者赋能 在谈及医疗现状时，讨论聚焦于深陷繁琐体制中的患者与医护人员。即便拥有最富关怀精神与同情心的医护团队，现实中依然没有足够的时间让每位患者都获得顶级的诊疗服务。AI 的介入并不是为了取代医生，而是为了在规模化普及高品质护理的同时，将医生的工作负担减轻到可控水平，从而实现真正的“患者赋能”（patient empowerment）。 为了说明这种个体体验的转变，Sam Altman 分享了发生在自己身上的真实经历。他最近做了一次包含上百项指标的常规血液检查，虽然没有重大疾病，但有几项指标轻微超标。当他向医生询问时，医生只是敷衍地表示“差不多都在正常范围内，没啥事”。随后 Sam 将整个血检报告上传到 ChatGPT，模型不仅给出了详尽解析，确认并无大碍，还精准指出了身体代谢的变化逻辑，建议他服用一种特定的膳食补充剂，并于一个月后复检。这种随时上传复杂体检数据即可瞬间获得清晰解答的体验，展现了 AI 在大幅削减医疗、养老和个性化教育成本方面的巨大潜力。 Adrien Ecoffet 补充指出，除了直接加速基础医学科研，AI 对“看护经济”（care economy）的革新关键在于剥离官僚化流程。医疗行业充斥着大量繁重而冰冷的行政文书工作，若能将这些交由系统自动化处理，就能将人类从业者彻底解放出来，把精力真正倾注到面对面的患者关怀与实际照护之中。 ```mermaid flowchart TD accTitle: AI 在医疗与照护体系中的双重赋能路径 accDescr: 展示 AI 如何通过处理数据与行政杂务，释放人类医护力量并赋能个体患者。 subgraph Care["医疗与照护革新"] A["复杂医疗数据与繁琐文书"] --> B["AI 自动化分析与行政减负"] B --> C["个体患者: 实时报告解读与个性化建议"] B --> D["医护人员: 卸下文书重担，专注于人本关怀"] end ``` ### [39:45](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=2385s) 能力悬置与“苦工盲区”的打破 对话随后转向所谓“能力悬置”（capability overhang）现象，即当前 AI 已经具备了极为强大的通用能力，但绝大多数人却根本没有主动去调用和挖掘。主持人援引了保罗·格雷厄姆（Paul Graham）提出的“苦工盲区”（schlepp blindness）概念，指出许多人已经习惯了现实世界的沉重与低效，甚至放弃去设想更美好的解决路径。唯有当大众真正迈出尝试的那一步，群体行为模式的改变才会像底层技术突破一样深刻地重塑整个社会。 面对“何时亲眼见证人们从墨守成规中顿悟”的提问，Sam Altman 讲述了他最喜爱的一个观察场景：拥有专业编程背景的技术人员父母，在旁看着从未写过代码的孩子第一次使用 Codex。孩子们完全不受传统软硬件工程框架与难易程度的认知局限，他们满脑子奇思妙想，直接对着电脑用语音描述自己构想中的电子游戏，而 Codex 就这样当场把游戏编写并运行了出来。 父母最初往往认定“这绝对不可能行得通”，但随着游戏画面在屏幕上实时展开，他们会陷入巨大的震撼：自己的下一代将天然成长在一个“只要说出想法，机器就会立刻实现”的世界里，而老一辈程序员却因为深知实现的“困难”而根本不会去尝试。这种戏剧性的代际反差，就如同 80 年代家里只有小孩子懂得如何熟练摆弄录像机（VCR）一样生动。 ### [42:23](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=2543s) 时间尺度错配与推理模型的认知鸿沟 与技术飞速迭代形成鲜明对照的，是公众认知的严重滞后。AI 的底层进化是以“数周和数月”为周期狂飙突进，而普通公众重新回访、测试产品的频率往往是“半年甚至更久”。由于普通用户极少主动将模型推向能力极限，甚至多数人完全不知道已经出现了具备深度思考链路的推理模型（Reasoning Models），依然停留在早期的基础对话模型上，导致社会上充斥着关于模型“幻觉”、“胡言乱语”或“生成垃圾内容”的刻板印象。 这种体感上的“认知鸿沟”（visceral belief gap）正在孕育剧烈转变。当全社会开始看到尖端机构与先锋群体以最高推理规格运行 AI、解决诸如秒级罕见病诊断或全流程内容构建等硬核难题时，公众的预期将被迫迅速刷新。技术正在彻底向富有创意的人群敞开大门——无论是天马行空的科学家、艺术家还是孩童，他们脑海中大量未曾落地的灵感正迎来爆发式落地的通道。 针对激发灵感本身，Sam 展望了一个即将到来的技术场景： - 未来的顶尖模型将扮演终极“思维伙伴”（thought partner）。 - 用户可以直接授权模型：“去翻阅我所有的短信记录、往来邮件和整台电脑，挖掘我曾经随口提及的零碎碎片或潜在感兴趣的方向，整理出具体的项目构想。” - 创作者无需再苦思冥想待办清单，只需对模型提炼出的灵感进行筛选并立即着手构建。 ### [44:46](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=2686s) 新产业政策倡议与开放研究资助 在论坛尾声，OpenAI 强调本次探讨的所有愿景虽然雄心勃勃，但整体仍处于探索初期，提出这些概念绝非公布既定蓝图，而是旨在作为公共对话的起点，邀请决策者、学术界与社会公众共同参与研讨与博弈。为了让利益成果广泛惠及社会，OpenAI 宣布了一系列推动新产业政策研究的具体扶持举措。 - **反馈征集机制**：开通专门沟通渠道（`newindustrialpolicy@openai.com`），广泛收集各界对 AI 产业政策与发展路径的建设性提案。 - **研究资助计划**：启动试点研究员项目与专项课题基金，单个项目最高可提供 10 万美元现金资助以及高达 100 万美元的 API 算力额度。 - **实体研讨空间**：OpenAI 将于 5 月在华盛顿特区设立全新的 Workshop 研讨空间，持续召集关于监管治理、社会适应与技术潜能的跨界闭门会谈。 ## Notable quotes > [00:05](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=5s) 超级智能已经离我们非常接近了 — 萨姆·奥特曼 > [00:05](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=5s) 我们可能会犯错，也可能会撞上一堵未知的技术高墙 — 萨姆·奥特曼 > [04:36](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=276s) 当下的感觉和当年完全一样：剧烈的变革已经发生，模型已经达到了某种临界水平，但整个社会尚未消化这一事实。我们觉得自己看得很清楚，正在拼尽全力告诉世界它即将到来，而这种信息极难传递。这感觉就像是回到了疫情初期的那个夜晚，独自穿行在街头。 — 萨姆·奥特曼 > [06:33](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=393s) 在一年内完成相当于过去数十年的科学突破 — 萨姆·奥特曼 > [30:39](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=1839s) 没有任何人应该因为失去工作而失去医疗保险，这种事情根本就不应该发生。 — 萨姆·奥特曼 > [32:27](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=1947s) 官方设定的目标是 2028 年 3 月 — 乔希·阿奇亚姆 > [33:58](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=2038s) 走进去在屏幕上按几下，让机器人完成动作端出一杯美味的咖啡，这完全是一次极度缺乏满足感、我再也不想经历的过程。我原以为自己根本不会在乎那些人际互动。 — 萨姆·奥特曼 > [42:23](https://www.youtube.com/watch?v=ZpUKNYcgM-E&t=2543s) 去翻阅我所有的短信记录、往来邮件和整台电脑，挖掘我曾经随口提及的零碎碎片或潜在感兴趣的方向，整理出具体的项目构想。 — 萨姆·奥特曼 ## Mentioned OpenAI, 克里斯·尼科尔森 (Chris Nicholson), 萨姆·奥特曼 (Sam Altman), 乔希·阿奇亚姆 (Josh Achiam), 阿德里安·埃科菲 (Adrien Ecoffet), 超级智能 (Superintelligence), Y Combinator, 亚马逊 AWS, ChatGPT, Codex, 保罗·格雷厄姆 (Paul Graham), 斯维特兰娜·罗曼诺娃 (Svetlana Romanova), 星巴克, 全民基础算力 (Universal Basic Compute), 无条件基本收入 (UBI) ## Video description AI is advancing faster than most people realize. In this OpenAI Forum conversation, Sam Altman joins Josh Achiam and Adrien Ecoffet to talk about what’s coming next. They discuss the pace of progress toward more capable AI systems, what these tools could unlock—from scientific breakthroughs to new ways of building companies—and the challenges society needs to prepare for. The conversation also looks at who gets access to AI, how to make sure the benefits are widely shared, and what it will take for institutions and individuals to keep up. [Source](https://www.youtube.com/watch?v=ZpUKNYcgM-E)

[0:05] 萨姆·奥特曼: 超级智能已经离我们非常接近了 [0:05] 萨姆·奥特曼: 我们可能会犯错，也可能会撞上一堵未知的技术高墙 [4:36] 萨姆·奥特曼: 当下的感觉和当年完全一样：剧烈的变革已经发生，模型已经达到了某种临界水平，但整个社会尚未消化这一事实。我们觉得自己看得很清楚，正在拼尽全力告诉世界它即将到来，而这种信息极难传递。这感觉就像是回到了疫情初期的那个夜晚，独自穿行在街头。 [6:33] 萨姆·奥特曼: 在一年内完成相当于过去数十年的科学突破 [30:39] 萨姆·奥特曼: 没有任何人应该因为失去工作而失去医疗保险，这种事情根本就不应该发生。 [32:27] 乔希·阿奇亚姆: 官方设定的目标是 2028 年 3 月 [33:58] 萨姆·奥特曼: 走进去在屏幕上按几下，让机器人完成动作端出一杯美味的咖啡，这完全是一次极度缺乏满足感、我再也不想经历的过程。我原以为自己根本不会在乎那些人际互动。 [42:23] 萨姆·奥特曼: 去翻阅我所有的短信记录、往来邮件和整台电脑，挖掘我曾经随口提及的零碎碎片或潜在感兴趣的方向，整理出具体的项目构想。

[34:35] Sam Altman: I thought I was someone that did not need like the barista at Starbucks to smile at me and say hi... It turns out that I really want that... walking in to like push on the screen and have the robot do the thing... was like a deeply unfulfilling I don't want this experience. [39:45] Sam Altman: The parent is just like, 'that's not going to work,' and then it works, and then they're like, 'wow, my kid is going to grow up in a world where he or she just like expects this happens.' [0:43] crazy people at OpenAI [30:54] Sam Altman: no one should lose their health care if they lose their job.

[0:05] Good afternoon, everyone, and welcome to [0:06] the OpenAI forum. I'm Chris Nicholson, [0:09] and I'm glad to be here with all of you [0:10] today. [0:11] The forum is a place for serious [0:13] conversation about how AI is being used [0:16] in the world, what we're learning from [0:18] that, and how more people can help shape [0:20] its trajectory. Today's conversation [0:22] focuses on one of the biggest questions [0:25] in technology, what it will mean as AI [0:27] systems grow dramatically more capable, [0:30] and how we should think about their [0:31] implications for science, work, our life [0:35] together, and governance. [0:37] To discuss that, I'm joined by Sam [0:39] Altman, co-founder and CEO of OpenAI, [0:41] Josh Acklin, [0:43] OpenAI's chief futurist, and Adrian [0:45] Akyol, [0:46] a long-time researcher here. So, let's [0:48] get started. [0:50] Sam, [0:51] the blueprint we released this morning [0:53] talks a lot about superintelligence. Um, [0:56] a big question on my mind is, [0:58] why are we doing that now, and what are [1:00] some things that you can see from the [1:01] inside that you wish everybody knew? [1:04] >> The The biggest reason is simply that [1:06] the rate of progress is continuing to [1:08] accelerate, and we believe we are very [1:10] close now. [1:12] And this won't be a one-time thing. This [1:13] will be a [1:14] you know, [1:16] over the next few years to powerful [1:17] models that will impact the world in [1:19] important ways. The researchers that are [1:21] working on these models that did and [1:24] I think an incredible job with this set [1:26] of ideas, and these are meant to be [1:28] early ideas to start a discussion. [1:31] You know, I'm sure we'll get to much [1:32] better ideas as the world debates all of [1:33] these. [1:34] Are staring down the models that are [1:36] coming, the pipeline in front of us. [1:39] And [1:41] you know, we may be wrong. We may hit [1:42] some wall. We We are imperfect, but our [1:45] But given what we see, [1:47] we expect to be in a world of extremely [1:50] capable models quite soon, and then for [1:53] the the ramp of the capability to [1:56] continue to increase. I think this will [1:58] have huge impact on the economy, on the [2:00] way we live, on what we can do. And [2:03] one thing I've observed from watching [2:05] the world go through some number of [2:07] transitions is that the more time the [2:11] the public, our leaders, the political [2:13] system has to debate ideas before you [2:15] really have to make a decision, the more [2:17] likely you are to make a good decision. [2:19] So, starting this now, given what we see [2:21] coming, so that as this becomes [2:23] a very large issue of public debate, I [2:25] think it's important. [2:26] >> For sure. Um and speaking of debate, we [2:29] brought in a lot of researchers in [2:31] during very early in this process. I [2:33] think it was a [2:35] maybe a unique exercise here in how many [2:36] researchers were working [2:38] with the folks who also think about [2:40] policy a lot. [2:42] It's closely tied. What was that like [2:44] for you and for the research group as a [2:46] whole? [2:47] >> Yeah, you know, I think it was Well, it [2:49] was a very interesting experience. Like [2:50] it was my first time [2:52] working like actively on a on a policy [2:55] doc. I think it was a little bit like [2:57] humbling in some cases, you know, uh [2:59] uh sometimes [3:01] as researchers, we can have these [3:02] abstract ideas of oh, we should really [3:05] be thinking about the economic impact of [3:07] this or or about like policy for for [3:09] safety, but you know, it's it's one [3:11] thing to think about this and another [3:12] thing to like actually put, you know, [3:14] pen to paper and um and think of [3:17] concrete [3:18] uh policy ideas that are going to be [3:19] debated by by your your peer [3:22] uh your your peer researchers. Um so, it [3:24] was an interesting experience for me. [3:25] Um I I hope there was [3:29] uh you know, very helpful for the final [3:31] product uh that we had the researchers [3:34] involved so early and so and so deeply [3:36] like for for something like this, uh [3:38] especially since it's so [3:40] forward-looking. [3:42] Uh you kind of need to to Sam's point, [3:44] like you need people who are um [3:47] who are dealing with this technology [3:48] every day, who know how to build it, [3:50] know how the safety stack works, [3:52] and also are seeing like [3:54] the speed of progress. Um and one thing [3:58] that, you know, that I remember as part [4:00] of this process in the in the past few [4:02] months that we were were working on it, [4:04] like [4:05] a lot of researchers went through a a [4:07] transition [4:08] uh [4:09] from, you know, writing most of their [4:11] own code to having AI most write most of [4:13] their code. And I think that I think to [4:16] some extent led them to bring a lot of [4:18] the urgency of like [4:19] this technology is is real and important [4:24] moving fast in a way that maybe um [4:27] not everyone can can see. Uh and and [4:30] that's part of also kind of your your [4:32] earlier question about why now, like [4:34] because we are seeing uh this urgency. [4:36] >> Can I tell a quick story that reminds me [4:38] of? Um [4:40] There was [4:42] a night [4:44] in [4:45] I guess it would've been [4:47] late January, early February of 2020, [4:49] when uh the OpenAI researchers kind of [4:52] got obsessed with COVID before the rest [4:55] of the world did. And we were talking [4:56] about it all the time, and we were [4:57] watching the numbers every day, and [4:58] we're like, "This is going to happen." [4:59] >> Yeah. [5:00] >> And, you know, we were like making plans [5:02] to go [5:03] uh [5:04] work from home, and there was this like [5:05] some article that came out like mocking [5:07] us because they're like these crazy [5:08] people at OpenAI. We had put like some [5:10] copper or something on some of the door [5:12] handles because people [5:13] >> I remember that. [5:14] >> You remember that? [5:15] And some journalist wrote about this, [5:16] and there's all all these things going [5:17] on. And I [5:21] And you know, we we kind of like [5:24] made our plans, and we assumed a [5:26] shutdown was going to happen. We're [5:27] like, "There's this thing coming. The [5:28] world's not paying attention." And for [5:30] whatever reason, uh something about like [5:32] working on exponential makes you [5:33] understand these things better. So, I [5:34] think we were like a group of people [5:36] that were primed to do it. Um [5:38] And then there was this one night, it [5:40] was like a very cold night. I lived in [5:41] the mission at the time. And I was like, [5:43] "I'm about to get locked in my house for [5:45] a while, and I'm going to go for a walk. [5:46] I want to one more time I want to like, [5:48] you know go for a walk cuz who knows [5:49] what what's going to happen. [5:51] And I went through this long walk [5:52] through the city, hours, cold cold [5:54] night. And I was like watching people, [5:55] you know, breathing in each other's [5:56] faces in restaurants and bars through [5:58] the windows. And I was like wearing my [5:59] mask looking all looking crazy. And [6:01] there was one other dude out wearing a [6:02] mask and kind of we nodded at each [6:04] other. And that was But other than that [6:05] it was just like [6:06] life felt totally normal. [6:08] I have not felt that so acutely as I do [6:11] again in this moment. [6:13] >> Yeah. [6:13] >> Um where it's like [6:16] there is this crazy change. The change [6:18] has already happened. Like the models [6:19] have already hit some level. Society has [6:21] not digested them yet. [6:22] >> Yeah. [6:23] >> We feel like we see it clearly. We are [6:25] trying to tell the world that it's [6:27] going to happen. It is hard to get this [6:29] across. But it feels like that night in [6:31] the very beginning of COVID walking [6:32] through the streets again. [6:33] >> That's really interesting. And and [6:34] you've talked a bit about the upsides to [6:37] unlock and a lot about science as a [6:39] means to do that. Um so so if we're [6:43] going to like contrast this with COVID, [6:45] what are some of the hugely positive [6:48] things in the pipeline for us? [6:51] >> I will answer that. I think the [6:52] positives are so positive we should talk [6:54] more about the other things that we're [6:56] thinking about here. But like it it you [6:58] mentioned science. If we can really go [7:00] make a [7:01] decades worth of scientific progress in [7:03] a year, if we can go cure a ton of [7:05] diseases, if we can come up with [7:06] personalized medicine for people, find [7:08] new materials to sort of make cheap [7:11] safe energy [7:13] there's [7:14] if we can make it such that anybody who [7:16] can come up with an idea for a startup [7:17] can have the AI implement it, write the [7:19] piece of software they want, [7:21] uh you know, have a custom video game [7:22] that's like the most fun for you to [7:23] play. Uh [7:25] this stuff is all wonderful. Now, [7:28] you know, part of the reason for putting [7:29] this out with urgency is there are going [7:30] to come there are a lot of things coming [7:32] that we'll need to mitigate. [7:34] I assume we'll figure it out. I I'm [7:36] always an optimistic person by nature. [7:37] But also [7:39] the benefits of this this [7:40] has change the [7:42] option space in front of us and what we [7:44] can do that one of [7:47] my first reaction to reading the [7:50] blueprint after the researchers wrote it [7:51] was like it is awesome but also a little [7:54] crazy that we can credibly be talking [7:55] about these kinds of things. [7:57] Um so we do have like [7:59] if [8:00] the AI is as good as we think [8:02] >> Yeah. [8:03] >> and all these wonderful things happen, [8:04] we also have an incredible new tool at [8:06] our disposal to help us mitigate the [8:08] potential downsides. [8:09] >> Yeah. And it's a tool for everyone. Uh [8:11] Josh, I'd like to ask you [8:13] So So I I think we've talked about this [8:15] as social infrastructure that changes [8:17] the way people work and learn and [8:19] participate in society. Um [8:22] what [8:24] what responsibilities do we have and do [8:26] institutions have to help society get [8:29] ready for that and how how do you think [8:30] about it? [8:31] >> Yeah. Um you know, one of the things [8:33] that I think about here, this does kind [8:34] of come back to the broad benefits for [8:35] everyone thing. [8:37] Uh for a long time in society, we've had [8:39] this great aspiration that everyone [8:41] could have, you know, food, shelter, [8:44] electricity, [8:45] uh health care and we've we've always [8:48] wanted [8:50] to see there be some kind of [8:53] new step in the direction of providing [8:54] these sorts of things for everyone so [8:55] that everyone can focus on the things [8:57] that are most important in their lives. [8:59] Um and we've often been told over and [9:01] over again like, you know, actually we [9:03] can't have that, it's too expensive. [9:05] There's no way to pay for it. Um [9:07] like it's it's nice but it would be too [9:09] burdensome to society to do all of that. [9:12] And I think what AI and [9:14] superintelligence will unlock is the [9:16] freedom to do all of it at a much lower [9:19] cost than has ever been possible. And [9:21] correspondingly for the folks of us who [9:23] are sort of the stewards of the [9:24] technology, we have a special [9:26] responsibility to make sure that we [9:28] actually do fully realize those benefits [9:30] and be a part of building policies and [9:34] systems that help working people, middle [9:36] class people, people in low-income [9:37] countries, really make sure that it [9:39] benefits everyone and not just the very [9:41] wealthy. So, I'm those are the things [9:43] that I think of as key responsibilities [9:45] here, and I'm very excited about it. [9:47] Plus, [9:48] the downside risks are also quite [9:50] serious, and there are systemic shocks [9:53] that could happen. We have to prepare. [9:54] We have to be thoughtful. We have to be [9:55] honest about them. [9:56] >> Yeah. Yeah, so that sounds like a [9:58] resilience question. Um [10:00] and and when I [10:02] look at the document and and speak with [10:04] you folks, it sounds like you think of [10:05] resilience in layers, right? There's um [10:09] kind of and and a lot of it is not [10:10] before the AI goes out, but it's also [10:12] after the AI goes out in terms of um our [10:15] responses to AI and how people are are [10:17] prepared for it. I'd actually like to [10:18] ask each of you how you think about [10:20] resilience. Um Adrian, why don't we why [10:22] don't we start with you? [10:23] >> Yeah. You know, I I think maybe you want [10:26] one distinction [10:28] that I would draw with uh [10:30] you know, classically we've thought [10:31] about about safety and about making sure [10:33] that we that we run, you know, safety [10:36] evals, that we red team our models, um [10:38] that we implement mitigations. And I [10:40] think like that's, you know, to your [10:41] point of resilience is a layer. This is [10:43] like a very important layer that we [10:45] should, you know, still keep doing and [10:47] keep expanding, but um [10:50] at the same time [10:52] uh [10:53] you want society to [10:56] uh be prepared for the possibility that [10:58] well, maybe there will be some actors [11:00] who do less safety testing, and like [11:03] what what happens what happens then? How [11:05] is society resilient to risks from AIs [11:07] uh in these cases? [11:09] Um maybe there will be [11:11] uh incidents in spite of our safety [11:13] testing or near misses uh in the [11:15] in the [11:17] uh blueprint, we talk, for instance, [11:18] about um [11:20] incident reporting that's modeled a [11:21] little bit of after um [11:23] how the uh aviation industry uh does [11:26] things when whenever there's kind of [11:28] like a near miss or any incidents, like [11:30] however minor, that kind of like gets [11:32] reported to um to a database so that, [11:35] you know, all the companies can kind of [11:37] know, okay, well, this is a risk and [11:39] perhaps uh you know, here are uh [11:41] mitigations that that you could [11:42] implement. And so, there's a lot of [11:44] things that could happen at a [11:45] society-wide uh [11:47] level. [11:48] Um another thing would be um [11:51] you know, defending against uh [11:53] against risks. Uh [11:55] we're talking about models that that can [11:57] code um [11:59] a lot better now. Uh that also implies [12:02] uh cyber capabilities. Like, could they [12:05] help uh bad actors uh run cyber attacks? [12:08] Uh and [12:10] I think part of resilience is ensuring [12:12] that we uh [12:14] that we make our software systems [12:16] uh more secure uh and use AI to do these [12:19] things. So, so again, there's as you [12:21] said, all these layers. So, that's kind [12:22] of how I take the the resilience [12:24] question. [12:25] >> Sam, before the show, we were talking [12:26] about how prosperity can be emergent [12:29] when everybody has access to AI. It also [12:30] strikes me that resilience can be [12:32] emergent. How How do you think about it? [12:34] >> Yeah, I [12:37] I think our original AI safety thinking [12:39] and the field as in general, like I'll [12:40] call it classical AI safety thinking, [12:42] was that there are going to be a very [12:45] tiny number of AIs in the world. The [12:47] only thing that matters is making sure [12:49] those AIs do the right thing and that as [12:51] long as you align them and they don't do [12:53] unsafe things, the world will be okay. [12:56] I think the picture now is [12:58] actually more stable but more complex [13:00] and there are going to be many AIs in [13:02] the world uh and the [13:06] it will not [13:07] be enough to just say like, you know, [13:09] this one company is going to make sure [13:10] the AI never does something it shouldn't [13:12] do. Um but there will need to be an [13:14] emergent response across society. Uh [13:19] you know, [13:20] Adrian talked a little bit about some of [13:21] this, but if we just take a few examples [13:23] of threats that we expect to be coming. [13:25] Um cybersecurity is definitely going to [13:27] become a huge issue. AI will be [13:29] incredibly good at finding [13:30] vulnerabilities in software, and I think [13:33] the world will find that their software [13:35] is much more brittle and much less [13:36] secure than we thought. And you know, [13:38] humans just have limited [13:39] capacity to find the exploits. Um, [13:43] it is not enough to say that, you know, [13:45] one or two or three model providers are [13:46] just going to make sure that their [13:48] systems won't do this because code is [13:50] like, you know, the same thing that [13:52] code is very useful to be good at, and [13:53] being good at writing code can also help [13:55] find security problems. And even if all [13:58] of us somehow could prevent our models [14:00] from ever being used for this, there [14:02] will be open-source models coming soon [14:04] that are that are good at code and thus [14:05] good at security exploits. So, what has [14:06] to happen is [14:09] the world has to use these models, and [14:10] there can be, you know, differential [14:11] access. You can give it to good known [14:13] trusted defenders first. We have a [14:14] program for that. Other companies will [14:16] do similar things. Um, [14:18] and you have to empower the companies [14:20] that defend software because there will [14:22] be some power plant that no one has to [14:25] the software for 20 years, and no one [14:26] can patch it, and there's a big problem [14:27] with it, and you have to do something [14:28] about that. Um, but [14:31] a resilience approach here is, okay, [14:34] there's this new thing in the world. [14:35] There's AI that is really good at [14:37] exploiting computer systems. Um, let's [14:39] use AI to defend it, and that is not a [14:41] one-company thing. That's going to [14:42] require this huge effort. Um, [14:45] if we go a little bit further, I I think [14:47] there will be a uh [14:49] bio version of this where, [14:51] you know, classically people have said, [14:53] well, we're just going to restrict our [14:54] models from being able to develop [14:56] pathogens. Someone at some point is [14:57] going to use some model to develop a [14:58] pathogen. And the world needs defense [15:01] shields against that, detection systems, [15:04] rapid response treatments, a whole bunch [15:05] of other things. [15:06] And this is not [15:09] This doesn't get us off the hook in any [15:10] way of aligning our systems and building [15:12] safe systems. We still have to do that. [15:14] Like, we get a time advantage there as [15:16] long as we stay at the frontier. [15:18] But, we do need the world to like do its [15:21] thing, society to have this emergent [15:23] magic and build these layers of [15:25] defensive shields. Um there are many [15:27] threats besides those two, but that's [15:28] enough given we have short time. [15:30] >> For sure. Um [15:31] Josh, you have shared some interesting [15:34] thoughts with me about how each huge [15:36] technological shift has produced new [15:39] institutions and new democratic [15:40] mechanisms. [15:41] >> Mhm. [15:42] >> And also you've uh been thinking a lot [15:43] about new institutions that might emerge [15:45] now. Um how are you [15:48] What are What are your kind of [15:51] What are your most exciting ideas and [15:54] what what do you really like to think [15:55] about in terms of the collective [15:56] response to superintelligence? [15:59] >> Yeah, certainly. Um so so one thing I'll [16:01] say first on the subject of resilience [16:03] just as grounding, a lot of the problems [16:05] that we're concerned about AI creating [16:07] new externalities for are problems and [16:09] vulnerabilities that exist in the world [16:11] regardless of whether AI is present and [16:13] AI just increases the urgency of action. [16:16] But coming back to COVID, we found back [16:18] then that everyone had a much deeper [16:20] dependency on the functioning of supply [16:23] chains than most people were previously [16:25] conscious of. [16:26] Um supply chains are super important, [16:28] supply chains for food, for goods, for [16:30] for everything that sustains [16:31] civilization. Um and also of course uh [16:34] you know, for for other types of [16:36] vulnerabilities like democracy. We've [16:37] been debating for years how um there are [16:39] real risks when people are influencing [16:41] society sort of inappropriately. [16:44] And we're we're worried that AI is going [16:46] to make these things potentially easier [16:48] to attack in the near term because there [16:50] will be tools at people's disposal that [16:52] they haven't had before. Um I'm excited [16:55] about the possibility that we can build [16:57] new institutions and state capacities to [16:59] use AI to rectify some of these [17:01] vulnerabilities. Um we can [17:03] systematically identify them with AI in [17:04] ways we couldn't in the past. We can [17:07] systematically close them in ways we [17:08] couldn't in the past and we can use AI [17:10] to scale up efforts to combat certain [17:11] types of issues uh in in ways where we [17:14] can potentially make it too expensive [17:16] for attackers to really do something. [17:18] So, uh you know, on on cyber and bio, I [17:21] am optimistic that maybe we can build an [17:24] ecosystem of defenders that all together [17:27] can make it so expensive for attackers [17:29] to try to, you know, run a cyber attack [17:31] that they there just won't be that much [17:32] of an incentive to do it. And if we can [17:35] fully implement the the bio resilient [17:37] side of things, um not just for [17:39] pathogens that might impact humans, but [17:41] uh especially for the food supply chain. [17:42] This is one where I have like a hobby [17:43] horse on this and I'm going to talk [17:44] about it every chance I get. I think [17:45] people under attend the food supply [17:47] chain bio risks. Um but we can use AI to [17:50] make that resilient at scale in a way [17:52] that today is cost prohibitive. Um so [17:54] I'm very excited about, you know, the [17:55] things we can build there. [17:56] >> Yeah. Neat. Um [17:58] one more question. [18:00] When we talk about kind of the [18:01] individual transition for many people, [18:04] um [18:04] what work looks like, where value shifts [18:06] to. I know you've been thinking about [18:08] that a lot. Um [18:09] how how what do you see happening as [18:12] folks kind of transition to other ways [18:15] of creating value in an AI economy? [18:18] >> Oh. Oh goodness. That's very broad. Um [18:21] I I think people will have [18:24] a lot more uh opportunity to exercise [18:27] agency. Um [18:29] you know, if you can [18:30] uh start a new business and have a team [18:34] of AIs that handle all of the functions [18:36] of putting the business together that [18:37] you have no expertise in yourself. Um [18:39] you can get something off the ground an [18:40] awful lot easier. And so, you know, [18:42] there there are a lot of ways that the [18:43] economy just fundamentally changes when [18:44] you get people access and tools to help [18:47] them do more things than they could have [18:48] before. [18:49] >> Yeah, for sure. Um [18:51] Sam, similar question. You have so much [18:53] experience with startups. [18:54] Running them, managing Y Combinator, or [18:56] just being in the ecosystem for so long. [18:59] What um how do you see startups changing [19:02] and like our potential to realize new [19:04] ideas changing with AI? [19:05] >> I'm obsessed with trying to explore the [19:07] space. I don't know exactly what it's [19:08] going to look like, but I this idea of [19:11] one person or a very small team being [19:13] able to create an entire startup [19:16] as Josh was talking about pretty [19:17] quickly. [19:18] Um [19:21] all my instincts are there's something [19:23] deep and important to figure out here. [19:25] Every time [19:27] in our industry that the [19:31] friction, cost, whatever you want to [19:32] call it of starting the startup has come [19:33] down a lot. [19:35] Amazing new things have happened. Um you [19:37] know, I remember the transition The one [19:40] of these transitions that I was doing uh [19:42] a startup during [19:43] uh was when like basically AWS came out [19:47] and all of a sudden there was like an [19:49] idea of a cloud and a small startup [19:51] didn't have to go do all the crazy [19:52] things you used to have to do. [19:53] >> Manage his racks in the closet. Yeah. [19:55] >> And that was like an amazing change to [19:59] what you could do as a small startup [20:01] with a few people. This one that's [20:02] coming is much bigger and there have [20:03] been several in between. But uh [20:07] I really want to find out what it looks [20:09] like when a startup is [20:11] you know, two or three people and a ton [20:12] of GPUs and you can just [20:13] >> Yeah. [20:14] >> you can really democratize who can start [20:17] a startup. [20:17] >> Yeah. Yeah. Yeah, it's that [20:19] democratization aspect and it gets down [20:21] to the widespread availability of AI. Um [20:25] how do you What's What's the best frame [20:27] for thinking about bringing more people [20:30] in, democratizing AI access? [20:35] >> I think when people talk about [20:36] democratization of AI, they they mean [20:37] two different things. One is share the [20:40] access [20:41] and making sure that everybody gets to [20:43] use sufficient AI to [20:45] improve their own lives, build things [20:47] for other people, all that. And the [20:48] other is a sort of a voice in where it's [20:50] all going to go. [20:51] Uh I think both are very important. Um [20:54] part of why we do things like this [20:56] blueprint are [20:58] if we're not debating the issues as a [21:00] society. Also part of the reason we [21:01] actually release products in the first [21:03] place. People don't have a feel for this [21:04] if we're not talking about this. [21:06] Um that is kind of a prerequisite to [21:08] people being able to have input, um, but [21:10] it's not enough. You also have to have a [21:11] way that people's input that we listen [21:13] and the input is captured back into the [21:15] into the system. [21:16] Um, so that's one that I think is really [21:19] important and the other is [21:21] we need to put [21:23] not just services like ChatGPT, but the [21:26] real deal high compute valuable services [21:28] where people can start a startup or make [21:31] a scientific discovery or whatever in [21:33] broadly into many people's hands. [21:34] >> Yeah. Yeah, and that's going to take new [21:37] economic models or much cheaper [21:40] uh, inference to to get it [21:42] wider spread, right? And that's an [21:43] infrastructure problem among many other [21:45] problems to solve. [21:47] >> You know, we we used to talk for years [21:49] in OpenAI about when we were going to [21:50] get through the compute crunch, when we [21:51] were going to build like enough compute [21:52] that we wouldn't be so strapped. I don't [21:54] think we ever get out of it. If we do [21:56] our job, if we keep driving the cost of [21:58] intelligence down and the capabilities [22:00] of intelligence up, I think uh, you [22:03] know, effectively if that happens, the [22:05] demand is uncapped. [22:06] And [22:08] the worlds where [22:11] we don't build enough [22:14] infrastructure, [22:15] I think you get a crazy concentration of [22:18] power and concentration of compute [22:20] because people will just bid the price [22:21] up and up and up. So, the only thing [22:23] that I really believe in as a long-term [22:25] democratization strategy is to just make [22:28] so much AI infrastructure available and [22:30] make the models so good and so capable [22:33] um, that we hope to get to a world where [22:34] people are like, [22:35] "I need help coming up with ideas for [22:37] what to use all this compute for." I [22:38] don't think we will in practice. [22:39] >> Yeah. [22:40] >> Um, but I certainly think that if [22:42] compute is very limited, [22:44] you know, the richest people and the [22:46] richest companies in the world will just [22:47] sort of bid up the price to a kind of [22:49] extreme degree. [22:50] >> It'll be another kind of scarcity that [22:52] that is um, monopolized. Where So, this [22:55] is a more data centers is actually a [22:57] very egalitarian [22:59] um, initiative [23:01] in the sense that they can make AI [23:03] access more widespread. [23:04] >> I I would look at many examples [23:06] throughout history. You know, one of the [23:08] best things that we ever did for really [23:10] increasing people's quality of life was [23:12] to drive the price of electricity way [23:14] down. [23:15] >> Mhm. [23:15] >> Price of energy broadly way down. [23:17] >> Mhm. [23:17] >> Um energy correlates incredibly well [23:20] with quality of life, or at least it has [23:21] for a long time. Maybe now it'll be more [23:22] about AI. [23:23] Um [23:24] and [23:26] by making energy abundant and shockingly [23:28] cheap relative to what it cost 50, 200 [23:31] years ago, [23:32] uh that has done like quite a bit for [23:35] lifting the entire world up. I think we [23:37] need to do the same thing with AI. [23:38] >> Mhm. [23:38] >> And I think that means you need a lot of [23:40] it, and just like with energy, you need [23:41] to innovate new ways to make it much [23:43] more affordable. [23:44] >> Yeah. [23:45] Can I [23:47] give give a thought on on this whole [23:48] issue as well? I think we've talked a [23:50] little bit about [23:51] uh you know, the broad access to AI, [23:53] which I think is very important and and [23:55] can create like [23:57] you know, potentially a lot of new [23:58] products that will be um very helpful [24:01] for the world, very useful for people. [24:03] Um I think one thing that, you know, we [24:05] try to think about as well [24:07] in in writing this blueprint is [24:12] you know, ensuring that, you know, [24:14] ordinary people who maybe, you know, [24:15] aren't going to uh start start start up [24:17] necessarily um all of them like aren't [24:19] left behind by this technology. Um I [24:22] think this is related to to uh something [24:25] that that we might talk about, which is [24:27] kind of [24:28] how does AI change the composition of [24:31] the economy? Does it uh move it more [24:34] towards [24:35] uh capital or towards like [24:38] labor, but labor that that is done by by [24:40] AI really. Um and and one of the things [24:43] that we talk about in the blueprint is [24:45] uh how do you monitor modernize the the [24:48] the tax base for an economy that is like [24:50] this uh uh and how do you distribute [24:53] kind of [24:54] uh this prosperity that, you know, will [24:56] be created by this technology? Uh how do [24:58] you make sure that this is prosperity [24:59] for everyone and not like wealth for, [25:02] you know, a relatively few few people. [25:05] And so, this is something that we try to [25:07] address in in the blueprint, but I'm [25:09] curious if we can [25:10] talk about it here as well. [25:12] >> we totally should. A couple of the big [25:14] ideas that I saw were [25:16] have workers co-authoring how AI is [25:19] deployed in their workspaces. I'd like [25:20] to to ask you about that, Josh, and I'd [25:22] also ask I'd like to ask all you about [25:24] more broadly how society can capture the [25:27] upside. What are the institutional forms [25:29] that'll [25:30] that'll take, but let's start with you, [25:31] Josh. [25:32] >> Sure. Just on the on this compute [25:33] question for 1 second, I want to say the [25:34] compute allocation problem, figuring out [25:37] what things we use compute to help [25:40] people do with AI will probably one of [25:42] the most important society-wide [25:43] questions to navigate over the next few [25:45] years while compute is relatively [25:47] scarce, and we should try to boost the [25:49] amount of compute in the world as much [25:51] and as quickly as we possibly can so we [25:53] don't have to have painful trade-off [25:54] questions where there's some [25:56] extraordinary good we could provide to [25:57] everyone, and we're stuck with a hard [25:59] question where someone says, "Well, how [26:00] are you going to pay for that?" because [26:01] the cost of compute is so high. Um on [26:04] getting workers involved in AI, [26:06] um I I actually I kind of want to back [26:08] up and just acknowledge an elephant in [26:09] the room, which is that a lot of workers [26:11] are concerned about AI. They're worried [26:13] about what AI means for them. They are [26:15] not immediately excited at the prospect [26:17] of of figuring out, all right, how are [26:18] we going to use AI in our workplace? And [26:19] they're thinking, "Oh my gosh, is the AI [26:21] going to replace me?" Um and what I what [26:25] I think is sort of the important first [26:26] step here is that those of us who are [26:27] working on AI and who are kind of the [26:29] stewards for this [26:30] have to be putting out things like this [26:32] this blueprint document where we say, [26:34] "Well, here's how we're going to [26:35] advocate for policies to make sure that [26:37] the economy is fair and that you are [26:38] supported, no matter what." And then, [26:41] given that we've made this level of [26:43] safety net, [26:44] now we can talk about, you know, [26:46] something where where you have [26:47] confidence in in us, we can talk [26:49] together, have a good conversation, Um, [26:51] we should figure out how do we empower [26:53] unions to make wise choices about where [26:55] and how to use AI. Um, how do we empower [26:58] workers to participate in conversations [26:59] about the acceptable use of AI in the [27:01] workplace. I think a lot of folks are [27:03] rightly concerned about AI surveillance [27:05] in the workplace, making sure that [27:06] workers are a part of the decisions [27:07] about those types of things feels very [27:09] important. And um, a big push on AI [27:11] literacy to ensure that folks um, [27:15] you know, get the tools they need to use [27:17] AI to make their lives better, uh, to [27:19] start small businesses, to to do all the [27:21] kinds of things that, you know, will [27:22] help them realize their potential. [27:24] >> For sure. Um, [27:26] what are some institutions that you've [27:28] considered that allow everyone to [27:30] capture some of the upside, Josh? [27:33] >> Um, [27:35] so [27:36] let's see. For for new kinds of [27:37] institutions, I think state capacity to [27:40] measure pieces of the economy in, you [27:42] know, greater granular depth so that [27:44] there can be responses if there's, you [27:45] know, like an economic shift in one [27:46] place or another. [27:47] >> yeah. [27:48] >> Um, I think AI's actually a very [27:49] exciting tool for making that more [27:51] scalable and less expensive than it [27:52] would have been in the past. [27:53] >> Yeah. [27:53] >> Uh, I I also think that there can be [27:55] more institutions that are kind of in [27:57] between corporations and governments, [28:00] which have very very different levels of [28:02] accountability governance-wise. There [28:04] maybe is like a need for something with [28:05] like more in between level of [28:07] accountability. Um, [28:09] that can [28:11] provide uh, [28:12] you know, like social safety net type [28:14] services or [28:16] >> Something between a corporate board and [28:17] the regulator. [28:18] >> Yeah, yeah, yeah. Like there's [28:21] um, [28:22] like this can't all just happen in [28:24] private companies that have very minimal [28:25] governance. Uh, and we also don't expect [28:28] that government, which moves fairly [28:29] slowly, is going to do all of it quite [28:31] quickly. And so we sort of maybe need [28:32] some in-between institutions that can [28:33] help us prototype things. Um, this is [28:36] like a little bit spitballing. Uh, [28:38] and I I I don't think this is like quite [28:39] covered in the blueprint, but uh, you [28:41] know, you asked and so that's I think in [28:42] the innovative new institutions that I [28:44] think we could do. [28:45] >> Cool. Um, Sam, institutionally, how are [28:48] you envisioning that people might [28:50] broadly participate in the upside? [28:56] >> I think [28:57] we talked earlier about very broad [28:59] access and giving a lot of people a lot [29:01] of compute. Um, I [29:04] you you hear these ideas like universal [29:06] basic compute or other things with nice [29:08] branding, but really what they mean is [29:10] just like [29:11] uh [29:13] instead of the traditional thinking of [29:15] we're going to give people, you know, a [29:16] monthly stipend or money or whatever [29:18] when AI does all the jobs, I think it's [29:19] way better to say, "Actually, people are [29:21] pretty good about knowing what they need [29:23] and pretty creative about figuring out [29:24] how to use things." But if people are [29:26] boxed out of access to this resource, [29:28] uh that will be [29:30] a a challenge. I do suspect that we're [29:32] going to have to make changes to how we [29:34] tax like in a world where AI is doing [29:37] most of the intellectual work in the [29:39] world or at least of the work of today. [29:41] Um, [29:41] you know, we probably are going to need [29:43] to explore some way to tax that instead [29:44] of taxing human income in the [29:46] traditional sense. [29:48] Um, I suspect that we will need to [29:50] provide new kinds of transition [29:52] assistance, unemployment insurance, [29:54] things like that. Um, and I suspect that [29:56] eventually we will need to think about [29:58] how people get to [30:01] be an owner in the upside of all of this [30:02] in new ways. Um [30:04] you know, capitalism is dependent on a [30:06] certain balance between labor and [30:08] capital and if that gets totally out of [30:10] whack, then the current system is not [30:12] going to work and there'll have to be [30:13] some sort of evolution. What that is I [30:15] think is a very open question and again [30:19] part of the goal here is to throw out [30:20] some ideas, but there's many more. [30:22] Um [30:24] and I will always leave some room and [30:26] say maybe we're wrong and maybe no [30:27] change at all is required and somehow [30:28] this just works differently than we [30:30] think. [30:30] But again, in a spirit of trying to use [30:34] the time we have to think and debate, it [30:37] seems like a good time to start putting [30:38] ideas here. [30:39] >> Can I say something on this maybe we're [30:41] wrong aspect? Like I I think one thing [30:43] that you know in the in the blueprint we [30:45] we have you know proposals around like [30:48] modernizing the the tax base and around [30:50] like uh [30:52] we talk about maybe even a 32-hour work [30:54] week and and these types of things. I [30:56] think there's there's something um [30:59] important here about [31:01] uh [31:02] trying to create counter-cyclical [31:04] measures where you know conditional on [31:07] disruption from uh from AI we have kind [31:10] of [31:11] uh additional like unemployment [31:13] insurance uh we have these measures like [31:15] the [31:16] um [31:16] 32-hour work week but that we um [31:20] it's I think fairly important to me that [31:22] you know some of these measures [31:24] uh [31:24] maybe are maybe some of these measures [31:26] are good in the current world but I [31:27] think many of them would be quite [31:29] disruptive and we're really talking [31:30] about a world uh that uh [31:33] that that changes a lot and so like to [31:34] me [31:35] institutionally something that we need [31:37] is kind of um some thinking about [31:41] what are potential disruptions that [31:43] could occur from AI and [31:45] what are things that we can implement as [31:47] we as we see those coming [31:50] uh to kind of uh [31:52] counter the disruptions and uh [31:54] distribute benefits uh [31:56] broadly uh at that time. Um [31:58] >> Yeah and I saw one of the ideas in the [31:59] report was portable benefits since [32:01] benefits are so linked to employers in [32:03] America now. It sounded like there were [32:05] a lot of [32:05] >> Correct. Yeah yeah. And this was this [32:07] one I think an example of the one of the [32:08] more kind of uh US-focused uh proposals [32:11] of course uh but um [32:14] uh but yeah [32:14] >> it's a great idea. I think it's insane [32:16] we don't already have that and that like [32:17] the way that the US benefits world has [32:20] evolved is really bad but I think it's a [32:22] great idea. [32:23] >> No [32:23] no one should lose their health care if [32:24] they lose their job. Like that just [32:26] shouldn't happen. [32:27] >> Agreed. Um so Adrian from inside the [32:30] research organization you're seeing this [32:32] since acceleration it's real for you you [32:34] can see some some [32:35] um scientific um kind of [32:39] progress that's being made with these [32:40] models. [32:42] Um [32:42] what's [32:43] what's our We want institutions. We want [32:45] society to keep pace with technology. [32:47] What what do you think the window is for [32:48] adaptation? [32:50] >> So, you know, it's it's a [32:54] Well, I was going to say it's hard to [32:54] put a number. I would say that there's a [32:56] lot of uncertainty about these numbers. [32:58] Uh we've talked uh I think about [33:01] uh having an automated researcher in [33:03] 2028 or late 2028. Uh and I I want [33:07] >> 2028 is the official goal. [33:08] >> What's that? [33:08] >> March of 2028 is the [33:09] >> March, thank you. Um [33:11] uh and [33:13] I I think one useful thing to to think [33:15] about here is once you have this [33:17] automated researcher, which is an [33:19] automated like AI researcher capable of [33:20] doing AI research, [33:22] you potentially have kind of like a [33:23] double whammy of disruption, I would [33:25] say. Like, first of all, you have a [33:27] model that's capable of advanced [33:30] cognitive work, clearly, which which AI [33:32] research is. And so, that in itself is a [33:34] disruption. It may accelerate [33:37] uh further AI progress. And so, um [33:40] you know, I I can't tell you, you know, [33:42] after that point exactly how much [33:44] progress we'll have made like a year [33:46] from then, but it's probably, you know, [33:48] more than the the pace of progress that [33:50] we've been making so far. And so, uh [33:52] this is kind of the the type of window [33:54] that uh that we're talking about, I [33:55] mean, you. [33:56] >> Yeah, thank you. Okay, so we are at the [33:58] point in the conversation where we open [34:00] up questions. It's not just me anymore. [34:02] It's uh members of the community. So, [34:05] I've got one here. Um it's from Svetlana [34:08] Romanova. As AI becomes more capable, [34:11] what which human qualities do you think [34:14] will matter most in the future? I'd like [34:15] to ask each of you one by one, Josh. [34:18] >> Um you know, I think there are human [34:20] qualities that are timeless and that [34:22] will always matter. Uh character, [34:24] commitment, effort, [34:26] um compassion. [34:27] I I think they're going to matter a lot. [34:29] >> Mhm. [34:30] Yeah. [34:31] Sam. [34:31] >> All of those, creativity, understanding [34:33] what other people want, but I will share [34:34] a recent anecdote that really struck [34:36] home for me that things don't quite go [34:37] the same way. I went to my first robot [34:39] cafe [34:40] for the first like I was so excited to [34:42] try. I thought I was going to love it. [34:43] >> Mhm. [34:44] >> And it was the most underwhelming [34:46] experience. And I thought I was someone [34:48] that did not need like the barista at [34:50] Starbucks to smile at me and say hi and [34:51] ask how my day was going. I really [34:53] thought I was like didn't care about [34:54] that. It turns out that I really want [34:57] that and walking in to like push on the [34:59] screen [34:59] >> Mhm. [35:00] >> and have the robot do the thing and give [35:01] you a delicious cup of coffee [35:03] >> Yeah. [35:03] >> was like a deeply unfulfilling I don't [35:06] want this experience. [35:07] >> Yeah. [35:08] >> And I thought I wouldn't have cared. [35:09] >> I totally just those [35:11] small interactions throughout the day. I [35:12] I really appreciate them, too. Okay. [35:15] Here Here's one. Should those [35:17] >> is [35:17] >> Oh, that's fine. That's fine. [35:19] >> No, go for it. [35:19] >> Well, I I mean I think what I was going [35:20] to say is a little bit along the lines [35:22] of of what Sam said. And And I believe [35:23] it's in the blueprint like uh [35:26] humans [35:28] need each other. Like I think we care [35:30] about other humans. Um in fact, you [35:32] know, to some extent uh one of the most [35:35] uh [35:36] uh scary things with the development of [35:38] of technology and and social media, [35:41] video games, and stuff like that is like [35:42] maybe uh that is gotten us to lose a [35:46] little bit of the connection that that [35:47] we used to have. But I I I still think [35:49] that it's something that matters [35:50] tremendously to people and that they [35:52] will recognize this more and more in [35:54] fact uh [35:55] as as AI becomes uh becomes more [35:58] advanced and can do some of some of [36:00] these other tasks that don't require [36:01] human connection. And so uh that's like [36:04] one big thing that I believe will um [36:07] will expand. It seems like a good thing [36:09] to me, right? That like this is um this [36:12] can be of course like a type of work, [36:14] right? [36:16] You know, nursing and and uh and all [36:18] these these types of works in the care [36:20] economy, teaching. Um [36:23] and but you know, it's also just [36:25] important to to our lives, right? And I [36:27] think that's to me like going to be the [36:30] the big quality, like how good you are [36:31] as a person to other people. [36:33] >> Yes, I agree. Um okay, so in America [36:36] things like health care, child care, and [36:37] elder care are very expensive. Um here's [36:40] a question. How can AI expand access to [36:43] those for everyday people? Josh? [36:47] >> Um one thing I am so excited about is [36:50] the way that AI can help provide [36:53] the best uh health care in the world to [36:55] everybody. [36:56] Um I've heard a lot of stories from [36:58] folks navigating the health care system [36:59] that they didn't know where to go, um [37:02] how to navigate the insurance, [37:04] uh what specialist to talk to, and you [37:07] hear stories of people bouncing around [37:09] for a diagnosis for years and years and [37:11] not getting anywhere. You hear about [37:12] folks who are who are stuck in a health [37:14] care system where even if you have the [37:16] most caring, compassionate nurses and [37:17] doctors, there's not enough time to give [37:20] everyone the best quality care. And I [37:22] think that AI is going to make it [37:23] possible to deliver the best quality [37:25] care at scale. I don't think it's going [37:27] to replace doctors. I think it's going [37:29] to make their workloads manageable, and [37:31] I think it's going to make it possible [37:32] for patients to get the best experience [37:33] possible. [37:34] >> Yeah. I agree with that. A lot of [37:35] patient empowerment out there. So, yeah. [37:39] >> Uh [37:40] we have all told stories many times of [37:41] things we've seen on uh social media [37:43] about people having this like amazing [37:45] health care experience. Uh [37:47] I don't have anything that amazing, but [37:48] I'll tell my own because at least it [37:49] happened to me. Uh I recently got a [37:51] blood test. Uh there were like nothing [37:53] seriously wrong, but a few markers are [37:55] just like kind of out of their range, [37:57] and you know, you kind of scan down. [37:59] There's like 100 things, and you just [38:00] look at the list of new tests and the [38:02] ones that are like a little bit out of [38:02] range. I asked my doctor, he was like, [38:04] uh you know, probably those are all [38:06] close enough, it's fine. I put it in [38:07] ChatGPT, it said, yeah, you're fine, but [38:09] like here's what's going on, take this [38:10] one supplement, get your blood test [38:12] again in a month. [38:13] >> Mhm. [38:13] >> You should be okay. And it did. Again, I [38:15] wasn't like that sick, I wasn't but like [38:18] the fact that I could just like upload [38:20] my blood test and instantly get the [38:21] right answer. For like a kind of complex [38:23] thing. [38:24] >> Yeah. [38:25] >> Was an amazing experience. Uh and I [38:26] think there will be many things like [38:27] that across health care education and [38:29] all these other areas, elder care. You [38:31] know, that's one where I think we want [38:33] people doing that. [38:34] >> Very much. Yeah. [38:35] >> But there will be a lot of ways we can [38:37] really drive down the cost of [38:38] >> Yeah. [38:39] >> um health care and education and things [38:40] like that. [38:40] >> For sure. [38:41] >> And even personalized learning like it [38:43] can still be humans but you can figure [38:44] out how to teach individuals, right? Um [38:47] you know, how do you think about it? [38:48] >> Yeah, I mean, you know, I would say [38:50] similar things of course. I mean, [38:52] another basic thing in terms of health [38:54] care, you know, that I hope at least is [38:56] is simply that, you know, AI will be [38:59] able to help medical research and that, [39:01] you know, that that's like a basic thing [39:03] that that would make [39:05] health care better for people. But also, [39:07] you know, [39:08] I just talked about [39:09] the care economy and and I think to some [39:11] extent, you know, to the extent that AI [39:13] can [39:14] um [39:15] facilitate some of the of the more, you [39:18] know, bureaucratic, frankly, aspects of [39:20] of health care and and help us have more [39:24] people actually, you know, providing the [39:27] the actual [39:29] the actual health care [39:31] to [39:32] to individuals. That that seems like [39:34] like a a positive thing. Like it's not [39:36] AI would be providing [39:39] would be doing this. We would be freeing [39:40] up [39:41] people to to actually do this work. [39:44] >> Yeah. [39:45] >> We can hope. [39:45] >> We we've talked a lot about the [39:46] capability overhang this year. How it [39:49] AI has [39:50] capable of so many things that most [39:52] folks are not leaning into it for. And I [39:55] sometimes suspect that it's because [39:56] they're resigned to the world like it [39:58] is, right? I think PG calls it schlepp [40:00] blindness, right? So they're they're [40:02] they have ceased to consider that [40:04] something better is possible. Um [40:06] so I think they're going to start [40:08] leaning in more, right? Um [40:11] and and I think when they lean in that [40:14] their behavior is actually what will [40:15] change society as much as any [40:17] technological breakthrough. Just that [40:19] moment of adoption. Um as you in your [40:22] lives [40:23] what are the moments where you've seen [40:25] somebody kind of light up and realize, [40:26] oh I I can do this. I don't have to live [40:29] resigned anymore to the old ways? [40:33] >> Um wait, I've taken a lot of questions [40:35] first. Adrian, do you want to take a [40:36] crack at this? [40:37] >> Now you get time to think about it. [40:39] >> Oh my god. [40:40] Wow. Well, [40:42] >> Yeah, putting you on the spot. [40:42] >> Yeah, maybe I maybe I maybe I do want [40:44] time to think about this, but [40:46] you know, I mean to your point, I think [40:48] it [40:50] I think it's definitely the case that [40:53] uh that people take a while to adapt to [40:54] new technologies. I think to some [40:56] extent, you know, the the capability [40:59] overhang might be large in terms of [41:02] capabil- capabilities versus how much [41:04] people are actually using them. [41:05] Um [41:06] I think there is an extent to which [41:08] that's a function [41:10] of how fast the capabilities are [41:13] improving versus how fast people are [41:15] used to, you know, things getting uh [41:18] things getting better. [41:19] >> I have a favorite to mention. [41:21] >> Yeah. [41:21] >> My favorite There are many options here, [41:24] but my favorite of all is watching [41:25] parents who are coders by training watch [41:29] watching the parent watch their kids use [41:31] Codex [41:32] >> Mhm. [41:32] >> for the first time. [41:33] >> Mhm. [41:34] >> And uh a kid who has a bunch of ideas [41:37] and no idea about like what the [41:38] traditional limits are or what would be [41:39] hard or easy, just start describing a [41:41] video game [41:43] and having Codex make it. [41:44] Uh And the the kind of like creative [41:47] journey the kid goes through. I mean, [41:49] often you see the kid doing this mostly [41:50] by voice. [41:51] >> Yeah. [41:51] >> And the parent is just like, that's not [41:53] going to work, and then it works, and [41:54] then they're like, wow, my kid is going [41:55] to grow up in a world where he or she [41:58] just like expects this this happens. [42:00] Like and I kind of still like I wouldn't [42:02] even thought to try that cuz I would've [42:03] been so certain it didn't work. [42:04] >> Yeah. [42:05] >> So like watching [42:06] watching it through the parent's eyes [42:07] the kid do it for the first time, [42:08] especially if the parent is a software [42:10] engineer by training, is awesome. [42:12] >> Yeah, especially. I mean, this is a tale [42:13] as old as time, right? Like the kids [42:15] always know how to use the the VCR in [42:17] the '80s or whatever, right? [42:19] Uh and yeah. Interesting. [42:20] >> That's That's funny. [42:23] >> Um I just on the capability overhang and [42:25] how fast people notice when a capability [42:28] has arrived, there's an interesting time [42:30] scale mismatch where people have like a [42:32] who who aren't super in the know on AI [42:34] have this like distant awareness that [42:36] something is happening. They know [42:37] there's a product out there. Once every [42:39] few months, they might check it out. [42:40] They don't immediately and instinctively [42:42] probe it to the maximal extent of its [42:44] capabilities. [42:45] Uh and they often don't put it on the [42:47] thinking setting. Like they don't know [42:48] that reasoning models have happened. [42:49] They stay on the default chat model [42:51] that's that's right out there. And so [42:53] they wind up with this misperception [42:54] that things aren't moving as fast as [42:56] they are. And you hear people talk [42:57] about, well, there's hallucinations, [42:59] it's slop, it's making mistakes, it's [43:01] inaccurate. Why would they ever tell us [43:02] that it's going to do, you know, these [43:04] great things? And uh this like visceral [43:07] belief gap is I think an issue that will [43:09] get overcome when they start to see [43:11] other folks and institutions very [43:13] successfully use AI at sort of the [43:15] maximal reasoning settings, at the most [43:17] capable settings, in ways that are [43:20] shocking to them. Um like this video [43:22] game example, but sort of like at scale [43:23] for society as a whole. Uh they're going [43:26] to see a lot of people get diagnoses [43:28] that, you know, they wouldn't have [43:30] expected that someone could get quickly. [43:32] And it's going to update them. And I [43:33] think I think that's just like an [43:34] interesting phenomenon that the the time [43:36] scale for AI progress is weeks and [43:38] months. And the time scale for people to [43:39] currently checking back is like every [43:41] half year or something. And uh [43:44] yeah, some big change will happen when [43:45] people um realize the maximal extent of [43:47] capabilities today. [43:48] >> Yeah. Um I agree. [43:50] The uh the point you brought up, Sam, [43:52] about the kids really illustrates for me [43:55] what an advantage [43:57] creative people have. Like there are [43:59] some folks that are just kind of a font [44:01] of ideas. Many of those ideas have never [44:03] been realized. Some are scientists, some [44:04] are artists, many are children. Um and [44:08] it feels like the floodgates are opening [44:11] for them to realize more things. [44:14] I think you mentioned that you would [44:16] actually burn through your Codex list, [44:18] and you're not having it run all night [44:19] anymore. [44:21] >> We just need to make a model that helps [44:22] you come up with good new ideas. I [44:24] actually think this will be one of the [44:25] most exciting things to do. I don't [44:28] think we're that far away from a model [44:29] where you can say, "Go look through all [44:31] my text messages, all my email, look at [44:32] my entire computer, [44:35] anything you can find about me, and just [44:36] suggest ideas that I've like gestured at [44:39] as fragments or that might be [44:40] interesting to me." [44:41] >> Yeah. [44:41] >> And then I'll build those. [44:42] >> Yeah. Yeah, agreed. A thought partner, I [44:44] see a lot of people using it like that. [44:46] Well, I think we're wrapping up here. [44:47] So, thank you all for joining us today, [44:50] and thank you, Sam, Josh, Adrian. This [44:53] has been pretty cool. [44:55] >> This is a lot of fun. Thank you. [44:56] >> Yeah. [44:57] So, the ideas discussed may sound [45:00] ambitious [45:02] to you, and they're meant to be. [45:04] But, we know that they're also early and [45:06] exploratory. We're offering them [45:08] not as a final plan, but as a starting [45:11] point in a very public conversation [45:14] with policy makers and everyone in [45:16] society to encourage more discussion and [45:19] research and debate. OpenAI wants this [45:23] conversation to continue. [45:26] The company, we are inviting feedback [45:28] through [45:30] a new email address. It's [45:31] newindustrialpolicy.openai.com. [45:36] Please send us your best ideas. [45:38] We're launching a pilot program of [45:41] fellowships and focused research grants [45:44] for up to $100,000 in funding and up to [45:46] a million in API credits. And we're [45:48] convening further discussions at the new [45:51] OpenAI workshop opening in Washington, [45:54] D.C. in May. So, thank you again for [45:56] being part of this forum. We appreciate [45:58] the time and thought that everyone is [46:00] bringing to this conversation, and we [46:01] look forward to continuing it.