AI and Human Life: Transforming Healthcare, Work, and Wellbeing

AI与人类生活：变革医疗、工作与福祉 主持人 Lindsay Knight（Blue Star Families）与 Peter Yoo（Synchron）及 Travis Hoppe（CDC）共同探讨了人工智能在神经技术、公共卫生监测和社会安全网领域以人为本的实际应用。与会嘉宾探讨了血管内脑机接口如何帮助瘫痪患者重获自主能力、CDC 如何在遵循严格联邦数据治理要求的同时利用生成式 AI 减少繁琐的行政流程阻碍，以及社区自然语言处理（NLP）分流机制如何及时识别军人家庭面临的严重困境。整场对话强调了劳动者尊严与“人在回路”（human-in-the-loop）安全机制，展现了经过审慎治理的 AI 如何能够增进人际联结、减轻系统性负担并保护弱势群体。 Synchron 研发出血管内植入式脑机接口（BCI），通过微创导管经脑血管直接抵达运动皮层附近，结合深度学习解码发音运动意图，有效规避了传统开颅手术带来的感染与创伤风险。 (2:24) 美国疾控中心（CDC）将机器学习深度融入公共卫生运营，不仅使 PCR 实验室检测速度提升多达 20 倍，其率先制定的生成式 AI 治理准则还被联邦人事管理局（OPM）推广至整个联邦系统。 (5:28) 军人家庭支持机构 Blue Star Families 利用自然语言处理监测 13 万成员的社区情绪态势，并通过自动化分流机制在识别到急性精神危机或家庭暴力隐患时，立即转由持证临床社工人工介入。 (10:11) 在数据合规方面，非营利组织通过部署沙盒环境防止敏感项目数据流入公共模型训练，而脑机接口领域的原始神经电信号因无法被人眼直观识别，为患者提供了天然的生理隐私屏障。 (19:14)

Artificial intelligence delivers its most meaningful human impact not by displacing human professionals, but by parsing complex computational workflows and high-volume data to restore personal autonomy, safeguard public health, and rapidly surface acute distress. Restoring Physical Agency Through Endovascular Neural Interfaces (7:35) Peter Yoo of Synchron explains that brain-computer interfaces (BCIs) restore personal agency to paralyzed individuals by bypassing severed physiological pathways. Rather than relying on invasive open-skull craniotomies, Synchron guides sensors through blood vessels into the brain's vasculature adjacent to the motor cortex. Deep learning models decode latent motor intentions from these electrophysiological signals—reconstructing intended phonetic units rather than reading semantic thoughts—which downstream generative models amplify to assist with communication and digital navigation. Yoo stresses that evaluating these Class III medical devices relies not on algorithmic complexity, but on practical patient independence. Restoring unassisted communication and computer access for individuals paralyzed by ALS, stroke, or spinal trauma delivers immediate functional equality while substantially alleviating the ongoing physical and emotional strain experienced by family caregivers. Accelerating Epidemiological Surveillance and Outbreak Response (5:28) At the Centers for Disease Control and Prevention (CDC), Acting Chief AI Officer Travis Hoppe distinguishes standard administrative tasks from high-velocity programmatic analytics in public health. The agency uses machine learning to ingest electronic health records nationwide, monitor microbial signals across approximately 1,000 wastewater stations, and accelerate laboratory PCR processing speeds by up to twenty times. In environmental investigations, computer vision models such as YOLO scan aerial imagery of industrial cooling towers to locate Legionnaires' disease sources, while retail point-of-sale receipts from grocers like Giant are analyzed computationally to isolate contaminated food sources during enteric disease clusters. Hoppe emphasizes that automating 95% of data collation does not replace public health epidemiologists. Instead, automated pipelines eliminate manual auditing bottlenecks and empower specialists to act more swiftly during community health crises. In international surveillance, autonomous research agents have similarly reduced the time needed to synthesize global reception of the CDC’s Morbidity and Mortality Weekly Report from several days of manual translation to minutes. Digital Triage and Community Connection in Dislocated Populations (10:11) To counter the persistent geographic disruption of military families—who relocate every two to three years, with roughly 70% living off-base in civilian neighborhoods—Blue Star Families developed 'The Neighborhood,' an online platform uniting over 130,000 individuals. Lindsay Knight describes how internal natural language processing models continuously analyze forum discourse to generate aggregate wellness indices, tracking fluctuations in community anxiety or resilience following major societal shocks and military deployments. The system also operates as an automated crisis triage pipeline. Algorithmic severity classifications route mild financial stress toward vetted resources like USAA, direct systemic needs to the Department of Veterans Affairs, and escalate acute distress directly to licensed clinical social workers. Knight notes that maintaining human oversight remains critical: during a soft launch of an administrative chatbot intended for routine operational inquiries, a community member used the interface to report active domestic violence, enabling human staff to intervene immediately. Navigating Federal Data Silos and Organizational Guardrails (14:59) Deploying artificial intelligence within sensitive public health and social sectors requires stringent governance to prevent surveillance overreach and maintain constituent trust. Hoppe highlights that the CDC operates under 300 to 400 tailored Data Use Agreements and statutory restrictions like CIPSEA, which legally prohibit divisions such as the National Center for Health Statistics from sharing data across the agency without explicit inter-agency agreements. To safely utilize large language models, the CDC deployed a secure internal Azure OpenAI environment—logging over two million sessions with an estimated 500% return on investment—establishing governance frameworks that were subsequently adopted government-wide by the Office of Personnel Management. Organizations outside HIPAA and FDA oversight face different exposure risks, particularly unmonitored employee use of consumer AI. Knight explains that Blue Star Families addressed this vulnerability by creating an isolated, sandboxed digital library. This allows staff to synthesize research and draft grant proposals efficiently while ensuring proprietary organizational and member data is never ingested into commercial model training sets. The panelists conclude that artificial intelligence fulfills its purpose only when designed to uphold human dignity and measured against real-world user flourishing. Realizing these benefits requires institutions to maintain strict data firewalls, continuous feedback loops with affected populations, and dedicated human oversight to manage edge-case crises that automated interfaces inevitably encounter.

来自美国疾病控制与预防中心（CDC）、脑机接口企业Synchron以及非营利组织Blue Star Families的与会代表指出，人工智能正在流行病预警、重度瘫痪功能康复与特定群体心理支持等领域释放实质效能；但要真正增进人类福祉与自主性，技术部署必须紧密依托严格的数据治理防火墙、临床级安全标准以及不可或缺的人工介入兜底机制。 疾控监测与高通量公共卫生分析加速 (5:28) 美国疾控中心代理首席AI官Travis Hoppe将机构AI应用划分为前后两端：前端主要处理会议纪要、合同起草和常规公文等日常行政工作，经Azure OpenAI测算取得约500%的投资回报率；而在承担核心使命的后端，机器学习正深层重塑流行病学情报与实验室计算能力。目前，CDC借助算法将聚合酶链式反应（PCR）实验室检测流程提速高达20倍，并实现对全美约1000个废水监测站微生物信号的常态化追踪。在突发环境卫生事件中，YOLO计算机视觉模型被用于扫描卫星与航空图像中的冷却塔以定位退伍军人病菌源，系统还可通过计算解析Giant等连锁超市的购物小票快速排查食源性疾病载体。 此外，Hoppe介绍了CDC针对100名员工开展的OpenAI“Deep Research”自主智能体测试。虽然科学办公室评估认为该工具无法替代严谨的学术文献综述，但在监测《发病率和死亡率周报》（MMWR）的全球传播反响时，该智能体数分钟内便完成了以往需要专人耗时四至五天进行的多语种舆情检索与翻译，切实减轻了研究人员的繁琐机械负担。 血管介入脑机接口与运动意图解码 (7:35) Synchron神经科学与算法高级总监Peter Yoo阐述了植入式脑机接口（BCI）在重建重度瘫痪患者自主性方面的进展。为规避传统开颅手术引发的高危组织创伤与感染风险，Synchron借鉴常规心血管介入与卒中手术路径，通过导管将微型传感设备经血管送入紧邻运动皮层的脑血管网内。作为受FDA最严格监管的III类医疗器械，该技术直接继承了深部脑刺激（DBS）和人工耳蜗等前代神经技术数十年积累的安全性与合规经验。 在信号处理层面，深度学习模型能够从时空神经电信号中解码出预期的发音单元（音素）等运动指令，而非读取抽象语义思维。解码后的神经意图进一步与生成式AI的上下文预测功能结合，为肌萎缩侧索硬化症（ALS）或脑卒中导致的重度瘫痪患者提供类似“动态自动联想”的操作界面。Yoo强调，该技术不仅旨在帮助闭锁状态患者重新接入数字世界与自主表达，其实质效果更在于系统性缓解家庭照护者与临床支持网络的长期身心负荷。 社区心理健康监测与人工分级干预 (10:11) Blue Star Families首席影响官Lindsay Knight指出，全美约有60万个军人家庭每两至三年面临异地重新安置，其中约70%居住在军事基地外的平民社区，高频异地搬迁严重削弱了社会支持网络与心理韧性。为此，该组织依托覆盖13万名成员的数字平台“The Neighborhood”，开发了自然语言处理模型，持续将论坛交流内容转化为社区健康指标仪表盘，动态评估群体情绪并将其与地缘冲突或突发枪击事件等外部冲击相关联。 在个体危机筛查上，平台建立了严格的分级分流转介机制：轻度心理或财务压力引导至USAA等外部合作资源；系统性医疗健康诉求转介至退伍军人事务部（VA）；当算法检测到重度心理危机或即刻危险信号时，系统会立即触发自动化转向，直接交由内部执业临床社工（LCSW）实施实时人工介入。Knight披露，该机构在一款原用于处理表格查询等行政琐事的网站聊天机器人试运行期间，仅一周半内便接收到一名家庭暴力受害者的实时求助；正是由于保留了人工监控与升级通道，团队得以及时介入处置，印证了在自动化前端保留人工安全网的关键意义。 联邦法定数据壁垒与非营利机构影子AI治理 (14:59) 围绕数据隐私与安全监管，Travis Hoppe指出联邦机构受到严格的法定限制约束。CDC内部管理着300至400项独立的数据使用协议（DUAs），涵盖全美各州及领地的出生与死亡登记；下属国家卫生统计中心（NCHS）受《机密信息保护和统计效率法案》（CIPSEA）制约，即便在机构内部也严禁未经专项授权的数据跨部门共享。面对繁琐的FedRAMP认证与系统运营授权（ATO）门槛，CDC在2023年落地首个合规大模型应用并出台首份生成式AI指南，其框架随后被美国人事管理局（OPM）采纳并推广至整个联邦政府。而在更深层的技术防护上，差分隐私与同态加密等隐私增强技术（PETs）正被视为化解数据合规冲突的核心途径。 与联邦层面的强监管环境不同，非营利组织面临的突出风险在于员工未受监管的“影子AI”使用习惯。Knight指出，非营利组织通常不受HIPAA规范约束，员工常直接将敏感成员资料或资助调研喂入商用大语言模型以辅助撰写立项申请，极易造成受助者隐私外泄或被模型吸收训练。为此，Blue Star Families构建了严密隔离的沙箱化内部知识库与成员退出机制，确保员工在调用AI提升撰写效率的同时，机构专有数据与弱势群体记录绝对不会流向公有模型的训练流水线。 与会专家一致认为，衡量AI成效的根本标准不在于算法本身的复杂度，而在于能否切实恢复人类个体的尊严、自主性与社区联结；然而，跨部门数据孤岛的行政阻碍、隐私增强技术的庞大算力开销，以及确保自动化工具不演变为冷漠技术隔离墙的治理要求，依然是相关方案深化应用过程中悬而未决的现实挑战。

## Background 该视频由 YouTube 频道 Ai4 于 2025 年 8 月 29 日发布，记录了一场题为《AI 与人类生活：重塑医疗、工作与福祉》的专题研讨会。会议汇聚了 Blue Star Families 的 Lindsay Knight、脑机接口公司 Synchron 的 Peter Yoo 以及美国疾病控制与预防中心（CDC）的 Travis Hoppe，共同探讨人工智能在改善医疗结果、减轻职场压力、激发创造力以及应对隐私与心理健康风险方面的深远影响。 ## TL;DR 本次圆桌论坛系统探讨了 AI 如何在神经科技、公共卫生与社会服务三大领域重塑人类福祉。Synchron 分享了微创血管内脑机接口借助机器学习解码神经意图与言语信号，为瘫痪患者重建生活自主权并减轻照护者负担；CDC 阐述了在严苛隐私合规与 Azure 私有化部署下，AI 如何赋能海量病历解析、实验室提速并把心智带宽归还给公卫专家；Blue Star Families 则展示了利用自然语言处理构建社区健康监测大屏、实现危机干预与高净推荐值社交匹配的经验。专家们一致认为，真正的“人类繁荣”必须建立在以人为本、尊重个体尊严、严格保障数据隐私以及紧密结合用户反馈闭环的基础之上。 ## Atmosphere & style 整场对话洋溢着务实、睿智且富有同理心的氛围，在严肃的技术监管探讨与生动幽默的趣事分享间取得了绝佳平衡。嘉宾们打破了传统公职人员或硬核科学家的刻板印象，既有对 500% ROI 和官僚流程的机智调侃，也有对瘫痪患者重获尊严、意外拦截家暴危机等真实场景的深情流露。整场交流既没有盲目的技术乐观主义，也无虚妄的末日恐慌，呈现出一种极其接地气、充满人文温度且兼具严密逻辑的真实探讨感。 ## Key takeaways - [02:24](https://www.youtube.com/watch?v=rBih5T7YooY&t=144s) Synchron 采用微创血管内介入导管技术植入脑机接口，避免了传统开颅手术的创伤与感染风险，在维持高信号质量的同时降低临床难度。 - [05:28](https://www.youtube.com/watch?v=rBih5T7YooY&t=328s) CDC 将 AI 划分为前台行政运营与后台公卫业务，在电子健康档案解析、PCR 检测提速 20 倍、全美 1000 处污水监测及食源性疾病溯源等领域实现了大规模自动化处理。 - [07:35](https://www.youtube.com/watch?v=rBih5T7YooY&t=455s) 现代机器学习正推动脑机接口从简单的单次点击动作映射，升级到基于发音器官神经动力学特征的言语生成与音素解码，赋能失语与瘫痪患者自主交流。 - [10:11](https://www.youtube.com/watch?v=rBih5T7YooY&t=611s) Blue Star Families 借助自然语言处理构建社区健康看板，既在宏观上追踪 13 万军属群体的每周情绪波动，又在微观上分级预警并联动执业临床社工进行危机干预。 - [14:59](https://www.youtube.com/watch?v=rBih5T7YooY&t=899s) 联邦机构面临着极其严格的数据使用协议和 CIPSEA 等隐私隔离壁垒，CDC 率先引入 ChatGPT 并牵头制定了首份全美公共卫生生成式 AI 使用指南。 - [19:17](https://www.youtube.com/watch?v=rBih5T7YooY&t=1157s) 植入式脑机接口作为三类医疗器械必须接受严苛监管，当前脑电信号天然的不可直观读取性构成了短期安全屏障，但长期仍需构建前瞻性数据安全架构。 - [22:04](https://www.youtube.com/watch?v=rBih5T7YooY&t=1324s) 非营利组织等非强监管行业普遍存在员工私下使用外部商业 LLM 的影子 AI 隐患，机构需通过搭建封闭安全专属库和制定数据准则来杜绝敏感数据泄露。 - [29:05](https://www.youtube.com/watch?v=rBih5T7YooY&t=1745s) CDC 在 Azure 专属云环境部署私有化大模型并分析了 200 多万次对话，通过‘聊天机器人冠军’竞赛实现高达 500% 的 ROI，将心智带宽还给公卫专家。 - [30:59](https://www.youtube.com/watch?v=rBih5T7YooY&t=1859s) 低心理门槛的 AI 界面曾意外截获真实的家庭暴力求助，而 NeighborLink 社交匹配算法则以满分 NPS 帮助军人家属消除成年人社交壁垒、建立深度连接。 - [33:41](https://www.youtube.com/watch?v=rBih5T7YooY&t=2021s) 真正的‘人类繁荣’不仅在于技术本身，更在于将联邦层面的 AI 成果与规范反哺给全美 3000 多个基层公卫机构，并减轻瘫痪患者背后整个照护生态的精神重负。 - [39:38](https://www.youtube.com/watch?v=rBih5T7YooY&t=2378s) CDC 针对 OpenAI Deep Research 进行实测，发现其虽无法胜任严谨的学术文献综述，但在跨国跨语种的 MMWR 报告全球社媒与博客舆情汇总中将耗时从数天缩短至数分钟。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=rBih5T7YooY&t=4s) 破冰开场与个人欢愉时刻：AI 如何融入日常生活 座谈会由 Blue Star Families 的首席影响力官 Lindsay Knight 主持开场。面对会场各处普遍弥漫的沉重议题（例如世界末日、经济转型、工作岗位流失等），Lindsay 决定用一个轻快的破冰问题开启探讨人类福祉、生活与人工智能的讨论：“在你的日常生活中，哪一种 AI 使用方式曾为你带来纯粹的快乐，或提升了你的幸福感？” Lindsay 率先介绍了自己的背景与经历。作为 Blue Star Families 的首席影响力官，她统筹该机构的数字化创新、数据科学与分析、全美项目、线下分支办事处（Chapters and Outposts）以及政府事务工作。Blue Star Families 是全美最大的军人及退伍军人家庭服务组织，目前拥有约 39 万名会员，每年直接影响约 150 万人，其核心使命是帮助军人家庭在所居住的社区中蓬勃发展。 提到 AI 带来的快乐，Lindsay 分享了一则商学院案例分析的趣事。当时她需要就一个长达 20 年历史、极为枯燥的企业并购（M&A）案例进行推介陈述，为了调动团队情绪，她使用 ChatGPT 模仿特定人物的口吻撰写公开讲稿。在不断输入“让内容更离奇一些”或“情绪再亢奋夸张一点”的提示词后，ChatGPT 生成了一份极其荒诞幽默的演讲稿。尽管那场并购方案并未说服任何人，但整个课堂的听众都被逗得前仰后合、捧腹大笑。 - **Blue Star Families 组织规模**：全美拥有约 390,000 名会员，每年服务与影响约 1,500,000 人。 - **趣味应用场景**：使用大语言模型（LLM）进行特定角色风格化模拟与极端夸张的情绪调优，打破枯燥商务陈述的刻板氛围。 --- ### [02:24](https://www.youtube.com/watch?v=rBih5T7YooY&t=144s) 神经科技与脑机接口：Synchron 的血管内介入方案 来自 Synchron 的神经科学与算法高级总监 Peter Yoo 接着作了自我介绍并分享他的见解。Peter 拥有博士学位与学术界背景，他指出，由于某些公众人物（如马斯克与 Neuralink）的广泛宣传，脑机接口（BCI）领域已经从过去极其小众的研究方向一跃成为全球焦点。Synchron 的核心差异在于避免进行传统的开颅手术（Open brain surgery）。要在帮助瘫痪患者的同时采集高精度的神经信号，传统方法必须极度贴近大脑，这造成了一种“既要高质量信号、又要降低开颅风险”的两难困境。 ```mermaid flowchart TD accTitle: Synchron 血管内脑机接口技术路径 subgraph Traditional["传统侵入式 BCI"] T1["开颅手术 (Craniotomy)"] --> T2["直接接触脑皮层"] T2 --> T3["手术创伤高 / 感染风险大"] end subgraph SynchronPath["Synchron 血管内介入方案"] S1["微创导管介入技术 (Catheter-based)"] --> S2["经由血管系统进入颅内近脑区"] S2 --> S3["创伤小 / 安全性高 / 维持高信号质量"] end ``` Synchron 借鉴了成熟的心血管与中风治疗介入方案，采用微创导管技术（Endovascular approach），让植入物经由血管网络进入颅骨内部、贴近大脑区域，从而在不切开颅骨的前提下获取清晰的脑电信号，既降低了临床风险，又为瘫痪患者恢复交互能力提供了可靠的数据基础。谈及 AI 带来的“情绪价值”，学术出身的 Peter 开玩笑说，在科研界提出新点子往往会遭到同行无情的否定和批评，而 AI 却始终像一个不知疲倦的“拉拉队长”，随时对你的各种想法给予热情认可与鼓励。 --- ### [03:51](https://www.youtube.com/watch?v=rBih5T7YooY&t=231s) 公共卫生前沿：CDC 的 AI 实践与儿童创意的点燃 美国疾病控制与预防中心（CDC）代理首席人工智能官 Travis Hoppe 随后发言。Travis 在联邦政府拥有深厚的技术任职背景，曾在美国国家卫生研究院（NIH）工作了 12 至 13 年，在 CDC 工作了 5 年。早在 AI 与机器学习成为流行概念之前、在 AlphaFold 等蛋白质折叠预测斩获诺贝尔奖之前，他就长期致力于该领域的研究与推广。 关于 AI 带来的快乐，Travis 分享了一个关于儿童创造力的动人场景。他提到，小孩子的绘画通常线条稚嫩、技法粗糙，但如果让他们讲出心中的故事，再借助生成式 AI 将那些粗糙的涂鸦转化为他们脑海中真正想象的画面，孩子们眼中瞬间点燃的惊喜与创造力火花是无比美妙的。尽管当下关于 AI 对儿童潜在负面影响的探讨非常必要且真实，但 AI 在激发和放大人类原生创造力方面的正面价值同样不可忽视。 --- ### [05:28](https://www.youtube.com/watch?v=rBih5T7YooY&t=328s) 从前台协作到后台疾控：CDC 的全场景赋能与自动化跃升 在进入核心议题——“AI 正在如何重塑人类生活体验”时，Travis Hoppe 将 CDC 的 AI 应用场景系统性地划分为“前台（Front of the house）”与“后台（Back of the house）”两大板块： ```mermaid flowchart LR accTitle: CDC 前台与后台 AI 赋能架构 subgraph FrontOffice["前台业务运营 (Business Ops)"] F1["合同管理与审批"] F2["会议纪要与文档生成"] F3["日常行政提效"] end subgraph BackOffice["后台公共卫生核心业务 (Public Health Operations)"] B1["海量电子健康档案 (EHR) 解析"] B2["实验室 PCR 提速 (20倍加速)"] B3["全美约 1,000 处污水监测站分析"] B4["基于 YOLO 的冷却塔与军团病图像识别"] B5["超市零售收据分析 (食源性疾病溯源)"] end ``` - **前台业务运营**：涵盖生成式 AI 在商务运作中的常规提效，如辅助 CDC 员工撰写与审查合同、自动整理会议记录、处理日常行政事务等。 - **后台公共卫生业务**：作为全美最大的公共卫生机构，CDC 利用 AI 实现了以往受限于人力而无法完成的超大规模数据处理： - **电子健康档案（EHR）解析**：以前所未有的宏观规模自动结构化并解析海量病历数据。 - **实验室 PCR 检测**：通过算法与流程优化，使实验室分子检测速率提升了 20 倍。 - **污水流行病学监测**：全天候分析全美约 1,000 个污水监测站点的数据，极早捕捉传染病暴发苗头。 - **军团病（Legionnaires' disease）防控**：在疫情暴发时，利用计算机视觉模型（如经典的目标检测网络 YOLO）对建筑冷却塔航拍或巡检图像进行自动化排查。 - **食源性疾病溯源**：当出现食源性疾病疫情时，分析大型连锁超市（如 Giant Food）提供的海量购物小票，快速定位致病污染源。 - **人机协同理念**：这些业务过去高度依赖人工或半自动化工具，现在能够实现 95% 的自动化处理；AI 的定位并非取代疾控专家与研究人员，而是大幅增强（Augment）他们的业务边界与响应效率。 --- ### [07:35](https://www.youtube.com/watch?v=rBih5T7YooY&t=455s) 意图解码与行动重建：脑机接口中的机器学习跃迁 Peter Yoo 探讨了脑机接口领域如何利用机器学习从底层重塑瘫痪患者的生存体验。瘫痪患者通常由于神经通路受损（如中风、脊髓损伤或肌萎缩侧索硬化症 ALS）而丧失运动输出能力，但在许多病症中，大脑皮层产生运动“意图（Intention）”的原始神经信号依然完好无损。患者在大脑中构思移动手部或采取行动的主观意志，只是无法跨越受损神经传递到肢体末端。 现代机器学习算法的核心价值，正是在于从极其复杂的脑电时序动态中“解码（Decode）”这些深藏的意图信息。Peter 阐述了 BCI 技术的演进阶梯： 1. **基础运动意图映射**：早期技术仅能进行粗粒度的二元转换，例如将患者在脑海中移动手部的意图识别并转换为屏幕上的“单次点击（Click）”，这种方式虽然简单，但具备高度的鲁棒性。 2. **高阶精细化意图解码**：随着深度学习模式识别能力的突破，当前前沿已经推进到**言语生成解码（Decoding speech production）**。Peter 强调，这并非科幻作品中读取人脑深层语义（Semantics）的“读心术”，而是精确捕获并解码大脑在指挥发音器官时产生的**音素（Phonemes）**神经动力学特征。通过将这些复杂的脑电模式还原为语言输出，BCI 正在赋予失语与瘫痪患者前所未有的自主表达与现实世界交互能力。 ### [09:34](https://www.youtube.com/watch?v=rBih5T7YooY&t=574s) 脑机接口与 AI：让瘫痪患者重新表达的“伟大平衡器” Peter Yoo 阐述了 AI 工具如何作为脑机接口（BCI）的信号放大器。瘫痪患者的输出能力极其有限，部分患者甚至完全丧失了言语能力（non-verbal）。以往即使为他们提供基础的辅助技术，患者能表达出的内容依然非常匮乏。然而，引入动态 UI 以及“强力版自动补全”（autocomplete on steroids）等 AI 技术后，患者能够更快速、更丰富地输出想法。 从这个意义上讲，AI 正在成为这些重度瘫痪患者的“伟大平衡器”（the great equalizer）。Peter Yoo 表示，团队的最终目标是借助这项技术逐步缩小差距，希望有朝一日，植入了 BCI 的患者在日常交流与表达能力上与健全人“没有任何区别”。 --- ### [10:11](https://www.youtube.com/watch?v=rBih5T7YooY&t=611s) 军人家庭的流动困境与 AI 赋能的社区健康网络 Lindsay Knight 随后从非营利组织和以人为本的服务交付角度切入，分享了 AI 在两大维度的巨大潜力：**建立人际连接**，以及**提升资源获取效率并早期识别心理健康危机、财务困境和食物短缺**。为了让观众理解这些举措的必要性，她介绍了军人家庭面临的严峻现实：全美约有 60 万个军人家庭每两到三年就必须随军调动一次，且 70% 的家庭不再住在封闭的军事基地内，而是散居在地方平民社区中。由于频繁搬迁，他们一次又一次被迫脱离原有的社交网络，导致社会资本和人际联系断裂，从而大幅增加了心理健康、身体健康和财务状况恶化的风险。 为了应对这一问题，Blue Star Families 构建了全美规模最大的军人及退役军人家属数字社区——“The Neighborhood”（邻里社区），目前已汇聚了约 13 万名成员。组织内的数据科学团队（Lindsay 特别致谢了在现场的数据科学家 Moen）利用自然语言处理技术，对社区中海量的日常对话进行分析，搭建起一套“社区健康大屏”（Community Wellness Dashboard）。 ```mermaid flowchart TD accTitle: 军人家庭社区健康监测与干预流程 subgraph S1["数据输入层"] A["13万成员数字社区<br/>(烹饪、旅行、心理健康等讨论)"] end subgraph S2["AI 分析与检测层"] B["NLP 文本情感与问题语言检测模型"] A --> B B --> C["宏观聚合分析<br/>(周度群体情绪追踪 / 外部冲击映射)"] B --> D["微观个体风险分级<br/>(轻度 / 重度危机预警)"] end subgraph S3["分级响应与资源对接"] D --> E["退伍军人事务部 (VA)<br/>(合规信赖伙伴)"] D --> F["USAA 机构<br/>(财务援助与支持)"] D --> G["执业临床社工 (LCSW)<br/>(直接人工干预)"] end ``` - **宏观群体情绪监控**：模型每周对各类群聊（从美食、旅行到专业心理疗愈小组）进行聚合分析，判断整个社区的情绪基调（如“感恩”、“恐惧”等）。未来，团队计划将这些群体情绪波动与外部突发事件（如政治事件、部队调派、甚至近期发生在 Fort Stewart 的枪击案）进行关联映射，量化外界冲击对军人家属的影响。 - **微观个体危机干预**：如果成员发布了令人担忧的言论，系统内部看板会实时标记并评估严重程度（从轻度困扰到紧急危机）。针对不同等级，团队会触发相应的支持流程：对接退伍军人事务部（VA）等受信赖的机构、引导至 USAA 寻求紧急财务援助，或在面临严重身心危机时直接调度内部的执业临床社工（LCSW）介入人工支持。 - **隐私与伦理发问**：在肯定 AI 连接人与资源的能力之余，Lindsay 也向在场专家抛出核心疑问：当机构掌握了如此聚合甚至直达个体大脑内部的极度隐私数据时，该如何负责任地处理伦理、知情同意、隐私和数据安全等边界问题？ --- ### [14:59](https://www.youtube.com/watch?v=rBih5T7YooY&t=899s) 联邦机构的合规高墙与生成式 AI 的破冰探索 作为疾控中心（CDC）的首席 AI 官，Travis Hoppe 率先回应了数据安全与隐私合规的挑战。他指出，联邦政府在数据收集与处理方面受制于极其严苛的法律规范。与他此前在国立卫生研究院（NIH）主要处理偏向知识产权类的基因数据不同，CDC 掌握的数据受保护程度远超外界想象。 仅在 CDC 内部，就存在 300 到 400 份各异的“数据使用协议”（Data Use Agreements）。例如，CDC 与全美每个州和领地都签有个性化的出生与死亡完整数据协议；下属的国家卫生统计中心（NCHS）则遵循独立的 CIPSEA（机密信息保护和统计效率法案）法规运作。这些极高标准的网络安全与隐私隔离导致了一个棘手现状：NCHS 甚至无法合法地与 CDC 内部其他部门直接共享数据。这种严格的数据保护虽捍卫了公众隐私，却给机器学习和 AI 应用带来了巨大的合规壁垒。 - **合规认证的现实阻力**：每次在行业会议上听到厂商推销炫酷的“Agentic AI 工具”时，Travis 首先要考虑的是合规硬指标：“这通过 FedRAMP 认证了吗？我能拿到 ATO（系统运行授权）吗？”他感叹如果从业者不知道这些繁琐的政府合规术语，“那真是太幸福了”，因为在联邦体制内推进这些审批往往需要耗费数月甚至更久的精力。 - **率先引入 ChatGPT 的制度拓荒**：2023 年，Travis 顶住压力推动 CDC 成为全美首个引入 ChatGPT 的联邦机构。为了解决“联邦聊天记录是否属于联邦档案”、“是否可以被《信息自由法》（FOIA）调取”等前所未有的制度空白，他组织了约 200 人团队撰写了首份公共卫生领域的生成式 AI 使用指南。这份指南后来直接被美国人事管理局（OPM，即联邦政府的人事总管机构）全文照搬作为通用规范。 - **打破闭门决策，拥抱公开透明**：履新该职位仅四个月的 Travis 表示，尽管做“拓荒者”必须在缺乏先例的情况下艰难决策，但他希望一改联邦雇员较少公开发声的常态，主动走向外部行业大会，向公众透明化阐述政府在患者隐私、数据保护和 AI 落地方面的考量与实践。 ### [19:17](https://www.youtube.com/watch?v=rBih5T7YooY&t=1157s) 脑机接口的医疗监管与神经数据隐私挑战 Peter Yoo 从脑机接口（BCI）开发者的角度，探讨了作为三类医疗器械（Class 3 medical device）所面临的严格监管与安全隐私标准。植入式脑机设备虽然属于前沿领域，但神经技术本身并非新生事物，深部脑刺激（DBS）和人工耳蜗（Cochlear implants）等此前早已积累了采集神经生物物理信号的经验，甚至由于涉及脑部电刺激，承担着更高的安全风险。Synchron 吸纳了大量拥有医疗器械与神经技术背景的资深专家，与美国食品药品监督管理局（FDA）等监管机构紧密配合。Peter 坦言，作为科学家虽然习惯于从第一性原理思考，但在三类医疗器械领域，遵守安全与隐私法规必须被视为“公理”（axiom）。这种严格监管尽管在一定程度上放缓了创新节奏，但对于保护患者而言是完全合理且必要的。 针对神经数据的隐私保护，Peter 指出，现阶段神经数据具有天然的“安全屏障”，因为脑电信号极其难以解码，普通人即便截获也无法直接理解其含义，这与 CT、MRI 甚至人脸图像等直观的医疗影像存在本质不同。但从长远来看，团队必须具备前瞻性思维，将数据安全架构从一开始就做到位。 - **三类医疗器械定位**：植入式脑机接口受 FDA 等机构极其严苛的合规监管，安全与隐私是不可逾越的底线。 - **成熟经验借鉴**：依托 DBS 和人工耳蜗等过往高风险脑刺激技术的监管与工程沉淀。 - **神经信号的不可读性**：当前脑信号尚未像图像数据那样能被人类直观解析，构成了一种天然的短期安全保护层。 --- ### [22:04](https://www.youtube.com/watch?v=rBih5T7YooY&t=1324s) 隐私增强技术（PETs）与非营利组织的 AI 治理底线 Travis Hoppe 补充强调了隐私增强技术（Privacy-Enhancing Technologies, PETs）在公共数据与 AI 领域的重要性。他呼吁行业应更多关注多方安全计算（MPC）、同态加密（Homomorphic Encryption）以及差分隐私（Differential Privacy）等技术方案。尽管例如美国人口普查局（US Census Bureau）已经采用差分隐私发布数据，但这些技术在现实中往往“各有各的痛点”——要么计算成本极高，要么实现难度极大，需要多方利益相关者深度配合，但它们依然是履行法律与伦理数据保护责任的关键途径。 ```mermaid flowchart TD accTitle: 组织级 AI 数据合规与安全架构流程图 subgraph NonProfitRisk["非营利组织/传统机构面临的盲区"] StaffUsage["员工私下使用外部 LLM<br/>（ChatGPT / Claude）"] LeakRisk["敏感研究数据与项目提案<br/>泄露/被纳入外部模型训练"] end subgraph DefenseArchitecture["Blue Star Families 安全治理方案"] InternalPolicy["建立内部数据准则与退出机制（Opt-out）"] LockedLibrary["搭建独立、严格锁定的安全专属库"] SafeDrafting["受控开展提案与基金撰写工作"] end StaffUsage --> LeakRisk InternalPolicy --> LockedLibrary LockedLibrary --> SafeDrafting ``` 随后，Lindsay Knight 代表非营利组织 Blue Star Families 阐述了非监管行业的 AI 治理挑战。作为 501(c)(3) 非营利机构，他们虽然不受 HIPAA 等严苛医疗法案约束，但 Lindsay 警示所有处于非高度监管领域的机构：必须立刻建立内部 AI 使用规范。很多机构的管理层完全不清楚员工正在私下使用哪些 AI 工具、喂入了何种数据。如果员工将涉及成员隐私的敏感数据直接上传至公共大语言模型，后果不堪设想。为此，Blue Star Families 在启动相关项目前便设立了严格的数据使用准则、存储边界以及用户退出机制（Opt-out），并专门搭建了绝对封闭、高安全等级的内部专属库，确保在利用大模型辅助撰写项目提案和基金申请（Grant Writing）时，绝不让外部模型抓取其成员数据。 - **隐私增强技术（PETs）**：包括多方安全计算、同态加密和差分隐私，计算开销虽大但对伦理合规至关重要。 - **非营利领域的“影子 AI”隐患**：许多社会部门员工在无监管状态下将未公开的研究数据输入商业 LLM。 - **数据隔离与权限管控**：通过自建隔离的安全知识库，在利用生成式 AI 提升文案效率的同时杜绝数据泄露。 --- ### [25:46](https://www.youtube.com/watch?v=rBih5T7YooY&t=1546s) AI 对工作福祉的影响：从重建患者自主权到政府级安全应用 在讨论 AI 究竟是提升效率、减轻压力，还是可能破坏人类福祉时，Peter Yoo 分享了脑机接口在改善患者生活方面的独特视角。对于严重瘫痪的患者，BCI 的设计初衷是“成为其身体的一部分，然后默默退至幕后”，核心在于重新赋予他们因疾病而丧失的自主能力（Autonomy）。目前全球仅有 Synchron 与 Neuralink 等极少数团队拥有真实人类长期植入案例；Peter 强调，技术的核心价值不在于秀算法，而在于患者能重新自主玩游戏、与照护者发消息沟通，切实重获生活乐趣与尊严。 ```mermaid flowchart LR accTitle: CDC 内部安全 LLM 部署与遥测分析架构 subgraph UserInteraction["CDC 员工交互端"] Staff["内部员工<br/>（撰写邮件/合同/代码）"] CustomUI["自研交互前端<br/>（界面与 API 访问）"] end subgraph MonitoringCloud["微软 Azure 专属云环境"] AzureOpenAI["Azure OpenAI 实例<br/>（安全隔离）"] Telemetry["全量遥测与审计监控<br/>（分析 200 万次对话）"] end Staff --> CustomUI CustomUI --> AzureOpenAI CustomUI --> Telemetry ``` Travis Hoppe 则以美国疾病控制与预防中心（CDC）的实践为例，介绍了 AI 在公共卫生机构“前台”（面向员工的 LLM 工具）的落地情况。CDC 于 2023 年通过微软 Azure 部署了专属的 OpenAI 实例，并完全自主开发了前端交互界面与 API 接口。由于政府系统内部对员工操作具有完备的监管权限且无个人隐私预期，团队能够全面收集遥测数据（Telemetry），对累积的 200 多万次内部对话展开实际用例研究。分析发现，在安全可控的环境下，员工广泛利用大模型辅助撰写电子邮件、起草合同提案，甚至调用 API 自行编写业务代码，在保障公共安全的前提下切实激发了内部生产力与创新活力。 - **恢复自主权（Autonomy）**：医疗 BCI 通过无感嵌入，帮助瘫痪患者重建与外界交流及娱乐的能力。 - **CDC 的 LLM 落地规模**：基于 Azure 搭建私有化前端，支撑了超过 200 万次内部交互。 - **全流程遥测与安全 API**：既保障了政府数据的合规可审计，又为全员提供了自主开发安全 AI 工具的底层能力。 ### [29:05](https://www.youtube.com/watch?v=rBih5T7YooY&t=1745s) CDC 的“聊天机器人冠军”与把心智带宽还给公卫专家 谈及 AI 的投资回报率，Travis Hoppe 直言团队曾做过一次 ROI 对比，计算投入了多少资金、节省了多少工时，结果完全契合行业预期——达到了惊人的 500% ROI。但他幽默地调侃道，政客们最喜欢这种光鲜的数据，“这太棒了，我可以把报告扔给他们，好让他们别来打扰我们，让我们能专心去做那些真正重要的事情。”对 CDC 而言，他们真正关心的是人们如何运用技术做出真正有意思的探索，这关乎一种“变革理论”（Theory of Change）。 为此，CDC 发起了一场听起来略带俗气却极为务实的内部竞赛——“聊天机器人冠军”（Chatbot Champions）。他们号召员工提交自己最棒的实际应用案例，寻找那些能够“带来喜悦”或显著改善工作职能的巧妙用法。团队收到了大约 200 份方案，并且全靠人工逐一手动审阅，完全没有借助 AI 偷懒。随后，他们把这些优秀案例的提交者请到了 CDC 的双周实践社区交流会上做分享——该会议每两周举行一次，常态化拥有约 3,000 名参与者。交流中呈现出的核心价值在于：当常规杂务被自动化分担、效率得以提升后，这些拥有公共卫生硕士（MPH）学位的流行病学家和统计学家，终于能够专注于他们最初来到 CDC 接受严格专业训练所要从事的核心科研与卫生事业，而不必整天被繁琐的合同审批、工时单填写或应付上级要求的五行待办摘要所消耗。AI 归还了他们的心智容量，让他们成为更好的公共服务者。 - **投资回报率指标**：达到了让行政管理者和政客极为满意的 500% ROI。 - **竞赛反馈规模**：收集到约 200 份员工自发的实用 AI 创新方案，均由人工筛选甄别。 - **公卫实践社群**：CDC 内部搭建的双周实践社区吸引了多达 3,000 人持续参与。 --- ### [30:59](https://www.youtube.com/watch?v=rBih5T7YooY&t=1859s) 意外的危机干预通道与 NPS 满分的破圈连接 Lindsay Knight 从非营利组织 Blue Star Families 的实践切入，分享了两个截然不同的 AI 应用故事。第一个案例同样关于聊天机器人。该项目最初仅为了给内部 HR 和行政团队分流减负，处理一些枯燥但高频的基础咨询，例如机构的资金来源渠道、990 税表下载路径或董事会成员名单。然而，就在机器人刚上线进行灰度内测的第一周半，他们收到了极其意外的紧急输入：一位面临严重危机的军属正在机器人对话框中实时求助一起家庭暴力事件。该用户明知屏幕另一端不是真人，但在极度孤立无援的崩溃状态下，这个小小的输入框成了她能抓住的唯一求助门户。由于系统尚处于测试期、未完全无人化生产上线，后台团队第一时间实施了人工干预，并迅速将求助者转接给了团队的持证临床社工（LCSW），化解了一场致命危机。 第二个案例则是充满温情色彩的“邻里连接”（NeighborLink）项目。该程序旨在利用 AI 算法，将彼此互不相识但有着共同兴趣或话语体系的军人家属进行匹配。例如在“闺蜜节”（Galentine's Day）活动中，系统根据兴趣偏好撮合素未谋面的军属结伴喝下午茶。Lindsay 指出，对于成年人来说，主动结交新朋友往往伴随着极大的心理包袱、尴尬感和高昂的时间成本。该工具提供了一种毫无压力的极简路径，让用户只需输入姓名即可被匹配到志趣相投的茶友。该项目在用户群体中获得了满分 100 分的净推荐值（NPS），用极轻量级的算法切实击碎了社交孤独感。 ```mermaid flowchart TD accTitle: Blue Star Families AI 落地双轨模式 subgraph S1["行政工具变身危机应急接入点"] A["官网基础咨询机器人<br/>（原用于查询 990 表/董事会）"] --> B["军属输入家暴求助信息<br/>（将 AI 视为低心理门槛的安全入口）"] B --> C["人工监测即时介入"] C --> D["无缝转接持证临床社工 (LCSW)"] end subgraph S2["NeighborLink 算法连接社群"] E["输入个人意向与兴趣偏好"] --> F["AI 智能偏好撮合"] F --> G["线下破冰喝茶会面<br/>（消除成年人社交尴尬）"] G --> H["社群凝聚力提升<br/>（NPS 达到 100 分）"] end ``` - **意外的家暴求助**：测试上线仅 10 天左右即拦截到真实家暴警情，验证了 AI 界面对处于危机中个体的低心理门槛价值。 - **惊人的 NPS 表现**：NeighborLink 军属社交匹配功能在消费者调研中收获了高达 100 的 NPS 满分。 --- ### [33:41](https://www.youtube.com/watch?v=rBih5T7YooY&t=2021s) 重新定义“人类繁荣”：温迪汉堡、有尊严的公卫与基层赋能 随后 Lindsay 抛出了圆桌的核心大议题：随着 AI 愈发深入日常生活，我们该如何定义“人类的全面繁荣”（Human Flourishing）？在 AI 驱动的世界里，怎样的生活才算美好？技术研发又该如何始终锚定以人为本？ Travis Hoppe 首先接话，并兴奋地提及了前一天大会上温迪汉堡（Wendy's）的演讲案例。许多人初听会觉得在快餐得来速（Drive-thru）车道部署生成式 AI 简直是最不近人情、最机械冷酷的做法，但主讲人介绍的设计理念却处处彰显着以人为本的温度与对员工尊严的呵护。Travis 指出，CDC 对健康福祉与繁荣的界定，不仅是没有身体疾患或减轻精神负荷，更关乎人是否拥有内在的尊严、在社区中是否受到尊重以及广泛的健康社会决定因素（SDOH）。 谈及公共卫生的宏观愿景，Travis 将视线引向了全美各地的基层网络。CDC 虽是联邦层面最大的公卫机构，但全美还有超过 3,000 个基层公卫机构——涵盖各州、领地、地方及部落管辖区（STLTs）。这些机构常年在幕后默默护卫公众健康。Travis 强调，联邦层面的繁荣绝非独善其身，真正的繁荣是将联邦探索出的技术经验、谈判下沉的采购合同、制定的指南与实践社区全面反哺给这 3,000 多个基层辖区。他还透露，CDC 正筹备今年秋季在亚特兰大举办一场大型“AI Day”，邀请地方伙伴汇聚一堂，让 AI 真正赋能整个公卫生态。 - **温迪得来速范式**：将自动化落脚于减轻基层操作工人的劳动负担，而非简单粗暴地取代人力。 - **全美公卫协作网络**：全美有 3,000 余个州、领地、地方与部落（STLTs）卫生部门急需联邦层面的技术成果与标准共享。 - **亚特兰大 AI Day**：CDC 计划于秋季举行线下专题日与学术论文摘要征集，向基层公卫机构开放交流。 --- ### [36:59](https://www.youtube.com/watch?v=rBih5T7YooY&t=2219s) 自主权的涟漪效应与以最终用户为中心的反馈闭环 针对人类繁荣的提问，脑机接口公司 Synchron 的 Peter Yoo 显得底气十足。他笑称，对于脑机接口行业来说，如果他们把工作做扎实了，“人类繁荣”几乎是水到渠成的附带产物，因为这项技术生来就是为了恢复瘫痪患者与外界沟通的自主权。但他随即补充，繁荣的受益者绝不仅限于患者个体。任何接触过瘫痪病患的人都明白，患者背后依托着一整个庞大的支持系统——日夜操劳的家庭看护人员、临床医护团队以及社区辅助力量。当 AI 算法解码神经信号、将生活掌控权重新还给患者本人时，这份自主权的恢复会产生涟漪效应，极大地卸下家属与照护者的沉重精神枷锁。 ```mermaid flowchart LR accTitle: 神经技术赋予患者自主权所引发的人类繁荣涟漪效应 A["AI 神经信号解码与脑机接口"] --> B["恢复瘫痪患者对外表达与行动的自主权"] B --> C["患者生活质量与尊严显著提升"] C --> D["照护家庭与临床陪护的精力/精神重负得以减轻"] D --> E["整个支持网络实现系统性的人类福祉与繁荣"] ``` Lindsay 打趣 Peter 是她见过的唯一一个敢直接宣称“人类繁荣？我们早就搞定了”的技术专家，引来全场笑声。收尾时，Lindsay 从公共服务与终端交付的视角升华了讨论：人类繁荣的本质，是确保个体能够平等获取那些能让生活变好的资源与人际连接。但每个群体在不同人生阶段的需求千差万别，这就要求开发者必须在产品设计中深度嵌入来自终端用户的即时反馈机制，确保技术方案是围绕受众真实渴望的方式来打造。倘若脱离了真实的社区呼声与用户需求，即便研发出纸面上再惊艳的技术，也绝不可能真正带来人类的繁荣。 - **照护生态的系统减负**：脑机接口的技术外溢效应直接改善了围绕瘫痪患者的家属与医疗看护团队的整体生存状态。 - **闭环反馈原则**：技术研发绝不能闭门造车，唯有建立面向最终受众的需求校准机制，AI 产品才能产生积极的社会效能。 ### [38:55](https://www.youtube.com/watch?v=rBih5T7YooY&t=2335s) 以用户为核心的闭环设计与问答环节开启 在数字化工具与项目落地的过程中，确保技术能够真正融入目标人群的日常生活至关重要。如果一项方案在设计时脱离了实际生活方式，或者盲目预设所有人都具备良好的数字化接入条件（Digital Access），那么公众将完全没有使用该产品的意愿。因此，从项目立项、数字化流程搭建到每一项研究推进，团队始终坚持将“设计与反馈闭环”（Feedback Loop）深度植入到每一个环节中。在阐述完这一原则后，主持人和嘉宾查看了剩余时间，发现距离本场结束仅剩最后约三分钟，随即提前开放现场观众提问。 ### [39:38](https://www.youtube.com/watch?v=rBih5T7YooY&t=2378s) 员工聊天趣事与 OpenAI Deep Research 的公共卫生实战 现场观众随即抛出了一个轻松有趣的问题：在后台观察员工日常使用 AI 聊天机器人时，有没有发现什么搞笑甚至滑稽的用例？来自美国疾控中心（CDC）的首席 AI 官 Travis Hoppe 笑着婉拒了直接爆料员工具体提示词的要求，表示必须严格尊重团队成员的个人隐私。不过，他转而分享了 CDC 内部刚刚进行的一项关于 OpenAI 全新智能体工具“Deep Research”（深度研究）的试点测试。Deep Research 是一种能够自主进行表层网络检索的 Agent 工具，虽然无法绕过付费学术墙，但可以跨网站聚合公开信息并快速生成一份详尽的长篇研究报告（约 9 页）。为了评估该工具对公共卫生业务的真实价值，CDC 组织了 100 名工作人员开展实测，检验它究竟是具备实际效用还是言过其实。 ```mermaid flowchart TD accTitle: MMWR报告国际传播分析流程对比 subgraph 传统人工模式 A1["发布每周 MMWR 报告"] --> B1["人工逐一检索多国社媒与博客"] B1 --> C1["人工翻译法/西等多语种内容"] C1 --> D1["耗时 4-5 天梳理汇总情绪与反馈"] end subgraph Deep Research Agent 模式 A2["发布每周 MMWR 报告"] --> B2["Deep Research 智能体跨国多语言爬取"] B2 --> C2["自动跨语言翻译与信息提炼"] C2 --> D2["自动生成全球舆情综述报告"] end ``` 实测反馈呈现出鲜明的两极化：CDC 科学办公室（Office of Science）明确指出，该工具在学术严谨度上依然不过关，根本无法胜任严密的文献综述（Literature Review）工作，“在学术文献梳理上表现依然很糟糕（it still sucks）”。然而在其他此前意想不到的场景中，它却展现出了惊人的效率。例如，CDC 每周发布的《发病率和死亡率周报》（MMWR），每期在学术界能够收获多达约 700 次引用，其全球传播影响力极其庞大。以往为了解海外公众与机构对周报的看法，CDC 必须安排专职人员花费 4 到 5 天时间，逐一翻阅法国、西班牙、英国等国家的社交媒体与博客，手动进行多语种翻译并撰写舆情反馈报告——这是一项耗时繁琐且令人厌烦的苦差事。而引入 Deep Research 后，AI 能够自主浏览多国博客和社媒、完成跨语种翻译并迅速整合出完整的全球舆情报告，直接将数天的枯燥苦工转化为即时可用的情报。 - **MMWR 报告影响力**：CDC 每周发布的 MMWR 报告单期引用量高达约 700 次，在公共卫生学界具有极高的权威性与全球关注度。 - **效率跃升对比**：传统跨国社媒与博客的检索、翻译及报告整理需要 4 至 5 个工作日，而自动化研究 Agent 可以在短时间内一键生成多语种舆情综述。 - **圆满结语**：在分享完这一前沿技术在公共卫生监测中的实际落地后，嘉宾与主持人在掌声中向全场参会者致谢并宣布本次研讨圆满结束。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=rBih5T7YooY&t=4s) 在你的日常生活中，哪一种 AI 使用方式曾为你带来纯粹的快乐，或提升了你的幸福感？ — Lindsay Knight > [09:34](https://www.youtube.com/watch?v=rBih5T7YooY&t=574s) 伟大平衡器（the great equalizer） — Peter Yoo > [14:59](https://www.youtube.com/watch?v=rBih5T7YooY&t=899s) 这通过 FedRAMP 认证了吗？我能拿到 ATO（系统运行授权）吗？ — Travis Hoppe > [19:17](https://www.youtube.com/watch?v=rBih5T7YooY&t=1157s) 遵守安全与隐私法规必须被视为“公理”（axiom）。 — Peter Yoo > [25:46](https://www.youtube.com/watch?v=rBih5T7YooY&t=1546s) 成为其身体的一部分，然后默默退至幕后 — Peter Yoo > [29:05](https://www.youtube.com/watch?v=rBih5T7YooY&t=1745s) 这太棒了，我可以把报告扔给他们，好让他们别来打扰我们，让我们能专心去做那些真正重要的事情。 — Travis Hoppe > [33:41](https://www.youtube.com/watch?v=rBih5T7YooY&t=2021s) 随着 AI 愈发深入日常生活，我们该如何定义“人类的全面繁荣”（Human Flourishing）？在 AI 驱动的世界里，怎样的生活才算美好？ — Lindsay Knight > [36:59](https://www.youtube.com/watch?v=rBih5T7YooY&t=2219s) 人类繁荣？我们早就搞定了 — Lindsay Knight > [39:38](https://www.youtube.com/watch?v=rBih5T7YooY&t=2378s) 在学术文献梳理上表现依然很糟糕（it still sucks） — Travis Hoppe ## Mentioned Lindsay Knight, Peter Yoo, Travis Hoppe, Blue Star Families, Synchron, 美国疾病控制与预防中心 (CDC), 美国国家卫生研究院 (NIH), Neuralink, Elon Musk, AlphaFold, ChatGPT, OpenAI, OpenAI Deep Research, YOLO, Giant Food, The Neighborhood, NeighborLink, 退伍军人事务部 (VA), USAA, 美国食品药品监督管理局 (FDA), 美国人事管理局 (OPM), 国家卫生统计中心 (NCHS), 美国人口普查局 (US Census Bureau), 微软 Azure (Microsoft Azure), 温迪汉堡 (Wendy's), 发病率和死亡率周报 (MMWR), FedRAMP, CIPSEA ## Video description Lindsay Knight, Chief Impact Officer, Blue Star Families Peter Yoo, Senior Director, Neuroscience and Algorithms, Synchron Travis Hoppe, Chief AI Officer, Centers for Disease Control and Prevention Abstract This session will focus on how AI is transforming human lives, from healthcare and mental health to work-life balance, leisure, and overall wellbeing. Panelists will explore both the positive and negative effects of AI on daily life—how AI can enhance healthcare outcomes, reduce stress in the workplace, and create new avenues for creativity and leisure, while also addressing the potential risks to mental health, privacy, and happiness. The panel will dive into what the world may look like as AI becomes more integrated into human life, discussing the future of human flourishing in an AI-driven world. [Source](https://www.youtube.com/watch?v=rBih5T7YooY)

[0:04] Lindsay Knight: 在你的日常生活中，哪一种 AI 使用方式曾为你带来纯粹的快乐，或提升了你的幸福感？ [9:34] Peter Yoo: 伟大平衡器（the great equalizer） [14:59] Travis Hoppe: 这通过 FedRAMP 认证了吗？我能拿到 ATO（系统运行授权）吗？ [19:17] Peter Yoo: 遵守安全与隐私法规必须被视为“公理”（axiom）。 [25:46] Peter Yoo: 成为其身体的一部分，然后默默退至幕后 [29:05] Travis Hoppe: 这太棒了，我可以把报告扔给他们，好让他们别来打扰我们，让我们能专心去做那些真正重要的事情。 [33:41] Lindsay Knight: 随着 AI 愈发深入日常生活，我们该如何定义“人类的全面繁荣”（Human Flourishing）？在 AI 驱动的世界里，怎样的生活才算美好？ [36:59] Lindsay Knight: 人类繁荣？我们早就搞定了 [39:38] Travis Hoppe: 在学术文献梳理上表现依然很糟糕（it still sucks）

[0:04] Lindsay Knight: make it weirder or make it more hyped [2:24] Peter Yoo: thanks to a certain someone [3:51] Travis Hoppe: before it was cool [5:28] Travis Hoppe: 95% of the way [7:35] Peter Yoo: autocomplete on steroids [16:20] Travis Hoppe: Is it FedRAMP? Am I going to get an ATO [Authority to Operate]? And if you don't know what those things are, God bless you, because I will spend hours working on these things. [25:50] Travis Hoppe: Chatbot Champions [39:40] Travis Hoppe: Deep Research

[0:04] Hi everyone. Thanks for joining us [0:06] today. Um, as you heard, I'm Lindsay [0:08] Knight. I'm with Blueest Star Families. [0:09] Uh, I figured we'd kick it off with just [0:10] some introductions first and also [0:12] because we are talking about well-being, [0:15] uh, human life and AI. Kind of a soft [0:19] icebreakerish. I know there have been a [0:21] lot of heavy topics happening in a [0:23] number of the rooms here. You know, the [0:24] end of the world, uh, the changing face [0:26] of the economy, where will jobs go, etc. [0:29] Um, so I'm going to ask a question and [0:31] also answer it because I totally just [0:33] sprung this on my fellow panelists [0:34] roughly 30 seconds ago, but um, I'd like [0:37] you to introduce yourselves and then [0:39] say, what is one way in which you use AI [0:41] that either brings joy to your life or [0:44] enhances your well-being and I will give [0:46] you a second to think and I will tell [0:48] you all who I am. Um so as chief impact [0:51] officer of blueest star families uh I [0:53] oversee all of our um uh digital [0:56] innovation data science and analytics [0:58] work our national programs uh our [1:00] chapters and outposts those are our [1:02] physical locations across the United [1:04] States um and then I'll do a lot of work [1:06] with our government affairs team as [1:08] well. Uh we are the nation's largest [1:10] military and veteran family serving [1:12] organization. We have about 390,000 [1:15] members so far and we impact around 1.5 [1:18] million people per year and our entire [1:20] mission is to have military and veteran [1:22] families thrive in the communities they [1:23] call home. Uh the way I use AI that [1:27] brings joy to me. Uh quick just like [1:30] audience check. Who here has ever done a [1:32] business style case study at like a [1:34] business school or like with a group [1:36] with yeah some hands? They're super [1:38] riveting, right? Um I was recently doing [1:40] one uh around mergers and acquisitions [1:43] and the case study was like 20 years old [1:44] and like really really dry. Um and I [1:47] used chat GPT. I also had to like sell [1:50] the position that the group that I was [1:52] working with to a group and it's just [1:54] really hard to amp yourself up for that. [1:56] Has anyone used chat GPT to create a [1:59] public script in someone else's voice? [2:03] You should. It is a riot. And if it's [2:05] not ridiculous enough, you can be like, [2:07] "Chat GPT, make it weirder or make it [2:11] more hyped and you just get the most [2:12] bonkers things to to read." And I didn't [2:14] convince anyone of the of the M&A [2:16] project, but I had the entire class like [2:18] rolling on the floor dying laughing. So [2:20] that brought me joy and I think the rest [2:22] of the room, but I'm going to toss it [2:23] over to you, Peter. [2:24] >> Sure. Hi, my name is Peter. I work at [2:26] Synchron and we're creating a brain [2:28] computer interface. So thanks to a uh [2:32] certain someone this field has now gone [2:34] from the niche to effectively uh to the [2:37] spotlight. Uh but our main point of [2:39] difference is that we don't want to [2:40] perform open brain surgery. But the fact [2:43] of the matter is if you want to record [2:45] good brain signals to help our paralyzed [2:47] patients, you have to get closer to the [2:49] brain. So then it's a bit of a catch 22. [2:52] But our founder and kind of our approach [2:54] is um taking the indescular approach. So [2:57] maybe a lot of you in this room know [2:59] someone who's had a stroke or a cardiac [3:02] um disease and that usually gets treated [3:04] through catheter based minimally [3:06] invasive methods instead of having to [3:08] perform an open brain surgery or open [3:10] heart surgery. So we're just taking that [3:13] innovation directly into the brain so we [3:15] can get inside of the skull closer to [3:17] the brain without having to perform an [3:19] open brain surgery. So it's minimally [3:21] invasive, less risky for the patients [3:23] and we still get good data so that we [3:25] can give more back to the patients. Um [3:28] in terms of the the AI, I mean I would [3:31] have prepped the joke if I knew, but um [3:33] other other than the fact that um it [3:36] really I I originally come from [3:37] academia, so I'm a PhD. Um usually when [3:40] you have a good idea that I really think [3:42] it's great, everyone shuts it down. [3:44] That's just how it works. So it's great [3:45] to have a cheerleader. Sometimes it says [3:47] everything's a good idea. [3:48] [Music] [3:50] off you. [3:51] >> Uh, good morning everybody. I am Travis [3:53] Hoppy. I am the acting chief AI officer [3:56] for the Center for Disease Control and [3:58] Prevention. It's the CDC. Um, I have a [4:02] long history in the federal government. [4:04] I've worked in the National Institutes [4:06] of Health, NIH, for what, 12, 13 years. [4:10] Spent five years at the CDC. I have been [4:12] a longtime proponent of machine learning [4:14] and AI before it was cool. did protein [4:16] folding again before it was cool, before [4:18] it was a Nobel Prize. I am deeply [4:20] interested in this field. But also, I'm [4:23] noting that all of you here today are [4:25] here on the ethics and the human impact [4:28] of AI and I just want to congratulate [4:30] all of you for thinking about these [4:32] problems along with us on the stage [4:34] here. I think this is really really [4:35] important. Um I I have a really deep [4:38] technical background, but I think these [4:40] things are some of the fundamental [4:41] questions we should be asking. So again, [4:44] awesome that we're all here and we're [4:45] carrying the torch forward on this. And [4:47] to answer your question that you sprung [4:48] on us 30 seconds ago before we stepped [4:50] on stage, um there's lots of really cool [4:52] things you can do with generative AI. [4:53] There's lots of cool things you can do [4:55] to predict. But one of the like the [4:57] thing that brings me so much joy is to [4:58] find a young child and let them like [5:01] tell a story and let their imagination [5:03] go a little bit further or have them [5:04] draw a terrible picture. Kids drawing [5:07] kind of suck. They're terrible pictures. [5:09] and then have AI imagine that as what [5:11] they see in their head and watch the joy [5:13] that sparks from that jump in creation. [5:16] And I think like we talk a lot about um [5:18] the negative impacts on children and I [5:21] think they're very valid and very real, [5:22] but watching the positive ways it can [5:24] reinforce their own creativity is just [5:26] like a small spark of joy. Back to you. [5:28] >> It's awesome. Thank you. And it's also a [5:29] perfect segue for the first question uh [5:31] which is AI is clearly informing and [5:34] changing so many different elements of [5:37] how human life functions whether that's [5:39] work, education, how we communicate, how [5:41] we detect illness, um how children can [5:45] experience creativity. So from your [5:47] perspective, what's one specific way AI [5:50] is reshaping the human experience for [5:52] better or worse? [5:55] >> I can start. Um, I'm not going to answer [5:58] the the one. I'm going to give you many. [5:59] Um, I I I think we should break it up [6:01] and we'll have time to discuss this into [6:03] things that are at the CDC, front of the [6:06] house and back of the house. So, front [6:08] of the house is the all the cool stuff [6:09] we're hearing about all the bisops [6:11] things, right? Like helping CDC staff [6:13] work on contracts or meeting nodes or [6:15] all of these like cool generative AI [6:17] pieces. Those are absolutely important [6:19] efficiency gains. Love them. They're [6:21] great. Um, but then there's the back of [6:23] the house stuff, right? CDC is the [6:26] nation's largest public health agency. [6:28] So, we do a lot of really kind of what [6:31] you would say programmatic work that I [6:33] think is really fascinating and is [6:34] changing the way we operate as a federal [6:36] agency. We can parse uh EHRs, electronic [6:40] health records on a scale we couldn't do [6:42] before. Uh we could do samplings in our [6:44] labs. There's really cool advancements [6:46] in um you know PCR where you can do this [6:50] 20 times faster in our laboratories. You [6:52] can do wastewater detection. We analyze [6:54] about a thousand wastewater uh stations [6:56] across the United States and being able [6:57] to process those quickly and in ways we [6:59] couldn't do before. We look at cooling [7:02] towers uh for Legionnaire's disease when [7:04] there's an outbreak. Um and that's like [7:05] a perfect task for YOLO for those of you [7:07] who are like back in the day of doing [7:09] image recognition. And all of these [7:11] things are things we were doing before [7:13] either by hand or semi-automated and now [7:16] they can be automated 95% of the way. We [7:18] look at receipts for uh grocery stores. [7:21] If there's food born illnesses that come [7:23] out, um Giant will then give them to us [7:25] to to analyze. And having the ability to [7:28] do that at a scale faster doesn't [7:30] replace our staff, but it really [7:32] augments all the work that they can do. [7:35] >> Great. Well, I come from a slightly [7:38] different angle of how we utilize AI um [7:40] and brain computer interfaces. So we [7:43] primarily use AI or primarily the [7:45] machine learning methods that power the [7:47] current models but that also directly [7:49] translates into how we can decode um a [7:53] patient's intention to do something. So [7:55] that's effectively what a BCI's value is [7:58] to our paralyzed patients. So you could [8:01] imagine you can be paralyzed for a [8:02] variety of conditions and a lot of them [8:05] actually have um that a lot of the [8:07] patients who suffer from paralysis still [8:08] have the ability to create that original [8:11] intention which is effectively neural [8:13] signals. So they can still think about [8:15] moving their hands, think about doing [8:18] you know wanting to do something be an [8:19] agent in the world except the signal [8:21] just doesn't end up where it should go [8:24] because of the condition. So the primary [8:28] uh way that we use is to use these [8:30] latest advancements in uh machine [8:32] learning and other techniques and we [8:34] apply that into trying to decode the [8:36] information that is embedded in neural [8:38] signals. So um long and short all that [8:42] means for the patients is that we can [8:44] give more back to the patients so they [8:45] can express more of their intention to [8:47] move. So before we would to precrude [8:50] things like you know um move your hand [8:53] and we'll turn that into a click and [8:54] that's still very useful and that's [8:55] robust. But as a field we're starting to [8:59] see we can really do incredible things [9:01] like decoding speech production. So [9:04] don't freak out. This isn't the [9:05] semantics in your head. This is your [9:07] ability to generate phone names. And if [9:09] you think about it everything that we do [9:12] is embedded in our dynamics of brain [9:15] activity. So these these machine [9:18] learning models are very powerful at [9:19] picking up patterns. So we can apply [9:21] these techniques and give back to the [9:23] patients and they can really start to [9:25] output more of their intentions out in [9:27] the world. And second fold is that we [9:30] can obviously utilize all of these [9:32] amazing quote unquote more traditional [9:34] AI tools now to then amplify that output [9:36] as well. So you know make the UI dynamic [9:39] or autocomplete um on steroids for [9:42] example. So these these patients have [9:44] pretty limited output because they're [9:45] paralyzed. Some are even completely [9:47] non-verbal. So then we can output more, [9:49] enable more. So effectively AI becomes a [9:51] great equalizer for these patients who [9:53] could, you know, even even if you gave [9:56] back a little bit of assistive [9:57] technology back to the patients, the [9:58] outputs look pretty limited. But with [10:00] this technology, we can really start to [10:02] ramp it back up. So they're really [10:04] hopefully we get to a point where there [10:05] is no difference between you and I. Um, [10:07] if you have if you have a BCI. So [10:11] >> awesome. Thanks. I I I'll give one or I [10:14] think the two main ways from uh a [10:17] nonprofit and uh human centric service [10:21] delivery perspective that I have seen AI [10:24] both do really good work but also have [10:26] incredible potential. One is around [10:28] connection and one is around um [10:31] increasing resource access or detection [10:34] for um uh mental health crisis, [10:37] financial instability, food instability. [10:39] Those are sort of the two areas that um [10:41] Blueest Star Families is exploring uh [10:43] the use of AI within our organization. [10:46] Um I should probably preface what I will [10:50] tell you we are doing with some [10:52] background as to why it's even a need in [10:54] the first place. So for those of you who [10:56] don't know um military connected [10:58] families uh move every two to three [11:00] years. It's actually about 600,000 [11:02] families move every two to three years [11:03] because they are deployed to different [11:05] spaces. uh and 70% of those families no [11:08] longer live on base anymore, which means [11:10] they're embedded in civilian [11:12] communities. If you think about what we [11:14] know about um social connection, social [11:17] capital, access to resources, what that [11:20] means for um downstream effects around [11:24] socio emotional health, um physical [11:26] health, um financial well-being, the [11:29] more you are disconnected from a network [11:32] that you would have previously embedded [11:34] in, your odds for having negative [11:37] effects downstream um increase [11:39] dramatically. So if you then think of [11:41] this population that moves every two to [11:43] three years and experiences [11:44] disconnection over and over and over [11:46] again, that's sort of the case study and [11:49] demographic that we're digging into. So [11:51] one of the things that we're doing to [11:52] approach that um is uh we have uh [11:56] currently built the uh nation's largest [11:59] uh digital community called the [12:00] neighborhood for military uh and veteran [12:02] connected families. Has about 130,000 [12:05] members on it. And what we are doing in [12:07] there is actually taking all of the [12:09] language that they use in that space um [12:12] and creating a community wellness [12:14] dashboard on it. And actually I can see [12:15] our data scientist in the audience. So [12:17] shout out to Moen for the work that he [12:19] does in that. Uh it's kind of awesome. [12:21] Uh particularly within a nonprofit. But [12:23] what we can do on a weekly basis is both [12:26] at the aggregate population level as [12:28] well as at the individual level. At the [12:30] aggregate population level, we take all [12:32] of the information that is in community [12:34] chat groups um you know around it could [12:37] be around cooking or travel or it could [12:38] be you know in a group that's devoted to [12:41] mental and emotional health or therapy. [12:43] Um and we have or I should say we Men [12:47] has created a model that uh detects uh [12:50] language that is problematic or [12:53] potential triggers for the outcomes [12:55] we're looking at increasing. So, socio [12:57] emotional health and well-being, social [12:59] connection. Um, and at the aggregate, it [13:03] shows what our community's sentiment is [13:05] as a week. Is it is it gratitude? Is it [13:07] fear? Is it um pick one? Uh, and what we [13:10] want to do in the future is actually map [13:12] that aggregate community sentiment with [13:15] say exogenous shocks that are happening [13:17] in the world. Um, did something [13:19] political happen? Was there a [13:20] deployment? um what there was recently [13:23] um uh a shooting at Fort Stewart. We are [13:27] going to be looking at how that impacted [13:29] our community as as a whole, but we're [13:32] also treating it as an individualized [13:34] tool with our members. So if someone [13:36] says something concerning, we actually [13:38] have an internal dashboard that is built [13:40] up that flags those comments and then we [13:43] have a process internally like [13:46] to what degree is that comment [13:49] concerning? Is it mild? Is it severe and [13:51] what is the warranted response? Do we [13:53] connect them with a vetted trusted [13:54] partner like the VA? Um do we connect [13:57] them with um a USAA that provides [14:00] financial support and help to families? [14:02] Um or is it an immediate mental or [14:04] physical crisis in which case we can [14:06] divert a licensed clinical social worker [14:07] internally to have direct human contact [14:10] there. So I am just incredibly intrigued [14:13] and excited about the capabilities of AI [14:16] to both connect resources and connect [14:18] people more effectively. Um but that [14:21] also raises a lot of questions around [14:22] privacy and guardrails and how one does [14:24] that responsibly. Um and also the the [14:27] data that um you all are dealing with [14:29] obviously borders on those questions [14:31] sometimes in even massively more [14:33] invasive ways if you're talking about [14:34] how the brain actually works inside of a [14:36] person. So can you speak a little bit to [14:39] how you are handling [14:42] questions around ethics, consent, [14:45] privacy, data security, um when you have [14:48] this aggregate data, when you have some [14:50] of the most intimate details of a [14:51] person, how do you treat it responsibly [14:53] and make sure it is being deployed for [14:56] well-being? [14:59] >> Yeah, I can go first. Um, so we are the [15:01] federal government. We have tons of laws [15:05] that regulate how we collect data, how [15:07] we process data in ways that um you know [15:10] before I had joined the CDC, I worked at [15:12] the NIH, right? And we we had some [15:14] genetic data um but most of it is closer [15:17] to like intellectual property, right? [15:19] Like I didn't appreciate the extent to [15:21] which CDC and the data they have they go [15:24] to protect this. And it's not just one [15:27] source of data. There are literally [15:30] three to 400 data use agreements that [15:32] you know I had to go through this [15:33] process where we look through them. Uh [15:35] we have an agreement with every state [15:36] and territory for their their births and [15:38] deaths and that is individualized and [15:40] and complete. We have agreements that [15:43] are specific to some of our centers in [15:44] CDC like the national center for health [15:46] statistics follows a different law [15:48] called sipsy which governs how [15:50] statistical agencies collect data like [15:52] uh BLS or census. Um, and each one of [15:55] these requires very specific safeguards. [15:58] And I could bore you with all the [16:00] details. Um, but a lot of these details [16:03] are making sure that only the people who [16:06] need access to this data have access to [16:08] this data. Um, and this presents [16:11] problems and challenges for if you want [16:13] to do AI because I am the AI guy or [16:15] before that I was a data science guy at [16:17] the CDC. And to the extent that we [16:20] protect the information from like an IT [16:22] and cyber security side, the National [16:25] Center for Health Statistics, for [16:26] example, cannot share data with the rest [16:28] of CDC because they don't have the right [16:30] agreements in place. So there's not even [16:31] data sharing within the agency um to [16:34] protect the the data use. And that is a [16:36] great thing for data and privacy. It [16:38] does present challenges for machine [16:40] learning. Uh it presents a lot of [16:42] challenges for AI because we cannot just [16:44] use random cool off-the-shelf things. [16:46] Like it's so fun to sit at these [16:49] lectures and just to listen to all the [16:51] vendors be like we have this super cool [16:53] agentic tool. We've got this awesome [16:55] thing and I'm like is it fed ramped? Am [16:59] I going to get a atto? And if you don't [17:00] know what those things are, God bless [17:02] you because I will spend hours working [17:04] on these things. U and there's a lot of [17:07] cyber security concerns that we have to [17:09] go through just to get anything within [17:10] the federal government. Um, one of the [17:12] things that's been useful with this new [17:14] administration is that they are working [17:17] to lower some of those barriers while [17:18] still keeping the cyber security things [17:20] in place. It is really hard to bring in [17:23] things into the federal government. Uh, [17:25] I brought in chat GPT to CDC. We were [17:28] the first federal agency to have it in [17:30] 2023 and that was like six months of my [17:34] life. You know, as soon as it as soon as [17:35] it launched in November, I was like, [17:36] "This is a thing. How do we do this? How [17:38] do we do this safely? How do we do it [17:40] and protect the the chat logs? Um, but [17:43] there's like a ton of things you have to [17:44] consider like are the federal chat logs [17:47] are they federal records? Can they be [17:49] foyed? Nobody asked these questions [17:51] before, right? And these are like the [17:53] things as the federal government that [17:54] we're kind of grappling with and coming [17:56] up with. And there's no there's no laws [17:58] governing it. There's laws that [18:00] obliquely cover some of these things, [18:02] but objectively it's really hard to get [18:04] to things that aren't specific answers [18:06] yet. And so we are writing guidelines [18:08] and we are figuring these things out. We [18:10] wrote the first generative AI guidelines [18:13] for all of the federal well we wrote it [18:15] for CDC. I got about 200 people together [18:17] and we wrote the guidelines. The first [18:19] thing like what does chat GBT mean for [18:22] public health and we wrote this and it [18:24] became the basis for OPM which is the [18:26] office of personnel management. [18:28] Literally the HR for all of the federal [18:31] government. They just copied and pasted [18:33] our our guidelines. And so like if we're [18:35] going to do this, I want to do this [18:36] responsibly. Um but where when you [18:39] trailblite is you have to like make some [18:41] of these decisions yourself and it's [18:42] it's really hard. Um so one of the [18:44] things I want to do, I've only been in [18:46] this role for four months is be a lot [18:48] more open to the rest of the community [18:51] what we're doing as a federal [18:52] government. This is why I'm here. This [18:54] is why I asked probably have not seen a [18:56] lot of like federal employees speaking [18:58] publicly. uh I asked specifically to [19:00] start coming to conferences and say [19:03] answers to these things that we're doing [19:04] around patient privacy or data privacy [19:06] or protections. I think it's really [19:08] important that we we let you know the [19:11] things that we are doing. [19:13] >> Over to you. [19:14] >> Yeah, I mean it's actually really good [19:15] to hear that the struggles that I face [19:17] because obviously my group is tasked [19:20] with creating these neural decoding [19:21] algorithms and deploying them that you [19:23] know obviously not to your extent I'm [19:25] sure um uh it's there. So we are a [19:29] medical device and it's an implantable [19:31] which means more or less effectively you [19:34] become a class 3 medical device. So [19:36] while we're not a uh federal agency, we [19:39] are regulated by uh these regulatory [19:41] bodies, you know, the FDA in the US and [19:43] other groups if if we were to go [19:45] overseas. So we work closely with um FDA [19:49] and and other agencies that regulate us. [19:51] So you have to abide by all of these [19:53] rules. So effectively if you're a [19:55] medical device company doing these um [19:57] class 3 medical devices you take that as [19:59] an axiom. So me as a data scientist um I [20:02] mean I'm a scientist neuroscientist by [20:04] train I always want to think in first [20:05] principles but I chose to uh I chose to [20:09] engage in this line of work which is a [20:11] medical device which has variety of [20:14] safety and um privacy um things that are [20:18] attached to it and you take that as XM [20:20] and we have um lots of talented veterans [20:23] of this medical device companies who've [20:25] done neurochnologies before so [20:27] neurochnologies is not a new thing so [20:28] DBS clear implants. All of these [20:30] things have collected biohysical neuro [20:33] data beforehand to provide even therapy, [20:35] right? So, it's even a higher level of [20:38] um risk there because you actually [20:40] stimulate into the brain. So, we've [20:42] there's big companies who've done uh [20:44] played in those fields for a long time [20:46] and they have the knowhow and they've [20:48] gone through and almost grown with the [20:50] regulatory bodies with these [20:51] innovations. um we have those people now [20:53] working for us um guiding us and but at [20:56] the same time if you want to be a trail [20:58] trailblazer you also have to make this [21:01] technology work because ultimately at [21:03] the end of it you can't really fit a [21:06] brand new technology into pigeon hole [21:08] and make sure they run efficiently or to [21:10] the best of this capability. So it's [21:12] balancing those two things but [21:13] effectively when it comes to privacy and [21:15] risk it's you take it as an axiom and we [21:18] are regulated and there's no other ways [21:20] about it and that's you know slow some [21:22] innovations down but I think rightly so. [21:25] So when it comes to data you know the [21:27] security all placed in and at least for [21:29] the time being [21:32] neural data is very hard to decode let [21:34] me tell you so it's not like other data [21:37] sets where it's directly human [21:38] comprehendable so if an image leaks or [21:41] even a CT scan or an MRI scan you can [21:44] see someone's face brain signals are [21:46] very difficult to do that because our [21:48] brains aren't designed to interpret [21:50] brain signals we have all these things [21:51] that we you know output different [21:54] signals that you and I can understand. [21:55] So we're protected by there for now. Uh [21:58] but of course we're thinking way ahead. [22:00] Uh so doing things right and being [22:02] regulated is how we how we do. [22:04] >> Can I add one more thing that that I [22:07] think it's really interesting for this [22:08] this group and this field to consider um [22:11] is is pets privacy enhancing [22:13] technologies. Like we've heard this for [22:15] a while and I I think there are some [22:17] talks that obliquely cover it. I there [22:20] should be more. Um these are things like [22:23] um multi-party communication, homorphic [22:26] encryption, um you know, differential [22:28] privacy. If you want to look at census, [22:30] they did a whole thing on differential [22:31] privacy. That's how the census is [22:33] released now. Um super important for [22:35] protecting data and they all kind of [22:37] suck in their own unique way. whether [22:39] they're computationally expensive, [22:40] they're hard to implement, you need all [22:42] the stakeholders to be engaged. But that [22:44] is one of the ways that we can satisfy, [22:47] you know, the laws that we have to do, [22:48] but also the ethical burden that we have [22:51] to protect the privacy of of everybody's [22:54] data. [22:56] >> I will answer the the question around [22:58] guide or guardrails a little bit [22:59] differently. Um, Blueest Star Families [23:01] is a 501c3. We are a recognized [23:03] nonprofit, but we are not a federal [23:05] entity or institution, and we're also [23:06] not a medical device company. So, we're [23:08] not governed by HIPPA. We're not asking [23:11] to, but like the what I would say to [23:14] anyone in this room, particularly if you [23:16] are outside of highly regulated [23:18] industries or spaces, if you haven't [23:21] thought about what the guard rails or [23:23] transparency around your organization's [23:25] AI usage ought to be, you need to now [23:31] because we we know it's being used. We [23:33] know it's coming. not necessarily even [23:35] in the product you are developing within [23:37] an organization but your staff are using [23:40] AI and the vast majority of folks [23:44] particularly I would say in the um in [23:47] the nonprofit or social sector have [23:50] absolutely no idea how their staff or [23:52] employees are employing AI in what [23:54] capacities and using what data. So, if [23:58] you have data that could violate [24:00] someone's privacy and someone feeds it [24:02] into an LLM and you're just like, I'm [24:05] sure it'll be fine. Um, we as a sector, [24:08] and I say this for all nonprofits or [24:10] social sector organizations need to do a [24:13] lot better than that. So one of the [24:15] things that Blueest Star families did [24:17] before we started all of this work is [24:19] come up with internal guidelines around [24:21] how we are using data, what data, where [24:24] is it stored, how do our members find [24:26] out about that, what is the opt out [24:29] process for those individuals that don't [24:31] want to be included. Um and then really [24:34] thinking at an organizational structural [24:36] level, what are our boundaries for [24:39] sharing information with any of the [24:42] larger external language models? So for [24:44] example, and I'm sure there are some [24:46] other nonprofits in the room um or even [24:49] government sector folks, you do a lot of [24:51] proposal writing and you share a lot of [24:53] information about what your past [24:55] research has been and um where you want [24:57] to go. [25:00] chat GPT and Claude and all sorts of [25:03] things make those a lot easier to draft [25:05] in the first draft and um I'm sure we [25:06] have all been there. But when you're [25:08] talking about specific data points or [25:10] research studies that might not be [25:12] external and you feed it into something [25:14] that is, you don't know what problems [25:17] that is potentially causing down the [25:18] road. So we have actually set up um our [25:21] own library that is absolutely locked [25:24] down, cyber secure and only features the [25:26] information we want it to for proposal [25:28] and grant writing things like that. So [25:30] we can we can feed it in and be [25:32] absolutely guaranteed um that no one [25:34] else has access to that information and [25:35] no one else's models are being trained [25:37] off of the information that our members [25:38] are voluntarily giving to us. Um so [25:41] yeah, if you haven't thought about [25:43] frameworks, you all should do so folks. [25:46] Okay, next question. Um, [25:50] when it comes to work and well-being, [25:52] the promise of AI is increasing [25:53] efficiency, reducing stress. I question [25:56] that one though. Um, and even increasing [25:58] leisure time. Uh, but is that actually [26:01] happening? And how do we ensure AI [26:03] supports and not replaces or overwhelms [26:06] human thriving? [26:12] Um again I come from a niche area so I [26:15] don't I don't know how much this uh [26:17] particular question applies to our [26:18] patients but um at least in the narrow [26:22] context of how we're utilizing AI to [26:24] enable our patients ability to re-engage [26:25] with the world. This is the whole point. [26:28] We want to deploy this technology and [26:30] get out of their way. So I always say [26:33] when you design a BCI product it should [26:36] be thought of as part of their body, [26:38] right? It should be with them at all [26:40] times. It's always available. And it's [26:42] all about giving back or restoring the [26:44] autonomy that is lost due to their [26:46] conditions. So they once had it, now [26:47] it's lost because the conditions and [26:49] we're trying to give restore that back. [26:51] So um if we if we've done our jobs [26:55] right, then all of those three things [26:57] that you said is it really happening [26:59] should really be happening. And you're [27:00] starting to see glimpses of this. So now [27:02] because this technology is become coming [27:04] to the mainstream, you see examples of [27:06] it on YouTube. We released a few neural [27:08] link releases a lot. It's just currently [27:10] probably us two um companies um with [27:13] real human implants, people utilizing it [27:15] in the real world and you see the story [27:17] isn't about the machine learning models. [27:19] It's you know they're using and you know [27:21] playing games and you know talking to [27:24] their caregivers and right so that is [27:26] effectively we're trying to drive too. [27:28] Um I think the question becomes a lot [27:30] harder for maybe you too because of the [27:32] way in which you're using AI but for us [27:35] luckily for children for me it's it's a [27:36] little bit easier to answer. [27:39] >> So I mentioned earlier there's you know [27:41] the front of the house and back of the [27:43] house stuff and back of the house is all [27:44] of our programmatic work that you [27:45] traditionally think of in a public [27:47] health agency. Front of the house is you [27:50] know all the cool LLM stuff that's [27:52] happening right now. Um and you know we [27:54] deployed a chatbot in 2023 and that is [27:57] we did it through you know Azure put it [28:00] out there open AI so we had access to [28:02] open AI and we had to write our own [28:05] interface and if we're writing our own [28:06] interface we can watch everything that [28:08] goes through it and as government [28:10] employees there's no expectation of your [28:11] own privacy on your government systems [28:13] and so we could use all the telemetry we [28:16] wanted to figure out what were people [28:17] using our chatbot for right we had two [28:21] million chats already that were in [28:23] there. Um, and we could do studies on [28:26] what people were using it for. And like [28:28] I feel like there's not a lot of real [28:31] world use cases in like maybe some [28:33] businesses have published it like the [28:34] government of um, Pennsylvania did a [28:37] really great study when they explored [28:39] chat GPT within their organization, but [28:41] there's not a lot of like civic [28:43] engagement. And so we we did it. I like [28:45] I started this role. I was like, what [28:46] are we what are we doing? What are we [28:48] using it for? And it's the things you [28:49] expect, right? like helping with emails, [28:52] writing contract proposals because this [28:54] is a safe and secure environment. Uh [28:56] doing all these really cool things. We [28:58] have API access to all of our staff so [29:00] they can write their own code that [29:01] utilizes this in a safe way. There's [29:03] some like really cool innovative things [29:05] but like that was only a sampling. So [29:07] like we did a ROI comparison like how [29:10] much money do we put into it? How much [29:11] time are we saving? Hit industry [29:13] estimates. It's like cool 500% ROI. [29:16] Politicals love that. That's great. I [29:18] could give that to them and let let them [29:19] ignore us while we did all like the [29:21] really important stuff. Um, but we [29:24] wanted to know like what were the like [29:27] really interesting things people were [29:28] doing and this is like a theory of [29:30] change sort of thing. So we sounds [29:33] really hokey, but we had chatbot [29:34] champions here at CDC where like we had [29:37] a a competition submit your like best [29:40] use cases for things that you were [29:41] doing, things that you know sparked you [29:43] joy or improved your job's function and [29:46] we got something like 200 responses and [29:49] we had to go through all of them. We did [29:50] it by hand, not with AI about things [29:53] that like they like really clever uses [29:55] of this. And in all of these cases of of [29:59] ways people were cleverly using the [30:01] technology that we had like enabled onto [30:03] them, what what we found, you know, when [30:06] we talked to them and we brought them [30:07] and got them to speak at our community [30:09] of practice, which has like 3,000 people [30:10] that show up every other week to this is [30:13] that [30:15] when they were freed up with this kind [30:16] of routine task that they had to do, [30:18] when they had efficiency gains, they [30:20] were able to do the things that they [30:21] came to CDC that they were trained to [30:23] do, right? Like people come to public [30:25] health with masters of public health. [30:27] They are epidemiologists. They're [30:28] statisticians. They are like highly [30:30] trained professionals and you know it's [30:33] a big organization. So you have to do [30:35] like fill out this contract or this time [30:37] sheet or you have to write five bullets [30:39] because somebody told you to do so, [30:41] right? Like there are things that you [30:42] have to do that are required and like [30:44] all organizations have this and this [30:47] frees up some of that and gives back [30:50] that mental capacity to them and they're [30:53] they're better humans for it and they're [30:55] better public servants in the [30:57] organization that they're in. [30:59] >> I will answer it from two ways. Uh but [31:01] it all one of them actually also relates [31:03] to a chatbot. Um, so our chatbot was [31:06] actually designed to like help with [31:10] efficiencies in HR and like alleviate [31:12] time for our staff. So like I'm not [31:13] going to say boring, but like yeah, not [31:16] terribly groundbreaking, right? Um, and [31:18] what it actually ended up doing is being [31:21] a really interesting open portal or flag [31:25] from our members who were experiencing [31:27] crises at times. Uh, so we launched did [31:30] a soft launch of a chatbot on our [31:32] website. um figuring the questions are [31:35] going to be like who is your funding [31:36] from? Where can I find your 990s? Like [31:38] who is on your board? Just like things [31:40] that don't really questions that don't [31:42] have to take up staff time. Um and [31:44] within the first week and a half of the [31:46] soft launch of that bot, we had a real [31:48] time domestic violence uh incident where [31:50] someone was reporting it to our bot [31:52] knowing full well it wasn't a human [31:54] being, but it was a portal and a like an [31:57] access point for someone who was in a [31:59] deep state of crisis. And because we had [32:02] not productionalized it yet and weren't [32:04] relying solely on AI or detective um uh [32:10] uh language to to do its work, we had an [32:13] immediate intervention um and then [32:15] directed that person again to a licensed [32:17] clinical social worker on staff. Um so [32:19] that's one example where it's kind of [32:21] light and boring on the front end, but [32:23] the implications are actually massively [32:25] important for our members on the back [32:26] end. The other side is that we have um a [32:29] program called neighborlink which is [32:31] like super light and fun and like kind [32:33] of fluffy on the outside. It's meant to [32:35] pair using AI to pair individuals [32:38] together that don't know one another um [32:40] but might have something in common or [32:42] want to talk to one another and have [32:43] like a we did one for Gallentine's Day [32:45] um where it was like meet a military [32:47] spouse you've never met before and get [32:49] together for tea and you were like [32:50] paired along different affinity [32:51] interests. Um, but what we're, the [32:54] reason we're actually doing it is [32:55] because it takes a lot of intense human [32:58] work to meet someone new. It's also can [33:02] be really awkward and strange, [33:04] particularly as an adult, to make a new [33:05] friend. And we just want people to have [33:07] a super efficient way to like, I'm going [33:09] to pop my name in here, and I'd love to [33:10] have tea randomly with someone I don't [33:12] know. And our members love it. It has an [33:16] NPS score of 100, which I don't know if [33:19] any of you have done consumer research [33:20] before, but it's off the charts. Um, but [33:23] it's creating community and it's making [33:25] it more efficient for people to meet one [33:27] another and and find genuine connection. [33:30] Um, and I think that's kind of fun. Uh, [33:33] last question and then I think we have [33:34] time to to open it up to the group and [33:37] I'm hoping we'll also end on a positive [33:39] note. Um, [33:41] as AI becomes more integrated into daily [33:43] life, uh, how do we define human [33:45] flourishing? The second part of this [33:48] question is what does a good life look [33:49] like in an a in an AIdriven world and [33:52] how do we center human beings as this [33:54] technology is further developed? [33:57] >> Yeah, I can I can start with that one. [33:59] Um, [34:00] you know, I'll say on the onset of this, [34:04] um, first of all, did anybody see the [34:06] Wendy's presentation yesterday? It was [34:09] just fantastic. It was a guy who [34:11] developed uh the generative AI chatbot [34:14] that happens on the drive-thru at [34:15] Wendy's. Um it seems like the most [34:17] inhumane of all uses and it was the most [34:20] humane and the most human- centric talk [34:22] I have seen in AI. So like if you get a [34:26] chance to see it, his story, his journey [34:27] is incredible. And you know at the CDC [34:30] we define well-being and flourishing as [34:33] you know having uh being free of [34:35] physical ailments uh being free or [34:37] reduced uh mental burdens for sure. Um [34:41] but also having dignity right having [34:43] dignity in yourself and your place [34:45] within the community and all these other [34:47] things these social determinance of [34:48] health. These are lots of things that we [34:50] define this and that was the same way [34:52] that that our our friend at Wendy's [34:54] described how he designed a generative [34:56] AI chatbot to go through. Um, and so as [34:59] we think about AI, I just kind of want [35:00] to reframe it. Uh, like flourishing [35:03] involves dignity. And the way I've heard [35:05] both of you talk centers the use of your [35:08] AI around some aspect of humanity or [35:11] some aspect of dignity whether it's [35:13] bringing um connection back to your [35:16] physical form so that you can speak [35:17] outwards or its connection to the [35:19] community like that is like the core [35:21] pieces of this. Um, and one last thing [35:24] be before I close my piece out is I do [35:27] want to give a big shout out to the rest [35:29] of public health. So CDC is the largest [35:32] public health agency in the United [35:34] States, but there are 3,000 other public [35:36] health agencies. These are the stilts, [35:38] these are the states, the territories, [35:40] the local, the tribal uh jurisdictions [35:44] that exist and all of them serve you in [35:47] some way, right? And they probably serve [35:49] you invisibly. And one of the things [35:51] that I'm really really trying to push [35:53] for um and public health gives dignity, [35:57] gives flourishing. It's like one of the [35:59] core tenants of public health um is to [36:01] increase their ability to do all the [36:03] things that we're doing at the federal [36:04] level. If there are things that we can [36:06] give back to the stilts, uh is great. If [36:08] we can partner with large AI companies [36:11] and bring the guidance that we learned [36:13] back to uh the stilts, great. If we can [36:17] help negotiate some of the contracts, [36:19] even better. if we can have this [36:21] community of practice, this is what I [36:23] see as human flourishing on the federal [36:26] to state and local levels. And that to [36:27] me is just such an exciting venture that [36:30] we might have this opportunity to do so. [36:32] And I cannot promise a date for this [36:34] yet, but there will be in the fall an AI [36:37] day in Atlanta. So, keep on the lookout [36:39] for it. There might be abstract [36:41] submissions coming through, things might [36:43] be livereamed. Uh, but this again [36:46] contributes to building the sense of [36:47] community. I really want to get the [36:49] stilts, these these jurisdictional [36:51] partners out here in Atlanta to talk [36:54] about all the cool work that they're [36:56] doing in AI and how it brings back to [36:58] the larger community. [36:59] >> Awesome. Again, lucky for me, if we do [37:02] our jobs right, the human flourishing, I [37:05] think it just naturally comes because [37:06] that's effectively what we're trying to [37:08] do. But that's a that's looking at it [37:10] from the patient specific perspective. [37:12] But if any of you have worked with um [37:16] paralyzed patients before, there's a [37:18] whole community of people supporting [37:20] that individual. So there there are the [37:22] caregivers, there are the clinicians, [37:24] there are you know many other people who [37:27] um spend a lot of their time as well. So [37:29] improving the quality of life of that [37:30] particular individual and giving back [37:32] the autonomy to the patients actually [37:34] also trickles down to the families and [37:36] the caregivers as well. So holistically [37:39] um this is at least for me it's a [37:41] no-brainer in terms of human flourishing [37:43] just coming um as part of AI helping um [37:47] for us to achieve our goal of restoring [37:49] autonomy to patients. Um [37:52] and yeah it's slightly easy for me. [37:56] >> I love the fact that you are the only [37:58] person I've met that's been like oh [37:59] human flourishing no we got that unlock [38:01] it's great. No, I I absolutely love [38:04] that. Um, human flourishing um I I think [38:08] from a public sector and particularly a [38:10] a service delivery or um member-facing [38:14] organization, human flourishing means [38:16] that individuals have the resources and [38:19] access to the people and the things that [38:22] make their lives feel better. Um that [38:25] means a very different thing to very [38:27] different people at very different [38:28] points in time. So I I think making sure [38:32] that you have the feedback mechanisms [38:34] built in from your end user so you know [38:37] that the thing that you are creating [38:38] meets and addresses the needs that [38:40] actually exist and the solutions are [38:43] designed in a way that that person or [38:46] individual or community wants to use [38:49] them. Um, if those are disconnected, you [38:52] can create the greatest thing in the [38:53] world that a community or individual [38:55] will have absolutely no desire to use [38:59] because it either doesn't fit into their [39:00] lifestyle or it assumes digital access [39:03] or any of those things. The design and [39:06] the feedback um needs to be in a loop [39:09] and we we try to bake that in to every [39:12] program as well as digital process as [39:15] well as piece of research that we do at [39:16] the organization for that reason. Um, [39:19] yeah, I think we have about five [39:21] minutes. Oh, we have three minutes left. [39:23] Um, but I still I've still finished [39:25] ahead of schedule. Uh, any questions [39:26] from the audience? [39:30] [Music] [39:36] [Music] [39:38] >> I have one. [39:40] >> Go for it. If there is none, when you're [39:43] uh checking out what the employees were [39:46] using the chatbots for, were there [39:47] anything funny? [39:49] Comically funny, you know, not [39:52] >> Yeah, not concerning funny, just fun. [39:55] >> So, I'm not going to answer that [39:56] question. Um, respecting the privacy of [39:59] our individuals that we have. But I can [40:01] tell you, we just did a pilot on deep [40:03] research. So for those of you who don't [40:05] know, deep research is kind of a new [40:06] tool that OpenAI put out where it it's [40:09] an agent that goes out and will do [40:11] surface level web queries. So things [40:13] that don't require a payw wall, but give [40:15] you like a nine-page report of what's [40:16] happening. Um, and I got a hundred staff [40:19] together be like, what does this mean [40:21] for public health? Like is it useful? Is [40:23] it bogus? It still can't do a literature [40:25] review. Our office of science was like, [40:27] it's not there. It still sucks. But what [40:29] it can do is like things I never [40:31] imagined, right? Like we release MMWRs [40:34] which are uh these like weekly reports. [40:37] They have like 700 citations every time [40:39] they come out once a week which if [40:40] you're in academia is insane, right? Um [40:44] but we want to know how they're being [40:45] read in other countries. And it was [40:47] somebody's job to go on like Fran French [40:50] social media or Spain social media and [40:52] be like how are they being perceived [40:53] because our global nature in public [40:56] health we want to know. There's lots of [40:59] good reasons for that. And so deep [41:00] research can do this. it can go through [41:02] and like read what's happening on the [41:04] blogs in France, read what's happening [41:05] in Spain, read what's happening in the [41:07] UK, translate all the languages, give a [41:09] little report alto together. This was [41:10] something that took four or five days [41:13] every time one of these things came out. [41:14] It was somebody's whole job that they [41:17] hated, right? And like they didn't want [41:19] to do this and they would like skim it [41:20] and put it together and now we have this [41:22] like report that can just go right into [41:24] it. So to be able to turn that into like [41:26] a new technology, super super cool. [41:30] >> That's awesome. Um, thank you for [41:31] joining us on stage. Thanks to everyone [41:33] uh who joined us here today. And uh if [41:35] you have any questions that relate to [41:37] any of the specific work that we do, [41:38] please uh track us down. [41:41] >> Thanks. Thank you all.