Coast Guard and States Shift to AI for Defense, Ransomware Battles

海岸警卫队与美各州转向利用AI应对国防与勒索软件威胁 州立法规定AI平台须对非法深度伪造承担民事责任，而国防领导人警告称，未能采用AI副驾驶将使国家安全从业人员面临淘汰。 专家小组探讨了AI在联邦国防机构及州政府中的业务整合，重点介绍了海岸警卫队利用计算机视觉进行搜救和非法船只追踪等实际应用。发言者强调了联邦采购中存在的业务摩擦，并主张私营部门积极参与以及推广州级创新模式。此外，讨论还涉及勒索软件防护、特定领域AI幻觉、价值观闭环工程（values-in-loop engineering）以及确立非法AI深度伪造民事与刑事责任的新州立法等关键政策关切。 美国海岸警卫队正将地理空间分析、计算机视觉与大语言模型结合，通过比对 AIS 应答机信号与卫星图像识别非法海上转移，并利用 EO/IR 传感器自动化搜救海面目标，摆脱对肉眼运气的依赖。 (9:39) 专家指出基于客观物理数据训练的 AI 模型更可靠，若将模型用于情报预测或跨领域法律调查等主观场景，容易因带入人类偏见和数据偏差而产生严重幻觉。 (15:06) 阿肯色州放弃传统昂贵的商业网络保险，转为全州机构与公立学校自负盈亏的自保基金，并通过修改法规允许在勒索软件谈判期间召开非公开会议以防黑客抬价。 (18:46) 阿肯色州出台新规打击深度伪造，规定受害者不仅可以起诉恶意生成者，还可起诉缺乏安全护栏的 AI 图像软件平台，促使厂商在架构中内嵌价值规范。 (39:22) 与会代表建议中小企业不要死等五角大楼或联邦繁琐的 RFP 招标，而应主动为政府制定技术标准，并强调无论行政或战场决策都必须坚持“人类在回路中”。 (26:24)

Deploying artificial intelligence across defense, intelligence, and state governance delivers clear tactical advantages in objective physical domains, but operationalizing these tools requires navigating severe model misalignments, federal procurement bottlenecks, and evolving state-level legal liabilities. Objective Computer Vision Outperforms Subjective Decision Models (9:39) Former Coast Guard CDAO Brian Erickson highlights operational successes where computer vision and geospatial analysis process verifiable physical signatures. In maritime enforcement, algorithms cross-reference satellite imagery against Automated Identification System (AIS) transponder signals to expose illicit cargo transfers when transponders are spoofed or disabled. Similarly, cockpit computer vision models linked to electro-optical/infrared sensors can detect anomalies such as a human head in varied sea conditions, overcoming severe human search limitations experienced during search-and-rescue flights. Domain Misalignment and Subjective Data Ingestion Litigation partner Christopher Swift cautions that machine learning tools degrade significantly when shifted from verifiable physical data to subjective human contexts. Off-the-shelf legal language models trained on standard corporate contract disputes or financial irregularities frequently hallucinate when applied to national security investigations, such as tracing illicit arms shipments to North Korea. Swift attributes these failures to selecting on the dependent variable and embedding subjective human assumptions into algorithmic workflows. Overcoming Federal Procurement Inertia Through Proactive Contracting Defense Alliance director Chip Laingen advises commercial vendors to define operational requirements for government buyers rather than waiting for formal agency requests for proposals (RFPs). Non-traditional commercial suppliers carry higher risk tolerance and faster iteration cycles than legacy defense primes, prompting agencies like the Department of Homeland Security and the Defense Innovation Unit (DIU) to leverage Commercial Solutions Openings (CSOs). Navigating public procurement often relies on regional innovation clusters, though Laingen notes federal budget cuts reduced the SBA-backed network from 25 regional clusters to just 3. State Regulatory Labs, Cybersecurity, and Software Liability Arkansas State Representative Stephen Meeks presents state-level policy interventions designed to address technological vulnerabilities and legal accountability. After cyber incidents doubled under Governor Sarah Huckabee Sanders, Arkansas replaced an inefficient commercial cyber insurance policy—which cost $10 million over five years for only $1.5 million in net payouts—with an internal self-insurance fund requiring strict security adherence. To prevent public negotiation compromises during ransomware responses, the state amended open-meeting laws so oversight boards can deliberate active incidents in executive sessions. Corporate Legal Accountability and Values-In-Loop Engineering (39:22) Addressing societal risks from generative tools and hostile foreign influence campaigns, Meeks details state legislation introducing criminal penalties for producing illicit deepfakes alongside civil causes of action against image generation hosting platforms. Comparing unregulated tools to leaving unsecured weapons on street corners, the framework imposes financial liability on tech vendors that deploy models lacking guardrails. While states aim to automate public benefits and administrative services, Arkansas statutory guardrails require a human to serve as the final decision-maker. The panelists agree that AI functions most effectively as a human-teaming tool rather than an autonomous substitute across defense and state administration. However, unresolved tensions persist between the rapid iteration demanded by commercial software developers and the bureaucratic, ethical, and legal safeguards required by national security institutions and state lawmakers.

国防、情报与应急领域的专家与立法者指出，人工智能在国家安全与公共管理中的落地已从概念验证转向实际部署，但其实战效能取决于能否区分客观物理检测与主观情报研判，并依赖于突破传统采购官僚体制、革新网络安全防御机制以及确立平台安全与人类决策的法定底线。 客观目标与主观研判：AI的效能边界与模型幻觉 (9:39) 前美国海岸警卫队首席数据与人工智能官（CDAO）Brian Erickson与诉讼律师Christopher Swift指出，人工智能在国家安全中的应用成效取决于处理数据的本质。在处理客观物理信号时，AI展现出极高的可靠性。例如，海岸警卫队利用地理空间计算机视觉模型，对比强制性自动识别系统（AIS）应答器信号与卫星图像，能够精准识别关闭应答器走私货物的船只，并开发出机载光电/红外（EO/IR）模型，解决了直升机搜救中肉眼难以在海浪中发现落水者头部的致命难题。Swift指出，这类基于可验证科学数据的物理特征极易被模型掌握。 然而，当AI被引入涉及人类主观推断与裁量的业务、情报与执法流程时，风险便会急剧增加。Swift警告，若将主观假设直接嵌入算法（例如预测特定组织是否会发动恐怖袭击或研判敌军动向），极易产生偏差。其实务中发现，现成的商业法律语言模型在分析涉朝非法运输等国家安全调查的大规模邮件时频繁出现幻觉，原因在于训练数据集中于商业合同与金融违规，无法适配国家安全语境。他强调，若在复杂的社会互动领域中存在“依据结果变量选样”（选择性使用数据）的偏差，无论底层技术多先进，AI都会产生失真输出。 打破政府采购僵局：非传统供应商与区域创新生态 国防联盟（Defense Alliance）执行董事Chip Laingen与Erickson共同分析了前沿技术进入公共部门的体制障碍。Laingen直言，传统国防采购遵循自身独特逻辑，外部供应商往往花费数年等待根本不存在的正式需求或招标书（RFP）。他建议创新型企业停止被动等待，主动为政府界定需求，利用非传统供应商具备的高风险承受力、快速迭代能力与跨行业经验推动国防部采纳新技术。Erickson证实，军方内部项目办公室同样在向采购部门施压，迫切希望加速商用软件准入，目前正通过国防创新单元（DIU）及国土安全部的商业解决方案开放计划（CSO）等新型渠道引入海事技术与大语言模型。 在落地生态方面，阿肯色州众议员Stephen Meeks提出，相比加利福尼亚州、得克萨斯州或佛罗里达州，体量较小、常被忽视的州官僚流程较少，能为初创企业提供更敏捷的测试与部署环境。Laingen补充指出，虽然在政府效率部（DOGE）的联邦预算削减下，由小企业管理局（SBA）支持的区域创新集群从25个骤降至3个，但幸存的集群和APEX加速器等机构依然为企业提供对接联邦与州级资金的免费初期咨询，并在复杂采购阶段协助企业过渡至法律、合规与专门会计等专业付费服务。 州级网络安全防御重构：自保基金与信息公开豁免 阿肯色州众议员Stephen Meeks分享了该州在网络安全与数据治理方面的制度改革。在州长Sarah Huckabee Sanders上任后，该州遭遇的网络攻击呈倍数激增。鉴于防御方必须确保百分之百安全而攻击者仅需一次成功，阿肯色州废止了低效的商业网络保险模式：在过去的五年中，该州累计支付了1000万美元的商业保费，却由于200万美元的免赔额门槛，最终仅获得150万美元的赔付。 为此，阿肯色州设立了覆盖全州机构、市政当局与公立学校的自我维持型网络安全自保基金。参与实体在严格遵守由州首席网络安全官与地方代表制定的安全标准后，即可获得全额应急响应支持与恢复赔偿。此外，Meeks列举了北小石城学区的勒索软件真实案例：州安全专家曾成功将25万美元赎金压减至16万美元，但因该州原有的公开会议法要求公开发表表决，学委会公开批准了“最高25万美元”的支付授权，导致黑客在一小时内重新勒索全额25万美元。为修补这一制度漏洞，阿肯色州迅速修改法律，允许政府机构在处理活跃网络安全事件及讨论系统漏洞时进入闭门执行会议。 安全底线与法律责任：认知战防范与模型开发者连带追责 (34:17) 面对AI技术的失控风险，与会代表主张将安全价值直接写入系统架构，并通过法律设定硬性边界。针对外国敌对势力部署AI聊天机器人蓄意激化社会意识形态对立、甚至策划引发线下动乱的信息战威胁，Meeks介绍阿肯色州已立法将数字素养纳入公立学校课程，教育学生辨别合成内容与机器账号，并与其余31个州一同在公立学校课堂中禁止使用手机，以减轻算法对青少年心理健康的负面侵蚀。 在法律追责机制上，阿肯色州出台了针对非法深度伪造（deepfakes）的标志性法案。该法案不仅对违法制作深度伪造内容的行为设立刑事处罚（初犯构成轻罪，累犯判处强制监禁），还建立了民事诉讼权，允许受害者同时起诉不良行为人与未设置防护栏的模型托管平台。Meeks以枪支制造类比：若武器制造商把无防护的枪支与弹药随意丢弃在街头，就必须承担连带法律责任；同理，未嵌入安全防护措施的AI软件平台必须为其工具造成的危害承担民事责任。此外，该州立法明文规定，政府所有公共服务决策必须保留人类作为最终责任人，坚决杜绝“计算机决定”的官僚推诿。 该研讨表明，人工智能在国防与公共事务中的价值不在于替代人类，而在于赋能作战决策与行政效率；然而，技术的飞跃并未消除人的核心作用——正如Swift所言，“硬件与软件依然离不开脑力（wetware）”。如何在缺乏成熟联邦监管的环境下，通过地方立法权衡创新激励、平台法律连带责任与人机协作的最终决策权，仍是未来国家安全体系演进的核心未决议题。

## Background 本视频由 Ai4 于 2025 年 8 月 28 日发布在 YouTube 平台上。视频记录了一场题为“AI 在国防、情报与战备中的变革”的专题研讨会，聚焦人工智能如何加速国防决策、增强威胁检测与优化资源配置，并深入探讨了公共安全领域面临的伦理审查与监管合规挑战。 ## TL;DR 专题讨论汇聚了前海岸警卫队首席数据与 AI 官、区域国防联盟负责人、州众议员及合规律师，从一线战术、小企业创新、地方立法及法律合规四大视角审视了 AI 在国家安全生态中的落地现状。嘉宾们指出，AI 在海面搜救、地理空间监测等客观物理任务中展现了显著战术价值，但在涉及主观意图推断的情报研判中仍面临因变量偏差与幻觉风险。针对传统国防采办体系僵化繁冗的痛点，专家建议初创企业主动为军方定义技术路径，并利用轻量化商业创新通道破局。同时，各州正扮演监管“实验室”的角色，通过网络安全自保基金、深度伪造平台追责法案以及强化“人在回路”底线，在制度层面平衡技术赋能与公民权益保护。 ## Atmosphere & style 整场研讨氛围务实且极具洞见，退役军官之间以直升机飞行员本能的风险感相互打趣，合规律师则用“咖啡让人更快做错决定”这类辛辣幽默的隐喻直击技术滥用痛点。对话摒弃了泛泛的概念炒作，穿插着勒索谈判因公开会议破局、波涛中依据手表反光搜救落水者等极具画面感的一线实战案例。在直面官僚体制惯性与采购挫败感的同时，讨论始终透露出推动跨界破局与主动承担安全底线责任的务实行动主义基调。 ## Key takeaways - [10:13](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=613s) 海岸警卫队等传统军事机构在引入 AI 时面临深刻的文化阻力，传统采办往往优先考虑直升机和舰艇等实体装备，而非无形的软件与 AI 算法。 - [14:30](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=870s) 在客观物理世界（如海事卫星遥感、AIS 轨迹反欺骗及直升机搜救中的计算机视觉目标检测），AI 表现稳健且易于落地验证。 - [16:26](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=986s) 主观意图推断与情报研判极易受到人类先验假设和“因变量选择偏差”影响，将商业领域大模型直接跨界用于制裁合规或战情分析极易引发幻觉。 - [20:25](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=1225s) 阿肯色州设立了全州范围的网络安全自保基金，将资金赔付与各机构严格遵守安全标准相挂钩，以此建立自下而上的安全激励机制。 - [21:10](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=1270s) 在勒索软件与网络危机事件中，立法允许举行闭门会议至关重要，能有效防止在公开表决中泄露赎金底线与安全漏洞。 - [22:46](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=1366s) 科技企业向立法机构沟通政策时必须在休会期提前介入并建立信任，因为会期内议员每天审议数十项法案，根本无暇现场消化复杂技术。 - [26:23](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=1583s) 前沿 AI 技术已超越政府现有的标准规范，非传统供应商不应被动等待 RFP，而应主动为军方和政府定义应用场景与技术指标。 - [31:55](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=1915s) 中小型州在推进前沿技术试点时决策链条更短、更敏捷，初创企业可利用 APEX 加速器与区域创新集群等资源精准切入政企合作。 - [34:49](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=2089s) 外国敌对势力利用 AI 驱动的聊天机器人加剧社会撕裂，各州正通过限制校园手机使用和推广批判性“数字素养”教育来构筑认知防御底座。 - [40:14](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=2414s) 针对非法深度伪造的立法正向平台追责延伸，若软件提供方未设置必要的安全防护机制，将直接面临民事连带诉讼赔偿。 - [42:28](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=2548s) AI 在国防和公共服务中的核心定位是赋能人类而非取代人类，必须始终确立“人在回路中”（Human in the loop）的底线责任制。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=4s) 会议开场与现场行业背景调查 主持人奥利维亚·泽特（Olivia Zetter）为本次专题讨论拉开序幕。本次会议聚焦于人工智能（AI）如何深度重塑国防、情报与战备状态，探讨 AI 在哪些领域已经产生了切实的变革、阻碍其落地的核心壁垒，以及在技术应用和部署过程中涉及的政策、伦理与工程技术决策。随后，她介绍了四位嘉宾：美国海岸警卫队（U.S. Coast Guard）首任首席数据与人工智能官（CDAO）布赖恩·埃里克森（Brian Erickson）、国防联盟（Defense Alliance）执行董事奇普·莱因根（Chip Laingen）、阿肯色州众议员斯蒂芬·米克斯（Stephen Meeks）以及福励律师事务所（Foley & Lardner LLP）合伙人克里斯托弗·斯威夫特（Christopher Swift）。 在正式进入议题前，主持人与现场观众进行了快速的互动举手调查以了解受众构成。现场不仅有来自初创公司和小微企业的创业者，还有大量来自大型科技公司及传统国防主承包商（Defense Primes）的代表，同时也有一部分来自政府及公共服务部门的与会人员。 ```mermaid flowchart TD accTitle: AI 在国防安全领域的四大观察视角 subgraph Viewpoints["国防安全 AI 观察视角"] A["军事运营与实战<br/>(Brian Erickson)"] B["小企业创新与供应链<br/>(Chip Laingen)"] C["州级立法与公民安全<br/>(Stephen Meeks)"] D["法律合规与人机协作<br/>(Christopher Swift)"] end ``` --- ### [01:40](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=100s) 美国海岸警卫队的 AI 早期落地实践 布赖恩·埃里克森（Brian Erickson）首先分享了他在美国海岸警卫队服役 33 年的经历。他退役前是一名上校，职业生涯早期是一名直升机飞行员，也驾驶过固定翼飞机，并在工程物流领域深耕多年，最终成为海岸警卫队历史上首位首席数据与 AI 官，直至约一年半前离任。他幽默地提到，政府部门对参加外部会议的审批越来越严格，因此很高兴看到今天现场依然有政府同行克服困难参会。 在谈到 AI 在国家安全和国防中最具影响力的应用场景时，埃里克森指出海岸警卫队在多个方向取得了实质进展： - **地理空间分析（Geospatial Applications）**：利用 AI 处理空间与地理信息，辅助态势感知。 - **计算机视觉（Computer Vision）**：用于图像识别、海上目标监测与巡检。 - **大语言模型（LLMs）与智能搜索**：早期阶段主要用于知识检索与调研，帮助不同垂直业务部门的人员高效获取所需信息，从而更好地执行日常任务。 --- ### [03:10](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=190s) 小企业创新、战场决策与供应链韧性 奇普·莱因根（Chip Laingen）接续发言，他透露自己也是一名退役直升机飞行员，曾在海军服役 21 年（2004 年退役）。他开玩笑说，台上有两名直升机飞行员，“只要是开直升机的，脑子里总觉得下一秒坏事就要发生”，并顺便调侃了前一天演讲中人工智能先驱杰弗里·辛顿（Geoffrey Hinton）博士对 AI 潜在风险的警告。在过去 20 年中，他一直在运营美国小企业管理局（SBA）旗下的区域创新集群“国防联盟”，致力于帮助小企业获取国防合同。 针对国防领域的 AI 创新，莱因根从小企业的视角总结了三大关键赋能领域： - **加速深度研发与创新周期**：利用 AI 工具大幅缩短小微企业的基础科研与方案迭代时间。 - **提升战场情报分析与决策速度**：将 AI 技术推向战术边缘，加速战场海量情报的处理，使前线指挥官能够更快做出反应。 - **增强供应链弹性与准入效率**：优化供应链的可见性与抗风险能力，保障整体国防基础设施的韧性。 --- ### [04:54](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=294s) 州级立法视角：在科技创新与公民权益保护间寻找平衡 阿肯色州众议员斯蒂芬·米克斯（Stephen Meeks）带来了州政府与地方立法的独特视角。他幽默地表示：“给政客一支麦克风和一个观众群体，往往是件很冒险的事。”他提到自己刚参加了在波士顿举行的全美州议员大会（NCSL），这是美国历史上规模最大的一次州级议员聚会，而 AI 政策在讨论中占据了绝对的核心地位。 米克斯强调，全美各州在 AI 领域展开了激烈的产业招商竞争，但与此同时，州政府必须承担起保护本地公民的重任，特别是在防范深度伪造（Deepfakes）和应对社交媒体滥用方面。他特别提到，州级层面的努力成功阻止了联邦试图出台的 10 年 AI 监管冻结期（10-year Moratorium）。他认为州是“民主的实验室”（Laboratories of Democracy），通过各州因地制宜的立法探索，当各州逐渐达成共识时，才是联邦政府出台全国性统一法案的最佳时机。 --- ### [06:51](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=411s) 法律合规与“人件”瓶颈：AI 的咖啡隐喻 克里斯托弗·斯威夫特（Christopher Swift）从诉讼律师与合规顾问的角度切入。他抛出了一个生动的历史类比：在 15 世纪之前，欧洲人早晨通常喝啤酒；而随着咖啡从中东传入，到 16 世纪城市居民早晨改喝咖啡后，欧洲的国民生产总值翻了四倍。在斯威夫特看来，AI 与咖啡极其相似：**“它能让你以更高的强度、更快的速度做出同样的错误决策。”** 斯威夫特强调，无论硬件（Hardware）与软件（Software）多么先进，系统终究离不开“人件”（Wetware，即人类大脑与认知）： - **技术无法解决纯粹的人类问题**：希腊语中的“Technology”（Technologos）本质是指“诀窍/工具”，工具是用来达成目标的，但界定核心使命和目标依然是人类的专属任务。 - **避免生搬硬套（Stapling AI）**：成功企业是将 AI 深度整合进既有核心使命中进行赋能，而那些盲目打上 AI 标签、缺乏对自身业务认知就贸然推进的公司，往往会引发管理混乱，甚至在面对美国政府监管与合规要求时“摔跟头”（get over their skis）。 --- ### [09:39](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=579s) 战术与多领域环境中的 AI 落地挑战 主持人奥利维亚对“AI 犹如咖啡”的比喻表示赞同，笑称这比通常听到的“AI 相当于电力或互联网革命”更为生动。随后，她将焦点重新引回埃里克森在海岸警卫队的实战经历，询问在面对海陆空多领域、高复杂度的多任务环境下，究竟哪一项具体的 AI 应用对战术和日常运营产生了最直接的实际影响，埃里克森随即展开了关于联邦机构在落地过程中面临挑战的深入剖析。 ### [10:13](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=613s) 传统机构的文化阻力与海岸警卫队的 AI 早期应用 美国海岸警卫队前首席数据与人工智能官（CDAO）Brian Erickson 谈到，在联邦政府部署 AI 面临的最大阻力其实是“人员与文化”。组织往往被传统观念束缚，难以在所有业务领域迅速铺开新技术。让具备合适技能的人才进入组织的关键岗位，协助一线作战人员、联合作战部队（Joint Forces）以及情报界（Intelligence Community, IC）协同作业，是一项关键且耗时的工作。 ```mermaid flowchart TD accTitle: 海岸警卫队技术与文化转型挑战 subgraph S1["传统文化与资源分配"] A["235年历史的传统观念"] --> B["倾向采购直升机、巡逻艇等物理实体装备"] A --> C["无形 AI 技术难以感知价值，面临预算与文化阻力"] end subgraph S2["AI 赋能海上任务"] D["地理空间卫星图像"] --> F["多源数据融合与异常检测"] E["AIS 自动识别系统轨迹"] --> F F --> G["识别船舶欺骗与公海非法靠泊"] H["直升机 EO/IR 光电红外传感器"] --> I["计算机视觉海面漂浮物检测"] I --> J["引导搜救队伍精准定位落水人员"] end ``` 海岸警卫队是一个拥有 235 年历史的组织，是美国历史最悠久的连续海上服役力量（Brian 打趣称海军中间可能“分手断档”过一段时间）。要改变根深蒂固的传统思维极为不易——当面对看不见摸不着的 AI 技术时，管理层往往更倾向于把预算砸在直升机、新型巡逻艇（Cutter）等实物资产上。 尽管困难重重，海岸警卫队仍在地理空间目标识别上取得了突出成果。例如，根据海事规定，一定吨位以上的船只必须定期发送 AIS（自动识别系统）应答信号。针对公海上船舶关闭或伪造 AIS 信号的复杂欺骗（Spoofing）行为，他们利用运行在地理空间平台上的 AI，将卫星光学图像与 AIS 轨迹进行交叉比对，能够自动发现公海中两艘大船异常靠拢等潜在非法行径。 - 海岸警卫队具备 235 年连续海上服役历史，文化惯性使无形软件/AI 的采办优先级落后于实体载具。 - 远洋大船每隔 15 秒至 10 分钟发送一次 AIS 信号，AI 结合卫星遥感可有效破除欺骗手段、识别可疑活动。 --- ### [12:53](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=773s) 泰碧岛海上搜救案例与计算机视觉搜救模型 Brian Erickson 分享了一起在佐治亚州泰碧岛（Tybee Island）附近的搜救亲历：一名父亲在离岸游泳时被暗流卷走，岸上的孩子报警后，搜救直升机在 45 分钟内赶到现场。搜救行动持续了整整 3 个小时，出动了两架直升机，直到夕阳西下依然一无所获。这名父亲水性极好，生还概率很高，但海面搜寻宛如大海捞针。 就在太阳即将落入地平线之际，漂在水中的父亲拼命挥手，其手表的表盘恰好反射了一缕夕阳强光，被机组人员敏锐捕捉到。获救后男子表示直升机从他头顶飞过了不下 20 次。在不同海况（Sea States）的波浪起伏中，肉眼很难看清浮在海面的头颅大小目标。海岸警卫队为此研发了集成在直升机光电/红外（EO/IR）传感器上的计算机视觉模型，能够在特定海况下自动检测异常漂浮目标并提示机组，将搜救从“全凭运气”转变为智能化、高效率的探测。 --- ### [14:30](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=870s) 动态威胁环境下的 AI 建模：客观追踪与主观推断的本质分歧 主持人 Olivia Zetter 提出，面对物理、动能（Kinetic）以及网络等复杂多变的威胁形态，模型开发者应如何设计出适应一线动态需求的工具。 ```mermaid flowchart LR accTitle: 客观实体追踪与主观意图推断的区别 subgraph Sub1["客观物理世界 (易量化/低幻觉)"] A["AIS 信号轨迹"] B["海面人头/漂浮物"] C["物理特征检测与跟踪"] end subgraph Sub2["主观社会交互 (高复杂/易偏见与幻觉)"] D["恐怖袭击概率预测"] E["跨国毒品网络判定"] F["商业纠纷模型错用于跨国走私调查"] end Sub1 --> G["数学/物理验证准确"] Sub2 --> H["易陷入因变量选择偏差"] ``` Foley & Lardner LLP 合伙人 Christopher Swift 指出，训练 AI 的核心问题在于“你在教它什么”以及“模型内置了哪些底层假设”。他将 AI 任务划分为两类截然不同的范畴： 1. **客观物理层面的识别与追踪**：例如追踪走私俄罗斯原油并伪造 AIS 的船只，或是搜寻海面上类似人头大小的物体。这属于有明确科学物理基础、可用客观数值验证的范畴，模型容易训练且结果可靠。 2. **主观流程与人际交互的推断**：一旦将 AI 应用于商业管理、执法或情报研判（例如预测某个组织是否会发起恐怖袭击、是否为毒品犯罪集团、敌国是否会在特定战区部署特定武器系统），便进入了主观判断领域。人类在训练和微调（Acculturate）模型时，会不自觉地注入自身的预期和先验假设。 Christopher 举了法律实务中的典型反例：律师团队常需要调取海量内部邮件进行事实调查，但市面上的法律大模型大多基于“商业纠纷”或“财务违规”语料训练。当律师试图用它调查“某批货物是否违规运往朝鲜”时，大模型便频繁出现严重的幻觉（Hallucinations）。他强调，这在政治学上被称为“因变量选择偏差”（Selecting on the dependent variable，即选择性挑拣数据），若将片面假设带入涉及人类行为的复杂领域，模型再先进也难免失真。 - 客观物理识别（AIS 轨迹、红外光斑）可验证性高，AI 落地相对容易。 - 主观意图研判（情报研判、制裁合规）极度依赖上下文与假设，通用领域微调模型直接跨界使用极易产生幻觉。 --- ### [18:46](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=1126s) 阿肯色州政府层面的 AI 规范与网络安全实战 阿肯色州众议员 Stephen Meeks 介绍了州政府层面的 AI 治理与网络防御现状。该州在最近的立法会议中通过法案，要求所有州属机构制定明确的 AI 使用政策（AI Use Policy）。在此之前约四年，该州曾要求各部门出台技术使用政策，甚至具体到“政府雇员午休时间是否允许在公用电脑上刷 Facebook”。 Stephen 提到，自州长 Sarah Huckabee Sanders 当选以来，阿肯色州遭遇的网络攻击次数在一夜之间翻倍。防御方必须做到“100% 成功防御”，而攻击者“只需得手一次”。在推进网络安全治理时，单纯下达行政命令（Mandates）往往会遭到各部门的抵触。 过去 5 年间，阿肯色州在网络安全保险（Cyber Insurance）上累计花费了 1000 万美元税款，但实际获得的赔付仅为 150 万美元，而保险起赔额（Deductible）就高达 200 万美元，导致这笔庞大的保险开支实际上被白白浪费。此外，基层在遇到网络安全漏洞或故障时，往往出于人性本能选择自行掩盖或私下排查，结果反而让安全漏洞越来越深。州政府正尝试通过更有序的政策与安全监督体系来纠正这种困境。 - 阿肯色州立法强制所有州政府机构制定 AI 使用合规政策。 - 州长当选后全州遭受的网络攻击量激增 100%。 - 5 年内该州支出 1000 万美元网络安全保费，仅获赔 150 万美元（免赔额 200 万美元），倒逼安全防护策略转型。 ### [20:25](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=1225s) 阿肯色州的网络安全自保基金与闭门会议立法 阿肯色州众议员 Stephen Meeks 详细分享了该州为应对勒索软件和网络攻击所建立的一套协同机制。在过去，地方机构在遭遇攻击时往往害怕寻求帮助；为此，州政府建立了一个全州范围的“自保基金”（Self-Supported Insurance Fund），要求各城市、州级机构以及学校共同注资参与。该基金相当于一份应对网络攻击的联合保险单，但获取资金赔付的前提是各单位必须严格遵守既定的安全规范。这些规范并非由行政长官单方面决定，而是由首席网络官（Chief Cyber Officer）联合所有参保实体共同组成的委员会制定，确保各方都有发言权。通过这种机制，网络安全合规的激励层层渗透，甚至细化延伸到了地方公共图书馆。 ```mermaid flowchart TD accTitle: 阿肯色州网络安全自保基金机制 subgraph RuleMaking["规则与参与"] A["首席网络官 & 各机构代表"] -->|"共同制定规则"| B["网络安全标准"] C["各城市 / 州机构 / 学校 / 图书馆"] -->|"注资并合规执行"| D["网络安全自保基金"] end subgraph Incident["应急响应与闭门谈判"] E["遭遇勒索软件攻击"] -->|"合规单位申请援助"| D D -->|"专家介入谈判 & 资金支持"| F["闭门闭门会议评估 / 支付"] end ``` Meeks 议员随后分享了一个真实而惨痛的案例：北小石城学区（North Little Rock School District）曾遭遇勒索软件攻击，攻击者索要 25 万美元赎金。州级技术专家介入后，成功将赎金压低至 16 万美元。然而，依据当时的法律，校董会批准资金支出必须召开公开会议；校董会在公开会议上授权了“最高 25 万美元”的支付额度。结果不到一个小时，黑客就打来电话，指明要求全额支付 25 万美元，最终学区不得不全额照付。这一教训促使阿肯色州迅速修改法律，允许地方实体以及州议会的技术委员会在处理此类网络安全漏洞和勒索事件时举行闭门会议（Private Sessions），从而在保护决策透明底线的同时，避免向犯罪分子暴露自身漏洞与谈判底牌。 - **全州自保基金**：各级政府部门、学区和公共机构统一注资，合规即可在受灾时获得专家支援与赔付资金。 - **北小石城学区事件**：勒索赎金由 25 万美元砍至 16 万美元，但因公开会议表决暴露了 25 万美元上限，导致黑客重新索要全款并得逞。 - **闭门会议立法豁免**：修改州法律，使技术委员会和地方机构能在涉密网络事件中闭门研判，防止安全弱点与底线外泄。 --- ### [22:46](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=1366s) 立法者的专业化挑战与给科技企业的游说建议 Stephen Meeks 议员进一步向与会者给出了极为务实的政策参与建议。他回忆道，当他最初担任科技委员会主席时，委员会讨论的核心议题还只是电话服务和宽带普及；而在过去十年里，技术发展的速度呈指数级加快。科技立法的专业门槛极高，就像医学政策需要依赖具有医疗背景的议员一样，科技政策也迫切需要愿意深挖细节、精通业务的专业立法者，以确保出台的政策基于专业知识，而非社会上的道听途说或末日论调（Hearsay and Doomsday）。 面对广大中小企业与初创团队，Meeks 强调了“提前沟通”的至关重要性。在立法会议（Legislative Session）正式开幕期间，议员每天清晨赶往小石城，全天需要研判辩论 40 到 50 项法案，往往筋疲力尽，根本没有精力和时间去消化复杂的全新议题。如果企业在开会当天临时找来游说，议员必然倾向于采纳平日里已建立信任关系的熟人的意见。因此，企业绝不能等法案审议时才临时抱佛脚，而应在休会期主动与负责科技政策的核心议员建立长期联系，甚至可以与其他中小企业联合成立州级 AI 联盟（AI Consortium），共同发声。Meeks 指出，当前的 AI 监管框架正在各州层面摸索成型，各州达成的广泛共识未来将反哺并推动联邦立法的落地。 - **立法节奏与现实**：议会开会期间每天需审理 40-50 项法案，议员无暇在现场临时学习复杂技术，临时游说极难奏效。 - **提前建立信任**：基于人性的决策逻辑，立法者在面对争议时会优先信任已有长期合作关系的对象。 - **中小企业破局策略**：建议无法单打独斗的小企业联合组建州级 AI 产业联盟，在休会期提前介入监管政策的讨论。 --- ### [26:23](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=1583s) 破局非逻辑的国防采购体系：非传统供应商的突围之道 主持人在总结 Meeks 的观点后顺势提问：初创企业与小微企业是技术创新的主要驱动力，但国防采购系统（Defense Procurement System）壁垒森严，Defense Alliance 执行董事 Chip Laingen 应如何看待非传统创新者为国家安全生态带来的独特价值？Chip 首先补充了一条给 AI 工程师的实战建议：当技术团队试图为政府寻找解决方案并查阅需求书（RFP）时，往往会发现政府的需求模糊不清甚至根本不存在。工程师绝不要耗费数年时间去空转寻找明确参数，因为在 AI 领域，前沿技术早已大幅领先政府的认知边界；企业应当具备决断力，主动为政府定义场景并设定技术指标——“大胆告诉政府你能做什么，为你自己设定舞台”。 谈及复杂的政府采购，曾作为海军飞行员退役并在该领域深耕 20 年的 Chip 坦言，即便从业多年，面对国防采购依然会感到挫败。他回忆起近 30 年前在海军部长（Secretary of the Navy）幕僚团队担任演讲撰稿人的一段往事：当时军方正在探讨是否允许女性进入潜艇服役，有幕僚困惑地质疑这在逻辑上是否合理，海军部长则意味深长地回答："我们要做符合美国价值观的事，而不一定是（传统眼光中）正确的事（We're going to do the American thing, not necessarily the right thing）。" Chip 借此指出，政府合同制度本身就不是一个纯粹理性的逻辑系统，企业必须首先接受这一现实。 与传统军工巨头相比，小型科技企业和初创公司作为“非传统供应商”（Non-Traditional Suppliers），具备更高的风险承受力、更快的迭代速度与创新能力，能将其他民用工业领域的尖端经验跨界引入国防工业。尤其在近八九年及现任政府的推动下，军方正主动打破诸多繁文缛节，积极吸纳体制外的非传统力量。因此，缺乏政府项目经验不再是致命短板，科技团队应当克服对联邦官僚的畏惧心理，自信而大胆地展示自身跨界技术实力。 - **主动定义需求**：AI 前沿进展远超政府现行标准，企业无需苦等现成的 RFP 参数，而应主动为军方定义技术路径。 - **采购系统的非逻辑性**：国防采购并非纯理性运转体系，企业需适应其独特的制度惯性与价值取向。 - **非传统供应商的核心优势**： - 更高的风险偏好与容错度； - 极快的产品迭代与响应速度； - 将民用与跨行业前沿技术（如商业 AI）反哺国防安全的能力。 ### [30:07](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=1807s) 政府内部的采购痛点与新兴商业准入通道 来自政府内部的视角往往与外部供应商同样充满挫败感。体制内的项目负责人和技术用户往往极度渴望尽快引入并应用前沿技术，但不得不面对繁重官僚体制的层层阻碍。内部各部门的业务人员常常在采购和采办部门的桌子上“拍桌子”，敦促他们开辟全新通道、加快采办速度，因为“世界正在飞速向前发展，而政府的反应显然过于迟缓”。 为了打破这种僵局，联邦政府正逐步推出更多精简流程的尝试。例如国防创新部门（DIU）以及美国国土安全部（DHS）新推出的商业解决方案征集公告（Commercial Solutions Opening, CSO）。在该机制下，企业只需提交视频和简短提案（Pitch）即可参与竞争。尽管该项目重点围绕海事技术，但若企业能为美国海岸警卫队（US Coast Guard）等机构的员工提供定制化的大语言模型（LLM）等支持方案，也能精准切入。政府内部正在积极寻求更多类似的创新采办机制，以缓解体制内的采办瓶颈。 ### [31:55](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=1915s) 州级政府合作优势：小型州的灵活性与敏捷性 对于初创企业与小型科技公司而言，参与政府采办并不一定要把目光完全锁定在联邦层面。各州政府在合作上往往更加便捷。全美拥有 50 个独立的州级实体，虽然像阿肯色州这样常被外界视为“飞地”或存在刻板印象的较小州资源不如加利福尼亚、佛罗里达或得克萨斯等大州雄厚，但其核心优势在于机制更加灵活、反应更为敏捷。在快速迭代的人工智能与新兴技术领域，选择机制扁平的小型州开展试点与落地，往往比直接死磕大州或联邦政府更容易获得突破。 ```mermaid flowchart TD accTitle: Government Procurement Routes for Small Tech Businesses subgraph Federal["联邦政府层级 (Federal Level)"] F1["国防创新部门 (DIU)"] F2["国土安全部 CSO 通道 (商业解决方案征集)"] end subgraph State["州级政府层级 (State Level)"] S1["大型州 (加州/得州等: 资源丰富但竞争激烈)"] S2["中小型州 (如阿肯色州: 机制更灵活、决策更敏捷)"] end subgraph Support["区域生态与扶持机构 (Regional Support)"] R1["APEX 加速器 (原 PTAC 采办技术援助中心)"] R2["区域创新集群 (Regional Innovation Clusters / SBA)"] end F1 --> Vendor["创新科技企业 / 中小企业 (Small Business & AI Tech)"] F2 --> Vendor S2 --> Vendor R1 -.->|"提供免费前期咨询与匹配"| Vendor R2 -.->|"提供法规、财税与深度对接"| Vendor ``` ### [32:53](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=1973s) 区域创新网络与中小企业支持生态 区域性组织在帮助企业对接政府需求方面发挥着关键桥梁作用。目前存在诸多面向小企业的扶持资源，例如由原采办技术援助中心（PTAC）更名而来的 APEX 加速器（Apex Accelerators），以及区域创新集群（Regional Innovation Clusters）。尽管全国原有的 25 个创新集群在经历“政府效率部”（DOGE）的削减后目前仅存 3 个（Defense Alliance 是幸存者之一），但美国小企业管理局（SBA）正计划逐步重建这些网络。 - **免费咨询与资源对接**：这类区域集群拥有熟悉本地环境的专业团队，掌握各州政策及能够匹配联邦资助的项目资金，且通常为企业提供完全免费的基础咨询服务。 - **由广博到深度的过渡**：区域组织能够跨领域提供多维度的指导，而在企业进入招投标深水区后，涉及深度法律事务、合规监管或财税会计时，集群也会精准引荐专业机构，转向市场化的有偿服务（Fee-for-service）。 ### [34:49](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=2089s) AI 伦理治理、认知对抗与数字素养教育 针对人工智能快速部署带来的伦理与监管挑战，政企之间的责任划分暴露出明显脱节。科技巨头过去在面对立法者关于伦理边界的质询时，往往给出“我们负责开发技术，你们来找出应对答案”这类推卸责任的回应。立法者认为企业绝不能置身事外，必须共同参与治理规则的制定。 当前各州在技术防护层面的紧迫课题正在延伸至认知安全领域。阿肯色州与犹他州等超过 32 个州近期立法禁止在校园内使用手机，背后的核心动因正是 AI 技术的爆发以及对青少年心理损害认知的加深。在更宏观的社会层面，美国政治体制虽然具有将分歧公开辩论的传统优势，但这种公开性也极易被外国敌对势力（Foreign Adversaries）利用。敌对势力通过部署由 AI 驱动的大规模虚假聊天机器人（Chatbots），在各大社交媒体平台上制造对立、煽动政治极化情绪，甚至可能借由特定突发事件迅速在现实街头掀起动乱。 面对这一新型认知安全威胁，阿肯色州在上一立法季通过法案，将“数字素养”（Digital Literacy）强制纳入学生课程体系，重点培养下一代批判性思维，使其能够分辨交互对象是真人还是 AI。这一挑战未来甚至可能逼迫社会走向“全网实名身份认证（National ID）”的监管极端，但此举同样伴随巨大争议。 立法者以历史上的“便携式计算器危机”类比——当年外界普遍担忧计算器的普及会导致学生不再学习数学，但人类最终成功适应了工具。对 AI 发展的长期前景应保持乐观，尽管试错不可避免，但通过企业深度参与地方议会、州政府与联邦代表团的政策制定，广泛的社会对话将有效减少重大决策失误。 ### [39:22](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=2362s) 价值内置与防御系统底座设计 在监管框架成型之前，AI 系统的价值导向本质上取决于开发者自身的设计抉择。正如 Geoffrey Hinton 博士在大会演讲中提到的，人类甚至需要探索为 AI 赋予类似“母性本能（Maternal Instincts）”的内置约束，以防止技术反噬人类。在面对能力远超人类总和的超级防御系统时，系统设计者必须将核心价值体系直接融入底层架构，而不能被动依赖滞后的外部监管机构来规范系统的伦理边界。 ### [40:14](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=2414s) 深度伪造立法与技术开发商的连带责任 讨论首先强调了国防与军事领域中恪守伦理标准的重要性。企业若想向军方提供技术解决方案，必须将合规与伦理考量置于核心地位，并在方案演示中明确体现对这些准则的尊重与承诺。 阿肯色州州众议员斯蒂芬·米克斯（Rep. Stephen Meeks）随即分享了州立法层面的最新举措。他在本届立法会议中推动通过了一项禁止非法深度伪造（illicit deepfakes）的法案。该法律规定，初犯者将被定为轻罪（misdemeanor）并处以警告罚款，再次违规者则面临监禁。更关键的是，该法案与社交媒体相关法案类似，创设了民事诉讼请求权（cause of action）：受害者若因深度伪造遭受经济或名誉损害，除政府罚款外，受害者本人有权直接起诉侵权者追讨全部损失；同时，托管或提供图像生成软件的技术公司若未设置合规防护措施，也将面临民事连带诉讼。 米克斯议员以枪支制造商作类比：通常情况下，若有人非法持枪伤人，在有枪支管控法规的前提下，制造商无需担责；但如果军火商随意将枪支弃置街头、任由任何人随意获取并提供无限量的免费弹药，那么一旦酿成惨剧，厂商显然必须承担法律责任。同样的逻辑也适用于生成式 AI 工具开发商——如果企业推出了能够生成危害性内容的工具，却没有建立完善的安全护栏（safeguards），导致不法分子利用其摧毁他人的生活，技术提供方就必须共同承担责任。他告诫所有 AI 开发者，在产品设计之初就必须嵌入安全防御机制，并预计这种平台追责立法将在全美各州迅速推广。 ```mermaid flowchart TD accTitle: 深度伪造侵权责任与追责链条 subgraph LegalChain["深度伪造法律追责与治理机制"] direction TB A["不法分子利用 AI 生成非法深度伪造"] --> B["受害者遭受经济或名誉损失"] B --> C["刑事惩戒：初犯罚款 / 再犯监禁"] B --> D["民事诉讼请求权 (Cause of Action)"] D --> E["起诉侵权个人索赔全部损失"] D --> F["追究 AI 软件/平台连带民事责任"] F -.-> G["前置条件：技术方未设置必要安全防护 (Safeguards)"] end ``` --- ### [42:28](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=2548s) 展望未来：AI 如何重塑人机协作、战场决策与政务效率 在圆桌会议的尾声，主持人抛出最后一个核心问题：**未来 AI 最具变革潜力的优先领域是什么？各方需要做些什么才能确保走在正确的轨道上？** 美国海岸警卫队前首席数据与人工智能官布赖恩·埃里克森（Brian Erickson）率先发言。他坚信 AI 的核心价值绝非取代人类，而是成为人类的智能伙伴（partners）。他指出，年轻一代的军人与技术人员已经展现出极强的适应力，他们积极利用 ChatGPT 等生成式工具快速攻克过去极其耗时的复杂问题。埃里克森打了个生动的比喻：就像当年磁共振成像（MRI）技术问世时一样，那些固守传统、拒绝使用 MRI 辅助诊断的医生很快便被时代淘汰；如今不论年轻一辈还是老一辈，都必须学会如何与 AI 深度协作，否则同样会面临被边缘化的风险。 随后，国防联盟（Defense Alliance）执行董事奇普·莱恩根（Chip Laingen）指出，在军事防务领域，AI 的终极目标是赋能一线作战人员（warriors），帮助他们在战场的高压环境下做出“更快、更好、更具判断力”的决策。为了强调主动性，他引用了电影《刺杀本·拉登》（*Zero Dark Thirty*）中的经典桥段：当 CIA 高官飞抵巴基斯坦对前方团队严厉训话时说道：“如果你以为在另一层楼还有别的行动小组在替你们解决这个问题，根本没有，解决问题的人就是你们自己（'If you think there's some other cell solving this problem on another floor, there isn't. It's you.'）。”莱恩根以此鼓励现场的技术开发者和企业，不要坐等五角大楼（Pentagon）或国土安全部（DHS）下达具体指令，业界最清楚技术的真实能力，应当主动将成熟可行的解决方案带到国家安全的谈判桌前。 米克斯议员最后从地方与州政府治理的角度补充道，AI 将在推动政府运作“更聪明、更迅捷、更高效”方面发挥无可替代的作用。在他主持制定的州 AI 安全护栏政策中，明确立下一条铁律——**人类必须始终是最终决策者（Human in the loop）**。没有任何公众希望在向政府部门求助时得到“抱歉我帮不了你，因为电脑程序就是这么判定的”这种冷冰冰的回复，立法者也绝不允许公务人员推卸责任。AI 的真正用途在于替公共部门分担繁重的流程，解决车管所（DMV）排长队、失业救济金审核缓慢以及基层人手短缺等痛点，显著提升公民与政府交互的体验。 - **年轻一代人机协作趋势**：初入军营的年轻官兵正在率先熟练运用 LLM 工具，以分钟级速度化解以往高耗时的信息检索与分析流程。 - **医学技术演进启示**：拥抱 AI 的过程犹如医学界全面接纳 MRI，掌握人机协同技能是未来专业人才立足的基石。 - **一线赋能与主动交付**：防务技术创新的重心在战场决策效率，创新源头在科技产业界，无需等待国防体系自上而下的缓慢规划。 - **政务服务与人类主导**：在提升 DMV、失业救济等公共服务效率的同时，严格保留“人在回路中”的底线责任制。 ## Notable quotes > [03:10](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=190s) 只要是开直升机的，脑子里总觉得下一秒坏事就要发生 — Chip Laingen > [04:54](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=294s) 给政客一支麦克风和一个观众群体，往往是件很冒险的事。 — Stephen Meeks > [06:51](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=411s) 它能让你以更高的强度、更快的速度做出同样的错误决策。 — Christopher Swift > [26:23](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=1583s) 大胆告诉政府你能做什么，为你自己设定舞台 — Chip Laingen > [26:23](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=1583s) 我们要做符合美国价值观的事，而不一定是（传统眼光中）正确的事（We're going to do the American thing, not necessarily the right thing）。 — Chip Laingen > [30:07](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=1807s) 世界正在飞速向前发展，而政府的反应显然过于迟缓 — Brian Erickson > [34:49](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=2089s) 我们负责开发技术，你们来找出应对答案 — Stephen Meeks > [42:28](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=2548s) 如果你以为在另一层楼还有别的行动小组在替你们解决这个问题，根本没有，解决问题的人就是你们自己（'If you think there's some other cell solving this problem on another floor, there isn't. It's you.'）。 — Chip Laingen > [42:28](https://www.youtube.com/watch?v=pzcYeDlUZuk&t=2548s) 抱歉我帮不了你，因为电脑程序就是这么判定的 — Stephen Meeks ## Mentioned Olivia Zetter, Christopher Swift, Brian Erickson, Chip Laingen, Stephen Meeks, Geoffrey Hinton, Sarah Huckabee Sanders, ARCH Venture Partners, Foley & Lardner LLP, Defense Alliance, 美国海岸警卫队 (U.S. Coast Guard), 美国小企业管理局 (SBA), 全美州议员大会 (NCSL), 国防创新部门 (DIU), 美国国土安全部 (DHS), APEX 加速器 (Apex Accelerators), 政府效率部 (DOGE), ChatGPT, 电影《刺杀本·拉登》（Zero Dark Thirty）, 阿肯色州 (Arkansas) ## Video description Olivia Zetter, Strategic Advisor & Venture Partner, ARCH Venture Partners Christopher Swift, Partner, Foley & Lardner LLP Brian Erickson, Former Chief Data and AI Officer, U.S. Coast Guard Chip Laingen, Executive Director, Defense Alliance Rep. Stephen Meeks, State Representative, State of Arkansas Abstracts This session will explore how AI is revolutionizing defense, intelligence, and preparedness, enabling faster decision-making, enhanced threat detection, and more effective resource allocation. Panelists will discuss how AI-enabled technologies are strengthening national security and enabling the next generation of strategic capabilities. The panel will also address the ethical considerations and regulatory challenges surrounding the use of AI in public security. Learn how AI is shaping a more resilient future and what governments and the private sector need to consider when deploying these powerful technologies. [Source](https://www.youtube.com/watch?v=pzcYeDlUZuk)

[3:10] Chip Laingen: 只要是开直升机的，脑子里总觉得下一秒坏事就要发生 [4:54] Stephen Meeks: 给政客一支麦克风和一个观众群体，往往是件很冒险的事。 [6:51] Christopher Swift: 它能让你以更高的强度、更快的速度做出同样的错误决策。 [26:23] Chip Laingen: 大胆告诉政府你能做什么，为你自己设定舞台 [26:23] Chip Laingen: 我们要做符合美国价值观的事，而不一定是（传统眼光中）正确的事（We're going to do the American thing, not necessarily the right thing）。 [30:07] Brian Erickson: 世界正在飞速向前发展，而政府的反应显然过于迟缓 [34:49] Stephen Meeks: 我们负责开发技术，你们来找出应对答案 [42:28] Chip Laingen: 如果你以为在另一层楼还有别的行动小组在替你们解决这个问题，根本没有，解决问题的人就是你们自己（'If you think there's some other cell solving this problem on another floor, there isn't. It's you.'）。 [42:28] Stephen Meeks: 抱歉我帮不了你，因为电脑程序就是这么判定的

[1:40] Chip Laingen: We always think something's really bad is about to happen if you're flying helicopters. [1:40] Rep. Stephen Meeks: I'm very glad that we were able to pull that 10-year moratorium away from the federal government because I can tell you the states do not want to stifle innovation, but we want to protect our citizens. [1:40] Christopher Swift: You get to make the same bad decisions faster with a higher degree of intensity. [1:40] Christopher Swift: The hardware and the software still requires the wetware. [21:34] Rep. Stephen Meeks: Within an hour they got a phone call from the bad guys. Guess what they wanted? They wanted $250,000 and that's what they ended up getting. [26:24] Chip Laingen: We're going to do the American thing, not necessarily the right thing. And every now and then those things coincide.

[0:04] Great. Well, welcome to our session [0:06] today on how AI is shaping defense [0:09] intelligence and preparedness. Um, today [0:11] we'll explore where AI is already making [0:14] a tangible difference, the barriers that [0:17] are preventing adoption and are still in [0:20] our way and the policy, ethical, and [0:22] technical choices uh that really impact [0:25] how we apply and deploy these [0:27] innovations. I'm joined by a fantastic [0:30] panel today. We have Brian Ericson, [0:33] former chief data and AI officer for the [0:35] US Coast Guard. Chip Langan, executive [0:38] director of the Defense Alliance, [0:41] Representative Steven Meeks, state [0:43] representative of Arkansas, and [0:45] Christopher Swift, partner at Foley [0:47] Lardner. And we'd like to start the [0:50] conversation today just by learning very [0:52] quickly a little bit more about who's in [0:54] the room. So, if you work for a small [0:58] startup or small business, could you [1:00] please raise your hand? [1:03] Great. And if you work for a large [1:04] company, a tech company or defense [1:06] contractor, please raise your hand. Wow, [1:10] lots of you. Okay. And if you work in [1:12] government or public service, please [1:14] raise your hand. [1:16] Great to see some of you still out [1:18] there. Um well to start this [1:21] conversation uh we often hear about AI [1:24] as a strategic advantage in defense and [1:26] intelligence. From your vantage point, [1:29] where are you seeing some of the most [1:31] impactful applications of AI for our [1:34] national security and defense? And [1:37] Brian, why don't we start with you and [1:38] we'll just go down the line. [1:40] >> Yes, sounds good. Thanks, Olivia. And uh [1:42] glad to be here at AI4 this year. Um am [1:45] surprised how many government people [1:47] made it. I uh I just recently retired [1:50] after 33 years in the US Coast Guard as [1:52] a captain and uh it was easy for me to [1:55] get here, but uh some of my friends [1:57] weren't able to make it uh just because [1:59] they kind of were tightening up on the [2:01] conferences. So, good to have uh [2:03] government folks uh here in the room [2:04] that were able to get here. Um so, you [2:08] know, I spent some time as an operator [2:11] primarily. I was a helicopter pilot uh [2:13] throughout my tours in the uh Coast [2:16] Guard. flew fixed wing a little bit. Uh [2:18] was in engineering logistics and then [2:19] and then made my way up to uh work in [2:22] tech as the Coast Guard's first uh chief [2:24] data AI officer. Um and then uh just [2:27] left that assignment just about a year [2:30] and a half ago. So what we were seeing [2:32] in in our um applications of AI were [2:36] some some good uses in uh geospatial uh [2:39] uses of AI and geospatial applications [2:42] um computer vision and then we were [2:44] starting to work with large language [2:46] models really uh just mo mostly um you [2:50] know search and research for um our [2:53] workforce um in different uh different [2:57] uh verticals within the organization and [3:00] uh and helping our employees get uh get [3:02] information that they needed to do their [3:04] job better. So, those were kind of some [3:05] of the areas for for me when uh at the [3:07] time when I was there. [3:10] Uh good morning everyone. I'm also a [3:12] retired helicopter pilot in the in the [3:14] sea services but Navy. So, uh [3:16] interesting to have two people on the [3:18] stage that uh we're always there's [3:20] another one, right? Uh we always think [3:22] something's really a bad really bad is [3:24] about to happen if you're flying [3:25] helicopters. So, hopefully that's not [3:27] the case uh with AI here. Although if [3:29] you heard uh Dr. Hinton the other day, [3:31] maybe there is something bad. Uh but [3:33] yeah, I did 21 years Navy. Uh retired in [3:36] 2004. Uh and I've been standing up [3:38] Defense Alliance for the past 20 years, [3:41] which is a uh regional innovation [3:43] cluster for the Small Business [3:45] Administration. We help small businesses [3:47] uh get contracts. So, it's great to see [3:49] a huge swath of small businesses and the [3:53] primes uh a lot of the primes in the [3:55] room as well. So, I was laughing because [3:57] we got to be careful who we're going to [3:58] insult here. It's nice to know they got [4:00] the whole the whole the mix. I may say [4:03] some things about the prime contractors [4:04] that you may not appreciate. So, maybe [4:06] you can talk to me uh in the hallway [4:07] later. Um what we're seeing in terms of [4:10] innovation in AI. Uh I think there's [4:12] kind of three areas for me and this is [4:14] again very small business centric. It's [4:17] the ability to do deep research with AI [4:20] and speed up the process of innovation. [4:23] uh that's been really great to see with [4:24] our small business membership. Uh [4:27] second, we're bringing things to the [4:28] battlefield that really help with uh [4:32] analyzing intelligence faster and making [4:35] decisions faster uh on the battlefield. [4:38] And then also we're starting to see some [4:40] impacts on uh helping with resiliency [4:43] and access to supply chains to improve [4:45] that process because all three of those [4:46] things uh matter in terms of uh making [4:50] our defense infrastructure more uh [4:52] resilient. [4:54] Okay. Uh so good morning again. My name [4:56] is State Representative Steven Meeks [4:57] from the state of Arkansas. Appreciate [4:58] AI for having me uh because as you all [5:01] know giving a politician a microphone [5:03] and an audience can sometimes be risky [5:05] business, right? Um but uh I'm going to [5:08] be bringing the state perspective this [5:09] morning. Uh and uh it's it's a very [5:11] important perspective. Last week I was [5:13] at the National Conference of State [5:15] Lawmakers in Boston. It was the largest [5:17] gathering of state lawmakers actually in [5:19] the history of our country. And as you [5:22] can imagine, AI was at the forefront of [5:24] uh a lot of those discussions. Um [5:27] there's a lot of regulations and policy [5:29] being done in the states. the states [5:31] understand that we're all competing [5:32] against each other in this space to try [5:35] to take the greatest advantage of it as [5:37] we can. And by the way, I I would be [5:39] missing my duties. If any of you are [5:41] looking at making future investments [5:42] anywhere, Arkansas is a great place and [5:44] I would love to speak with you [5:45] afterwards about how you can make that [5:47] happen. Um but but anyways, the states [5:50] at the same time are also wanting to [5:52] make sure we're protecting our citizens [5:53] and that's everything from uh deep fakes [5:56] to uh you know social media uh because [6:00] as part of the national defense that [6:02] goes all not only the federal government [6:04] but also on the home front and that's [6:05] the important place that that the uh [6:07] role that the states play and I'm very [6:10] glad that we were able to pull that [6:12] 10-year moratorium out of away from the [6:14] federal government because I can tell [6:15] you the states do not want to stifle [6:18] innovation, but we want to protect our [6:20] citizens. And there's a lot of [6:22] protections that would have gone away [6:23] had that bill been uh had been passed as [6:27] part of the bill. Uh and so what my hope [6:30] is is that as the states, the [6:32] laboratories of democracy, as we figure [6:34] out a lot of these AI policy questions, [6:37] that when we start to see consensus [6:39] among the states, that's usually a [6:41] pretty good sign to the federal [6:42] government, hey, it's time to take [6:43] federal action on these issues. Um, and [6:46] so again, glad to to be here and be part [6:49] of this conversation. [6:51] >> Good morning everyone. I hope you can [6:52] hear me. My name is Christopher Swift. [6:54] I'm a partner uh and a litigator at the [6:56] law firm Foley and Lardner. We're a [6:58] 200-year-old White Shoe Law Firm and our [7:01] job is to make your problems go away. [7:04] Um, could we raise the question again if [7:06] that's okay? What was the question [7:08] presented? [7:10] So if you want to just tell a little bit [7:12] more about your work and then also where [7:14] you are seeing some of the most [7:15] impactful applications of AI um with the [7:19] companies that you're working with. [7:21] >> Okay. So the most impactful applications [7:23] of AI. How many of you had coffee this [7:25] morning before you came? Hands up coffee [7:28] drinkers. Okay. Did you know that before [7:31] 1400 [7:32] people drank beer in the morning [7:35] and then coffee was brought in from the [7:37] Middle East and by 1500600 people were [7:39] drinking coffee in the morning at least [7:41] in cities. Europe's gross national [7:44] product quadrupled after the [7:47] introduction of coffee. [7:50] What I see is a lawyer counseling [7:52] clients who are either trying to take AI [7:54] and plug it into what they're already [7:55] doing or trying to leverage AI to do [7:59] what they do better or who are dealing [8:01] with the fallout of the government [8:02] trying to figure out what in God's good [8:04] name to do with AI and the technology [8:06] around it is a lot like coffee. You get [8:09] to make the same bad decisions faster [8:12] with a higher degree of intensity. [8:15] That's supposed to be a joke. The point [8:16] I'm trying to make here is the most [8:19] important thing that we're seeing with [8:21] respect to AI as as lawyers practicing [8:24] in this space is that technology doesn't [8:28] solve the human problem. [8:31] The hardware and the software still [8:34] requires the wet wear. And so if you're [8:37] using the technology in a way that [8:38] amplifies or augments what your mission [8:42] set already is and we tend to see that [8:45] those companies are doing really really [8:46] well. But in those instances where [8:49] people are sort of stapling an AI [8:51] capability or AI branding onto something [8:53] without truly understanding what the [8:55] core mission of the organization is, we [8:58] see a lot of tension and we see a lot of [8:59] mistakes and we see mistakes that [9:01] sometimes get companies over their skis [9:04] when it comes to the US government. Um, [9:06] why should you have a lawyer on the [9:08] stage? Well, you've got two helicopter [9:09] pilots and a politician from Arkansas, [9:12] so you might as well have a lawyer up [9:13] here. Um, the bottom line is people are [9:16] always going to be integral to whatever [9:18] your AI solution might be, right? The [9:21] technology, the Greek word technology [9:23] means technologos or knowhow. The [9:26] technology is the tool you use to [9:28] achieve your objective. But determining [9:31] what the objectives are, that's still a [9:33] human endeavor. And I think that's [9:34] probably the most important thing that [9:36] working with AI has taught us over the [9:38] last few years. [9:39] >> I've heard AI compared to electricity or [9:42] to the internet revolution, but I [9:44] haven't heard it compared to coffee yet. [9:46] And I like that analogy. I might have to [9:47] steal it. Brian, in your dayto-day with [9:52] uh the Coast Guard, you had to apply AI [9:54] and these datadriven technologies across [9:56] a multi-domain [9:58] um highly diverse environment. What is [10:01] one application that really had a [10:04] tactical operational impact? [10:07] >> Yeah. So, so we like the federal [10:11] government, I there's a lot of struggles [10:13] across federal government with legacy [10:15] technology. So, you know, first of all, [10:18] we always said that our our biggest [10:20] challenge was going to be people and [10:22] culture. And I think that I think that [10:24] truthfully if we really pull the thread [10:26] on what was you know preventing us from [10:29] you know rapidly deploying artificial [10:31] intelligence across all of our domains [10:33] that's probably what it was. But it was [10:35] also it was also hard to deploy [10:38] technology. Um getting the getting the [10:40] people in place was very much um an [10:43] important part of my role. getting the [10:45] talent in the right locations in the [10:47] organization so that they could um help [10:49] the frontline operator do perform their [10:52] duties better um work with the joint [10:54] forces better uh work with the IC better [10:58] um and then the culture you know are we [11:01] going to we're a 235 year old [11:04] organization it's the it was the it's [11:06] the longest uh continuous we use that [11:09] word continuous seagoing ser the oldest [11:11] continuous seagulling service in America [11:14] Navy maybe had a little breakup at one [11:15] point in time, but uh they're back [11:17] together. Um and so a 235 year old organ [11:21] old organization is going to you know [11:23] have some legacy ideas and trying to [11:25] break those legacy ideas and bring um [11:28] you know new ideas to the fold and and [11:30] resource you know technology artificial [11:33] intelligence when you can't even feel it [11:35] or you don't even know exactly what it's [11:37] going to provide you in this work over a [11:39] physical asset like hey buying a new [11:41] helicopter or a new coastg guard cutter [11:44] that was a a cultural challenge uh while [11:47] I was there and I think that my my peers [11:49] that are still there continue to [11:51] struggle with that um an application. So [11:54] we we had some great um AI use cases in [11:58] geospatial um targeting uh as an example [12:02] uh many ships over a certain tonnage are [12:05] required to to um push out a transponder [12:08] signal. So you know every uh you know 15 [12:12] seconds to 1 minute to you know even 10 [12:15] minutes there's a signal and and you can [12:17] you can locate these ships everywhere um [12:20] across the world. um but they can spoof [12:22] them. And so using artificial [12:24] intelligence on top of a geospatial [12:26] platform, we could identify, you know, [12:29] very very um uh complex spoofing um [12:33] efforts where um taking a visual [12:37] satellite image, we would locate vessels [12:40] and where they were traveling versus [12:42] where their AIS were and then identify [12:44] where bad things might be happening. You [12:46] know, two large ships that come together [12:48] out in the middle of the ocean. That's [12:50] interesting, you know. So things like [12:51] that were very valuable. I'll give you [12:53] one example. Um as we have been [12:56] continuing to try to um advance this [12:57] effort. I was doing a search and rescue [12:59] case off of Tybee Island one time in a [13:01] in a helicopter and um the case was um a [13:05] a child was watching his father swim and [13:08] the father was a surfer but he was just [13:11] out swimming offshore Tybee Island and [13:12] he got swept away. The child reported to [13:15] it. We got the call and we went out [13:17] there. We were probably there within an [13:20] hour, but maybe 45 minutes of of the [13:22] loss of sight of this guy. And so the [13:25] sun was setting. It was it was probably [13:28] 6:00 by this time. We had about 3 hours. [13:30] We searched for 3 hours. We put two [13:32] helicopters on it. We looked everywhere [13:34] for this guy. We knew that there's a [13:36] possibility that he survived because he [13:38] was a strong swimmer. It just didn't [13:39] make any sense. You know, it would have [13:40] been an anomaly if he he wasn't there. [13:43] By the grace of God, the sun was setting [13:45] really low on the horizon and of course [13:48] he's down there, you know, waving [13:49] because he can see the helicopter, but [13:50] his watch reflected the sunlight and [13:53] someone saw it and they um turned [13:57] towards him. He said we flew over him 20 [13:59] times. [14:00] >> It's you know it is so hard to see a [14:02] head especially on in different sea [14:04] states. So we were working on a computer [14:07] vision model that inside the helicopter [14:09] using our EOIR system would detect [14:12] anomalies in up to um you know certain [14:15] sea states where we could find this and [14:17] and cue um you know our our our teams to [14:21] look at certain areas like hey and you [14:23] know maybe we would have saved that guy [14:24] uh a little bit earlier. We're lucky on [14:26] this one but maybe we're not lucky on [14:27] the last on the next one. Yeah, these [14:30] threats that we face today, um whether [14:33] they come from emergencies like that, [14:36] whether they're cyber, physical, and [14:38] kinetic, uh new or novel, they're [14:41] incredibly diverse and nuanced. When we [14:45] think about developing these [14:46] technologies that can be most useful to [14:49] operators and to users, how can [14:52] engineers and model developers really [14:55] build to fit a dynamic and nuanced [14:58] threat landscape? And Chris, you've [15:00] counseledled companies across the [15:02] application space. How are you helping [15:04] them think about this challenge? Sure. [15:06] >> Well, one of the most important issues [15:09] when teaching AI how to learn is what [15:13] you're teaching it. and what assumptions [15:15] are built in. So when we talk about [15:17] identifying a particular target set like [15:19] you know chips that are I don't know [15:21] smuggling Russian oil for example that [15:23] are spoofing their AIS um that's an [15:26] objective [15:28] factual physical thing that you can [15:31] track right is the ship sending out a [15:33] signal right is there something about [15:35] the size of a human head or a seal head [15:38] in the in the ocean right there's a [15:39] physical thing that you can track and [15:42] folks teach the AIA the AI to track [15:45] That's easy. When it comes to managing [15:48] business processes or law enforcement [15:51] processes or god forbid intelligence [15:54] processes, that's not objective anymore. [15:57] That's subjective, which is to say it's [16:00] really complicated because everybody is [16:03] different. We're talking about human [16:04] interactions. So when you're teaching [16:06] AI, when you are culturate AI, [16:10] you are sharing your own expectations or [16:13] your own assumptions about how certain [16:16] things in the world work. If you have a [16:18] scientific basis for that that's [16:20] verifiable with things like numbers, [16:22] then that's a relatively easier problem [16:25] set. When you start to say, okay, tell [16:27] us whether that particular organization [16:30] is going to engage in terrorist organ in [16:32] terrorist activity or tell us whether [16:34] this organization over here is more [16:37] likely or less likely to be a narcotic [16:39] syndicate or tell us whether this [16:41] foreign adversary overseas is going to [16:43] deploy or not going to deploy a [16:45] particular weapon system in a particular [16:47] theater of battle. Right? There's a [16:49] whole lot of judgments in there and [16:51] assumptions in there. And when you plan [16:54] you when you take those assumptions and [16:56] judgments and you put them into the AI [16:58] model, you can wind up with some really [17:00] weird stuff. Let me give you a great [17:01] example of it. As as lawyers, we often [17:04] will do large investigations where we're [17:06] looking at, you know, the evidence that [17:07] the client may have done something wrong [17:09] so that we can either respond to a [17:11] subpoena or figure out what the story is [17:12] to get Uncle Sam to calm down and go [17:14] away. Apologies to my former my current [17:17] government colleagues. Um, we get all [17:20] kinds of hallucinations just looking at [17:22] email traffic because the AI has been [17:25] trained to look for evidence of [17:27] commercial disputes, right? Evidence of [17:30] some instance where one company is [17:32] ripping off another or where there's a [17:34] financial irregularity. But in our [17:36] investigations where we're looking for [17:38] something like did that shipment go to [17:40] North Korea, the AI is completely [17:42] useless. It hallucinates all over the [17:44] place because the language model that [17:45] it's been you that it's used in order to [17:48] try to solve all of our problems as [17:50] lawyers is actually a different area of [17:52] law. It's a different problem set. So [17:54] when you're thinking about how to manage [17:57] AI, you've got to get back to what is [17:59] your core mission, right? What's the [18:01] organization of your what's your [18:03] organization's mission? What task have [18:05] you been asked to solve? And what [18:07] assumptions are you bringing to the [18:08] problem? In political science, we have [18:10] this concept called um selecting on the [18:12] dependent variable, right? Also known as [18:14] cherry-picking the data. If you [18:16] cherrypick the data and it involves [18:19] human interactions, whether it's law, [18:21] the military, intelligence, law [18:22] enforcement, or whatever, you're going [18:24] to get hallucinations. And the result, [18:26] no matter how good your technology is. [18:29] And Stephen, at the state level, how are [18:32] you applying AI to government use cases, [18:35] whether cyber security, uh, critical [18:37] infrastructure management, preparedness [18:39] or otherwise? And how do you think about [18:42] managing the human oversight of these [18:44] technologies? [18:46] >> Uh, great question. And there's still a [18:48] lot of questions there to be to to be [18:50] honest, we're still trying to figure [18:51] that out. Um, so the the the last [18:53] legislative session I passed legislation [18:55] that requires all of our state entities [18:57] to actually develop an AI use policy. In [19:01] the past, um, I don't know, it's been [19:03] about four years ago, I had them develop [19:05] a technology use policy, something as [19:08] simple as are employees permitted to be [19:11] on Facebook on government computers [19:13] during their lunch hour. Is that good? [19:15] Is it not? And so all that is being [19:17] developed through our office of uh [19:20] technology. Uh as y'all probably know, [19:22] we have a fairly high-profile governor. [19:24] When as soon as Governor Sanders got [19:26] elected, the number of cyber attacks on [19:28] our state doubled overnight. And as you [19:30] all know, as a state, we've got to be [19:33] right 100% of the time. The bad guys [19:34] have only got to be right once. And so [19:37] we've done several things in in our [19:39] state to try to uh promote cyber [19:42] security because what ends up happening [19:44] is if you send out mandates, what's the [19:47] response to a mandate? Typically [19:50] push back, right? And so we took a an [19:52] entirely different tact. We actually in [19:54] the state um over the last 5 years we [19:57] spent $10 million on cyber insurance and [20:01] we only had $1.5 million in payout and [20:04] our deductible is $2 million. So we [20:07] spent $10 million for an insurance [20:09] policy that we never used. Basically it [20:11] was a waste of taxpayer dollars. And [20:13] what we found is is and and again human [20:15] nature. If you have a problem instead of [20:18] going to somebody for help, what do you [20:20] typically do? Well, you try to solve it [20:22] yourself and end up digging your hole [20:24] deeper and deeper and deeper, right? [20:25] That's typically what happens. And so, [20:26] we had to create a system in in our [20:28] state where people felt comfortable when [20:31] they ran into one of these situations [20:33] that they felt comfortable going to the [20:35] experts for help. And so, what we did is [20:37] we we told all the cities, all the state [20:39] agencies, uh, or and all the schools, [20:43] we're going to put you into this fund, [20:45] if you will, this insurance fund. It's [20:46] going to be a self self, uh, uh, [20:49] supported fund. you everyone's going to [20:52] participate in this. You're all going to [20:53] put money into it. And what this fund is [20:55] going to be, this is going to be your [20:56] insurance policy against a cyber attack. [20:59] But in order to be able to access that [21:01] fund, you have to be following all the [21:04] rules. Now, who establishes the rules? [21:06] So, our chief cyber officer and then all [21:09] of those entities all have a say so on [21:11] that board that makes what these rules [21:13] are. So, everybody has a say and then [21:15] everybody is told to follow the rules. [21:17] as long as you follow the rules. If you [21:19] get hit by a cyber attack, then we're [21:21] going to help you fix it and you So, [21:23] we're we're creating incentive all the [21:26] way down to the local library to be [21:28] involved in making sure you've got the [21:30] right um cyber security in place. And [21:34] here's just a real real simple example [21:35] of how this comes into place. North [21:37] Little Rock School District got hit with [21:39] a uh a ransomware attack and there were [21:44] bad guys wanted a $250,000 ransom. Our [21:48] experts went in. We got it down to [21:50] $160,000 ransom. Well, in order for the [21:53] school board to approve the payout, they [21:56] had to have a school board meeting. That [21:58] school board meeting had to be open to [21:59] the public. The school board approved [22:02] 250 thou up to a $250,000 payout. Well, [22:06] within an hour they got a phone call [22:07] from the bad guys. Guess what they [22:08] wanted? They wanted $250,000 and that's [22:11] what they ended up getting. So, we had [22:13] to modify our laws within our state to [22:15] allow those local entities to go into [22:18] private sessions to be able to handle [22:19] these situations. And on our technology [22:22] committee, same thing, to allow us to go [22:25] uh and handle these things in private so [22:27] that the bad guy so that we as decision [22:29] makers can hear where the [22:30] vulnerabilities are at, but we don't [22:32] tell the bad guys where the [22:33] vulnerabilities are are at. and um and I [22:37] think that's about the only time that we [22:39] can do things in private and as a [22:41] legislature out of uh the public site is [22:44] in this space. And the other thing I'm [22:46] know I know I'm kind of going down a [22:47] rabbit trail here, but I I want to give [22:51] you all some practical advice going [22:53] forward because I know a lot of the [22:54] stuff that we speak about here tends to [22:56] be theoretical high level. [22:59] A lot of the decisions that we're having [23:02] to make on the state level [23:04] are we're trying to feel our way and [23:06] figure this stuff out, right? When I [23:08] first became chair of the technology [23:10] committee, the biggest technology [23:11] decisions we were making was around [23:13] phone service uh and trying to get [23:15] broadband out. That has accelerated [23:18] dramatically in the last 10 years. And [23:20] it's becoming such a specialized field [23:23] that lawmakers, you're having to find [23:25] lawmakers who are actually interested [23:26] and specialized in this area in order to [23:29] try to put forth these policies to put [23:30] these protections in place. For example, [23:33] I have colleagues who are, you know, [23:35] they're knowledgeable. They're experts [23:36] in the medical field. I am happy to [23:38] leave that area to them. I I'm I'm [23:41] that's just an area I'm not interested [23:42] in, but I'm glad we I have colleagues [23:44] who are experts in that area to help [23:46] drive that policy. Same way technology [23:49] is becoming one of those places where [23:52] you've got to have lawmakers who are [23:54] willing to take the time to dig in to [23:56] get to know the issues so that as we're [23:58] developing policies, it's made on [24:00] knowledge rather than all the hearsay [24:02] and doomsday and all that stuff is going [24:04] out. And so here's the the the practical [24:06] advice I want to put out there to you. [24:08] You as business leaders and and small uh [24:11] small medium, it doesn't matter. [24:14] Your local officials need to hear your [24:16] voice on where we need to go with these [24:18] technologies. [24:20] And that voice needs to happen before [24:23] we're in a legislative session. When we [24:26] start our legislative session, I get up [24:29] in the morning. I drive to Little Rock. [24:31] I am there. I'm learning debating 40 50 [24:34] bills. I get home just in time to go to [24:36] bed. Maybe get a couple hours sleep. I [24:38] get up the next day. that is not the [24:40] time to try to come and persuade me on [24:44] very complex issues. At that point, it's [24:47] too late. And if I'm in a situation [24:49] where I've got, you know, I've got Bill, [24:52] I've gotten I've got a relationship with [24:54] him. I, you know, I trust him and he [24:57] comes up and he's advocating a certain [24:59] position when it comes to these [25:01] technologies. And then I have, you know, [25:03] Tom comes up the day of and says, "Hey, [25:05] you know, you can't do this." [25:08] All things being equal, who am I going [25:09] to listen to? I'm going to listen to [25:11] human nature. I'm going to listen to [25:12] that guy who I've got that relationship [25:14] with. So, I would encourage you, you [25:17] know, if you're a small business, maybe [25:18] you can't be down there all the time. [25:20] Maybe you can't hire. Get with other [25:22] small business. Create an AI consortium [25:24] in your state. Start interacting with [25:27] your lawmakers now so that way when we [25:30] get into our legislative sessions, you [25:32] can be influential. If you're waiting [25:33] till the last minute to be influential, [25:35] it it's it's it's already too late. So, [25:38] uh, that's going to be my biggest piece [25:39] of advice to you. Be a part of the [25:41] conversation. But to be a part of the [25:42] conversation, you need to start acting [25:44] with your lawmakers now. Figure out who [25:46] those lawmakers who are the ones that [25:48] are in technology and driving the [25:50] technology policy and get to know them [25:53] and develop those uh, relationships. So [25:56] very long- winded to your question, but [25:59] I I'm hoping that that will help get the [26:01] people in this room involved in the [26:03] conversations at the state levels [26:04] because right now the state levels is [26:06] where a lot of the um figuring this [26:09] stuff out on a regulatory framework is [26:11] taking place and then I think once we [26:13] get large scale consensus on the state [26:15] level, hopefully that's when the federal [26:16] government will step in and they will [26:18] start saying, "Hey, all the states think [26:20] this is a good idea, so it's probably a [26:21] good idea overall for the whole nation." [26:23] That's fantastic advice and a great [26:25] segue to our next question as well. We [26:28] see smaller companies and startups [26:30] really driving innovation, but [26:32] navigating uh the defense procurement [26:35] system is really, really hard. Um, Chip, [26:40] the Defense Alliance has helped elevate [26:43] startups for the last 20 years and help [26:45] get these voices heard. What [26:47] non-traditional expertise do these [26:50] innovators bring to the national [26:52] security ecosystem? [26:54] >> Yeah, thank you. Good question. If I can [26:55] just add one thing to the last question, [26:57] there absolutely practical tip. The [26:59] question was, if you're an engineer [27:00] trying to design a solution for an AI [27:03] solution for something in the [27:04] government, uh, one of the first things [27:06] you do is read the RFP or something, see [27:08] what are the requirements. A lot of [27:10] times you can't find the requirements. [27:12] You might get something general and then [27:14] a next logical step is okay what do you [27:16] want it to do what are the parameters [27:18] how much [27:20] you can run around for years trying to [27:22] find that don't do that uh especially in [27:25] the AI world you are already so far [27:27] above the government requirements you [27:30] you're in the driver's seat for a lot of [27:31] these things set the stage for the [27:33] government tell them what you can do and [27:35] said yes yes you did your due diligence [27:37] to try to find the requirements you [27:39] couldn't find them I'm going to set the [27:40] stage and define them for for you. So, [27:42] be bold in that way. So, back to the [27:45] whole government procurement thing for [27:46] small business. Yes, it's really tough. [27:48] Uh it's really hard. I've been doing [27:50] this for 20 years uh since I retired uh [27:53] as an aviator and I still don't quite [27:55] understand it. Every day there's [27:56] something different. It's really really [27:58] frustrating. Uh one of the things I I [28:00] always fall back on is when I was uh on [28:03] the Secretary of the Navy staff, I was a [28:05] speech writer and he was he was writing [28:07] a speech about putting women in [28:08] submarines. This is 20 25 years ago, [28:11] almost 30 years ago now. And somebody on [28:14] the staff said, "Well, sir, does that [28:17] make sense?" And you know, this is 25 30 [28:20] years ago. It's actually worked really [28:22] well. But back then, people were [28:23] scratching their heads saying, "That [28:25] doesn't make any sense." The Secretary [28:27] of the Navy looked at the staff and he [28:28] said, "Guys, uh," and it was all guys [28:31] then. He said, "We're going to do the [28:34] American thing, not necessarily the [28:35] right thing." And every now and then [28:38] those things coincide and and I just say [28:41] that to set the stage because government [28:44] contracting is not a logical system. So [28:48] get over that part of it to begin with, [28:50] right? And then fall back on the the the [28:52] luxury of the fact that you are as a [28:54] small business a non-traditional [28:57] supplier to the federal government [29:00] because you have a better risk [29:02] tolerance, you can move faster, you [29:04] innovate faster. [29:06] uh and and all of those things are [29:08] accelerated in the AI world. So by [29:11] definition you are non-traditional [29:14] and the other good news is now [29:15] especially in the past eight or nine [29:18] years and certainly under this [29:20] administration there's a concerted [29:22] effort now to say let's throw a lot of [29:24] the rules out the window and seek those [29:26] non-traditional companies. So if you [29:30] don't have prior expect expertise in [29:32] federal government, okay, that can be a [29:34] handicap but less so now. Speak about [29:37] what you can do, speak boldly about it. [29:40] Uh and and you know, use those [29:43] non-traditional things that you do. You [29:45] bring expertise from other industries [29:47] into the defense industry. That's that's [29:50] wanted now much more than ever. So, you [29:53] know, champion those things, say them [29:54] out loud, and be bold uh with these [29:58] federal procurement folks that you're [29:59] dealing with. And the last thing I'll [30:01] say about that is don't be afraid of [30:03] these guys if you're in the federal [30:05] government. Um, [30:07] they work for us, right? We pay their [30:09] taxes. So, don't hesitate to call them, [30:12] email, pester them. Obviously, be [30:14] respectful uh of their time, but uh be [30:18] demanding of those requirements that you [30:20] need to do the best possible job to get [30:22] them what they want. And let me say one [30:25] thing about that, too. Like from the [30:27] consumer on the inside, I'm just as [30:30] frustrated as you are. You know, I'm [30:32] trying I'm trying to get your technology [30:35] uh applied as fast as possible. And [30:37] there's, you know, bureaucratic hurdles [30:39] that I'm up against or I I say I was up [30:42] against, you know, when I was on the [30:44] inside. So, um, the and the programs, [30:48] you know, where I where I sat and where [30:49] all my peers sat, the programs, you [30:51] know, were pounding the desk of [30:53] procurement and and acquisitions, find [30:55] new ways to, you know, make us fast. [30:58] find new ways to get us the things that [31:00] we need because, you know, the world is [31:03] is running by us and the and the and the [31:05] government is, you know, obviously slow [31:07] to to capture. So, you know, uh there's [31:11] there's new opportunities like DIU and [31:14] and um I just saw for the first time [31:17] that I recall seeing it um the [31:20] Department of Homeland Security just [31:21] issued a commercial solutions opening [31:24] where you know, you provide your videos [31:26] and and uh your pitch and it's for um [31:30] maritime technology, but that could be [31:33] you know, if you're supporting the US [31:35] Coast Guard employees with a large [31:38] language model of some sort that might [31:41] fit. So, you know, take that's like a [31:43] little pitch for that, but just take a [31:44] look at that. And there's other [31:45] opportunities popping up across [31:47] government. They're trying to make it [31:49] more uh streamlined because, you know, [31:52] inside we're frustrated. So, [31:55] >> and and kind of further that on the [31:57] state level, something I'll throw out [31:58] there, especially if you're a small [31:59] business looking to interact with [32:01] government, a lot of times the states [32:03] can be a lot easier to work with. Now, [32:06] of course, there's 50 different [32:06] entities. And I'm from Arkansas. I know [32:10] that to a lot of people outside of our [32:12] state, Arkansas is like flyover country. [32:14] We still have some of that, you know, [32:16] oldfashioned uh country bumpkin, you [32:19] know, kind of. That's not Arkansas [32:21] today. [32:22] What I want to put out to you is [32:24] sometimes the smaller states, which [32:26] often get look overlooked in these [32:27] conversations because we don't have the [32:29] resources like a California, Florida, [32:31] Texas. But a lot of times the benefit of [32:34] the smaller states is we can be a lot [32:35] more flexible, quicker and nimble than [32:38] our larger cousins are. And so I would [32:40] encourage you in this space where that [32:43] can be a major advantage that looking to [32:45] the smaller states could be a better fit [32:48] for you than always going after the [32:49] larger states or the federal government. [32:53] And Chip, could you tell us a little bit [32:55] more about how a regional entity like [32:58] the Defense Alliance can support [33:00] companies? [33:02] >> Yeah, so there are there's a lot of [33:04] resources out there uh for small [33:06] business. You may be aware of them, but [33:08] a lot of folks aren't. There's Apex [33:10] Accelerators. Uh used to be the [33:12] procurement technical assistance center. [33:13] Now it's Apex. Uh there's regional in [33:17] innovation clusters like ours. Um well [33:19] there there were 25 now there's three [33:21] left after Doge got through with us. Uh [33:23] we're one of the surviving three in the [33:25] country but there SBA is going to build [33:26] those back up. So that that brings you [33:29] regional expertise uh folks like like [33:32] our cluster that know what the state [33:35] environment is like where the funding is [33:37] that can match federal funding and [33:39] oftentimes these resources are [33:41] absolutely free. Uh so so by all means [33:44] reach out to that. Um people are a [33:47] little scared when they say you're a [33:48] free contract or a free consultant. That [33:50] can't be true. Um it's absolutely true. [33:54] Uh in our case, we have a lot of [33:55] expertise across a wide variety of [33:58] topics, but often times that expert [33:59] expertise doesn't go very deep. uh and [34:02] in this procurement system you do need [34:04] deep knowledge at some point in legal [34:07] affairs, regulatory, accounting and [34:10] groups like ours can point you to those [34:12] people and then it does become you know [34:14] fee for service at that point. [34:17] >> And switching gears slightly um the [34:21] rapid deployment of AI in this space [34:24] raises important and ethical questions [34:28] that many have been discussing for the [34:30] last 5 years. I think there's been a [34:32] really robust discussion in the industry [34:34] but fewer decisions coming out of that. [34:38] What should policymakers and government [34:40] leaders be thinking about now to ensure [34:43] responsible and transparent use of AI [34:45] going forward? And Stephen, let's start [34:48] with you. [34:49] >> I'm glad that you asked that question. [34:50] And I was actually at a a legislative [34:53] conference a couple years ago and I [34:55] actually broached that very same [34:57] question to one of the largest um [35:00] companies, tech companies in the world. [35:02] They were doing a presentation and the [35:05] response I got back was basically we're [35:08] developing the the technology. Y'all [35:10] figure out the answer to those [35:11] questions. And unfortunately, a I think [35:15] that's the wrong way to do it. I think [35:16] they obviously need to be involved in [35:17] that. I think that's changed somewhat. [35:20] Um [35:22] but [35:23] still a lot of still a lot of questions [35:25] around that. We're still trying to feel [35:27] our way through it. Um kind of switching [35:29] away from the national defense aspect a [35:31] little bit. Um you know Utah and [35:34] Arkansas and a few other states are [35:36] looking into how do we protect our [35:39] students from social media? Um in the [35:42] last two years Arkansas along with 31 [35:45] other states have banned cell phones in [35:47] schools. Now, cell phones have been [35:48] around for a long time, but what's [35:50] changed in the last couple years? Well, [35:52] a AI and B, we're finally beginning to [35:56] understand the damages that it's doing [35:58] to our our children. And so, the states [36:00] are trying to put reasonable safeguards [36:03] in place to protect our kids and even [36:05] coming to to adults. [36:08] Uh, all of us in here, I think if I were [36:10] to were to ask you, how many of y'all [36:12] get frustrated with politics and what [36:13] goes on in Washington DC? I think most [36:15] of us would probably raise our hand, [36:17] right? But the United States, we have [36:20] the oldest government currently in the [36:23] world. There are many cultures that are [36:25] older than us, but we have the oldest [36:27] government active in the world. And one [36:29] of the reasons why we do, I believe, one [36:31] of our strengths is that we air all our [36:33] grievances in public. We fight it all [36:35] out in public. And so you don't have [36:37] these private groups festering that [36:39] leads to revolution. [36:41] By doing that, that makes things very [36:43] messy and can also very uh often lead to [36:47] us being very divisive with each other. [36:49] I think it was President Lincoln who [36:50] said that America will never be [36:52] destroyed from uh without. It'll be [36:54] because we're destroyed from within. And [36:56] I've seen videos in the last couple of [36:59] uh days where um foreign adversaries [37:04] were creating fake accounts run by chat [37:07] bots that go onto these social media [37:09] platforms. and their whole purpose is is [37:11] to create division amongst us. So, you [37:13] could be on a social media platform [37:15] arguing with a chatbot whose sole [37:17] purpose is to get you to argue with it [37:19] to inflame both sides of the political [37:21] spectrum. And theoretically, you could [37:25] have a foreign adversary [37:27] find some event that's taking place in [37:29] our country, unleash these chat bots, [37:33] get our citizens all fired up, and by [37:36] evening you've got a riot in the street [37:37] on Main Street, all driven by foreign [37:40] adversaries trying to trying to divide [37:43] us. And so I think one of the biggest [37:45] things that we're doing in our state is [37:48] we passed legislation this last session [37:50] to require uh our students in our state [37:53] to begin learning digital literacy. [37:54] We're actually going to put that into [37:56] the curriculum so that students learn [37:58] how to use these tools, but more [38:00] importantly they have the critical [38:01] thinking to understand, hey, am I [38:04] talking to a human? Am I talking to an [38:06] AI? And I think that's going to become [38:07] increasingly more difficult. And I don't [38:10] know how you get around that unless you [38:12] get to the point where we're going to [38:14] have to start, you know, doing a [38:15] national ID and everyone's going to have [38:16] to do a uh, you know, an ID in order to [38:20] get onto these social media platforms. [38:22] And I don't know that that's a good [38:23] answer either. So, uh, still a lot of [38:26] questions going on. And again, that's [38:28] why I encourage y'all that are in the [38:29] industry, be engaged to help us try to [38:33] figure out what the answers to these [38:34] are. And just like I told my legislative [38:36] colleagues in Boston, I'm old enough to [38:38] remember when the pocket calculator came [38:40] out. And I remember that the fear at [38:43] that time was nobody was ever going to [38:45] learn math anymore because of the pocket [38:46] calculator. Well, we figured that out. I [38:49] think we will figure this out as well. [38:51] And AI, I'm going to be the eternal [38:53] optimist here. Are there going to be [38:54] issues? Are there going to be bumps in [38:56] the road? Are we going to make wrong [38:57] decisions? Uh, definitely. But I think [39:01] the more voices that we have involved [39:03] hopefully the less mistakes we'll make. [39:05] And so again, just encourage you get [39:07] involved in what's going on in this [39:09] space. Don't just develop the [39:11] technology, be involved in the [39:13] regulation side by getting involved with [39:14] your, you know, your city council, your [39:17] state leaders, and also your federal [39:19] delegation. [39:22] >> And then Chiff or Brian, did you have [39:24] anything you wanted to add on the [39:25] regulatory side? [39:29] Well, I I don't know about regulatory, [39:31] but I I will say um [39:34] well, our values matter. Uh and so [39:38] you're you're designing systems and it's [39:41] going to be up to you to design the [39:42] right values into them. If if you saw um [39:45] Dr. Hinton's uh welcome address on the [39:47] first day of the conference, he talked [39:49] about, you know, building in maternal [39:51] instincts to uh to AI so it doesn't kill [39:53] us, right? Um, and similarly with [39:56] defense systems, it's going to be up to [39:58] you, not the regulatory bodies to figure [40:00] out what you're designing into these [40:02] systems that are way way smarter than [40:04] all of us combined. [40:07] And to kind of echo that, I I would just [40:10] say that you're, you know, you're [40:11] working with um people who have high [40:14] ethical standards in the US military. [40:17] And if you are not coming to the table [40:19] with your solutions that just have at [40:21] least a nod to that, um, you should and [40:24] you should think about how you present [40:25] that to them. [40:27] >> And and I'll throw this out kind of you [40:29] brought a a thought to my mind. One of [40:31] the pieces of legislation that I passed [40:33] this session was to outlaw elicit deep [40:36] fakes within our state. The first [40:38] offense, misdemeanor, you to slap on the [40:40] rest. After that, it includes jail time. [40:43] But one of the things that I added to [40:45] that legislation was a cause of action. [40:48] Uh we also one of my other colleagues [40:50] did the same thing cause of action on [40:52] the social media. And so if a [40:56] uh you know if he makes an illicit deep [40:58] fake of me and it causes me monetarily [41:01] damage, not only is he going to get the [41:03] the the fine from the government, but I [41:05] can also sue him for all my damages. On [41:08] top of that, I also added where the uh [41:12] company that hosted the image generating [41:15] software, they can be sued as well as [41:18] part of this. The reason why I did that, [41:21] the kind of a simple analogy is with gun [41:24] manufacturers. If someone uses a gun in [41:27] a harmful way, we're generally not going [41:29] to hold them accountable because we have [41:31] laws in place that uh limit access to [41:34] weapons. However, if the uh gun [41:37] companies just had guns laying all over [41:39] the place, anybody could get a hold of [41:41] them, they had all the free bullets you [41:42] could use. Well, then we're obviously [41:44] going to hold them liable for any [41:45] outcome that comes from that. So, the [41:47] same thing with these image generations. [41:49] If you put out a product to generate [41:52] these images and help the the bad guys [41:55] to create these image and destroy [41:57] people's lives and you don't put [41:59] safeguards in place, then I think you [42:01] should be partially responsible for that [42:03] outcome. So, as you're developing these [42:06] products, uh, as these gentlemen said, [42:08] you've got to make sure you're putting [42:10] these safeguards in place because if [42:12] not, you're going to share part of the [42:14] liability, uh, of whatever the user [42:18] does, uh, with that product, if that [42:21] makes sense. And I think on and I think [42:23] you'll probably start see I'm hoping [42:25] that will spread across the the country. [42:28] >> And as our last question today, this has [42:30] been a fantastic conversation. Thank you [42:33] all. What is one priority area where you [42:35] believe AI will play a transformative [42:38] role in the future and what do we need [42:40] to do to get it right? Ryan, we'll start [42:42] with you and go down. [42:45] >> Yes. So, I I I don't think that uh AI is [42:48] going to, you know, replace us. I think [42:49] that AI is going to be our partners and [42:52] and what already what I think I'm seeing [42:54] are um the youth of our organizations [42:57] are partnering with AI and just look at [43:00] chat GBT. I mean there's they're they're [43:03] coming up you know our young service [43:05] members are coming up with um answers to [43:07] questions that otherwise would have been [43:09] very you know quite timeconuming quite [43:11] quickly they're understanding how to be [43:13] partners I think that we um you know the [43:16] the older generation as well need to do [43:18] that as well those who don't partner [43:20] with AI think about the MRI there were [43:23] doctors one day that didn't use MRIs and [43:26] then the MRI came out all the doctors [43:28] that don't use MRIs they're useless [43:30] they're gone. So, the same thing's going [43:32] to happen if we don't understand and [43:34] figure out how to partner with AI. [43:37] >> And for me, it's making uh have giving [43:39] warriors the ability to make better [43:41] decisions on the battlefield, faster, [43:43] quicker, better judgment. If you can [43:45] bring that to them, that's huge. And the [43:46] last thing I'll say is anybody see the [43:48] movie Zero Dark 30? There's that one [43:51] scene where the CIA chief comes from DC [43:53] to Pakistan to meet with his team in [43:55] Pakistan on the ground and he he goes [43:57] off on him. It's a great scene if you [43:58] ever seen it. and he says, "If you think [44:00] there's some other cell solving this [44:02] problem on another floor, there isn't. [44:05] It's you." And and that's that's the [44:08] thing here. Don't look to the Pentagon. [44:10] Don't look to DHS. I mean, you have to [44:12] to some degree, but the solutions are [44:14] here. You all know what the technology [44:17] can do and bring. So, don't wait for [44:19] somebody to bring it. Bring it to the [44:20] table yourselves. [44:22] And kind of echoing that on the state [44:24] level where I see AI coming into is [44:26] making government smarter, faster, more [44:29] efficient. One of the one of the laws [44:30] that I that I in uh passed in that AI [44:35] guard rail policy that I told the state [44:37] agencies to set up [44:39] is I also included legislation that said [44:42] a human has always got to be the final [44:44] decision maker because none of us wants [44:45] to call a government agency and say, [44:47] "Hey, sorry I can't help you. The [44:48] computer said so." And I as a lawmaker [44:50] don't want to have to go to uh a uh [44:53] employee and have them say it's not my [44:55] responsibility the computer set. We [44:57] always want a human in the loop. And I [44:59] think where AI comes in is AI will help [45:01] make their jobs quicker, smarter, faster [45:03] so that as as our citizens are [45:05] interacting with state government that [45:07] it's a much cleaner, more efficient [45:09] process. Uh, you know, we've all stood [45:11] in the line at the DMV wondering why [45:13] we're having to waste our time here or [45:15] if you call, you're trying to get [45:17] unemployment or some other benefit. [45:19] Sometimes it can take forever. And part [45:22] of the problem is trying to get enough [45:23] employees to even fill those positions. [45:26] And so, if y'all have AI solutions that [45:28] can come help those employees work [45:31] better, smarter, faster, easier, I think [45:34] that would be, you know, the states [45:35] would would welcome and eat that up. [45:37] >> Well, thank you so much for joining us [45:39] today. Thank you all for joining and we [45:42] appreciate AI4 for having this [45:43] conversation.