Israel Aerospace, Loft Orbital Push Tactical AI to Battle Edge

以色列航太与 Loft Orbital 推动战术 AI 迈向作战前沿 国防领袖正将 AI 从云端系统转移至轨道与战术边缘硬件，以消除通信瓶颈并加快战场决策。 国防与航天专家深入探讨了国防 AI 从实验性技术向战术、机载及太空领域战备级关键任务部署的转变。专家们详细阐述了多项实际落地应用，例如用于边缘指挥的“盒装参谋”（staff in a box）计算平台、实时轨道图像处理，以及人在回路的自主蜂群系统。此外，他们还探讨了操作员信任机制、确定性防护边界、克服军方体制惯性等现实挑战，并展望了量子与 AI 融合以及自主编排的主权卫星星座等未来前沿方向。 国防领域正加速将 AI 推向战术边缘与指挥控制（C2）系统，Legion Intelligence 的 Centurion 等紧凑型硬件能够在通信受阻的极端战场环境下充当“盒中参谋”，协助指挥官在本地独立完成作战推演与决策。 (5:02) Loft Orbital 等机构将边缘计算与视觉语言模型（如 Gemma 3）部署至在轨卫星，打破了传统数小时的极地地面站下传延迟，实现数十分钟内自主识别目标并实时下发轻量化战术警报。 (20:34) 为建立战场操作员对 AI 的信任，以色列航太工业（IAI）结合真实作战数据建立混合仿真与可解释性验证管道，并在自主无人蜂群任务中严格恪守“人在回路”（HITL）原则，确保动力打击由人类最终授权。 (23:50) 未来防务智能的关键在于软硬件周期的解耦与模型量化，指挥官未来可通过自然语言统一调度分布式卫星星座，同时推动确保数据与模型自主可控的“主权 AI”。 (41:43)

Defense AI is shifting from speculative software experimentation into resilient, edge-deployed mission infrastructure, where operational advantage depends on compressing sensor-to-decision timelines across disconnected tactical nodes and orbital constellations. Edge Computing in Contested Tactical Environments (5:02) Modern dense battlefields expose warfighters to severe cognitive saturation from multi-domain threats, including loitering munitions, anti-tank guided missiles, and electronic warfare. Eytan Eshel of Israel Aerospace Industries (IAI) explains that embedding AI into Command and Control (C2) architectures serves as critical decision support, synthesizing vast sensor feeds into a coherent situational picture to identify adversary positions and plot safer routes. To ensure operational continuity when external communications are severed, Chris Hume of Legion Intelligence highlights the deployment of its Centurion platform onto compact hardware—such as NVIDIA Jetson- or IGX Orin-class devices with 128 GB of unified memory. Tested at the U.S. Army National Training Center during Project Convergence, these localized 'staff in a box' systems execute the Military Decision-Making Process (MDMP) directly at the battalion level, falling back on local data stores and quantized models when cloud links are jammed. Orbital Edge Inference and Constellation Orchestration (20:34) Conventional Earth observation workflows face severe latency bottlenecks, typically constrained by a six-hour Service Level Agreement (SLA) while satellites await polar ground-station passes to downlink hundreds of gigabytes of raw data. Paul Lasserre of Loft Orbital explains that executing AI inference directly on orbit reduces alerts to mere kilobytes of telemetry—such as coordinates and cropped image thumbnails—transmitted instantaneously over geostationary relays. In dual-use missions like French wildfire tracking, this slashes notification latency from two hours to tens of minutes. Lasserre also points to a transition from rigid Convolutional Neural Networks toward Vision-Language Models (VLMs), citing orbital validation of Google's Gemma 3 with NASA and planned testing of Gemma 4. This compute layer decouples software cycles from satellite hardware, allowing modern foundation models to fly on legacy sensors. In broader space operations, functional models act as orchestration layers: an initial low-resolution satellite detects an anomaly, which automatically cues high-resolution assets to verify and cross-fuse target data based on natural language queries. Provenance, Explainability, and Calibrated Operator Trust (23:50) Operator trust remains the decisive gate for military AI fielding. Eshel emphasizes that defense systems cannot rely on synthetic commercial data; IAI utilizes decades of proprietary operational datasets and hybrid simulation environments where pilots stress-test algorithms against synthetic threat injects. Even within highly coordinated autonomous drone swarms, IAI preserves a strict human-in-the-loop (HITL) architecture, requiring explainable confidence metrics before an operator authorizes kinetic engagement. Lasserre frames mission trust as an unbroken 'chain of custody from photon to alert,' comparing the defense sector's adaptation to probabilistic models to healthcare's decade-long journey adopting AI in radiology. Hume argues that deployment speed must be calibrated against operational risk: low-consequence tasks can automate quickly, whereas high-consequence operations require transparent audit trails and deterministic rules wrapping probabilistic model outputs. Institutional Inertia, Staff Silos, and Sovereign Architecture (31:26) Institutional culture often lags behind technical readiness. Lasserre draws an analogy to 19th-century naval forces that maintained canvas sails on steamships for nearly a century due to doctrinal habit, though active conflicts in Ukraine and the Middle East are accelerating the adoption of autonomous capabilities under wartime pressure. Hume observes that AI synthesis is beginning to dismantle the centuries-old Napoleonic staff system, eroding the institutional boundaries separating intelligence (J2) and operations (J3) by piping fused insights directly into live tactical decision cycles. Looking toward long-term deployment, Hume emphasizes growing demand for 'sovereign AI,' where nation-states insist on total control over their data, model weights, and compute stacks. Panel moderator Christine Keung of J2 Ventures notes that legacy prime contractors no longer maintain a monopoly over defense innovation, with commercial and open-source models offering agile solutions to manage data proliferation—such as processing the one petabyte of data generated daily by a single aircraft carrier. While high-stakes conflict accelerates the real-world fielding of edge and orbital AI, broader military institutionalization remains constrained by legacy doctrines, procurement inertia, and strict validation requirements. The key technological challenges going forward center on deploying performant quantized models on edge hardware, ensuring absolute data provenance, and orchestrating distributed autonomous constellations without severing ultimate human command authority.

在此次 Ai4 研讨会上，来自 J2 Ventures、Legion Intelligence、以色列航空工业公司（IAI）和 Loft Orbital 的与会专家指出，国防人工智能正从实验性技术转向实战关键能力，其核心在于将算力推向战术与太空边缘，并在通信受限与高危作战环境中建立可验证的人机信任与主权控制。 算力下沉战术边缘缓解战场认知负荷 (5:02) 现代高烈度战场中，密集的无人机蜂群、反坦克导弹与简易爆炸装置使士兵面临严重的认知过载。IAI 执行副总裁 Eytan Eshel 表示，将 AI 整合到指挥控制（C2）系统中，能够实时消化多传感器输入以生成态势图并绘制更安全的行动路线。Legion Intelligence 首席业务官 Chris Hume 介绍了其 Centurion 产品线，该系统采用仅有小型烤面包机尺寸的紧凑硬件（配备 128 GB 统一内存，如 NVIDIA IGX Orin 或 Jetson 级系统），为战术指挥官提供“盒装参谋”能力。在美军欧文堡国家训练中心（NTC）的“融合工程”（Project Convergence）测试中，该系统证明了即使在云连接断开的拒止环境下，也能通过本地量化推理模型支持营级参谋执行军事决策流程（MDMP）。 星载边缘计算与协同打破卫星情报延迟 (20:34) 传统低轨对地观测卫星面临严重下行瓶颈，过去需等待飞越极地地面站才能回传数百 GB 原始数据，使多国防务部门的情报服务等级协议（SLA）延迟长达约 6 小时。Loft Orbital 太空 AI 总经理 Paul Lasserre 指出，通过在轨运行边缘计算模型，卫星可实现机载数据处理，并借助地球同步轨道（GEO）低带宽中继仅下发包含目标分类、坐标及置信度的几 KB 遥测警报，将野火检测等通知时间缩短至几十分钟，并实现对无人机或飞机的跨域实时引导。此外，卫星任务正从传统 YOLO 这类刚性卷积神经网络转向视觉语言模型（VLM）；Loft Orbital 与 NASA 合作在轨验证了 Gemma 3 并推进 Gemma 4 部署，使指挥官能用自然语言直接向太空资产下达抽象巡逻任务。Lasserre 还强调，通过解耦卫星硬件与软件更新周期，五年前发射的旧传感器亦可搭载最新基础模型运行分布式编排任务。 以严格验证与人在回路重塑操作员信任 (23:50) 与追求部署速度的民用软件不同，防务 AI 在生死关头以弹性、安全性和战备状态为衡量标准。Eshel 强调，建立操作员信任必须依赖严格的验证与确认（V&V）、可解释 AI（XAI）以及数十年的机密实战数据集训练，而非仅依赖合成或开源数据；即使自主蜂群能完成搜索与机动，IAI 仍坚持在实施动能打击的最终决策中保持“人在回路”（HITL）。Lasserre 将操作信任概括为“从光子到警报的完整监管链”，必须确保每一个战术提示都能溯源至原始传感器通道、算法架构与置信度；他认为防务领域对概率性输出的接受过程类似于医学影像从确定性诊断向 AI 辅助转变的长达十年的适应期。Hume 则补充，系统部署速度必须依据任务后果进行分级，低后果任务可快速迭代，高后果动能行动则必须设置严密的人工审查节点。 打破编制壁垒与技术收敛下的未来战争 (31:26) AI 工具的深度应用正在冲击军事机构的传统惯性。Hume 分析，经典拿破仑式参谋体系长期将情报（J2）与作战（J3）明确分离，但随着 AI 工具直接将分析融合至实时决策闭环，情报与作战之间的边界正在模糊，推动指挥机关组织结构的重组。在技术演进层面，Eshel 指出未来十年 AI 与量子计算的融合将成为战略倍增器，量子技术将在气动设计与新材料发现、超灵敏传感器以及抗破译的量子密钥分发（QKD）通信网络中发挥核心作用。与此同时，防务客户对确保数据和模型全权自主的“主权 AI”以及能按需发射定制星座的“按需卫星”需求正在显著提升。研讨会主持人 Christine Keung 总结认为，传统大型防务总承包商的技术垄断正在被打破，面对单艘航空母舰每天产生高达 1 PB 数据的算力压力，军方正日益转向开源技术与商业创新团队提供的新型加密及边缘计算方案。 本场研讨表明，边缘 AI 正从根本上重塑战术指挥与太空侦察的响应速度，但军事机构在摆脱历史教条、适应概率性决策以及构建严密监管链方面仍处于早期阶段。如何在战时快速部署需求与维持人类伦理审查之间取得平衡，仍是各国防务力量需要长期面对的制度与工程挑战。

## Background 本内容由 YouTube 频道 Ai4 于 2026 年 8 月 21 日发布，题为《Command, Control & Compute: AI at the Defense Frontier》。视频由 J2 Ventures 通用合伙人 Christine Keung 主持，汇聚了以色列航天工业公司（IAI）、Loft Orbital 和 Legion Intelligence 的防务与航天领袖，探讨 AI 如何重构任务规划、边缘战术作战、在轨太空计算与指挥控制系统，同时兼顾复杂环境下的安全与合规挑战。 ## TL;DR 研讨会深入探讨了 AI 从实验阶段迈向实战化国防与太空核心支柱的演进历程。嘉宾们分享了在拒止对抗环境下的边缘端部署成果，包括将推理模型与业务流整合的 Centurion 便携式“智能参谋”，以及将低轨卫星海量数据在轨提炼为秒级告警的太空边缘计算和视觉语言模型（VLM）。针对前线部署，讨论重点剖析了“人在回路”、全流程可审计的监管链与可解释性对建立士兵信任的重要性。最后，专家们展望了由软硬件解耦的“提示与排队”（Tip and Queue）分布式卫星星座、量子 AI 以及打破传统军工巨头垄断的非纯硬件创新所驱动的未来国防格局。 ## Atmosphere & style 整场研讨会节奏紧凑且充满战略紧迫感，交织着前线战术实战经验与硬核前沿科技的深度碰撞。主持人敏锐且富有穿透力地串联话题，三位具备退役军官与顶级军工高管双重背景的嘉宾发言沉稳专业、直击痛点，毫无浮夸的商业吹嘘。现场氛围既展现了对战争复杂残酷现实的深刻敬畏，又洋溢着对边缘多模态模型、太空分布式计算等颠覆性技术突破的务实雄心，给人以置身现代国防科技最前沿指挥部的沉浸感与思辨深度。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=4s) 国防 AI 已跨越单纯实验阶段成为决定胜负的战略优势，其衡量标准不在于交付速度，而在于极高风险环境下的可信度、系统韧性与实战即战力。 - [01:57](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=117s) 国防与航天领域的三大核心优先级聚焦于：在通信拒止环境中部署边缘 AI、将开源前沿成果安全迁移至本地环境，以及推动卫星从传统传感器向自主智能节点转型。 - [05:06](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=306s) 轨道边缘计算与星上直接推理彻底改变了太空情报时效，通过将数百 GB 的遥感原始数据瞬间浓缩为 KB 级告警元数据，将情报传输与响应延迟从 6 小时压缩至秒级。 - [05:06](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=306s) 部署在紧凑硬件上的“装在盒子里的参谋部”实现了推理模型与军事决策工作流（MDMP）的本地运行，即使在网络被干扰断开时也能平滑降级至本地小模型维持指挥研判。 - [14:14](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=854s) 边缘计算在战术层面构建了通用作战图景（COP），不仅显著减轻前线指挥官的认知负荷，更将原属于营旅级参谋部的深度决策分析能力直接下沉赋能至一线战术班组。 - [18:25](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=1105s) 建立人机信任是防务 AI 真正落地的根本前提，必须依托严苛的形式化验证确认（V&V）、可解释 AI（XAI）以及融合真实数据的混合模拟器，让作战人员充分理解算法决策依据。 - [20:34](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=1234s) 卫星边缘智能正从单一特征检测模型跃迁为在轨部署的视觉语言模型（VLM），支持操作员下达模糊自然语言指令并由模型自主完成多模态推理与异常告警。 - [23:50](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=1430s) “人在回路”（Human-in-the-loop）是自主作战不可逾越的底线，AI 主要负责触发线索（Cueing）和并发管理作战单元，而最终交战与开火权限必须始终由人类掌握。 - [23:50](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=1430s) 构建信任依赖“从光子到告警”的全流程可审计监管链，同时前沿模型在权威防务数据上的锚定正在加速打破传统参谋体系中情报（J2）与作战（J3）的职能壁垒。 - [32:58](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=1978s) 一线人员的采纳逻辑存在战时与平时分化：和平时期需通过低后果任务展示成效以积累信任，而在战时高压与生存威胁下，迫切需求会倒逼前沿自主技术的超常速部署。 - [39:24](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=2364s) 量子计算与 AI 的深度结合（Quantum AI）被视为未来十年的战略高地，将大幅压缩武器研发与材料模拟周期，并重构绝对安全的抗截获通信网络。 - [41:43](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=2503s) 基于“提示与排队”（Tip-and-Queue）机制的分布式太空计算与软硬件解耦架构，使得数年前发射的硬件载荷可无缝更新并运行最新模型，LLM 则充当跨星座任务编排层。 - [43:17](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=2597s) 防务 AI 重心正由盲目追求大参数转向边缘模型量化与微型化部署，同时各国军队对确保数据所有权与核心分析算法自主可控的“主权 AI”诉求日益迫切。 - [44:45](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=2685s) 国防采办格局正在打破传统军工巨头的垄断，向更广泛的科技生态、开源架构以及利用新型密码学等非纯硬件途径解决海量数据瓶颈的创新方案全面开放。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=4s) 研讨会开场：AI、太空与国防的深度交汇 会议在充满紧迫感与前沿碰撞的氛围中拉开帷幕。主持人 Christine Keung 说明了自己的身份，她是 J2 Ventures 的通用合伙人，专注于投资兼顾军用与民用的双重用途技术（Dual-use technologies），而太空领域正是其核心投资板块之一。 Christine 强调，当前时刻选择举办这场研讨会堪称“完美”。过去一周的行业动态清晰表明了技术演进的烈度： * 中国公司月之暗面（Moonshot AI）发布了全新的开源大模型 Kimi，再次在全球范围内点燃了中美之间的 AI 军备竞赛； * 就在会议开始前 9 小时，SpaceX 完成了 IPO 后的首次财报电话会议，虽然其股价已从高点回落了约 50%，但市场极其剧烈的波动恰恰凸显了资本与军工界对该领域的高度关注。 Christine 总结指出，AI 在国防领域早已脱离了单纯的实验阶段，成长为一种决定胜负的“战略优势”。它帮助指挥官加快决策、预测战机故障、提升卫星情报质量，并强化复杂环境下的作战执行。在残酷的战争现实中，系统部署的成功与否不取决于“交付速度”，而取决于系统在极高风险下是否**足够可信、韧性强、安全且具备实战即战力**。 --- ### [01:57](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=117s) 嘉宾自我介绍与核心 AI 战略优先级 随后，三位来自顶尖军工企业、国防技术初创公司和前沿航天平台的专家依次介绍了个人背景及各自机构当前最核心的 AI 发展优先级： 1. **Chris Hume（Legion Intelligence 首席业务官）：** * **背景：** 拥有 7 年 AI 与国家安全交叉领域的深耕经验；服役出身，曾先后担任美国陆军步兵官（Infantry Officer）及陆军特种部队官（Special Forces Officer）。 * **当前核心优先级：** 致力于将**具备 AI 能力的使命应用部署至边缘端（Edge）**，重点保障在拒止及整合对抗环境（Denied Integrated Environments）下的通信与系统韧性。 2. **Eytan Eshel（以色列航天工业公司 IAI 执行副总裁兼首席技术/研发官）：** * **背景：** IAI 是以色列最大的航空航天与国防企业；其个人曾任以色列国防部（MOD）军用研发负责人，退役前为准将（Brigade General），现为预备役人员。 * **当前核心优先级：** 面对两大核心挑战——一是如何将通用开源生态中所有的 AI 技术优势安全地迁移到**本地部署（On-Premise）环境**中；二是如何打造高效且高可靠性的**人机协作（Manned-Unmanned Teaming, MUM-T）**体系，实现自主系统与有人装备的深度战术协同。 3. **Paul Lasserre（Loft Orbital 太空 AI 事业部总经理）：** * **背景：** 拥有横跨美国陆军、以色列空军以及法国海军（Ex-French Navy）的独特跨国军旅背景。 * **公司概况与优先级：** Loft Orbital 是一家成熟的 Series 级 Scale-up 企业，已累计融资超 3 亿美元，去年营收突破 1 亿美元，手握 5 亿美元订单，在美、法、阿联酋拥有 350 名员工。公司核心业务是在标准化卫星上搭载第三方的载荷或软件。其 AI 优先级在于**彻底颠覆传统卫星仅作为“遥感相机/雷达”的旧模式，打造能够在轨道边缘端直接进行自主思考与分析的“智能卫星”**。 --- ### [05:06](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=306s) 真实部署的 AI 实战能力与解除战术痛点 在讨论实操案例时，嘉宾们分享了已经投入生产或正处于核心实战评估阶段的 AI 技术，阐述了其解决的战场痛点及实际成效： * **IAI 的 AI 增强型指挥控制（C2）系统：** * **解决的战场痛点：** 现代化密集的现代战场充斥着来自四面八方的威胁，包括无人机、轻武器、反坦克导弹及简易爆炸装置（IED）。战士需要消化极其海量的数据。 * **实际成效：** 引入 AI 技术驱动的 C2 系统后，能够自动消化和处理极其繁复的信息，自动研判敌方位置与安全通行路线，为战士及空中飞行员提供清晰的态势感知与实时决策支持。这不仅显著提升了战术效率，更直接成为了“挽救生命”的实战利器。 * **Legion Intelligence 的 Centurion 硬件与软件整合平台：** * **硬件形态：** Centurion 产品线将专用硬件与软件平台结合，不仅能部署在 Intel 笔记本电脑上，也能运行在拥有约 128GB 统一内存、仅有“微波炉/面包机”大小的 NVIDIA GPX Spark 边缘设备上。 * **功能与实战演练：** 该设备本质上是一个**“装在盒子里的参谋部”（Staff in a box）**。它将推理模型（Reasoning Models）与确定性工作流相融合，让军事决策过程（MDMP）在边缘端本地运行。当云端连接断开时，系统能够无缝平滑降级（Graceful fallback）至本地小模型与本地数据，保障作战参谋持续研判变化背后的“原因”而非仅仅是“事实”。 * **演练案例：** 团队近期参加了在欧文堡（Fort Irwin）美国陆军国家训练中心举办的“融合计划”（Project Convergence）演练，成功支持了营级参谋部在高度拒止环境下的连续作战研判。 ```mermaid flowchart TD accTitle: Centurion 边缘端混合架构决策流程 subgraph Cloud["云端/高阶层 (Cloud / High Echelon)"] A["大语言/推理模型与全局数据库"] end subgraph Edge["边缘端/战术单元 (Edge / Tactical Unit)"] B{"通信状态检测\n(Connectivity Check)"} C["高性能推理模型\n(例如：NVIDIA GPX Spark / 128GB)"] D["本地轻量化小模型与本地数据\n(Graceful Fallback)"] E["Centurion 智能参谋系统\n(Staff in a Box)"] F["指挥官战场态势感知与战術决策\n(MDMP Workflow)"] end A -- "高带宽连接时同步" --> B B -- "网络正常" --> C B -- "网络被干扰/降级" --> D C --> E D --> E E --> F ``` * **Loft Orbital 的太空边缘计算与实时 C2 革命：** * **传统模式痛点：** 过去几十年，低轨（LEO）卫星采集到数百吉字节（GB）的数据后，必须等待飞越极地地面站（Ground station）才能下传，地面分析师或 AI 拿到数据时往往已有 6 小时的延迟（SLA）。 * **边缘 AI 模式革新：** 结合**边缘 AI 芯片**、**地面高带宽通信**与**地球同步轨道（GEO）卫星低带宽实时链路**（如 Inmarsat）。边缘 AI 模型在卫星端直接完成图像处理与目标识别，将数百 GB 的原始图像瞬间浓缩为仅有几千字节（KB）的核心元数据——包括时间、坐标、风险类型、置信度及裁剪压缩的目标小图。 * **实际影响：** 下游战术网能够实时获取关键预警，并立即调度无遥控无人机、战机或舰船进行跟进，彻底改变了太空情报到战术指挥（C2）的响应时效。 * **卫星数据情报的软硬结合（IAI 补补充说明）：** 在后方的算力中心，面对来自光电（EO）、合成孔径雷达（SAR）及信号情报（SIGINT）卫星下传的浩瀚数据，AI 算法能够迅速精准地定位目标并标记给分析师，实现精确定位与快速分发。 - **百亿级太空产业数据：** Loft Orbital 现已完成超过 3 亿美元融资，去年营收突破 1 亿美元，累计签约合同达 5 亿美元。 - **边缘算力规格：** Legion 部署的 Centurion 平台支持 128GB 统一内存的微波炉大小边缘硬件，实现了“参谋部降级”本地计算。 - **情报时效突破：** 太空边缘 AI 成功将原本长达 6 小时的数据传输与分析周期（SLA），缩短为秒级的千字节（KB）级别实时预警。 --- ### [14:14](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=854s) 边缘计算在极端对抗环境下的实战价值 Christine Keung 结合 J2 Ventures 的投资经验引导大家深入讨论边缘 AI。她提到，J2 投资组合中有三分之一集中在太空和高级计算领域（如致力于自由空间光学通信的 Yria 以及开发替代性 PNT 芯片级原子钟的 Mech Quantum）。Christine 提问：对于那些不了解军事场景的人来说，边缘计算在战术层面究竟意味着什么？ **Chris Hume 的战术视角（以作战小队为例）：** Chris 回顾了自己多年前担任特种部队小分队指挥官（Detachment Commander）的经历。在真实的战场上，信息极其有限，而所有的作战决策本质上都是**风险管理**。 * 指挥官常常在移动中、低能见度、甚至受到火力压制的极其恶劣的环境下，同时指挥多方战术力量和盟军。 * 边缘 AI 的价值在于建立**通用作战图景（Common Operating Picture, COP）**，显著降低指挥官的认知负荷（Cognitive load），且无需依赖远程后方数据中心（Reachback dependencies）。 * Centurion 平台通过自动处理输入的无线电语音流和动态数据，帮助指挥官将抽象的“指挥官意图”（Commander's Intent）快速转化为具体的战术行动，甚至将原先仅属于营级（Battalion）或旅级（Brigade）参谋部的复杂决策分析能力，直接“下沉”下放给最前线的战术班组。 --- ### [18:25](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=1105s) 可解释 AI、人机信任与复杂模拟环境的构建 **Eytan Eshel 谈“建立人机信任”（Trust & Explainability）：** Eytan 指出，在生死攸关的国防应用中，核心关键词是**“信任”（Trust）**。如果系统无法获得战士或飞行员的信任，再先进的技术也形同虚设。 * **飞行员的真实困境：** 驾驶高科技战斗机的飞行员面临极其复杂的威胁环境——空对空导弹、地对空导弹、高强度电子干扰（EW）及雷达锁定。战术决策支持系统必须高度可信，才能帮其分担压力，使其专注于核心任务。 * **验证与可解释性（V&V and Explainable AI）：** 为获得使用者的信任，必须经过极其严苛的验证与确认程序，同时引入**可解释 AI（XAI）**，让飞行员清楚了解算法为何做出此项判断。 * **混合模拟器（Hybrid Simulators）：** 为了在安全环境下测试 AI，IAI 打造了将真实战场数据与模拟数据深度融合的“混合模拟器”。飞行员直接参与到 AI 功能的迭代开发中，在逼近真实战场的环境下亲自测试和验证算法，从而在正式进入战场前确立对 AI 系统的绝对信任。 --- <ElicitationsGroup message="如果您想深入了解本场研讨会的后续内容或特定主题，请选择以下方向："> <Elicitation label="了解 Part 2 研讨会内容" query="请提供该研讨会 Part 2 录音文本的详细整理与解读。"/> <Elicitation label="深入分析军用边缘 AI 架构" query="请详细拆解国防领域中边缘 AI (Edge AI) 硬件、模型压缩及抗干扰通信的具体技术架构。"/> <Elicitation label="探索太空 AI 与智能卫星应用" query="请深入介绍 Loft Orbital 等公司如何利用太空边缘计算 (Space Edge Computing) 改变遥感与 C2 流程。"/> </ElicitationsGroup> ### [20:34](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=1234s) 边缘 AI 与太空视觉语言模型的前沿演进 Paul Lasserre 紧随 C2（指挥与控制）系统的话题指出，可解释性是指挥控制的核心所在。几天前 Google 模型在特定生成数据上出现幻觉的事件再次提醒人们，如果没有透彻的理解与信任，作战人员就无法将关键威胁放心交给系统自动处理。Loft Orbital 在过去两三年间与 NASA 紧密合作，在两颗全 AI 驱动的卫星上运行边缘智能（AI at the edge）。在民用和军民两用领域，这种边缘推理能力带来了翻天覆地的变化——例如在山火探测中，传统方式无法在每片森林部署摄像头或哨兵，卫星边缘检测将发现时间从两小时缩短至几十分钟；类似的应用还包括甲烷泄漏排查与海洋污染监测。在军事领域，他们与 Helsing 开展公开任务，涵盖 GPS 干扰源检测、射频（RF）异常检测以及海上态势感知（MDA），实现在轨实时发现、识别目标并驱动决策流程。 更具突破性的是，Loft Orbital 与 NASA 联合发表了关于首个在轨运行的视觉语言模型（VLM）的研究，已成功部署 Gemma 3 模型，并正在推进 Gemma 4 的下一代版本。Paul 强调，过去的边缘 AI 范式主要依赖轻量级 CNN 或针对特定目标训练的 YOLO 模型，只能回答预设问题（例如判断画面中是否有山火或伊朗的弹道导弹发射架）。而 VLM 具备通用人类知识与多模态推理能力，使得操作人员能够提出高度模糊的宏观指令，例如“帮我监控这条边境，发现任何可疑情况立即汇报”。模型能够自主理解卡车集结或部队异动等潜在异常并生成告警，再由操作人员调度立方星进行高分辨率跟进侦察。 ```mermaid flowchart TD accTitle: 卫星边缘 AI 架构演进 subgraph Traditional["传统卫星边缘推理 (CNN / YOLO)"] T_In["卫星传感器捕获数据"] --> T_Model["轻量级单一目标检测模型"] T_Model --> T_Out["固定规则输出 (是/否存在特定目标)"] end subgraph Frontier["新一代多模态推理 (在轨 VLM 如 Gemma)"] F_In["多源遥感 / 高光谱传感器数据"] --> F_VLM["太空边缘 VLM (具备世界知识与推理能力)"] F_Query["操作员模糊指令 (如 '监测边境异常')"] --> F_VLM F_VLM --> F_Alert["自主态势评估与异常告警 (Cue)"] F_Alert --> F_Action["调度高分辨率侦察资源进行复核"] end ``` - 传统边缘识别模式依赖针对单一特征训练的专用小模型（如 CNN/YOLO），泛化能力极其受限。 - 引入太空部署的 VLM 后，卫星具备了上下文推理能力，支持自然语言模糊查询与多源态势综合判断。 - 军民双重应用场景包括火灾侦测、射频干扰源（GPS Jammer）定位、甲烷外泄与海上舰船识别追踪。 --- ### [23:50](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=1430s) 部署速度与信任构建：验证、责任链与人在回路 主持人 Christine Keung 将讨论引向核心矛盾：在战略需求日益迫切的背景下，军方应如何在快速部署新技术与维护系统信任、弹性之间取得平衡。Eytan Eshel 指出，防务领域的自主系统（如无人机蜂群或导弹）远比商业系统复杂，涉及严格的伦理、安全性与严格的边界控制。为了确保模型高度可靠，不仅在开发阶段需要严苛的网络安全防护，在训练时更必须依赖数十年积累的真实作战数据，而非仅依靠开源或仿真数据。尽管形式化验证与确认（V&V）工具的测试周期远长于民用领域，但这是不可逾越的底线。Eytan 重申，无论自主化程度多高，“人在回路”（Human-in-the-loop）是绝对前提，AI 负责支撑并发管理成倍增长的作战单元（如单个操作员控制整个无人机蜂群），但最终的交战与开火权限始终由人类掌控。 Paul Lasserre 进一步从数据架构角度阐述，信任本质上是“从光子到告警的监管链”（A chain of custody from photon to alert）。指挥官在收到告警时，必须能够追溯其原始传感器、卫星过境批次、所用模型版本、训练数据集以及置信度评分，并能够随时审计底层原始数据。同时，军事领域正在经历医疗行业曾历经十年的范式转变——从确定性的人工判断转向概率性系统的辅助。操作人员需要建立“AI 提供线索，人类进行决断”（AI should cue and humans decide）的认知，并建议军方从“无悔行动”（No-regrets actions，如火灾与污染监测等低后果场景）切入，逐步积累对系统的信心。 Chris Hume 与 Paul 产生共鸣，强调信任构建应根据任务后果的严重性采取分级控制（Graduated controls）。对于后果可逆、边界明确的低风险任务，应当加速迭代；对于高风险行动，则必须引入更高频的人工干预。Legion Intelligence 的策略并非盲目从头训练模型，而是确保通用前沿模型锚定（Grounding）在权威防务数据源上，并在全流程保留审计追踪机制。Chris 指出，技术扩散正在打破百年来拿破仑时期的参谋体系界限（如 J2 情报与 J3 作战的传统割裂），实现情监侦与行动执行的深度融合。作为前海军军官，Paul 补充了一个经典的历史隐喻：在19世纪蒸汽机技术成熟后，由于传统水手对风帆的情感与信任惯性，带有风帆的混合动力军舰在世界上存在了近一个世纪之久；防务 AI 的采纳也将经历类似的新旧范式长期共存。 - **监管链追溯**：要求从最初的光子信号输入到最终生成的作战警报，全程具备模型版本、置信度与原始数据的可审计性。 - **人在回路控制**：将 AI 定位为“线索触发器”（Cueing），由人类指挥官最终核准开火及重大行动。 - **组织结构重塑**：AI 辅助决策正在加速打破传统 J2（情报）与 J3（作战）之间的职能壁垒，推动跨部门协同。 --- ### [32:58](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=1978s) 推动一线采纳：实战化赋能与 AI 素养培养 讨论转入一线作战人员对 AI 的实际采纳与反馈。Eytan Eshel 强调，CTO 的职责绝不仅是研发先进系统，如果前线部队不用，那就是彻底的资源浪费。战场充斥着高度不确定的“战争迷雾”，IAI 通过深入实战现场向指挥官证明，AI 系统能显著增强态势感知并缩短对威胁的响应时间。一旦飞行员和地面部队亲身体验到系统带来的生存率与杀伤力提升，他们就会像告别风帆一样，彻底放弃旧式操作模式。 Chris Hume 观察到当前军队内部对 AI 的理解存在广泛分化，但一股“AI 原生”的新生代力量正在快速成长。例如美军内部推广的 NIPRNet 级安全环境应用（如 genai.mil），让官兵在合规框架内接触到大语言模型，有效普及了基础 AI 素养。防务技术的演进目前正处于从“通用聊天机器人”向“任务级应用 AI”（Applied AI）跃迁的关键节点。通过引入具备幻觉护栏、逻辑推理能力并直接对接作战工作流的专用工具，官兵们能够直观看到数据源头，从而由点及面地推动部队整体的数字化转型。 Paul Lasserre 补充分析了和平时期与战时状态下 AI 采纳机制的巨大差异： - **和平时期**：必须遵循“展示而非说教”（Show, not tell）的原则。例如法国海事部门原先使用巡逻艇盲目搜寻违规排污船只，效率极低；当通过高光谱卫星实时锁定海面油污轨迹并直接引导飞机精准拦截后，一线操作人员迅速建立起对 AI 的信任。 - **战时状态**：在乌克兰战场以及针对伊朗弹道导弹发射架的侦测等高压环境下，对生存与反击速度的极度渴望迫使各方以远超平时的速度试验并部署自主化与实验性技术，极大地加速了技术迭代周期。 --- ### [39:24](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=2364s) 未来十年最具战略颠覆性的前沿 AI 技术 在圆桌会议的尾声，Christine Keung 邀请各位嘉宾展望未来十年最具战略优势的颠覆性技术。Eytan Eshel 旗帜鲜明地指出，**量子计算与 AI 的结合（Quantum AI）** 将引发防务领域的全方位革命。从利用量子计算和 AI 加速飞行器结构设计与新材料发现，到基于量子技术的先进传感器，再到利用量子密钥分发（QKD）构建绝对安全的抗截获通信网络，这种技术融合将大幅压缩武器系统的研发与部署周期，并以远低于当下的算力成本训练出前所未有的新型复杂模型。 Paul Lasserre 则从自身领域出发，断言**星座级高级卫星智能（Constellation-level Satellite Intelligence）** 将成为核心战略高地。未来的自主作战单元将不再局限于单颗卫星的边缘计算，而是跃迁为由成百上千颗卫星构成的协同网络，实现全球范围内的秒级自主重构、协同感知与即时下行分发。 ### [41:43](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=2503s) 太空分布式计算与软件优先的星座协同 嘉宾们深入探讨了卫星星座与边缘智能的协同工作机制，重点强调了“提示与排队”（Tip and Queue）的工作流能力。在这种模式下，大视场、低分辨率的卫星首先负责广域侦察与目标发现，随后引导更高分辨率的卫星对目标进行精确分类、数据融合与负载均衡。从本质上讲，这相当于一个部署在飞行资产上的分布式计算系统，尽管其在太空中运行的特性增加了实现难度。在这种架构下，大型语言模型（LLM）或更通用的功能模型充当整个系统的编排层（Orchestration Layer），使得前线指挥官能够直接使用自然语言提问，而系统会自动将需求分解为任务下发、感知采集和数据融合等具体步骤。 ```mermaid flowchart TD accTitle: Space Distributed Intelligence Workflow subgraph Natural_Language_Orchestration ["指挥与编排层"] A["指挥官自然语言指令"] --> B["LLM / 功能模型编排层"] end subgraph Distributed_Space_Assets ["空间资产与任务分解"] B --> C["广域低分辨率卫星<br/>(Tip / 目标初筛)"] C --> D["高精度卫星<br/>(Queue / 精细分类)"] D --> E["数据融合与负载均衡<br/>(边缘计算)"] end E --> F["态势感知结果与行动闭环"] ``` 长期以来，全球军方在将硬件与软件解耦方面一直面临挑战，但国防科技（Defense Tech）与“新太空”（New Space）的发展彻底改变了这一现状。现代体系带来了极快的创新迭代周期，军方甚至可以利用五年前发射上天的卫星传感器，搭配今年最新部署的 Foundation Model（基础模型）协同工作。这种由各类新兴企业引入的“软件优先”（Software-First）思维正在重塑行业。未来的竞争关键将不再局限于谁拥有更多数量的卫星或无人机，而是谁能够让跨域资产星座高效联动，精准响应问询并采取行动。 - **Tip-and-Queue 机制**：利用低分辨率卫星大范围搜索触发目标线索，再调度高分辨率资产完成精准识别与分类。 - **软硬件解耦**：太空资产支持快速软件升级，数年前发射的硬件载荷可即时运行最新的前沿 AI 模型。 - **分布式空间计算**：将动态在轨星座视为浮动算力网络，通过 LLM 编排层自动化处理任务规划与数据融合。 --- ### [43:17](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=2597s) 边缘模型量化、主权 AI 与按需卫星发射 在谈及 AI 落地实践时，嘉宾指出技术的重心正在发生深刻转移：行业将不再盲目追求算力规模更大、参数更强的极致模型，而是更加聚焦于如何在更小的尺寸规格（Form Factors）上压榨出模型性能，包括模型量化（Quantization）、边缘设备部署以及将 AI 无缝嵌入日常工作流。当前人类将 AI 融入日常作战行动的进程仍处于极早期阶段，核心挑战在于如何将能力充沛的推理模型与确定性的业务流程优雅结合，最终把系统的控制权完整交还给一线操作员和组织本身。与这一技术趋势相伴的是各方对“主权 AI”（Sovereign AI）的强烈呼声，各国与军队不仅要求对自身数据拥有绝对控制权，也要求对用于分析和决策执行的底层模型具备完全自主掌控力。 来自以色列航太工业（IAI）的 Eytan Eshel 随即补充了“按需卫星”（Satellite on Demand）的关键概念。未来的太空支援不仅依托于纳米卫星（Nano-satellite）星座，更在于具备根据即时作战需求、在指定时间精准发射入轨至指定空域的响应能力，这一变革正在快速变为现实。 - **边缘优化与模型量化**：将复杂的推理模型轻量化并部署至边缘终端，结合确定性规则构建可靠闭环。 - **主权 AI 诉求高涨**：确保数据所有权与核心分析/执行模型的自主可控成为国防采办的共识。 - **按需快速响应发射**：结合纳米卫星星座，实现战术级“按需随选”的快速入轨补网与特定任务覆盖。 --- ### [44:45](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=2685s) 打破传统军工垄断：非硬件方案与不对称创新 在对全场讨论进行总结时，主持风格转为充满激情与洞察的收尾。主持人指出，当前国防领域最令人振奋的变化在于：下一代国防创新正逐渐摆脱由传统军工巨头（Primes）主导的局面。过去十年的演变表明，军工创新的垄断格局已经被打破，防务技术的采购方正积极向传统集中网络之外寻找解决方案。这为开源技术的应用与不对称作战能力的开发打开了全新空间。 主持人分享了一个极具想象力的前沿方向——新型密码学在海量数据计算中的应用。以现代航空母舰为例，单艘航母每天产生的数据量高达 1 PB（Petabyte），如何在不单纯依赖堆叠硬件算力的情况下高效处理这些数据，是软件与密码学能够大显身手的关键领域。主持人最后表达了对 defense tech 领域创新潜力的热切期待，并向全体参会嘉宾致谢，宣布座谈正式结束并在台前预留 10 分钟供现场参会者自由交流。 - **创新去垄断化**：国防技术采办跳出传统军工巨头垄断，向更广泛的创新科技生态与开源架构开放。 - **海量数据挑战**：单艘航母日均生成 1 PB 数据，推动了利用密码学等非纯硬件途径解决计算瓶颈的新探索。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=4s) 系统部署的成功与否不取决于“交付速度”，而取决于系统在极高风险下是否足够可信、韧性强、安全且具备实战即战力。 — Christine Keung > [01:57](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=117s) 面对两大核心挑战——一是如何将通用开源生态中所有的 AI 技术优势安全地迁移到本地部署（On-Premise）环境中；二是如何打造高效且高可靠性的人机协作（Manned-Unmanned Teaming, MUM-T）体系。 — Eytan Eshel > [01:57](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=117s) 彻底颠覆传统卫星仅作为“遥感相机/雷达”的旧模式，打造能够在轨道边缘端直接进行自主思考与分析的“智能卫星”。 — Paul Lasserre > [14:14](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=854s) 在真实的战场上，信息极其有限，而所有的作战决策本质上都是风险管理。 — Chris Hume > [18:25](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=1105s) 如果系统无法获得战士或飞行员的信任，再先进的技术也形同虚设。 — Eytan Eshel > [20:34](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=1234s) 过去几十年的边缘 AI 范式主要依赖轻量级 CNN 或针对特定目标训练的 YOLO 模型，只能回答预设问题……而 VLM 具备通用人类知识与多模态推理能力，使得操作人员能够提出高度模糊的宏观指令。 — Paul Lasserre > [23:50](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=1430s) 无论自主化程度多高，“人在回路”（Human-in-the-loop）是绝对前提，AI 负责支撑并发管理成倍增长的作战单元，但最终的交战与开火权限始终由人类掌控。 — Eytan Eshel > [23:50](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=1430s) 信任本质上是从光子到告警的监管链（A chain of custody from photon to alert）。 — Paul Lasserre > [23:50](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=1430s) AI should cue and humans decide — Paul Lasserre > [32:58](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=1978s) CTO 的职责绝不仅是研发先进系统，如果前线部队不用，那就是彻底的资源浪费。 — Eytan Eshel > [39:24](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=2364s) 量子计算与 AI 的结合（Quantum AI）将引发防务领域的全方位革命。 — Eytan Eshel > [41:43](https://www.youtube.com/watch?v=tZWv4Hnx_V4&t=2503s) 未来的竞争关键将不再局限于谁拥有更多数量的卫星或无人机，而是谁能够让跨域资产星座高效联动，精准响应问询并采取行动。 — Paul Lasserre ## Mentioned Christine Keung, J2 Ventures, Chris Hume, Legion Intelligence, Eytan Eshel, Israel Aerospace Industries (IAI), Paul Lasserre, Loft Orbital, Moonshot AI, Kimi, SpaceX, Centurion, Intel, NVIDIA GPX Spark, Project Convergence, Fort Irwin, Inmarsat, Yria, Mech Quantum, NASA, Helsing, Gemma 3, Gemma 4, Google, YOLO, CNN, genai.mil, NIPRNet ## Video description Eytan Eshel - Executive VP Chief Technology and R&D Officer, Israel Aerospace Industries (IAI) Paul Lasserre - General Manager, AI for Space, Loft Orbital Christine Keung - General Partner, J2 Ventures Chris Hume - Chief Business Officer, Legion Intelligence National defense and aerospace leaders are harnessing AI to rethink how missions are planned, executed, and supported. This panel explores real deployments, from edge intelligence in combat zones to predictive maintenance in aircraft, and how the defense sector is navigating the intersection of innovation and oversight. Attendees will gain insights on how AI improves speed, precision, and survivability, while aligning with complex security and compliance mandates. [Source](https://www.youtube.com/watch?v=tZWv4Hnx_V4)

[0:04] Christine Keung: 系统部署的成功与否不取决于“交付速度”，而取决于系统在极高风险下是否足够可信、韧性强、安全且具备实战即战力。 [1:57] Eytan Eshel: 面对两大核心挑战——一是如何将通用开源生态中所有的 AI 技术优势安全地迁移到本地部署（On-Premise）环境中；二是如何打造高效且高可靠性的人机协作（Manned-Unmanned Teaming, MUM-T）体系。 [1:57] Paul Lasserre: 彻底颠覆传统卫星仅作为“遥感相机/雷达”的旧模式，打造能够在轨道边缘端直接进行自主思考与分析的“智能卫星”。 [14:14] Chris Hume: 在真实的战场上，信息极其有限，而所有的作战决策本质上都是风险管理。 [18:25] Eytan Eshel: 如果系统无法获得战士或飞行员的信任，再先进的技术也形同虚设。 [20:34] Paul Lasserre: 过去几十年的边缘 AI 范式主要依赖轻量级 CNN 或针对特定目标训练的 YOLO 模型，只能回答预设问题……而 VLM 具备通用人类知识与多模态推理能力，使得操作人员能够提出高度模糊的宏观指令。 [23:50] Eytan Eshel: 无论自主化程度多高，“人在回路”（Human-in-the-loop）是绝对前提，AI 负责支撑并发管理成倍增长的作战单元，但最终的交战与开火权限始终由人类掌控。 [23:50] Paul Lasserre: 信任本质上是从光子到告警的监管链（A chain of custody from photon to alert）。 [23:50] Paul Lasserre: AI should cue and humans decide [32:58] Eytan Eshel: CTO 的职责绝不仅是研发先进系统，如果前线部队不用，那就是彻底的资源浪费。 [39:24] Eytan Eshel: 量子计算与 AI 的结合（Quantum AI）将引发防务领域的全方位革命。 [41:43] Paul Lasserre: 未来的竞争关键将不再局限于谁拥有更多数量的卫星或无人机，而是谁能够让跨域资产星座高效联动，精准响应问询并采取行动。

[5:02] Eytan Eshel: In the battlefield and current battlefield... warrior face lots of threats coming from above—drones, small arms, anti-tank missiles, IEDs... By implementing AI on those C2 systems, it became very, very essential... actually it's a life-saving decision-making support. [5:02] Chris Hume: Think of being able to bring mission applications to a device... about 128 gigs of unified memory, about the size of a small toaster oven. In doing so, you're essentially providing the user, the operator, with a 'staff in a box.' [5:02] Paul Lasserre: We've basically been using satellites... capture data, hundreds of gigabytes... but then in low Earth orbit you have no real-time connectivity... most MODs in the world have about a six-hour SLA. And so we're totally changing this paradigm the minute we combine AI at the edge with real-time low-bandwidth connectivity.

[0:04] So, good morning everyone and welcome. Um,  as I mentioned, I'm Christine Young. I'm a [0:09] general partner at J2 Ventures. My partners and  I invest in dual use technologies. I think most [0:14] of you who are here might know what dual use is,  but basically means that we invest in technologies [0:18] that serve both large commercial and national  security market. Space is very much part of that [0:23] portfolio. And honestly, I feel like we couldn't  have timed this panel better. I think for any of [0:29] you that have been watching or reading the news  this past week, right, we've seen the buzz last [0:36] week around Moonshot AI's Kimmy released, right,  which this new open source model is now reigniting [0:41] the global AI race with China. I think that's  going to be a hot topic. Um, at the same time, [0:45] nine hours ago, right, we just had SpaceX's first  earnings call since the IPO and Wall Street is [0:51] certainly responding. Um, I think SpaceX shares  have fallen roughly 50% since its peak IPO price. [0:58] And really, again, like none of this is like  financial advice, but I hope it kind of shows [1:02] that the interest and the volatility in the market  underscores how AI, space, and national security [1:08] are increasingly converging. AI has moved well  beyond experimentation in defense. Today, it's [1:14] helping commanders make faster decisions, enabling  aircraft to predict failures before they happen, [1:20] improving satellite intelligence, and supporting  operators in increasingly complex environments. [1:25] AI is also no longer a future capability when it  comes to defense. It is a strategic advantage. [1:31] And I think um what you're going to hear today  from our panelists is in context of war, right? [1:37] Um it's you know AI deployment isn't measured  by how fast you ship. It's measured by whether [1:42] these systems are trusted, resilient, secure,  and mission ready when stakes are high. So, [1:47] that's exactly what we're going to talk about  today. What it actually takes to move AI from [1:51] promising technology to operational capability.  To kick us off, I'd love to hear from each of you. [1:57] Please introduce yourself in a sentence or two and  tell us what the single most important AI priority [2:03] um that's driving your organization today.  And Chris, I'll start with you. Sure. Thanks, [2:07] Christine. And uh good morning everyone. My name  is Chris Hume. I'm the chief business officer for [2:12] Legion Intelligence based out of San Francisco  and I've been working in at the intersection of [2:17] AI and national security for about the past seven  years. I started my career in the army. I was an [2:22] infantry officer then special forces officer and  our most pressing critical priority for this year [2:30] really matches what the military is is pursuing as  well and that's how do we push AI enabled mission [2:36] applications to the edge and really focus on the  resilience in denied integrated environments. [2:44] Thanks for being here. Okay, thanks Christine. I'm  Etonel. I'm EVP CTO of Israel Aerospace Industry, [2:55] the largest aerospace and defense industry in  Israel. I used to be the head of military R&D [3:01] in the Israeli MOD and defense as a brigade  general and now I'm in the reserve and [3:07] uh we have two major challenges not only one  and one is to be able to bring all the good [3:16] that we have in the open in the ecosystem of AI  bringing into working on prem which is very very [3:24] very demanding I think we will talk about it later  on and second to create men on men teaming working [3:30] in reliable manner and to bring the to be to be  able to use those autonomous capabilities together [3:36] with the man system. These are two challenges  that we face and we'll talk about that later [3:45] and hi everyone uh polair um US Army Israeli  Air Force and ex French Navy. So what's a nice [3:53] uh panel distribution. Um I today I lead AI uh for  space uh business unit at loft orbital. Um loft [4:02] orbital uh is a series scale up. Uh we've raised  over $300 million. We've crossed the $100 million [4:10] revenue last year. Uh booked half billion dollar  of of contracts. Uh 350 people roughly across the [4:16] US, France, and the UA. And what we do basically  is we allow Ozers, we fly Ozers's payload or [4:25] software on standardized uh satellites. And over  the the past 3 years or so, we've increasingly [4:32] used AI at the edge. And without killing the  suspense of uh our discussion, space has mostly [4:39] been about putting or using satellites as remote  cameras or radars for Earth observation. And today [4:47] we we we are turning these assets into intelligent  uh satellites that can think at the edge. [4:57] So I'm I'm really glad that we have such great  like diversity ac from the operator's perspective. [5:02] And so for the rest of this conversation, I also  really want you guys to bring that perspective [5:06] not just as builders, right, but also as users of  the technology. Um so there is a lot of excitement [5:11] around AI, but what matters most is actually  what's being deployed and where the bottlenecks [5:14] are. Can each of you share one AI capability  that's already in production or actively being [5:19] considered and explain the operational problem  it solves and the impact that you've seen? May [5:26] I start? You can start. You know in the in the  battlefield and current battlefield and future [5:33] battlefield warrior face lots of threats coming  from above drones small arms anti-tank missiles [5:43] ID improvised explosive devices and many threats  and lots of elements that he need to digest in [5:49] order to operate. By using a command and control  system powered by AI, we're able to ease it for [5:56] the warriors to have a a very clear awareness to  decide to be able to understand the battlefield [6:04] to understand the threats, the environments and  the way we need to to move. By implementing AI on [6:12] those C2 system, it became very very essential in  this dense conflict that we facing. And actually [6:20] it's a lifesaving and these are the type of system  that in Israel aerospace we implement and bringing [6:28] to the field and give it as a decision making  support to the warriors uh by giving C2 system [6:36] with AI that enable to digest and process a huge  amount of elements and knows the where the enemy [6:45] supposed to be where the safeware to to to go  through and uh this is a type of system that we [6:53] use. We have it also in the air force to support  the pilots, but we can talk about it later on. [7:02] Well, I think you probably picked up on a common  theme is how do we do more and more with less, [7:08] especially when we go into these edge  environments? And then um where we're [7:12] we're proud to to really focus is combining what  we're calling the Centurion product line which we [7:18] launched at the beginning of the year which is  really the combination of a of a purpose-built [7:22] hardware form factor with our platform. And so  think of being able to bring Mission applications [7:30] to a device. We've deployed it onto Intel laptops.  We've deployed it onto an NVIDIA GPX Spark, [7:37] which is um about 128 uh gigs of of unified  memory. And um and it's about the size of a [7:45] small toaster oven. And um and in doing so, you're  essentially providing the user, the operator, [7:52] with a staff in a box. And so it it allows um  that resilience. that can connect to the cloud, [7:59] can use larger cloud-based models when available,  but gracefully fall back. And um and so that's [8:05] really the core of whether it's supporting  a battle captain running current operations, [8:10] supporting the J35 on planning and future  operations and integrating all of the the data [8:18] that is proliferating on the battlefield. There's  not really a challenge of of data. It's more [8:24] of making sense of that data. And by providing  reasoning models on top of deterministic steps in [8:30] these purpose-built mission application workflows,  you can start to run your military decision-m [8:37] process with that reasoning in hand. Uh, as things  change, it's it's less about what's happening, [8:42] but why is that change matter? And how do we then  um sync between cloud-based data and resources and [8:50] and what's available locally? And so just as a  quick example, uh we recently were participating [8:55] right down the road in the um at Fort Irwin, which  is a US Army National Training Center as part of [9:01] Project Convergence. We're supporting battalion  level staff and again demonstrating that yes, [9:06] we can work when we have connectivity, but  in the in the current operating environment, [9:12] that's that's never guaranteed and in many  in many ways that is a a liability. And so [9:16] demonstrating that graceful fall back to smaller  models to locally available data and ensure that [9:22] the staff can continue to plan and and manage the  the battlefield and the battle space um at pace. [9:30] Yeah. So mostly so operational AI use  in satellites. So I I leave aside the [9:37] uh data center in space conversation. If if you if  you if anyone is curious about that topic, let's [9:42] have a chat over beer tonight. uh for the concrete  AI in space use cases today there are mostly three [9:49] uh operational uh one is the one maybe I'm going  to to dive deep deeper a little bit in a sec it's [9:55] AI at the edge that unlock real time and autonomy  uh which is kind of critical for these new C2 [10:03] approaches uh campaign management uh we are not  in the single satellite tasking anymore but how [10:09] do you optimize for an entire constellation with  multiple participants in your work and the last [10:15] one is to basically lower the bar in the in the  space domain by allowing a commander in plain [10:21] English to task a satellite to answer a questions.  Um on the first one um by a quick show of hand who [10:29] is not from the space industry in the room that's  okay I was not there in Yeah. Okay. So most people [10:36] so let me maybe give you the maybe the dumb down  version of what what's happened for decades. [10:41] We've basically been using satellites unlike  what we see in Hollywood movies. We've been using [10:46] satellites with like sensors whether they are  camera RF sensors dic sensors radars we we bring [10:56] put them in a satellite we t the satellite in an  area of interest because we have intelligence or [11:00] whatever we capture data over the area of interest  and then it's a huge amount of data like hundreds [11:07] of gigabytes collected usually during an orbit.  Uh but then you have in lower low orbit you have [11:12] no real-time connectivity. So you need to wait  to fly over a ground station which are usually [11:19] located near the poles because that's where your  satellites fly uh over the most often. And so at [11:25] best you get your data on the ground that you can  an analyze with analyst or AI hours after the fact [11:32] like most mods in the world have about six hour  SLA. And so we totally changing this paradigm the [11:39] minute we combine three things. First is obviously  uh AI at the edge. So how can you process the data [11:45] out of the sensor? Um two uh you want to have  realtime connectivity. uh you can in fact couple [11:55] high bandwidth intermittent connectivity with  your ground station which everybody use today like [12:01] or Amazon or what have you or military uh ones  with real time low bandwidth connectivity with [12:08] geo stationary satellites in um we use in Mars for  instance but you no way you send like huge images [12:17] uh heavy data but once your model has processed  the data you're not talking about gigabytes [12:22] anymore. only a few kilobytes or hundreds of  kilobytes, time, location, nature of the risk, [12:29] confidence, maybe a cropped picture compressed of  your ship or SUV or fire over a forest and that [12:35] you have it in real time and you can then reorder  uh your entire process because you you might still [12:41] want to have the raw data later on, but you know  when something likely happened and because you [12:46] know something happened, you can now task a  drone, a plane, anything uh uh downstream uh [12:52] to And so that's a massive change in space right  now that is really changing the nature of of C2. [12:59] We'll talk more about that. Chris May one sentence  following Paul and very interesting description. [13:08] What happened down on the the data data center?  So you have a huge amount of data that coming for [13:13] electoptical satellites, star satellites, sig  satellites, huge amount of data that needed to [13:21] process and then when you use AI algorithms  in order to pinpoint the exact targets on a [13:30] huge amount of data and to be able to assist the  analyst where to look and where which is the right [13:38] target. This is something that it's essential and  we use it and since I is producing those type of [13:46] satellites we know the nature of singance very  well and the ability to train the models in order [13:53] to pinpoint their exact target and those targets  are almost I I cannot say in real time but very [14:00] shortly is disseminate to to the forces that is  air force and some others. This complimentary [14:14] I want to stay on this topic around edge  computing and edge AI especially since I [14:18] mean all three of you have taken as to why  edge AI provides that decision advantage [14:23] um at least at J2 I mean we've had a thesis and a  prepared mind around telecom around edge computing [14:28] since our inception 5 years ago a third of our  portfolio is in space it's in advanced computing [14:33] um but specifically around delivering um  telecom connectivity capabil abilities in commu [14:39] communications eroded environments right so when  you guys were even sharing about your capabilities [14:43] it made me think about our companies that are  complimentary right to what you guys are building [14:47] we've done everything from companies like yria  which is focused on free space optics to investing [14:53] companies like me quantum that's you know building  alternative position navigation timing systems [14:58] so but but I I think sometimes like for folks who  are not familiar with defense who are not familiar [15:04] with the environments that each of you have been  in right to have to make decisions. I I'd love [15:09] to just maybe for each of you, not just talk about  the technology, but just ground it in the example, [15:15] right? Like like especially for me as an investor  when I'm explaining to my investors like why we [15:20] made an investment in chip scale atomic clock.  We ground it in the Ukraine war. We help explain [15:24] why that technology matters. I'd love to hear  from each each of you have talked about your [15:28] organization's tech capabilities, but I'd love  for you to just go back to that and just give [15:33] one example of, you know, why that matters and  where that impact would happen. I I'll start. Um, [15:41] so this is u this will date me, but when I was  a detachment commander several years ago, um, [15:48] the the access to information uh was was  limited. you you made decisions the best [15:54] you could. And at the end of the day, decision-  making in in combat scenarios is an exercise in [16:00] risk management. And when you think about  how to improve that decision-m ultimately, [16:06] you're looking to add context. Okay, I understand  that there's gunfire coming from this direction [16:11] or I understand that there's a a village over here  and and oftentimes you're doing that on the move. [16:16] you're doing that uh potentially in limited  visibility and you're also trying to manage [16:22] a number of different assets and and um and your  your personnel as well as allied personnel. And [16:28] so the more you can develop this common operating  picture, right? We hear about this in in military [16:34] both technology and and uh doctrine, the the  better decisions you can make. And so the more [16:40] um we can assist those decision makers with  reducing the cognitive load to try to make [16:45] sense of all these different inputs and do that  locally so there's no delay in reachback. There's [16:50] no dependencies outside of their uh immediate  area. Um the better decisions that those [16:57] uh those commanders can make. And so ultimately  from a technology perspective, we can enable [17:02] that by helping essentially with putting those  some of those decisions and some of that context [17:09] management into into our technology. And um and  so the Centurion device I mentioned earlier, [17:16] that's how we've manifested that to provide  this context layer to manage incoming streams [17:21] of radio traffic, putting that into just think of  that as the repository of of known information. [17:26] And as things change, you can uh track  those changes. It also ensures that as [17:32] the consequences of decisions increase, there's  more and more oversight and auditability and human [17:38] uh intervention. But the I guess long story short,  the the ultimate goal is how do you change how do [17:45] you convert the command improve the commander's  situational awareness while also helping enable [17:52] the conversion of a commander's intent into into  that outcome. And there's many different ways to [17:58] do that using deterministic steps by controlling  and really ensuring that the applications are [18:03] purpose-built for the need and the mission. And  with configurable Aentic orchestrator workflows, [18:08] you can do that very quickly. You can do that  locally. And you just improve and really push [18:15] down the capabilities of what was previously a  battalion or brigade level. So think very very [18:20] large echelon organizations down to the more  tactical units. And at the end of the day, that [18:25] leads to better, safer, more effective decisions.  It's all a matter of trust. And uh this is uh the [18:35] whole thing in order to gain trust of the warrior  or the pilot in the system there is a special way [18:41] how to develop it and how to enable it to to the  user. So take for example a pilot that is flying a [18:50] fighter pilots in a very complicated environments  with airto-air missiles, ground to air missiles, [18:57] many threats, jamming and need to use the EW and  more lots of elements that affect his ability [19:07] to do the mission by create creating a trusted  decision making support system enabling to focus [19:17] on the mission and to do it in the right manner.  But in order to gain its trust in the system, [19:25] we need a very intensive rigorous process of  validation, verification and explainable AI [19:33] in order to him to use it in the right manner.  And sometimes uh we use the operators the pilots [19:40] through the process of development the AI  capability enable them to be able to test [19:46] it in an environment that is very closely to  the real operational environment that they will [19:51] face in the and this is why we created kind of  a hybrid simulators that beside bringing putting [19:59] simulated data doing it in a real environment.  So we have a real data to get a simulated data [20:06] and enable to test the algorithms and to enable  him to understand that this is trusted algorithm. [20:14] But before we enable it, it's pass a huge way in  order to gain its reliability in order to be able [20:22] to to be used in the field. So it's a single  pilot but it's a pilot that is also working in [20:28] a team and all the team need to have the data and  the capabilities and the lots of data that coming [20:34] from outside and all of it is supported by a a  situ system that is powered by AI but for them [20:42] to use it and to ignore some threats that they  know the system will automatically handle they [20:48] need to trust the the system. Yeah, it's even more  like we follow the news with with what happened [20:54] a few days ago with this Google model that like  hallucinated on like uh special generated data. [21:01] Uh I think explanability for C2 is obviously at  the heart of everything. Uh use cases maybe let [21:07] me give you a few current use cases and where we  see uh the world going. Uh especially with some of [21:13] the research we've done with with NAZA. uh we've  been running like AI at the edge for the past like [21:19] 2 three years in one then the second fully uh AI  powered uh uh satellites um with different sensors [21:27] on board and it's mostly used today for you know  that France unfortunately is burning right now [21:34] um the difference in detecting a wildfire like in  a few tens of minutes and two hours but just day [21:40] and night and usually you can't put cameras or  people watching a forest everywhere satellites [21:46] can do that quite easily. Uh so having the  insight right away when there is a wildfire, [21:51] when there is a metal a metal leak, when there  is like a maritime pollution are kind of dual [21:57] uh use cases in the military. You can imagine  every single use cases that require like uh real [22:04] time or end autonomy uh from damage assessment.  We've been running missions uh with Helsing that [22:10] are publicly available about GPS jammer uh  detection uh RF anomaly detection in general [22:18] and any type of like maritime domain awareness use  cases where you can identify ship at least detect [22:24] identify and then uh run your processes. Um we've  published a paper uh a few weeks ago with NAZA [22:33] uh with the first VLM flying uh ever. It's a GMA  3 model. We're working on the next version with [22:38] a GMA 4 now. uh and it's it's uh it's the  next step in using AI at the edge before as [22:47] um Ean was mentioning the state-of-the-art was  really to use lightweight CNN's uh YOLO models [22:54] or like really models you've trained to recognize  something that are going to when the inference [23:01] happens tell you yes there is a wildfire or yes  there is a ballistic missile launcher in Iran [23:07] because they have seen some before with a vision  language model, you can push this paradigm to [23:12] the next level where these models have have seen  virtually the entire human knowledge before. So [23:18] you don't need to to teach them specifically and  they have reasoning meaning that you can now have [23:25] uh ambiguous queries such as monitor this  frontier, this border for me and let me know [23:31] if you see anything suspicious. Very ambiguous,  something you could not do with a classic CNN. But [23:37] now the model is going to give you alerts when you  believe that maybe some SUV movement or troops or [23:43] whatever is suspicious and let you know and then  you can go and investig investigate with the cube. [23:50] Thank you Paul. So I want to go back to Eton's  earlier point about trust. Um and let's have a [23:56] conversation around balancing speed of deployment  versus trust. Right? I think hopefully what's [24:01] clear to everyone now is that the stakes are high.  The demand is definitely there and if anything the [24:06] demand is growing. So I think just for each of  you like how do you think the military should be [24:13] balancing um speed of deployment right adoption  of new technologies while also maintaining the [24:19] mandate of trust and resiliency. Maybe I'll start.  Okay. Yes. So when we talk about the autonomous [24:27] system and specifically if we talk about swarm  or missiles and so on you need to create a very [24:36] very trusted AI models in order to to be able  to use it. So it's it's much more complicated [24:46] in the commercial world. When you talk about  defense is a matter of ethics, safety and being [24:52] able to know exactly what the system will do in  the boundaries that is allowed. And this is why [24:59] uh the level of uh testing is become very very  uh important and the ability to create a rigorous [25:09] process to be able to take it in the all levels  from the the development phase with the handling [25:17] all the cyber security of course but through the  way that you train the model and you need to train [25:26] the model as much as possible. with real data  which is operational data. It's couldn't be only [25:32] simulated data or some others that you bring from  the open source. So this is why and our advantage [25:39] is that we have a huge amount of operational data  that been used from our operational system that we [25:45] collect through the years decades of collection  of those data that enable us to train the models [25:51] in a way that create reliable models. Yet the  validation and verification process need a special [26:00] tools and special assets in order to to do to do  so. And we are doing this but it's taking much [26:07] more time that is adapted in the civilian area  and uh and then when we bringing at the end to [26:15] the to the field it's always monitored by human.  It's not that you handle the system that is freely [26:23] working or fully autonomous. Always there is a  man in the loop that monitoring at the end and [26:30] enable it to shoot or not to shoot and we brought  it to almost ultimate autonomous capabilities but [26:39] yet in the end we need a human to control it and  this is what I saw in the current future but of [26:48] course the capability that we give enable him to  handle a a much more amount of system in at the [26:56] same time. So when we talk about drone, it can  handle a swarm but understand exactly what they [27:01] will do and not enable it for work freely. So keep  many in the loop validation verification process [27:10] and explainable. You need to exact understanding  what the system will do, what are the source of [27:16] the data and how it react and to predict how  it react. May maybe I'll build on that because [27:23] we are like building the the peaks and shovels to  enable like operational decision maker to to work [27:30] this way. Uh I'd say that trust is a chain of um  of custody from photon to alert. Um you really [27:39] need to uh be able to tell uh from each insight,  each alert you receive uh where does it originate, [27:47] which sensor, which pass, which model did use,  which uh training data set, what is the confidence [27:55] um and be able to audit that back uh at  the raw uh data level. And by the way, [27:59] it's not because you get real-time insights that  you don't want the raw data maybe two hours later. [28:04] uh most in fact military customers today tend  to do both. I believe personally one day you [28:11] you could not care less about the road data but  today we are not there quite there yet. Uh second [28:15] I'd say that AI should cue and and humans decide.  uh I don't see a world tomorrow where it's fully [28:22] automated and we are not inventing anything like  we've seen that happening in the real world owners [28:28] for a decade like most today hospital um uh CT  scan or others are AI powered and it took 10 [28:38] years for the p practitioners to to understand how  to use that to go from a determinist kind of like [28:44] human assessment to probabilistic system and the  same is true here and it requires some training [28:49] for the operators to understand that you have a  system that you optimize for the rate of uh maybe [28:56] false positive or whatever you want to to detect  and it just give you like a higher probability [29:02] to look somewhere for instance and it's not  something maybe most operators are used to in the [29:07] uh quote unquote uh old world and last I'd  say that um we need to build confidence in [29:13] the system uh by starting with like no regrets  uh action we talked about wildfire detection, [29:21] pollution. These are the ones where you need  to really get people to trust the system before [29:26] you go to higher stake uh military use cases.  Yeah. To to build on what Paul just mentioned, [29:33] it's it's about applying the controls uh and  um according to those consequences. And so the [29:40] lower consequence activities you can uh that  are reversible that are bounded, we can move [29:46] faster there. where there's higher consequences,  move a bit more slowly. But we we see this in in [29:52] the commercial space as well. The trust piece,  the change management piece, that's often the [29:56] longest poll in the tent. And what does that take?  It takes incremental demonstrations of confidence, [30:04] transparency, being able to show where the  information is coming from. Our approach is [30:09] effectively we don't need to train models. We need  to ensure that the models at hand are pointing at [30:14] the authoritative data and that we can then audit  those outputs or audit those tasks along the way [30:21] and that again based on the um quality of the  consequence we uh inject more and more frequent or [30:29] higher level uh human intervention. And so as we  think about how do we accelerate the adoption it's [30:36] really about being on the ground. It's being with  the the operators, understanding the pain points, [30:41] customizing and conforming the workflows around  the existing way of doing business and and or [30:48] are executing operations. And then as we can  demonstrate and build that trust, then you [30:53] can increasingly change just how the nature of how  staff officers uh operate. Right? We we have these [31:00] Napoleon Napoleonic um codes. J1 is personnel, J2  is intelligence, J3 is operations. They've been [31:08] around for hundreds of years. As new technologies  injected into those staff functions, you start to [31:15] erode the distinction between intel and operations  and you can start to converge and re reform how [31:22] teams and staff are even organized. But first,  you have to build build their trust. You have [31:26] to replicate and automate in certain steps. So  it's if if we go too fast and you lose trust, [31:32] then it's it's off for not just like a a new  radio, a new vehicle, a new weapon system. So [31:38] it's really no different there. The consequences  can be much greater and it can become a bit more [31:44] opaque given the complexity of the technology. But  again, you can you can build into the the you can [31:50] build that trust by building into the product and  interaction that transparency, that ability to to [31:57] audit and uh and validate. Maybe just one thing as  a former like uh Navy officer historical reference [32:05] about this trust, whatever you do, it will take  time. Uh I often take the analogy of the the the [32:11] Navy using like sails uh with like then um actual  uh steam engines. Um you you might have seen this [32:21] these these paintings where for about almost  one century uh during the 19th century while [32:27] the technology the steam engine was totally mature  you could not find a single ship that had no sales [32:34] because for a sailor a ship should have sales and  did not matter if you could prove that the engines [32:41] was was good enough it took like almost a century.  So I think for many of these questions we should [32:46] anticipate that these new technologies and the new  way to think about these conflicts will coexist [32:52] with the more like legacy approaches. Yeah, I  I really appreciate you guys bringing this up [32:58] and and we this didn't even come up in our prep  call. I'm actually surprised and delighted that [33:02] we're actually very aligned on I don't know like  I guess when like when slash if autonomous fully [33:09] autonomous like weapons or defense systems will  ever be deployed. Um, I think what I'm hearing [33:13] from each of you though is defense systems,  right? When you're building your technologies, [33:17] you're still building it around the judgment of  the war fighters, right? Like you're building [33:21] tools and technologies that's augmenting it, but  you're not replacing that of the war fighter. Um, [33:26] I I'd love to hear and you know, I think often  times you come to conferences like AI4, we spend [33:31] a lot of time talking about tech capabilities,  but the reality is like human adoption is so [33:36] critical in each of your industries. And so  would love to get even just hear some case [33:40] studies or stories of how you've been building  confidence amongst the pilots, the commanders, [33:45] the operators that are using the technologies and  what are some of the feedback that you're getting [33:49] and maybe what's something that has surprised you?  You know when you educate the actually you need to [33:58] educate the the the forces to use those system and  when you prove them that the system assist in a [34:07] way that they get a better situation awareness  very quick response to threats and enhance [34:17] dramatically their capability in a timely manner.  they will understand how to use the system and [34:23] uh myself as a city of II it's not only the  development of the system and you can do the [34:30] the state-of-the-art system but if they are not  using them then it's a waste so we work very [34:36] closely with the forces and enable to bring the  system to the field and educate them that by using [34:43] those system as a decision support and intero  intelligence and better understanding of what's [34:51] happening because in the in the battlefield a huge  cloud of I mean cloud that you don't understand [34:58] what's going on and those type of system enable  them to have a clear understanding what's going [35:03] on and by bringing it again it's take a a bit a  bit more time but when they educated to be able [35:12] to use those system in the way that enhance their  capability they will start to use it and they will [35:18] not will go back to the old system like the sales  that Paul mentioned. So it takes some times but [35:27] after they get understanding of the advantage  that it bring they will not go back and will [35:34] continue to use it and so that way you gain trust.  So you need to bring it to the field and to have [35:41] a clear understanding what this system is able to  do and how it's enrich the understanding and the [35:49] way they they they will fight. So it's applicable  to ground forces but the same for air force to the [35:56] pilots when he gets a clear understanding  what the system is enabling to do. Uh, [36:05] I'd say unsurprisingly there's a wide spectrum of  comfort and understanding and literacy around AI [36:12] and um and national security and um surprisingly  there's a a growing um end of the spectrum that [36:22] is increasingly aware and increasingly joining  the force AI native in many ways. And what we're [36:28] seeing is there's this proliferation  of access to AI on government systems. [36:34] Think of ji.mill for those that are familiar which  is essentially providing in a unclassed but secure [36:40] government sanctioned environment access to a  chat chatbot and and that's great for people's [36:45] just comfort and and building that literacy where  and and when it comes to continuing to build that [36:52] trust in education. it really uh comes down to  identifying those people that are comfortable [36:57] and and are literate and they can in turn build uh  build programs and they're often quite motivated [37:03] to share what they've done, what they've  learned, what they've built with their with [37:06] their colleagues. And so encouraging that and and  um giving more and more tools to to enable that [37:12] is important where I'd see really the the phase  that we're in now is is really translating from [37:19] AI and access to chatbots and really the applied  AI piece. So as people understand okay this is [37:25] what AI can do and what it cannot do you start to  again build into applications these a AI steps. uh [37:34] you can start to incorporate the reasoning models  that are guardrailed against hallucinations. You [37:39] start to ensure that the the user can see where  the source data where it's coming from. And um [37:44] and by doing so, you combine that base kind of  foundational AI literacy with an appreciation [37:50] for how when applied to actual workflows and  and mission applications, you you really start [37:56] to unlock the true power and and you build upon  that that trust and and that change management. [38:04] quickly because we often like a step uh remote  from the the end user. We usually uh work with [38:11] a little place lab in the US housing uh quantum  system uh for before the military user. I would [38:18] still make the difference between peace and  wartime uh conditions. In in peace time what [38:24] we've seen working is to kind of unlock to show  not tell to show typically like in France for [38:30] maritime pollution you would have the prefer team  it's called who is sending like patrol boat to try [38:35] to detect on the spot maritime pollution. It's  if you think about it that's quite ineffective [38:40] because you need to go across large areas to find  like the guilty ship. If you can show in real time [38:48] where you have with hyperspectral measurements  some like oil in the sea and direct the the [38:53] plane then people tend to uh to trust you pretty  quick. um in in wartime what we've seen is that [39:01] with Ukraine and Iran recently we've seen like the  progress of autonomy uh going much faster than uh [39:08] ever uh before and uh whether that's in Ukraine  whether that's detecting like m ballistic missile [39:17] launcher before they they fire you're more willing  to test any experimental technologies and and prog [39:24] this way So, we're down to our last five minutes  and I'd like to just wrap by, you know, maybe just [39:32] reflecting on where we think our industry is  going, right? Like each of you have identified [39:37] or just spoke to frontier technology that is  just growing rapidly in the commercial sector, [39:43] right? We talked about generative AI, we've talked  about um autonomous systems, edge compute, AI [39:47] enabled satellite networks, and you're just seeing  rapid rapid commercial adoption and improvements. [39:52] um which which emerging AI capability do you  believe will create the greatest strategic [39:58] advantage over the next decade and why? I think  the combination of AI with the quantum computing [40:06] will uh make a dramatic change all over and  specifically to the defense. the ability to design [40:14] air structure much faster in a very different  way that we are doing and to be able to analyze [40:23] a new material materials that you use using  quantum computing and AI together that will take [40:29] a shortening time will make a revolution in the  way that we design the system and implement this [40:36] capability. So it can goes from a sensor build  it over a quantum capabilities to communication [40:46] using QKD quantum key distributor to create a  very uh safe networking and communication. It [40:54] started already, but it also goes to computing  and the ability to create new models in a faster [41:02] way and maybe cheaper way than we use today and  create a a a capability that we are not able to [41:10] yet to develop in a timely manner. So I think this  in the next decade will influence a lot not only [41:17] in the commercial but also in the defense system.  I I preach for my own kind of like area. Um I'd [41:27] say advanced uh satellite uh intelligence. Um not  at the satellite level, but at the constellation [41:35] level. Uh where the the unit of of autonomy  really stops being the one satellite and become [41:43] the ability to tip and queue to have like a large  swap low resolution satellite detect something. [41:48] the next satellite have maybe higher resolution,  more precise ability to uh to to classify, [41:55] do uh data fusion, do load balancing. It's  pretty much like a distributed compute problem [42:02] uh just with flying assets which makes it a little  bit harder. You've got your LLM or your let's say [42:07] more generally your functional model as the  orchestration layer which allows a commander [42:12] to ask uh in plain language a question and get the  constellation decompose that in terms of tasking [42:22] sensing and data fusion. Um and what we see which  is something the military traditionally across the [42:30] the world has not been very good at is decompose  hardware and software. What we see with uh defense [42:36] tech and uh and the new space is that you now have  the ability to have uh quick cycles of innovation [42:43] you can use a five years old like sensor flying  on a satellite with this year latest uh foration [42:50] model which uh really is a software first mindset  that all types of of companies uh bring uh which [42:58] is quite interesting. So again I don't think the  future is going to be was the more satellites or [43:02] drones but who has the ability to have this  constellation of assets not only satellites [43:07] uh answer questions and take action with the right  fulfillment and um as as they say where you stand [43:17] often determines is determined by where you sit um  I believe that being able to it's it's it will be [43:24] less and less about um getting access to more  and more performant models. It'll be more and [43:30] more about squeezing the performance out of these  models on for smaller form factors, being able to [43:37] quantize them, being able to put them in into  edge devices, and being able to integrate them [43:42] into just our daily work. I think we're very very  early innings at actually incorporating AI into [43:49] our day-to-day operations and finding those ways  to elegantly combine the reasoning models that [43:56] are more more than performant for most of the task  today with those deterministic steps and really at [44:01] the end of the day giving that control back to  the operator back to the organization. We hear [44:06] more and more around the the calls for sovereign  AI and control of both not only your data but but [44:12] the models that are that are used to analyze and  and execute and and I think that trend will only [44:17] accelerate and the the technology will enable that  acceleration. It's may I may I add one sentence to [44:24] what Paul mentioned it's also satellite on demand.  goes to nano satellite constellation but also on [44:31] demand the ability to launch it exactly when you  need it to the point that you need it with some [44:38] constellation and this is something that we see  it's coming we're at time but I'm going to give [44:45] you guys my quick take as well um because why not  um I I would say I I agree with all of you right [44:53] um but I I also think it's really important that  like hopefully what you hear by this panel is I [45:00] think the next set of defense innovations aren't  going to be coming from the prime that's one of [45:04] the most exciting things right about the last  10 years even from the last week is there is no [45:09] longer a monopoly over defense innovations right  like you are seeing buyers of technologies looking [45:14] at solutions outside of a concentrated network  which hopefully means just you know potentially [45:21] more adoption of open source potentially more  development of asymmetric capabilities who knows [45:27] Right. I'd say one thing that's been capturing my  imagination has been looking at new applications [45:31] of cryptography. Right. An aircraft carrier  collects one pabyte of data a day. I spend a lot [45:36] of time thinking about non-h hardware solutions on  how to compute that data. Um I I I mean I hope you [45:43] guys all just feel excited and energized by these  capabilities. Um I also want you all to just join [45:48] me in thanking our fabulous fabulous panelists.  We didn't leave time for questions. I'm sorry [45:53] about that. But we're going to be sticking  around for about 10 minutes at the front and [45:56] we'd love to come up and you know we'd love to  talk to you. So thank you all. Thank you guys.