Optimum, AOI Executives Detail Telecom Shift From AI Hype to ROI

Optimum与AOI高管详解电信业从AI炒作转向ROI 电信运营商正放弃对实验性人工智能的大规模支出，转而青睐能够带来可衡量成本节约的针对性自主智能体与预测性维护。 电信行业高管探讨了从投机性AI炒作向基础设施及客户体验中务实、以ROI为导向的实际生产应用的转变。核心议题包括用实时座席指引取代静态推荐系统、将预测性维护应用于网络运营，以及通过分阶段区域推广来推进传统硬件升级。专家组还探讨了自主行动与人工监管之间的平衡、工程角色向模型管理方向的演进，以及对AI治理和安全护栏的迫切需求。 现代电信客户体验正从基于规则的引擎转向上下文AI，以生成实时座席指引和个性化方案。 (4:07) 数字孪生使电信运营商能够在开展实际营销活动之前，模拟客户对新定价和服务套餐的反应。 (4:07) 为管理核心网络基础设施中的运营风险，部署采用了分层模型，范围涵盖从完全自主的低风险任务到针对高风险操作的人在回路（human-in-the-loop）建议工作流。 (6:25) 电信企业正在推行“AI节食”（AI diet），将投资集中于可衡量的ROI、网络可用性和主动留存用户，而非单纯追求Token消耗量指标。 (8:21) 尽管L5级完全自主、自愈的网络仍是长期目标，但近期进展主要依赖于针对高成本、高频网络问题的专用AI智能体。 (13:29)

Telecommunications providers are transitioning artificial intelligence from speculative pilots into pragmatic production, prioritizing customer experience personalization and tiered network automation. According to industry panelists, realizing fully autonomous operations remains constrained by legacy hardware telemetry gaps, operational risk, and the need for rigorous governance. Customer Experience and Tiered Decision Automation (4:07) Customer service operations have delivered immediate production value by replacing static, rule-based scripts with contextual models that evaluate subscriber sentiment and generate dynamic recommendations in real time. Beyond live agent guidance, operators utilize digital twins to simulate customer persona reactions to revised pricing and packaging tiers prior to commercial deployment. To safely delegate authority, Sundar Krishnan outlines a value-versus-effort matrix that isolates automation risks. Routine, low-stakes administrative tasks are offloaded to autonomous tools—such as an internal '50-cent analyst' bot that resolves database investigations previously requiring an hour of manual analyst labor. Conversely, high-value strategic workflows remain bound to human-in-the-loop oversight and traceable evaluation frameworks to prevent unmonitored decision-making and model hallucinations. Risk-Stratified Autonomous and Self-Healing Networks (6:25) In core network infrastructure, operators are moving from reactive fault management toward predictive maintenance. Because physical hardware such as cable and fiber amplifiers cannot dynamically reroute traffic during failures, predictive anomaly detection is essential. Sahil Yadav describes an operational framework categorized by risk: low-risk tasks execute autonomously, medium-risk actions follow a human-on-the-loop model requiring operator sign-off, and high-risk tasks restrict AI to diagnostic guidance for field engineers. Anne Lee notes that while Level 5 fully autonomous, self-evolving networks remain a distant milestone, near-term progress relies on specialized agents resolving high-frequency, localized network problems. By automating routine troubleshooting, operators can redirect engineering capacity toward higher-value services, including the development of advanced emergency response infrastructure featuring live video and interactive whiteboards for first responders. Legacy Hardware Hurdles and Phased Deployment Deploying machine learning models across cable and fiber networks is heavily bottlenecked by legacy equipment. Older amplifiers, nodes, and routers lack the integrated sensors required to generate continuous telemetry feeds for inference engines. Upgrading these physical components demands substantial capital expenditure from major carriers like AT&T and Comcast. To mitigate financial risk and prevent operational disruptions, providers utilize geographically phased rollouts. This approach limits the potential blast radius of malfunctioning algorithms while collecting real-world operational data to refine model performance. Anne Lee compares current pilot failure rates to previous shifts spanning 1G through 5G, emphasizing that an initial minimum viable product merely validates baseline functionality in a lab, whereas scalable deployment requires extensive capacity and reliability engineering. ROI Discipline, Edge Models, and Governance (8:21) Addressing industry maturity, panelists describe an 'AI diet' that replaces broad experimentation and token-tracking metrics with measurable ROI, such as proactive churn reduction and model cost-per-inference evaluations. Operators also apply predictive signal-to-noise ratio models to diagnose connection issues remotely, avoiding costly field service dispatches ('truck rolls') by enabling subscriber self-troubleshooting via visual diagnostics. At the edge, open-source 1-bit, 2-bit, and 4-bit quantized models enable lightweight co-pilots to run inference locally on mobile devices without network connectivity. Concurrently, Sundar Krishnan expects manual prompt engineering to fade as autonomous agentic architectures handle reasoning natively, pointing to Anthropic's reported elimination of 80% of system prompts during Claude Code development. However, Sahil Yadav and Lee caution that safety containment and explainable AI (XAI) guardrails must be prioritized to manage systemic operational risks as agent autonomy expands. While telecommunications operators are capturing measurable returns in customer support and localized diagnostics, achieving fully autonomous, self-healing networks remains an incremental, capital-intensive endeavor. Long-term success hinges on modernizing physical field hardware, enforcing rigorous evaluation harnesses, and establishing robust safety guardrails before removing human oversight from critical systems.

在Ai4研讨会上，来自AOI、Optimum及前沿咨询领域的专家指出，电信行业已越过单纯炒作阶段，正通过价值与风险分级框架推进AI落地，将重点转向客户体验动态决策、网络预见性维护以及兼顾基础设施现实的审慎投资。 从流量狂热转向ROI与精准分级 (8:21) Sundar Krishnan指出电信业正在推行“AI节食”（AI diet），企业从单纯追踪Token调用量与使用排行榜，转为严格评估具体业务功能的实际投资回报（ROI）与单次推理成本。为兼顾创新与风控，Optimum建立了价值与难度的四象限评估体系：对于低价值、低难度的事务，允许自动化系统自主探索，例如其内部开发的“50美分分析师”机器人，能以每条查询约50美分的成本自动检索后端数据，直接替代数据分析师原本耗时约1小时的调研工作；而对于高价值与高风险的战略决策，则必须引入“人在回路”（Human-in-the-loop）机制与可追溯的评测框架，确保模型输出严密受控。 客户体验重构与端侧离线智能 (4:07) 在客服与用户运营领域，传统静态、基于固定规则的推荐引擎正被实时上下文AI取代。AI能够动态评估网络故障或升级意图等用户情绪，为坐席实时生成定制话术与优惠方案；运营商还构建了数字孪生系统，在向真实用户发布促销策略前先在合成人群画像上模拟反应以降低试错风险。Anne Lee预测在5年内，AI客服将实现流畅的自然语言理解，彻底淘汰按键式交互语音应答（IVR）。 针对网络优化与现场运维，Sundar Krishnan说明预测性信噪比建模与结合视觉AI的用户自洽App可帮助诊断家庭网络故障，从而减少代价高昂的现场技术人员上门服务（truck rolls）。在边缘计算方面，1比特、2比特与4比特等高度压缩的开源小模型配合轻量化智能体架构，使离线移动设备也能在严格的算力限制下运行智能助理，而计算效率的提升正如“杰文斯悖论”（Jevons' paradox）所示，正加速推动AI的整体应用普及。 自愈网络推进与分阶段基建改造 (6:25) 在底层基础设施方面，AOI的Sahil Yadav指出有线电视与光纤硬件（如放大器）无法像IP路由器那样在故障时自动重路由，因此AI的异常检测和预测性维护尤为关键。针对工程团队对生产环境控制权交接的谨慎态度，业内采取分级部署策略：低风险任务由AI自主执行，中风险任务采用“人在环上”（Human-on-the-loop）审批后自动部署，高风险任务则仅由AI担任诊断辅助顾问。Anne Lee补充表示，按照电信网络自动化的五个等级标准，实现完全自主、持续演进的Level 5网络仍是长期目标，当前重点是针对高频、高成本痛点部署专用自愈智能体。 实现自愈网络的最大物理瓶颈在于老旧硬件缺乏传感器与远程遥测能力。Sahil Yadav说明，全面更换智能硬件需要AT&T与Comcast等运营商承担高额资本支出（CapEx）。为了平衡成本并控制模型失常可能导致的“爆炸半径”（blast radius），运营商普遍采取按地理区域分阶段升级的方式，既为模型在受限环境中积累真实运行数据，又平滑了改造成本与运营风险。 规模化工程障碍与长期治理风险 针对外界热议的“90%至95%的AI试点项目面临失败”，Anne Lee结合其横跨1G至5G长达四十年的行业经验指出，任何新技术在早期原型或最小可行产品（MVP）阶段的淘汰率都很高，实验室环境下的成功不等于具备商业规模所需的网络容量与现场可靠性。Sundar Krishnan补充说，许多概念验证（PoC）项目停滞不前，是因为团队只专注于炫酷的原型，而忽视了生产环境所需的基于角色的访问控制（RBAC）、数据流水线、安全护栏以及延迟优化，且未能从一开始对齐核心战略目标与投资预算。 在关于技术未来走向的讨论中，Sundar Krishnan认为随着智能体原生自主推理能力的增强，需要人工调试的“提示词工程”将逐渐退场。与此同时，Sahil Yadav则对业界轻视AI治理与安全风控提出警告，引述有关不受控模型可能带来极端破坏力的言论，强调必须加大对可解释性AI（XAI）的投入，在智能体日益自主化之前构建严格的控制与解释机制。 本场讨论的共识在于，AI不会直接削减电信工程师岗位，而是将其从繁琐的日常维护中解放出来，转向监管机器学习模型以及开发下一代应急通信等高价值公共服务。然而，传统无感知硬件升级所需的高额CapEx投入，以及完全自主网络在安全治理与可解释性上的长远挑战，仍是电信行业必须持续权衡的现实约束。

## Background 该视频由 Ai4 于 2026 年 8 月 22 日发布在 YouTube 平台。根据页面简介，本次圆桌论坛聚焦电信运营商如何将 AI 技术全面应用于核心网络基础设施、客户体验及后台流程，由 CEW Communications 创始人 Cathy White 主持，邀请了 Optimum 的 Sundar Krishnan、AOI 的 Sahil Yadav 以及资深电信顾问 Anne Lee 共同探讨实时诊断、预测性维护、降低用户流失率以及应对 5G 与边缘规模化部署等关键议题。 ## TL;DR 论坛深入剖析了 AI 从概念验证（POC）走向电信现网规模化部署的落地路径与挑战。嘉宾们重点分享了自愈网络的分级风险控制矩阵、数字孪生与超个性化推荐在客服中心的实战应用，并指出了企业从盲目跟风转向注重实际业务产出与合规的“AI 节食”趋势。此外，讨论还涵盖了对老旧遗留基础设施的区域化渐进改造、轻量级边缘智能体的发展，以及 AI 如何推动一线工程师向系统治理与高价值公共安全服务转型。 ## Atmosphere & style 整场研讨会专业务实且兼具风趣幽默，主持人 Cathy White 以轻松活跃的现场互动开场，迅速调动起行业受众的参与感。跨越多个通信代际的资深专家与一线技术领袖在对话中既有对网络物理瓶颈与落地风险的直言不讳，也有对自动化技术演进的从容定力。从调侃传统电话菜单的普遍痛点到最终针对 AI 治理失控风险的深刻预警，整场对话在生动风趣的技术共鸣与严肃严谨的产业反思之间形成了极具张力的平衡。 ## Key takeaways - [09:44](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=584s) 电信行业对 AI 的投资正告别前期的盲目跟风，进入理性务实的“AI 节食”（AI Diet）阶段，企业更聚焦实际业务功能产出与 ROI。 - [13:29](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=809s) 现网落地自愈网络面临极高可用性挑战，必须引入“低风险 AI 闭环、中风险人在回路旁、高风险人在回路中”的分级风险矩阵。 - [13:29](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=809s) 电信工程师的角色正从繁重的设备日常排障运维，转型为训练、优化与监管自愈 AI 系统的核心管理者。 - [16:33](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=993s) 迈向 L5 级全自主网络仍需漫长时间，短期更务实的路径是针对高停机损失、高频痛点故障构建垂直专有 AI 智能体。 - [16:33](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=993s) 基础设施降本释放的工程力量与资金应赋能公共安全领域，推动 911 紧急救援系统向实时视频与多模态数据共享升级，实现深远的“人类价值回报”。 - [20:34](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=1234s) 企业放权 AI 需依据“精力投入 vs 业务价值”二维象限矩阵，低价值低精力任务优先全面放权（如 50 美分分析师），高价值决策则必须保留严格的人工介入与评估体系（Evals）。 - [23:33](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=1413s) 大量 AI 试点（POC）无法规模化量产的核心根源不仅是工程护栏不足，更在于立项时与企业的核心业务战略与财务 ROI 脱节。 - [27:58](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=1678s) 面对老旧“哑设备”缺乏遥测能力的困境，电信运营商采取分区域渐进式部署策略，以严格控制故障爆炸半径并在实测中迭代模型。 - [31:42](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=1902s) 轻量化小型模型、模型量化与微调技术的突破，使得专事专办的离线边缘智能体（Edge AI）在设备端脱机运行成为行业演进必然趋势。 - [39:29](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=2369s) 未来五年内，具备精准意图理解的 AI 智能客服将无缝取代传统多层 IVR 按键菜单与漫长的人工排队等待。 - [39:29](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=2369s) 随着全自主智能体（Autonomous Agents）能力的演进，“提示词工程”（Prompt Engineering）等依赖人工微调 Prompt 的开发方式将逐步退出历史舞台。 - [41:57](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=2517s) AI 治理（AI Governance）与可解释性 AI（XAI）是当前最被低估但至关重要的领域，必须及早为自主模型建立严密的安全护栏与合规边界。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=4s) 论坛开场与现场观众互动 主持人 Cathy White（CEW Communications 创始人兼 CEO）以轻松幽默的口吻为论坛拉开序幕。她笑着提到自己的欧洲口音一听就“不是本地人”，并透露这是她第一次来到拉斯维加斯参加行业峰会，一路收获颇丰。在正式引入主题之前，Cathy 与现场观众展开了即兴互动，调查在场嘉宾的行业背景与 AI 落地现状：“如果你在电信行业工作，请举手？”看到台下热烈的反应，她打趣道：“谢天谢地，你们来对房间了。”随后她继续追问有哪些企业已经将 AI 部署到了生产环境中，以及哪些人还在观望且不知从何下手，从而迅速将讨论聚焦于本次核心议题——AI 在电信领域的当下实践与未来走向。 ```mermaid flowchart TD accTitle: 观众AI采纳现状调研 subgraph S1["现场受众画像调研"] A["电信行业从业者"] --> B{"AI 实施阶段"} B -->|"已在生产中部署 AI"| C["推进核心业务与基础设施变革"] B -->|"正在考虑/不知从何入手"| D["探索具体应用场景与ROI"] end ``` ### [01:05](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=65s) 嘉宾背景与核心 AI 战略举措 三位来自不同细分领域的专家依次介绍了各自的背景以及目前最为核心的 AI 项目。AOI（Applied Optoelectronics Inc）软件与 AI 负责人 Sahil Yadav 表示，AOI 是一家为 Comcast 等大型电信服务商制造硬件放大器并开发全套软件基础设施的领先厂商。Sahil 指出，他当前最聚焦的战略是将 AI 从单纯的数据监控大盘（Dashboards）推向现场实操（In the field in action），构建具备自愈（Self-healing）与自主优化能力的自动化网络。 随后发言的 Anne Lee 拥有近 40 年的深厚电信从业背景，亲历了从 1G 模拟通信、2G 数字通信、3G 宽带、4G 移动语音（VoIP）到如今人工智能时代的完整演进。她开玩笑说自己刚刚退休，目前最重要的事情就是来参加这场圆桌论坛，引得全场哄堂大笑。谈及退休前的重点研究，她介绍自己一直致力于打造一个“高能效的持续学习智能体 AI 框架（Energy-efficient continual learning agentic AI framework）”，以实现完全自主的网络调度与管理。 Optimum AI/ML 总监 Sundar Krishnan（此前在 Comcast 任职多年）则分享了来自一线运营商视角的重点举措。他所带领的团队核心发力于三大方向：指导客服坐席进行精准销售的超个性化推荐系统、用于行为预演的数字孪生（Digital Twins）仿真平台，以及大规模智能体（Agentic）解决方案。 - **Sahil Yadav (AOI)**：专注于自愈与自优化网络，推动 AI 从被动图表展示走向网络前沿的自主决策与干预。 - **Anne Lee (行业顾问)**：跨越 1G 至 4G/AI 的行业老兵，主攻支持全自主运行网络的低能耗持续学习智能体架构。 - **Sundar Krishnan (Optimum)**：主导客服超个性化推荐、数字孪生模拟预测与大规模智能体协同方案。 ### [04:07](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=247s) 客服场景中的超个性化推荐与数字孪生仿真 谈到 AI 如何从早期的概念验证（PoC）真正跨入生产环境并创造商业价值，Sundar Krishnan 详细拆解了客服中心（Customer Operations）的演进路径。过去客服系统依赖死板的基于规则（Rule-based）的推荐引擎，坐席只能机械地按流程提示向用户推销套餐；随后演进为机器学习模型辅助决策；而今天，生成式与智能体 AI 不仅能够为客服实时生成话术，更能精准理解用户的来电情绪与底层诉求。无论是准备升级套餐的高兴客户，还是遭遇网络故障而怒气冲冲的投诉用户，系统都能在通话建立的瞬间完成情绪和意图模拟，帮助坐席在极短的时间窗口内做出最恰当的应对。 ```mermaid flowchart LR accTitle: 坐席超个性化与数字孪生工作流 subgraph Input["数据与状态输入"] U["客户来电/历史数据"] DT["客户数字孪生模型"] end subgraph AI["AI 决策层"] P["意图预测与情绪模拟"] R["超个性化话术与优惠匹配"] end subgraph Action["执行层"] A["坐席实时交互"] C["促成留存/升级转化"] end Input --> AI AI --> Action ``` 更具前瞻性的是数字孪生在客户运营中的应用。Sundar 解释，团队通过构建客户的“数字孪生智能体”，在向真实客户推送套餐或营销方案前，先在模拟环境中进行反应测试与压力推演。这种做法避免了盲目向真实用户试错带来的体验损伤，让运营商在客户主动致电之前，就能够在规模化层面上精准制定出高度个性化的服务策略。 ### [06:25](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=385s) 基础设施层的自主网络管理与自愈挑战 在网络基础设施层面，Sahil Yadav 深入剖析了 AI 为网络运营管理带来的底层变革。从预测性维护（Predictive maintenance）、异常检测（Anomaly detection）到真正的网络自愈，AI 正在从“发现问题”进化为“主动部署解决方案”。然而在实际落地中，有线电视（Cable TV）与光纤网络面临着独特的物理约束——与无线网络中可以通过不同路由器灵活绕行（Reroute）流量不同，有线网络依赖放大器（Amplifiers）级联，一旦放大器损坏便无路可绕。 这种物理特性的不可替代性使得网络高可用要求变得极其苛刻（Mission-critical）。Sahil 直言，尽管自动化是迈向下一代网络的必经之路，但在真实生产环境中全面放权给自动化系统仍面临着来自一线运维团队的巨大阻力：“自动化是会让人们感到恐惧的（Automation is something that scares people）”。运营团队对断网和停机的极度敏感，要求 AI 系统必须在准确性和可靠性上达到极高标准。 Anne Lee 随后从宏观产业视角对此表示赞同。她指出，电信业投资 AI 最明智的两个方向正是“网络运维管理”与“客户留存与个性化”： - **故障阻断与硬件告警**：利用 AI 迅速识别隐患并触发自愈，或在需要物理更换板卡（Swap out a card）时第一时间通知工程人员前往现场，最大限度缩短甚至消除停机时间。 - **商业价值闭环**：通过个性化体验牢牢绑定存量订阅用户，防止客户在激烈的市场竞争中流失到竞对网络。 ### [09:44](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=584s) 告别盲目狂热：主动式运营与“AI 节食”趋势 针对“行业是否已经走出 AI 炒作周期（Hype）”的提问，Sundar Krishnan 坚决表示，当前的 AI 已经全面进入业务实操阶段。过去的运营商往往处于被动响应状态，等待客户打电话进来投诉；而如今，海量网络和用户数据让运营商能够提前预测客户何时可能流失（Churn）或何时产生套餐变更需求，从而能够先发制人（Proactively）主动联系客户并提供预制选项。网络自愈、主动关怀与前置决策相结合，正在真正改变企业的商业回报（ROI）模型。 在谈到行业对 AI 投资的态度转变时，Cathy 引出了一个内部探讨过的概念——“AI 节食（AI Diet）”。Sundar 证实了这一现象：在今年年初，各大科技与电信企业不计成本地大举投入 AI，甚至像 Uber 那样在内部设立排行榜（Leaderboards）来追踪员工的 AI 工具使用率；但随着时间推移，企业管理层正在从前期的盲目跟风和过度支出转向理性评估，更加严格地审视 AI 项目的实际生产力产出与商业回报率。 ### [12:36](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=756s) 企业对 AI 的开支回归理性：“AI 正在节食” 在今年年初，不少企业陷入了“为了在 AI 上花钱而花钱”的狂热之中，但眼下风向已然逆转。行业内部现在流行一个形象的说法——“AI 正在节食”（AI is on a diet）。企业不再不加节制地追逐一切新技术噱头，而是开始大幅收缩战线，在技术选型上变得异常克制。当下企业关注的焦点已切实转向两个核心命题：第一，能否通过 AI 真正产出的功能来精准衡量其业务价值；第二，能否通过追踪企业内部的 AI 使用情况，确保系统始终在做正确、合规的选择。这种转变标志着企业开始用更加务实聪明的策略替代盲目投资，不再紧盯大模型排行榜上的 Token 消耗量，而是关注具体的业务产出与投资回报。 - 企业策略从盲目跟风转变为严格控制技术选型范围与预算。 - 核心评估指标从大模型 Token 消耗量排行榜转向实际交付功能所产生的业务价值。 - 普遍引入内部 AI 使用追踪机制，对算法决策的准确性与合规性进行持续审计。 --- ### [13:29](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=809s) 自愈网络的现实落地：风险矩阵与工程师角色的重塑 在电信行业的各种技术会议上，“自愈网络”（Self-healing Networks）是一个被频繁讨论的高频词，但在实验室与实际现网部署之间，存在着巨大的现实鸿沟。在实验室的受控测试环境中，自愈的定义极为纯粹：AI 自行发现问题、自行制定修复方案并全自动下发部署。然而一旦投放到纷繁复杂的现网，性质就彻底变了——错误的配置修复或误操作可能会引发大规模现网瘫痪（Outage），带来无法估量的商业灾难。为了在安全与自治之间取得平衡，现网自动化必须引入“风险矩阵”（Risk Matrix）机制：对于风险级别最低的基础任务，可以直接放权给 AI 闭环执行；对于中等风险的运维操作，采用“人在回路旁”（Human-on-the-loop）的被动监督模式；而对于涉及核心基础设施的高风险操作，则必须坚持“人在回路中”（Human-in-the-loop）的人工审核与最终拍板。 ```mermaid flowchart TD accTitle: 现网自动化风险控制流程 subgraph S1["风险分级 (Risk Matrix)"] L["低风险操作"] M["中风险操作"] H["高风险操作"] end subgraph S2["自动化执行策略"] A["AI 自主执行闭环"] B["人在回路旁 (Human-on-the-loop)"] C["人在回路中 (Human-in-the-loop)"] end subgraph S3["反馈与演进"] F["现网实测暴露"] FB["工程师反馈"] TR["建立信任 & 模型演进"] end L --> A M --> B H --> C A & B & C --> F F --> FB FB --> TR ``` 在现阶段的技术成熟度下，AI 的真实定位更像是现场工程师与技术专家的“工作伴侣”（Companion）。它负责快速汇总多源网络信息并向工程师提出排障建议，但远未达到全面无人自主化的阶段。这种人机协同形成了一个良性的双向进化回路：工程师借助 AI 辅助提高排障效率，同时工程师的每一次反馈与确认又在反哺 AI，使模型脱离纯粹的实验室温室，在真实复杂的现网环境中经受打磨并建立起工程信任。随着技术逐步走向更深层次的自愈，电信工程师的岗位职责也将发生剧烈转变——他们将从疲于奔命的“网络运维者”蜕变为“运行该网络的 AI 系统管理者”，其核心技能树将全面转向机器学习模型调优、人机交互接口管理与策略治理。 - 自愈网络现网落地采用分级风险矩阵，按低风险（AI 自治）、中风险（人工旁路监督）、高风险（人工回路审批）推进。 - AI 当前在运维中的核心角色是智能助手与副驾驶，尚不具备全自动化现网代管能力。 - 工程师角色正从底层网络设备排障转向训练、调优与监管负责网络自愈的 AI 系统。 --- ### [16:33](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=993s) 自动化演进阶段：从特定痛点代理到公共生命救援的“人类 ROI” 面对“我们距离真正的自愈网络究竟还有多远”的追问，产业界的预期正在走向务实。电信标准组织（如 TM Forum）虽已定义了从 L1 到 L5 的五级网络自主化框架，其中第五级是完全自主、持续自学习与自我演化的高级网络，但要达成这一终极愿景仍需相当长的时间。当下更为可行且紧迫的技术路径，是针对那些最令运维人员痛苦、导致巨大财务损失的高频重大网络故障，构建专门的 AI 智能体（AI Agents），实现局部的特定领域自愈。 这种局部自治所带来的核心红利是工程生产力的解放。过去，电信服务商的工程师被庞大且繁杂的基础设施运维死死缠住，许多构思已久的高附加值创新业务长期搁置。通过自愈自动化将工程师释放出来并进行技能重塑，企业可以将工程力量转向能够直接创造新增营收的高价值业务。更为重要的是，AI 带来的降本效益不仅应考量商业维度的财务投资回报率（Financial ROI），还具有深远的“人类价值回报率”（Human ROI）。 以当下社会的紧急救援求助系统（如 911）为例，目前绝大部分公众拨打报警电话时依然只能与公共安全应答中心（PSAP）的话务员进行纯音频通话。从纯技术能力来看，现代通信网络早已完全具备传输实时视频、共享白板、高解析度图片和现场遥测文件的能力。如果在危及生命的重大紧急事故中，第一响应者能够即时看到现场视频或交互图像，救援效率将获得质的飞跃。这一关键能力未能普及的瓶颈不是技术本身，而是巨大的改造成本。倘若能利用 AI 在网络底层基础设施中大幅节省运营开支，并将节省下来的资金转移注入到事关公众生命安全的关键通信能力建设中，电信技术才能真正实现对整个人类社会的深远回馈。 - 电信网络自主化标准划分为五级，完全自学习自我演进的 L5 级全自动网络仍属长期愿景。 - 短期可行的突破口是构建垂直领域专有 AI 智能体，优先解决高痛点、高停机损失的核心故障。 - 基础设施降本释放的工程预算可转向公共安全等领域，推动 911 等应急救援系统向实时视频与多模态数据共享升级。 --- ### [20:34](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=1234s) 决策权移交界限：从“5毛钱分析师”到高价值决策的评估体系 对于究竟敢把多大程度的决策权放手交给 AI，企业内部必须建立一套清晰的二维量化评估体系。坐标的一维是完成该项工作所需的“工作量/精力投入”（Effort），另一维则是该项工作所创造的“业务价值”（Value），由此划分出四个决策象限： | 业务象限 | 典型任务类型 | AI 决策与自治策略 | 典型应用范例 | |:--- |:--- |:--- |:--- | | **低精力 / 低价值** (Low Effort, Low Value) | 基础报表查询、常规数据检索 | **高度放权给 AI**：允许独立查询与输出结果 | Optimum 的“50 美分分析师”Chatbot 工具 | | **高精力 / 低价值** (High Effort, Low Value) | 重复性跨系统数据清洗、流程整理 | **构建工作流自动化**：由 AI 驱动端到端流水线 | 后台日志初筛、重复文档归档 | | **高精力 / 高价值** (High Effort, High Value) | 关键架构变更、现网核心配置调整 | **极度谨慎，人在回路**：AI 仅充当建议与辅助分析 | 核心网路由调整、灾备切换决策 | | **低精力 / 高价值** (Low Effort, High Value) | 紧急阻断指令、高权限业务准入审批 | **严格的人工主导**：配合自动化防御规则熔断 | 关键安全拦截、合规性终审批准 | 在大模型固有的幻觉隐患和潜在错误输出难以彻底根除的前提下，最适合立即交付给 AI 独立执行的是“低价值、低精力”领域。Optimum 团队在此落地了一个名为“50 美分分析师”（50 Cent Analyst Tool）的典型产品。以往业务人员提出一个基础数据查询请求，人工数据分析师需要花费大约一个小时编写脚本、连接数据库并生成结果；而现在，基于底层数据源打造的对话式机器人可以在几秒钟内自主探查后端系统并返回答案，单次查询的调用推理成本大约仅需 50 美分。 相反，一旦任务进入“高价值”象限，直接让 AI 自主拍板将伴随极高风险。高价值决策必须严格保留人类专家的介入与引导。为了推动高价值场景逐步向自动化过渡，团队必须建立极为严苛的自动化评估框架（Evals），对 AI 的每一次中间决策与推理路径进行全链路追踪（Tracing），量化衡量每一次尝试的准确率，唯有在成熟度指标完全达标后，方可逐步扩大自主决策的权限边界。 - 采用“精力投入 vs 业务价值”二维象限矩阵决定放权策略，低风险低价值任务优先全面自动化。 - Optimum 研发的“50 美分分析师”工具通过让 Chatbot 自主探查后端数据源，替代了耗时 1 小时的人工作业。 - 进军高价值决策领域的前提是建立成熟的 Evals 评估体系与全流程决策链路追踪系统。 --- ### [23:33](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=1413s) 跨越概念验证（POC）鸿沟：紧扣企业战略目标才能实现规模化 在推进 AI 落地的过程中，许多企业普遍陷入了“POC 丰富，规模化为零”的困境。搭建一个酷炫的概念验证（POC）原型进行内部演示并不困难，但要真正将其推向生产环境，需要建立严密的护栏机制（Guardrails）、测试控制工具（Harness）、底层算力环境与工程技术底座。许多团队在这些工程化脏活累活上投入严重不足，急于求成，最终导致项目胎死腹中。 然而，比纯技术工程挑战更深层的阻碍在于战略脱节。企业每年都有明确的核心目标、战略优先级与 ROI 指标。如果一个团队所专注的 POC 原型从一开始就未能紧密锚定公司的核心战略或可量化的 ROI，那么它在本质上就无法为企业创造真实的商业价值。这种脱离业务主航道的尝试哪怕技术概念再新潮，最终也只能沦为自娱自乐的边缘试验。唯有从立项之初就将技术探索与企业增长蓝图和硬性财务指标深度绑定，AI 项目才有可能争取到足够的组织资源，真正跨越概念验证走向规模化现网落地。 - POC 演示与生产级部署存在巨大工程鸿沟，生产环境极度依赖完备的安全护栏与测试支撑底座。 - 大量 AI 项目无法规模化的根源并非单纯的技术瓶颈，而是立项时与企业的核心业务优先级和战略目标脱节。 - AI 项目立项必须优先锚定明确的 ROI 增长指标，否则难以获得持续的资金支持与跨部门协同。 ### [25:27](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=1527s) 从试点走向生产：数据管道、权限管控与工程落地挑战 Sundar Krishnan 指出，将 AI 项目从试点（Pilot）成功推向生产环境，首要前提是全面对齐资源并配备能够搭建数据管道的专业人才。很多团队往往把注意力单纯放在数据输入和概念验证（PoC）上，却忽略了后续一系列繁杂的中间工程步骤。这包括如何实施基于角色的访问控制（RBAC）、如何确保正确的人员以合规方式访问数据、如何优化系统延迟，以及如何确保终端应用能够稳定做出正确决策。 正因为缺乏对这些工程细节的通盘考量，许多企业才会在试点阶段陷入停滞，甚至在取得初步进展后直接将产品束之高阁，指望未来有人来接手收尾。Sundar 强调，企业必须在立项之初建立起端到端的全局视野，提前筹备好所有配套资源，才能确保试点项目真正平稳过渡到生产环境。 ```mermaid flowchart TD accTitle: AI 从试点走向生产的落地路径 subgraph S1["试点与验证 (Pilot & PoC)"] A["数据接入与模型原型"] --> B["PoC 功能验证"] end subgraph S2["工程化与合规关键要素 (Engineering Core)"] C["数据管道 (Pipelines)"] D["角色访问控制 (RBAC)"] E["延迟优化 (Latency Control)"] F["决策与应用逻辑校验"] end subgraph S3["生产规模化 (Scale to Production)"] G["商业化部署与持续监控"] end B --> C C --> D D --> E E --> F F --> G ``` --- ### [26:28](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=1588s) 电信历史视角的规模化：从 1G 到 5G 的试点与 AI 落地 主持人 Cathy White 向拥有 40 年电信行业经验的行业顾问 Anne Lee 提问：媒体和业界经常渲染“90% 到 95% 的 AI 试点最终走向失败”，纵观电信技术数十年的演进，究竟是什么决定了组织能否成功将新技术规模化落地？ Anne Lee 回应称，试点的失败率高并非 AI 所独有，这在整个电信史乃至所有工业领域都是常态。回顾移动通信从 1G、2G、3G 到 4G 的每一次代际跃迁，或是推出任何一项全新产品，行业总是从原型设计（Prototype）起步。原型成功仅仅意味着新技术在实验室或小范围可行，但要让其成为可行的商业产品，必须解决规模化、网络容量（Capacity）和极高可靠性（Reliability）等一系列严苛考验。 将一个技术原型或最小可行产品（MVP）打磨成可规模化商用部署的系统，其工程量和复杂度往往远超前期的技术验证本身。Anne 表示，电信行业早已在历次技术变革中积累了极其丰富的规模化落地经验，“在过去，我们面对的是其他通信技术，但我们成功做到了；面对 AI，我们也同样能够做到。” --- ### [27:58](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=1678s) 遗留基础设施改造与区域化渐进式 AI 部署 谈及电信网络的现实困境，主持人将话题引向长期存在的遗留基础设施（Legacy Infrastructure），探讨企业如何在不彻底推倒重构整个网络的前提下实现现代化。 AOI 的 Sahil Yadav 分析道，尽管有线电视（Cable TV）与光纤网络的底层架构与其他电信网络略有差异，但遗留资产的包袱是全行业共通的痛点。AI 运行的基础是数据，网络需要源源不断地从节点（Nodes）、路由器（Routers）、放大器（Amplifiers）等设备中提取遥测数据（Telemetry Feed）并输入数据管道。然而，现网中大量旧设备既缺乏捕获数据的传感器，也没有远程管理能力，沦为事实上的“哑设备”（Dumb Devices）。 将这些老旧哑设备替换为具备感知和联网能力的智能设备，对于 Comcast、AT&T 等大型运营商而言意味着庞大的资本支出（CapEx）。因此，行业普遍采取分阶段、按区域的渐进式升级策略： - **限制故障爆炸半径（Blast Radius）**：先在特定区域升级硬件并部署 AI，若算法发生异常，其负面影响被严格限制在局部范围。 - **获取真实外场实测数据（Field Experience）**：在真实网络环境中验证并迭代 AI 模型，待系统成熟后再全网铺开。 --- ### [31:42](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=1902s) 现场提问：网络优化诊断与离线边缘 AI（Edge AI）协同 现场观众提问聚焦于两个核心方向：一是 AI 如何具体改变网络优化（包括传统的路测 Drive Testing）；二是边缘 AI（Edge AI）在无网络连接的极端情况下，能否直接在移动设备本地充当 Copilot，并在网络恢复后同步数据。 Sundar Krishnan 首先就网络优化做了解答。他透露，当前运营商正利用 AI 实时预测信噪比（SNR）以及特定片区内联网家庭的服务受损情况，从根本上预判断网风险并分析诱因。这种预测能力直接重构了工单派发逻辑——传统模式下用户稍有投诉就必须派出工程车（Truck Roll），成本极其高昂；如今系统能智能判断是否真正需要外勤排障，并优先引导用户通过 App 执行自愈操作或利用视觉 AI（Visual AI）拍摄调制解调器（Modem）指示灯获取排障指导。 针对离线边缘 AI，Sundar 表示这正是行业演进的大势所趋： - **模型量化与微调**：1-bit、2-bit、4-bit 量化技术以及模型蒸馏（Distillation）技术日趋成熟，开源社区已涌现出大量超轻量级小型语言模型（Tiny LLMs），完全能够脱机部署在设备端进行本地推理与决策。 - **杰文斯悖论（Jevons Paradox）在 AI 领域的体现**：随着资源利用效率的大幅提升与边缘模型的极致压缩，技术的普及并不会减少算力消耗，反而会加速创新并激发出成倍增长的新应用场景，形成良性循环。 Sahil Yadav 进一步补充指出，**智能体架构（Agentic Architecture）**极大地释放了边缘计算的潜力。工程师可以针对特定任务构建体积极小、专事专办的轻量化智能体（Lightweight Purpose-built Agents），在极其有限的边缘硬件算力与存储空间内，以高度紧凑的形式并发执行复杂业务动作。 --- ### [35:51](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=2151s) 人力价值转型：从拉斯维加斯出租车对话看电信工程师的未来 会议接近尾声，主持人 Cathy White 分享了一个生动的插曲：前一天她乘出租车前往威尼斯人酒店（The Venetian）时，司机得知她是来参加 AI 大会的，便立刻焦虑地追问“未来的工作岗位会变成什么样”。Cathy 借此将讨论引向 AI 浪潮下电信行业员工与工程师的未来角色定位。 Anne Lee 幽默地回应：“我甚至在想咱们是不是坐了同一辆出租车，还是说拉斯维加斯所有的出租车司机都在思考 AI 的问题。”随后她严肃指出，AI 的普及实质上是将工程师从低效繁琐的重复性劳动中解放出来，企业应当主动推动工程师的技能重塑（Upskilling）： - 转向高附加值服务（Value-Added Services）的开发，例如为应急通信服务（Emergency Services）打造更强大的增强通信功能，在关键时刻挽救更多生命。 - 改造并革新现有的传统交互式语音应答系统（IVR）等客户触点，利用 AI 重塑电信交互与运维体验。 ### [37:51](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=2271s) 传统 IVR 的痛点与 AI 驱动客服的信任构建 探讨从传统交互式语音应答（IVR）系统的痛点切入。现有的客服电话体验经常令人抓狂：用户不得不按照语音提示“按 1 选择此项、按 4 选择该项、再按 5”，在层层嵌套的菜单中不断深入，流程冗长且极其繁琐。如果能引入具备 AI 能力的智能 IVR 系统，用户便无需再进行这些机械的按键操作。 虽然目前用于客户服务的 Chatbot 仍不完美，但这恰恰指明了未来的优化方向与就业机遇。从人力与技能转型的角度来看，转向 AI 并不意味着工作岗位的消失，而是催生了新的需求——行业需要专业人员去开发、完善并调优这些新型 AI 服务，从而显著提升交互质量。对于任何面向客户的业务而言，这类 AI 服务既能为企业降本增效，也能大幅改善用户体验。然而，面向公众部署 AI 的另一大核心挑战在于如何跨人群建立普遍信任，任何新技术的普及都会面临阻力，电信行业未来如何应对这一信任门槛将至关重要。 ```mermaid flowchart TD accTitle: Traditional IVR vs AI-Enabled Smart Customer Experience subgraph Traditional["传统客服交互痛点"] A["多层级语音菜单 (按1、按4、按5)"] --> B["漫长等待与繁琐按键"] B --> C["用户体验差 & 效率低下"] end subgraph AIEvolution["AI 智能客服演进"] D["AI 智能语音 / 语义理解 IVR"] --> E["精准意图识别与即时响应"] E --> F["AI 研发、优化与服务打磨岗位"] F --> G["建立公众信任与流畅客户体验"] end ``` ### [39:29](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=2369s) 闪电问答：未来五年习以为常的电信能力与退出历史的 AI 热词 进入小组讨论尾声的“闪电问答”（Lightning Round）环节，主持人 Cathy White 向嘉宾们抛出快问快答。针对“未来五年内哪项电信能力会被大众视为理所当然”这一问题，Anne Lee 认为最直观的变化将发生在公众接触面：未来的客服交互中，AI 将变得无比自然，以至于用户完全不会意识到后台是 AI 在运转，更不会产生“真希望现在能转接人工”或者“千万别让我按 1、按 3、按 6 走多层菜单”的无奈感。人们将能直接与 AI 流畅对话，且系统绝大部分时间都能准确理解意图，这种无缝体验将变成一种日常习惯。主持人 Cathy 对此深有共鸣，并打趣道自己十分期待那一天的到来，这样就再也不用在联系英国税务海关总署（HMRC）时被迫听上 20 分钟的电梯等待音乐了。 随后，主持人询问 Sundar Krishnan 最希望淘汰哪一个 AI 行业热词。Sundar 明确指出是“提示词工程”（Prompt Engineering）。他结合近期从 Anthropic 团队关于 Claude Code 的技术分享中提到的进展解释道，开发团队在构建 Claude Code 系统时已经移除了 80% 的提示词设定。这表明随着全自主智能体（Autonomous Agents）的普及，依赖人工精细调整 Prompt 的阶段很快将成为过去，系统将具备自主决策能力，提示词工程将逐步退出历史舞台。 - **未来五年普及体验**：用户与能够精准理解人类意图的 AI 客服无缝对话，彻底告别传统多层 IVR 按键与漫长的人工排队等待。 - **即将被淘汰的流行词**：提示词工程（Prompt Engineering）；随着智能体自主能力的提升，依赖人工调试 Prompt 的开发方式将被自动化决策取代。 ### [41:57](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=2517s) 被低估的 AI 治理危机与圆桌结语 面对“当今哪项技术或领域被严重低估”的问题，Sahil Yadav 指出公众与行业最容易忽视的其实是“AI 治理”（AI Governance）领域。近期关于模型越狱、失控乃至恶意攻击多家企业的相关报道，已经凸显了安全隐患。他引用了 Elon Musk 的知名警告指出：如果 AI 模型失去控制，其潜在危险性甚至可能超过核导弹。 Sahil 强调，AI 治理在未来必须获得极高重视。目前学术界和工业界已经展开了相关防御研究，例如可解释性人工智能（XAI, Explainability AI）分支，其核心目标就是为 AI 模型建立安全护栏（Guardrails）与合规框架。 在小结阶段，主持人 Cathy 幽默地调侃原本设计闪电问答是为了以“高昂且充满活力”的氛围收尾，没想到最终结论落在了“AI 可能比核武器还危险”的沉重预警上。最后，主持人向与会观众致谢，提醒大家可移步会场外继续与嘉宾交流探讨，并在热烈的掌声中为本场电信与 AI 变革圆桌研讨会画上句号。 - **被低估的领域**：AI 治理（AI Governance）与安全控制体系。 - **关键技术研究分支**：可解释性 AI（XAI），致力于为自主模型设立安全边界与防护栏。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=4s) 谢天谢地，你们来对房间了。 — Cathy White > [01:05](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=65s) 高能效的持续学习智能体 AI 框架 — Anne Lee > [06:25](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=385s) 自动化是会让人们感到恐惧的（Automation is something that scares people） — Sahil Yadav > [09:44](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=584s) AI 节食（AI Diet） — Sundar Krishnan > [12:36](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=756s) AI 正在节食（AI is on a diet） — Sundar Krishnan > [26:28](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=1588s) 在过去，我们面对的是其他通信技术，但我们成功做到了；面对 AI，我们也同样能够做到。 — Anne Lee > [35:51](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=2151s) 我甚至在想咱们是不是坐了同一辆出租车，还是说拉斯维加斯所有的出租车司机都在思考 AI 的问题。 — Anne Lee > [39:29](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=2369s) 真希望现在能转接人工 — Anne Lee > [39:29](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=2369s) 提示词工程（Prompt Engineering） — Sundar Krishnan > [41:57](https://www.youtube.com/watch?v=Adp2YgYJ0As&t=2517s) 如果 AI 模型失去控制，其潜在危险性甚至可能超过核导弹。 — Sahil Yadav ## Mentioned Cathy White, CEW Communications, Sundar Krishnan, Optimum, Comcast, Sahil Yadav, Applied Optoelectronics Inc (AOI), Anne Lee, Uber, TM Forum, 911, AT&T, Claude Code, Anthropic, Elon Musk, 英国税务海关总署 (HMRC), 威尼斯人酒店 (The Venetian) ## Video description Cathy White - Founder & CEO, CEW Communications Sundar Krishnan - Director AI/ML, Optimum Sahil Yadav - Head of Software and AI, Applied Optoelectronics Inc (AOI) Anne Lee - Consultant, Consultant This panel highlights how telecoms are applying AI from core infrastructure to customer experience and backend processes. Learn how AI enables real-time diagnostics, predictive maintenance, and smarter CX—while reducing churn and boosting margins. Leaders will also explore how they’re preparing for AI-heavy 5G and edge deployments at scale. Join this panel to hear from leading technology executives and telecom leaders who are shaping the future of the industry. [Source](https://www.youtube.com/watch?v=Adp2YgYJ0As)

[0:04] Cathy White: 谢天谢地，你们来对房间了。 [1:05] Anne Lee: 高能效的持续学习智能体 AI 框架 [6:25] Sahil Yadav: 自动化是会让人们感到恐惧的（Automation is something that scares people） [9:44] Sundar Krishnan: AI 节食（AI Diet） [12:36] Sundar Krishnan: AI 正在节食（AI is on a diet） [26:28] Anne Lee: 在过去，我们面对的是其他通信技术，但我们成功做到了；面对 AI，我们也同样能够做到。 [35:51] Anne Lee: 我甚至在想咱们是不是坐了同一辆出租车，还是说拉斯维加斯所有的出租车司机都在思考 AI 的问题。 [39:29] Anne Lee: 真希望现在能转接人工 [39:29] Sundar Krishnan: 提示词工程（Prompt Engineering） [41:57] Sahil Yadav: 如果 AI 模型失去控制，其潜在危险性甚至可能超过核导弹。

[8:21] Outcome Metrics Over Usage [8:21] Proactive Subscriber Retention [8:21] Selective Scaling [13:29] Shifting Resources to Growth [13:29] Human and Social ROI [21:42] Sundar Krishnan: 50-cent analyst [35:51] Cathy White: press 1 for sales, press 4 for support [39:29] 5-Year Prediction [39:29] Buzzword to Retire [39:29] Underestimated Field [42:39] they can be more dangerous than nuclear missiles themselves [42:39] Cathy White: lightning round [42:39] Cathy White: high energy [42:39] Cathy White: AI will be more dangerous than nuclear weapons [43:27] Cathy White: Don't come to me because I'm not the expert.

[0:04] So, my name is Kathy White. I'm the founder  and CEO of CEW Communications. You can tell [0:09] by my accent I am not from around these here  parts. And I'm uh I'm visiting from Europe, [0:15] first ever conference in Vegas, learning a lot  along the way. I'm very delighted to be here and [0:20] joined by my fabulous panel. And today we're  going to be talking about AI and Telkom. and [0:26] we're going to talk about where we're at right now  in terms of the present and where we're going in [0:30] the future. And before I get my lovely panel to do  a brief introduction to themselves, we would like [0:37] to have you do an introduction to us. We'd love to  understand who you are in the room. So, would you [0:43] mind raising your hand for me if you are working  in telco? Thank god you're in the right room. All [0:52] right. Who is already deploying AI within their  telco? Okay. And who is considering it and doesn't [0:59] quite know where they're beginning? Okay. They're  not here. They probably should be. All right. So, [1:05] I'm going to start with Sahil. Sahil, are you  happy just to do a very brief introduction to [1:10] who you are, what your organization does, and my  bonus question for you is what's the single most [1:16] important AI initiative you're focused on today?  Sure thing. mic is working. Yeah. Uh I'm Sahilia [1:24] Dove. I work for a company called AOI. Uh we are  manufacturers of amplifiers for telecommunication [1:33] service providers such as Comcast and others and  uh we along with building the hardware we build [1:40] all the software infrastructure as well and and we  are one of the foremost vendors for that. Um the [1:46] question you asked was uh if you can ask one more  time. Single most important AI initiative you're [1:52] focused on today. Sure. Uh I lead AI in my company  and uh one of the products that we're building [2:01] uh relates to self-healing or self optimization  as we call it. And uh the biggest initiative is [2:07] AI is out there on dashboards. We want to put AI  in the field in action um driving some of those [2:15] autonomous actions in the networks. Awesome. Thank  you. An hi, my name is my name is Ann Lee. Um [2:25] I've been in the telecom industry for almost four  decades and um starting with analog 1G to digital [2:35] 2G then broadband 3G then voice over IP 4G and  then I started working on artificial intelligence. [2:43] Uh I I mentioned I'm retired. [laughter] [gasps]  So I just just retired. Uh so I guess the most [2:53] important thing that I've been working on I  guess at at the moment is uh I'm doing this [2:57] panel. Uh but the the the research project I was  working on before I retired was to um the the [3:08] project was was to create a uh energyefficient  continual learning agentic uh AI framework uh [3:17] to enable fully autonomous networks. Yeah. And  I'm I'm impressed we've pulled you away from [3:23] retirement to join this panel today. So thank you  for that. Sunda. [snorts] Hi everybody. Uh this is [3:29] Sunda. I've been working in the telco industry  for more than seven years. Started at Comcast, [3:36] been there for a while and then after that  recently I'm leading a team AML team at Optima. [3:42] So to answer your question, we focus more on like  the hyperpersonalization recommener systems that [3:49] guide how agents make offers to customers when  they're calling us. That's one of them. And then [3:54] like digital twins if you if you have heard that  topic those bus buzzwords these these years and [4:00] also the H&TK solutions at scale. So those are  some of the work that we focus on at Optimum. [4:07] Glad to be here. Nice to meet you all. Thank you.  So we're going to start with where AI is creating [4:13] real value today in Telcom and and Sundra I'm  going to start with you. So you know at this point [4:19] we've been talking about AI for many many years  hence there's this massive conference in Vegas. [4:23] Um where have you actually seen AI move beyond  experimentation into production if you have? So [4:31] I think the one uh I mean the work use case that  we are working on right so hyperpersonalization [4:38] um previously we used to have like rulebased  recommener system that was like guiding agents [4:43] to tell what to tell c customers when they're  calling us hey this is the issue that they're [4:48] calling about then you can recommend this. But  now after years of like experience being in this [4:54] field what I see is like we started with machine  learning models guiding our agent decisions but [4:59] now we have moved towards AI giving the talking  points for the agents to talk to customers plus [5:05] telling them how to make the offer and how to  understand what the customer is talking about [5:11] where they are coming from. So we we simulate all  those parts for the customer. So you could be a [5:17] happy customer calling us to upgrade a product or  you could be like a frustrated customer talking [5:22] about your service issue or an internet issue and  we need to know beforehand why you are calling us [5:28] so that we can make the right choices for you  in a limited amount of time and that's where [5:32] the hyperpersonalization aspect of AI has come  a long way and that's that's one of the things [5:38] that we are actually like you know very cool about  like putting putting things together and then the [5:43] other thing is actually also like as I was telling  about the digital twins and simulations, right? [5:50] We actually are thinking what the customers will  take will do or will react to how we kind of put [5:58] an offer in front of the customer, right? So like  that we can do based on the digital twins like [6:02] we're not asking our real customers and we are  not seeing how they will react to certain things [6:07] rather we are asking our agents kind of customers  like I can say the digital twins and then making [6:13] those decisions beforehand. So that way we can  make those personalizations and other things at [6:20] scale before the customer calls us. So glad that  could answer the question. I'm glad you could [6:25] answer the question too. Excellent. Um Sahil,  you're much closer to infrastructure. Where is [6:31] AI fundamentally changing the way telecom networks  operate and not just making existing processes [6:38] faster? Absolutely. On the infrastructure end,  we're looking at massive transformation in terms [6:46] of uh automating network management. Uh we're  looking at we've been talking about several use [6:52] cases over the years from predictive maintenance,  anomaly detection, uh self-healing if you will. [6:57] Uh but we're getting closer and closer into uh  managing networks more autonomously. What that [7:04] means is um not only AI is able to identify  issues whether proactively or reactively [7:11] but also has the ability to identify solutions  that can be deployed. Um it's not an easy road [7:18] to take especially if you were to draw some uh  comparisons against uh uh uh the world that we [7:25] operate in which is uh cable TV and and fiber uh  which is slightly different from wireless in a way [7:32] that we have amplifiers instead of routers and uh  you can route traffic around a router when there's [7:39] something wrong but if if an amplifier breaks  there's no rerouting. So um our AI capabilities [7:46] are very unique to our industry. Uh at the same  time they make it more mission critical as well [7:53] cuz nobody likes to have downtimes. Nobody likes  to have outages. So um the road that we're taking [8:01] in is important from the standpoint that uh we're  getting a lot of push back from customers as well [8:06] to put in the live environment because automation  is is something that that scares people. But at [8:12] the same time it is it is the evolution into  the next generation of the networks. Yeah. So [8:21] an I'm going to get you to do a bit more of  the kind of zoomed out view looking across [8:25] the industry today. Where do you think telecom  companies are making the smartest AI investments? [8:33] So um one of the areas as Hill was saying uh  is really in the area of network operations [8:44] and management. Um, we all don't like it when  the network goes down, right? So, we'd like to [8:53] minimize that as much as possible, even eliminate  it if if possible. So if it's poss if we can use [8:59] AI in order to recognize that something has gone  wrong and then also take action to self-heal if [9:08] it's possible to to do so or at least alert a  human engineer to go swap out a card [snorts] um [9:15] and do that as quickly as possible. [laughter] Uh  that that's important. Um also using it for the [9:24] purposes of co customer retention and uh you know  doing personalization for customers so that the [9:32] um the subscriber wants to stay on your  network and doesn't want to switch to another [9:38] uh another provider. That's very important  from a business perspective. Um so those are [9:44] some areas in which it makes sense to apply  AI. Okay. Sunda are we are we kind of moving [9:54] beyond AI hype at the moment you know or is  most of the industry still experimenting. [10:01] Now I think we have come a long way right now.  Right. So first I think like I mean if you're [10:10] if you're asking me we are moving along the AI  hype the first thing that I'm seeing here is like [10:15] customers are now they know what they want but  then they don't want to go through the tedious [10:21] process that they have to go through to understand  or to explain to an agent what they want right [10:27] and so essentially what I'm seeing now is we have  this lot of data that is being used by companies [10:34] to make proactive decisions on on behalf of the  customers. So some of the things could be like [10:41] you could have um one of the things like uh where  where do you think the customer is going to make a [10:50] product choice or when when they're going to churn  all of all of those things we now know beforehand [10:55] proactively. So what we are now doing is instead  of like waiting for the customer to call we are [11:02] actually letting calling them and giving them  choices that they could take and that is only [11:08] possible because of all the things that has come  a long way as I mentioned before right and so that [11:14] AI is not a hype it is actually things in action  that actually turns your business around right so [11:20] essentially you're not sitting and waiting for  things that you want to implement to plan for [11:26] an ROI right for your company or not. Rather than  that, what you're doing is you're taking actions [11:30] and proactive steps like like both of these  people said like the network maintenance and [11:35] other things. All of them are like uh things that  you can do proactively to make sure that one your [11:41] network is always up. Two, your customers are  always happy and then three um you are taking [11:47] the choices on behalf of the customer before even  they call to us and tell us what is the what is [11:51] their problem. when all of this is actually  happening because of not because of AI hype [11:56] but because we can make choices and things in  action that can be done using AI actually does [12:03] it help that does help but I'm going to I'm going  to give you a prompt AI diet we briefly touched [12:10] on this I know I know I know so one of the things  that we also I was telling them before which is [12:16] actually like companies were actually spending a  lot of money on AI in the start of the year. So [12:23] you might have seen like you know Uber spending  uh creating lead leaderboards and tracking usage [12:30] of employees across EI and other stuff right. So  you would actually see a lot of companies spending [12:36] on AI just because they want to spend on AI. That  was the start of the year scenario but now if you [12:41] go and read articles around AI what you're seeing  is AI is on a diet. And then what people are doing [12:48] now is like they're scaling back on choices that  they have to make. And especially they are trying [12:53] to make sure one can we actually measure the value  of AI through the features that are produced from [13:00] AI right and then the second thing is actually  also how can we make sure that the AI is taking [13:06] the right choices and the right decisions by by  tracking the AI usage across the company and so [13:12] all of this is possible just because companies are  now thinking smarter not just because they want to [13:17] spend on AI and also like trying to make sure  that they're investing their choices wisely by [13:22] create by looking at the features and not just  the to leaderboard on token usage. So, so I'm [13:29] I'm going to switch us a little bit and talk more  about smarter networks and and automation and sil [13:34] go back to one of the things that you were talking  about previously around self-healing networks. Um [13:39] can you explain a little bit more about what that  actually might look like and also how that AI is [13:45] going to change the role for engineers within  telco? big question there. Oh, you know, it's [13:53] an it's a conference about AI. Um, self-healing  uh is is the industry term. It means um a lot of [14:02] things in different contextes. But at a broader  level, I what when we talk about self-healing, [14:08] what we mean is AI is able to identify issues. It  is able to identify resolutions and deploy that [14:15] those resolutions in autonomous way. Um, when we  think about running tests in the lab, that's the [14:23] definition that you'd go with. But when you talk  about deploying in the field, it's a completely [14:27] different picture cuz you talk about all sorts  of things in terms of uh the implications of of [14:34] deploying wrong fixes, uh, implications  of of causing outages and and all that. [14:41] Um usually we would use a something called as a  risk matrix in terms of um deploying AI in the [14:48] field starting with lowest risk uh profile giving  those jobs to AI uh medium risks you keep human on [14:57] the loop and then high risk you keep human in the  loop and that's how the that's the that's the way [15:04] of deploying automation in the field. Um having  said that with all of this out there AI providing [15:13] all this information AI is as I see today is a  companion to the sales engineer. Uh it can provide [15:22] information it can guide the sales engineer but  it's not at a level where it can automate things. [15:29] So um and that is a natural progression into  consuming AI cuz the other way to think about this [15:36] is uh AI is getting feedback from the engineer  as well while he's using it as a companion and [15:44] AI is also getting some field exposure as opposed  to just running in the labs. So that is improving [15:49] AI and building trust on AI. Now with armed with  all of this information if if and when we get to [15:57] self-healing part um at that point in time AI will  be able to do much more. So the role of the sales [16:04] engineer while having a lot more information now  will uh will change uh to being able to not just [16:12] manage the networks but also manage the AI that is  running that network. So the skill set will evolve [16:19] over time. uh skill set will include learning  about machine learning models, interfacing with [16:27] AI and so on so forth. And I'm going to ask  you a question that I asked you earlier and [16:33] you didn't like me asking the question, but I'm  going to ask the question anyway. Um how close [16:36] are we really to self-healing networks? Is there  a estimated time? Um is there an estimated time? [16:47] So one of the things we did discuss was that um  we may not get to so there is a standards forum [16:56] uh for telecom that has defined these five levels  of automation networks and the and the fifth one [17:03] is fully automated self- evvolving uh continually  learning network that's going to be a while before [17:11] that happens. We did have this discussion about  how um what may be realistic and maybe narrower [17:19] term is really focusing on some of the top issues  that does cause a lot of um pain and maybe costly [17:32] from a financial perspective that we should create  AI agents that are capable of recognizing and [17:39] then taking action to resolve those issues. So  that kind of self-healing of specialized areas [17:45] um I think is close closer term um and and once we  do have some form of you know further self-healing [17:56] than what we have today uh Sahil had mentioned  the the the change to the evolution to the way [18:05] engineers are doing their job. I think that we  should also recognize that um one of the reasons [18:14] that people state that they want to also be able  to reduce the amount of work for these engineers [18:21] is that they can then shift the engineers focus,  reskill them, shift their focus to more value [18:27] added type services. And those value added type  services could be services that make more money, [18:34] bring more revenue to the customers or to the  service provider, which has been, you know, in the [18:40] to-do list for a while, but just haven't gotten to  it because there's costs and you're busy managing [18:48] the network as opposed to creating these value  added services. Um and in some cases maybe there [18:54] isn't necessarily um a financial ROI but there's  a human ROI to these new services. For example, [19:02] um emergency services we talked about today when  you call 911 um for the most part you basically [19:11] talk to the uh the PAP operator. But if you look  around you uh you you do know that communication [19:21] services are capable of more than just audio.  They can do video uh live video or recorded video, [19:30] shared images, shared whiteboard, shared files.  Imagine if in a emergency ser emergency use case [19:39] scenario um where there it might be life critical.  It might be helpful to the person who needs help [19:52] if if the per if the if there were images or  videos that could also be conveyed to the first [19:58] responder or to the to the 911 operator. But that  doesn't exist right now. And part of the reason is [20:05] costs. Uh right, it's not technical because that  capabil those capabilities exist technically. [20:13] So if we are able to reduce cost somewhere  else and then shift the funding then to do [20:19] some of these other capabilities including  life critical services I think that would be [20:26] um a win-win for for society. So I we are going to  touch a little bit more on the the human element [20:34] of this all in a little bit. Um before we  get there though Sundu I wanted to to ask [20:39] you how much decision making you're comfortable  handing over to AI today you know is AI supporting [20:47] engineers is it making decisions itself what's  your what's your comfort level so I think that's [20:53] why we are all here right um if you have to think  of your work and kind of put it in a two-dimension [21:00] perspective right where where one dimension is  like effort that it takes takes to complete that [21:06] work and then the other dimension is like the  value that effort creates right there's also [21:12] another dimension which I'm missing obviously the  priority aspect of it which you can ignore for now [21:16] but when you put that into those two dimensions  you have these four quadrants which is the low low [21:22] high obviously high high and then high low and  so because you we all know that AI systems has [21:30] a tendency to hallucinate and it can also produce  sometimes the long wrong results without the right [21:35] evils it will not work right. I think where I feel  comfortable today handing off my decisions to AI [21:42] is on the low low area where I can actually let  the AI systems come up with its own solutions at [21:48] certain points because there obviously the value  is also low and I don't want to spend resources [21:55] um spending there. So essentially what we did back  then at Optimum was we created something called [22:01] like a 50 cent analyst tool. So essentially any  questions that you want to ask an analyst would [22:07] have to spend like an hour or so to create that  solution. We now have a chatbot that can actually [22:12] go and explore the data sources behind the scenes  and get that answers and it spends 50 cents for [22:17] answering those questions roughly give or take.  So that's the kind of capability I would say you [22:22] can put there. But if that is going to be used  to challenge some of the decisions that you're [22:27] going to make which is I would say then it falls  into the high high value right high value bucket [22:34] that's where you have to be very cautious about  letting an AI make its own decisions versus you [22:39] there's a human in the loop aspect of it right so  I'd be pretty comfortable putting the AI systems [22:44] I would say low low and low uh value again high  effort there you can create some workflows to [22:51] actually let the AI do those decisions. But  then when it when the value shifts to high, [22:57] it should always there there should always be a  human-guided effort to make the AI systems get to [23:03] a point where it is mature. So how do we get to  that mature state is based on the evas that you [23:08] put in place tracing each of those decisions that  it is making. Um trying to make sure that we are [23:15] able to measure its goal measure its success each  time and then making it to the either moving into [23:22] the low low or the low low value high effort  kind of thing. So that's that's how naturally [23:28] things should progress and I think today it's a  kind of like a hybrid mix especially in the high [23:33] high value bucket. So that brings us to scaling  AI right the big challenge right so we know many [23:41] many organizations seem to have pilots few scale  well um I'm starting with you again so be ready [23:51] uh bring it on what is what is usually the  biggest obstacle right to success in terms [23:57] of scaling these pilots is it the technology is  it the data is it leadership what is it so um [24:03] most of the time what I see And I think this is  not just with respect to telecom right it is an [24:08] industrywide issue which is most of the time  people start spending time on building a P. [24:14] Um that is always cool to showcase as a demo but  then when it has to translate to production you [24:22] need to put certain guardrails around it which  you can call it as harness you can call it as [24:27] the environment the technology whatn not right and  that's where people spend little less time and uh [24:35] try to rush and that that that's not going to make  it happen so before answering that scalability [24:41] aspect of it right and if you step back and  think about from your company perspective, [24:46] you would always have an ROA associated with each  of the efforts that you're going to that you're [24:52] going to pursue this year or like I should say  the priorities, right? And so if that pilot or [24:58] the P that you're working on doesn't stick to  any of your goals or priorities, then I don't [25:05] think it's producing value for your business at  this point in time. So you might have to think [25:09] yourself or ask the question again is this the  right P to pursue right it can be fancy it can be [25:16] like things that you want to explore that you can  always do separately but again it doesn't support [25:21] your company's growth or priorities and so tie  that to your ROF first that way you can get the [25:27] resources and everything that you need to scale it  to production that's the first first thing and the [25:32] second thing is actually having the right talent  for you to kind of build that pipelines for you [25:38] It's not just that like you have your data coming  into that and then you get the PC done but rather [25:43] than that how do you put the the role based access  control and everything how do you make sure that [25:49] the right people have right access to the data  and then the latency aspect of it and what are [25:55] the applications that you're going to use to make  sure that you know the application is making the [26:00] right choices actually right at the end of the  day. So all of these things require intermediate [26:05] steps which people don't always think and that's  why most of the people get stuck um at the pilot [26:11] process or even like they make some progress  and then put their product in shelf and wait for [26:15] someone to come it up come and take care of the  rest of the thing. So get that holistic picture [26:21] first align your resources and that way I think  you can always commit your pilot to production. [26:28] And I'm going to turn to you. So, so you shared  this with me earlier. Obviously, you're retired, [26:32] but somehow back on a panel [laughter] and and  you've got four decades of experience across the [26:38] sector and you've seen a whole mix of different  technologies come into play with telecom. [26:44] Um we are talking specifically about AI obviously  but you know looking at at what you've seen across [26:50] your career what separates the organizations  that successfully scale a pilot and whether [26:57] it's AI not AI from those that don't. Yeah.  Um I was thinking a little bit about this and [27:05] from a scaling perspective and the whole reason  why there's this concern about 90 to 95% of AI [27:14] pilots failing is because of the headline news  that 90 to 95% of pilots fail. But I think that [27:23] um in all industries if you think about it  including telecom and over the generations [27:31] when we have a new let's say we have a new G 1 G  2G 3G 4G or we have a new product we always do a [27:40] prototype and many times the prototype may succeed  or fail but when they succeed it means that you've [27:51] gotten that new technology, that new technology  to work, but you know that in order for it to [28:00] be viable, it needs to scale and be deployable,  commercially deployable. Um, and so it needs to [28:10] address capacity, it has to address reliability,  has to, you know, all those things for a scaled [28:17] commercial deployment is important. And we all  are familiar with that. And that is probably [28:24] even more work than getting the new technology  to to uh to MVP or to a pilot. And so I think [28:34] that if you think of it from that perspective  that we have this experience of getting pilots [28:41] to scale. In this case, we're talking about AI.  But in the past, it was a different technology, [28:50] but we've done it. And so, we can do it again. We  can do it again for AI. Sahil, it's been a while [28:58] since I've asked you a question. I'm coming  back to you. Um, legacy infrastructure not [29:02] going away for now. Are you all right? You keep  you raised. I'm sorry, that was a question. Oh, [29:08] I'm sorry. Can I come back to you in a second?  Thank you. Um, how do organizations modernize [29:13] without rebuilding their entire network? And I'm  just brief pause. Can you come closer and then I [29:20] will get your question in a second. Thank you very  much, Sil. I just enjoying my moment of quietness, [29:25] but thanks for the question. Give you an extra  beat. There you go. Um, good question again. Uh, [29:32] you talk about infrastructure. As I was  saying before, uh the cable TV and the fiber [29:38] infrastructure is slightly different compared  to the rest of the industry. But I think uh the [29:42] challenge of legacy infrastructure is uh is is is  faced by by all sides of telecom. And I say that [29:50] because when you talk about scaling uh especially  uh the first problem that you got to solve is if [29:57] you want to run AI you got to feed data to AI and  for that purpose you need to build a data pipeline [30:04] to be able to collect information out of your  devices um nodes routers amplifiers what have you [30:12] uh to be able to feed into AI so that AI is  providing useful results. Now the challenge [30:20] with the industry is that because of the legacy  infrastructure either the infrastructure is [30:25] incapable uh such such that there are no sensors  to capture that data translate that into uh a tele [30:35] telemetry feed or perhaps uh the devices are just  so old that they are not capable of being managed [30:43] remotely and so on so forth. So the biggest  problem that the industry is facing right now is [30:50] um there is a need to replace these devices for  the lack of a better term dumb devices with smart [30:56] devices and that involves a capex expenditure from  from from companies like Comcast AT&Ts and so on. [31:05] So it's a big expense of course um changing  your entire network but at the same time the [31:11] way companies are approaching this they're doing  it in phases they're doing it in in in parts in [31:15] regions and that is a good way for AI to scale as  well cuz uh once you upgrade infrastructure in a [31:22] particular region then you can deploy AI uh you  can uh limit the uh the blast radius of the AI as [31:30] well in case something goes wrong but at the same  time you can get some real field experience for [31:34] the AI learn and then scale across your entire  network. Thank you. Where's the gentleman that [31:42] raised his hand? Hello. Um do you want to borrow  my microphone? You can ask a question quick. [31:52] Hi there. Uh I was asking about uh how  is AI impacting network optimization [31:58] specifically drive testing and then the  second question is uh what about edge [32:03] like AI at the edge bringing it down to  the actual mobile device and without the [32:08] presence of internet can you still at  least have a companion and a co-pilot [32:13] on a device and then have it upload as you  do retain or regain internet uh connectivity [32:21] I I can go I mean I can share my view on that  right on the network optimization side of it [32:28] right so we from our from our side what we are  doing is actually kind of like understanding [32:34] when the network is going to go out of network  right so essentially we predict the signal to [32:40] noise ratio for a particular thing and how  many like what do you call the connected [32:45] um homes are actually working right in a given  area like what is the impact of the network why [32:51] is it being cost. All sorts of informations  are being fed to us and we are using that [32:56] information to actually determine whether we need  to even dispatch a truck roll to get it resolved. [33:01] Right? And the efficient you can you can think  about this right like every company um if they [33:06] have to just schedule a truck roll for just every  issue that the customer is calling about that's [33:10] going to incur a lot of cost there. But rather  than that, how can we be more smart about like [33:16] where we deploy our truck rolls where we don't and  how do we kind of like reroute customers to like [33:22] take selfexloratory steps to resolve their network  issues. That's something that most companies have [33:27] already invested in and you can see that like even  if you go to your apps today who whoever your ISP [33:32] provider might be, they automatically have like  u reset network stuff or like you know take some [33:38] simple actions to resolve your network issues.  That is one of the things that I'm seeing growing [33:43] on and on. And the other thing is like also you  can use uh visual AI. So for example like you can [33:48] take an image of your modem or like whatever that  is and then like you know then let the AI explain [33:54] to you the steps you have to take to resolve your  network issue. That is something that I'm seeing [33:59] a lot of movement that is happening. Right. And on  your second question around the AJ AI, I I mean I [34:05] believe that's where the industry is also going  towards right like you can see one bit models, [34:10] two bit models and four bit models that are coming  up and especially like the every provider that you [34:16] know of has an open source alternative that is a  tiny LLM that you can that can sit on your device [34:21] and can make that that that can run inference on  that edge system helping you make those decisions [34:27] like your co-pilot or things like that right and  I think um it is going to get smarter smart and [34:32] smarter, right? And everyone is working on  distillation process making it better. And [34:37] so once that actually knows how to best compress  your models, put that models into your devices. [34:43] And I think this is where I want to uh talk  about Javan's paradox, right? So if you if you [34:49] heard about it, once you start utilizing certain  things, right? The increase in the efficiency of [34:58] those resources will start to accelerate. meaning  like you will start to see people u coming up with [35:05] a lot of innovations to make that system more  efficient especially in AGI or any other systems [35:10] of that sort but then what ultimately it will lead  to is the more increased use of that system itself [35:15] so it is kind of like a paradox you get cyclical  and you become more efficient and that's what is [35:19] happening today in EI overall so hopefully that  helps I'll just add one quick thing to a great [35:25] answer but um another thing that's happening is  uh the agentic architecture is also uh helping [35:32] the edge computing use case a lot is uh in in that  architecture you can create a purpose-built agents [35:39] that do small things but they're lightweight. So  as a result you can have so many of those agents [35:45] with given the amount of compute space that you  have but they can do certain certain actions in a [35:51] much uh compacted format. So we're running a bit  tight on time now. I'm going to skip a section [35:59] that we talked about earlier because there's one  thing that I do want to make sure we discuss with [36:04] the time that we have left. So, you know, when I  arrived in Vegas yesterday, I got a taxi ride to [36:10] the Venetian and I mentioned to my driver that  I was here for an AI conference and immediately [36:17] the questions start coming up, right? Like what  do the future of jobs look like? We we know all [36:22] of these questions, right? So let's talk about the  the human component of AI within telecom. Um and [36:29] obviously you mentioned earlier redeploying people  in in new ways for potentially new solutions and [36:35] we're more specifically they're talking about  engineers. What should those working in telecom [36:41] really be thinking about in terms of the future  of the humans involved in their companies? Yeah. [36:48] And I was wondering whether we had the same taxi  driver or whether all taxi drivers in Las Vegas [36:53] are thinking the same way about AI. Um but uh  but yeah from a human perspective I think as [37:01] as I had mentioned earlier about um freeing  up and upskilling engineers so that they can [37:07] start building more value added services that uh  like for emergency services getting more enhanced [37:16] communications capabilities that might save  more lives. Now there are other uh areas that [37:25] uh could use more help and and work um for example  I think you know we talked a little bit about [37:34] uh IVR not not us but we talked before this it's  the pre- panel panel the pre- panel part uh talked [37:42] a little bit about um IVR is one area that could  use um some revolution some some enhancements or [37:51] or a change because I I think we're all we  all get annoyed when we have to press one [37:58] for this and then press four for that and then  press five and how deep do you have to go? Um, [38:07] and so in if we could instead have an AI enabled  IVR that's smarter that you don't have to do all [38:16] that pressing. Um, and of course the chat bots  that exist today for customer service isn't [38:26] perfect yet, right? So we need people to work on  that to improve that customer service interaction. [38:35] um for any kind of business that that would be  very helpful for businesses. It's great for the [38:40] people that are using it as well. Uh so so from a  human perspective I think from a customer service [38:49] perspective that would help and also from from  the other side a skills perspective and a job [38:56] perspective there is work to be done once we move  over to AI. You just need to work on creating [39:03] those new AI services and then perfecting those  AI services so that they work well. And I guess [39:10] the other element there with those new AI services  as well as for anything that is customerf facing. [39:15] It's also building that general trust across the  population, right? You know, any new adoption [39:21] of technology has its sticking points. Um be  interesting to see how telecom deals with those. [39:29] So, we're coming to the end here now. I did  promise you all a lightning round and uh please [39:37] don't hate me if you hate the questions and please  don't hate me if you hate the questions as well. [39:41] Um so, an what's one telecom capability that we  might take for granted in the next five years? [39:50] So maybe I could go out on a limb and um  since you're asking this question I think [40:00] from a standpoint of what the public would  probably see versus something that you have [40:09] no idea that it's AI in the system then um  maybe in the future in the near future you'll [40:18] be you will be dealing with an AI um for  customer service that you don't think, oh, [40:26] I wish I could talk to a human [laughter] or  that I wish I didn't have to go down several [40:35] layers of press one, press three, press six um  through that uh IVR interface. So, in the future, [40:44] you'll be able to actually talk to an AI and  it actually understands you nearly most of the [40:51] time and it will be like it's normal. I think I  I'm looking forward to the stage of life where [40:58] I no longer have to listen to elevator music  for 20 minutes while I talk to her majesty's [41:03] revenue and custom. [laughter] Um, Sunda, one AI  buzzword that you would happily retire. Happily [41:13] retire. Okay. Um, I mean I'm going based on the  recent developments that I've heard from Claude, [41:19] right? So I believe what I think we are going to  see is prompt engineering is going to go away. [41:24] That's probably something that is going to retire  very soon because um I was hearing to the head [41:30] of AI who built the clot code talk uh recently  and he mentioned that they took away 80% of the [41:36] problems that they put in the cloud code system.  So that makes me believe or makes me think like [41:41] you know prompt engineering as a core thing that  we are doing today in our agent solutions. Yeah, [41:46] I think that'll be gone. Hopefully we won't  have to tweak prompts anymore and we will let [41:51] the system decide because we are going to  build autonomous agents all over. Nice and [41:57] sill a different question for you. What's one  technology everyone is underestimating today? [42:05] Well, let me answer this from a not from  a technology standpoint, but more from a [42:12] um u from a field standpoint that people are  underestimating and it's been all over the news [42:18] as well is AI governance. Um there have been some  recent examples for example a couple of days ago [42:24] news about claude escaping the lab u attacking  three different companies. Jad GBD did that. So [42:31] it's it's it's becoming more and more dangerous.  Um Elon Musk famously tweeted tweeted well all of [42:39] his tweets are famous but this one being about  AI governance. If AI models are not controlled [42:46] they can be they can be more dangerous  than than nuclear missiles themselves. So [42:52] um I strongly believe that there will be a  lot more attention to AI governance over time. [42:58] uh there's already research that's going on. Uh  for example, XAI is a branch, explanability AI [43:05] is a branch which is already focusing on putting  guardrails around AI and so on so forth. Um so [43:11] um yeah, that's the uh that's the next thing. I'd  just like to point out I did say to the panel that [43:17] I was going to give them a lightning round so that  we could end with high energy. We've ended with AI [43:21] will be more dangerous than nuclear weapons.  [laughter] Thank you for that happy note. Um, [43:27] but on that note, we've reached the end of  the panel. If you would like to ask any of [43:32] the lovely people up here a question. Don't come  to me because I'm not the expert, but I'm going to [43:36] suggest that you exit this room and we don't block  the stage for whoever is on next. So, speakers may [43:42] be out here. Um, guys, thank you so much for a  very interesting panel. I have learned a lot. Um, [43:50] and thank you for joining us. A big round of  applause for the panel, please. [applause]