Kaspar Companies and Sumitomo Rubber Cut Downtime via AI Repair Tools

Kaspar Companies与住友橡胶利用AI维修工具减少停机时间 工业AI只有与精益制造原则直接挂钩，并在工厂车间以一周为周期快速推行落地，才能产生可衡量的投资回报。 来自Kaspar Companies、Viaduct和Olsen Capital的行业领袖分享了生产实践经验。 住友橡胶（Sumitomo Rubber）在宫崎工厂引入 Viaduct 开发的 AI 维修助手，结合十年历史维护日志与实时遥测数据推荐检修方案，月均处理 400 多起维修，将平均停机时间缩短约 40 分钟。 (2:15) Kaspar Companies 将精益制造（Lean Manufacturing）作为 AI 战略基石，通过双目视觉系统 Mammoth Vision 自动核对卡车底盘五金包，并规定所有 AI 项目必须通过改善（Kaizen）活动验证业务价值，避免脱离实际场景的技术堆砌。 (5:25) 针对 AI 冲击就业的顾虑，专家指出一线工人正逐步转型为督导多个自主智能体的管理者；Kaspar Companies 亦确立了严格的提效不裁员政策，受影响员工全员转入改善推广办公室并晋升至更高薪岗位。 (9:20) 为打破一线人员对预测性维护的抵触，企业需采用可解释性 AI（XAI）展示底层传感器指标与历史故障判例以建立信任；同时若全员生产维护（TPM）等传统精益流程已能解决根因，则无需强行堆砌复杂预测算法。 (13:04)

David Hallac of Viaduct and Peter Robinson of Kaspar Companies argue that industrial AI delivers sustainable returns only when anchored in lean manufacturing principles, rapid shop-floor experimentation, and operator trust rather than custom algorithmic complexity. Grounding AI in Lean Problem Solving (5:25) Technological deployment failures in industrial environments typically stem from misdiagnosing operational gaps rather than software shortcomings. When Kaspar Companies launched KAI Labs in November 2025, it pursued an aggressive target of deploying one new AI solution weekly. Although the unit delivered over 36 implementations, early projects frequently targeted unvetted problems without measurable business value, yielding an initial success rate where only 40% to 50% produced direct value and roughly 30% delivered none. To correct this, Kaspar mandated that every project originate through a structured Lean Kaizen workshop to define the root issue before writing code. Robinson notes that existing frameworks like Total Productive Maintenance (TPM)—which transfers basic maintenance directly to line operators—had already stabilized machine uptime across their plants. Consequently, management avoided forcing complex predictive AI models into workflows where low-tech process discipline had already eliminated the bottleneck. High-Impact Deployments and Model Explainability (2:15) Targeted factory deployments demonstrate significant operational gains when designed around specific bottlenecks. At Sumitomo Rubber's Miyazaki tire plant, Viaduct deployed an AI repair assistant trained on a decade of maintenance records and Japanese shorthand logs. By analyzing real-time telemetry to recommend the top three repair actions, the tool handles over 400 monthly repairs and cuts average downtime by approximately 40 minutes per incident, while reinforcement learning loops improved recommendation accuracy by nearly 50% over initial production runs. Similarly, Kaspar deployed 'Mammoth Vision,' an inline dual-camera system that scans serial numbers across 180 truck bed SKUs and verifies accompanying hardware packs against bill-of-materials specifications, flashing red alarms to prevent misconfigured shipments. Adopting predictive maintenance requires overcoming operator skepticism toward taking running machinery offline. Hallac emphasizes that black-box alerts fail on the plant floor; models must provide Explainable AI (XAI) diagnostics. Presenting the exact telemetry anomalies alongside empirical precedents—such as showing 18 identical historical sensor patterns that ended in total failure—provides the transparent rationale operators require to trust algorithmic recommendations. Workforce Safeguards and Human-in-the-Loop Reskilling (9:20) Sustained shop-floor adoption requires mitigating worker displacement anxiety through concrete institutional protections. Kaspar Companies enforces a strict corporate policy guaranteeing that no employee will be laid off due to efficiency gains or AI automation. When automation replaces manual roles, impacted personnel transition to the internal Kaizen Promotion Office (KPO), where they train in continuous improvement methodologies and lead future optimization projects; Robinson reports that 100% of transitioned workers have advanced into higher-paying, higher-responsibility positions. Hallac contends that AI will not replace human workers, but rather elevate them into managerial oversight roles where a single operator directs multiple autonomous agents. Because algorithmic agents lack real-world factory context, human domain expertise remains essential to validate diagnoses and steer continuous process refinement. Agile Sprints and Uncovering Secondary ROI (17:44) Hallac estimates that 98% of manufacturing firms do not require custom model development and should instead expect tangible ROI within 90 days using existing tools. To maintain velocity and avoid the sunk-cost fallacy, Kaspar utilizes small 'two-pizza' teams of four developers working in dedicated physical war rooms (Obeya rooms) with strict one-week sprint boundaries. If an initiative fails to show traction within a week, it is paused or discarded, preventing resource drain while technical teams co-locate with shop-floor workers to iterate solutions directly. Both panelists observe that practical deployments often generate unexpected secondary financial gains. Kaspar's initial algorithmic truck-loading model was designed to prevent stranded yard inventory caused by flat-rate freight caps; unexpectedly, it revealed available capacity in certain configurations, prompting sales reps to upsell additional beds over the phone. Similarly, Hallac noted that predictive analytics deployed strictly to hit automotive production volume targets indirectly elevated the manufacturer's third-party J.D. Power quality ratings after two years of plant-wide stability. Looking ahead, the panelists anticipate inference and infrastructure costs falling by 10x to 100x within a year alongside advancing humanoid robotics. As general foundational models consolidate among major tech providers, the industrial competitive advantage will belong to nimble domain teams that apply simple, off-the-shelf tools to real factory pain points, rather than organizations building redundant internal software architectures.

Peter Robinson与David Hallac指出，现代制造业应用AI的核心逻辑在于将技术严格绑定精益制造（Lean）原则并聚焦切实的生产瓶颈，而非盲目自研复杂的黑盒模型。企业通过高频敏捷迭代、现场数据验证以及兼顾可解释性与员工保障，能够以最低的试错成本稳定提升生产力并获得超预期的业务增益。 精益驱动与痛点筛选：拒绝盲目自研模型 (5:25) Kaspar Companies与Viaduct的实践表明，工业AI落地受阻通常源于对核心业务问题的误解，而非底层技术本身的局限。David Hallac提出，企业决策层应明确区分“止痛药”（亟待解决的急性瓶颈）与“维生素”（锦上添花的功能），并强调98%的普通制造企业完全无需组建高薪团队定制算法，而应优先在90天内获得可量化投资回报。针对低效或能力不足的项目，企业宁可暂停6个月等待新一代基础模型演进，也不应在旧架构上过度定制开发。 Peter Robinson回顾指出，Kaspar旗下KAI Labs在追求每周部署一个AI方案的早期阶段，曾出现向未经验证的问题盲目投放算法的偏向。为纠正这一偏差，公司将AI战略彻底锚定在精益制造文化上，强制要求所有项目必须由改善（Kaizen）活动发起，由跨职能团队在写代码前先行明晰业务差距与消除流程浪费。在全员生产维护（TPM）等传统精益机制已通过基层日常检修有效稳定设备稼动率的领域，Kaspar坚决避免强行套用复杂的预测性AI，坚持让技术严格服务于未被解决的核心瓶颈。 产线落地实践：视觉防错与可解释性维修 (2:15) 在生产现场，针对具体工序痛点定制的实用型工具已产出实质效益。在住友橡胶（Sumitomo Rubber）位于日本宫崎的轮胎工厂，设备故障导致的停机曾长期拉低产能。Viaduct利用该厂十年的维修日志、技术文档及老员工隐性经验构建知识图谱，开发出AI维修助手。该系统结合实时遥测与强化学习算法，在故障时为技术人员推荐前三项维修方案，使平均单次停机时间缩短约40分钟，单厂每月支持超400次维修，数月内推荐准确率提升近50%。针对操作员对预测告警的抵触心理，David Hallac强调系统必须提供可解释性证据，具体展示异常传感器数据并列举历史上引发停机的18起相同先例，以经验证据打消工人的怀疑。 在Kaspar旗下生产180种卡车车斗SKU的Bedrock工厂，过去常因配件螺栓包错发而导致分销商安装中断。KAI Labs研发了名为Mammoth Vision的双摄像头内联视觉质检系统：第一台摄像头扫描车斗序列号以调取物料清单规格，第二台摄像头比对螺栓包实物。一旦检测到错漏配，系统立即触发红灯警报并下发更换指令，用轻量级的视觉闭环彻底杜绝了发货缺陷。 人机协同机制：员工赋能与零裁员承诺 (9:20) 针对自动化带来的裁员恐慌，两位专家均证实AI在工业场景的定位是劳动力杠杆而非替代性工具。David Hallac引用行业共识指出，“AI不会取代人类，但使用AI的人将取代不使用AI的人”。AI助理使操作员的角色从纯体力执行者跃升为能够同时指挥调度10个自主智能体的功能管理者；缺乏人类的现场领域专长与真实语境判断，智能体本身无法独立解决现实问题。 Peter Robinson公布了Kaspar实行的制度性保障：公司承诺绝不因流程改善或AI实施而解雇任何员工。当自动化项目消除某条产线上的手工岗位时，受影响员工会被转入改善推进办公室（Kaizen Promotion Office, KPO），带薪接受持续改进方法论培训并负责主导后续的精益优化。Robinson证实，所有经此机制转岗的员工均100%晋升至更高薪资和更高权责的管理或专业岗位，从组织文化上彻底消除了基层对新技术落地的戒备与阻力。 敏捷小团队迭代与现场催生的溢出价值 (17:44) 在组织推进模式上，Kaspar借鉴“两张披萨团队”理念，仅设立由4名开发人员构成的小型敏捷团队，通过大屋（Obeya）作战室目视化面板追踪每日冲刺瓶颈。团队设立每周部署一项新工具的高压学习目标，并严守“单周冲刺限制”：实验性项目若在一周内未能展现价值便立即止损封存，通过高频迭代快速积累经验。数据显示，部署的37至38个项目中约40%至50%产生了直接业务价值，约30%未达预期，但高频试错为企业建立了成熟的软件交付规范。 两位专家强调，工程师必须扎根车间一线与工人共同测试系统（“吃狗粮”），这一过程常能发掘预案之外的次生商业价值。例如，Kaspar为解决按单定制车斗因固定运费超量积压在货场的问题，开发了装车图形优化系统；客服人员在接听电话时实时发现部分卡车仍有装载冗余，顺势向客户提出加购建议并获得广泛同意，意外将库存控制工具转化为了直接拉动顶层营收的增销引擎。类似地，David Hallac团队在北美汽车制造厂针对排产目标部署预测性分析两年后，发现其下游制造稳定性的提升间接推动了第三方J.D. Power质量排名的显著上升。 总体而言，工业AI的核心竞争力正在从通用大模型算法转向深入具体工业工作流的务实应用。尽管未来一年内开源模型和算力成本有望降低10至100倍、人形机器人也将逐步进入物料搬运领域，但企业仍面临工具泛滥与一次性微应用沦为“技术新奇玩具”的风险。如何在模型加速商品化的浪潮中维持严格的问题导向和精益流程约束，仍是传统制造企业数字化转型中未完结的治理考验。

## Background 该视频由 Ai4 频道于 2026 年 8 月 22 日发布在 YouTube 平台，主题为“更智能、更快速、更精益：现代制造业中的 AI”。根据页面简介，来自 Kaspar Companies、Viaduct 与 Olsen Capital 的行业专家在此次研讨会上，共同探讨了制造企业如何利用预测性分析、计算机视觉与机器人技术提升生产力与供应链韧性，并分享了围绕 ROI、安全性和设备正常运行时间展开的工厂一线实战经验。 ## TL;DR 本场研讨会深入探讨了 AI 在现代制造业中的真实落地路径与精益融合策略。嘉宾们结合智能维修助手、计算机视觉质检及三维装载优化等实际案例，强调企业必须坚持“问题导向”而非拿着技术找场景，优先打造直击痛点的“止痛药”方案。讨论重点剖析了如何通过 Kaizen（改善）机制与“绝不因改善裁员”的承诺化解一线员工抵触，并通过深入车间现场挖掘出拉动销售与提升质量评级的隐性收益。最后，嘉宾一致认为未来通用横向平台将加速整合，深耕特定场景的垂直方案将迎来大爆发，具身智能与大幅降本的开源模型将成为下一阶段的核心驱动力。 ## Atmosphere & style 对谈整体氛围务实接地气且充满一线实战经验，彻底摒弃了空洞的宏大叙事与 AI 炒作，洋溢着硬核制造业的严谨与活力。嘉宾与主持人之间的互动幽默而坦率，不仅毫无保留地分享了早期盲目试错的项目淘汰率，还穿插着调侃管理层预算和代码编写现状的轻松笑料。整场讨论节奏紧凑、层次清晰，既有对车间工人信任机制与组织文化的深刻剖析，又展现出敏捷迭代、拥抱具身智能的激进探索欲，带给听众极强的现场代入感与启发性。 ## Key takeaways - [02:15](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=135s) 住友橡胶宫崎工厂部署的 AI 维修助手基于十年历史维修数据推荐最佳方案，单次停机时间缩短约 40 分钟，每月支持超 400 次维修并推广至全球工厂。 - [02:15](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=135s) Kaspar 旗下 Bedrock 车厢业务部署 Mammoth 视觉系统，实现 180 种车厢 SKU 与对应配件包 100% 精准匹配，彻底杜绝发货漏装与错装。 - [05:41](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=341s) 从盲目推 AI 转向回归精益制造（Lean）与 Kaizen（改善），通过精准定义痛点与差距后再决定技术选型，显著提升了项目效益。 - [09:25](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=565s) AI 不会取代人类，但使用 AI 的人将取代不使用 AI 的人；员工正从独立贡献者（IC）跃升为能同时调动多个 AI Agent 的管理者。 - [10:46](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=646s) 企业设立改善推进办公室（KPO）并确立“绝不因 AI/流程优化裁员”的铁律，让富余员工学习精益方法论并获得晋升，有效消除了全员抵触心理。 - [13:04](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=784s) 工业维保场景拒绝黑盒模式，模型必须具备可解释性（如列举 7 项前置征兆与 18 个历史相似案例），现场工程师才能建立信任并果断行动。 - [16:11](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=971s) 卡车车厢 3D 装载优化系统不仅解决了成品库存积压的初衷，还因实时挖掘货车剩余载重空间实现了追加增销，成为拉动营收的意外引擎。 - [23:32](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=1412s) 传统制造企业构建 AI 能力的前 90 天应坚持“问题优先”，拒绝制定长达两年的宏大计划，并在透明作战室（Obeya）中高频暴露问题与迭代。 - [26:13](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=1573s) 团队应“站在错误的边缘”快速试错，设定每周交付解决方案的激进目标，严格执行单周冲刺与快速止损机制，以高频失败换取认知沉淀。 - [31:40](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=1900s) 立项必须挑选直击业务痛点的“止痛药”项目而非可有可无的“维他命”，并遵循准时制（JIT）交付，避免技术提前上线形成数字化库存浪费。 - [37:33](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=2253s) 通过 4 人组成的“双披萨”敏捷小团队保持高效交付，利用内部统一平台、演示短视频以及一线员工在年度峰会现身说法来常态化传播成效。 - [41:19](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=2479s) 横向通用平台将向少数模型巨头高度集中，而深度融合垂直业务场景、能端到端解决完整痛点的专用方案将迎来大爆发。 - [43:14](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=2594s) 业务落地遵循最简原则，优先选择非 AI 或现成模型，仅靠 AI 快速套壳的同质化工具生命周期极短，深耕行业流程才是真正的护城河。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=4s) 嘉宾介绍与制造业业务背景 研讨会开场，主持人 Joe Olsen 说明原本应到场的嘉宾 Royal Sequiera 缺席，本次讨论由台上三位嘉宾展开。三位嘉宾依次进行了自我介绍并阐述了各自的业务范畴： Peter Robinson 担任 Kaspar Companies 的首席运营官（COO）及创新部门 KAI Labs 的执行负责人。Kaspar Companies 位于得克萨斯州夏纳（Shiner, Texas），是一家拥有 128 年历史的第五代家族企业，旗下拥有四大业务平台： - **卡车车厢制造（Truck Beds）**：主要品牌为 Bedrock Truck Beds。 - **马鞍制造（Horse Saddles）**。 - **户外用品（Outdoors）**：主营栓动步枪等装备。 - **贵金属业务（Precious Metals）**。 David Hallac 是 Viaduct 的 CEO，本科/博士毕业于斯坦福大学，主攻大规模时间序列 AI。Viaduct 位于加州帕洛阿尔托（Palo Alto），专注于利用制造与维护数据，通过 AI 在设备故障发生前进行预测与修复。该公司于去年被日本住友橡胶（Sumitomo Rubber）收购；目前他正致力于将住友内部验证成熟的 AI 部署方案推向外部非住友企业。 Joe Olsen（主持人）来自 Olsen Capital 家族办公室，同时是一名 AI 与基础设施领域的顾问和工程师，热衷于在周末亲自编写代码构建试点项目。 --- ### [02:15](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=135s) 制造业 AI 落地实战：智能维修助手与视觉质检 在探讨具体的 AI 应用实例时，David Hallac 与 Peter Robinson 分别分享了已在工厂实际部署并取得可衡量成效的案例。 David 介绍了在住友橡胶位于日本南部宫崎（Miyazaki）大型轮胎工厂落地的 **AI 维修助手（AI Repair Assistant）**。该工厂此前因设备意外停机频繁导致无法完成生产配额。维修人员排查和解决故障常常耗时数小时，造成整条轮胎产线大面积积压。Viaduct 基于该工厂过去 10 年的历史维修记录，构建了能实时推荐“前三项最佳维修方案”的智能助手。该系统将单次维修停机时间平均缩短了约 40 分钟，目前在该工厂每月支持超过 400 次维修任务，并正在向住友橡胶全球所有工厂推广。 Peter 分享了 Kaspar Companies 旗下 Bedrock 卡车车厢业务中部署的 **Mammoth Vision** 计算机视觉系统。卡车车厢在出厂发货时，必须随车附带专用的螺栓配件包（bolt packs）。由于卡车车厢型号多达 180 种 SKU，若发货时漏装配件包或型号配错，下游改装厂（job shops）拆卸旧车厢后便无法安装新车厢，引发客户与分销商的大量投诉。 ```mermaid flowchart TD accTitle: Mammoth Vision 工作流程 subgraph S1["产线移动与双摄识别"] A["卡车车厢沿产线移动"] --> B["相机 1：扫描车厢序列号 / SKU"] A --> C["相机 2：俯视检测配件包类型"] end subgraph S2["系统判断与反馈响应"] B & C --> D{"配件包与 180 种 SKU 是否匹配？"} D -- "完全匹配" --> E["正常放行至下一工序"] D -- "配件包型号错误" --> F["提示操作员应更换的具体型号"] D -- "配件包缺失" --> G["触发红色警报灯闪烁并报警"] end ``` - **停机时间减少**：住友宫崎工厂单次维修停机时间减少约 40 分钟。 - **高频应用**：单厂每月处理 400+ 次维修。 - **防错设计**：Mammoth Vision 实现 180 种车厢 SKU 与对应配件包 100% 准确匹配。 --- ### [05:41](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=341s) 探索期的阵痛与破局：精益思想（Lean）与本体数据构建 主持人提到，智能工厂的理想终态备受关注，但早期探索往往极其混乱，询问两位嘉宾遇到过的核心瓶颈与应对策略。 Peter 回顾了 KAI Labs 团队于去年 11 月成立时的曲折历程。为了快速试错，团队设定了“每周上线一个 AI 解决方案”的激进 10 倍速（10x）目标，截至目前已部署了 36 个解决方案。但初期遇到了严重的混乱——团队在没有搞清楚业务痛点的情况下盲目乱扔 AI 方案（"Throwing out AI solutions"）。为此，Kaspar 决定回归自身的“精益制造（Lean）”根基，通过开展 **Kaizen（改善活动）** 召集相关一线人员，精准定义问题和差距（gap）。只有在明确问题本质后，才决定是否采用 AI，这一转变彻底盘活了 AI 项目的实际效益。 David 则从数据角度剖析了挑战的演变。15 年前他在斯坦福攻读时间序列 AI 博士时，核心痛点是“清洗数据（clean data）”；而在今天，智能体（Agentic systems）已能够高效组织数据。以住友的维修指导系统为例，知识主要分散在两个极其难以利用的地方： 1. **老旧系统中的日本技术员速记简写（Japanese shorthand comments）**，外人几乎无法看懂； 2. **资深技师的大脑经验**。 Viaduct 重点打造了能够从每次维修中自主吸收反馈的强化学习（RL）系统，构建起严谨的数据本体（Ontology/Platform）与层级结构。系统部署初期随着数据流通达到临界体量（Critical Mass），预测准确率在数月内提升了近 50%。 --- ### [09:25](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=565s) AI 与人类工人的协作关系：“从独立贡献者到管理者” 针对“AI 是否会取代工人”的行业顾虑，David Hallac 引用了一句经典行业观点： > "AI won't replace humans, but people using AI will replace people not using AI."（AI 不会取代人类，但使用 AI 的人将取代不使用 AI 的人。） David 指出，AI 目前如同“世界上最聪明的高校实习生（the world's smartest intern）”，但在业务机制、机构专有知识（institutional knowledge）、法律责任及人道伦理方面，依然离不开人类设置安全护栏（guardrails）。 在实际制造管理中，积极拥抱 AI 会带来显著的生产力倍增效应。AI 正在促使一线员工发生角色跃迁——从单打独斗的“独立贡献者（Individual Contributor）”转变为“管理者（Manager）”。员工不再受限于自己双手的打字和操作速度，而是能够像同时指挥 10 个 AI Agent 一样高效输出，极大地激发了创意的实现速度与产线效率。 ### [10:46](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=646s) AI 落地不裁员：持续改进与改善推进办公室（KPO）机制 关于人脑知识与 AI 结合的讨论仍在继续，Peter Robinson 结合 Kaspar Companies 的实践分享了企业在面对 AI 舆论恐慌时的应对之策。在 AI 概念大火、媒体每天都在鼓吹“AI 将取代人类工作”的背景下，公司决定退一步思考：如何以员工能够欣然接纳的方式引入 AI？答案最终回归到企业自身的运营管理体系。 ```mermaid flowchart TD accTitle: Continuous Improvement and Kaizen Flow subgraph S1["发现问题与推进改善"] A["业务痛点与问题"] --> B["开展 Kaizen 改善活动<br/>(全员平等参与，无职级界限)"] end subgraph S2["人员保障与转岗机制"] B --> C{"岗位是否因优化而缩减？"} C -- "是 (如缩减 3 个岗位)" --> D["进入改善推进办公室 (KPO)"] D --> E["学习持续改进方法论"] E --> F["参与下一轮改善 / 获得晋升与更好岗位"] C -- "否" --> G["原岗位效率提升，产能扩展"] end subgraph S3["AI 项目准入原则"] B --> H["AI 赋能项目接入 (KAI Labs)"] H --> I["严禁因改善/AI裁员"] end ``` Kaspar Companies 拥有一套指导所有业务板块管理者的经营体系手册，其核心原则就是“持续改进”（Continuous Improvement）——无论有没有 AI，都要日复一日地追求卓越、创造更多价值，实现“零额外成本的业务增长”（grow for free）。这种改进释放出更多产能并带来全新机遇。最关键的是，公司立下一条铁律：“绝对不会因为流程改善而辞退任何一名员工。” 公司专门设立了改善推进办公室（Kaizen Promotion Office，简称 KPO）。如果某项流程改善后导致某个区域减少了 3 个岗位需求，这 3 名员工会被调入 KPO。在那里，他们系统学习持续改进的方法论，并直接参与下一轮改善项目；实践证明，在每一种情况下，这些员工最终都转岗到了更好的职位或获得了晋升机会。在 KAI Labs 推进 AI 时，也严格遵循这套逻辑：在启动任何 AI 项目之前，团队必须先做一次 Kaizen 改善，而 Kaizen 的铁律就是绝不裁员。员工切实体会到这一点后，对 AI 的抵触情绪便烟消云散。 - **持续改进的核心逻辑**：不依赖 AI 也能每天变得更好，通过优化释放产能并创造新增长，且绝不把优化成果变成裁员手段。 - **KPO（改善推进办公室）的运作机制**：冗余岗位的员工转入 KPO 进修精益改进方法论，随后投入新的改进项目中，全员最终均实现职级晋升或转入更具前景的岗位。 - **AI 项目准入红线**：在 KAI Labs，所有 AI 项目必须依托 Kaizen 流程展开，必须严格恪守“改善绝不裁员”的承诺。 --- ### [13:04](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=784s) 拒绝“拿着锤子找钉子”：从痛点出发与模型可解释性 讨论转向制造业从“事后被动维修”（Reactive Maintenance）向“预测性维护”（Predictive Maintenance）转变所面临的文化阻力。主持人 Joe Olsen 提问：要让现场工人放弃多年的直觉经验、转而信任 AI 和数据，最困难的地方在哪里？ Peter Robinson 直言，成功的秘诀在于“千万不要从 AI 开始”。如果手里拿着 AI 技术到处去找应用场景，最终只会误入歧途，不仅无法创造实际价值，还会招致一线工人的强烈抵触。几年前，Kaspar 旗下工厂的机器频繁发生故障停机，每天都在忙着给设备供应商打电话求援。他们并没有盲目上马 AI，而是推行了“全员生产维修”（Total Productive Maintenance, TPM），把设备维护的责任和主导权交到一线机台操作工手中。仅仅依靠这一项管理变革，工厂就彻底解决了停机问题，至今机器运行良好。既然停机已经不再是主要痛点，公司也就没有必要为了用 AI 而去强行引入 AI。 David Hallac 对此深表赞同，并补充道：“不能因为手里有锤子，就把一切都看作钉子。”不过，对于确实需要技术介入的复杂设备维护场景，建立信任的关键在于“模型的可解释性”（Explainability）。市面上的通用大模型（如 ChatGPT 或 Claude）大多只是像黑盒一样直接给出回答，无法清晰呈现推导逻辑。但在工业维保场景中，如果发动机表面看起来完好无损、没有冒烟，操作工很难仅凭黑盒模型的报警就去拆解设备。Viaduct 的做法是提供详尽的归因证据链： > “AI 会告诉工程师：我认为这个排气歧管很可能会发生故障，你应该立刻处理，因为当前已经出现了这 7 项前置征兆，并且在近期历史记录中，还有 18 台其他设备/车辆在经历完全相同的征兆序列后发生了排气歧管故障。” 这种基于历史相似案例和明确归因的解释，能够彻底打破“黑盒”带来的不信任感。 - **问题导向而非技术导向**：不要为了技术而技术；Kaspar Companies 依靠 TPM（全员生产维修）让操作工当家作主，未写一行 AI 代码就彻底消除了困扰多年的设备停机顽疾。 - **可解释性是工业 AI 的信任基石**：黑盒预测无法说服现场工程师拆修外观完好的设备；必须依靠多重前置条件归因与历史相似案例样本（如“7 项特征 + 18 个历史案例”）来辅助决策。 --- ### [16:11](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=971s) 扁平化共创与意外之喜：卡车装载优化带来的营收增长 在探讨如何向内部推广 AI 方案时，Peter Robinson 强调了推行方式的演变。早期的几个项目往往是自上而下地向现场“推销方案”，效果并不理想；后来团队全面转向“以问题为核心”，并在 Kaizen 活动中让一线操作工全程参与。在这些改善研讨会上，所有人抛开职务头衔（No titles, everyone has a voice），平起平坐地共同梳理痛点。由于一线人员参与了方案的设计与推导，AI 落地时的推行阻力自然迎刃而解。 针对主持人关于“隐藏财务收益”（如节能、降低质保成本等未写在初期商业计划书里的回报）的提问，Peter 分享了他们第一个 AI 项目歪打正着的成功案例： 当时，卡车车厢制造业务（truck bed system）面临严重的成品库存积压问题，堆场甚至不得不扩建。该业务原本是按订单生产（Make-to-Order），按理不该有大量成品滞留。经过深入排查，团队发现了症结所在：公司的直接客户（中间商）每次订货都是按整车皮（full truckload）采购，因为无论装多装少运费都是固定的。由于车厢尺寸繁多，订货时谁也拿不准一辆平板货车究竟能塞下多少个车厢，客户通常会盲目下 10 到 12 个车厢的订单。工厂把货造好后，全公司只有一名经验丰富的老师傅懂得如何像玩俄罗斯方块一样“投机取巧”（jimmy-rig）把车厢尽量塞进货车。如果最后发现只能装下 10 个，客服打电话询问客户是否再叫一辆车运走剩下的 2 个时，客户往往会说：“算了，那 2 个不要了，就发这 10 个吧。”结果，多做出来的 2 个车厢只能堆在堆场里变成积压库存。 ```mermaid flowchart LR accTitle: Truck Bed AI Order Optimization Flow subgraph S1["传统下订单与排产"] A["中间商盲目下单<br/>(例如预估 12 个车厢)"] --> B["车间按单全数制造"] B --> C["装车老师傅现场尝试装载"] C --> D{"货车空间超限<br/>(实际只能塞进 10 个)"} D --> E["剩余 2 个车厢遭客户放弃<br/>堆场积压大量死库存"] end subgraph S2["AI 介入后的图形化实时优化"] F["客户来电选购车厢"] --> G["AI 系统实时图形化<br/>三维装载优化计算"] G --> H{"算力反馈装载空间"} H -- "超出容量 (如选了12个只能装11个)" --> I["客服当场核减 1 个<br/>从源头消除废弃库存"] H -- "尚有富余 (如选了12个还能塞2个)" --> J["客服建议追加 2 个<br/>促成增销，直接拉动营收"] end ``` 为了解决库存堆积问题，团队开发了一套 AI 图形化装车实时优化系统。当客户打电话下单选品时，系统会实时根据三维尺寸动态计算货车空间： 1. 如果客户订了 12 个但系统显示只能装 11 个，客服在通话时就能当场把多余的 1 个去掉，避免了生产后的库存浪费。 2. 更有意思的是意外收获（Hidden Financial Win）：很多时候客户订了 12 个，系统计算后发现“其实还能再塞 2 个”。客服顺势建议客户追加订单，客户往往欣然同意。 这个原本旨在解决库存积压的 AI 优化工具，最终直接转化为了强大的销售增长引擎。 现场随即爆发出一段轻松幽默的互动。主持人打趣问道：“你们有没有把这笔额外的销售提成截留一部分放到 AI 实验室的预算里？”Peter 笑着打趣回应：“没门，如果你们认识我们总裁 Greg 的话就知道，他管钱抠得很，一分钱都不会多给的！”引发全场会心大笑。David 接过话茬总结道，大众往往只盯着最直观的顶线数据（Topline）和可量化指标，但在制造业 AI 实践中，软性收益与业务协同所带来的潜在价值往往同样重大，甚至更加关键。 - **消除沟通盲区**：通过在下单前端引入 3D 几何装载优化算法，使原本不可知的装车容量在订单录入环节实现透明化。 - **库存优化反哺销售增量**：AI 算法不仅解决了堆场成品车厢积压的痛点，更通过挖掘卡车剩余载重空间实现了附加销售（Upsell），成为撬动顶线营收的意外杠杆。 ### [21:19](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=1279s) AI 落地的意外收益与深入一线现场的价值 在制造业部署 AI 往往会带来意料之外的连锁正向效应。David Hallac 分享了 Viaduct 与一家北美知名汽车制造商的合作经历：团队最初的工作重点是利用预测性分析协助工厂达成生产目标，然而经过两年的运行，该厂商在 J.D. Power 的新车质量评级（IQS）得分竟出现了持续显著的提升。令人玩味的是，Viaduct 从未直接与负责 J.D. Power 指标的部门对接，但对制造与下游系统的全局优化，却最终体现在了外部机构与媒体对端到端流程改善的认可上。正如 David 所言，这种整体性的收益往往无法在项目初期的商业计划书中精确量化，但当 AI 真正击中业务核心痛点时，整条价值链都会随之受益。 与此呼应，Peter Robinson 也提到业务线负责人 Greg 将 AI 带来的线索与其客服团队的销售提成机制挂钩，通过增销奖励极大地调动了一线员工的积极性。两位嘉宾一致指出，这类最具价值、ROI 最高的副产物绝不可能仅凭工程师坐在硅谷的办公室里凭空设想出来，研发人员必须深入工厂车间、亲自“吃自己的狗粮”（dogfooding）、与一线工人同桌吃午饭，才能敏锐捕捉到现场的实际反馈并将其转化为系统性价值。 ```mermaid flowchart LR accTitle: 制造业 AI 端到端价值外溢与现场反馈闭环 subgraph S1["现场深度融合 (Floor Presence)"] A["工程师深入车间"] --> B["一线实地测试与反馈"] end subgraph S2["核心系统优化 (AI Deployment)"] B --> C["产线预测性分析"] C --> D["生产目标达成"] end subgraph S3["衍生连锁收益 (Value Spillover)"] D --> E["J.D. Power 质量评级攀升"] D --> F["客服增销与提成机制优化"] end ``` --- ### [23:32](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=1412s) 企业构建 AI 能力的前 90 天：以问题为先与系统化作战室 当被主持人 Joe Olsen 问及传统制造企业 CEO 若想从零构建 AI 能力，在起步的前 90 天应当做什么时，David Hallac 提出了两点关键原则： 1. **坚持“问题优先”（Problem-first）思维**：对于市场上 98% 的制造企业而言，花重金聘请所谓“顶尖 AI 算法大牛”来从零自研模型完全是没有必要的。前期的核心精力应该放在对企业当前最严峻业务痛点的梳理、权衡与优先级排序上。 2. **拒绝冗长规划，追求极致速度**：生成式 AI 和现代 AI 工具的爆发已经将执行速度提高了十倍。企业最容易犯的错误就是制定一份长达两年的宏大数字化转型计划。“如果你在明确问题之后的 90 天内无法创造出实际业务价值，那你就没有做好自己的本职工作。” Peter Robinson 对此高度赞同，并补充了构建**流程体系**与**信息透明化**的重要性。KAI Labs 团队曾专门花整整一周时间脱产建立工作系统，将每时每刻的研发与推进流程标准化。为了保证高度透明，他们设立了被称为“大部屋”（Obeya / War room）的作战指挥室，管理层可以实时掌握每位开发人员当天的进度、当前冲刺（Sprint）的需求细节以及差距指标。通过快速暴露问题，这套协作体系几乎每周都在迭代进化。 --- ### [26:13](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=1573s) 站在错误的边缘：每周交付的 10 倍目标与快速迭代止损 在谈到研发节奏时，Peter 引用了戴森（Dyson）创始人的一句名言——“你必须站在错误的边缘（stand at the edge of error）”。在探索性极强的 AI 落地过程中，团队必须在心理上做好犯错和方案不及预期的准备，而快速学习的前提就是快速试错。 - **10 倍速交付目标**：为了打破“招一个人负责 AI 随便做做”的狭隘思路，Peter 给团队设定了极为严苛的目标——“每周上线交付一个 AI 解决方案”。在这一高压机制下，团队自去年 11 月以来迅速完成了 37 至 38 个项目。 - **高淘汰率与经验沉淀**：在这近 40 个项目中，大约只有 40% 到 50% 真正产生了明确的商业价值，有超过 30% 完全没有产生价值，但团队通过高频的失败周期积累了宝贵的工程经验。 - **严格的单周冲刺与止损机制**：在过去，团队曾因在一个贵金属公司的计算机视觉项目上过度死磕而影响了整体进度。为此他们重塑了工作机制：严格限定单周冲刺，项目负责人仅在次周提供技术支持，一旦确认无效立即停掉并总结教训，决不恋战；待后续通过“改善研讨会”（Kaizen）沉淀出更成熟的视觉方案体系后，再回过头去重新审视并攻克此前搁置的项目。 --- ### [29:03](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=1743s) 识别转向时机：模型演进、重塑问题定义与“准时制”交付 针对企业客户如何在庞大预算下决定技术转向（Pivot）的痛点，David Hallac 强调了“保持智识上的诚实（intellectually honest）”：AI 绝不是一个耗时数年的纯研发（R&D）黑洞。如果一个项目无法产生价值，原因通常只有两个： | 失败归因 | 应对策略与行动逻辑 | |:--- |:--- | | **基础模型能力不足** | 与其自行投入巨额资源死磕底层算法，不如**选择等待 6 个月**，直接调用最新一代能力更强的大语言模型（LLM）或开源模型来重新评估。 | | **问题定义存在偏差（更常见）** | 即使技术上攻克了该算法，它也无法解决业务的真问题。此时必须**退回到问题定义阶段**，重新梳理能真正创造商业价值的痛点。 | 在探讨“等待时机”与大型企业节奏管理时，Peter Robinson 结合精益生产（Lean）理念提出了对 AI 开发的深刻反思。他注意到，如果团队过早开发出某个 AI 工具，但业务部门和一线员工尚未做好承接准备，软件就会被闲置积压。因此，现代制造企业引入 AI 必须遵循**准时制（Just In Time, JIT）**原则——技术开发必须与业务接收端紧密协同，做到“即交付、即上线、即使用、即反馈”，避免产生数字化库存与研发浪费。 ### [31:40](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=1900s) 选对痛点：做“止痛药”而非“维他命”的快速迭代 在制造场景中推进技术方案时，找准时机和核心痛点至关重要。如果某个项目的启动时机不对，或者没有在早期发现问题并迅速修复，方案往往会逐渐失去关注甚至夭折。与其过早强推，不如在前期就明确各方是否已做好准备。对此，现场讨论提出了一个极为贴切的隐喻：“这是维他命（Vitamin）与止痛药（Painkiller）的区别”。企业在立项时必须挑选那些属于“止痛药”的项目——只要执行得当，这类直击痛点的方案几乎必然会被一线迅速采纳。 最苛刻的用户往往是最好的客户。开发团队最希望看到的是系统刚搭建完成，就有用户迅速挑出毛病，直言“这没有解决 X 问题，也没有解决 Y 问题”。这种反馈能够倒逼出以天为单位的高频迭代节奏。例如与住友（Sumitomo）等内外伙伴合作时，技术团队经常会直接进驻现场与一线人员同坐一周；周初时系统可能只能解决 30% 的问题，但到了周末就能解决 90%。一线用户被解决痛点时展现出的紧迫感与热情，正是驱动项目落地的真正源动力，而“AI 只是一个能显著加快挖掘速度的铁锹”。 ```mermaid flowchart LR accTitle: 止痛药型项目的现场高频迭代循环 subgraph Iteration["驻场迭代机制（以周为周期）"] A["锁定真实痛点<br/>（止痛药项目）"] --> B["一线挑刺与反馈<br/>（周初解决30%）"] B --> C["每日快速微调与迭代"] C --> D["现场高效交付<br/>（周末解决90%）"] end D -.->|"激发用户热情与动能"| A ``` --- ### [33:03](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=1983s) 用户消化能力与以用户为中心的产品思维 虽然制造业 AI 领域的讨论常常充斥着各种底层技术参数，但落地的真正核心往往在于文化认知与组织的消化吸收能力。一位讲者回忆起自己早期的工程生涯：当时他向客户交付了一整套完善的解决方案，对方却反而询问“能不能先只给我们三分之一的方案？”因为终端客户根本无法一次性消化全部功能，必须按月分批逐步推进。组织对于新技术的吸收存在一个固有的“胃口”和节奏。 在产品构建层面，传统工业 AI 供应商往往更像是一套“包裹了少量产品外壳的庞大销售引擎”，而真正具备长期生命力的做法则恰恰相反——必须让工厂车间的一线操作员主动“乞求”使用这款产品： - **拒绝过度销售**：不盲目追求在大型展会上搭建气派的展位，而是将重心放在赢得车间工人的心智上。 - **贴合现场体验**：优先考虑产线工人和现场工程师真正想要解决的日常痛点，设计门槛极低、人人都能上手的交互界面。 - **让用户爱不释手**：只有让实际使用者产生真实依赖，技术的长期留存与推广才会自然发生。 --- ### [34:45](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=2085s) 如何向高管层与全员有效传播 AI 成效 在推动 AI 落地一年左右、技术逐渐褪去新鲜感之后，如何向董事会与高管层汇报并维持组织内部的兴奋度便成了关键课题。Peter 分享了 Kaspar Companies 的具体实践： ```mermaid flowchart TD accTitle: AI成果内部传播与推广体系 subgraph MultiChannel["多维成果展示与共享机制"] V["标准化演示短视频<br/>（每个方案均配发）"] P["统一 AI 成果平台<br/>（集中展示状态与教程）"] S["年度商业峰会<br/>（一线员工现身说法）"] K["常态化改善成果汇报<br/>（每年100+场 Kaizen）"] end V --> P P --> S S --> K ``` - **制作成果短片与搭建统一平台**：团队为上线的每一个 AI 解决方案都录制了简短的视频，并搭建了统一的企业内部平台。全公司任何员工都可以随时访问，查看各个项目的进展状态并观看操作教程，使技术落地全流程透明可见。 - **让一线员工在年度峰会上现身说法**：在公司每年的商业峰会（Business Summit）上，核心展示由一线操作员亲自讲述的视频，分享 AI 如何改善了他们的实际工作。比起技术团队自夸，来自使用者的真情实感更能打动人心。 - **融入 Kaizen（改善）常态化汇报机制**：公司内部每年推行超过 100 场改善（Kaizen）活动。每次改善项目都会组织 30 分钟的线上（Zoom）或线下汇报会，标准化复盘“所解决的问题、经历的探索、吸取的教训、采取的对策以及创造的价值”，促进全组织的共同学习与互信认可。 --- ### [37:33](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=2253s) 组织协同：自上而下的包容与“双披萨”敏捷小团队 优秀的 AI 转型不仅需要自下而上的用户支持，同样需要自上而下的战略引导与文化包容。高管不仅要公开宣传 AI 的成功案例，更应当坦然谈论失败的项目，公开表明“我们尝试了某件事，虽然没有成功，但这完全没有问题”。组织管理的关键在于在“中央赋能”与“业务自主”之间找到平衡，既要在局部注入原生的 AI 能力，又要赋予工厂各节点尝试新工具的自主权，避免中心化的 AI 团队沦为压制创新的“阻碍者（Gatekeeper）”。 在团队架构上，应当坚决避免盲目扩张 headcount。Peter 引用了杰夫·贝索斯（Jeff Bezos）著名的“双披萨团队（Two-pizza teams）”理念，强调专职 AI 团队必须保持极度精炼： - **小团队高频冲刺**：团队仅由 4 名开发者组成，每周都能高频交付并迭代新项目。 - **摆脱官僚泥潭**：过大的团队容易陷入层层审批和沟通内耗，精简的编制能带来极高的试错和学习速度。 - **从任务组到学习机制**：相比于冷冰冰的“AI 专班/突击队（AI Task Force）”，构建一个开放、灵活的“学习型日程（Learning Agenda）”更能激发全员的自驱力。 --- ### [40:01](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=2401s) 未来一年展望：具身智能机器人与模型成本剧降 在展望未来一年的制造业风口时，嘉宾们分别给出了极具前瞻性的判断： 1. **具身智能与人形机器人（Humanoids & Robotics）**： Peter 认为机器人将是制造业的下一个核心突破口。正如过去大家探索 AI 一样，接下来全行业都将全力探索机器人如何与制造深度结合。他的团队已经将其纳入绩效考评目标（Scorecard），并且全员“迫不及待地想要搞到一台机器人并开始探索落地”。 2. **开源模型普及与算力成本出现 10 至 100 倍的下降**： David 指出，随着 AI 算力供给瓶颈逐步缓解以及开源生态的爆发，一年后模型的使用成本预计将比现在便宜 10 到 100 倍。成本门槛的大幅降低将使制造业的智能化转型变得更为激进，同时也对企业在前期规划、痛点挖掘以及落地敏捷度上提出了更高的战略要求。 ### [41:19](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=2479s) 通用横向平台整合与垂直业务场景大爆发 在对谈接近尾声时，主持人向嘉宾抛出了一个关于行业格局的核心问题：未来的 AI 市场究竟会走向极度细分的垂直利基市场，还是向少数通用型方案收拢？主持人观察到，如今走在展厅里，大约四分之三的参展商在产品功能与解决方案上高度重合。嘉宾对此指出，横向通用平台（Horizontal Platforms）必然会经历剧烈的兼并整合，如果一家公司不能切实解决具体的商业问题，将极难与 OpenAI、Anthropic 等少数头部巨头竞争。相反，深耕具体业务场景与专门问题的垂直应用将会迎来大爆发，未来会有成千上万家企业依靠熟练运用这些基础模型工具，打造能够即插即用、解决企业具体痛点的专用方案。 其他嘉宾对此深表赞同，并补充道，未来最终能够脱颖而出的企业，一定是那些以解决实际业务问题为中心，而非仅仅兜售现成单一功能（如单纯做数据采集或监测 AI 缺陷）的厂商。很多现场参展商兜售的单点工具，企业内部的研发团队其实早已实现并完成了功能整合。如果外部供应商提供的仅仅是拼图中的零散碎片，其附加值将非常微弱；唯有能够整合多种底层能力、端到端解决完整业务问题的方案，才真正具备长久价值。 ```mermaid flowchart TD accTitle: AI市场演进格局 subgraph Horizontal["横向通用层 (高集中度)"] G1["OpenAI / Anthropic 等模型巨头"] G2["底层通用 AI 基础设施平台"] end subgraph Vertical["垂直应用层 (大爆发)"] V1["深入特定业务场景"] V2["多组件端到端整合"] V3["即插即用解决实际痛点"] end Horizontal -->|"赋能与整合"| Vertical ``` - **横向平台趋势**：通用模型层加速整合，缺乏业务场景壁垒的中间层工具极易被巨头挤压淘汰。 - **垂直利基趋势**：围绕特定业务场景、能快速插拔进现有企业工作流的应用生态将迎来数以万计的增长。 - **价值核心判断**：单纯提供数据收集或单点测试的供应商价值有限，能够组合多模块直接交付业务结果的产品才是胜出者。 --- ### [43:14](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=2594s) 观众提问：AI 是构建手段还是最终产物？ 在最后倒计时的问答环节，一位现场观众拿过麦克风提出了耐人寻味的疑问。该观众表示，巡视整个展厅时到处都在空谈宏大的 AI 治理（AI Governance）或 AI 安全，描绘的愿景仿佛是云端有个类似贾维斯（Jarvis）的超级智能在接管一切；而台上 Peter 的团队却非常务实，由一群硬核工程师日夜运用 AI 高效交付方案。观众由此发问：这些最终交付的解决方案本身到底包含多少 AI 属性？还是说它本质上只是一个被 AI 工具快速敲定出来的普通优秀方案？ Peter 回应称现实情况是两者的结合，企业的目标始终是解决实际业务痛点，绝不会为了用 AI 而用 AI。他透露团队交付过一些“业务价值最高的解决方案，其底层逻辑甚至完全与 AI 无关”，但整个方案的开发构建过程却深度依赖了 AI。 其他嘉宾进一步拆解了这一逻辑，指出面对业务问题时应当厘清两重维度： 1. **开发阶段**：利用 AI 提高编程效率早已是必然趋势，正如嘉宾调侃的“现在人类都不自己写代码了，都是 Claude 在写代码”。 2. **落地阶段**：根据业务需求选择最简方案，通常分为三个层级——完全无需 AI 的方案、使用现成即用型 AI 的方案，以及定制研发 AI 的方案。原则永远是“选用能解决问题的最简单方式”。 嘉宾们最后谈到，如果市面上的应用（如使用类似 Lovable 等无代码/低代码生成工具）都在极速复刻相同的功能，缺乏深度壁垒，这种企业的生命周期可能只有 90 天到一年。企业无需为眼下不成熟的噱头焦虑，如果某个细分场景确实有真实市场，自然会有团队快速构建出适配的成熟工具。 - **构建工具 vs 业务产物**：AI 既是加速代码开发的杠杆，也是解决特定问题的可选技术栈，不可本末倒置。 - **方案极简主义**：业务落地遵循最简原则，优先选择非 AI 或现成模型，非必要不搞过度复杂的自研算法。 - **同质化与壁垒**：仅靠 AI 快速套壳生成的同质化工具生命周期极为短暂，真正的护城河在于对行业业务流程的深度理解与落地结合。 ## Notable quotes > [09:25](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=565s) AI won't replace humans, but people using AI will replace people not using AI. — David Hallac > [09:25](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=565s) the world's smartest intern — David Hallac > [10:46](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=646s) 绝对不会因为流程改善而辞退任何一名员工。 — Peter Robinson > [13:04](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=784s) 千万不要从 AI 开始 — Peter Robinson > [13:04](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=784s) 不能因为手里有锤子，就把一切都看作钉子。 — David Hallac > [13:04](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=784s) AI 会告诉工程师：我认为这个排气歧管很可能会发生故障，你应该立刻处理，因为当前已经出现了这 7 项前置征兆，并且在近期历史记录中，还有 18 台其他设备/车辆在经历完全相同的征兆序列后发生了排气歧管故障。 — David Hallac > [16:11](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=971s) No titles, everyone has a voice — Peter Robinson > [23:32](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=1412s) 如果你在明确问题之后的 90 天内无法创造出实际业务价值，那你就没有做好自己的本职工作。 — David Hallac > [26:13](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=1573s) 你必须站在错误的边缘（stand at the edge of error） — Peter Robinson > [31:40](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=1900s) 这是维他命（Vitamin）与止痛药（Painkiller）的区别 — David Hallac > [31:40](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=1900s) AI 只是一个能显著加快挖掘速度的铁锹 — David Hallac > [43:14](https://www.youtube.com/watch?v=BNrBCkh8B6s&t=2594s) 现在人类都不自己写代码了，都是 Claude 在写代码 — David Hallac ## Mentioned Peter Robinson, David Hallac, Joe Olsen, Royal Sequiera, Greg, Kaspar Companies, KAI Labs, Viaduct, Sumitomo Rubber（住友橡胶）, Olsen Capital, FGF Brands, Stanford University（斯坦福大学）, Bedrock Truck Beds, Mammoth Vision, ChatGPT, Claude, J.D. Power, James Dyson, Jeff Bezos, OpenAI, Anthropic, Lovable, Zoom ## Video description Peter Robinson - COO and Executive Lead of KAI Labs, Kaspar Companies Royal Sequiera - VP, AI & Digital Transformation, FGF Brands David Hallac - CEO, Viaduct Joe Olsen - Managing Partner, Olsen Capital This panel showcases how leading manufacturers are using AI to supercharge productivity and resilience. Learn how predictive analytics reduces unplanned downtime, how computer vision catches defects in real time, and how AI orchestrates supply chains at scale. Explore how robotics will shape the future of manufacturing. Panelists will share practical lessons on deploying AI on the factory floor—with an eye toward ROI, safety, and uptime. Join this panel to understand the cutting edge of AI within manufacturing. [Source](https://www.youtube.com/watch?v=BNrBCkh8B6s)

[9:25] David Hallac: AI won't replace humans, but people using AI will replace people not using AI. [9:25] David Hallac: the world's smartest intern [10:46] Peter Robinson: 绝对不会因为流程改善而辞退任何一名员工。 [13:04] Peter Robinson: 千万不要从 AI 开始 [13:04] David Hallac: 不能因为手里有锤子，就把一切都看作钉子。 [13:04] David Hallac: AI 会告诉工程师：我认为这个排气歧管很可能会发生故障，你应该立刻处理，因为当前已经出现了这 7 项前置征兆，并且在近期历史记录中，还有 18 台其他设备/车辆在经历完全相同的征兆序列后发生了排气歧管故障。 [16:11] Peter Robinson: No titles, everyone has a voice [23:32] David Hallac: 如果你在明确问题之后的 90 天内无法创造出实际业务价值，那你就没有做好自己的本职工作。 [26:13] Peter Robinson: 你必须站在错误的边缘（stand at the edge of error） [31:40] David Hallac: 这是维他命（Vitamin）与止痛药（Painkiller）的区别 [31:40] David Hallac: AI 只是一个能显著加快挖掘速度的铁锹 [43:14] David Hallac: 现在人类都不自己写代码了，都是 Claude 在写代码

[9:20] David Hallac: AI will not replace humans, but people using AI will replace people not using AI. [23:37] Peter Robinson: stand at the edge of error [37:33] Peter Robinson: I go back to Jeff Bezos... keep your team small. We have a small team. We have four developers. That's it. And we're... sprinting out new projects every single week. [43:14] David Hallac: Humans don't write code. Claude writes code these days... frankly, you should use the simplest one that solves the business problem.

[0:04] Um, Royal is supposed to be with us, another  uh, one of our panelists. Um, so it's just [0:08] going to be us three today. Um, we might have  some time for questions at the end now, but uh, [0:13] before we get started, why don't we do quick round  of intros. I'll let you guys kind of introduce [0:17] yourselves and businesses and awesome. Uh, my  name is uh, Peter Robinson uh, out of Shiner, [0:24] Texas. Uh, our company is Casper Companies.  It's a 128-year-old uh family-owned company, [0:30] fifth generation. And uh I've uh been at Casper  Companies for about uh almost 8 years now. And [0:37] uh we have multiple different platforms. So our  business is not just one business. Uh there's [0:42] multiple platforms with uh various different  brands underneath those. Those platforms are [0:48] three manufacturing platforms. One is around  truck beds, the other is in uh horse saddles and [0:54] then the third one is uh outdoors. uh mainly uh uh  boltaction rifles and and so forth. And then our [1:01] our fourth platform is precious metals. And so uh  yeah, that's a little bit about Casper Companies. [1:07] I'm uh David Hack. I'm the CEO of Vioaduct. We  are an AI startup that takes manufacturing and [1:15] maintenance data and uses AI to fix problems  before they occur. We are based in PaloAlto, [1:22] California. Um, we actually sold our company  last year to Sumitomo Rubber. So much of the [1:28] last year has been spent deploying integrating  our technology into the Sumitomo portfolio. [1:32] Sumitomo is a conglomerate in Japan with almost  2 million employees across the various companies. [1:38] And a lot of my focus this year is on taking the  playbook that we've been working on this last year [1:43] and starting to scale that externally with with  non-sumitomo companies as well. Awesome. Yeah. Uh, [1:49] my name is Joel Wilson. um here representing a  family office I run but uh I'm a consultant in [1:54] the AI uh and infrastructure space and have been  uh an engineer in the past still write code on [2:00] the weekends weekend warrior code writing pilots  and things like that and uh we do a lot of work [2:05] in the AI infrastructure space so I'm I'm excited  to talk about this as it relates to you know being [2:10] faster leaner smarter manufacturing and those  kind of things I guess maybe a place to start [2:15] just so everybody gets some context on where you  guys are working in the space maybe talk about one [2:21] of the AI initiatives either you finished or you  have in development right now and then you know [2:25] try to get as specific as possible so people get  a sense of it. Yeah. So one one good example of [2:31] this is actually within Sumitomo we deployed our  technology in one of their largest factories which [2:36] is in Miyazaki in the south of Japan. It's an AI  repair assistant. So they were having problems [2:42] missing production targets. And a lot of this  was due to equipment downtime and the maintenance [2:47] team was overwhelmed because the average time to  repair from the time a a um a problem occurred on [2:54] the machine to when the m technician could fix it  would be often measured in hours and it would back [2:58] up the the entire tire production process. We  built an assistant that gave them the top three [3:04] recommendations and recommended repairs based on  10 years of historical data from this factory. [3:10] and helped reduce downtime by um 40 minutes or so  per um per incident. And we now use this on over [3:19] 400 repairs just in this factory per month and  are scaling it across all of Sumitomo's factories. [3:24] Wow. Yeah. So, we have we have several different  u uh solutions. One I'll talk about is uh in our [3:33] bedrock uh which is our truck bed business.  Uh, one of the challenges they were having, [3:38] one of the problems they came came to us with  is they're having problems with uh, their actual [3:43] boat packs that goes along with the beds leaving  with the bed and actually getting shipped with [3:49] the bed. Not only making sure those boat packs  actually got shipped, but also making sure that [3:55] there were the right ones. Why was that important?  because when they get to the job shops and they're [4:01] uh you know they're looking to actually put that  bed on a uh retrofit a truck and and they take [4:07] the old truck bed off and put the new one on and  now all of a sudden they don't have the actual [4:11] hardware for it or it's the wrong hardware. Uh  our customers are not happy about that. And so [4:17] uh many of our distributors uh have uh complained  about this and what we decided to do was to do uh [4:22] what we call Mammoth Vision. And so what Mammoth  vision is is it's a opportunity for us to uh at [4:29] a certain location in a manufacturing process as  that bed is actually being uh um you know being uh [4:35] moved to the next phase. It actually has a camera  that looks at two parts of the bed. The first [4:41] camera what it does is it takes the serial number  of that bed. We have 180 different bed uh SKs. So, [4:49] it'll take it'll take the SK uh that skew number  and it'll know exactly the type of boat packs that [4:55] needs to be on there. The second camera looks  down and what that camera does is it looks at [5:00] what the boat packs are. And if the boat packs  are run wrong, what it will do is it'll tell the [5:06] uh the operator what boat packs to take off and  which ones to put on or if if it's not there, [5:12] it'll be a big flashing red light in the alarm  saying there's a problem. And uh that's how we [5:17] make sure now that every bit leave with the right  boat B packs. Well, those are both big uh uh you [5:25] know sort of straightforward solutions I guess  driven by AI but solving big problems. I know [5:30] the the end state of AI seems to get a lot of uh I  guess smart factory seems to get a lot of the air [5:37] time but the beginning is really usually pretty  messy when you're trying to figure that stuff out. [5:41] Um, I'm curious if if you guys could share maybe  one bottleneck time or one time you got stuck or [5:48] something that you had to overcome maybe in one  of the examples you gave or a different one. [5:52] Yeah. So, uh, just to put in perspective, give  you a little bit background. Just one though, [5:57] not not give you a little this morning. Give you  give you a little background. We actually started [6:02] our internal AI team, Kylabs, uh, last year  around November. And uh the reason we started [6:09] that team was because we felt like that we need  to learn faster. And so knowing that we had to [6:15] learn faster and we knew we were going to make a  lot of mistakes as we went through that process, [6:19] we gave ourselves a 10x goal. And that 10x  goal was that every week we will deploy an [6:25] AI solution. So to date, we've we've deployed  36 AI solutions. And uh I can tell you at the [6:32] beginning it was rough. We're talking about  the messiness. There was a lot of messiness. [6:36] And what we found out with the messiness was that  we did not understand the problem we were trying [6:42] to solve. What we were doing was throwing out  AI solutions. You know, someone said, "Oo, this [6:47] will be good or that'll be good." And what we had  to do is go back to our roots. And our roots for [6:52] our our company is we're a very lean uh uh centric  company. Uh if you don't know what lean is, that's [6:58] uh a a methodology of how you think and how you  uh how you solve problems and how you reduce waste [7:04] and maximize value. And so what we did decided  to do is go back to our roots and actually do [7:10] uh what they call Kaizen, which is Kaizons is  continuous improvement events. You get the right [7:15] people in the room. You identify your problem and  you actually go after the problem and the clear [7:20] the clear gap that you're trying to close. And  so when we did that, everything changed because [7:26] now we made sure that the AI solutions if it was  needed is based on the problem that we're trying [7:32] to solve. And so that was the big shift for us  and that shift really changed everything. Yeah, [7:40] I would say for me I think that has changed a lot  over the last couple years. So my background is in [7:44] AI. I got my PhD from Stanford in large scale time  series AI about 15 years ago. I think back then it [7:51] was all around clean data. I think honestly over  the last couple years that's become much less of [7:56] a problem because frankly agentic systems know how  to organize data as long as you have a sufficient [8:02] platform that can combine it. So for this repair  guidance we had all of the knowledge was in in one [8:07] of two places. It was this legacy system that was  Japanese shorthand comments from technicians that [8:14] was pretty much unintelligible and it was in the  the heads of a lot of these senior technicians. So [8:19] more or less the challenge was how do you extract  these two data sources into essentially a unified [8:24] data system and data platform. And there again  you can leverage technology that gets better every [8:29] week these days. And a lot of what we actually  focused on was more or less building a like some [8:36] reinforcement learning system that could integrate  and and learn and grow from every repair over [8:42] time. So for example, when we first deployed it,  we we're meas measuring accuracy and generally our [8:48] accuracy has gone up by almost 50% over the last  few months just as more data has flown through [8:53] and we we reached a critical mass similar to  frankly uh these large language models. Once [8:58] you you reached a critical mass of data, you just  sort of intuitively get it right and understand [9:02] the the data hierarchy and the data structure. A  lot of this is some people call it the ontology, [9:08] some people call it the platform. sort of how  all this data relates and organizes itself. I'd [9:13] say increasingly that system needs to be done very  thoughtfully, but new tools are making that a lot [9:19] easier than 5 10 years ago. For sure. Yeah. So, I  noticed uh just to deviate for a second, I noticed [9:25] both of you brought up non-technical problems and  there were more human related problems. There's a [9:30] lot of conversation when you get into AI about  AI replacing people and things of that nature, [9:36] but I'm just curious for you guys like how do you  think about the relationship between AI and people [9:41] and and how have you used those types of examples  to change your implementation strategy going [9:47] forward, right? Like how does that factor into  the you know the the technical solution, I guess. [9:52] Yeah. Yeah. And I I don't remember who said this,  but there's a quote around the fact that AI won't [9:57] replace humans, but people using AI will replace  people not using AI. Yeah. So, I think in many [10:02] ways AI is going to change the way we work, but it  is really um you know, everyone calls it whether [10:08] it's like the world's smartest intern or something  along those lines. It's something that really [10:12] needs to be harnessed and provided guardrails that  only humans both for uh institutional knowledge [10:20] perspective even liability perspective humanity  like there's a lot of reasons why but for us at [10:25] least what we found is it is a productivity boost  for people who embrace it and a lot of what it is [10:31] is it turns everyone from a individual contributor  to a manager because what you're doing is instead [10:36] of just being able to move as fast as your  fingers type you can move as fast as 10 agents [10:41] fingers can type. And that really supercharges  a lot of the the work and the the ideas people [10:46] have in their brain. But fundamentally without  the ideas in people's brain, the agents don't [10:50] know what to do and they're not actually solving  the problems. Yeah. Yeah. Yeah. Peter, what do you [10:54] think? Yeah. This is interesting because when we  first started going on this journey, we knew that [10:59] AI and all the hype about it, all the challenges  people have been hearing about how it's going to [11:04] take jobs away every day on the news and so forth  like that. And so we had to just really step back [11:10] and say like how do we approach AI in the way  that our people start to embrace it and the way [11:17] that we did that was very much how we operate as a  company. Uh we have a business system our company [11:22] has a business system how we do things. It's a  book on and that every leader in our organization [11:28] across multiple companies and how we how we do  business. One of those things that we talk about [11:33] is is continuous improvement. Getting better each  and every day regardless of AI. This is about how [11:38] do we get better each and every day and add more  value where we can grow and grow for free. And so [11:44] that concept is all about helping people see that  as they improve it allows us to add more capacity, [11:52] grow, allow them to go into new opportunities and  so forth like that. And so we've made a decision [11:57] as an organization is we never let an individual  go because we improve something, right? Uh we even [12:04] have what they call a Kaizen promotion office.  And what that office is is let's say that we [12:08] improved an area and we have three roles that no  is no longer needed in that area. We actually take [12:15] them into the Kaizen promotion office, learn our  uh learn our continuous improvement methodologies. [12:21] They get to be a part of the next improvement  and they end up in in every scenario, those [12:26] individuals got better roles and they were able  to get promoted into to new opportunities. And [12:31] so that's how we approach it. And we said we're  going to do the exact same thing with uh uh with [12:37] our uh AI uh focus. And the reason it works is it  just integrates into our system. No one can do an [12:44] AI project in in uh Kaiabs before they do what we  call a Kaizen and if we do a Kaizen we do not let [12:53] people go when we do Kaizons. So it's very clear  and and our team understands that. Yeah. Well, [12:59] both those approaches are awesome. So um that's  great to hear. Um I guess getting back to the [13:04] manufacturing stuff. So predictive versus  reactive maintenance just as like a category [13:11] um is probably a fundamental cultural shift  for for companies um whe whether you know uh [13:18] whether it's Caspar or a gigantic company in Japan  which I have to imagine is insane. Um but uh what [13:26] was the hardest part about about getting people  to sort of trust AI and data versus what their [13:31] guts telling them? Like you talked David about  trying to pull stuff out of these guys' heads or [13:36] the chicken scratching that they're doing on the  sides, you know, and um you had your own examples, [13:40] Peter. I'm just curious how you got them over that  hump. Well, well, I think it's a couple of things [13:47] for us. We don't start with AI. I I I keep saying  that and that is so important. The one thing that [13:53] we've learned is starting with AI just gets you  caught up. uh you you go down paths that doesn't [13:58] actually add value for you and then it's you get  more resistance and so forth like that. We start [14:03] with the problem for us uh several years ago.  We had a lot of problems with machines breaking [14:09] down in multiple areas of our business and we  started to do uh uh total productive maintenance [14:16] uh which is about putting ownership in our  operator's hands so that they can actually [14:21] uh be a part of the solution and making sure we  take care of our uh our machines and so forth like [14:26] that. That in itself was the big win for us. Like  our our machines haven't been going down since [14:32] then. I remember day in and day out where we were  calling people constantly calling vendors just [14:38] trying to keep machines up and we just don't have  that problem anymore. And so honestly we have not [14:44] went down a path as of yet to say hey let's go  and get AI involved in the process because it's [14:50] not our major problem. I I completely agree with  the don't use AI unless you need to. I would say [14:57] I think plenty of problems need AI to be solved  but don't you know just because you have a hammer [15:02] doesn't mean everything is a nail. I think  specifically for the equipment maintenance [15:05] work that we've done I agree there's this big  change management that needs to occur. It's you [15:10] know if they used to only react when the engine  is smoking now the engine looks great and they [15:14] don't want to touch it. Um I think ultimately this  is where explanability really comes into play. I [15:20] think a lot of you know what people in the in the  audience probably use with chat GPT or or Claude [15:26] is just you talk and it's a chatbot and doesn't  really explain its thought well. I think these [15:31] reasoning models are starting to get a little bit  better. But what's really powerful is if you say [15:35] hey I think this exhaust manifold is likely to  fail. I think you should act on it because these [15:42] seven things have occurred and here's 18 examples  in the in the recent history of other equipment, [15:47] other vehicles, other whatever it may be that  had this similar sequence of events occur and led [15:52] to an exhaust manifold failure. So you can point  to reasons behind your thinking rather than what [15:58] I think is a little more dangerous, just a pure  blackbox model. And I think that's something that [16:02] needs to be solved on a case- by case basis. don't  think off-the-shelf models solve that well, but [16:07] fundamentally those are the types of things that  I think are critical to to building trust. And do [16:11] you guys incorporate like when you're selling  these solutions or bringing these solutions to [16:15] the table internally or however it works, like  when you're bringing these things to the table, [16:19] are you proactively thinking about some of these  challenges with operators and things beforehand? [16:25] Or are you selling the technical solution, the  financial sort of impacts from that, etc., etc., [16:29] and then going, "Oh, by the way, and this is how  we're going to get them to like it." or does that [16:33] come up in the front? Yeah. Again, like for us at  the beginning, like our first handful of projects, [16:39] uh we I feel like we were trying to sell them the  solution. Yeah. And and the reason that was is we [16:45] were not starting with the problem. But when we  start with the problem and they're part of the [16:50] process, so again, when I talk about Kaizons, we  have the people who are doing the work in those [16:55] Kaizons. No one has a a title in those Kaizons.  Everyone has a a voice. And when we do those, [17:02] everyone's walking along together. And so when we  do get to the solutions, they're now part of it. [17:06] And it just makes it much much easier to to follow  through. I completely agree with what Peter said [17:12] there. I think really starting with the problem,  focusing on the pain point and in many ways [17:19] exhausting all the ways you can solve the painoint  until you come to the right solution. And very [17:23] often this you know complex AI based solution is  necessary because it is a a real PL pain painoint [17:29] blocking production blocking quality blocking  whatever it may may be but I think uh without that [17:34] you're you're in trouble to to not get the buy  in you need and like behind some of the obvious [17:39] stuff you know um you talk about thing predictive  maintenance sequels reduce downtime and things [17:44] like that. what's what's a hidden financial win  that you guys have seen um in some of the projects [17:51] that you've done whether it's like energy savings  or warranty cost reduction or waste and all this [17:56] other stuff that's not very shiny in the pitch  deck. Yeah, I I I want to talk about our first one [18:01] like you know how sometimes you get lucky on the  first one. We actually got lucky on our first one. [18:06] Uh we we did a uh a project where we were solving  a problem of inventory and our truck in our truck [18:13] bed system. So, we had so much inventory in the  yard. They were expanding the yard. And the the [18:19] issue as we started to look at what was causing  and driving this extra uh inventory because we [18:25] really make to order. So, I'm a little confused  why we're just constantly getting more and more [18:29] inventory of finished goods. And so, as we went  through that problem, what we identified was that [18:36] uh a a customer, so our customers are not the  end customer, right? It's the middle customer. [18:41] And so our customer would order a full truckload  of beds. And the whole reason for that is that [18:47] a truckload you pay for the the amount no matter  what you're going to if you have one bed on there, [18:52] you have 10 beds on there, you're going to pay the  the price for it. Now, we have so many different [18:56] sizes of beds. So, no one ever really knew how  many beds you needed. So, they'll or they'll [19:02] they'll order 10 or 12. Let's say they order 12  beds. We make those beds. We manufacture them. [19:07] And then when the truck comes, we have one guy in  the whole in our whole uh business that knows how [19:14] to like, you know, Jimmy rig it and get all the  beds in as best as he could. And so he he would [19:21] go through that process and let's say he only can  get 10 on. We'll call the customer and say, "Hey, [19:27] we need a second truck." What do you think the  customer says? No, forget forget the two. We'll [19:32] just go with the 10. Right? And so we saw that  problem. And so now we bait the two and now we [19:38] put them in, you know, in a yard and extra  extra inventory. So we were trying to solve [19:43] that problem. We came up with an AI solution where  basically when the customer calls in the order, [19:50] they would order that they would pick their  beds and then the system will automatically [19:56] uh graphically optimize the truck as they're  choosing their beds. Now what's interesting [20:02] about it is what we found is as they actually uh  chose their beds the the customer service person [20:08] on the other end would say hey you know what uh  right now you have uh you know you have 12 beds on [20:15] only could take 11 they'll take the one off but in  many cases what happened is the customer have 12 [20:21] beds and you can get another two on this was the  hidden financial thing we did not notice that hey [20:27] you know what like when we when we reach out to  the customer and say you can get another two on, [20:31] they'll say absolutely. And so they'll put an  extra two and we started seeing uh sales growth [20:36] from that. Uh and uh it was pretty cool because  we were solving one problem, but we saw financial [20:42] uh gain on a revenue end on a uh uh because of  the solution we came up with. Did you put those [20:47] commissions back into your AI lab? No, not No,  no, I couldn't do that. Uh if if you know 20% on [20:53] those things. I'm I'm gonna give I'm gonna give a  shout out to our president, Greg. If you know him, [20:57] he ain't gonna give you anything, man. He's  tight tight. Craig's not here so go ahead. Yeah, [21:06] I think for us a lot of this is everyone is very  focused on topline and quantifiable benefits. I [21:13] think very often the soft benefits are are just  as important if not more important. Um again this [21:19] is I wouldn't call it necessarily unexpected but  it's sort of by a byproduct of the type of work we [21:26] did. We were doing uh where we'd been working with  this North American car manufacturer and we were [21:31] doing a bunch of work on predictive analytics etc.  We were helping them hit production targets and [21:35] all of a sudden after a couple years they noticed  that their JD Power quality scores kept going up. [21:41] We were not even working with the team that cared  about JD power scores. And again, I'm not going [21:45] to take a ton of credit for that, but we can  take some of the credit for just broadly helping [21:49] optimize some system downstream helps, you know,  the design team, helps the manufacturing team, [21:54] helps external auditors like JD or not auditors,  external journalists like JD Power um recognize [22:01] the the endto-end process improvements that have  been uh implemented and executed over the over [22:07] the many years. And again, it's not like you can  put a dollar amount on that. It's not like you [22:11] can put a business case or a plan behind that. But  the benefit is if you're really using AI well and [22:16] driving it to solve frankly the biggest problems  of a business, the the entire endto-end solution [22:21] gets improved. And how did you track that back  to the like what what's what sort of like h how [22:27] did you become in both cases? I guess I'm curious  how did you become alerted to those two things? [22:33] Like did somebody who's in charge of revenue  come over and say, "Hey, how this is tied to [22:37] that or do you guys look for that stuff?" Well,  well, Greg told me about it. He was like, "Hey, [22:41] I'm getting this extra benefit." I was like,  "That's really cool." And he created a whole [22:44] system with his customer service reps where they  got a commission based on every upsell they got. [22:50] So, that's how he he got them motivated about it  as well. Oh, that's cool. And yeah, to chime in, [22:55] I think there's no, it was similar with us. It  just happened incidentally. I think there's really [22:59] no way of knowing these benefits and and a lot of  the challenge of these types of deployments. It [23:04] sounds like you guys are doing this as well, is  really being on the floor with with the the team [23:09] using these products. I think at the end of the  day, uh the the coincidental things tend to be the [23:15] biggest value and the biggest ROI. And you can't  learn that from, you know, from your computer in [23:19] Silicon Valley. You really need to be working  with the team, using the product, dog fooding [23:23] the the types of solutions you're deploying. and  and frankly you might hear, you know, a friend [23:28] at lunch at at the factory tell you that some  coincidental thing happened and you're the one who [23:32] can tie these two and two together. Okay. So, when  you guys are out when you're pitching this stuff, [23:37] uh let's say you're talking to uh like the CEO of  a of a business about AI and why they should use [23:42] it, just to create a very generalized example. um  if they wanted to build AI capability from scratch [23:52] um you know and and and wanted to move into kind  of modern manufacturing um organization uh what [23:59] would you tell them the first 90 days they should  be doing? I think that's a great question and [24:04] I think there's different potential blockers but  probably the most important things to focus on are [24:12] twofold. One of them is really having a problem  first mindset. I think as Peter said earlier, [24:19] coming in with uh hiring some fancy AI name who  can come in and build some model is probably not [24:26] necessary for 98% of companies out there.  So really coming in and and understanding [24:30] and ranking and prioritizing the biggest problems  of the business. And then on the execution side, [24:35] I almost have the opposite advice which is I  think speed is critical and really important. [24:40] I think AI has even over the last year 10xed  all expectations of speed. And the worst thing [24:47] that companies do is they build some two-year  plan for large scale transformation. Like I'm [24:52] sort of of the mindset that within 90 days once  you you should spend a long time identifying the [24:57] problem. Spend a long time knowing what to do and  then if you're not creating value within 90 days, [25:01] you're not doing your job. Man, I I think you hit  all the major points. Ultimately, uh I would say [25:09] with everything you said, the the other thing that  I would be doing is building out a system and how [25:15] you approach solving the problems. If you don't  have one today, if you have one, great. How do [25:20] you incorporate AI solutions within that versus  trying to do an AI thing by itself? There needs [25:28] to be a system built out. Uh our team, we spent a  whole week away just to build out a system. Every [25:35] aspect of what we do each moment is built out.  Secondly, making things visible is so important. [25:41] We have a war room. We call it Albea room. We make  things very very visible. I know what's going on [25:47] every single day with every single developer. I  know what you know we we have our requirements [25:52] and our gap of what sprint we're working on that  day that week and it is very clear. So I think [25:58] just building some of those system things out and  how you approach it really matters because you [26:04] can improve from there and we've we've changed our  system since day one probably change it every even [26:09] the last couple of weeks we've changed it right so  we're always getting better the second thing is is [26:14] that going back to the speed uh I like to say um  I think it was uh Dyson uh Dyson he had said um [26:24] you have to stand at the edge of error right and  and Really what that's saying is like you're going [26:29] to make mistakes. Things aren't going to work out  like you want it to. You have to be prepared for [26:34] that mentally. You have to get your team prepared  for that mentally, but you have to learn fast. And [26:40] learning fast mean experimenting fast. And this is  why we put a 10x goal out. When when I was first [26:45] doing this, I was going to hire one person to be  our AI guy and just run around and do things. And [26:50] I realized that that was not good thinking. What I  realized is that I have to push myself farther to [26:56] build a an excellent team. And the way I pushed  myself is I said, "We will deploy AI solutions [27:03] every week. We will deploy a solution every week."  When I pushed myself to that 10x goal, I realized [27:11] I had to have a whole different system. I had to  think about it differently. And by doing that, [27:17] we have learned so much. if I did like in a since  November if we only did four or five or a handful [27:22] of projects like only I know so much but we've  done over like I said I think we're on 37 38 [27:28] right now projects and we've learned so much in  those cycles only about 40 to 50% of them have [27:36] actually produced value there's a good 30% that  hasn't produced any value but we've learned a ton [27:43] from those projects and that's why it's important  how long PE on that sorry to interrupt you on that [27:48] on at run like how how long do you let something  like that go before you sort of say hey no more [27:52] time wasted on that we're going to move on to the  next thing or do you fully complete things and [27:57] then just go oh they never got adopted or never  implemented or didn't like like so I I was having [28:03] a problem with that up front uh with my team and  us like you know we're in we're in the middle of [28:07] something was like let's keep going let's keep  going we're trying to do a vision project for [28:11] our precious metals uh company and we kept going  kept grinding at it and what we found got in the [28:16] way is we couldn't focus on the next thing the  right way. So we had to change our mental models [28:21] on it and say we have a sprint in that week. We  do the sprint in that week. We take the learnings [28:27] from that. The lead for that sprint does support  the follow the next week. But beyond that like we [28:34] move on. If it's not working, we we'll call it  out. We'll get our learnings from it and we'll [28:39] move on to the next thing. And then sometimes  we'll come back to it. So, we did a vision [28:43] Kaizen because we've done several vision projects  right now and we did a vision uh Kaizen, my team [28:48] did to talk about what's been working and what  hasn't been working. And so, now we're refining [28:53] the the system around how we approach vision and  now we're going to go back to some of the uh the [28:58] the the vision projects that we never got success  out of. And David, on your side, like how do you [29:04] help how do you help customers understand when  to pivot or when your team decides, hey, we've [29:09] gone down a road, we need to turn left or turn  right. I need to help the customer understand, [29:15] you know, they've got big investments. The company  you mentioned is massive. So, probably not a lot [29:20] of room for wayfinding uh in the budget. So, how  do you how do you manage that? Yeah, I think uh [29:26] again similar to Peter like really it's important  to be intellectually honest as I said I think [29:33] spending a lot of time on the discovery on the  problem finding stage is is reasonable and smart [29:38] but within the deliverable and the execution like  this is not a AI is not a multi-year R&D project [29:45] and I think if things are not creating value it's  one of two reasons one is the models aren't good [29:49] enough or your models aren't good enough there  you can decide to invest more heavily in that [29:53] often it's honestly better just to wait six months  for the better LLMs to come out and then you can [29:57] re reassess it or the problem statement is not the  right problem statement meaning you can solve the [30:02] problem but it doesn't actually create value. I  think that's more often the the issue. So very [30:06] often if we try something it doesn't work we go  back to the problem statement and find the the [30:10] right problem statement that we actually think  we can create value for. Yeah. I know I noticed [30:15] so when you talked about uh waiting um I found  similarly especially if enterprise companies that [30:21] maybe move a little slower where you can't pilot  something get it to get it out into hands get it [30:26] to market whatever in 45 to 60 days or whatever  that oftentimes like waiting is the right answer [30:31] sometimes uh from a roadmap perspective do you  do you do you h how do you manage I guess how do [30:38] you manage the knowledge on whether waiting is  right or wrong how do you keep up with all the [30:42] change that's happening you know, uh, not just in  your own industry, but like what's going on with [30:47] AI in general to be to be able to make those kind  of strategic recommendations. Yeah. I, you know, [30:51] this is actually interesting because it's one that  I've been, uh, uh, reflecting on a lot lately in [30:58] terms of are we putting something out there before  it's ready to go out? And what I mean by that is [31:05] uh we've had uh a few projects where we would go  and actually build something and it was sit. It [31:13] was sit because the business isn't ready for it.  And what I found out is that's not helpful for us. [31:20] So I started going back to my lean thinking and  saying that we really need to think about just in [31:23] time. And what just in time is about is like we  have to coordinate a little bit better with our [31:30] projects and the things that we're doing. If  we're going to develop it, you don't want it [31:34] to sit. Once you develop it, you want to have  people working with it, experimenting with it, [31:40] understanding what bugs are there, getting  those things addressed immediately or what [31:45] happens is it dies away and then uh no one's  really focused on it like they need to. And so, [31:50] uh, for me, it's really more about, um, if  something is not the right timing for it, [31:55] let's figure out that upfront and let's not even  start on it. Let's wait until everyone's ready to [32:01] move forward with it. Yeah, completely agree.  This is uh the difference between a vitamin [32:05] and a painkiller. Like, pick your projects that  are painkillers, and if if you execute it well, [32:11] there's no way it won't be picked up immediately.  I can't wait to tell to ask somebody if this is [32:16] a vitamin or a painkiller. By the way, that's a  fantastic metaphor. But anyway, keep going. Sorry. [32:20] No, I I think I think it's exactly that. And and  what you want in a sense is like the the hardest [32:28] customers are the best customers, meaning you  want to build some system and then have someone [32:32] break it immediately and say it doesn't solve X,  it doesn't solve Y. And then that iteration speed [32:36] can happen on a daily cadence. Very often our  our team will go and sit with with our partners [32:40] either internally at Sumitomo or externally and  just sit with them for a week and by the end of [32:44] the week you know maybe at the beginning of the  week it solves 30% of their problems by the end [32:47] of the week might solve 90% of their problems  and that iteration and the urgency and the [32:53] enthusiasm coming from the person whose problem  you are solving is the the energy source that [32:58] drives this and AI is just like a a significantly  faster shovel to to execute whatever is needed to [33:03] to be solved. Yeah. I think um what what I find  fascinating about uh you know uh the conversation [33:11] we've had today is that like while they're while  all the stuff in AI and manufacturing you know it [33:18] seems to be a lot of the conversation is about  technical stuff but a lot of the stuff that you [33:21] guys are talking about is culture and adoption and  all of those types of things. the the example you [33:27] brought up earlier, Peter, reminded me of early on  in my engineering career, I delivered something to [33:33] uh a client that was going to sell it to another  client. And he kind of came back and said, [33:36] "Can I have a third of the solution?" And I was  like, "I don't understand, but we can do it all [33:40] right now." And he was like, "They're not they're  not ready for it. Let's do a third a day. Let's [33:43] wait a month. Let's give them a little more." But  so there's like a appetite for the ability to even [33:48] consume the thing. I I agree. One thing I found is  and for our products at least, we really emphasize [33:54] the users and in particular, I think a lot of  traditional manufacturing, you know, quote unquote [34:00] AI companies are are more or less like uh sales  forces or like sale sales engines with a little [34:05] bit of product mixed in. And I think we we try to  come with the opposite perspective, which is like [34:09] at the end of the day, you need the users at the  factory to be begging for this product. And if [34:15] they are begging for this product, all the right  things will happen. Again, you might not be able [34:18] to compete with them at these big conferences like  this and get a get a bigger booth, but but really [34:23] the the long-term durability comes from building  a product that people love. And really, in our [34:28] case with AI, it's prioritizing use cases and  applications and even just like a broader design [34:33] system that someone on the manufacturing line,  someone in the field can use, feels accessible, [34:38] and lets them actually solve the problems that  they're trying to tackle on on a day-to-day basis. [34:45] on under Peter maybe a question for you. One  question is how do you socialize the success of AI [34:51] at the executive and board levels to continue to  get people you know bought into the narrative that [34:57] we've been doing this we're doing it for longer  now you know a year in it's like hey we've been [35:01] doing it for 12 months it starts to become normal  um and then David maybe that on the same thing on [35:05] yours like h how how do you see clients you're  working with doing that like taking the wins [35:10] that you're getting turning them into you know  narrative uh that they that they use to fuel those [35:16] investments. Yeah. Yeah. You know, this is always  the hard part, right? Like just trying to figure [35:20] out how do you communicate in an effective way uh  to make sure that uh people are excited about it, [35:26] getting motivated, they want more. Uh I think  part of it is first of all doing a good job, [35:31] right? Getting good product out there. Uh people  start talking themsel uh and I love when I get [35:36] proactive feedback from uh the leaders of the  businesses and so forth like that. Secondly, uh [35:42] one of the things that we've incorporated is for  every AI solution we create, we actually create [35:48] a like a little video on it, you know. So, and  we've we've created a uh a a um overall platform [35:56] where our teams can go to anyone in the in the uh  in our business can go to that platform. They can [36:01] see every project that we've done on there, what  the status of it is, they can do the tutorials, [36:07] everything. So, we're trying to make that part  visible. that's part of our deployment plan. So [36:11] that works for every project is going to have  that. Secondly, we do have things in place as [36:16] a company. Uh so for instance, uh we we have a  business summit every year and this year we we [36:22] put a big focus on AI and we actually created  videos where the users were talking about what [36:27] they got out of the AI experience and how it made  their lives better. That's even more, you know, [36:33] heartfelt because it's actually the people who are  using it. They're seeing the value in it. Uh and [36:38] so that's another thing that we've been doing as  well. And then uh beyond that, part of our system, [36:43] our Kaizen system is we have across our company,  we do over 100 Kaizons a year. And in our company, [36:50] we have this thing uh where we uh have a  report out like tomorrow morning I'll be [36:55] on a report out that I'm actually doing for one  of our subsidiaries. Uh and so uh in that's like [37:00] a zoom call where everybody So typically it's  either in person, there is an in person part, [37:04] but then if you're not there, you do it Zoom.  And so it's a 30inut deal where they talk about, [37:09] hey, we did this Kaizen. Here's the problem we  solved, we're trying to solve. Here's uh what we [37:14] went through. Here's the learnings we got out of  it. And here's the counter measures we're putting [37:18] in place. And here's the value that we're going to  get out of it. So uh it's pretty standard and what [37:23] that looks like, but what it does, it allows the  organization to learn together, see what's going [37:28] on together, you know, recognize each other and  things like that. So that's another mechanism we [37:33] have in place for it. I I think those are spot  on. I think if I look at some of the companies [37:39] we work with that are doing the best job adopting  this this transformation versus some of the ones [37:44] that are really struggling, it is a combination of  openness and willingness to try new things. I also [37:51] think I know in the last question I talked a lot  about how you need this bottom up and user level [37:55] support. I actually think a lot of this actually  is also very top down. Meaning I think executives [38:00] need to be communicating openly about the wins of  AI, even the losses of AI. I think it's just as [38:06] valuable to say, hey, we tried this and it didn't  work and and that that it's okay. And I think [38:12] very often it is um providing the right balance  between centralization and autonomy between the [38:21] different businesses or the different nodes in  the org chart to try new things but also have [38:27] sufficient support internally to you know be able  to answer and solve any questions they have. And I [38:32] don't think there's a a one-sizefits-all solution  for that. But I think generally both bringing [38:38] some AI nativeness and AI capacity into some small  part of the company sounds like what you're doing, [38:44] but also allowing the the factory, allowing the  sales team, allowing the different teams to try [38:48] new things on their own without being gatekeep  by by a central AI force that is sort of blocking [38:53] people from trying new things and discouraging uh  experimentation. Yeah. And I I just want to add [38:57] something to that like you don't want to go and  hire a bunch of people to try to have a big team [39:02] doing AI. I don't think that's the idea. I go back  to Jeff Bezos. He used to talk about two pizza [39:06] teams back in the days. And the idea of that is  like keep your team small. We have a small team. [39:11] We have four developers. That's it. And we're  like again, we're sprinting out new projects every [39:17] single week, right? And what by doing that, it  allows you to learn faster. You don't get caught [39:22] up like you said in this like bureaucracy where  you know you got all these people and you can't [39:26] really move. Uh and so I just really encourage  you if you're going to have a dedicated team, [39:31] I do believe having a dedicated team makes a  difference for us. Uh but if you're going to [39:36] have a dedicated team, keep it very small. Yeah.  It's like the difference between you hear some [39:40] of these companies have like AI an AI teamed AI  task force versus like a learning agenda where one [39:46] feels a little more open, a little more fluent.  You talked about celebrating losses. I think [39:50] that's huge. Not just look how great AI is doing  all the time. people don't tend to, you know, [39:56] uh, react to those things very well. So, yeah. Um,  well, we've got a few more minutes, I guess. Uh, [40:01] you know, you can't do these talks without sort  of saying, let's look into the crystal ball a [40:06] little bit. So, if we're sitting here next year  on the stage having this conversation, you know, [40:11] what's what's what what are we talking about?  What's one thing you think will be, you know, [40:16] at the top of everybody's list next year? I  personally think it's going to be robotics. [40:20] I'm trying to figure out how to get a robot  myself. So, if anybody can help me with that, [40:23] please let me know. My team is already chomping  at the bit. It's it's part of uh our scorecard. [40:28] Uh and I really just think robotics is the next  level. I don't think I think it's going to be just [40:33] like AI. Like people are going to try to figure it  out. Uh and so humanoids, I'm just really excited [40:38] about it. I think for me it is solving the supply  crunch in AI and the proliferation of open source [40:45] models. I think people are experimenting with  these new tools. They're using them more and more. [40:50] I think what's going to happen a year from now is  everything will be 10 to 100x cheaper. And I think [40:55] the implications of that are that this transition  that's been happening is going to be even more [40:59] aggressive. I think doing it right the way that  Peter and I were were talking about earlier is [41:03] going to be even more important. And I think the  need for speed, the need for thoughtfulness in the [41:08] planning and the problem discovery and the general  um need to transform companies based on this new [41:15] technology will become even more important. Yeah.  Do you think we'll see uh maybe just because we [41:19] have a couple more minutes left, do you think  we'll see a uh what would I call it? I guess a a [41:26] further niching of sort of more niche AI services  versus kind of the generalized version you see out [41:33] there today. Like if you w walk the expo floor  today, I noticed you see threequarters of the [41:38] companies are probably overlapping in different  ways in things that they provide or solutions [41:41] that they provide. Do you think that will get  increasingly that way where you kind of have to be [41:47] you you have to work across all of AI to kind of  help or do you think that that will go the other [41:51] way which is people will kind of find the one or  two things they're really good at and the and the [41:54] marketplace will just open up for new things.  Yeah, I think uh the horizontal platforms will [42:00] be increasing consolidation. I think if you're not  solving actual business problems, it's going to be [42:04] harder and harder to compete with anthropic and  open AI and and a handful of giants. I think on [42:09] the specific use cases, the specific business  problems, that's going to explode. And there's [42:13] going to be hundreds, thousands, tens of thousands  of companies getting really, really good at using [42:18] these tools to solve very specific problems that  can immediately and easily plug into every company [42:24] out there. Yeah. No, I I 100% agree with you.  think is going to be the companies that I believe [42:29] is going to come to the top are going to be the  ones that are solving actual business problems [42:33] not a offtheshelf thing that's focusing on data  data gathering or some aspect of it uh you know [42:40] I saw some out there that was on like focusing on  like looking at AI bugs and like I can tell you [42:45] like my walk today out there a lot of the things  I saw my team is already doing and they've they've [42:51] consolidated those things together and so to have  a a vendor that as literally just one of those [42:58] pieces to the puzzle doesn't really help. It's  not as going to be as value added. But if you're [43:02] going to be solving a business problem and it  may take it may have uh it may align multiple of [43:08] those components to solve a business problem, I do  believe that will be valuable for uh great. Well, [43:14] we have 20 seconds left. So, we have one one lucky  winner in the crowd has a question. I think we can [43:19] manage it or Oh my gosh. All right. I'm gonna go.  Well, that he has the microphone. Hey, thank you. [43:28] Uh I think this for me has been one of the most  useful talk and you're the last ones guys. You can [43:34] keep going longer. Uh but uh what I what I find  kind of touching especially from Peter's talk is [43:42] that when you walk the floor, everybody's talking  about the AI governance, AI safety. Uh it's [43:49] almost like there's a Jarvis that oversees your  company, runs it from, you know, Google Cloud, [43:55] but Peter just has a team of uh you know, die hard  software engineers that crank out AI, you know, uh [44:04] accelerated solutions and I'm thinking are those  solutions that you make, is there anything to do [44:10] with AI once the solution is ready or it's just a  great solution for a problem that was done quickly [44:17] because of, you know, with the tool tools that  you have which are AI. That's a very very good [44:22] question. The reality is is a mixed bag, right?  You're solving the problem but you're not going to [44:27] say let's use AI when there's a better solution.  Like uh you know we've we've created multiple one [44:33] of our highest value solutions has nothing to  do with AI has nothing to do with AI but we use [44:40] AI to create it. Maybe to chime in there I think  there's two phases of this. One of them is is his [44:47] team or is any team using AI to write code faster?  I think 100% like humans don't write code. Claude [44:53] writes code these days. I think the other part is  does the business problem require AI? And I think [44:58] there's three levels. There's certain ones that  don't have any AI. There's certain ones that might [45:02] just use off-the-shelf AI and there's certain  ones that might need custom AI. And frankly, [45:06] you should use the simplest one that solves the  business problem. And there's, you know, to your [45:09] point too, it's it's interesting like it would  be really interesting to go into like lovables [45:13] database and figure out how many people have built  the same lovable solution. You know what I mean? [45:17] Because they'll say like we have two million  websites and you're like I think 1.7 million [45:20] are exactly the same because they can solve the  problem really quickly. It's it's it's pretty [45:24] much narrow and then by the time they've solved it  it's kind of like okay I've got my own version of [45:28] it. And I feel you know those those those kind of  companies out there that are building those kind [45:32] of solutions it's like are you building a business  that's 90 days long? Does it last for a year? [45:36] Like, you know, to your point earlier, the  advice you're giving businesses, sorry, [45:40] the advice you're giving businesses are like, hey,  if it if it's not here today and it's not ready, [45:44] just wait because if there's a market in it, those  guys will find it and they'll build it, right? So,