Cox Builds 5,000 AI Apps by Decentralizing Workforce Training

Cox推行去中心化员工培训，构建逾5000个AI应用 只有当企业用安全的本地化工具与多层级同伴互助学习网络取代中心化IT瓶颈和静态HR课程时，企业级AI应用才能真正实现规模化普及。 David Lee深入剖析了Cox Communications历时数年的转型历程：从2023年初仅有四至六名具备能力的AI构建者，发展到其企业总部25%至30%的员工均熟练掌握AI。Cox绕过了僵化迟缓的传统企业培训，转而建立了一套多层级的基层学习生态，涵盖基础训练营（Foundations cohorts）、爱好者开放问答（Enthusiast open Q&As）、部门先锋倡导者（departmental Pioneer champions）、实操践行者（hands-on Practitioners）以及战略专家（strategic Experts）。在社区建设的同时，该机构还部署了安全的本地GPU集群以及连接至60多个内部系统的易用工作流工具（如n8n）。这种去中心化的“传帮带”（each one teach one）蓝图使非技术员工能够安全构建超过5000个定制应用程序，并执行数百万次智能体工作流。 企业在AI应用落地过程中的主要瓶颈在于人员能力和实际熟练程度，而非获取模型或资金的渠道。 (0:04) Cox Communications的AI骨干群体从2023年初的仅4到6名构建者，发展到截至2026年年中的2000多名熟练运用AI的员工。 (0:04) 非正式的实操同伴群体（即“践行者”Practitioners）打破了组织壁垒，比自上而下的行政指令能更快加速技术模式的共享。 (3:09) 传统的静态HR培训无法跟上AI的发展节奏，因为模型能力和工具每6到8周就会更新迭代一次。 (4:11) 非正式且免于评判的问答环节营造了心理安全感，鼓励员工积极提问并采用安全合规的企业级实践，避免诉诸影子AI（shadow AI）。 (4:11)

According to David Lee, Head of Applied AI R&D at Cox Communications, enterprises must replace static, top-down instruction with decentralized, peer-led learning tiers supported by secure internal tooling to achieve meaningful AI adoption. By cultivating grassroots champions rather than relying on a centralized bottleneck, Cox scaled its AI-capable workforce from fewer than a dozen individuals in early 2023 to more than 2,000 fluent employees across thousands of deployed automations. Multi-Tiered Peer Learning Over Static Training (4:11) In early 2023, only four to six individuals across Cox's 28,000-person workforce (including 8,000 corporate staff) possessed the skills to solve enterprise problems using AI. Lee notes that conventional corporate HR programs consistently fail in this domain because underlying model capabilities and tooling evolve every six to eight weeks, rendering published curricula obsolete almost immediately. To address this gap, Cox established continuous, self-paced peer cohorts that grouped personnel by technical ability and immediate operational needs. Hands-on technical builders joined the "Practitioners" group to evaluate model releases and collaborate on system architecture. Meanwhile, non-technical staff entered the "Enthusiasts" cohort, which combined monthly presentations of approved use cases with confidential, off-the-record Q&A sessions designed to eliminate fears of corporate discipline. As Enthusiast participation surged past 800 members, Cox introduced "Foundations," a structured four-month rolling onboarding series covering large language model basics, prompt engineering, agentic loops, and workforce transition models to systematically educate newcomers. Mobilizing Departmental Champions and Autonomous Councils (6:42) To translate central research into business operations, Lee mobilized "Pioneers"—passionate, non-technical power users in operational roles such as project managers, call center leads, and analysts. Operating on an "each one teach one" model, these departmental advocates demystified technology for colleagues and established six or seven customized training initiatives within their respective units, preventing the central 35-person AI R&D team from becoming an adoption bottleneck. Governance matured alongside workforce adoption. Cox instituted the "Experts" group, a steering body of senior enterprise architects and technical leads charged with defining cross-departmental strategy, privacy guardrails, and deployment standards. Concurrently, organic specialty user groups emerged around dedicated platforms like Glean, deep research tools, and workflow automation, enabling functional teams to exchange prompts and operational playbooks without central managerial oversight. Secure Infrastructure for Enterprise Experimentation (10:10) Lee emphasizes that workforce upskilling fails unless employees are provided secure, explicitly sanctioned environments that replace blanket corporate bans with defined operational parameters. Leveraging its capital flexibility as a privately held enterprise, Cox deployed an on-premises GPU cluster hosting open-weight models behind the corporate firewall to prevent sensitive data exposure to external commercial APIs. This infrastructure supported six internal platforms, including private conversational GPT interfaces, document retrieval tools ("Chat with My Docs"), deep research agents, and an on-premise installation of the n8n automation engine. By wiring n8n into more than 60 enterprise systems—including Microsoft Outlook, Oracle, and Atlassian Jira and Confluence—Cox enabled employees to automate processes directly touching personally identifiable information, internal payroll, and customer records safely. Measurable Scale in Workflow Automation (17:45) By late 2025 and 2026, the five-tiered learning framework expanded to over 2,300 active participants, representing roughly 25% to 30% of Cox's 8,000 corporate employees. The active user base grew to encompass over 1,500 community members, more than 2,000 regular internal tool users, and upwards of 2,500 weekly active users deploying Claude and Claude Code across business and engineering operations. This decentralized literacy generated substantial technical output: between 800 and 900+ employees leveraged n8n to build more than 5,000 custom applications across dozens of business units. Operating within central governance guardrails, the platform achieved 2.3 million to 2.5 million cumulative runs, processing approximately 60,000 fully autonomous agentic workflow executions per week. Cox views its expandable "circus tent" model—scaling organizational capability by elevating internal practitioners—as a replicable roadmap for enterprise transformation, though maintaining this decentralized momentum will face practical tests as the company navigates an upcoming corporate merger.

考克斯通信（Cox Communications）应用人工智能研发负责人 David Lee 指出，传统的自上而下人力资源培训无法跟上快速迭代的人工智能技术；企业必须通过多层级点对点学习社群、基层倡导者网络与私有化安全工具链，建立去中心化的规模化应用能力。该策略使这家拥有约 28,000 名员工的企业在三年内将具备 AI 开发能力的人员从不足 6 人扩展至逾 2,300 人，构建了超过 5,000 个定制应用。 突破传统培训局限的分层学习社群 (3:09) 在 2023 年初 ChatGPT 推出约六个月后，考克斯通信全公司约 28,000 名员工（含约 8,000 名企业总部员工）中，只有不到 5 人理解生成式预训练 Transformer（GPT），且全公司仅有 4 至 6 人能切实利用 AI 解决实际业务问题。面对这一能力瓶颈，研发负责人 David Lee 认为，由传统人力资源主导的静态培训模式已完全失效，因为现代 AI 模型与工具链每 6 到 8 周就会更新一次，传统课件发布之时往往已经过时。为此，团队从基层技术人员切入，首先组建了约 30 人的“实践者”（Practitioners）月度交流群体，供数据科学家、机器学习工程师和软件开发人员分享最新模型动态、概念验证成果及架构解决方案。 到 2023 年年中，随着非技术员工在业余时间广泛接触 AI 但在办公环境中因数据隐私政策而不敢尝试，Lee 团队设立了第二层级的“爱好者”（Enthusiasts）社群。该社群每月举行会议，前 30 分钟展示获批的企业用例，后半程则关闭录制进行保密答疑，允许员工在没有“AI 监管警察”惩戒顾虑的环境中自由提问。随着爱好者群体在后期扩展至 800 人以上、新进员工难以跟进，团队进一步制度化了为期四个月的轮转“基础”（Foundations）入门课程，依次涵盖 AI 与大语言模型基础、提示工程与上下文记忆、自主智能体与循环机制，以及 AI 对未来工作形态的影响，从而形成了可持续的人才梯队引入机制。 基层先锋与点对点传帮带机制 (6:42) 到 2024 年初，爱好者社群增至约 200 人，实践者增至约 80 人，同时催生出第三类关键群体——“先锋者”（Pioneers）。该群体由项目经理、呼叫中心主管、运营主管和数据分析师等非技术业务骨干组成，他们在各自部门极具好奇心与人际影响力。Lee 援引一项研究指出，每周活跃使用 ChatGPT 的用户中有 40% 信任模型输出胜过人类建议，而在先锋者群体中有一半人对此表示认同。受 Derek Sivers 演讲中“首位追随者将孤独舞者转变为集体运动”观点的启发，中央团队将先锋者确立为影子意见领袖网络，为其提供直接辅导以破除技术虚假宣传，并借此在各业务部门孵化了 6 至 7 个定制化培训项目。 Lee 强调，一个约 35 人的中央 AI 部门绝不可能承担整个企业的所有智能化需求，规模化的核心在于“各教一人”（Each One Teach One）的金字塔架构。员工从探索可能性的爱好者成长为测试用例的先锋者，再晋升为部署工作流并指导同事的实践者，促成自下而上的知识传递。在治理层面，团队进一步设立了由资深企业架构师和技术主管构成的“专家”（Experts）委员会，负责制定全公司 AI 战略、架构治理与合规标准，同时辅以围绕 Glean、n8n 和文档对话等平台自发形成的专项用户组，让一线员工自主交流自动化模版与操作手册。 私有化基础设施与受控工具生态 (10:10) Lee 指出，单纯进行技能培训而不提供受控的工具支持毫无意义；企业推动 AI 采纳的关键在于明确告知员工“可以使用什么”，而非实施一刀切的禁令。考克斯通信利用其私有控股公司在长期资本配置上的灵活性，在企业防火墙内部署了基于本地 GPU 集群的基础设施，专门运行开源权重与开源大语言模型，完全避免了将敏感数据传输给第三方外部 API 的合规风险。 依托该基础设施，中央团队搭建了六大内部平台，包括私有对话界面、企业文档检索系统（Chat with My Docs）、私有深度研究智能体，以及本地部署的工作流自动化平台 n8n。通过将 n8n 安全接入微软 Outlook、Atlassian Jira/Confluence、Oracle 等 60 多个内部企业系统，员工能够在严守合规的前提下，编排涉及个人身份信息（PII）、客户账户及企业薪酬数据的复杂业务流程，为跨部门的大规模自主应用构建了安全沙箱。 规模化落地指标与马戏团帐篷蓝图 (17:45) 截至 2025 年末至 2026 年中，考克斯通信在 8,000 名总部员工中实现了 25% 至 30% 的 AI 实际熟练度。具体落地数据表明：企业内部拥有超过 1,500 名活跃社群成员，2,000 多名定期使用内部 AI 工具的员工，以及超过 2,500 名每周使用 Claude 和 Claude Code 的工程与业务人员。此外，已有 800 至 900 多名员工使用 n8n 自动化平台累计构建了超过 5,000 个定制应用程序；该智能体平台已累计完成 230 万至 250 万次运行，每周处理约 60,000 次完全自主的智能体工作流执行，全部运行在受监控的集中防护体系之内。 Lee 将此转型框架归纳为“马戏团帐篷”（Circus Tent）模型：起初规模紧凑，但随着核心实践者不断提升能力并立起支撑柱，帐篷的覆盖面积随之扩大，最终将更多员工纳入保护与赋能伞之下。他总结该蓝图包含四大通用原则：按学习阶段划分的同侪社群、深入业务单元的导师冠军、具备事实依据且易于传播的成功经验，以及具备严格监控与明确安全边界的操作沙箱。 David Lee 表示，尽管考克斯通信面临后续的企业并购，但这套涵盖社群分层、倡导者网络与私有沙箱的采纳蓝图可直接迁移至其他企业。然而，这种自下而上的生态能够持续运转，高度依赖企业私有基础设施的资本投入与跨部门的自发维护；在未来更大规模的企业合并或重组过程中，该去中心化网络能否维系既有的组织动能仍有待实践检验。

## Background 该内容由 YouTube 频道 Ai4 于 2026 年 8 月 15 日发布，记录了考克斯通信（Cox Communications）应用人工智能研发副总裁 David Lee 的主题演讲。视频简介指出，演讲聚焦于企业如何攻克人才瓶颈，通过自下而上的愿景锚定、梯队化学习社群以及低门槛工具底座，将掌握 AI 构建能力的人数从最初的不足十人扩展到两千人以上。 ## TL;DR 面对传统培训周期滞后与企业 AI 认知断层的挑战，考克斯通信摒弃一刀切的禁令与枯燥课程，创新性地组建了践行者、爱好者、先锋者与基础班等分层互助学习社群，构建起民间的影子影响力网络。配合本地私有化 GPU 集群与 n8n 等自动化安全工具底座，公司在保障数据合规的前提下为基层业务员工全面赋能。历经三年实践，仅有 35 人的核心 AI 团队带动全公司 25%–30% 的员工跨越了 AI 门槛，非技术人员自主构建了超过 5,000 个业务应用并常态化运行海量 Agent 工作流。 ## Atmosphere & style 演讲整体氛围风趣幽默且务实坦诚，讲者开场借由自嘲与弟弟的 TED 播放量差距迅速活跃了现场气氛，让沉重的企业组织变革话题显得平易近人。其叙事既有从 4 到 6 人摸索起步的真实细节，又有对同侪自组织裂变与安全心理机制的敏锐洞察，生动展现了将前沿技术火种转化为全员自驱创新浪潮的落地质感与启发力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=MwTTF9qjnns&t=4s) 企业推进 AI 落地的最大瓶颈并非模型算力或预算，而是员工的实际应用素养；考克斯早期在 2.8 万名员工中仅有 4 至 6 人具备 AI 构建能力。 - [03:09](https://www.youtube.com/watch?v=MwTTF9qjnns&t=189s) 传统自上而下的 HR 培训体系在快速演进的技术浪潮中极易过时，以同侪互助为核心的敏捷学习社群是提升组织 AI 素养的高效路径。 - [03:09](https://www.youtube.com/watch?v=MwTTF9qjnns&t=189s) 通过双轨机制分流培养：技术骨干加入“践行者”切磋工程落地，非技术员工加入“爱好者”并在不录制、不惩罚的安全空间里探索合规用法。 - [06:51](https://www.youtube.com/watch?v=MwTTF9qjnns&t=411s) 吸纳狂热的业务骨干组建“先锋者”，将其打造成遍布各业务线的“影子影响力网络”，自发在部门内部孵化培训并推动组织协同。 - [10:10](https://www.youtube.com/watch?v=MwTTF9qjnns&t=610s) 推动全员使用 AI 的前提是提供确定性：管理者不能只下一道禁止禁令，而应构建私有化安全底座并公布官方允许使用的工具清单。 - [15:09](https://www.youtube.com/watch?v=MwTTF9qjnns&t=909s) 针对零基础员工设计 4 个月滚动周期的“基础班”，系统化讲授大模型机理、提示词工程、智能体机制及未来工作形态，补齐人才培养闭环。 - [16:23](https://www.youtube.com/watch?v=MwTTF9qjnns&t=983s) 核心 AI 团队无法承接所有业务需求，必须依托“一人教一人”的金字塔晋升机制，赋能全员实现 25% 至 30% 的跨部门能力裂变。 - [17:45](https://www.youtube.com/watch?v=MwTTF9qjnns&t=1065s) 安全可控的工具平台大幅释放了非技术员工的生产力：900 多名基层员工自发搭建 5,000 余个应用，Agent 平台每周执行 6 万次全自动工作流。 - [18:34](https://www.youtube.com/watch?v=MwTTF9qjnns&t=1114s) 落地企业 AI 赋能的“马戏团大篷车”方法论：按认知阶段分层组建社群、在各部门培养传播骨干并赋予易传播的故事、下沉低门槛工具与明确规范。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=MwTTF9qjnns&t=4s) 现场互动与开场：企业内部的 AI 素养困局与讲者背景 演讲一开场，David Lee（考克斯通信应用人工智能研发副总裁）向全场抛出了一个直击痛点的问题：“在座的各位，你们公司今天有多少比例的人真正懂得如何用好 AI？整体的 AI 素养（AI fluency）究竟如何？”现场举手表决的结果颇为现实：认为公司内有超过 50% 的人精通 AI 的寥寥无几，仅有两三个人举手；至于 25% 左右的也不多；绝大多数观众所属企业的 AI 熟练度都在 25% 以下。这正是 David 过去几年一直在思考并尝试攻克的企业级难题。 为了介绍这一探索的渊源，David 分享了自己的职业经历：考克斯通信（Cox Communications）是全美第三大有线电视与宽带服务商；在加入考克斯之前，他曾担任 UPS 全球创新主管，更早之前则是美国一家大型银行的创新负责人。2017 年，他曾做过一场关于“智能机器时代如何重构工作流程”的 TED 演讲，意外成为当年观看次数最多的 14 场 TED 演讲之一，斩获约 250 万次播放。幽默的是，他打趣称自己“依然是亚裔父母眼中的不肖子”，因为他的弟弟 Johnny Lee 早在 2008 年就做了一场极其轰动的 TED 演讲，播放量高达 700 万次，跻身 TED 历史前 25 名之列，因此他调侃台下观众“如果看我的演讲，千万别去看我弟弟的”。 回到业务本身，早在 2013 年他在银行工作时便组建了数据科学部门，2018 年在 UPS 搭建了机器学习部门，并在 2023 年正式在考克斯创立了 AI 研发部门。2023 年初，即 ChatGPT 问世大约六个月后，他清醒地意识到，大企业亟需让更多员工掌握 AI 的使用方法。而在当时，整栋办公大楼里真正懂得 GPT 是什么的人屈指可数，可能不到 5 个人，幸运的是其中有 3 人就在他的团队。 - 考克斯通信企业规模：全美第三大有线电视公司，拥有约 28,000 名员工，其中总部企业白领人群约为 8,000 人。 - 早期能力现状：2023 年初，全公司能接下业务难题并用 AI 去解决的专业人员仅有 4 到 6 人。 --- ### [03:09](https://www.youtube.com/watch?v=MwTTF9qjnns&t=189s) 学习社群的起步：践行者（Practitioners）与爱好者（Enthusiasts） 只有四到六个人真正懂行，另有几个人渴望学习，如何让整家大企业聪明起来？考克斯的做法没有诉诸冗长的传统自上而下式培训，而是选择“把人聚在一个屋子里直接交流”，打造以同侪为中心的高频、低试错成本的学习社群（Learning Communities）。 第一批被组织起来的是“践行者”（Practitioners）。这群人主要由数据科学家、软件工程师等重度极客组成，属于键盘敲代码的“核心构建者（Builders）”。David 笑着调侃，这群人平日里“除了星际迷航（Star Trek）和 LLM 模型之外几乎聊不来别的”，在 2023 年的派对上大家最怕被他们堵在角落滔滔不绝地讲 AI 有多酷。最开始这个群体约有 30 人，每月固定开会交流一次，抛出具体的技术新突破或者实操碰到的技术壁垒，让这群第一线敲键盘的技术人员能够面对面切磋解法。 到了 2023 年年中，AI 的狂热蔓延到了非技术业务岗位。许多员工在下班回家后，用 ChatGPT 制定度假路线、做健康规划，每周在家自发使用 5 到 20 个小时。然而，当他们跨进公司大门时却陷入了迷茫：“我不知道在工作中我被允许做什么，哪些能传，哪些不能传。”面对这种灰色地带，考克斯没有采取一刀切的封杀禁令，而是成立了第二个社群——“爱好者”（Enthusiasts）。该社群每月开展一次活动，前 30 分钟由内部专家分享 AI 基础逻辑以及合规的企业级落地实践，后半截则直接“关闭录制按钮”，让大家在一个完全放松、不被惩罚的安全空间里畅所欲言。员工可以放心承认“我前天上传了某个文档”，组织者则会耐心引导“这种做法不可取，你应该换成这种方式”。他们不当“AI 警察”，而是为员工建立清晰的行为边界与心理安全感。 ```mermaid flowchart TD accTitle: 考克斯早期社群演进路径 accDescription: 展示从极少数技术人员到建立践行者社群与爱好者社群的演化路径 subgraph Origin["2023 年初：火种期"] Seed["4-6 名核心技术先锋"] end subgraph Mid2023["2023 年中：双轨社群并行"] P["践行者社群 (Practitioners)<br/>• 约 30 人起步<br/>• 数据科学家 / 软件工程师<br/>• 月度技术切磋与代码实操"] E["爱好者社群 (Enthusiasts)<br/>• 快速扩张至数百人<br/>• 非技术业务人员<br/>• 30分钟合规知识 + 闭门问答"] end Seed --> P Seed --> E ``` - 传统培训的困局：HR 或组织发展部门往往面临“课程刚开发完毕，内容就已经在一个半月内过时”的尴尬节奏，唯有敏捷互助的学习社群能跟上技术迭代。 - 社群规模扩张：截至 2023 年末至 2024 年初，践行者扩展到约 80 人，爱好者社群迅速吸纳了约 200 人。 --- ### [06:51](https://www.youtube.com/watch?v=MwTTF9qjnns&t=411s) 先锋者（Pioneers）：构建民间的影子影响力网络 进入 2024 年初，David 敏锐地察觉到需要创立第三支特殊力量，他将其定义为“先锋者”（Pioneers）。这群人大多完全不懂代码，是项目经理、数据分析师或呼叫中心主管等非技术业务带头人，但他们对 AI 的信任和痴迷程度甚至超越了常理。在第一次先锋者会议上，David 分享了一项调查——约 40% 的每周活跃 GPT 用户表示，相比生活中身边的任何真人，他们更信任 GPT 给出的回答。当时 David 觉得这简直不可理喻，没想到会场里有一半的先锋者当即表示：“这说的就是我。” 在各个业务部门里，先锋者往往扮演着“当地那位狂热的 AI 专家”角色。同事们遇到 AI 问题，不会去找遥远的技术中心，而是会去敲这名业务同事的桌子。David 敏锐地抓住了这种社群自组织机制，将这些非正式意见领袖正式吸纳并赋予“先锋者”的身份认同。一旦戴上这枚无形的荣誉徽章，他们便自发将自己视为代表公司正确路线的“思想领袖”。 通过每月的交流，技术团队帮助先锋者厘清流言与事实，纠正关于模型底层原理的错误认知，指导他们如何化解普通员工对“AI 窃取公司数据”的恐慌。这群人迅速演变成了分布在企业各个角落的“影子影响力网络（Shadow Influencer Network）”。令人惊喜的是，这些先锋者随后回到各自的呼叫中心或业务线，自主孵化建立了六到七个定制化的部门正式培训项目，形成了遍地开花的自驱态势。David 借用 Derek Sivers 著名的 TED 演讲《如何掀起一场运动》（How to Start a Movement）做类比：先锋者就是那群最初在荒野中独自起舞的狂人，正是因为他们的带动，个人的疯狂才最终演化成了波及全公司的协同浪潮。 --- ### [10:10](https://www.youtube.com/watch?v=MwTTF9qjnns&t=610s) 安全底座与多级社群体系的成型 光有社群的宣导还不够，若没有一套合规安全的企业级技术底座，员工终究会陷入寸步难行的境地。由于考克斯是一家私有控股公司，享有长期战略投资的资本自由度，David 的团队在 2024 年年中直接斥资搭建了本地化 GPU 算力集群（On-prem GPU system）。他们全面拥抱开源与开放权重模型，先后自主搭建了私有化版本的 Private GPT、“Chat with my Docs”（私有知识库对话）以及“Deep Research”（深度调研助手）。 为了打通企业业务骨干，他们部署并私有化集成了工作流自动化开源工具 n8n，将其与底层的 Outlook、Atlassian 套件、Oracle 等 60 多个核心业务系统打通。因为所有推理和数据流转都在企业防火墙内完成，员工可以毫无顾虑地处理包含薪资数据、客户敏感信息（PII）的高安全等级任务。David 强调，对于大多数没有充裕算力预算的企业而言，最核心的启示在于：**想要员工积极使用 AI，管理者不能只下一道冷冰冰的“禁止指令”，而是必须清晰明了地公布一张“官方允许使用的安全工具清单”**。 ```mermaid flowchart TD accTitle: 考克斯多层级社群矩阵与技术底座 accDescription: 说明自上而下的战略专家、专业用户组、先锋者、践行者、爱好者及基础班体系，依托底层私有化基建运行。 subgraph Governance["治理与战略层"] EXP["专家群 (Experts)<br/>架构师与各部门技术领袖·制定企业 AI 战略与合规路线"] end subgraph ToolGroups["工具专长组 (Specialty Groups)"] SP["围绕具体工具自发成立 (Glean, n8n, Chat with my docs, Deep Research)"] end subgraph CommunityNetwork["核心学习与扩散网络"] PIO["先锋者 (Pioneers)<br/>业务骨干与影子影响力网络·跨部门扩散理念"] PRA["践行者 (Practitioners)<br/>一线工程师与数据科学家·技术研发与落地"] ENT["爱好者 (Enthusiasts)<br/>跨职能广泛员工·从初阶成长为中高阶用户"] end subgraph Onboarding["新人导入层"] FND["基础班 (Foundations)<br/>4 个月滚动周期·入门通识四步走"] end subgraph Infra["企业私有化安全底座"] GPU["本地私有化 GPU 集群"] --- Models["开源开放权重模型"] Models --- Tools["Private GPT / 私有知识库 / n8n 连接 60+ 系统"] end EXP --- ToolGroups ToolGroups --- CommunityNetwork CommunityNetwork --- Onboarding Onboarding -.运行在.-> Infra ``` 随着工具的快速普及，原有社群开始产生分化，演变出了更加精细的五层矩阵： 1. **专家群（Experts）**：并非纯编程高手，而是由各部门资深架构师、政策合规与技术带头人组成的智囊团。该社群本质上是全公司的 AI 战略核心，负责统筹全企业范围的战略推进与落地准则。 2. **工具专长组（Specialty Groups）**：完全由业务部门自发建立的工具拥护者俱乐部。针对企业搜索 Glean、工作流引擎 n8n、文档助手和深度调研工具，高阶用户每月碰头交流使用妙招，堪称企业内部低成本推动落地最快的一步棋。 3. **先锋者群（Pioneers）**：持续承担业务部门内部的技术布道与答疑角色。 4. **践行者群（Practitioners）**：专注算法、工程落地及技术挑战攻坚的代码极客圈。 5. **爱好者群（Enthusiasts）**：规模逐渐滚雪球般攀升至 800 人以上。随着两年的沉淀，这一群体整体知识架构大幅提升，早已不再是初学者。 --- ### [15:09](https://www.youtube.com/watch?v=MwTTF9qjnns&t=909s) 补齐最后拼图：基础班（Foundations）与两千人跃升 由于老牌“爱好者群”的平均认知水平已经相当深厚，新加入的零基础员工往往感到门槛太高。于是在演进的最终阶段，团队补齐了闭环的最后一环——面向完全零起步人群的“基础班”（Foundations）。 基础班不同于长期持续运作的俱乐部，它被设计为一个滚动式的为期 4 个月的入门培养项目。学员每月参加一次 60 分钟的学习（40 分钟主题讲授 + 20 分钟互动答疑），内容结构清晰明了： 1. **第一讲：AI 与大语言模型基础通识（Foundations of AI & LLMs）**：拆解底层技术运转机理与基本概念。 2. **第二讲：提示词工程、上下文与记忆机制（Prompting, Context & Memory）**：掌握与模型高效沟通的实操技巧。 3. **第三讲：智能体机制与系统循环（Agents & Loops）**：理解复杂自动化与 Agent 的运作逻辑。 4. **第四讲：工作与个人未来的前瞻视野（The Near and Far Future of Work）**：建立对未来组织变革的宏观认知，指导个体在 AI 浪潮中持续进阶。 在这一系列体系化社群与基础设施的推动下，到了 2026 年年中，考克斯内部常态化参与各类 AI 社群与研讨的人数已从三年前的 20 余人狂飙至约 2,300 人。在全公司约 8,000 名办公室总部员工中，具备良好 AI 素养的员工占比实现跨越式增长。回顾这一历程，David 感慨这套最初连自己心里都没底的探索，最终证实了一条颠覆传统培训模式的可行路径：在剧烈演进的技术浪潮中，企业必须依靠自治、互信的梯队社群和安全的本地基础设施，才能将凤毛麟角的专家火种，迅速燎原为整支企业的实质战斗力。 ### [16:23](https://www.youtube.com/watch?v=MwTTF9qjnns&t=983s) AI 普及率与“金字塔”式社群裂变效应 无论是为了团队、企业、家庭、学校还是国家的生存与发展，尽可能提高全员的 AI 熟练度（AI Fluency）都至关重要。目前在 Cox Communications 内部，已有大约 25% 到 30% 的员工掌握了不同程度的 AI 能力。这一比例意味着整个团队与员工个体都具备了穿越技术周期的能力。在过去三年的社区建设中，最令人欣慰与充满成就感的时刻，莫过于许多普通员工走上前来感叹：“天哪，你彻底改变了我的生活，我现在竟然能做到这么多以前根本无法想象的事情。” 这种规模化的普及并不是单靠某一个人的力量实现的，而是得益于“金字塔”式的层级跃迁机制。随着员工从 AI 爱好者（Enthusiasts）进阶为先驱者（Pioneers），再成长为实践者（Practitioners），他们自然而然地践行“一人教一人（each one teach one）”的理念，带动各自部门的同事共同学习。对于一个仅有 35 人的 AI 核心团队而言，不可能承接全公司所有 AI 项目的需求；最有效的战略是赋能尽可能多的人，为他们提供完备的工具、扎实的基础知识以及安全的实验环境，从而实现自主繁荣与规模化裂变。 ```mermaid flowchart TD accTitle: AI Community Growth Pyramid accDescription: Flowchart illustrating how AI literacy scales through community progression and peer learning. subgraph Growth["金字塔式层级晋升与社区裂变"] direction TB A["AI 爱好者 (Enthusiasts)"] --> B["AI 先驱者 (Pioneers)"] B --> C["AI 实践者 (Practitioners)"] C --> D["各自部门内导师 / 传播者 (Each One Teach One)"] end subgraph Empowerment["核心 AI 团队赋能 (35人团队)"] E["提供基础知识与培训"] F["提供低门槛工具与平台"] G["构建安全受控的实验环境"] end Empowerment -.-> Growth ``` - **团队规模对比**：仅靠 35 人的核心 AI 部门无法包揽所有需求，全员赋能是实现规模化的唯一路径。 - **普及率现状**：全公司约 25%–30% 的员工已具备一定水平的 AI 熟练度。 - **裂变逻辑**：从爱好者、先驱者到实践者层层进阶，由先学者反哺各部门新手，形成自驱型学习生态。 --- ### [17:45](https://www.youtube.com/watch?v=MwTTF9qjnns&t=1065s) 自动化工作流与 Agent 平台落地成效 经过三年的深耕，Cox 在工具使用与应用构建上取得了极其可观的数据。目前，活跃在 AI 社区的成员已超过 1,500 人，日常使用内部 AI 工具的员工达到 2,000 人，每周使用 Claude 及 Claude Code 的活跃用户更是突破了 2,500 人。在自动化工作流平台（如 n8n 等自动化系统）上，已有 800 至 900 多名员工参与，他们在这个许多人眼里的“传统老牌有线电视公司”中，自发构建了超过 5,000 个应用程序。 此外，公司的 Agent 平台（Agentic Platform）已累计完成了约 230 万至 250 万次任务执行，每周运行的全自动 Agent 工作流达 6 万次。这 900 多位非传统开发背景的员工跨越数十个部门，自主编写并部署了各类 Agent 流水线。尽管核心团队无法一一了解每一个具体工作流的业务细节，但所有运行都在受到严格监控、保障合规且安全的框架内稳定运转，真正释放了基层业务的创新活力。 - **社区与用户规模**：社区成员 1,500+ 人，内部工具常规用户 2,000 人，Claude / Claude Code 周活跃用户 2,500+ 人。 - **应用构建数据**：900+ 员工在自动化平台上搭建了 5,000+ 个应用。 - **Agent 执行量**：累计完成任务 230 万至 250 万次，每周全自动 Agent 工作流执行量约 60,000 次，覆盖数十个业务部门。 --- ### [18:34](https://www.youtube.com/watch?v=MwTTF9qjnns&t=1114s) “马戏团大篷车”心智模型与企业落地建议 尽管公司正经历合并等组织层面的变革，未来走向充满未知，但这种以人为本、社区先行的赋能模式在绝大多数企业组织中都具备高度可复制性。这一策略的核心心智模型可以形象地比喻为“撑起马戏团大篷车（Circus Tent）”：刚开始时棚顶很低、空间很小，随着团队不断探索与学习，大棚被不断向外、向上撑大，越来越多的员工被接纳进大篷车内，并沿着学习路径逐步走向更高级别的安全应用。 对于希望在自身组织中推广 AI 能力的企业，可以遵循以下落地路径： 1. **按学习阶段组建社群**：连接处于相近认知阶段的员工，让他们相互交流解惑，并为向更高阶段跃迁的人持续提供成长空间。 2. **在各部门培养传播骨干**：发掘并培养能够在本部门传授知识的带头人，赋予他们“易于转述的故事（retellable stories）”，让基于事实的最佳实践在组织内部自发传颂。 3. **提供安全工具与清晰指引**：在合规与受控的前提下，将低门槛工具与明确规范下沉到基层，让员工敢于动手、能够成事。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=MwTTF9qjnns&t=4s) 在座的各位，你们公司今天有多少比例的人真正懂得如何用好 AI？整体的 AI 素养（AI fluency）究竟如何？ — David Lee > [00:04](https://www.youtube.com/watch?v=MwTTF9qjnns&t=4s) 我依然是亚裔父母眼中的不肖子 — David Lee > [00:04](https://www.youtube.com/watch?v=MwTTF9qjnns&t=4s) 如果看我的演讲，千万别去看我弟弟的 — David Lee > [03:09](https://www.youtube.com/watch?v=MwTTF9qjnns&t=189s) 我不知道在工作中我被允许做什么，哪些能传，哪些不能传。 — David Lee > [06:51](https://www.youtube.com/watch?v=MwTTF9qjnns&t=411s) 相比生活中身边的任何真人，他们更信任 GPT 给出的回答。 — David Lee > [06:51](https://www.youtube.com/watch?v=MwTTF9qjnns&t=411s) 这说的就是我。 — David Lee > [10:10](https://www.youtube.com/watch?v=MwTTF9qjnns&t=610s) 想要员工积极使用 AI，管理者不能只下一道冷冰冰的“禁止指令”，而是必须清晰明了地公布一张“官方允许使用的安全工具清单” — David Lee > [16:23](https://www.youtube.com/watch?v=MwTTF9qjnns&t=983s) 天哪，你彻底改变了我的生活，我现在竟然能做到这么多以前根本无法想象的事情。 — David Lee > [16:23](https://www.youtube.com/watch?v=MwTTF9qjnns&t=983s) 一人教一人（each one teach one） — David Lee ## Mentioned David Lee, Johnny Lee, Derek Sivers, 考克斯通信 (Cox Communications), UPS, TED, ChatGPT, GPT, Private GPT, Chat with my Docs, Deep Research, n8n, Outlook, Atlassian, Oracle, Glean, Claude, Claude Code, 《如何掀起一场运动》（How to Start a Movement） ## Video description David Lee - AVP, Applied AI Research & Development, Cox Communications Three years ago, fewer than a dozen people at Cox knew how to build with AI. Today, that number is 2,000 and climbing. Every company is now racing to turn AI hype into real capability, but the bottleneck isn't models or budget—it's people who know how to use them. This session walks through how Cox scaled AI fluency across the organization: the vision that anchored it, the learning communities that spread it, and the tooling and platforms that made building accessible to non-experts. Walk away with a practical path for growing human AI capability from a handful of pioneers to a workforce. [Source](https://www.youtube.com/watch?v=MwTTF9qjnns)

[0:04] David Lee: 在座的各位，你们公司今天有多少比例的人真正懂得如何用好 AI？整体的 AI 素养（AI fluency）究竟如何？ [0:04] David Lee: 我依然是亚裔父母眼中的不肖子 [0:04] David Lee: 如果看我的演讲，千万别去看我弟弟的 [3:09] David Lee: 我不知道在工作中我被允许做什么，哪些能传，哪些不能传。 [6:51] David Lee: 相比生活中身边的任何真人，他们更信任 GPT 给出的回答。 [6:51] David Lee: 这说的就是我。 [10:10] David Lee: 想要员工积极使用 AI，管理者不能只下一道冷冰冰的“禁止指令”，而是必须清晰明了地公布一张“官方允许使用的安全工具清单” [16:23] David Lee: 天哪，你彻底改变了我的生活，我现在竟然能做到这么多以前根本无法想象的事情。 [16:23] David Lee: 一人教一人（each one teach one）

[0:04] David Lee: a disappointment to my Asian parents [3:09] David Lee: real nerdy people that basically are not interesting to hang out with unless you're talking about Star Trek or LLM models. [4:11] David Lee: dumb questions [4:11] David Lee: AI police [6:42] David Lee: crazy person [10:10] David Lee: Chat with My Docs [16:23] David Lee: Holy cow, you changed my life because now I can do all these things that I didn't know I could do before. [16:23] David Lee: each one teach one [17:45] David Lee: sleepy cable company [18:34] David Lee: circus tent

[0:04] I thought uh I'd spend a little time talking about  what we've been doing at Cox Communications over [0:07] the last three years uh because I think we've  done something remarkable. Um the uh the question [0:14] I want to start with is the following. Uh which is  thinking about your company today, what percentage [0:20] of the population knows how to do AI well? Like  what's your AI fluency today? So, just by show of [0:26] hands, who thinks that their company has like  more than 50% of your people are AI fluent? [0:33] One hand. That's two, three. That's pretty  good. Pretty good. Okay. Um, how about 25%. [0:39] All right. What about less than 25%. About most  of the room. So this is the challenge that that we [0:46] that I've been thinking about um for a couple  of years which is uh just as for background um [0:53] I run applied AI research for Cox Communications.  Cox is the third largest cable company in the [0:59] United States. Um prior to this I was the head  of global innovation for UPS and prior to that [1:04] I was the head of uh innovation for a large  um US bank. Uh, in 2017, I gave a TED talk [1:11] about how to re-engineer work for an age of  intelligent machines. And by my surprise, [1:16] it was one of the 14 most watched TED talks of  the year with about two and a half million views. [1:21] Um, don't get too excited. I am still a  disappointment to my Asian parents. Uh, [1:26] because my younger brother also gave a TED talk,  but in 2008, and he got about 7 million views [1:31] and was one of the 25 most watched TED talks  of all times. So, if you do watch my TED talk, [1:36] I'll ask you not to watch my brother's TED talk.  Uh, his name is Johnny Lee. It's actually a very [1:40] good TED talk, but it's a little bit dated  because it's a couple years old. So, anyway, [1:44] in 2023, about 6 months after Chat GBT came out,  we really realized that it was time to sort of get [1:51] this large enterprise geared around how do I get  more people to know more about how to use AI? um [1:59] in the sense that uh like there were probably  fewer than five people who understood what GPT [2:09] was prior to chat GPT inside of the building.  Uh luckily three of them were on my team. [2:13] Um and what we've been doing some early research  around how do you use generative AI tools, both [2:19] text generation and image generation in sort of  new product development, right? Um so looking at [2:25] oh you can't even see it. there's a tiny little  purple triangle underneath the thing. There were [2:28] four to six people inside of the company in early  2023 who you could give them a problem and say, [2:33] "Can you use AI to solve this problem?" Which was  not enough. And our hypothesis was at the time um [2:39] that AI was going to be big and that turned  out to be correct. I had a little bit of a [2:43] head start though uh I mentioned I used to run  uh innovation for a bank. As part of that, we [2:47] launched the data science department in 2013. And  then when I went to UPS, we launched the machine [2:52] learning department there in 2018. And then I  formed the AI department inside of Cox in 2023. [2:59] So the question is, if I have five to six people  who really know what they're doing and five to six [3:04] more who want to be good at what they're doing,  how do I get everybody smarter? So we said, well, [3:09] let's just get them in a room and just talk.  So we for formed our first learning community [3:14] that we called our practitioners. And  these are typically like data scientists or [3:18] or sort of software engineers, like  real nerdy people that basically [3:22] are not interesting to hang out with unless  you're talking about Star Trek or LLM models. [3:28] Um, so these were our hands-on keyboard people.  Uh, these were builders. Um, like I said, they [3:33] were data scientist stuff. Uh, they would not shut  up about AI. They're like, "Oh my god, this stuff [3:38] is so cool. It can generate that and this and the  other." Uh my recommendation is certainly in 2023 [3:43] you did not want to get cornered at a party by  these people. Uh it's probably still true. Many [3:47] of them are here. Uh so I'll be careful about  that. Yeah. So avoid the party tomorrow night. [3:53] Um so this was the logo for our practitioner  meetup. It was about 30 people to start with. [3:57] We just met on a monthly basis and it was just  like hey what's a cool thing that's going on and [4:02] and uh here's a problem that I'm working on. How  would you guys solve it? It was meant for hands-on [4:06] keyboards practitioners to talk to other hands-  on keyboards practitioners to meet each other. [4:11] All right, [snorts] so now we're in about the  middle of 2023 and this sort of excitement around [4:17] use of AI started to go beyond the practitioner  level, right? So people uh at home were using [4:22] chat GBT. I don't think maybe perplexity had come  out, but there weren't many other tools available. [4:27] Um, and what we started to find is that there were  people who were non-technical but were using AI to [4:32] plan vacations and do health plans and all kinds  of things like that. And the question was, how [4:37] do I engage these people in bringing their sort  of passion for AI into the workplace? And so we [4:44] built the second learning community that's called  enthusiasts, right? These are people who just uh [4:49] um would, you know, were learning how to use  it for vacation planning or whatever it is. [4:54] and they didn't know how to say, "Well, I use it  for probably 5 hours a week to 20 hours a week at [4:59] home, but I don't know what I'm allowed to do  at work. I don't know what I'm supposed to do. [5:02] I don't know what I'm allowed to upload and not  upload." And so, as opposed to them just saying, [5:07] "No, don't use it at work." What we said  is, "Come be part of this monthly community. [5:11] We're going to teach you the foundations of how AI  works. We're going to show you which tools you're [5:15] allowed to use, which ones you're not allowed  to use." But we were not the AI police. So, [5:18] people can just come ask us dumb questions. So  the way the meetup worked is that there were [5:23] uh I think let's see it's on this slide. Yeah  30 minutes each month of like prepared content [5:28] where we had somebody leading an AI something who  would come and teach them about how to use it uh [5:33] and how we were using it in a corporate and  professional environment and then we'd turn [5:36] the recording off and then we had a safe place for  people to ask questions. Oh, I used this thing. [5:41] I uploaded this thing and we go you shouldn't  do that. Uh do it this way instead. Right? Uh [5:46] and so uh we recorded it and this group has since  grown to kind of a lot. And so by 2023 we had our [5:52] the late 2023 we had two learning groups. One for  hands-on keyboards practitioners and the second [5:58] for set of slightly less technical enthusiasts who  were learning how to apply their excitement around [6:02] AI at work. Um and then what we found is that  the way to get more people to understand was to [6:10] create learning communities. Right? I don't know  how many people in the room are in HR or work in [6:14] training or personal development, but  the real problem with H with training AI [6:18] is that by the time you finish the training, it is  outdated, right? Like it's literally changes every [6:23] month and a half. And so the only way to sort  of keep uh sort of your entire population kind [6:30] of educated is to sort of give them peers to learn  from on a regular basis and a low stakes and sort [6:36] of high temporal pace uh environment. So learning  communities was our answer to how we got there. [6:42] Um and so now we're talking early 2024. I  can't I don't remember how many enthusiasts [6:47] had grown to about 200 people, about 80 people  in our practitioners group. And then I realized [6:51] there was a need for a third group. Um and the  third group was these people that I called uh [6:59] pioneers. And these were the non-technical crazy  people. Like I I remember distinctly I had a [7:06] my first pioneers call. We were talking about this  survey that said something like 40% of active GPT [7:13] users on a weekly basis trusted the answers of  a GPT more than any human being in their life. [7:19] And I thought this was the most insane thing I'd  ever heard. And half of the people in my pioneers [7:23] group were like, "That's me." Uh, and I didn't  understand why, but then they told me why, and [7:28] then I understood why. Um, and I don't have enough  time to tell you, but it's really remarkable. [7:33] Um anyway, so at the end of the day, what we  realized is that learning and awareness and [7:40] adoption of AI tools was happening all the time  inside and outside of work. And what I needed was [7:45] a way to talk to the social leaders who were the  not the engineers, right? So there are project [7:51] managers, data analysts, call center managers,  all kinds of things like that. And what happens is [7:55] they were the local crazy person. So in any given  department, they would go to that person and say, [7:59] "Hey, I know you're like the AI person. and how  do you use this or show me how to use this? And [8:03] what I wanted to be able to do was do two things.  One, connect these early leaders to help them [8:09] understand that they were the most early adopters  going on in the company. Like they may know it, [8:14] but then they were recognized and then uh by me  kind of saying, "Hey, you're a pioneer. You should [8:19] become part of this community." They sort of put  this badge on and goes, "Wow, gosh, I am a thought [8:24] leader inside of this company and now I represent  the right way to do this inside the company." So [8:29] I would work with them on a monthly basis and we  talk about things like you know uh you know how [8:34] do you answer this question or somebody is scared  about AI stealing their information or whatever [8:38] and I wanted to get them grounded in facts and so  that they could share good information back to the [8:43] people that were around. So to me the pioneers  was basically like this shadow uh influencer [8:49] network that I built so that people could uh get  the right information and the right information [8:53] would be dispersed across lots of communities.  uh what ended up happening over time is that [8:57] each of these pioneers ended up building  sort of formal training programs in all of [9:01] their departments. And so there were like six or  seven different training programs that came from [9:04] these different pioneers because they then felt  empowered that they had the right to do it. Um [9:10] this uh this actually reminds me there's a very  good TED talk from about 15 years ago and I should [9:17] shouldn't bring it up on stage and I forget the  guy's name. Uh Ed Severs Tim Severs or something [9:21] like that but it's called how to start a movement.  Uh, and if you're thinking about how do you get [9:26] lots of people to go do something, I really  recommend this thing. It's about uh, it basically [9:30] tells the story of how you go from being a maniac  dancing alone in a field to starting a movement. [9:36] Uh, Derek Severs, I think, is his name. It's  super good. So, anyway, these pioneers were my [9:41] maniacs dancing out in the field. Uh, and so we  would get together on a monthly basis and talk [9:45] about these kinds of things. Uh but broadly  speaking it was about taking non-technical [9:52] leaders not engineers not programmers uh not  data scientists and giving them a place where [9:57] we could talk about technology so they get smart  about how it's actually working. Uh there was a [10:01] lot of separating truth from fiction of things  that they had heard and then I could me and my [10:05] engineering team could could have correct their  mental models for how these AI things worked. [10:10] So now we're kind of middle to late 2024. We've  got three different communities. Um, I don't know, [10:14] we're up to maybe, I don't know, 500ish people  involved in these monthly learning communities. [10:22] Uh, but what became clear is that just teaching  people wasn't enough. What we needed to do was to [10:26] give them a safe environment. And so what my team  was doing in this time period was we were building [10:31] uh several here. I've listed six different  official tools that they were allowed to use. [10:37] Uh, I had a bit of a luxury that I work for  a privately held company and so we had extra [10:40] capital. So we we built an on-prem GPU system.  We used a bunch of openw weight models. We [10:45] uh used open source solutions and we had a  private GPT. We had a private chat with my [10:49] docs. We had a private deep research tool.  Um if anybody's heard of N8, this is like a [10:55] a almost open-source like automated  workflow system. We installed that inside, [11:00] you know, securely connected it to 60 different  systems so that it could talk to Outlook, [11:04] it could talk to Atlassian, it could talk to  Oracle, it could do all these things, but we [11:08] weren't outsourcing things to GPT. We weren't  outsourcing things to third parties. And so all of [11:12] the data and all the conversations were happening  behind our firewall privately and securely. So you [11:16] could have conversations around paychecks and  customer informations and PII and all kinds of [11:20] things like that and not have to worry about it.  So this was a huge strategic advantage for us. [11:24] I know most people do not have this advantage,  but it was something that we're proud of. But [11:28] I think the the lesson I would take home for  people who don't have this is the key is if [11:32] you want more people to start using AI, you have  to tell them what tools they're allowed to use [11:36] and not and not just say please don't do anything  and definitely don't put you know work information [11:41] in GPT which you shouldn't do but you got to tell  them what things they're allowed to use. Okay. So [11:46] also during this time because we had launched all  these tools and we've got these three communities [11:51] what became clear is that the sophistication  level that was going on in the pioneers group um [11:55] needed to be split in the in the in the  in the uh uh the pioneers group and the [12:03] the practitioners group is that we actually needed  to sort of split practitioners into three levels. [12:08] And so I added an experts group and that's  uh every I can tell you if you raise a room [12:13] and you're like hey do you want to be  part of the experts group everybody [12:16] wants to be part of the expert group.  So they need to be very like handpicked [12:19] and the way that I positioned is it was not  actually the technical experts. The experts were [12:23] actually the people who were driving technical and  architectural and policy and compliance decisions [12:28] across the different parts of the company. And  it ended up becoming basically a strategy team [12:33] where different leader technical leaders not  the sort of boss not the VPs necessarily but the [12:40] like our senior architects and the thought  leaders among different departments would get [12:44] together and talk about what they were trying to  do. And this experts group was more about what is [12:48] our companywide enterprise AI strategy.  And it allowed different organizations [12:53] to have strategic conversations about how to  deploy AI in a safe and private environment. [12:57] And then also these specialty groups started to  showing up. Um I actually don't manage any of [13:01] these. What these were were community groups  around glean and nin and deep research and [13:06] chat with my docs where power users and fusers  groups started to meet on a monthly basis to [13:10] talk about what they were doing. Um and so these  specialty groups were tool specific. Okay. So uh [13:17] I think I already covered this but basically the  experts group became an AI strategy group more so [13:22] uh than actually a group of experts. Uh  and then specialty group started showing [13:26] up around each of the platforms and tools that  we gave them. So this this is if there's a cheap [13:31] win, this is the easiest win you could do. For  every tool that you have AI tool that's been [13:35] deployed in your company, you need to build a  user group so that they can get together uh and [13:40] let them do it organically. Uh, and I think that  the last thing is make sure there's a safe place [13:44] for people who are doing things that probably  are against policy to talk about those things [13:49] offline, right? So, stop the recording  and let them have private conversations. [13:53] Okay, so now we're talking late 2025. Uh, we've  got five different layers of people. I think you [13:58] add this up, we're at about 1,300. Uh, and first  off, I'm pretty pumped about this. Like the idea [14:06] that we could go from 20 people who knew what  they were doing to more than that of any level [14:11] was remarkable. And now we've literally multiplied  that times 500 uh and is that right? Uh 50. And um [14:19] and we've gone from uh a population so Cox  uh communications has about 28,000 employees [14:25] of which about 8,000 are sort of corporate office  people. And so when I think about our population [14:30] of corporate office people at this point I'm up to  about uh 5% something like that of people who um [14:37] I shouldn't do math on stage. This is another  reason why my Asian parents are disappointed in [14:41] me. Uh but the anyway we're talking about,300  out of 8,000 have some level of AI fluency at [14:48] this point and we're talking sort of uh  the middle of summer of last year. So [14:52] um and the enthusiast group, this orange  block at the bottom had gotten gigantic, [14:56] like 800 people in it. Um and so there was  something that was interesting because the [15:00] enthusiast group had now been meeting for 2 years  and the people who had joined this working group [15:04] for 2 years are now like pretty smart about it.  They are no longer beginners. And what we found [15:09] is as the group getting kept getting bigger,  there was a whole group of people and this uh [15:16] close captioning is covering it uh of foundations  at the bottom. We realized that there needed to be [15:21] one more group and this is it. I promise  it's the last group that we built. Um, [15:25] and it's the foundations group. And the  foundations group is not a long-term group. It's a [15:31] fourmonth program for people who are just starting  their AI learning journey. And it gives them the [15:35] chance to come together and have a speaker talk  about four topics over the course of four months. [15:40] So each of these is one month meeting. Basically  40 minutes of content, 20 minutes of Q&A, [15:45] foundations of AI and LLMs, how to prompt and  understanding context and memory. Number three [15:50] is understanding agents and loops. And number four  is the near and far future like helping them get a [15:54] mental understanding of probably what is going  to happen to the future work. How are things [15:58] going to be affected? And then as a human being  who needs to who wants to learn more about AI, [16:02] how do you do that? And those are the sort of  the four foundation classes that are going on. [16:06] So here we are in the middle of 2026. um probably  up to like 2,300 people that come to all of these [16:12] meetings at this point. I'm incredibly proud of it  because I just wasn't sure it was going to work. [16:17] U but it was built really on this hypothesis that  in order for any organization at any scale, your [16:23] team, your company, your family, your school, your  country to survive, you need to have the highest [16:29] percentage of AI fluency as possible. Right?  And we are now at about 25 30% of our people [16:35] have some level of AI fluency. And I think that  means they're going to make it. Like whether the [16:41] company makes it or the people make it, the we're  going to make it. Um it's been very very very very [16:47] fulfilling for me because across all of these  communities over the last 3 years, I have like [16:52] literally people walk up to me and go, "Holy cow,  you changed my life because now I can do all these [16:56] things that I didn't know I could do before." And  it wasn't just me. I like there's a there because [17:00] of the way this pyramid works. As people move  from enthusiasts to pioneers to practitioners, [17:06] then they have then you know each one teach one  they've brought other people in their departments [17:09] along. And so what happens is that you are able  to get scale by creating communities that then [17:15] allow the community itself to grow.  So I don't need to be the one uh [17:19] teaching people because I've taught others and  then they'll teach people down below. The same [17:22] philosophy is how the whole department works which  is look if we have an AI department of 35 people [17:27] we cannot do all of the projects and do all of  the things that you should be doing with AI. [17:31] What you need to do is empower as many people  as possible, give them all the tools, give them [17:35] the foundational knowledge to get stuff done and  then give them a safe environment in which they [17:39] can experiment. And that is what has happened.  So over the last 3 years, we now have you know [17:45] what's this uh 1500 plus community members  uh 2,000 people using our internal tools on [17:51] a regular basis. Uh I'm really proud of this over  2500 weekly users of Claude and Claude Code. Um, [17:58] we have that in8 platform, this automated workflow  system that we talked about. There are 800 plus [18:03] 900 plus people that use it. They've built  5,000 applications and this, you know, sleepy [18:10] cable company that probably most of you have never  heard of. Uh we have an agentic platform that has [18:15] completed you know 2.3 million probably this  week we're up to 2.5 million it's about 60,000 [18:21] executions a week fully agentic workflows running  written by 900 different people running in dozens [18:26] of different departments I don't even know what  they all do but I do know that they're safe and [18:30] they're monitored and they're controlled uh and so  that's what we're doing so we have done that over [18:34] the course of the last three years and I think  uh we're actually going through a merger so it's [18:38] unclear to me where this goes in the future um  but I firmly believe this will work most places. [18:45] And the key is like how do you connect people who  are at similar stages in their learning journey [18:50] so that they can learn from each other and then  keep giving space for people who move through the [18:55] next ring and get more and more advanced with  how to use and how to apply AI in a safe way. [19:00] Allow them to teach other people and bring people  I so that this this triangle we've talked about [19:05] I've always thought about as kind of a circus tent  that starts really small and the course is over [19:09] time you keep lifting the circus tent. you keep  learning search that the tent keeps getting bigger [19:13] and more people start becoming part of the tent.  Uh and so that's the mental model that I use. [19:18] So if you want to try to do this at your company  or your organization, I would first start with [19:22] build communities of people of similar learning  stages. Cultivate the leaders who are teaching [19:27] in each department so that you know uh the story  that is told. You give them retellable stories [19:32] and you know that your your grounded facts  become the ones that are repeated uh through [19:37] throughout the organization. and then providing  tooling and guidance for other people to work.