OpenAI, Chip Ganassi Test Track AI as Data Wars Reshape Racing

数据战重塑赛车运动：OpenAI携手Chip Ganassi测试赛道AI 人工智能的普及将削弱大型车队在软件技术上的既有优势，使人类车手的直觉与快速提示词工程成为决胜的关键因素。 本期内容深入探讨了如何利用人工智能和Codex等大语言模型，将高带宽时序遥测数据、非结构化车手定性记录以及传统Excel表格整合统一。嘉宾Joyce Ruffell与Chase Holden详细说明了AI如何通过扩充工程处理能力、优化物流及加速假设验证，成为小型车队的平衡器。他们还探讨了OpenAI内部深厚的赛车文化、竞技赛车相比自主机器人更趋向以人为本的未来，以及随着AI工具在NASCAR和IndyCar中普及数据分析，行业专业知识如何成为最核心的竞争优势。 赛车运动目前正陷入激烈的“数据战”阶段，能够以最快速度整理、解析海量赛道数据并付诸行动的车队将占据获胜优势。 (0:00) OpenAI与Chip Ganassi Racing正展开合作，使IndyCar数十年积累的历史遥测数据实现即时可查，从而为赛车工程师加速假设推导与模拟循环。 (0:43) Chase Holden在没有计算机科学专业背景的情况下，利用ChatGPT和OpenAI Codex完成了从NASCAR播客主播到创办RaceTek Systems的转型。 (2:43) AI模型通过将长达数小时的技术日志消化为通俗易懂的汇报，并协助工程师构建精准的提示词，从而弥合了车手与工程师之间的沟通鸿沟。 (10:00) 通过摄取定性记录并将其与ECU定量遥测数据结合，AI成功将“软性”的人类反馈与以往受困于传统电子表格中的“硬”数据融为一体。 (15:31)

OpenAI researcher Joyce Ruffell and RaceTek Systems co-founder Chase Holden argue that modern motorsport has become a data war where artificial intelligence provides a competitive edge by synthesizing dense telemetry with qualitative driver feedback and accelerating strategic decisions for lean engineering teams. Integrating Telemetry with Legacy Spreadsheets (15:31) Motorsports organizations routinely generate vast streams of high-frequency time-series telemetry alongside qualitative, unstructured notes from drivers, tire specialists, and pit crews. Ruffell observes that race shops historically run on massive, multi-tab Excel workbooks filled with custom macros and intricate formulas, creating isolated data silos between individual engineers and mechanics. Standardizing this information requires format conversions because complex raw spreadsheets do not natively map into modern machine learning architectures. By using language models to ingest and unify quantitative chassis metrics with unstructured subjective notes—such as a driver reporting that the rear end feels loose on corner entry—teams can query decades of historical race archives. According to Ruffell, this unified searchability accelerates the hypothesis loop for active programs like Chip Ganassi Racing's IndyCar and INDY NXT teams, allowing engineers to simulate setup scenarios and maximize the value of limited on-track practice sessions. Virtual Engineering Support for Lean Teams (19:17) Significant budget and personnel disparities exist across NASCAR and IndyCar, where premier outfits deploy specialized engineering squads to monitor individual sub-systems while lean organizations operate with smaller staffs. Ruffell notes that Chip Ganassi Racing employs fewer than 200 total personnel, requiring team members to handle multiple roles under tight schedules. Holden and Ruffell explain that AI tools act as integrative connective tissue across departmental silos, enabling smaller outfits to replicate the analytical output of larger competitors without adding headcount. Beyond track telemetry, race organizations apply AI to manage equipment logistics, supply chains, and component wear across global travel schedules. Automating repetitive data consolidation frees technical personnel to explore deprioritized edge cases that can unlock critical lap-time advantages. For software development, Holden reports that utilizing coding agents allows solo builders and small teams to turn garage feedback into working prototypes within a single week. Translating Driver Feedback and Prompting Precision (10:00) A recurring operational friction in motorsport is the language gap between drivers, who experience car behavior through sensory intuition, and race engineers, who require quantifiable technical data. Language models help reconcile this divide by processing technical logs and practice reports to generate concise debriefs for drivers, while simultaneously translating driver commentary into structured inputs for the timing stand. Ruffell emphasizes that effective adoption requires demystifying artificial intelligence so engineers treat models as exact tools rather than black boxes. Holden highlights that extracting reliable strategic recommendations depends on prompt structure, deliberate intent, and sustained multi-turn context. Development workflows have similarly shifted from manually copying code snippets into web interfaces to utilizing embedded assistants powered by OpenAI Codex directly within developer environments. Preserving Human Judgment in High-Speed Decisions (36:17) Holden and Ruffell dismiss the idea that autonomous vehicles or robotics will replace human racing drivers, distinguishing experimental self-driving exhibitions like those in Abu Dhabi from traditional series. Spectators follow motorsports primarily to watch human skill, physical grit, and competitive execution under pressure, ensuring drivers remain central to the sport. As advanced AI models and coding agents become universally accessible, traditional analytical software moats will erode across the paddock. The speakers argue that when technological access equalizes, competitive advantage shifts back to human intuition, operational taste, and prompt efficiency. On the pit wall, where an analytical delay of ten seconds can alter a race result, execution speed and human judgment remain the decisive performance factors. While AI lowers the barrier to software development and streamlines data synthesis for racing teams, its strategic value is constrained by the need for real-time precision and human oversight. Ruffell and Holden conclude that universal access to AI will eliminate traditional technological advantages, leaving domain expertise, prompt communication, and human intuition as the primary competitive differentiators.

OpenAI 研究员 Joyce Ruffell 与 RaceTek Systems 联合创始人 Chase Holden 指出，现代赛车运动已进入高密度信息对抗的“数据战”阶段，人工智能正通过统一多源数据、降低研发门槛以及优化运营协作重塑车队生态；但随着技术工具逐步普惠，决定赛道胜负的终极优势仍将取决于人类专家的领域直觉、沟通效率与即时决策能力。 整合遥测数据与非结构化人类反馈 (15:31) Joyce Ruffell 表示，传统赛车团队高度依赖庞大且复杂的 Excel 电子表格记录数据，工程师经常需要在多层嵌套公式、自定义宏与条件格式中繁琐检索。尽管既有计算程序能高效处理来自电子控制单元（ECU）的高频数值遥测数据（如车速、油门开度及悬挂行程），但车手的定性主观反馈（例如“在3号弯入弯时车尾发飘”）、轮胎工程师的现场观察以及维修站的手写记录等“软数据”长期处于孤立分散状态。 OpenAI 与拥有35年历史的 Chip Ganassi Racing 展开合作，利用大语言模型将这些非结构化文本与硬件遥测数据融合为统一的查询系统。这一整合旨在打通历史档案，加快工程师的假设验证与模拟测试循环，确保在练习赛、排位赛和正赛中的每一分钟赛道时间都能发挥最大效益。此外，OpenAI 约每三个月对其前沿模型的能力进行一次基准评估，以探寻如何减少人工预处理负担并直接摄取未经格式化的复杂原始表格。 弥合车手与工程团队的沟通壁垒 (10:00) 在车队的日常运转中，赛车手与技术工程师之间普遍存在专业语言障碍。车手通常依据身体直觉感知车辆的机械动态，但往往难以使用严密的工程术语进行表述；而工程师则专注于传感器指标，容易忽视驾驶员的感官细节。Ruffell 与 Holden 指出，AI 工具能够有效充当两者的转译中介，将长达 4 至 5 小时的密集练习赛技术报告与工程日志快速浓缩为清晰简要的总结，帮助车手在重新驶入赛道前迅速把握核心调校信息。 与此同时，车队正在加强针对技术人员的提示词工程（Prompt Engineering）培训。Ruffell 强调，向工程师阐明模型底层逻辑有助于消除对 AI 系统的“黑箱”或“魔法”误解，使技术人员能将其视作精密仪器，通过构建高质量、目标明确的提示词，在计时台上快速提取高度精确且具备可操作性的调校方案与战略决策指令。 消除预算劣势与加速策略创新迭代 (18:26) Holden 与 Ruffell 观察到，传统赛车竞赛的优势往往与资金规模直接挂钩，顶级 NASCAR 或 IndyCar 车队能在维修区配备庞大的专属工程团队分别监控各个子系统，而中小型车队受限于预算无法维持同等编制。例如 Chip Ganassi Racing 作为一个精简机构，总员工通常不足 200 人，成员必须在高压赛程下同时承担多重职责。AI 充当了“虚拟工程支持”的角色，使小团队能够自动化捕捉与整理次要细节及边缘案例，从而以极少的人力达到接近行业巨头的分析产出。 更深层次的影响在于研发成本的骤降。以往提出一项新设想需要召集整个专家团队进行可行性论证，而现在领域专家能够先借助 AI 工具独立完成概念验证与方案推演。Ruffell 认为，这种模式使得赛车竞技由单纯依赖资本投入的研发周期，逐步转向由创意与领域专长驱动的创新周期，大幅缩短了战术调整从构想到落地的时间跨度。 全流程工程代理与赛道外运营自动化 (39:26) Holden 介绍了 RaceTek Systems 如何将 AI 深度融入软件开发与赛车商业运营的各个环节。在研发流程上，团队已从早期在网页端向 ChatGPT-4o 手动复制粘贴代码，演进为在集成开发环境（IDE）中直接调用基于 OpenAI Codex 的代码生成代理。在结束比赛周末并收集实地反馈后，Holden 能在一周内利用代码代理快速搭建并测试下一站赛事的原型软件，但他也强调人工审查仍是修复软件缺陷与确保执行精度的必要最后防线。 在赛道之外，AI 承担了广泛的行政与物流管理工作。从为车手规划赞助外联公文、设计媒体内容方案，到跨区域统筹运输日程与追踪比赛中损耗、折断的零部件库存，自动化工具显著降低了运营间接成本。Holden 曾凭借该工作流在北卡罗来纳州成功签约首个商业客户——征战 NASCAR O'Reilly Auto Parts 系列赛的 SS-GreenLight Racing / BRK Racing 车队及其车手 Garrett Smithley。 技术普惠背景下人类判断的核心地位 (36:17) 针对自动化系统是否会取代人类车手的讨论，Joyce 与 Chase 一致认为赛车本质上是一项以人类竞技、胆识和高压表现为核心的运动；尽管阿布扎比等地已开展无人驾驶开轮赛车测试，但这些项目目前属于工程展览性质，无法取代主流的人类职业赛事。赛车团队在应用 AI 时并不寻求宽泛的指引，而是要求在极短时间内提供确定、高精度的具体输出，这也倒逼模型开发朝着更高响应速度的方向演进。 两位受访者预测，随着技术向通用人工智能（AGI）演进且高级分析工具全面普及，各车队之间的传统技术壁垒将被逐步抹平。届时竞争重心将重新回归人类特质，由车队的直觉判断、专业审美以及提示词沟通效率决定优劣。Holden 强调，在维修区战术执行中，仅 10 秒钟的数据分析延迟就可能导致整场比赛失利，最终的胜负分水岭仍将取决于人类决策者构建端到端快速响应工作流的能力。 该节目表明，AI 在高水平赛车领域的价值在于辅助人类专家快速推演与整合杂乱数据，而非替代车手或工程师的主观能动性。不过，对话也反映出当前技术落地的现实局限：传统密集的复杂电子表格仍需针对 AI 进行格式转换与适配，代码代理生成的方案必须依赖严格的人工调试，且在争分夺秒的维修站极端高压场景下，算法输出的延迟与准确度仍面临严苛检验。

## Background 本期内容由 OpenAI 官方频道于 2026 年 8 月 3 日发布在 YouTube 平台，为播客节目《The OpenAI Podcast》第 22 期。根据官方简介，赛车是一项由微小差距与海量数据构筑的运动，OpenAI 研究员 Joyce Ruffell 与 RaceTek Systems 联合创始人 Chase Holden 在节目中深入探讨了如何利用 ChatGPT 与 Codex 激活赛道、车间及车手背后的数据潜能，并介绍了 OpenAI 与 Chip Ganassi Racing 的科研合作、AI 对中小车队的赋能以及 OpenAI 内部深厚的汽车文化。 ## TL;DR 节目围绕赛车运动正在打响的“数据之战”展开，探讨了 AI 如何替代繁重庞杂的电子表格工作流，成为连接赛车硬核遥测与车手主观感受的智能纽带。Joyce Ruffell 详述了 OpenAI 与传奇车队 Chip Ganassi Racing 合作处理高带宽时间序列数据的实战经验，而 Chase Holden 则分享了自己借助 Codex 从播客主播跨界创立赛车智能系统、帮助中小车队抹平技术资源壁垒的创业故事。讨论进一步揭示了赛车调校与 AI 研究在追求边际收益上的精神同构，并明确指出无论自动驾驶如何演进，赛车运动的核心始终是“以人为本”，在 AGI 时代人类的直觉、品味与即时决断力仍将是制胜的终极壁垒。 ## Atmosphere & style 整场对话洋溢着硬核工程理性与赛道竞速激情的奇妙碰撞，主持人与嘉宾之间配合默契，叙事既有前沿技术探讨的深度，又不乏生动接地气的幽默感。从吐槽无限滚动的巨型 Excel 迷宫，到回忆五岁时在塔拉迪加看台被引擎轰鸣震撼的童年瞬间，对话的情绪弧线真挚而炽热。聆听过程仿佛带领观众在代码闪烁的 IDE 界面与油压轰鸣的维修区指挥台之间穿梭，极富现场沉浸感与感染力。 ## Bulletin - [00:43](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=43s) OpenAI 与 35 年传奇车队 Chip Ganassi Racing 展开深度科研合作，利用前沿大模型处理车载传感器产生的严苛高带宽时间序列数据，为维修区指挥台提供深度赛前决策推演。 - [02:43](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=163s) RaceTek 联合创始人 Chase Holden 凭借 ChatGPT 与 Codex 等工具，从 NASCAR 播客主播成功跨界创立赛车智能公司，为中小车队搭建等效于“虚拟工程师天团”的辅助系统以抹平资源劣势。 - [15:31](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=931s) 赛车界正深陷“数据之战”，AI 不仅用于打通孤立繁杂的 Excel 工作流，还将车手的主观体感、工程师手写笔记与客观遥测数据融合，帮助车手将数小时的练习赛报告提炼为即时洞见。 - [33:01](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=1981s) AI 在赛车领域的真正价值并非打造无人驾驶替代人类，而是重塑成本结构并赋予一线工程专家自主构建工具的能力，在 AGI 时代将竞技胜负手重新交还给人类直觉与临场决断力。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=0s) 现代赛车运动已进入“数据之争”（data wars）时代，毫秒级的胜负取决于团队能否高效梳理整合数据，实现人、机器与工程团队的高度同频共振。 - [00:43](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=43s) 车载传感器产生的高带宽时间序列数据极度严苛，OpenAI 针对 Chip Ganassi Racing 拥有的数十年历史数据，聚焦解决底层数据清洗、检索效率与实时战术推演闭环。 - [02:43](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=163s) RaceTek 定位于赛车智能技术，通过聚合比赛周末维修区与车间内零散隐蔽的上下文数据，为技师和策略组提供统一的全景决策视图。 - [03:32](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=212s) 生成式 AI 和 Codex 工具降低了软件开发与系统构建的门槛，使得非传统计算机科班的从业者也能快速将创新赛车设想转化为可交付的实战软件。 - [11:32](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=692s) 赛车工程引入 AI 的关键在于“通识教育”与“去魅化”，帮助硬核工程师理解输入决定输出的 Prompt 工程机制，将应用重心放在赛前的严密战备推演。 - [13:20](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=800s) AI 能够打破工程团队与车手之间的沟通壁垒，将过去耗时四到五个小时的晦涩工程报告瞬间重构为直观、易懂的高价值操控摘要。 - [15:31](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=931s) 赛车行业对 Excel 依赖极深，AI 正在扮演粘合桥梁，将车手无线电反馈等主观非结构化“软性数据”与 ECU 遥测硬核数字深度融合。 - [18:08](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=1088s) AI 相当于为预算有限的中小型车队配备了“虚拟工程师天团”，大幅降低边际成本，使其能够拾起常被遗漏的细节数据并挑战传统行业巨无霸。 - [22:50](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=1370s) 将赛车数据重构为“AI 原生格式”（AI-native format）虽需前期投入，但能打破车队内部各工程师专属表格形成的壁垒，让全员步入同一战术认知频道。 - [31:36](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=1896s) OpenAI 内部盛行赛车文化的本质原因在于：赛车与大模型科研本质同构，都是将极其复杂的系统工程极致浓缩为单一指标（圈速对比测试损失）的微小边际收益追求。 - [33:01](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=1981s) 赛车运动具有不可替代的“以人为本”属性，顶尖车队使用 AI 时保持着极其严苛的精准度标准，AI 的角色是超级协同辅助而非无人化取代人类车手。 - [36:14](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=2174s) 当未来 AGI 和 AI 工具全面普及平民化后，赛道核心竞争力将回归人类本源——即更敏锐的直觉、审美底蕴以及稍纵即逝的临场沟通决策力。 - [38:02](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=2282s) AI 改变了赛车研发的投资与成本结构，使得一线专家无需经过冗长的评估流程，即可自主低成本验证创意原型，极大提升了创意的边际溢价。 - [39:26](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=2366s) AI 技术不仅赋能赛道分析，还深入延伸到车队的赛后代码快速迭代、零部件物流追踪、赞助商商务联络等全局运营流程。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=0s) 赛车运动进入“数据之争”时代 主持人 Andrew Mayne 开启了本期播客，引出这一期探讨赛车与人工智能交汇的核心主题。赛车本质上是一项由毫米级差距和海量数据构筑的极致运动，每一次过弯、换挡和策略博弈都依赖背后庞大的信息支撑。在开篇中，嘉宾们直言不讳地指出，当前的赛车运动已经全面迈入所谓的“数据之争”（data wars）。在这场较量中，决定胜负的不仅是引擎的马力，更是人与机器之间如何实现天衣无缝的协同——谁能以最高效的方式梳理和整合资源，让工程团队、车手与赛车同频共振，谁就能在格网起点占据绝对优势。 - 本期对话围绕 OpenAI 研究员 Joyce Ruffell 与 RaceTek Systems 联合创始人 Chase Holden 展开。 - 讨论聚焦于如何运用 ChatGPT 和 Codex 等 AI 工具，推动赛车运动从赛道物流、赛车调校到车手表现的全方位进化。 ### [00:43](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=43s) 打造赛道数据引擎：OpenAI 与 Chip Ganassi Racing 的合作 Joyce Ruffell 详细介绍了自己在 OpenAI 负责的研究工作，过去几年她一直专注于解决最底层、最核心的数据难题，涵盖数据来源、数据清洗质量、数据流水线以及数据处理效率等领域。正是基于这种对数据基础设施的深耕，她牵头推进了 OpenAI 与传奇车队 Chip Ganassi Racing（简称 CGR）的前沿研究合作。该车队目前的核心重心在 IndyCar（印第赛车）和 INDY NXT 赛事上，历史上也曾征战 NASCAR 和各类跑车耐力赛，在长达 35 年的建队史中积攒了难以估量的深厚底蕴。 Joyce 指出，赛车数据具有极度罕见且苛刻的特性：它是由密密麻麻的车载传感器产生的“高带宽时间序列数据”（high-bandwidth time-series data）。车队不仅需要在赛后进行详尽的回溯分析，更需要在比赛周末现场实时处理这些数据流，争分夺秒地做出战术判定。虽然 Chip Ganassi 拥有跨越数十年、涉及多名车手、覆盖全美各大赛道与不同赛车规格的海量专属知识，但真正的瓶颈在于“可访问性与检索效率”。工程团队需要能够闪电般地提出假设、在历史数据中验证假设，并据此在每次练习赛、排位赛乃至正赛前快速微调赛车设置。数据反馈的闭环越短，车队能进行的实验就越多，从而把每一次出站练习都转化为通往正赛胜利的有效积累。 ```mermaid flowchart LR accTitle: 赛车数据处理与迭代闭环 accDescr: 展示从赛道传感器高带宽数据采集，到 AI 检索分析，再到策略假设验证与赛车微调的闭环流程。 subgraph S1["数据采集与沉淀"] A["车载高带宽时间序列数据"] B["35年历史数据库<br/>车手/赛道/赛车规格"] end subgraph S2["AI 智能分析引擎"] C["实时数据清洗与流处理"] D["快速假设验证与模式对比"] end subgraph S3["赛道实战输出"] E["排位/练习赛参数微调"] F["正赛战术策略决策"] end A --> C B --> D C --> D D --> E E --> F F -.反馈积累.-> B ``` ### [02:43](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=163s) RaceTek Systems：为车队打造赛车智能系统 随后，Chase Holden 分享了自己创立 RaceTek Systems 的业务愿景。RaceTek 的定位是一家“赛车智能”（racing intelligence）技术公司，目前主要为各级车队定制专门的软件工具与智能系统。很多时候在紧张繁忙的比赛周里，车间与维修区会有大量细碎、隐蔽的信息被忽略或遗漏，而 RaceTek 当前正在研发的一站式智能平台，就是为了将这些零散的数据与线索汇聚一处，确保整个团队能在统一的视图下工作。 Chase 强调，这套智能工具并不是要凭空替人类做决定，而是给车队技师、工程师和策略组提供一副“更聪明的透镜”。通过实时整合与深度提炼赛道上下文，帮助车队在分秒必争的比赛周末看清全局，做出更高质量、更具确定性的临场决断。 ### [03:32](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=212s) 从塔拉迪加看台到播客主持，再到 AI 驱动的创业 回顾自己走进赛车世界的奇特历程，Chase 坦言自己的经历完全不同于传统圈内人。他在路易斯安那州南部长大，当地其实并没有什么现成的赛车场环境，但他父母的一群好友是铁杆车迷，热衷于结伴前往阿拉巴马州的塔拉迪加超级赛道（Talladega Superspeedway）。在那座以狂野奔放、传统民风著称的赛道圣地，年仅 5 岁的 Chase 站在著名的塔拉迪加大道旁，亲眼见证轰鸣呼啸的重型房车（stock cars）从眼前掠过。那份结合了家庭温情、极致速度与舍身拼搏的震撼感，瞬间在他幼小的心灵中埋下了对赛车的终身痴迷。 然而，长大后的 Chase 最初并没有直接投身赛车行业，而是在金融和保险界从事一份朝九晚五的普通工作。直到 2017 年左右，他内心深处的一股渴望被唤醒，发觉按部就班的生活无法实现人生抱负，于是毅然转向自己从小钟爱的电台与播客领域。他白手起家创办了一档专注探讨 NASCAR 比赛与体育博彩的播客节目，逐步积累了极高的专业度和忠实社群。到了 2022 年前后，播客办得风生水起，而以 ChatGPT 为代表的新一代 AI 技术横空出世，彻底击中了他从小的另一个狂热爱好——科幻与人工智能。儿时的 Chase 曾一遍又一遍地循环观看电影《机器管家》（Bicentennial Man）和《人工智能》（A.I.），对未来充满遐想。 面对大语言模型和 Codex 等代码工具的爆发，Chase 敏锐地意识到人生的质变点来了。过去，作为没有传统计算机工程背景或机械制造资本的局外人，他纵使对改进赛车系统有无数宏大构想，在向别人倾诉时也往往只换来“这太异想天开、太不切实际”的冷水。但现在，借助大模型的多轮交互与代码生成能力，他能够独立把头脑中的蓝图亲手构建出来，直接交付给赛车团队验证。 - 传统赛车从业路径往往是从竞技一线退役走向媒体解说，而 Chase 走通了一条罕见的“从媒体发声切回一线竞技研发”的逆向路径。 - 他特别关注那些资金与技术资源相对匮乏的中小型车队，希望利用轻量、智能的 AI 软件抹平技术鸿沟，为赛车这项给予他无数热爱的运动带来全新血液。 ### [07:45](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=465s) 破圈连接：从车迷日常到深入印第安纳波利斯车队车间 话题转回到 Joyce 身上。Andrew 提到自己认识 Joyce 已经很久，却直到一年前才惊讶地发现她竟然是个狂热的车迷。Joyce 笑着回应，自己其实从未刻意隐瞒过这一爱好——她甚至在 OpenAI 内部创办并主持了公司的汽车兴趣社群。她坦言，在加入 OpenAI 之前自己就曾在汽车产业工作，始终对汽车身上那种机械工程精密性、工业美学与大众情感纽带的融合深深着迷。 真正让她彻底“深陷赛车泥潭”的契机是新冠疫情期间。由于突然多出了时间，加上伴侣也极有兴致一起下赛道，Joyce 开始高频次地奔赴各大赛车场，广泛关注从跑车耐力赛、印第赛车到 NASCAR 的各类组别，深度爱上了围场社区与底层硬核工程学。作为一名在赛车与科技领域同时深耕的女性，她后来前往大部分 IndyCar 车队的大本营——印第安纳波利斯参加行业会议。 令她意想不到的是，会议主办方与当地车队渊源深厚，竟然直接安排他们参访了包括 Chip Ganassi Racing 在内的多家顶级车队车间。Joyce 感慨，赛车围场一向以极度封闭、戒备森严和商业保密著称，哪怕是业内人士也很难获准走入核心厂房，这次经历被她视为“极具启发性的殊荣”。在车间里她惊讶地发现，早在 2024 年底大多数人还在观望时，Chip Ganassi 的工程师们就已经在严肃探索如何将前沿 AI 技术融入赛道工作流。这种对先进生产力的渴求瞬间点燃了 Joyce，促使她积极推动 OpenAI 与 CGR 达成前沿科研合作。 ### [10:00](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=600s) 速度不仅在赛道上：信息组织与决策效率的革命 当 Andrew 询问究竟是在哪一个瞬间让大家真切意识到“引入 AI 会带来实质性竞技优势”时，Chase 从赛队运营深处给出了深刻的见解。他指出，很多车队数十年如一日地遵循固定习惯来管理数据，而真正决定冠军归属的底牌，往往落脚在“谁能将复杂的赛车信息以最高效率组织并提取出来”。 Chase 分析道，赛车比赛涉及极为庞杂的变量网络：特定赛道的沥青抓地力、特定弯角的路肩特性、不同车手的驾驶微习惯以及车队自身软硬件的极限短板。这些碎片化信息如果依靠人工翻找或传统流程整合，往往耗时过长甚至错失良机。谁能够最快、最准确地将所有这些维度拼凑在一起，并在赛道窗口期迅速交付精准洞见，谁就能在漫长艰苦的赛季中看到比赛成绩的爆发式增长。紧接着，Chase 向 Joyce 抛出了一个直击要害的问题：在与车队的深度对接中，AI 究竟在哪些极度精细、看似微不足道的具体环节，做到了人类原本渴望却难以企及的高精度交付？ ### [11:32](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=692s) 揭开模型面纱：赛车工程中的 AI 通识与 Prompt 工程 当被问及有哪些交付给车队的成果看似微不足道、却在赛道实际运作中产生决定性影响时，Joyce Ruffell 直言，最具冲击力的工作其实是“通识教育（Education）”。这种教育的核心是向车队工程师拆解底层模型的工作原理，揭开 AI 的神秘面纱，向他们证明“引擎盖下运转的只是一台机器，根本不是什么玄学魔术”。由于赛车团队本身全由硬核工程师组成，一旦他们理解了机器的底层构造，就更容易构思该如何高效利用它们。这种认知转变最直接的落地形态便是 Prompt 工程（提示词工程）。 Joyce 调侃同场的 Andrew 是“初代资深提示词专家”，并指出模型交互的关键就在于输入：只要给出恰当的输入，就能得到预期的输出，因此责任其实落在使用者的提炼能力上。OpenAI 与 Chip Ganassi Racing 的合作重点放在了比赛日之前的深度整备上。作为赛车界的顶尖车队，Ganassi 的团队一旦走上赛场，对战术调度的把握就极其敏锐。目前双方仍在探索能在策略决策关头提供哪些辅助，但现阶段的大量工作是提前做足功课，确保坐在计时台（Timing Stand）上的策略人员在面对每条赛道时，都能拥有最充分的技术与信息储备。 - **去魅化认知**：将大语言模型拆解为可理解、确定性的计算机械，帮助工程师卸下心智负担。 - **输入决定输出**：指导车队掌握精准的 Prompt 构造技巧，把领域问题转化为模型最擅长处理的输入格式。 - **赛前战备重于赛时干预**：当前的落地重心在于赛前针对特定赛道建立深度的信息推演，为维修区指挥台（Timing Stand）赋能。 --- ### [13:20](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=800s) 打破车手沟通壁垒：将数小时的数据转化为即时洞见 Chase Holden 对教育的重要性深表认同，但他将视角延伸到了座舱内的车手身上。他指出，工程师团队在车辆调校上深耕多年，外人很难去教他们怎么做事；然而在车手端，AI 却能成为打破沟通壁垒的绝佳杠杆。现实中并非每位车手都是工程科班出身，往往难以完全消化技术团队传递的高深参数，有时他们为了不耽误进度只能频频点头说“好，我明白了”，但实际上并未真正吸收。 针对这种痛点，AI 能够将过去车手需要耗费四到五个小时反复钻研的练习赛报告，瞬间压缩并重构为直观、易读且符合直觉的摘要。车手在练习赛结束、准备正式重返赛道去争夺更好名次的关键窗口期，可以立刻抓住核心操控要领。Chase 表示，在 NASCAR 赛场上，一旦工程端的工作步入正轨，他的重心就会立刻转向车手服务。Joyce 也补充道，赛车比赛归根结底是一场信息的竞速：工程师能更快提炼数据，就能更快地将洞见同步给车手，从而在电光石火间做出胜算更高的决策。 ```mermaid flowchart LR accTitle: 车手端数据消化与决策流程 accDescr: 展示原始练习赛数据如何通过 AI 提炼，并在极短时间内赋能车手决策 subgraph DataPrep ["数据收集与提炼"] A["复杂练习赛报告与工程遥测"] --> B["AI 语义提炼与格式重构"] end subgraph DriverAction ["车手吸收与实战"] B --> C["高可读性洞见<br/>(由4-5小时缩短至即时)"] C --> D["赛前快速理解与策略微调"] D --> E["赛道精准操控与完赛圈速提升"] end ``` --- ### [15:31](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=931s) 软硬件数据的融合：连接主观体感与硬核遥测 赛车运动中存在大量的赛道特质与隐性知识，许多宝贵经验常年锁在老车手或技师长的脑海中。针对如何将这些非结构化认知转化为 Codex 或 ChatGPT 可读取的格式，Joyce 指出赛车世界的一个残酷现实：“全世界大部分行业都运行在 Excel 上，而赛车界更是彻底活在 Excel 里。”尽管模型可以解析表格，但电子表格并非 AI 的理想输入。赛车运动有着悠久的“人车合一（Man and Machine）”传统，这种软硬件协同不仅要求 AI 去迁就人类现有的工作习惯，更要求人与模型共同进化。 在实际研发中，最棘手也最具价值的挑战在于处理所谓的“软性界面（Soft Interface）”。传统的数字、遥测数据向来有专属脚本、绘图算法和经典软件工具去清洗与呈现；但诸如车手在无线电中随口描述的“车辆抓地力体感”，或是工程师在各测试轮次（Session）中随手记下的手写笔记，属于极其难以检索与归纳的软性资产。AI 正在发挥桥梁作用，将这些难以结构化的主观文字与 ECU（发动机控制单元）等传感器吐出的绝对硬核数字强力粘合，融合成统一的数据集，供下游的策略推演模型使用。 - **人机共生进化**：超越传统棒垒球（Stick-and-ball Sports）纯靠肉身竞技的逻辑，赛车从诞生起就是人与机器的协同进化，如今 AI 正在成为机器阵营的新变量。 - **软硬数据弥合**：打破传统单一依靠 ECU 遥测数字的局限，将车手主观感受、试车笔记与客观物理量全面对齐。 --- ### [18:08](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=1088s) 抹平资源鸿沟：AI 如何为中小车队赋能与降本增效 在探讨 RaceTek 如何与历史悠久的大型赛车分析机构竞争时，Chase 坦言头部 NASCAR 豪门在指挥台上配备了极其庞大的工程师天团，每位工程师分工极细，各盯一类遥测数据。中小型车队根本无力承担这种人力开销。而 AI 系统的介入，实质上相当于为中小车队配备了一组“虚拟工程师队伍”，将那些平时因人手不足而被遗漏的细节边缘数据捡起来并拼成完整图景，极大地拉近了独立车队与行业巨无霸（Juggernauts）之间的差距。 Joyce 对此高度赞同，并以 Chip Ganassi Racing 为例：尽管战绩辉煌，但 Ganassi 整体规模不足 200 人，每个人都身兼数职，时间是他们唯一的衡量基准。AI 的核心价值在于解放高负荷运转的关键人员： - **释放探索潜能**：通过为工程师单项任务省下时间，让他们能去深入钻研此前因优先级不够而被搁置的技术疑问，而赛道上千分之几秒的微弱优势往往就藏在这些被忽略的细节里。 - **全链路后勤支持**：赛车涉及极为繁重的跨国与全美巡回物流，AI 被用于赛车零部件追踪，统筹每场比赛撞损、掉落或需要周期性更换的昂贵部件。 - **容错率提升**：在 Ganassi 的各个业务部门中，AI 正在承接最具重复性且最容易出现人为差错的环节，帮助身兼数职的个人筑牢防错底线。 --- ### [21:43](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=1303s) 赛车界的“电子表格迷宫”：无穷滚动与图灵完备的 Excel 谈及赛车界对电子表格的依赖程度，Joyce 分享了一个有趣的幕后插曲：OpenAI 内部全员配发 MacBook，而苹果系统在打开巨型 Excel 表格时的表现并不尽如人意，因此她在与车队合作时，往往要求对方提前将数据清洗转码为更友好的格式。但在日常远程屏幕共享时，工程师们调出的表格依然令人叹为观止。 这些工作表不仅数据量极大，常常需要“向左向下一路无休止地滚动”，更充斥着极其复杂的视觉高亮体系、嵌套公式以及单元格之间错综复杂的钩稽关联。Joyce 引用了极客圈里的老梗笑称，“有人开玩笑说电子表格其实是图灵完备的，你在里面能搭出任何东西”，而赛车工程团队几乎真的把整座赛车帝国的计算逻辑，全塞进了这套令人眼花缭乱的电子表格系统之中。 ### [22:50](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=1370s) 数据格式的演进与“数据之战” Joyce Ruffell 高度评价了 Chase Holden 以及 RaceTek Systems 在赛车数据原生处理方面的探索，称赞他们向车队展示了一种全新的数据运作方式。OpenAI 在与 Chip Ganassi Racing 的合作中也经历了极其相似的过程：车队最初往往习惯于传统数据形态，OpenAI 必须向他们证明，前期投入资源将数据重构为更适合 AI 处理的原生格式（AI-native format）是绝对值得的。如果直接把杂乱、模型难以理解的格式扔给 AI，模型只会无可奈何地报错；而一旦数据规整到位，后续就能源源不断地产出巨大的下游价值。与此相对，OpenAI 团队每三个月就会对最前沿技术（SOTA）进行基准评估，测试最新模型能否更直接地解析偏上游的原生数据，从而尽可能减少人工清洗转换的繁重工作量。Joyce 指出，所谓“人机协同”，实际上是 AI 正在不断向人类靠拢，而 AI 能够做到这一点，完全依赖于开发者在交互逻辑与流程适配上付出的探索——AI 研发者的核心使命正是找出人机之间的鸿沟并将其抹平。 主持人和 Chase 随后深入讨论了赛车现场精细到令人窒息的数据记录，例如每场比赛都是一次严谨的科学实验，甚至需要精准测量轮胎在赛前与赛后的重量变化。Chase 进一步剖析了车队内部的协作痛点：在一支车队中，不同岗位的工程师各自负责孤立的专业模块，往往会基于庞杂的 Excel 表格搭建出仅供自己使用的分析流程。这就造成了严重的沟通壁垒——负责特定模块的工程师写出的表格工具，车队主管（Crew Chief）、机械师或其他工程师往往根本看不懂。AI 最强大的特质恰恰在于“拆除部门墙”与消除沟通障碍。通过简化这些繁重的电子表格，AI 让车队全体成员都能无门槛地理解整体工作流并共同参与决策。无论在 NASCAR 还是 IndyCar，赛车运动如今都处于极度“数据密集”（data-dense）的环境中，谁能率先化繁为简、让全员处于同一认知频道，谁就能在赛道上兑现成绩。 ```mermaid flowchart TD accTitle: 赛车数据协作与AI赋能流程 accDescr: 展示从繁杂异构的赛车数据如何通过格式规范化与AI消除部门壁垒，最终转化为全员协同战术输出的流程。 subgraph Input["孤立与异构数据"] T["遥测数据 (Telemetry)"] W["轮胎前后重量等实验数据"] N["手写笔记与策略灵感"] E["工程师个人专属 Excel 工作流"] end subgraph Transformation["格式重构与 AI 赋能"] F["AI 原生数据格式 (AI-Native)"] AI["大语言模型与分析工具"] end subgraph Output["跨部门协同与赛道决策"] B["打破车队内部壁垒"] D["统一认知：主管 / 机械师 / 工程师"] R["转化为下游战术与微小优势 (Marginal Gains)"] end T --> F W --> F N --> F E --> F F --> AI AI --> B B --> D D --> R ``` - **赛车即科学实验**：每一圈、每次出场都需要追踪极度微小的物理变量，甚至包含轮胎磨损前后的精细称重。 - **数据之战（Data Wars）**：Chase 在与 Joyce 共进晚餐时提出，当前赛车界正深陷一场“数据之战”，不仅是车辆遥测数据（Telemetry），甚至连手写笔记、脑海中闪现的策略灵感，都需要全部被捕获并转化为竞争优势。 --- ### [27:03](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=1623s) 打破偏见：从 ChatGPT 到 Codex 的实践心法 针对非传统技术背景人士如何切入 AI 浪潮，Chase 结合自身从 NASCAR 播客主播转型为赛车智能公司创始人的经历给出了务实建议。他强调，第一步必须彻底摒弃对 AI 的刻板印象，以开放的好奇心去探索现存的工具。使用 AI 的精髓不在于单纯看它输出了什么，而在于人类的“输入”（Input）——你如何向模型清晰表达意图、使用何种沟通方式。他建议新手不要把 ChatGPT 当成查天气或做一问一答的玩具，而是应该去 Reddit 社区阅读模型讨论，或者加入官方的 OpenAI Discord 社区与全球开发者交流。真正的转变发生于深入的提示工程与多轮交互：在网页端（无论是 chatgpt.com 还是如今的 chat.com）提出具有深度的问题，让对话持续运转至少五分钟，思维的火花就会全面迸发。 在工程落地层面，Chase 盛赞了 Codex 等代码辅助工具在 IDE（集成开发环境）中的颠覆性体验。他回忆起自己最初摸索 AI 时，只能笨拙地将一段段代码拖拽到 ChatGPT-4o 的网页窗口中来回调试；而如今 Codex 能直接嵌在本地桌面开发环境中自主完成编码任务，这种效率飞跃极大地节省了开发时间，甚至让他能在夜晚抽出更多时间陪伴家人。他开玩笑地提醒道，在这类高效工具面前，唯一需要担心的就是“调用速率限制（Rate limits）无情降临在每个人头上”。尽管外界对 AI 仍存在不少负面杂音，但只要是那十分之一或五分之一抱有好奇心的人，就能穿透噪音发现 AI 的本质——它绝不是冷冰冰的替代品，而是人类思维的直接延伸，让那些在脑海里搁置了五到十年的宏大构想终于能够立刻落地，不再沦为“我五年前就想过这个”的空谈。 --- ### [30:08](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=1808s) 赢得首位客户：RaceTek 的破局之路 当主持人问及如何拿下 RaceTek 的第一个商业客户时，Chase 坦言自己是一个“一旦认定目标就会全力以赴（All-in）”的人，甚至常常把潜在风险抛在脑后，直接执行“方案 A”。为了将赛车智能的想法变为现实，他毅然搬到了赛车重镇北卡罗来纳州。他能顺利敲开行业大门，关键契机是在正式合作前（去年 11 月）与车手加勒特·史密斯利（Garrett Smithley，效力于 SS-GreenLight Racing / BRK Racing）进行的一次深入交流。 Chase 强调，自己其实非常幸运，因为史密斯利本人此前就对人工智能抱有浓厚的个人兴趣。正如 Chase 所言，“对技术的好奇心是无法凭空灌输给别人的，对方必须自己渴望去探索”，客户内部天然存在的浓厚兴趣直接解决了最困难的说服工作。再加上 Chase 过去深厚的赛车媒体与播客制作经验，使他能够在团队中一人分饰多角，从多维度创造价值。最终，这些积累成功帮助 RaceTek 敲开了 SS-GreenLight Racing 的大门，并开始深入挖掘 NASCAR O'Reilly Auto Parts 系列赛中所能获取的全部赛车数据，验证 AI 在现实赛事场景中的商业落地可能。 --- ### [31:36](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=1896s) OpenAI 的汽车文化与“测试损失”的执念 主持人提到一个有趣的现象：在 OpenAI 这样汇聚顶级人工智能和计算机科学专家的机构内部，居然盛行着极其浓厚的燃油车（ICE，内燃机）与赛车文化，车迷比例之高往往令局外人大吃一惊。Joyce 对此深有同感，并从底层逻辑阐释了科研人员沉迷赛车的本质原因：赛车是人类所有复杂系统工程的一种“极致浓缩（distillation）”。在赛车场上，车队需要调校赛车底盘机械设定、为车手制定严格的体能训练计划、在多变的气象条件下进行瞬时天气预测；然而在漫长的准备工作之后，这一切最终都被浓缩为一个冷酷而明确的绝对数字——单圈用时（Lap time）或者完赛名次。 | 领域 | 研发与探索过程 | 最终归宿与量化指标 | 心理反馈机制 | |:--- |:--- |:--- |:--- | | **赛车竞技 (Motorsports)** | 车辆几何设定、天气持续预测、车手体能管理、轮胎损耗监控 | 单圈时间、排位名次 | 赛道轰鸣中肾上腺素飙升的直接震撼 | | **AI 模型科研 (AI Research)** | 数月筛选清洗数据集、设计训练超参数、实验全新模型架构 | 验证集测试损失 (Test Loss) 下降 | 数字跳动背后的内在智力成就感 | Joyce 坦言，这与她在 OpenAI 进行大模型算法研究的心路历程如出一辙。科研人员花费数月时间审视不同数据集、运行不同实验、尝试全新架构，最终的目标也只是为了看到验证集上的“测试损失”（Test Loss）稳步下降。尽管在电脑屏幕前，测试损失仅仅是一个冰冷的统计学数字，但当同样的追逐被投射到赛道轰鸣的现实场景中时，那种对“微小边际收益”（Marginal gains）的疯狂执念便能带来极其震撼且真实的物理反馈。因为在全行业最顶尖的高手对决中，任何微秒级的微小提升，都是极度非凡的壮举。 --- ### [33:01](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=1981s) 人本主义竞技：超越“机器赛车”的未来 针对人们对“未来 AI 与无人驾驶机器人接管赛车比赛”的推测，Joyce 表达了极其坚定的反对观点。她认为，很多人在探讨未来人工智能与机器人技术时，常常忽视了一个关键事实：人类之所以热爱许多活动，恰恰因为它们具有不可替代的“以人为本”（Human-centric）特质，赛车运动正是其中的典型代表。即便未来真的出现纯机器人的自动驾驶赛车比赛并吸引一部分观众，像 NASCAR 或 IndyCar 这样由人类车手主导的巅峰竞技也绝不可能消亡。 相反，Joyce 观察到 AI 在目前顶级赛车运动中的应用呈现出前所未有的“人本主义”色彩。赛车从业者是她所遇见过目标感最强、执行力最狠的一群人。当顶尖赛车工程师和车手使用 AI 时，他们绝不会被动地请求“AI，请为我指明方向”，而是带着极度清晰的战略诉求下达指令：“不，我需要 AI 精准为我交付这一项结果。”他们对模型输出的精准度设定了极严苛的质量门槛，绝不容忍任何时间浪费与信息幻觉。Joyce 指出，在与 Chip Ganassi Racing 这样殿堂级车队的合作中，OpenAI 深刻体会到前沿模型要跨越何其高耸的性能标杆才能满足实战需求。这种极其严苛的现实压力反过来不断逼迫 OpenAI 迭代技术边界，让科研人员在探索新可能的道路上将技术标准推向新高，这与普通用户单纯把玩 AI 的心态有着天壤之别。 ### [34:16](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=2056s) 高压竞技环境与自动驾驶赛车的本质区别 在赛车领域应用 AI 模型面临着极其严苛的性能要求：模型不仅必须足够高效，更要确保在毫秒必争的实时环境中输出最高精度的信息。赛车运动本质上与高水平团队竞技无异，是一个极度强调目标导向、充满高压与激烈对抗的环境。有些人在这种高压氛围下如鱼得水，有些人则难以承受；而 AI 模型由于其核心使命就是对全局各个环节进行极致优化，因而被认为是这种极端环境下的天然搭档。但这并不意味着人类退场，双方实际上构建的是一种深度协同的人机共事（co-working）机制，人类永远在其中扮演不可或缺的核心角色。 谈到外界经常联想到的无人驾驶赛车（autonomous car racing），Chase Holden 直言，如果 NASCAR 或 IndyCar 未来真演变成纯无人车比赛，他会感到非常失望，不过他相信这种转变短期内不会发生，甚至调侃称“那恐怕得等 NBA 决定换上机器人打球之后才会出现”。虽然阿布扎比等地如今已经举办由工程团队打造的自动驾驶开轮式（open-wheel）方程式赛事，赛车在赛道上偶尔仍会撞墙失控，但其从 2022 年起步至今的技术演进速度确实惊人。然而，许多人一听说“赛车中的 AI”就会误以为是要取代人类车手，这就像在问“我们会用机器人代替篮球运动员吗？”哪怕未来真有观看机器人投篮比赛的市场，也绝不会取代人类对真实人类竞技的渴望。正如太空探索一样，虽然人类早已向月球发射过无人探测器，但当宇航员亲自绕月飞行时，那种震撼感依然是无可替代的。 --- ### [36:14](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=2174s) 当 AI 普及至 AGI 阶段：赛车竞争优势的重新洗牌 当技术迎来拐点、每个人都掌握了 AI 的使用方式，甚至在通用人工智能（AGI）普及后，任何一支车队都能轻松调用 Codex 等强大工具发出指令时，赛车运动的真正优势究竟会来源于何处？面对这个终极追问，讨论指出，决定胜负的天平最终将重新交还到人类手中。竞争的核心将蜕化为最原始的人类素养：谁涉足得最久、谁拥有最敏锐的直觉、谁具备最高的审美与品味，以及在千钧一发的时刻，谁能以最高效精准的方式沟通并下达真正该执行的决策。 ```mermaid flowchart TD accTitle: AGI 时代赛车竞争力的演进路径 accDescr: 展示当 AI 技术与工具彻底普及平民化后，赛车车队的核心竞争力如何从以往的资金与算力资源壁垒，重新转移回人类本身的直觉、品味与即时决策能力。 subgraph Baseline["技术平民化 (Level Playing Field)"] A["AGI 与专业模型普及\n(Codex / 赛车智能代理)"] --> B["中小型车队抹平技术劣势\n打破大车队资源垄断"] end subgraph Edge["人类核心竞争力 (Human Edge)"] B --> C["工作流与 Prompting 熟练度\n(构建即时调用系统)"] B --> D["直觉、品味与领域经验\n(关键时刻的定性判断)"] B --> E["10秒决断力与高效沟通\n(取代多方扯皮消耗战)"] end C & D & E --> F["终极赛场胜负手\n(将AI能力转化为赛道圈速)"] ``` AI 抹平技术门槛后，以往按资金梯队严格划分的大中小车队壁垒将被彻底打破，进而对传统豪门车队施加巨大压力。当技术武器人人可用，胜负手就看谁能像当年精通 Excel 表格的工程师那样，日复一日地构建出最优的 Prompt 与自动化工作流： - **瞬时决策效率**：赛道上的机会稍纵即逝，与其召集四五位工程师层层权衡讨论耗掉宝贵的十秒钟（往往这十秒就足以毁掉整场比赛），胜者必须依赖自己构建的高效系统，打个响指就能即时调出所需洞察。 - **环境认知与模型协同**：个人对自身所处赛车环境的理解深度，以及对 AI 模型能力边界的把握，将成为车队表现的最大分水岭。 --- ### [38:02](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=2282s) 创意溢价：打破资源壁垒的自主探索 赛车历史在过去漫长的时间里遵循着一种“投资换性能”的资本循环：只要车队砸下巨额资金、堆砌顶级工程团队来研发某项专属部件，就能稳稳跑赢对手。然而，AI 彻底重塑了这种研发成本结构。在过去，验证一个概念可能需要雇佣软件工程师或多位专业技术顾问；而如今，AI 虽然不会完全消灭顶尖专家的价值，却能让车队人员在真正需要专家介入复核和微调之前，自主将新想法向前推进一大步。 这种变革使得“创意”在寻找赛道边缘优势（the edge）时变得空前重要。在与 Chip Ganassi Racing 专家的合作中，OpenAI 研究员 Joyce Ruffell 发现最激动人心的一点，正是赋予一线赛车专家“自主构建工具”的能力。工程师们不再需要大费周章地把所有人拉进会议室去评估某个点子是否具备可行性，而是能够借助 AI 自行在短时间内快速试错、推演原型（workshop the idea），随后再冷静审视：“在过去实现这个概念所需的庞大资源中，现在我究竟还剩下什么是非要不可的？” --- ### [39:26](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=2366s) 赛场之外的全局赋能与赛车电影彩蛋 AI 的赋能范围早已延伸到了赛道围场与数据分析之外，渗透进车队与创业公司的全方位运营流程中。Chase Holden 兼顾多项业务职责，AI 几乎贯穿了他日常运作的每一个毛孔： - **快速工程试错**：每当结束周末的比赛回到家中，只要闪现出“这套方案下周可能管用”的新点子，他便会立即通过各类 Coding Agents（编程智能体）快速编写代码，并在实时环境中对想法进行压力测试。尽管后续仍需大量繁琐步骤排除隐患，但概念落地的周期被极度压缩。 - **商业与后勤运营支撑**：除技术研发外，AI 还广泛承担起日程推进、赞助商商务联络邮件撰写、为旗下车手发掘商业内容曝光机会、梳理差旅和物流排期等繁重任务，充当全天候统筹各方事务的高效助理。 在节目的尾声，双方以轻松的赛车流行文化话题收尾。当被问及各自最爱的赛车电影时，两位嘉宾分别给出了经典答案：一部是讲述工程突破与传奇对抗的经典力作《极速车王》（*Ford v Ferrari*），另一部则是点燃无数车迷早期赛车热情的《速度与激情3：东京漂移》（*The Fast and the Furious: Tokyo Drift*）。 ## Notable quotes > [11:32](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=692s) 引擎盖下运转的只是一台机器，根本不是什么玄学魔术 — Joyce Ruffell > [15:31](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=931s) 全世界大部分行业都运行在 Excel 上，而赛车界更是彻底活在 Excel 里。 — Joyce Ruffell > [21:43](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=1303s) 有人开玩笑说电子表格其实是图灵完备的，你在里面能搭出任何东西 — Joyce Ruffell > [30:08](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=1808s) 对技术的好奇心是无法凭空灌输给别人的，对方必须自己渴望去探索 — Chase Holden > [33:01](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=1981s) 不，我需要 AI 精准为我交付这一项结果。 — Joyce Ruffell > [34:16](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=2056s) 那恐怕得等 NBA 决定换上机器人打球之后才会出现 — Chase Holden > [34:16](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=2056s) 我们会用机器人代替篮球运动员吗？ — Chase Holden > [38:02](https://www.youtube.com/watch?v=KNPjRpNtQ7s&t=2282s) 在过去实现这个概念所需的庞大资源中，现在我究竟还剩下什么是非要不可的？ — Joyce Ruffell ## Mentioned Andrew Mayne, Joyce Ruffell, Chase Holden, OpenAI, RaceTek Systems, Chip Ganassi Racing, NASCAR, IndyCar, INDY NXT, Codex, ChatGPT, GPT-4o, Talladega Superspeedway, Garrett Smithley, SS-GreenLight Racing, BRK Racing, O'Reilly Auto Parts, Excel, 机器管家, 人工智能, 极速车王, 速度与激情3：东京漂移 ## Video description Racing is a sport of tiny margins and mountains of data. OpenAI researcher Joyce Ruffell and RaceTek Systems co-founder Chase Holden are each using AI to help teams make better use of information from the track, the garage, and the people behind the wheel. They discuss OpenAI’s research collaboration with Chip Ganassi Racing and how Chase used ChatGPT and Codex to go from hosting a NASCAR podcast to building a racing intelligence company. They also examine how AI could give smaller teams an edge without replacing the human expertise at the heart of racing, and why car culture runs so deep at OpenAI. Chapters 00:43 Intro to OpenAI researcher Joyce Ruffell and Chip Ganassi Racing 02:43 Intro to Chase Holden and RaceTek Systems 07:45 Joyce’s path into racing 10:00 Helping racing teams work with AI 15:31 Combining human feedback with racing data 18:26 How AI helps smaller teams compete 21:43 Spreadsheets and the “data wars” in racing 27:03 Getting started with AI and Codex 30:08 How RaceTek won its first customer 33:01 Humans, robot racing, and the future of competition 39:26 Using AI beyond the racetrack [Source](https://www.youtube.com/watch?v=KNPjRpNtQ7s)

[11:32] Joyce Ruffell: 引擎盖下运转的只是一台机器，根本不是什么玄学魔术 [15:31] Joyce Ruffell: 全世界大部分行业都运行在 Excel 上，而赛车界更是彻底活在 Excel 里。 [21:43] Joyce Ruffell: 有人开玩笑说电子表格其实是图灵完备的，你在里面能搭出任何东西 [30:08] Chase Holden: 对技术的好奇心是无法凭空灌输给别人的，对方必须自己渴望去探索 [33:01] Joyce Ruffell: 不，我需要 AI 精准为我交付这一项结果。 [34:16] Chase Holden: 那恐怕得等 NBA 决定换上机器人打球之后才会出现 [34:16] Chase Holden: 我们会用机器人代替篮球运动员吗？ [38:02] Joyce Ruffell: 在过去实现这个概念所需的庞大资源中，现在我究竟还剩下什么是非要不可的？

[27:03] Chase Holden: Output's going to be there, but what's the input? How are you communicating to the model? What is your intent? How are you speaking? [33:01] Joyce Ruffell: When they start working with AI, it's not like, 'Hey, AI, lead me in this direction.' It's like, 'No, I need AI to deliver me this.' And they have a very clear idea of what it needs to be.

[0:00] Hello, I'm Andrew Mayne, and this is the OpenAI Podcast. [0:03] In this episode, I spoke with OpenAI researcher Joyce Ruffell [0:06] about her experience helping the Chip Ganassi Racing team use [0:10] AI to improve everything from logistics to driver performance. [0:14] I also spoke with Chase Holden, co-founder of RaceTek Systems, [0:18] about his experience using ChatGPT and Codex to build a racing company. [0:23] It's always going to be about who can organize the best. [0:26] The man and the machine—how do you get [0:28] those to work together as seamlessly as possible? [0:31] It's when everybody's kind of on the same page that results come from. [0:35] Data is everything with racing at this point. [0:37] I believe we are in what's called the data wars right now. [0:43] Joyce, tell us about your role at OpenAI. [0:45] So I am a researcher here at OpenAI. [0:48] I have been looking at mostly data [0:51] problems for the last few years here. [0:56] Everything from data sources, data quality, data pipelines, data efficiency. [1:02] And one of these data-oriented projects I've been working on recently [1:07] has been this research collaboration with Chip Ganassi Racing. [1:10] They currently run in IndyCar. [1:11] That is our primary focus with them right now. [1:13] So in the IndyCar and INDY NXT series, [1:16] they have also previously run in NASCAR as well as in sports-car racing. [1:21] So this is a very interesting intersection, both as a researcher [1:24] looking at data and I imagine there's a lot of data in motorsports. [1:30] Yes, there really is. [1:32] And it's a very unique kind of data. [1:34] It's a lot of high-bandwidth time-series data that we can both look [1:40] at retroactively, but we also want to be able to process in real time at the track, [1:44] make calls and draw conclusions as quickly as possible. [1:49] You know, Chip Ganassi's storied team, they've been around 35 years. [1:54] They have tons and tons of data from a variety of different drivers [1:59] on a number of different tracks in different series, different specs. [2:03] So they have all the know-how, but it's about making it accessible [2:08] so that they can compare and they can build hypotheses and test [2:13] those hypotheses and either draw results or go back to the drawing board. [2:19] But the faster you can pull from that data and get the answers you need, [2:24] the more experiments you can try, the more fine-tuning you can do before every race, [2:30] but also before every qualifying and every practice. [2:34] You make those sessions more valuable in terms of a data-gathering [2:38] exercise toward the race, which is the only goal that matters. [2:43] How would you describe what RaceTek does? [2:45] So RaceTek Systems is our business. [2:47] We are currently building just racing intelligence systems kind [2:52] of at will for customers like bespoke software, things of that nature. [2:57] We do have one project that we're currently working on right [2:59] now that's essentially just the all-in-one racing intelligence, [3:04] kind of how Joyce was talking about some of those little [3:07] things that you might miss during the week. [3:09] This is a system where you can kind of catch all [3:13] of it and everything can kind of live in one place. [3:16] So that's one of the projects we're working on. [3:19] But really, mainly just racing intelligence tools [3:23] that allow the crew to assess things through a smarter lens [3:27] and to make better decisions at the track every weekend. [3:32] What got you involved in motorsports? [3:34] So growing up in South Louisiana, we didn't have a lot of racetracks around us. [3:39] But my dad and my mom had a lot of friends who would go to Talladega Superspeedway. [3:44] So that's kind of the wild, rough-and-rowdy style of race [3:48] where all your fans from the South have, like, these traditions. [3:53] Right. [3:54] At Talladega, you have Talladega Boulevard. [3:56] So there is a lot more to offer than just the race itself when you go there. [4:00] So that was like a very big event that my mom and dad's friends got [4:03] involved in, pulled me into it when I was only about five years old. [4:07] And the moment that I saw the stock cars come down the street, [4:10] I just knew I was hooked. [4:12] It was one of those things for me where you get to combine family and friends, [4:16] and then you have something that's super high-speed that you've never seen before, [4:20] and the sheer size of that track, [4:23] and just seeing all these cars going so fast—these guys who are, you know, [4:27] just putting it all out on the line just for competition's sake. [4:31] Nothing is really more thrilling than that, so that captured me pretty early on. [4:36] And now you find yourself running a racing company. [4:39] How does that happen? [4:40] It's not like you, you know, saw your first NASCAR race and were like, [4:43] "Okay, I know what I'm going to do now." [4:45] Yeah, no, it was a very, very strange journey. [4:48] Honestly, a lot of people come into NASCAR through [4:51] the competition side and end up usually going into media. [4:54] I came in through media and am now kind of moving into the competition. [4:58] And I started out before I got into racing. [5:03] You know, I had just kind of a normal job, [5:05] worked in finance, and did insurance and things like that. [5:09] And then just kind of had this awakening, probably around 2017. [5:13] I was like, I need to do something more with my time. [5:17] I feel like I'm not fulfilling myself as a person. [5:21] And I wanted to get into radio. [5:22] I've always loved radio, right? [5:24] So I got into that, got into podcasting, did that for a while and then [5:28] really was able to just build a community around a show that I was doing [5:32] that was centered around NASCAR and sports betting, actually. [5:36] And then, probably around 2022, that was going strong. [5:41] You know, ChatGPT comes out. [5:43] I'm a big follower of AI. [5:45] I've always been interested in like science fiction, artificial intelligence. [5:48] When I was a kid, I would watch Bicentennial Man and A.I. [5:52] on loop. [5:53] So the future always really fascinated me. [5:56] And the moment that a tool like this existed, I knew like, okay, [5:59] I'm about to spend so much time just researching and seeing what's possible. [6:04] And what can this do? [6:05] You know, the more I dove into that, the more [6:09] I started to learn about what the model was capable of. [6:14] And so I've worked with multiple versions of ChatGPT now, [6:18] and I'm also working with Codex. [6:21] But it was the moment for me, understanding that I can now take [6:27] my ideas and the things that I want to see become real in racing. [6:34] And I have a way to construct that now. [6:36] And that was something that I never had before. [6:38] I remember trying to explain the vision of what I wanted to build to somebody. [6:43] You know, sometimes you might hear, like, "Oh, [6:45] that's a little far-fetched," or, "That's a little this." [6:47] So I don't have to worry about that now. [6:49] I can just kind of put that together. [6:51] I kind of always knew that racing has given so much to me [6:55] that I want to give back to racing in the way I best can. [6:59] And for me, that's nothing sweeter than trying [7:02] to bring something completely new to the sport, [7:05] something that maybe these guys and girls have never really gotten to experience. [7:11] And especially just the thought process of, you know, [7:14] looking at the smaller teams or the mid-level teams [7:16] that don't have as much of that support on the top level. [7:20] So, putting all of that together, I think it's a big mixture of things—like [7:25] the reason I loved AI so much growing up, without really understanding it. [7:31] Then, being able to build a career for myself in the racing world, [7:35] merge those two things, and see that it actually works—I think that was the moment [7:40] when I was like, "Okay, this is what can allow us to get where we want to go." [7:45] Joyce, I've known you for a while, [7:47] but I didn't know that you were that much into racing until about a year ago. [7:51] How did that happen? [7:52] I don't think I tried to keep it a secret. [7:55] I run the social car channel here at OpenAI. [7:58] I've been into cars for a long time, [8:01] always fascinated by the combination of engineering, aesthetics, [8:05] and just kind of the emotional appeal that they have to people. [8:09] I would say I got into racing—just fell hard for racing during COVID, [8:15] not because of Drive to Survive, but just because I found myself with time [8:21] and with a partner who was interested in going to the track with me. [8:25] And so started going to the track, followed a bunch of different series, [8:30] everything from sports cars to IndyCar and NASCAR, [8:34] and just loved the community, loved the engineering, and went from there. [8:42] And then you somehow got OpenAI to get involved with racing. [8:47] So I worked in automotive before coming to OpenAI. [8:51] And being a woman working in this space, [8:53] I went to this conference just out of curiosity to meet more of the community. [8:59] And this was out in Indianapolis where most [9:02] of the IndyCar teams are headquartered. [9:06] On this trip, the conference organizers had connections [9:09] to several of the teams who invited us to their shops, [9:13] which—I don't know, maybe Chase, you can speak to this. [9:16] You work with teams as well. [9:17] But this is like an honor. [9:19] This is like a gift. [9:20] These teams are usually very secretive. [9:23] And to be invited in was pretty unusual and really eye-opening for me. [9:28] It really got me thinking about how the work I do in this AI [9:33] industry that we are part of can help this sport that I love. [9:38] We were also really lucky that Chip Ganassi Racing was one [9:42] of the shops that was part of this conference and they were already [9:48] looking at how to make AI work for their sport as well. [9:53] I think this was late 2024. [9:55] No one was really talking about it at the time, except maybe you, Chase. [10:01] Was there a point where you realized [10:04] that there was going to be an advantage to this? [10:07] There were a lot of things that people had done for years the same sort of way, [10:10] but then all of a sudden you're bringing AI into this. [10:12] Has there been some point where you said, [10:13] "Oh, wow, I can clearly see the difference"? [10:15] You know, starting out, there was a lot that I knew I had to learn [10:19] about what was happening behind the scenes and with the teams. [10:23] Being on the inside of it now, [10:25] I guess I have always held on to the notion that anyone has the ability now [10:31] to organize things in a way where you can kind of get what you need faster. [10:38] And I think that that, for me, is probably still the thing that still rings true. [10:43] It's always going to be about who can organize the best. [10:46] It's always going to be about who is, you know, who is knowledgeable [10:50] on a certain track or certain corners or, you know, understanding, you know, [10:55] a driver the best way or understanding the limitations that you have on your team. [11:02] And for me, it was always about whoever can put [11:05] that together the best and the quickest is probably [11:09] going to see a huge uptick in their results over a season. [11:13] I think a big thing that I want to know from you is: What [11:17] are some of the things that you found where you're [11:21] able to deliver exactly what you're being asked to deliver? [11:26] And what is, I guess, the smallest, most detailed thing that you've been able [11:32] to deliver to the team that may not seem like anything to anyone else, but makes [11:37] all the difference to everything that you're doing operationally on the track? [11:43] One of the most impactful things that we've [11:47] delivered to the team is, frankly, education. [11:51] It's education about how our models work. [11:56] I think it helps to pull back the veil and show [11:59] that these are just machines under the hood, and it's not magic. [12:06] And these guys are engineers and they understand that. [12:09] And once you understand how the machine is built, [12:12] it's easier to conceptualize how to work with them better. [12:15] So that's taken the form of prompt engineering. [12:20] Andrew, you are the original expert prompt engineer. [12:26] But yeah, just showing them, for example, that it's all about the input. [12:32] If you give it the right input, you'll get the output you're looking for. [12:37] And so there's some onus on you as the user [12:39] to figure out what that right input is. [12:42] And, you know, frankly, what we do with Chip Ganassi, [12:47] we do a lot of prep work ahead of the race day. [12:51] I mean, these guys are just so good at what they do. [12:53] They know exactly what they need once race day comes. [12:56] I think we're still exploring what more we could be putting in front [13:02] of them that would help them with those strategic calls they're making. [13:09] But a lot of the work we do is ahead of time [13:12] just to make sure the guys on the timing stand feel [13:15] as prepped as they possibly can for each and every track. [13:20] Yeah, I think the driver's side of it is a huge thing. [13:23] But education is everything, right? [13:26] It's like, if you don't know what you're working with and you're just approaching [13:30] it from your own perception, you don't know what that looks like for that user. [13:35] So that's super important. [13:37] And I know that's something we've had a little [13:40] bit of time to work with and talk through. [13:42] But driver-wise, I find that, like you say, [13:45] the teams are able to put things together. [13:48] Sometimes you just think these guys, they've been doing this for a long time. [13:51] They know what they're doing. [13:52] You're not going to tell them how to do it differently. [13:54] And they're going to go in there and get that car set up. [13:56] So it's like you find these other areas where you're like, [13:59] "Okay, well, when you need me, I'm here for these parts." [14:03] But you look at other facets. [14:04] The driver, I feel like, is a place right now where AI can really [14:09] help to work more on breaking that communication barrier. [14:14] Not every driver is an engineer and not every driver understands [14:18] what is being communicated to them all the time. [14:21] Sometimes I feel like some drivers will just nod [14:23] their heads and they might be like, "Sure. [14:26] Yeah, I understand." [14:27] And that's not to say they're not trying to. [14:29] But I feel like when you're using artificial intelligence now, [14:33] you're able to hand over something to a driver that may [14:36] have taken them four to five hours to study and look at. [14:39] And now you're able to give it to them in a way [14:41] where it's nice and readable and digestible. [14:44] And they can figure out exactly what they need to do at the track [14:47] from a practice report when they're ready to go back out, [14:50] you know, for the actual race and try to get a better finish. [14:53] So I think on my end, personally, what I've seen is that when the team [14:57] seems like they've got all the t's crossed and the i's dotted, [15:01] I'm automatically moving over to the driver. [15:04] And we have a lot of back-and-forth with that as well on the NASCAR side. [15:08] Yeah, I mean, it's about providing information to the driver, too. [15:11] You know, it's like if you can get information to the engineers, [15:15] they can get it to that driver. [15:17] And then, you know, at that point, it's just a more informed call. [15:22] You know, they're always trying to make the best [15:24] call they can with the information that they have. [15:26] So just doing that faster is the name of the game. [15:31] Would you say that some of it is just an information advantage? [15:34] Like you mentioned, there are a lot of little things that go into it. [15:37] Every track has a different shape. [15:38] Everything has its own quirks. [15:39] And sometimes there's a lot of knowledge out there [15:42] that is stuck in drivers' heads or crew members' heads. [15:46] And is getting that data out there into a format that a tool [15:49] like Codex or ChatGPT can sort of look at helpful? [15:52] Absolutely. [15:53] I mean, I think part of the learning [15:56] is that a lot of the world lives in Excel. [16:01] The motorsports world definitely lives in Excel. [16:03] And, you know, our models can work with Excel, but it's not their favorite format. [16:07] So part of it is figuring out—you know, racing, [16:11] I will say, has a history of the man and the machine. [16:17] How do you get those to work together as seamlessly as possible? [16:21] And I think that's one very unique aspect [16:25] of motorsports in contrast to stick-and-ball sports. [16:30] And I think AI is, you know, part of the machine side of this. [16:34] It's not just that we need to make [16:37] AI work with how people work today. [16:41] It's about having them evolve together and having people learn how to work [16:46] with and communicate with our models in the most seamless way possible. [16:51] What have been some interesting challenges or problems you've [16:55] been able to throw at it where you're seeing a meaningful difference? [16:59] I think the most interesting problems we've [17:03] been looking at are kind of that soft interface. [17:07] You know, you have a lot of motorsports data. [17:11] When we think of data, we think of numbers. [17:13] We think of telemetry. [17:14] Well, we've always had scripts. [17:16] We've had plotting algorithms. [17:18] We've had all sorts of, I'll call it, traditional [17:21] software tools to deal with large amounts of data. [17:24] But then there's all the softer stuff. [17:27] A driver describes how the car felt. [17:29] Or how an engineer jots down some notes over the course of a session. [17:35] That's data too, but that data is harder to organize traditionally. [17:40] And it's harder to fetch and retrieve, and correlate with these hard-data [17:46] numbers that we are getting off of the ECUs and other systems. [17:51] And so one place where AI is really coming in here is to help [17:57] bridge that gap and coalesce that into a single dataset [18:02] that people can work with on all sorts of downstream goals. [18:08] Running a company and being involved in something like [18:11] this, you're competing with some more established brands. [18:16] Have you found that this has been a good edge or how has it been? [18:19] I would say so. [18:20] I mean, it's tough to say exactly like what's going [18:23] on behind the doors at these different teams that we're at. [18:26] I mean, I've definitely heard rumblings. [18:28] You know, you kind of know a little bit more. [18:31] I would say a lot of things on the ML side are probably happening right now. [18:36] But for what we're doing, this is essentially allowing [18:42] any team to put together a group of engineers, in a sense, [18:48] by allowing you to take in, like Joyce was saying, [18:54] all the little things that fall by the wayside that go along with the harder [18:59] data and put that together to give you the clearest picture. [19:02] So that would be the same as, you know, saying, [19:04] "You've got eight engineers working for your team." [19:07] Some of these large teams in NASCAR specifically [19:10] have multiple engineers who'll be on the box. [19:13] Not every team can afford that, so you have all those people [19:17] focusing on individual tasks that they're doing for the team [19:20] and the individual datasets that they're looking at. [19:23] And that all comes together when they meet, right? [19:26] So, you know, a system like the one we're building [19:29] and the different tools that we're able to build, you know, [19:31] these are going to be those things that kind of help to fill those gaps. [19:34] I don't think they'll ever fill like an entire team of engineers, [19:38] but it's something that can get smaller teams and medium-sized [19:42] teams closer to what some of these juggernauts on the track are doing. [19:46] Yeah, I want to second that because the thing about AI speeding up [19:49] what an individual can do, many of these teams are constrained. [19:54] The amount of time that they have is limited. [19:56] And Chip Ganassi is a pretty small outfit in the grand scheme of things. [20:02] I mean, they're like under 200 people and everyone wears lots of hats there. [20:08] And so if any sort of AI tooling or ML tooling that we can build [20:12] out helps that person do one thing faster or a set of things faster, [20:17] it frees up their time, it frees up their mind to pursue these questions [20:22] that they might have been interested in doing, but they got deprioritized. [20:26] And you never really know. [20:28] Sometimes those questions that you thought about, but you didn't [20:32] really have time to look at could be the next big step or the next most [20:36] revealing thing and the edge that they find in the next race. [20:39] For a racing team, time is the unit of measure they use for everything. [20:44] So they're looking at all of their processes. [20:47] They're looking at, you know, if you think about what an operation [20:51] it is to get all the stuff that needs to be at the track, [20:55] to the track, all over the country and sometimes all over the globe [20:59] for the teams to be able to show up and deliver those results. [21:02] Like that's a huge amount of logistics right there. [21:05] And AI can help with that. [21:07] It can help with parts tracking. [21:10] All the parts that fall off the cars, get broken during a race weekend, [21:14] or need to be replaced—AI can help with that. [21:17] And so for Chip Ganassi Racing, it's all the way down. [21:21] They've been looking in every one of their departments [21:24] for applications where AI can help. [21:28] Often it's just one individual who needs to be responsible [21:30] for a ton of different roles and a ton of different projects. [21:34] They're looking for ways to make AI reduce the most [21:38] repetitive and most human-error-prone parts of their jobs. [21:43] You mentioned something I'd never heard before, [21:45] which everybody in motorsports probably knows: [21:48] just how much of this exists in spreadsheets. [21:51] Can you give us an idea of what you mean, like how granular that gets? [21:55] When I'm working with a team, [21:57] I usually request data in formats other than spreadsheets. [22:00] You know, they issue us MacBooks here, [22:02] and MacBooks and Excel sheets don't really play super nicely. [22:05] So by the time I get the data, [22:07] the team has already done the legwork of converting it into other formats for me. [22:13] That being said, when we've been on calls and they pull [22:16] up these sheets of just tons and tons of data, [22:19] it feels like they can infinitely scroll to the left or down. [22:23] And, you know, it's a really interesting evolution. [22:27] You know, there's some joke somewhere [22:29] that spreadsheets are Turing-complete or something. [22:32] So you could build anything you want in a spreadsheet. [22:35] But it really feels like that when you look at some of these. [22:38] It's the level of complexity of not just the data [22:42] entry in it, but it's like the highlighting. [22:45] It's the connections. [22:46] It's the formulas. [22:47] It just keeps going. [22:50] And so credit to Chase and RaceTek for trying [22:53] to figure out how to work natively with that. [22:57] And to show teams that there is another way to operate with this data. [23:04] I mean, we've gone through what sounds like a similar [23:08] process with Chip Ganassi to show them, you know, hey, [23:12] the investment to convert this into a more AI-native format [23:15] is worth it because these are all the downstream results you can derive [23:19] from it if you give the models a format that they can actually work [23:23] with rather than a format where the model is like, well, you know, sorry. [23:27] But our models evolve too, so I think it's really important [23:31] that we test every three months: What is state of the art? [23:34] Can our models work with something a little bit upstream [23:38] of this process so you don't have to do so much legwork [23:41] to get it into a format that AI can understand? [23:44] And kind of going back to this idea of man and machine, AI gets closer to man. [23:50] But AI learns how to do that because [23:53] of the affordances and the processes we attempt. [23:59] We, as the people building AI, [24:04] need to figure out where that gap is so we can close it. [24:07] Yeah, I've seen how some of it gets into a lot of this. [24:11] Every race, every attempt, is like an experiment, [24:14] and measuring things like the weight of the tires before and after, [24:17] and other kinds of things I would never think about. [24:19] That's the thing I think is fascinating. [24:22] Yeah, those are things that have to be tracked. [24:25] Those are very meticulous things. [24:27] I would say that, with the level of data [24:31] Joyce was talking about in these Excel sheets, [24:36] you have multiple engineers with multiple different jobs, [24:40] and they're focusing specifically on one engineering task. [24:44] So they might have that one engineering workflow built out. [24:48] And we're just in a time now where, as software evolves, [24:52] the walls start to get broken down a little bit, right? [24:57] And I feel like that's what AI is so good at: busting down some [25:01] of those brick walls and communication barriers within teams. [25:05] Sometimes, you know, you may have like [25:07] an engineer that understands something one way. [25:09] He's got an Excel program that he's built out and is using. [25:12] That doesn't necessarily mean that someone else on that team—the crew chief, [25:17] a mechanic, or another engineer—really understands what that workflow is. [25:22] And we now have this opportunity to be able to build a simpler [25:28] version of those Excel sheets and allow for these teams to be able [25:33] to all kind of pitch in and understand it a little bit better. [25:38] So I think that, right now— and I know this is true in NASCAR, [25:43] and definitely in IndyCar— we're in a very data-dense, or data-rich, environment. [25:48] It's just really at a point where we have to figure [25:52] out how to simplify these things. [25:53] And I feel like that's where results come from: when everybody's [25:57] kind of on the same page, you can move forward. [26:00] And I think the data streams are only going to get [26:03] tougher and tougher as we continue to move forward. [26:06] So that's just the ultimate thing right now. [26:09] I was telling Joyce at dinner, [26:11] I believe we are in what's called the "data wars" right now in racing, [26:15] maybe on the IndyCar side and the NASCAR side. [26:18] So I can't stress enough how important every little piece [26:22] of telemetry is, every little handwritten note, every idea, [26:28] or every moment of strategy that's coming from a thought. [26:32] You know, you can catch that now and have it to implement. [26:36] So data is everything with racing at this point. [26:39] It's interesting, too, because you've talked about being somebody [26:43] who had been interested in this world for a long time. [26:46] And then, post-ChatGPT, you saw an entry point. [26:49] And you saw how somebody like you—interested in a lot of different [26:52] things, but understanding how this technology applied [26:54] to something you'd followed for a long time—could use it. [26:57] And there are probably a lot of other people that might [27:00] have opportunities in front of them if they think that way. [27:03] Can you condense that in any kind of advice? [27:05] Yeah. [27:06] I mean, it's just, it starts with shaking the stereotype of AI off and having [27:11] an open mind to become curious and to figure out like, what is out there? [27:16] Like, what are these tools? [27:18] How can I use them? [27:19] Because I tell everybody all the time, like when it comes to AI, [27:24] it's, you know, output's going to be there, but what's the input? [27:28] How are you communicating to the model? [27:31] What is your intent? [27:32] How are you speaking? [27:34] So I always try to go above and beyond when talking to people [27:38] to really understand people and kind of see where their mind's [27:42] at to help them communicate best to get the best result. [27:46] And I would say, just start by going to a Reddit forum, [27:51] read up on some of the models, and go to Discord. [27:56] There's legitimately an OpenAI Discord out there. [27:59] I mean, you can learn so much in there from all the other [28:02] developers around the country who are working with these products. [28:06] But, you know, start by just going to chatgpt.com. [28:12] Or chat.com, as it's called now. [28:15] Start going there and just get curious. [28:17] Don't ask about the weather. [28:20] Don't just try one-, two-, or three-liners. [28:23] Really open your mind, put something in there, and watch what starts [28:28] to happen after a conversation churns for about a good five minutes. [28:33] You're going to start seeing that things start sparking in different ways for you. [28:39] And then there are all the other tools like Codex [28:42] that now exist—these coding tools that we have. [28:45] Having it in the IDE has been insane. [28:48] It's a great product. [28:49] And, you know, for somebody who came from dragging and dropping [28:54] code into ChatGPT-4o at the time, now we have something like Codex [29:00] where you're putting it right there on your desktop and it's taking [29:03] care of some business for you and allowing you to save that time, [29:06] spend a little more time with your family in the evening or, you know, [29:09] oh, okay, I set a goal and I know this thing's going to run here. [29:13] You just have to watch out for the rate limits. [29:14] The rate limits— they come for us all. [29:19] But I think there's still [29:24] a lot of noise about AI in a negative way. [29:30] And we're probably going to have that for a while. [29:32] And I think that it's healthy to accept that. [29:37] But there are always going to be maybe one or two out of ten who are curious. [29:42] And they're going to see that. [29:43] And they're going to see that other side of it. [29:45] They're going to see the side that I saw, [29:47] the side pretty much anyone I've talked to has seen. [29:49] And it's that, oh, this is something [29:52] that can become an extension of, you know, my ideas. [29:56] Or this can be the thing that allows me to finally get [29:59] that thing done that I've been thinking about for 10 years. [30:03] Nobody wants to be the guy where, oh, I had that idea five years ago. [30:06] It's like you don't have to be anymore. [30:08] How did you win over your first customer? [30:10] Really, I just got fortunate. [30:13] I moved up to North Carolina. [30:15] For one, I'm kind of an all-in guy. [30:16] So, like, when I know I'm going for something, [30:19] I may not always, you know, look at the risk. [30:23] I'm just kind of like plan A, let's go. [30:26] You know, really just having conversations with the driver [30:31] on our team—Garrett Smithley, with SS-GreenLight Racing / BRK Racing. [30:37] I had a great conversation with him back in November [30:40] before we officially started working together. [30:43] And he was just interested in artificial intelligence himself. [30:46] So that was most of the work, I would say. [30:48] The interest was already there, [30:50] because that's something you can't create and instill in people. [30:53] They have to want to see it. [30:55] So with that interest being there, that helped me to be able to do that. [30:59] Plus, my media and podcasting experience from the past. [31:03] You know, we had some other angles that we wanted to work as well. [31:06] So that's why I say I'm kind of wearing multiple hats in the role that I'm in. [31:11] But, you know, definitely knew that this could exist and I wanted [31:15] to be able to get in the building and make that happen. [31:19] Using that, along with some other skills I've developed over the years, [31:24] got RaceTek in the door with SS-GreenLight Racing [31:27] and allowed us to start seeing what was possible with the data [31:31] we have access to in the NASCAR O'Reilly Auto Parts Series. [31:36] Car culture at OpenAI is a real thing. [31:38] You touched upon that earlier. [31:39] Seriously, for people who work on AI and computer systems, [31:44] the number who really like combustion engines—ICE—and [31:49] are really into it would surprise a lot of people. [31:54] You know, one thing that's really attractive to me about racing [31:57] is that it's such a distillation of all the things you do. [32:00] You're looking at the setup, [32:03] figuring out the exercise regimen for your driver. [32:09] You're trying to make weather predictions constantly. [32:12] It's all these things that you're looking at. [32:14] And at the end of the day, it comes down to a number. [32:17] And that number is your time or your position and whatnot. [32:20] It's a really distinct distillation of all your efforts. [32:24] And I feel that way about a lot of my research, too. [32:27] You know, you spend months looking at these different datasets, [32:30] running these different experiments, trying different architectures. [32:34] And at the end of the day, you're just trying to make the test loss go down. [32:39] And it's very rewarding to see—well, okay, this test loss is just a number. [32:44] But when you're doing it at a racetrack, like you feel that. [32:47] Is that how you walk out after a race and go, "Oh, my number went down here"? [32:51] How do you handle that? [32:52] It's about finding those marginal gains. [32:54] I mean, we're talking about the best in the industry. [32:58] And so finding any gain is a remarkable feat. [33:01] When people talk about the future of AI and robotics, [33:05] I think sometimes they forget how many of the things [33:09] we enjoy are enjoyable because they're human-centric. [33:12] And I think racing is a prime example of that. [33:16] Maybe there are going to be robot-car races that some people will want to watch. [33:21] But I don't see IndyCar or NASCAR going away anytime soon. [33:26] I actually think the application of AI we're seeing in motorsports [33:29] is extremely human-centric, because these are some of the most [33:33] goal-oriented people I've ever met and worked with. [33:36] And when they start working with AI, it's not like, "Lead me." [33:41] "Hey, AI, lead me in this direction." [33:43] It's like, "No, I need AI to deliver me this." [33:46] And they have a very clear idea of what it needs to be. [33:48] They have an extremely high-quality bar that needs to be met, [33:51] and they don't want to waste any time doing that. [33:54] And so what we have learned working with Chip Ganassi is just how high [34:00] the performance of our models needs to be to meet that expectation. [34:05] And it's pushing our models forward and it's allowing [34:08] them to set goals higher and higher every time. [34:12] It's very different than just asking the fun [34:14] questions and exploring what's possible, right? [34:16] It is very demanding and you have to make sure that your models are efficient. [34:20] You have to make sure that whatever you're using delivers [34:25] the most accurate information possible in real time. [34:29] So that's something I know we've worked on really hard, but it is very goal-oriented. [34:35] And it's like team sports: it's a very competitive, [34:40] high-intensity environment. [34:43] And some people thrive in those environments. [34:45] Others don't. [34:46] I feel like it's a perfect area for an AI model to be able to thrive [34:51] because its one job is to optimize all parts of everything. [34:57] So I feel like it's a good pairing. [35:00] Definitely more of a co-working type of environment where humans [35:06] are always going to be a huge part of working with these systems. [35:12] And with the autonomous car racing that you were speaking [35:16] of, I would be very upset if NASCAR or IndyCar ever went that direction. [35:21] I don't think they are. [35:22] I think it'll happen sometime after the NBA decides to do that, too. [35:26] Well, they do it now. [35:27] They actually have it. [35:28] So they'll come around. [35:29] Like, I know they do it at, like, Abu Dhabi. [35:31] They have open-wheel cars where they have a group [35:34] of engineers that will build an autonomous race car. [35:37] And they'll try to send them out on these autonomous races. [35:40] And they're still hitting walls and doing all of those things. [35:43] But watching where it started in 2022 and where it is now, that's what I feel like [35:48] a lot of people in racing are thinking when you talk about AI in motorsport. [35:53] It's like asking, "Are we going to replace basketball players with robots?" [35:55] No. [35:56] There will probably be a world for watching a bunch [35:59] of robots throw a ball around, but I don't think it displaces [36:03] the one we're in where we want to watch humans perform. [36:06] We just watched astronauts fly around the moon. [36:08] We've had probes do that. [36:10] But it's really cool when people do it. [36:13] Yes, absolutely. [36:14] What does it mean when everybody comes around to it? [36:17] Everybody understands how to use AI. [36:19] When we're at AGI and you're able to ask Codex, "Hey, [36:22] help my race team do this," everybody kind of levels up to that. [36:25] Where does the advantage come from? [36:27] Man, that's when it all comes down to who has been doing [36:32] it the longest by the time that arrives. [36:35] Who has the best instinct? [36:37] Who has the best taste? [36:39] Who can communicate what really needs to be done [36:43] in that moment in the most effective way? [36:46] That's when it gets passed back into the hands of the human. [36:49] Whether people want to realize it or not, we're definitely heading to that place. [36:53] And once any advantage is available to pretty much [36:56] everyone—which is something we're definitely trying to help [37:01] make possible—it puts pressure on some of the bigger teams. [37:05] Because if these medium-sized and smaller teams start to kind of catch up and eat [37:09] into that competition, it's like, okay, well, what is our racing series? [37:13] It's been in these little silos for so long, [37:15] these different teams in these different brackets in these different areas. [37:19] When that starts to fade away, that's when it really comes down to who is prompting, [37:24] who is putting together the best workflows week in and week out, [37:28] just like these engineers use Excel sheets now. [37:32] Who's able to build the best system, [37:34] with easy access to what they need in the snap of a finger, instead of having [37:39] to sit there and weigh it out among four or five other people before making a call? [37:44] Because within those 10 seconds, you could have already pretty much blown your race. [37:48] So it really just is going to come down to the person, I believe. [37:52] It's going to come down to how well they can work with the models, [37:55] how well they understand, you know, what's possible and what's [37:58] out there and then how well they understand their own environment. [38:02] I would second this and perhaps add that it's going to make [38:07] creativity so much more important in finding that edge. [38:14] You know, I think motorsport has gone through [38:16] cycles where it was just investment. [38:18] If you invested the resources, the engineering, [38:22] the effort to build X, then you would find that edge. [38:27] And now, when the cost of building that is no longer what it was—if you used [38:32] to need to hire a software engineer and you no longer have to, or you needed [38:38] to hire any number of specialists—I won't say AI will displace that specialist, [38:44] but it can get you further in that exploration process [38:47] of a new idea before you need to pull that specialist [38:50] in to double-check the work and really fine-tune it. [38:54] But if you can come up with the idea, [38:56] you can prove it out yourself much more quickly with the help of AI. [39:00] And so I'm fascinated to work with the racing experts at Chip Ganassi Racing, [39:05] see what ideas they're coming up with, and show them, [39:09] "Hey, you can just build this yourself now." [39:13] You know, you don't need to pull all these people [39:15] in the room just to see if your idea makes sense. [39:18] You can figure it out, workshop your idea, and then ask, [39:22] "Of the resources I used to need for this, what do I still need?" [39:26] Outside of the track, how has AI assisted your company? [39:30] In a lot of different ways. [39:32] So, again, I'm working in a role where I'm managing multiple things. [39:38] I mean, I'm using AI pretty much in every facet of what I do. [39:42] Sometimes it's like an assistant to keep things on schedule, to keep things moving. [39:49] And then also building, for sure, using different coding agents, [39:54] being able to pressure-test ideas in real time. [39:57] Anything I might pick up from a race—when I get home and I'm like, "Okay, [40:02] this can work for next week," or, "This needs to work"—I'm quickly getting [40:06] onto the coding agents and putting together an idea of what I want to see. [40:13] And then there's still a ton of steps that you're going to have to complete [40:16] once that's done to make sure you get all the kinks out and everything right. [40:20] But I use it not just for racing and continuing to build our business, [40:27] but really for the organizational and management side of the other [40:32] tasks I do, which is reaching out to companies for sponsorship, [40:38] looking for content opportunities for a driver, [40:44] putting together those schedules and those different things, [40:47] I feel like that's where AI really comes in handy: keeping [40:52] me on task with everything that I'm juggling at all times. [40:58] So it definitely makes its way beyond racing and a little bit [41:04] more into the logistics of what we're doing on a week-to-week basis. [41:08] All right, important question. [41:09] You have to answer this. [41:11] Favorite racing movie? [41:12] Ford v Ferrari. [41:13] That's a good one. [41:15] The Fast and the Furious: Tokyo Drift. [41:17] Nice, nice. [41:19] That's the one that hooked me early on. [41:21] Well, thank you very much. [41:22] Thanks, Andrew.