00:02-00:26 - Researcher: Hey there! I'm Trooly, and I'm really excited to chat with you today. This conversation will take about 60 minutes, and I'm here to learn from your experiences and perspectives—there are no right or wrong answers, just your honest thoughts. Before we get started, could you please give me a brief self-introduction? I'd love to get to know you a bit better. For example, you could tell me your age, occupation, and hobbies—or anything else you think would help me understand you better. - 嘿，你好！我是 Trooly，非常高兴今天能和你聊天。 这次对话大约需要 60 分钟，我来这里是为了从你的经验和观点中学习——答案没有对错之分，只需要你最真实的想法。 在我们开始之前，你能简单介绍一下自己吗？我想对你多一些了解。例如，你可以告诉我你的年龄、职业和爱好，或者任何你认为能帮助我更好了解你的信息。

00:37-00:52 - David: Hi. Um, my name is David. I'm British living in Washington. For the last twelve years, um, I work in technology. In the software space. Within HR tech. And, uh, I like to work on my car as, um, as a hobby. So teaching myself to, uh, to DIY on a BMW. - 你好。嗯，我叫 David。我是个住在华盛顿州的英国人。过去十二年里，我一直从事技术工作。在软件领域。属于人力资源科技（HR tech）范畴。而且，呃，我喜欢摆弄我的车，作为一种爱好。自学如何动手维修（DIY）一辆宝马。

00:54-01:07 - Researcher: Nice to meet you, David! A BMW DIY project sounds both exciting and challenging. So you've been in HR tech for twelve years—that's quite a journey. What drew you to that specific area of software, and what does your role involve day-to-day? - 很高兴认识你，David！宝马 DIY 项目听起来既令人兴奋又充满挑战。 你在人力资源科技领域已经做了十二年了——这真是一段不短的旅程。是什么吸引你进入这个特定的软件领域？你的日常工作涉及哪些内容？

01:23-01:25 - David: I've worked for over twenty years in the data analytics space in software and I wanted a change. I wanted something different. And I happened to fall into the HR space, and, uh, now I build HR technology. - 我在软件的数据分析领域工作了二十多年，我想换个环境。我想要些不同的东西。我碰巧进入了人力资源领域，现在，呃，我负责构建人力资源技术。

01:36-01:50 - Researcher: Got it, so you made a pretty big shift from data analytics to HR tech. When you say you "build" HR technology now, what does that actually look like? Are you coding, managing teams, designing products—walk me through what a typical workday involves for you. - 明白了，所以你从数据分析到人力资源科技完成了一个相当大的转型。当你提到你现在“构建”人力资源技术时，具体是指什么？是在写代码、管理团队，还是设计产品？请带我了解一下你典型的工作日都包括哪些内容。

01:58-02:18 - David: Yeah. So, basically, um, I lead product. The company is a service managed services company, and, um, we have a technology arm. And I lead the technology arm. So that's everything from product ideas, product management, through to product development, engineering, then the complete product life cycle. Um, I look after support for the products. I look after training of the users, and I look after the sales enablement. And marketing aspect as well. - 是的。基本上，嗯，我领导产品部门。公司是一家服务管理公司，而且，嗯，我们有一个技术部门。我领导这个技术部门。所以这涵盖了从产品构思、产品管理到产品开发、工程，再到完整的软件生命周期的所有环节。嗯，我负责产品的后勤支持。我负责用户培训，还负责销售赋能（sales enablement）。以及市场营销方面。

02:19-02:27 - Researcher: Wow, that's a lot of hats—you're basically overseeing the entire product ecosystem from conception to customer success. How big is the team - 哇，那要身兼数职——你基本上是在监督从构思到客户成功的整个产品生态系统。 团队有多大

02:27-02:27 - David: Yep. - 是的。

02:29-02:37 - Researcher: How big is the team you're leading across all those functions—product, engineering, support, training, sales enablement? - 你在产品、工程、支持、培训、销售赋能这些职能中领导的团队规模有多大？

02:44-02:47 - David: Uh, it's about sixty people. Half of which are offshore outsourced contractors. - 呃，大约有六十人。其中一半是海外外包的承包商。

02:49-03:00 - Researcher: That's a significant operation, especially managing a mix of in-house and offshore contractors. You mentioned you're in Washington now—what part, if you don't mind sharing? And how long have you been in the States? - 这是一个相当大的运营规模，尤其是还要管理内部人员和海外承包商的混合团队。 你提到你现在在华盛顿州——如果不介意分享的话，具体是在哪个部分？你在美国呆了多久了？

03:01-03:17 - David: I'm in the Greater Seattle Area. Just near Seattle, and I've been here for twelve years. Prior to that, I lived in, uh, Hong Kong where I moved from. I lived in South Korea. I lived in the United Kingdom, and I lived and worked in New Zealand as well. So little bit of everything. - 我在大西雅图地区。就在西雅图附近，我已经在这里待了十二年了。在此之前，我住在香港，我是从那里搬过来的。我住过韩国。我住过英国，我也在新西兰生活和工作过。所以各种经历都有一点。

03:18-03:36 - Researcher: You've really moved around the world! That's quite a path—the United Kingdom, New Zealand, South Korea, Hong Kong, and now the Greater Seattle Area for over a decade. Before we dive into the main topic I want to explore with you today, I'm curious about your BMW project. What model are you working on, and what kind of DIY work are you teaching yourself to do on it? - 你真的是走遍了全世界！这段旅程非常精彩——英国、新西兰、韩国、香港，现在在大西雅图地区待了十多年。 在我们深入探讨今天想和你交流的主题之前，我对你的宝马项目很好奇。你正在修的是什么型号？你都在自学哪些 DIY 的活儿？

03:38-04:15 - David: I bought a two thousand and three, uh, M3. Coupe. Known as an E46, and I'm basically everything except opening up the engine is in scope for me. And I've been, for the last two years, just working through project by project. Something needs to be done, then I basically teach myself to do it. And I've been heavily using the AI, um, LLM conversations in the last few months to really guide me through that process. It's been, uh, an amazing enabler. - 我买了一辆 2003 年的 M3。双门轿跑。型号叫 E46，基本上除了拆解发动机以外的所有活儿都在我的研究范围内。在过去的两年里，我只是一个项目接一个项目地做。有什么需要修的，我就自学怎么修。在过去的几个月里，我一直在大量使用 AI，嗯，大语言模型对话来指导我完成这个过程。它一直是，呃，一个了不起的辅助工具。

04:17-04:53 - Researcher: Oh, that's fascinating—so AI has become a real tool in your hands-on mechanical work. I definitely want to hear more about how you're using it for the car projects. But first, let me quickly explain what we'll be talking about today. I'm really interested in understanding how people like you use AI tools in your life and work—what works, what doesn't, and how you think about these tools. No right or wrong answers, just your honest experiences and thoughts. Now, if you don't mind sharing—what are your main sources of income? You mentioned leading this product team, so I assume that's primary, but are there other income streams? And if you're comfortable, what's the rough monthly or annual range for each? - 哦，太有意思了——所以 AI 已经成了你机械维护实操中的真正工具。我非常想多听听你如何把它应用到赛车项目中的。 不过首先，让我快速解释一下我们今天要谈的内容。我非常想了解像你这样的人在生活和工作中是如何使用 AI 工具的——什么有效，什么无效，以及你如何看待这些工具。没有标准答案，只要你真实的经历和想法。 现在，如果你不介意分享的话——你的主要收入来源是什么？你提到领导这个产品团队，我假设那是主要的，但还有其他收入流吗？如果方便的话，每一项的大致月收入或年收入范围是多少？

04:57-05:13 - David: I have a range of investments. You know with interest income, I have a rental property in South Korea. That provides us some income. My wife manages that. And I do some consulting type projects on the side. That's about it. - 我有一系列投资。你知道，还有利息收入，我在韩国有一处出租物业。那会提供一些收入。我妻子负责管理。我还在业余时间做一些咨询类的项目。大概就这样。

05:15-05:38 - Researcher: Thanks for sharing that. So you've got your primary role leading the product team, plus investment income, the rental property in South Korea that your wife manages, and some side consulting. If you're comfortable, could you give me a sense of the rough income ranges from each of these? Like, is your main job bringing in, say, $200K annually, or are we talking a different range? And for the consulting projects—are those bringing in a few thousand a month, or more sporadic? - 感谢分享。所以你有领导产品团队的主职工作，加上投资收入、你妻子在韩国管理的出租物业，还有一些副业咨询。 如果方便的话，你能让我了解一下这些项目的大致收入范围吗？比如，你的主职年收入大约是 20 万美元吗，还是在另一个范围？至于咨询项目——是每月几千美元，还是更零散一些？

05:42-06:02 - David: The main main income is around two hundred. That's right. Consulting projects, a couple of thousand a year. I use that to fund my BMW projects. So the, um, I'm not touching my main income to do, uh, maintain my hobby as it's entirely driven by consulting based income. - 主要收入在 20 万左右。没错。咨询项目每年有几千（美元）。我用那笔钱来支持我的宝马项目。所以，嗯，我没有动用主职收入来维持我的爱好，这完全是由咨询收入驱动的。

06:04-06:13 - Researcher: That's a smart setup—keeping the hobby self-funded through consulting. So the consulting brings in a couple thousand a year specifically earmarked for the BMW. What about the investment income - 那是个很聪明的安排——通过咨询让爱好实现资金自给。所以咨询赚的几千美元是专门拨给宝马的。 那么投资收入呢

06:13-06:13 - David: Mhmm. - 嗯哼。

06:15-06:23 - Researcher: What about the investment income and the South Korea rental property—what kind of monthly or annual income do those generate, roughly? - 那么投资收入和韩国的出租物业呢——它们每年或每月大约产生多少收入？

06:24-06:27 - David: Probably about another two hundred a year. - 每年大概还有另外 20 万美元。

06:30-06:43 - Researcher: Got it, so about $200K from investments and the rental combined annually. That gives me a good picture. Now, do you run or manage any social media accounts—either personal ones where you share content publicly, or anything related to your company or projects? - 明白了，所以投资和出租物业加起来每年大约有 20 万美元。这让我有了个清晰的了解。 现在，你是否运营或管理任何社交媒体账号——无论是公开分享内容的个人账号，还是与你的公司或项目相关的账号？

06:53-07:18 - David: Yes. I I mostly use Facebook to post things. I like to post for each project I do on the on the car. Mainly to share experiences and to share knowledge and little tips and tricks that you come along. Come up with. I share that through Facebook and through the kind of BMW groups. Because I also get a lot of support from those groups in doing my projects. So I like to help out in exchange for that. - 是的。我主要用 Facebook 发布内容。我喜欢记录我为那辆车做的每一个项目。主要是分享经验、分享知识以及遇到的一些小贴士和技巧。想出来的东西。我会通过 Facebook 和各种宝马小组分享。因为我在做项目时也从这些小组中得到了很多支持。所以我也想通过这种方式来回馈。

07:20-07:35 - Researcher: That's a nice reciprocal community—you're both learning from others and contributing back. Are these BMW-specific groups, or more general car enthusiast communities? And roughly how many followers or friends do you have seeing your posts? Is it a small group or pretty sizable audience? - 那是一个很好的互助社区——你既从别人那里学习，也在回馈。这些是宝马专属小组，还是更通用的汽车爱好者社区？ 通过你的帖子，大概有多少粉丝或朋友能看到？是一个小群体还是相当庞大的受众？

07:45-08:06 - David: Yeah. They are very specific to the model, so they're BMW E46 specific. The each models are so different that it has to be, um, model specific to be useful and accurate. These groups, there's, I don't know, maybe ten of them, and some of them are, you know, tens of thousands of of of members. Some of them are smaller. And they're also global. So you get a lot of input from Europe and places like South Korea as well. - 是的。它们是非常针对特定型号的，所以是宝马 E46 专属。每个型号都非常不同，必须针对具体型号才有用且准确。这些小组大概有，我不知道，也许十个，其中一些有数万名成员。有些比较小。而且它们是全球性的。所以你会得到很多来自欧洲甚至韩国等地的建议。

08:08-08:26 - Researcher: That's really helpful context—so you're active in about ten different E46-specific groups, some with tens of thousands of members globally. Have you ever monetized any of this content or activity? Like, do people pay you for advice, or have you considered creating paid guides, or anything like that? Or is it purely community contribution at this point? - 这个背景信息非常有帮助——你活跃在大约十个不同的 E46 专属小组中，其中一些在全球拥有数万名成员。 你有没有把这些内容或活动变现过？比如，有人付钱请你提供建议吗，或者你考虑过制作付费指南之类的吗？或者到目前为止纯粹是社区贡献？

08:37-08:55 - David: Purely community at this point. I'd love to monetize it, but I just haven't I didn't have time. Um, my main job is pretty time consuming. So just getting the time to do the work on the car is difficult enough. Spending the extra time to properly document it and video it and all that sort of stuff with the side of it would double the time it takes to do the work. So I just don't have the time. - 目前纯粹是社区性质。我很想把它变现，但我只是没有，我没有时间。嗯，我的主职工作相当耗时。所以仅仅抽出时间来修车就已经够难了。再花额外的时间去妥善记录、拍视频之类的事情，会增加一倍的工作量。所以我只是没时间。

08:58-09:21 - Researcher: That makes total sense—documenting everything properly would basically double your workload, and time is already tight. Alright, now let's shift gears to AI. You mentioned earlier that you've been heavily using AI LLMs for your BMW projects in recent months, which sounds really interesting. What are the main AI apps or web tools you use most often in your life—whether for work, the car, or anything else? And what's your experience been like with them? - 这完全理解——妥善记录一切基本上会使你的工作量翻倍，而你的时间已经很紧张了。 好了，现在让我们把话题转到 AI 上。你之前提到，近几个月你一直在大量使用 AI 大语言模型来做宝马项目，这听起来很有趣。在生活和工作中，你最常用的 AI 应用或网页工具是什么？你的使用体验如何？

09:22-10:34 - David: So, um, for work, I have a chat GPT enterprise license. And we also use Microsoft Copilot. So they're the tools that I use officially for work. However, for private projects, I use more Gemini and Claude. And the main reason for that is I wanna experience the different models. You don't know which one is gonna dominate there. They keep kind of outpacing each other. So I wanna as a software person and as a user experience, professional, I like to experience a range of the different models that are available and, you know, make choices about which ones I use. I do have a paid Gemini license as of a few weeks ago, and my goal is to use that specifically for the car. To try to develop some gem type application that I don't have to keep telling it information about my car. I want it to automatically know all about what car I'm talking about, and so I can just ask questions I believe that would be a lot faster. - 所以，嗯，工作上，我有 ChatGPT 企业版许可证。我们也使用 Microsoft Copilot。这些是我官方用于工作的工具。不过，对于私人项目，我更多地使用 Gemini 和 Claude。主要原因是我想体验不同的模型。你不知道哪一个会占据主导。它们不停地相互超越。所以作为一个软件从业者和用户体验专家，我喜欢体验各种不同的可用模型，并据此决定使用哪一个。几周前我买了一个付费的 Gemini 许可证，我的目标是专门把它用于那辆车。尝试开发某种类似 Gem 的应用，这样我就不必一直向它输入我那辆车的信息。我希望它能自动了解我谈论的是哪辆车，这样我就可以直接问问题，我相信这样会快得多。

10:37-11:04 - Researcher: Interesting—so you're strategically spreading across different models to experience how they each evolve, and you're thinking about building a custom gem that already knows your car's specs so you don't have to re-explain context every time. Let me make sure I've got the paid ones clear: you mentioned a paid Gemini license you got a few weeks ago. Are you also paying for Claude, or just using the free version? And with ChatGPT, the enterprise license is through work, but do you have any personal paid ChatGPT subscription too? - 有意思——所以你很有策略地分散使用不同的模型，以体验它们各自的进化过程，并且你正在考虑构建一个自定义的 Gem，它已经了解你车的规格，这样你就不用每次都重新解释背景。 让我确认一下付费项目：你提到了几周前购买的付费 Gemini 许可证。你是否也为 Claude 付费，还是只使用免费版本？至于 ChatGPT，企业许可证是通过公司获得的，那么你个人还有付费的 ChatGPT 订阅吗？

11:08-11:47 - David: The Claude is the free license. The ChatGPT enterprise is paid through work. We have an enterprise license for about fifty seats, and we're giving them out to certain people in the business. To encourage them to start experimenting with productivity type tasks. And the Copilot license is also paid through work as part of our Microsoft arrangement. And I do have a personal free chat GPT account as well that I, uh, I don't really use anymore. Now that I have the Gemini. - Claude 是免费许可证。ChatGPT 企业版是由公司付费的。我们有一个大约五十个席位的企业许可证，我们正把它们发放给业务部门的某些人。以鼓励他们开始尝试提高生产力的任务。Copilot 许可证也是作为我们与微软协议的一部分由公司支付的。我也有一个个人的免费 ChatGPT 账号，但我，呃，现在已经不怎么用了。因为有了 Gemini。

11:49-11:57 - Researcher: Got it. So for paid AI tools, you personally pay for Gemini, and work covers ChatGPT Enterprise and Microsoft Copilot. Can you tell me the specifics on - 明白了。所以对于付费 AI 工具，你个人支付 Gemini 的费用，公司负责 ChatGPT 企业版和 Microsoft Copilot 的费用。 你能告诉我也许具体的

11:57-11:58 - David: Yes. - 是的。

12:00-12:08 - Researcher: Can you tell me the specifics on your Gemini subscription—what tier is it, and how much are you paying monthly or annually for it? - 你能告诉我 Gemini 订阅的具体情况吗——它是什么档次的，每月或每年支付多少钱？

12:10-12:48 - David: I don't know exactly. It's bundled into a Google One package that I recently upgraded to. There was an AI add on So I combined my my Google account with my Nest Home account and added the AI option into a single subscription now about two weeks ago. That's what I'm using. So I have my Gmail. I have my Nest, and now I have my Gemini all in the one I believe it's nineteen ninety nine a month. But I am on a a promotional rate for a few months. - 具体多少我不知道。它是捆绑在我最近升级的 Google One 套餐里的。有一个 AI 附加组件，所以我把我的谷歌账号和我的 Nest 家庭账号合并了，并在大约两周前将 AI 选项加入了一个单一的订阅包。这就是我现在使用的服务。所以我现在有我的 Gmail、我的 Nest，现在还有了 Gemini，都在一起。我想是每月 19.99 美元。但我现在前几个月是享有促销价的。

12:51-13:11 - Researcher: Okay, so it's bundled into Google One at $19.99month (with a promo rate right now), and that includes Gmail storage, Nest, and Gemini Advanced all together. Besides Gemini, are there any other AI tools or apps you're personally paying for? I'm thinking about things like—chat assistants, content generation tools (like image or video AI), productivity tools with AI features, or any other category? - 好的，所以它捆绑在 Google One 中，每月 19.99 美元（目前有促销价），其中包括 Gmail 空间、Nest 和 Gemini Advanced。 除了 Gemini，还有其他你个人付费的 AI 工具或应用吗？我想的是类似——聊天助手、内容生成工具（如图像或视频 AI）、带有 AI 功能的效率工具，或者任何其他类别？

13:31-14:05 - David: Have not done anything really with image or video generation. I played around a little bit. With image. I I don't remember which tool I was using at the time, but I did generate some interesting BMW E46 logos. That just for fun. But otherwise, the only other one I really use I paid for a Notion AI license for about a year. Because they were a very early um, that was one of the earliest, uh, AI enabled tools that there was. And I wanted to experience that. And I use that for content generation. Marketing, content, social media, posting content type type stuff for work. - 我在图像或视频生成方面还没怎么做。我稍微玩了一下。图像方面。我不记得当时用的是哪个工具了，但我确实生成了一些有趣的宝马 E46 标志。纯属好玩。除此之外，我唯一真正用过的就是订阅了一年左右的 Notion AI 许可证。因为他们是很早期的，嗯，那是最早具备 AI 功能的工具之一。我想体验一下。我把那个用于内容生成。市场营销内容、社交媒体、工作方面的发帖内容之类。

14:08-14:16 - Researcher: So you paid for Notion AI for about a year to experiment with it for work content generation. Are you still paying for that now, or did you stop at some point? - 所以你付费订阅了一年的 Notion AI，尝试将其用于工作内容的生成。你现在还在付费吗，还是已经在某个时间点停止了？

14:19-14:40 - David: I stopped. I got a lot of benefit out of it. We changed kinda strategy at work a little bit, so there was not as much need for the content external, um, social media content generation. So I stopped paying for it, and I stopped using Notion. I kind of consolidated everything back into Microsoft OneNote, which the company pays for. - 停了。我从中受益匪浅。但后来公司的工作策略变了一点，不再需要那么多对外的，嗯，社交媒体内容生成。所以我就不再续费了，也不用 Notion 了。我把一切都重新整合到了 Microsoft OneNote 中，那是公司付钱的。

14:51-15:11 - Researcher: Got it—so you used Notion AI for about a year, found it valuable, but then work strategy shifted and you consolidated back to OneNote. So just to confirm: right now, the only AI tool you're personally paying for is Gemini Advanced (bundled in Google One at $19.99month). Everything else—ChatGPT Enterprise, Copilot—is covered by work. Is that right, or am I missing any other paid AI subscriptions? - 明白了——你用了一年 Notion AI，觉得很有价值，但后来由于工作策略转变，你又回到了 OneNote。 所以确认一下：目前你个人付费的唯一 AI 工具是 Gemini Advanced（打包在每月 19.99 美元的 Google One 中）。其他所有工具——ChatGPT 企业版、Copilot——都由公司承担。对吗，还是我遗漏了其他付费 AI 订阅？

15:11-15:12 - David: No. That's correct. - 不。完全正确。

15:14-15:26 - Researcher: Perfect. Now, thinking about your experience with these paid AI tools—has there been any AI purchase or subscription that you've been particularly satisfied with recently? Something where you felt like, "This was really worth the money"? - 太棒了。现在，回想一下使用这些付费 AI 工具的经历——最近有没有哪次 AI 购买或订阅让你特别满意？让你觉得“这钱花得真值”？

15:29-16:38 - David: I haven't, um, done enough with the Gemini advanced license yet to really to really experience the difference. Between the free version and the paid version, but I'm pretty sure that will happen in the next couple of weeks as I get deeper into the gem creation The ChatGPT enterprise license, though, has been a real eye opener. The ability to create GPTs for particular work tasks things that people do manually and that are time consuming today. We've picked a set of those tasks and are generating GPTs. That will basically do that work. And was able to build two of them in in just a couple hours. I now have five that are kind of in a tuning phase, and the results have been quite amazing. So that has been a very good experience. - 我目前，嗯，对 Gemini Advanced 许可证的使用还没多到能真正体验出差别。在免费版和付费版之间还没怎么对比，但我非常肯定就在接下来的几周内，随着我深入研究 Gem 的创建，这种差别会显现出来。不过，ChatGPT 企业版确实让我大开眼界。它能够为特定的工作任务创建 GPT——那些现在人们手动操作且非常耗时的事情。我们挑选了一系列这类任务并正在生成 GPT。它们基本上能完成那些工作。而且能在短短几个小时内就建立起其中两个。我现在有五个处于微调阶段，结果相当惊人。所以那次体验非常好。

16:39-16:52 - Researcher: That's a strong endorsement—building five custom GPTs in just a couple hours and seeing real productivity gains. Can you give me a specific example of one of those GPTs? Like, what manual task does it replace, and how much time does it actually save people? - 这是一个非常有力的认可——短短几小时内就建立了五个自定义 GPT，并看到了真正的生产力提升。你能给我举一个关于这些 GPT 的具体例子吗？比如，它取代了什么手动任务，实际为人们节省了多少时间？

17:02-18:21 - David: A couple of good examples related to contracts. We do a lot of work with HR vendor contracts. Signing a contract with a vendor to purchase something. And quite often, one of one of my SaaS products that we developed and that I own is a contract data tool. A contract data repository and an an analysis tool. And the biggest challenge is getting contract data into that tool. Because the contract data lives in a unstructured format in a PDF file. So getting that data out of a PDF file and into a database type structure. If it's done manually, one contract will take about twenty five minutes. To do by hand. I used to do it by hand, so I have calibrated that. Uh, now with the GPT that I built, still tuning, but it it kinda happens in, like, a minute. So the ability to process, um, a large number of contracts very in a very short period of time and now spend that time that I used to spend copy and pasting data. I can spend that time on making sure the quality of the ingested data is much better. - 有几个和合同相关的典型例子。我们在 HR 供应商合同方面有很多工作。比如和某个供应商签订合同购买东西。通常情况下，我开发并拥有的一个 SaaS 产品就是一种合同数据工具。一个合同数据存储库和分析工具。最大的挑战是将合同数据录入该工具。因为合同数据以非结构化的格式存在于 PDF 文件中。手动将数据从 PDF 文件提取到数据库类型的结构中。如果手动操作，一份合同大约需要 25 分钟。纯手工操作。我以前亲手做过，所以我算过。呃，现在用我建立的这个 GPT，虽然还在微调中，但它大概只要一分钟就能搞定。所以，在极短的时间内处理大量合同的能力，让我可以把以前用来复制粘贴数据的时间，花在确保摄入数据的质量，使其变得更好。

18:23-18:44 - Researcher: Wow—from 25 minutes per contract down to about a minute. That's a massive time shift, and you're reinvesting that saved time into quality control instead of just copy-pasting. So that GPT is built on ChatGPT Enterprise. When you think about the value you're getting from that enterprise license for your team—those 50 seats—do you have a sense of what the company is paying for it, or roughly what the cost per seat is? - 哇——从每份合同 25 分钟缩短到大约一分钟。这是一个巨大的时间跨度转变，你把省下的时间重新投入到了质量控制中，而不是仅仅重复劳动。 所以那个 GPT 是构建在 ChatGPT 企业版上的。当你想到这个企业版授权为你团队（那 50 个席位）带来的价值时——你知道公司为此支付了多少钱，或者大致的单价是多少吗？

18:49-19:12 - David: I don't off the top of my head know. Um, I know somewhere in in the in my files, I've got it. But don't know exactly what it is. I think we it's hard to calculate because we do have some subsidization from Google, a relationship we have with Google Cloud. So the actual cost is, um, is reduced or offset by some discount like that we have. - 我脑子里一下子想不起来具体数字。嗯，我知道在我的文件里的某个地方有。但我确实不知道具体是多少。我想这也是因为我们得到了谷歌的一些补贴，我们和 Google Cloud 有合作关系。所以实际成本是，嗯，通过类似那样的折扣被降低或抵消了。

19:17-19:33 - Researcher: That makes sense—enterprise pricing often has those kinds of offsets and custom deals, so it's not always straightforward. Alright, flip side: has there been any AI app or product purchase that left you feeling disappointed or unsatisfied? Something where you paid for it and thought, "This wasn't worth it" or "I expected more"? - 这说得通——企业版定价通常包含这类抵扣和定制协议，所以并不总是那么直接。 好，反过来看：有没有哪次 AI 应用或产品的购买让你感到失望或不满意？比如你付了钱却觉得“这不值”或者“我期望它更高”？

19:47-20:20 - David: Not yet. I I expect there will be. I've been very conservative about paying for anything. Um, I my basic approach is to use the free version until I need more functionality than I have to pay for. Honestly, I was able to use Gemini and Claude and ChatGPT for about two years. Without paying anything. You know, my the sophistication of what I was trying to do with them was though they were enough. To to to do what I needed to do. Now as I need to do higher level things, I just have to start paying for the license. So so far, no disappointment. - 还没有。我预感会有的。我对付费一直很谨慎。嗯，我的基本策略是先用免费版，直到我需要付费才能实现的功能。诚实地说，Gemini、Claude 和 ChatGPT 我免费用了大约两年。没花一分钱。你知道，当时我想通过它们实现的事情，免费版也够用了。足以完成我需要做的事。现在随着我需要处理更高层级的事情，我才不得不开始支付许可费。所以到目前为止，还没有失望。

20:22-20:51 - Researcher: That's a really disciplined approach—use free until you hit the ceiling, then upgrade only when you need to. Let's dive deeper into each of these tools you're using for work and productivity. You've mentioned ChatGPT Enterprise, Copilot, Gemini, and Claude. Let me start with ChatGPT Enterprise since you've had such strong results with it. You already told me about the contract data extraction GPT. Can you walk me through one more recent example—something you've used ChatGPT Enterprise for in the last week or two? What was the task, and what kind of prompt or input did you give it? - 这是一个非常自律的方式——先用免费的直到触及天花板，只有在真正需要时再升级。 让我们深入探讨一下你在工作和效率方面使用的工具。你提到了 ChatGPT 企业版、Copilot、Gemini 和 Claude。让我从 ChatGPT 企业版开始，因为你用它取得了显著的成果。 你刚才谈到了合同数据提取 GPT。你能带我回顾一下另一个最近的例子吗——你在过去一两周内用 ChatGPT 企业版做过的事情？任务是什么，你给了它什么样的提示词（prompt）或输入？

20:59-22:53 - David: Yeah. Another good one is again, working with contracts. Over a period of time in a relationship with between our company and another vendor. You get a series of different type of contract documents that are put in place and signed. You might have some master agreement and a couple of amendments over time, maybe some statements of work or some order forms. After a couple of years, you could have six or eight different signed different types of signed contracts, between the same two companies. And your relationship is not clear based on looking at any one of those documents. You really have to look at all of those documents together. And one of the things that we need in my company when we are negotiating contracts for our customers. Is to get a very clear view of what that relationship looks like. So the other great GPT that I built, you basically, I feed it a bunch of contracts from between the same two companies. That represent that relationship And what the GPT does is it merges all of those documents together. And creates simulates a single document that represents everything in that relationship. So if a if a particular term was updated, by an amendment, the updated version is what would show in that document not the original version of that term. And I was quite surprised at how accurate that type of work can be. Very good results, and now I'm working on tuning that. And then I'm gonna make it available to the business so that they can start using it. - 好的。另一个不错的例子还是和合同相关的。在我们的公司和另一个供应商的长期合作关系中。你会签署并积累一系列不同类型的合同文件。你可能有一个主协议，过段时间会有几个修正案，也许还有一些工作说明书或订购单。几年后，两家公司之间可能会有六到八份不同的签约合同。如果你只看其中任何一份文件，你的合作关系并不清晰。你真的必须把所有这些文件放在一起看。我们在为客户协商合同时，正好需要这一点。就是要对这种关系有一个非常清晰的认识。所以我建立的另一个很棒的 GPT，基本上是你喂给它两家公司之间的一堆合同。这些合同代表了那段关系。GPT 所做的是将所有这些文件整合在一起。并模拟出一份代表该关系中所有内容的单一文件。因此，如果某个特定条款后来被修正案更新了，那么这个文件中显示的就是更新后的版本，而不是该条款的原始版本。我对这类工作的准确性感到非常惊讶。结果非常好，我现在正在进行微调。然后我将把它开放给业务部门，让他们开始使用。

22:56-23:16 - Researcher: That's incredibly sophisticated—merging multiple contract versions, tracking amendments, and showing the current state of each term. That's exactly the kind of complex, high-stakes work where accuracy really matters. When you say you're "tuning" it now, what does that tuning process actually look like? Are you feeding it test cases and correcting mistakes, or adjusting the prompts, or something else? - 这非常复杂且专业——整合多个合同版本、追踪修正案，并显示每个条款的现状。这正是那种准确性至关重要、复杂且高风险的工作。 当你提到你现在正在“微调（tuning）”它时，那个微调过程具体是什么样的？你是给它测试案例并纠正错误，还是在调整提示词，或者是别的什么方式？

23:18-24:28 - David: Yeah. I'm adjusting the prompts. And it's quite cool how you can kind of have two windows, and you can adjust the prompt and then update it and run it again and compare the output. So I'm in that kind of tuning iteration cycle. I'm also looking to add different files to the knowledge in the GPT so that it actually has examples of what good looks like that it can refer to. So I'm probably halfway through that tuning. There's a lot of things I haven't done yet that I expect I'll get done in the next week. But that's the that's what I call tuning anyway. Just looking at the output, changing something, rewording the prompt, adding some knowledge. You know, I have not yet started plugging in other applications or data sources via like, a connection of some kind, like an API. That's something I wanna do next. But for now, it's working pretty well. - 是的。我正在调整提示词。而且你可以并排开两个窗口，调整提示词，更新它，再次运行并对比输出，这非常酷。所以我处于那种微调迭代循环中。我也在考虑在 GPT 的知识库中增加不同的文件，这样它就有了“标杆案例（good looks like）”可以参考。所以我现在的微调进度估计完成了一半。还有很多事没做，我预计下周能搞定。但这就是我所说的微调。只是观察输出，改变一些东西，重写提示词，增加一些知识储备。你知道，我还没开始通过像 API 连接之类的方式插入其他应用或数据源。那是我接下来的计划。但就目前而言，它运行得相当不错。

24:30-24:50 - Researcher: So you're building up a knowledge base of "good examples" to guide it, iteratively refining prompts, and planning to connect external data sources next. That's a pretty methodical build process. When you think about the contract merging GPT versus the data extraction one—are these tools you're building primarily for yourself to use, or are they meant for your team of 60 to use across the board? - 所以你正在建立一个“优秀案例”知识库来引导它，不断改进提示词，并计划下一步连接外部数据源。这是一个非常系统化的建设过程。 当你想到合同整合 GPT 和数据提取 GPT 时——这些工具主要是为你自己建立的，还是打算供你那 60 人的整个团队使用的？

24:53-26:19 - David: They are primarily proof of concepts So I'm a I'm a while I was an engineer thirty years ago, I'm not an engineer anymore. I'm a product person. And my main job right now is to prove that this technology can do the types of things that we need it to do. So what I'm building are prototypes of that capability. So the GPT that I build is just a very easy way for me to rapidly prototype functionality. How that functionality then gets delivered to the users in the company. Is the next step. I believe there's different form factors that we can use to deliver. To different audiences. That same basic functionality, might be a GPT. For some people, it might need to be evolved into being a web app. For some other people, it may need to become a function inside of the SaaS product that we have. That's accessed via the SaaS user interface. There's all kinds of potential form factors for delivering that functionality. But what I'm focused on right now is can the technology do the job? And if it can do the job, we move on to the next step. - 它们主要是概念验证（PoC）。虽然我三十年前是个工程师，但我现在已经不是了。我是一个产品人。而我目前的主要工作是证明这项技术可以完成我们需要它完成的任务。所以我建立的是这种能力的原型。因此，我建立的 GPT 只是让我能快速原型化功能的一种非常简便的方式。至于这些功能随后如何交付给公司的用户，是下一步的事。我相信有不同的形态（form factors）可以用来向不同的受众交付。同样的基础功能，对于某些人来说可能是一个 GPT。对于另一些人，它可能需要演变成一个 Web 应用。对于其他人，它可能需要成为我们 SaaS 产品内部的一个功能，通过 SaaS 的用户界面访问。交付这些功能有很多潜在的形态。但我现在关注的是技术能否胜任这项工作？如果能胜任，我们就进入下一步。

26:21-26:41 - Researcher: Got it—you're the R&D arm proving feasibility before the engineering team figures out how to productize it. That makes a lot of sense given your role. So when you prove a concept works—like the contract merging GPT—what happens next? Who decides which form factor it becomes, and how long does it typically take to go from your prototype to something the team or customers can actually use? - 明白了——你相当于研发部门，在工程团队决定如何将其产品化之前，先通过原型证明可行性。考虑到你的角色，这非常有道理。 当你证明一个概念可行时——比如合同整合 GPT——接下来会发生什么？谁来决定它变成何种形态？从你的原型到团队或客户真正能用的东西，通常需要多长时间？

26:44-27:35 - David: I would I'll engage my product management team assign a product manager to take over the give them the the proof of concept and take over now the next step, which is thinking about who the user personas are, And there may be different parts of the company that are actually gonna use that tool in different ways. So we treat them as different personas, and then we start mapping out what the what the workflow actually looks like for different personas. Because that's what's gonna really dictate what the form factor is, whether a GPT is good enough. Or whether they may use this thing, you know, all day every day to do their job, in which case, a GPT is probably not good enough. We probably need something more. More of an application. - 我会启动我的产品管理团队，指派一名产品经理接手这个概念验证，现在接手下一步，也就是思考用户画像（user personas）是谁。公司的不同部门可能会以不同的方式使用该工具。所以我们将他们视为不同的画像，然后我们开始描绘不同画像的实际工作流是什么样的。因为这才是真正决定形态的因素，是 GPT 就足够好了，还是他们可能整天都在用这玩意儿工作，如果是那样的话，GPT 可能就不够好了。我们可能需要更像一个应用的东西。

27:38-27:55 - Researcher: That's a really clear handoff—you prove it works, product management figures out the personas and workflows, then that dictates the delivery method. When you think about the people who might use this "all day every day"—like the contract merging tool—who in your company would that be? What's their role, and how would this change their daily work? - 这是一个非常明确的交接过程——你证明它行得通，产品经理确定画像和工作流，然后决定交付方式。 当你想到那些可能“整天”使用它的人时——比如合同整合工具——公司里会是谁呢？他们的职位是什么，这会如何改变他们的日常工作？

28:11-28:33 - David: The main target for that would be an HR manager, somebody who is running vendor integrations and they may run twenty projects simultaneously. And so if they could use a tool like that to prepare for those twenty projects, their preparation time would be much, much shorter. And it it would be more accurate. They may even be able to do more projects at the same time. If the prep if the prep time is, uh, is reduced. - 这方面的主要目标受众是人事经理（HR manager），或者是负责供应商整合的人。他们可能同时运行着 20 个项目。因此，如果他们能使用这样的工具来为这 20 个项目做准备，他们的准备时间会大幅缩短，而且会更准确。如果准备时间减少了，他们甚至可能同时处理更多的项目。

28:36-28:57 - Researcher: So HR managers juggling 20 vendor integrations at once—if prep time drops significantly, they could either handle more volume or just have way more headspace for the actual negotiation and strategy work. Right now, without this tool, how long does that prep work typically take them per project? And what does that prep process look like—are they manually reading through all those contract documents themselves? - 所以，那些同时处理 20 个供应商整合的人事经理——如果准备时间显著下降，他们要么可以处理更大的工作量，要么能有更多的精力去进行实际的谈判和战略工作。 现在，如果没有这个工具，每个项目通常需要他们花多长时间做准备？准备过程是怎样的——他们是自己在手动阅读所有这些合同文件吗？

29:17-29:50 - David: In most cases, they're not doing the full prep because they don't have time. I used to do that role. And when I was doing that role, I actually did this. I made sure I did it, and it would take me a couple of hours, um, maybe even half a day. To get through reading you know, five to ten contract documents and tracing back, which which clauses had been, you know, superseded and what what the current in force set of t's and c's might look like. So I could easily cut three to four hours of work down to half an hour. And in some cases, it's just not getting done. So that would actually doing it would actually improve the process even further. - 在大多数情况下，他们并没有做充分的准备，因为他们没有时间。我以前也做过那个岗位。当我在做那个工作时，我确实会去做准备。我会确保自己做到位，但这会花掉我几个小时，甚至半天的时间。你需要读完 5 到 10 份合同文件，并追溯哪些条款已被取代，以及当前生效的条款和条件（T&C）到底是什么样的。所以我可以轻松地将 3 到 4 小时的工作量缩减到半小时。而在某些情况下，这项工作根本就没人做。因此，如果能真正把这件事做起来，将进一步改进流程。

29:52-30:12 - Researcher: Wow, so you're talking about cutting 3-4 hours down to half an hour—and in many cases people are just skipping that prep entirely because they don't have the time. That's a massive unlock, especially if it means they're going into negotiations better informed. When you were doing that role and you did take the time to do the full prep—what advantage did that give you compared to colleagues who skipped it? - 哇，所以你是在说将 3-4 小时缩减到半小时——而且在很多情况下，人们因为没时间就完全跳过了准备环节。这是一个巨大的突破，特别是如果这意味着他们在谈判时能掌握更充分的信息。 当你还在那个岗位上，并且花时间做了充分准备时——与那些跳过准备的同事相比，这带给你了什么优势？

30:17-30:51 - David: I was able to walk into the vendor negotiation much better armed. I have one very good example where I did the work. It was it was, like, fifteen contracts over ten years, and I took half a day. I did the work properly. I walked into the negotiation. The vendor was totally off guard. Um, did not expect us to have a good understanding of the contracts, and we were easily able to negotiate a much better deal for for our client. Because of that preparation was done. So huge value. - 我走进供应商谈判现场时，武装得更充分。我有一个非常好的例子，那次我做了功课。那是横跨 10 年的大约 15 份合同，我花了半天时间。我扎实地完成了准备工作，然后走进谈判室。供应商完全措手不及。他们没料到我们对合同有如此深入的了解，因此我们能够轻松地为客户谈成条件好得多的交易。这完全归功于那次准备工作。所以价值巨大。

30:54-31:15 - Researcher: That's a perfect example—the vendor wasn't expecting you to have done your homework, and that asymmetry gave you real negotiating leverage. Sounds like it translated directly into better commercial outcomes. When you think about rolling this out to your HR managers, do they recognize that gap in their process? Like, do they know they should be doing this prep but can't, or have they just adapted to working without it? - 这是一个完美的例子——供应商没想到你做了功课，这种信息不对称给了你真正的谈判筹码。听起来这直接转化为了更好的商业成果。 当你考虑向你的人事经理推广这个工具时，他们是否意识到了流程中的这个缺口？也就是说，他们知道自己应该做准备但做不到，还是说他们已经适应了不做准备的工作方式？

31:15-31:40 - David: I think most people tend to take the easy path there. They're getting away with not doing it. I don't think anyone's noticing because I don't know that the leadership is fully aware either. The upside is actually, is even higher because I don't think most people are aware that there's a big opportunity here. - 我认为大多数人倾向于走捷径。他们不做准备也混过去了。我不认为有人注意到这一点，因为我不知道领导层是否也完全意识到这个问题。其实好的一面是，潜力甚至更高，因为我觉得大多数人还没意识到这里存在着巨大的机遇。

31:42-32:07 - Researcher: So it's not even a recognized pain point yet—people have just normalized skipping the prep because no one's measuring the cost of not doing it. That's interesting, because it means when you roll this out, you're not just solving a known problem, you're revealing an opportunity they didn't know they were missing. How do you plan to surface that value to leadership or the HR managers themselves? Like, what would convince them this is worth adopting if they don't currently see it as a gap? - 所以这甚至还不是一个公认的痛点——人们已经把跳过准备视作常态，因为没有人去衡量不这样做的成本。这很有趣，因为这意味着当你推广它时，你不仅是在解决一个已知的问题，你还在揭示一个他们以前没意识到自己正在错失的机会。 你计划如何向领导层或人事经理展示这种价值？如果他们目前不认为这是一个缺口，什么能说服他们这值得采用？

32:17-32:35 - David: That you have to do that top down. So the the results of this type of POC work, uh, gets shared with the executive team first. And the executive team will be very interested in anything that drives productivity in our business model. So it's gotta go top down. I tried to do it bottom up for years, and people just don't don't wanna do the extra work. So gonna have to go top down. - 你必须采取自上而下的方式。所以这种概念验证（POC）工作的结果，首先会分享给执行团队。执行团队会对任何能提高我们业务模式生产力的东西非常感兴趣。所以必须自上而下。我尝试过多年自下而上地推动，但人们就是不想做额外的工作。所以必须走自上而下的路线。

32:38-32:55 - Researcher: That makes total sense—if people are already "getting away with it," bottom-up adoption is a non-starter. Executive mandate based on productivity gains is the only way it actually gets used. You mentioned you tried bottom-up for years—was that with AI tools specifically, or other process improvements? What did you learn from those failed attempts? - 这完全说得通——如果人们已经“得过且过”了，自下而上的采用根本行不通。基于生产力提升的行政指令是让它真正被投入使用的唯一途径。 你提到你尝试自下而上推动了很多年——那是针对具体的 AI 工具，还是其他的流程改进？你从那些失败的尝试中吸取了什么教训？

33:00-33:05 - David: No. There was no AI. That was roll up your sleeves and do the work manually. The way that I used to do it - 全都没有。那时候还没有 AI。那就是撸起袖子干体力活，手动去做。就像我以前做的那样。

33:08-33:13 - Researcher: Ah, so you were trying to get people to adopt the manual best practices - 啊，所以你当时是想让人们采用手动的最佳实践。

33:13-33:13 - David: Yes. - 是的。

33:15-33:30 - Researcher: —the ones that took you half a day per project. No wonder that didn't stick. People saw the effort, not the payoff. Now with the AI doing the heavy lifting, do you think adoption will be easier just because the time investment is so much lower? Or will you still face resistance even if it only takes 30 minutes? - ——就是那些每个项目要花掉你半天时间的实践。难怪推行不下。人们只看到了付出的努力，没看到回报。 现在由 AI 来承担繁重的工作，你认为仅仅因为时间投入大大降低，采用起来会更容易吗？或者，即使只需要 30 分钟，你仍然会面临阻力？

33:37-33:50 - David: No. I think the adoption will become obvious because once it's clear to the senior leadership, then it will it will be measured And if people aren't doing it, then you know, that that would be that would be revealed in the measurements. - 不。我认为采用会变得顺理成章，因为一旦高层领导明确了这一点，它就会被衡量。如果人们不去做，那么，你也知道，在衡量数据中就会暴露出来。

33:52-34:07 - Researcher: So it becomes a KPI—leadership tracks who's using it, and suddenly it's not optional anymore. The measurement creates the accountability. What metrics do you think leadership will care about most? Is it time saved, number of projects handled, negotiation outcomes, or something else? - 所以它变成了 KPI——领导层追踪谁在使用，突然之间它就不再是可选的了。衡量产生了问责制。 你认为领导层最关心的指标是什么？是节省的时间、处理的项目数量、谈判的结果，还是别的什么？

34:25-34:57 - David: I think it's a mixture of all of those. Some of those are quantitative. Some of them are qualitative. And there's there's potentially the consistency of good quantitative results that would be one measurement, not just you know, good results here and there, but consistently getting good results. We measure results in the savings that could be generated from a negotiation. So it's pretty easy to measure. And then from a more qualitative perspective, can kind of keep track of, you know, who's using the techniques and who's not. And I think it it'll be a mixture of quantitative and qualitative KPIs. - 我认为这些都会涉及。其中一些是定量的，一些是定性的。可能还会有一个指标是良好定量结果的一致性，而不仅仅是这里或那里偶尔出现的好结果，而是持续获得好结果。我们通过谈判产生的节省金额来衡量结果，所以这很容易通过数字衡量。然后从更定性的角度，可以跟踪谁在使用这些技术，谁没有使用。我认为这将是定量和定性 KPI 的结合。

34:59-35:19 - Researcher: That's a really solid measurement framework—savings per negotiation is concrete, and consistency over time shows it's not just luck. The qualitative tracking of who's actually using the techniques adds the behavior layer. When you think about the contract merging GPT specifically—have you shown it to any HR managers yet informally, or is it still just in your hands at the POC stage? - 这是一个非常扎实的衡量框架——每次谈判的节省是具体的，而长期的连贯性说明这不仅仅是运气。对实际使用技术的人员进行定性跟踪，则增加了行为层面的考量。 当你具体想到“合同合并 GPT”（contract merging GPT）时——你是否已经非正式地向任何人事经理展示过它，还是它目前仍然只是在你的 POC 阶段？

35:25-35:53 - David: It's POC right now. The next um, I'm going to review it with our legal team actually first. Because they have a more pure contract perspective, and I'd like to make sure that it's not just my little technique. I wanna make sure that it's kind of validated by the legal team. And if they think it's best practice as well, and could actually potentially use it, then that gives me the extra endorsement when I take it to the leadership - 现在还是 POC。下一步，实际上我要先和我们的法务团队进行评审。因为他们有更纯粹的合同视角，我想确保这不仅仅是我个人的小技巧。我想确保它得到了法务团队的验证。如果他们也认为这是最佳实践，并且实际上也可能使用它，那么当我把它提交给领导层时，就多了一份背书。

35:56-36:14 - Researcher: Smart move—getting legal to validate it first gives you credibility and de-risks it from a compliance perspective. If they endorse it as best practice, that's a much stronger case than just "this worked for me." What's your sense of how legal will react? Are they generally open to AI tools, or is this going to be a harder sell given the sensitivity of contract data? - 聪明的做法——先让法务验证能给你带来公信力，并从合规角度降低风险。如果他们背书这是最佳实践，那比起仅仅说“这对我有用”要有力得多。 你觉得法务会如何反应？他们对 AI 工具通常持开放态度吗？还是说考虑到合同数据的敏感性，这会是一个很难说服的过程？

36:24-36:39 - David: Legal are very scared of AI tools. That's why it would also be a good a good test because if if they can see how useful and beneficial this is, for their own work. Then I think that could help the overall resistance to adopting AI that you tend to get from a conservative legal team. - 法务非常害怕 AI 工具。这也是为什么这是一个很好的测试，因为如果他们能看到这对于他们自己的工作是多么有用和有益，那么我认为这有助于缓解你通常会从保守的法务团队那里遇到的那种对采用 AI 的普遍抵触情绪。

36:42-36:59 - Researcher: So legal is actually one of your toughest audiences—if you can win them over, that's a strong signal the tool is solid and it helps break down broader AI resistance across the company. What specifically are they scared of? Is it accuracy, confidentiality, liability if the AI gets something wrong, or just general discomfort with the technology? - 所以法务实际上是你最难打交道的受众之一——如果你能争取到他们，这就释放了一个强烈的信号，表明该工具是可靠的，并有助于打破公司范围内更广泛的 AI 阻力。 他们具体在害怕什么？是准确性、保密性、如果 AI 弄错了要承担的法律责任，还是仅仅是对这项技术普遍感到不安？

37:06-37:58 - David: It's confidentiality of the data that's getting put into the AI. And I think also the liability aspect. Remember that in order to do this, we have to feed in confidential documents contracts between two companies into the AI in order for it to do its work. Whether it's extracting the data into fields or whether it's merging the the different contract documents. We're actually we have to expose you know, confidential somewhat confidential documents to the tool. So it took it took over a year to get to the point where we could even get an enterprise license, and we didn't we held off doing this type of POC work until we had a more private environment we get with the enterprise license to actually start testing this stuff. - 是输入 AI 的数据的保密性。我认为还有责任归属方面的问题。记住，为了做到这一点，我们必须将两家公司之间的机密文件、合同喂给 AI，以便它进行工作。无论是将数据提取到字段中，还是合并不同的合同文件，我们实际上都必须向这个工具公开一些半机密的文档。所以，我们花了一年多时间才达到可以获得企业版许可证（enterprise license）的程度，在拥有企业版许可证提供的更私密的环境之前，我们一直没有进行这种类型的 POC 工作，之后才真正开始测试这些东西。

38:01-38:21 - Researcher: Ah, so the year-long wait for ChatGPT Enterprise was specifically about getting a contractually protected environment where you could safely feed in confidential contract data without it leaking into training data or being exposed. Now that you have that enterprise license with the privacy guarantees—does legal feel comfortable with it, or are they still nervous even with those protections in place? - 啊，所以为了得到 ChatGPT 企业版等了一年多，专门是为了获得一个受合同保护的环境，这样你就可以安全地输入机密合同数据，而不用担心它泄露到训练数据中或被曝光。 既然你现在有了带隐私保证的企业版许可证——法务觉得放心了吗？或者即使有了这些保护措施，他们还是很紧张？

38:21-38:35 - David: They're still nervous. They'll always be nervous. But I think over time, they will become less resistant. They will have to And yeah, I think it's, um, it's only gonna get better. - 他们还是很紧张。他们永远会紧张。但我认为随着时间的推移，他们的抵力会变小。他们不得不接受。是的，我认为情况只会越来越好。

38:38-38:56 - Researcher: So it's a gradual trust-building process—they're never going to be enthusiastic, but showing them concrete, safe use cases like the contract merging GPT chips away at the resistance over time. When you eventually do roll this out to HR managers after legal signs off, what does success look like six months in? How will you know it's actually working? - 所以这是一个逐渐建立信任的过程——他们永远不会表现得充满热情，但向他们展示具体的、安全的用例（如合同合并 GPT），会随着时间的推移逐渐瓦解阻力。 在法务批准后，当你最终向人事经理推广这个工具时，六个月后的成功标准是什么？你如何知道它真的起作用了？

39:00-39:27 - David: Adoption, people using it, and then I think also engagement around people wanting to refine it, tune it further. Make it do different things or make it work in slightly different ways. So I think once if we get that sort of feedback, from the users, about not just using it, but wanting to contribute to how it works. I think that's a good sign of adoption. - 采用率，人们在使用它，然后我认为还有关于人们想要改进它、进一步调整它的参与度。让他们想让它做不同的事情，或以稍微不同的方式工作。所以我想，如果我们收到用户这样的反馈，不仅仅是关于使用它，而是想为它的工作方式做出贡献，我认为这就是采用良好的标志。

39:29-39:45 - Researcher: So it's not just usage metrics, but people actually caring enough to suggest improvements and customizations—that's when you know it's become part of their workflow, not just a mandated checkbox. Have you thought about how you'll collect that feedback? Will it be structured, or more organic conversations? - 所以不仅仅是使用指标，而是人们真的在意到足以提出改进和定制建议——这时你就知道它已经成为他们工作流的一部分，而不仅仅是一个被强制要求的复选框。 你有没有想过如何收集这些反馈？是结构化的，还是更随机的对话？

39:46-40:13 - David: I haven't thought about that yet. It's a challenge even with our SaaS products. It's a big enough challenge to keep track of who is using them and getting feedback from those users. We we're not really a software company, so we haven't really engineered a lot of that stuff into the technology when we build it. We have to go back later and add it in. So, don't know how we're gonna do it. It's gonna be a challenge. - 我还没想过这个问题。哪怕是对我们的 SaaS 产品，这都是一个挑战。要追踪谁在使用它们并从这些用户那里获得反馈已经是一个足够大的挑战了。我们并不是真正的软件公司，所以我们在构建技术时，并没有在其中设计很多这类的功能。我们必须稍后再回头添加。所以，我不知道我们会怎么做。这将是一个挑战。

40:15-40:37 - Researcher: That's a real gap—you're building powerful tools but don't have the feedback loops baked in from the start. It sounds like you're learning as you go, which makes sense given you're a managed services company, not a product company. Let me shift gears a bit. We've talked a lot about ChatGPT Enterprise and the custom GPTs. Earlier you mentioned you also have Microsoft Copilot at work—how does that fit into your workflow compared to ChatGPT? - 这是一个真正的缺口——你正在构建强大的工具，但从一开始就没有把反馈闭环设计进去。听起来你是一边做一边学，这也是可以理解的，因为你们是一家管理服务公司，而不是产品公司。 让我们换个话题。我们聊了很多关于 ChatGPT 企业版和自定义 GPT 的话题。早些时候你提到你在工作中也使用 Microsoft Copilot——与 ChatGPT 相比，它如何融入你的工作流？

40:44-42:21 - David: Copilot's used by the developers. So the engineers I have, we have a few different technology stacks. We have six different SaaS products. Across four different technology stacks. We've evolved over time. And so the development team you know, they can't really easily move between one stack and the other stack. The Java developers you know, can't go and start coding in Ruby. Ruby guys go and start coding in Python. Python guys maybe they can do some SQL, but the the Ruby guys can't do the SQL. So there's a there's a fragmentation of the development resources. That we have and enabling them with a a copilot We have copilot in a few different forms. We have GitHub Copilot for the developers, Microsoft Copilot for other, maybe data analysts who aren't really, you know, into the software development side. But enabling all of those people to become more productive using Copilot to generate code test code, critique code, document and comment code, write documentation about what it does, all that stuff. We've already was one of the earliest use cases for us using the the LLM technology. We started probably two and a half years ago. In that area. So we're quite mature. - Copilot 是开发人员使用的。我手下的工程师们，我们有几个不同的技术栈。跨越四个不同的技术栈，我们有六个不同的 SaaS 产品。我们是随着时间的推移演进过来的，所以开发团队并不能轻易地在一个技术栈和另一个技术栈之间切换。Java 开发人员不能直接去用 Ruby 写代码，Ruby 的人不能直接去写 Python。Python 的人也许能写点 SQL，但 Ruby 的人不能写 SQL。所以我们的开发资源是碎片化的。 而为他们配备 Copilot——我们有几种不同形式的 Copilot。我们有给开发人员用的 GitHub Copilot，还有给其他人员，比如可能不从事软件开发的数据分析师用的 Microsoft Copilot。但是，让所有这些人都能通过使用 Copilot 来生成代码、测试代码、评析代码、为代码编写文档和注释、编写关于代码功能的说明文档等等，从而变得更高效。这已经是我们使用大语言模型（LLM）技术最早的用例之一了。我们大概在两年前就开始在那个领域探索了。所以我们在那方面相当成熟。

42:24-42:43 - Researcher: So Copilot has been in play for two and a half years across your dev team—that's early adoption. It sounds like it's become essential for bridging those technology stack silos, letting people work outside their native languages and be more versatile. Have you seen measurable productivity gains from the developers using Copilot, or is it more qualitative—like they can handle more diverse tasks? - 所以 Copilot 已经你的开发团队中应用了两年半——这算得很早的采用了。听起来它已经成为连接那些技术栈孤岛、让人们能够跨越母语限制工作并变得多才多艺的关键。 你是否在使用 Copilot 的开发人员身上看到了可衡量的生产力提升？或者这种提升更多是定性的——比如他们可以处理更多样化的任务？

42:49-43:34 - David: It's a bit of both. So we're we're able to hire less experienced people and have them operate at a higher level. So, like, junior people and have them operating as intermediate people. Because they're augmented by the copilot. That's been a huge gain because it's you know, getting less experienced people obviously costs less money. And also, just with you know, consistency of code that's being generated, rather than have people just make up stuff themselves. They're getting, you know, guidance from the same Copilot system, then can get a a level of consistency across the coding as well. - 两者皆有。我们可以招聘经验较少的人，并让他们在更高的水平上运作。比如，让初级人员像中级人员一样运作，因为他们得到了 Copilot 的增强。这是一个巨大的收获，因为招聘经验较少的人显然成本更低。此外，还可以提高生成代码的一致性，而不是让人们自己凭空捏造。他们从同一个 Copilot 系统获得指导，那么在代码编写层面上也能获得一定程度的一致性。

43:37-43:55 - Researcher: That's a huge unlock—you're essentially using Copilot to level up junior devs to intermediate capability at a lower cost, plus you get more consistent code quality across the team because everyone's guided by the same AI. Has that changed your hiring strategy? Are you now actively seeking out less experienced people knowing you can augment them, or is it more opportunistic? - 这是一个巨大的突破——你本质上是在使用 Copilot 以更低的成本将初级开发人员提升到中级能力，而且由于每个人都受到同一个 AI 的指导，团队的代码质量也更加一致。 这改变了你的招聘策略吗？你现在是在主动寻找经验较少的人，因为你知道可以增强他们的能力，还是说这更多是顺势而为？

44:00-45:19 - David: It we were. It it did impact it, and we were hiring less experienced people. Kind of maybe a year and a half ago. I think what we're doing now is keeping the more experienced people and hiring less of the lower experienced people because a lot of that work now is gonna get automated. Not just not just using the tools to augment the work, but actually start develop I mean, just creating the code from scratch. We've had some good examples recently where experimenting with things like like Claude to just develop an entire application generate the code. We used Claude to do some vibe coding recently as a different POC. And we're basically able to generate all of the Python code for the the back end of the product and all of the JavaScript code for the interface. Without us writing any code. Uh, the person doing it is actually not a programmer. But was a data analyst. And that was a fantastic example. - 我们以前是这样做的。它确实产生过影响，大概一年半以前，我们在招经验较少的人。我认为我们现在的做法是留住更有经验的人，而减少招聘经验较少的人，因为现在很多工作都会被自动化。不仅是使用工具来增强工作，而是开始直接生成代码，我是说从头开始创建代码。 我们最近有一些很好的例子，比如尝试使用 Claude 之类的工具直接开发整个应用程序，生成代码。最近作为另一个 POC，我们用 Claude 做了一些“氛围编程”（vibe coding）。我们基本上能够在不编写任何代码的情况下，生成产品后端的全部 Python 代码和接口的全部 JavaScript 代码。做这件事的人实际上不是程序员，而是一名数据分析师。这是一个极好的例子。

45:22-45:39 - Researcher: Wait, that's a massive shift—from augmenting junior devs to potentially not needing them at all because a non-programmer data analyst can now generate full-stack applications using Claude. Walk me through that specific example. What was the application they built, and how long did it take them compared to what a traditional dev would need? - 等等，这是一个巨大的转变——从增强初级开发人员，转变到可能根本不需要他们，因为一个非程序员的数据分析师现在可以使用 Claude 生成全栈应用程序。 带我了解一下那个具体的例子。他们构建的应用程序是什么？与传统的开发人员相比，他们花了多长时间？

45:50-47:15 - David: The application was working with company names and normalizing those company names and then retrieving information via the LLM that told us what industry the company worked in. And then using a lookup table that we control to enrich that data to basically categorize that company into a three level vendor segment taxonomy so that we knew what vendor segments a particular company works in And that was something that used to be done manually. Someone would basically use Google search to look up each company name. And interpret a bunch of links and then make that decision themselves. So consistency of doing the work was not good. The amount of time it took, it took weeks. Days, or weeks to work through. A file full of company names. That work now is with this application, basically, automated. It's more consistent. And then this is a tool that we will probably have a GPT version of for some people We'll have a web app version of for some people, and we'll probably build it into our SaaS applications so it's used behind the scenes. To do things so that whenever this particular work gets done, it gets done in the same way. - 那个应用程序是对公司名称进行处理和规范化，然后通过 LLM 检索信息，告诉我们该公司从事什么行业。接着，使用我们控制的查找表来丰富该数据，基本上将该公司分类为三级供应商细分分类法，以便我们了解特定公司在哪些供应商细分市场中运营。 这在以前是手动完成的。有人会用谷歌搜索来查找每个公司名称，解读一堆链接，然后自己做决定。所以工作的连贯性并不好。费时费力，需要几天甚至几周的时间才能处理完一个充满公司名称的文件。现在有了这个应用程序，这项工作基本上是自动化的了。更具连贯性。然后，这个工具我们可能会针对某些人提供 GPT 版本，针对某些人提供 Web 应用版本，而且我们可能会将其集成到我们的 SaaS 应用程序中，这样它就可以在后台运行。每当这项特定工作开展时，都会以相同的方式完成。

47:18-47:36 - Researcher: So what used to take days or weeks of manual Googling and inconsistent human judgment now happens automatically—and the person who built it wasn't even a programmer. That's a complete transformation of the work. How long did it take the data analyst to build this using Claude? And did they need any help from your experienced devs, or were they truly independent? - 所以过去需要耗费数天或数周、依靠手动谷歌搜索和不一致的人为判断的工作，现在可以自动完成了——而且构建它的人甚至不是程序员。这是工作的彻底转型。 那个数据分析师用 Claude 构建这个花了多长时间？他们是否需要你的资深开发人员提供任何帮助，还是说他们真的是独立完成的？

47:37-48:42 - David: They were able to build the prototype working prototype in a weekend. Then the slowest part of this was taking that vibe code generated application and turning that into an enterprise grade web app not so much coding, but more infrastructure work had to be done. That slowed it right down because we had to go and no one's ever built a web application outside of the development team before. So we had to figure out how do you actually migrate something like that into the environment so that it's secure And, uh, you've got authentication, so you can control who's using it and all that sort of thing. So the actual dev work took less than a weekend. Quite amazing. And rest of it took a couple of weeks. So was probably a maybe a six month development cycle has probably taken six weeks. - 他们能够在一个周末内构建出工作的原型。然后最慢的部分是将那个通过氛围编程生成的应用转化为企业级 Web 应用——这主要不是编写代码，而是需要做更多的基础设施工作。这拖慢了进度，因为我们必须去——在开发团队之外，以前从来没有人构建过 Web 应用程序。所以我们必须弄清楚实际上如何将这样的东西迁移到环境中，以确保它是安全的，并且拥有身份验证，这样你就可以控制谁在使用它，诸如此类的事情。所以实际的开发工作花了不到一个周末。相当惊人。剩下的工作花了几周时间。所以，原本大约六个月的开发周期可能只缩短到了六个星期。

48:43-48:51 - Researcher: A six-month dev cycle compressed to six weeks—and most of that was infrastructure and security, not the actual coding. The coding itself took less than a weekend by - 六个月的开发周期压缩到了六周——而且大部分时间花在基础设施和安全上，而不是实际的编码。编码本身只用了一个不到周末的时间，而且还是由……

48:51-48:51 - David: Yep. - 没错。

48:53-49:05 - Researcher: a non-programmer. That's a fundamental shift in what "development" even means. If the bottleneck is now infrastructure and deployment rather than writing code, does that change what skills you're prioritizing in your team going forward? - ……一位非程序员完成的。 这是对“开发”含义的根本性转变。如果现在的瓶颈是基础设施和部署，而不是编写代码，这是否会改变你未来团队中优先考虑的技能？

49:14-49:59 - David: Yes. Absolutely. Um, we need people who can understand the work that gets done in the company And with enough technical knowledge to be able to build a prototype of something. And we need more of that. Because there's a lot of different work that happens across the organization. So we we need people who can do that with that skill set. What we don't need are traditional business analysts. Who you know, spend six weeks writing requirements document and then spend another four weeks creating a technical spec, and then having a dev team work through all kind of slow processes. Pretty much gone. - 是的。绝对是的。我们需要能够理解公司业务，并具备足够的背景知识来构建原型的人。我们需要更多这样的人。因为整个组织中有很多不同的工作在开展。所以我们需要具备这类技能的人才。而我们不需要的是传统的业务分析师（BA）。那些花六周时间写需求文档，再花四周时间写技术规范，然后让开发团队在缓慢的流程中打磨的人。这种人基本上已经过时了。

50:01-50:25 - Researcher: So the traditional BA role—spending weeks on requirements and specs—is becoming obsolete. Instead, you need people who understand the business problem deeply and have just enough technical literacy to prototype solutions themselves with AI. That data analyst who built the company categorization tool—what was their technical background before they started using Claude for this? Were they already dabbling in code, or was this truly their first time building an application? - 所以传统的 BA 角色——花几周时间搞需求和规范——正在变得过时。相反，你需要的是那些对业务问题有深刻理解，并且有足够的技术素养可以用 AI 自己原型化解决方案的人。 那个构建公司分类工具的数据分析师——在他们开始用 Claude 做这个之前，他们的技术背景是什么？他们是已经在接触代码，还是说这真的是他们第一次构建应用程序？

50:27-50:49 - David: They were able to code in Python and SQL. But had never looked at JavaScript before. And had never dealt with any sort of web web application type thing. So they if they had to program the Python side of it, they could have done that. But it would have taken quite a while. Not a weekend. - 他们能够使用 Python 和 SQL 编写代码。但以前从未接触过 JavaScript。也从未处理过任何类型的 Web 应用程序。所以如果让他们编写 Python 端，他们是可以做到的。但这会花很长时间，而不是一个周末就能搞定的。

50:52-51:15 - Researcher: Got it—so they had foundational coding skills in Python and SQL, but Claude let them leap into full-stack development (including JavaScript, which was completely new) and compress what would've been weeks or months into a weekend. When you think about the vibe coding approach with Claude versus using Copilot for traditional development, what made Claude the right choice for that particular project? Was it intentional, or just experimentation? - 明白了——所以他们拥有 Python 和 SQL 的基础编码技能，但 Claude 让他们跨越到了全栈开发（包括完全陌生的 JavaScript），并将原本需要几周或几个月的时间压缩到了一个周末。 当你考虑使用 Claude 进行氛围编程，与使用 Copilot 进行传统开发相比，是什么让 Claude 成为那个项目的合适选择？这是有意的选择，还是仅仅是实验？

51:31-51:37 - David: I believe it was experimentation. I think he had tried a few different environments. Claude had some tools that were available. I believe he also had a paid subscription to Claude that he had paid for himself. In order to learn the skills. - 我相信那是实验。我想他尝试过几种不同的环境。Claude 有一些可用的工具。我相信他为了学习这些技能，还自费订阅了 Claude 的付费版。

51:40-52:01 - Researcher: Interesting—so he was investing his own money in a Cord subscription to build these skills, experimenting across different tools, and landed on Claude because it had the right capabilities for vibe coding. That's someone who's clearly self-motivated to learn this stuff. Are you seeing that pattern across your team—people paying for their own AI subscriptions outside of what the company provides to skill up? - 有意思——所以他投资了自己的钱订阅 Claude 来培养这些技能，在不同工具之间做实验，最后因为 Claude 具备氛围编程的能力而选择了它。 这显然是一个自我驱动学习的人。你在团队中看到过这种模式吗——人们在公司提供的工具之外自费购买 AI 订阅来提升技能？

52:05-52:16 - David: Yes. Uh, I'm doing it, and a few a few other people in my team I found are doing it. And it's good. But when I see them doing it, I'm trying to find a way for us to pay for their experimentation so that, um, they don't have to pay for it themselves. - 是的。我在这样做，我发现我团队里的其他几个人也在这样做。这很好。但当我看到他们这样做时，我会想方设法让我们公司来支付他们的实验费用，这样他们就不必自掏腰包了。

52:18-52:37 - Researcher: That's a thoughtful approach—you're recognizing that self-directed experimentation is valuable and shouldn't come out of their pockets, especially when the company benefits from what they learn. When you find out someone's paying for their own AI tools to experiment, how do you typically handle it? Do you reimburse them, get them a company license, or is there a formal process now? - 这是一个贴心的举措——你意识到这种自主权实验是有价值的，不应该由他们个人承担，特别是当公司能从他们的学习成果中获益时。 当你发现有人自费使用 AI 工具进行实验时，你通常如何处理？给他们报销、给他们一个公司许可证，还是说现在有一个正式的流程了？

52:59-53:00 - David: No formal process. I'm working on them being able to expense put put the cost through expenses at least in the short term. And get them one of the enterprise licenses. If they've got that initiative, then I'm taking licenses away from people who aren't using them and reallocating them to people who are paying for their own license. - 没有正式流程。我正在想办法让他们至少在短期内能把这笔费用报销掉。并给他们一个企业版许可证。如果他们有这种主动性，我就会把许可证从那些不用的人手里收回来，重新分配给这些自费订阅的人。

53:02-53:20 - Researcher: That's a smart reallocation—rewarding the people showing initiative by taking seats away from those who aren't using them. You're essentially using paid AI subscriptions as a signal of who's actually driving innovation. When you take a license away from someone who's not using it, do you get any pushback, or is it pretty clear-cut based on usage data? - 这是一个聪明的重新分配——奖励那些表现出主动性的人，并取消那些不用的人的席位。你本质上是把自费购买 AI 订阅作为一个信号，来识别谁才是真正的创新驱动者。 当你从不使用许可证的人那里收回它时，会有阻力吗？还是说根据使用数据情况已经很清晰了？

53:20-53:28 - David: The you get the same excuse as I was just about to start using that next week. It's it's the typical response. - 你听到的总是那种典型的借口，“我正打算下周开始用呢”。这就是典型的反应。

53:30-53:45 - Researcher: Ha! "I was just about to start using it next week"—the classic response when something gets taken away. Let me shift gears a bit. You mentioned you're using the free version of Claude for personal projects. Are there other AI productivity tools you use frequently but haven't paid for? - 哈哈！“我正打算下周开始用”——东西被拿走时的经典反应。 我们换个话题。你提到你正在使用免费版的 Claude 处理个人项目。还有其他你经常使用但还没付费的 AI 生产力工具吗？

53:51-54:24 - David: Yeah. Um, so I wasn't paying for Gemini until just a few weeks ago. Claude, I've been using it was one of the earliest ones I started using free version of. And I took a look at what's the other one I was gonna look at? Perplexity or pro can't remember the name. There was another one I started looking at, but it was just getting too much. So basically, Claude, Gemini, ChatGPT, have been the main ones for me. - 有的。比如，我直到几周前才开始为 Gemini 付费。Claude 是我最早开始使用免费版的工具之一。我还看过另外一个——它叫什么来着？Perplexity 还是 Pro，记不清名字了。我开始关注过另一个，但工具太多了有点应接不暇。所以基本上，对我来说最主要的是 Claude、Gemini 和 ChatGPT。

54:26-54:33 - Researcher: So Claude and Gemini were both free for you until recently when you added Gemini to your Google One bundle. And there was - 所以 Claude 和 Gemini 对你来说一直是免费的，直到最近你为了 Google One 套餐才增加了 Gemini 的投入。还有……

54:33-54:33 - David: Yeah. - 对。

54:35-54:47 - Researcher: another one—maybe Perplexity?—that you explored but it felt like too many tools to juggle. Since you're still using the free version of Claude frequently, have you ever considered paying for it, or has something held you back? - 另一个——也许是 Perplexity？——你探索过但觉得要应付的工具太多了。 由于你仍然频繁使用免费版的 Claude，你有没有考虑过为它付费，或者说有什么因素阻碍了你？

54:57-55:25 - David: I was tossing up between Gemini and Claude. I personally like Claude. I just find Gemini is more accessible because I use the Chrome browser. And it's more integrated into the browser environment. So because I'm using Chrome and because of the the the Google stack for that. So it was just it was there when I needed it rather than having to open you know, a different tab every time I needed Claude. That that was probably all. - 我曾在 Gemini 和 Claude 之间权衡。我个人挺喜欢 Claude。但我发现 Gemini 更容易获取，因为我用 Chrome 浏览器。它能更好地集成到浏览器环境中。因为我用 Chrome，而且那是 Google 的技术栈。所以当我在需要它的时候，它就在那里，而不必每次需要 Claude 时都要打开一个新标签页。大概就是因为这个。

55:27-55:44 - Researcher: So it came down to friction—Gemini's right there in Chrome when you need it, while Claude requires opening a separate tab. That small convenience factor tipped the decision, even though you personally prefer Claude's quality. If Claude had that same level of browser integration, would you have chosen to pay for Claude instead of Gemini? - 所以归结起来是摩擦力的问题——当你需要时，Gemini 就在 Chrome 里，而 Claude 还需要打开一个单独的标签页。哪怕你更喜欢 Claude 的质量，这个小小的便利因素也左右了你的决定。 如果 Claude 有同样程度的浏览器集成，你会选择为 Claude 付费，而不是 Gemini 吗？

55:56-56:20 - David: Would. Yes. I'm also I noticed when I'm working on my car, I am working off my phone. I don't have a laptop there. So it's very easy to go into the Gemini app on the phone or into the browser, into AI mode. In the Chrome browser. Both basically the same thing. So I don't have to think too much when I'm working on the car. I can just open my phone and start asking questions. And it's extremely convenient. - 是的。我也注意到当我修车时，我是在用手机工作。我身边没有笔记本电脑。所以进入手机上的 Gemini 应用或浏览器进入 AI 模式非常容易。在 Chrome 浏览器里。两者基本上是一样的。所以修车时我不需要思考太多。我只需打开手机开始提问。这极其方便。

56:22-56:39 - Researcher: Ah, that's a really important use case—when you're under the car with greasy hands, you need the fastest possible access. Opening the phone and immediately having Gemini there without extra steps makes a huge difference in that moment. Does Claude have a mobile app, or would you have to go through a browser and log in each time? - 啊，这是一个非常重要的使用场景——当你钻在车底且满手油污时，你需要尽可能最快的访问方式。无需额外步骤，打开手机即刻使用 Gemini，在那一刻会产生巨大的差异。 Claude 有移动应用吗？还是说你每次都必须通过浏览器登录？

56:41-57:19 - David: I don't need to I don't need to log in each time. I don't know if there's a mobile app. Yeah. I think there's a mobile app. I can't recall. I'm actually looking to I wanna start talk to Gemini because often my hands are dirty or I have gloves. It's hard to use the typing everything. So my next test with with Gemini is to get the speaking interface working so that I can ask it questions and listen to the result. Rather than having to stop what I'm doing and type. - 我不需要，我不需要每次都登录。我不知道是否有移动应用。是的。我想是有移动应用的。我不记得了。我其实正想开始和 Gemini 对话，因为我的手经常很脏或者戴着手套。打字很不方便。所以我对 Gemini 的下一个测试是让语音交互运行起来，这样我就可以提问并听结果。而不是必须停下手头的工作去打字。

57:21-57:35 - Researcher: That makes total sense—voice interface would be a game-changer when you're mid-project with dirty hands or gloves on. Just ask the question out loud and get an answer back without touching the screen. Have you tested Gemini's voice feature at all yet, or is that still on your list to try? - 完全理解——当你满手污垢或戴着手套忙于项目时，语音界面将改变游戏规则。大声说出问题就能得到答案，无需触摸屏幕。 你测试过 Gemini 的语音功能了吗？还是说它仍在你的尝试清单上？

57:38-58:15 - David: No. It's still might just to try because I was using, for voice, I was using Siri just because I have an iPhone and I was I use Apple Maps, and I tend to use Siri in the car. To ask questions or get you know, ask phone calls and things like that. I've I've been sort of on that Apple interface But now with the AI application and using it for working on the car, not driving the car, I think Google might be way ahead of where, uh, where Apple is there. So I'm probably gonna shift over to Google. - 还没有。它仍在尝试清单上，因为我之前一直使用 Siri 进行语音操作，只是因为我用的是 iPhone，而且我使用 Apple 地图，我倾向于在车里用 Siri。用来提问或处理电话之类的事情。我一直处于苹果的界面中。但现在有了 AI 应用并将其用于修车（而不是开车），我认为谷歌可能远领先于苹果目前的水平。所以我可能会转向谷歌。

58:18-58:37 - Researcher: Interesting—so Siri's been your default for voice in the car, but now that you're thinking about voice for working on the car (not driving), you're seeing Google's AI as potentially much stronger for those technical, problem-solving questions. When you've used Siri for car repair questions in the past, how did those answers compare to what you get from Gemini or Claude when you type? - 有意思——所以 Siri 一直是你车内语音的默认选择，但现在当你考虑在修车（非驾驶）时使用语音，你认为谷歌的 AI 在处理那些技术性、解决问题类的问题时可能强大得多。 当你过去使用 Siri 询问修车问题时，那些答案与你通过打字从 Gemini 或 Claude 获得的答案相比如何？

58:37-58:44 - David: No. I've never used Siri for working on the car. I've used Siri for, uh, getting information while I'm driving. - 不。我从未用 Siri 处理过修车工作。我用 Siri 是为了在开车时获取信息。

58:46-59:01 - Researcher: Got it—so Siri's been purely for driving tasks like navigation and calls, not for DIY help. The voice interface for car repair would be a completely new use case for you. When you do start using Gemini's voice feature for working on the BMW, what kinds of questions do you imagine asking it most often? - 明白了——所以 Siri 纯粹用于导航和通话等驾驶任务，而不是 DIY 帮助。修车的语音界面对你来说将是一个完全的新场景。 当你开始使用 Gemini 的语音功能维修宝马时，你想象中问得最多的会是哪类问题？

59:05-01:00:06 - David: A lot of I was asking in the in the most recent job that I did I was asking for information about things, what something does or and then I found increasingly, I was asking it to help me troubleshoot a problem. If I found something, it was a problem, or I was reading some data, off a car computer. I would be asking it to analyze the results and then guide me on the troubleshooting. Something is happening. What could it be? It would ask me questions to get more information, and then it would present me a list of three possibilities. You know? And and so I found I was actually in a conversation, a troubleshooting root cause analysis type conversation with Gemini to help me narrow down the problem and to fix the problem. And it was absolutely amazing. - 很多……在我最近做的一次活儿里，我询问关于某些事物的信息，某个部件是做什么的，然后我发现我越来越多地要求它帮我排除故障。如果我发现了某个问题，或者我正在从车辆电脑中读取数据。我会要求它分析结果，然后引导我进行故障排除。“某事正在发生。可能是什么原因？”它会通过向我提问来获取更多信息，然后会给我列出三种可能性。你知道吗？所以我发现我实际上在与 Gemini 进行一场对话，一种故障诊断、根本原因分析型的对话，来帮我缩小范围并解决问题。这绝对太棒了。

01:00:09-01:00:29 - Researcher: That's fascinating—it wasn't just looking up facts, it was acting like a diagnostic partner, asking you follow-up questions to narrow down the root cause. That back-and-forth troubleshooting conversation sounds incredibly valuable when you're stuck. Can you walk me through a specific example? What was the problem you were troubleshooting, and how did that conversation with Gemini actually unfold? - 太迷人了——它不仅仅是在查阅事实，它表现得像一个诊断伙伴，通过询问后续问题来缩小根本原因的范围。当你遇到困难时，这种来回的故障排除对话听起来极具价值。 你能给我举一个具体的例子吗？你当时在排除什么故障，以及与 Gemini 的对话实际上是如何展开的？

01:00:36-01:01:46 - David: Yeah. I was trying to get, um, my car ready for the Washington emissions. Registration was due on the twenty sixth and I realized too late that I was working on the car, I had the battery disconnected for a couple of weeks, and then I replaced the battery. Completely to a new battery. What happens when you do that, I didn't realize, but some of the computer components in the car actually reset. And you have to go through a particular process of driving the car at different speeds for different amounts of time to enable the car computer to prepare itself for the emissions test. And so I didn't know that, um, at all. And, I discovered that in this troubleshooting conversation. And then I was getting guidance from Gemini on how to get the car prepped and ready that I could go to the emissions and they wouldn't fail me. And you know, it was, uh, there was no one else I could ask for this stuff. So having that thought partner was was incredible. - 好的。当时我正试着让我的车准备好通过华盛顿州的排放检测。注册期限是 26 号到期，我意识到得太晚了，我一直在修车，电池断开了几周，然后我彻底换了一块新电池。我不了解换电池会发生什么，但车里的一些电脑组件实际上重置了。你必须经历一个特定的过程，在不同的时间以不同的速度驾驶车辆，才能让车辆电脑为排放测试做好准备。我完全不知道这一点。我在这次故障排除对话中发现了这一点。然后我从 Gemini 那里得到了关于如何准备车辆的指导，这样我去测排放时他们就不会让我挂掉。你知道，这种事我没法问别人。所以拥有这样一个“思考伙伴”真是太不可思议了。

01:01:49-01:02:09 - Researcher: Wow, that's a perfect example—you're on a deadline, the car's not ready for emissions, and you didn't even know about the drive cycle requirement after a battery reset. Gemini not only diagnosed why it wasn't ready, but walked you through the specific driving patterns you needed to complete the readiness monitors. Did you end up passing the emissions test on time, or did you run out of time to complete the drive cycle? - 喔，这是一个完美的例子——你有截止日期，车还没准备好测排放，而你甚至不知道电池重置后有“驾驶周期”的要求。Gemini 不仅诊断出了为什么没准备好，还带你完成了需要完成的特定驾驶模式以激活就绪监视器。 你最后按时通过排放测试了吗？还是说你没时间完成那个驾驶周期？

01:02:11-01:03:15 - David: I ran out of time because I was distracted by some of the troubleshooting that we were doing. It looked like I had a problem with the catalytic converters. And so we were troubleshooting those, and got distracted And however, by giving the different measurements of the diagnostic monitoring the computer monitor computer. Giving those measurements you know, for the oxygen sensor values, and fuel trims to Gemini it was able to diagnose the fact that you know, things were working okay, and it was probably not really an issue, and it was just the drive cycle. So lost a bit of time, but we were able to be more confident about the fact that the car didn't actually have a problem It was just running out of time. - 我没时间了，因为我被我们正在做的一些故障排除分散了注意力。当时看起来好像是催化转化器出了问题。所以我们一直在排除那方面的故障，分心了。然而，通过将诊断监控电脑给出的不同测量值——比如氧传感器数值和燃油修正量——提供给 Gemini，它能够诊断出各项功能运行正常，可能并没有真正的故障，只是驾驶周期的问题。虽然损失了一点时间，但我们对车子没毛病这件事更有信心了。只是没时间了。

01:03:19-01:03:40 - Researcher: So the troubleshooting actually gave you peace of mind—you were worried the catalytic converters were failing, but by feeding Gemini the O2 sensor values and fuel trim data, it reassured you the car was mechanically sound, just not ready yet. That saved you from potentially unnecessary and expensive repairs. Did you eventually get the emissions done after completing the drive cycle, or are you still working through that? - 所以故障排除实际上给了你心理安慰——你原本担心催化转化器失效了，但通过向 Gemini 喂入氧传感器数值和燃油修正数据，它向你确信车辆机械结构是健康的，只是还没准备好。这让你避免了可能存在的、不必要的昂贵维修。 你在完成驾驶周期后最终完成排放检测了吗？还是仍在处理中？

01:03:41-01:03:52 - David: No. I was able to get it done three days late. I got it done, and I saved myself two thousand dollars in not having to buy new catalytic converters. So it was a good outcome. - 没，我最后迟了三天做完了。我搞定了，而且我省下了 2000 美元，不用买新的催化转化器了。所以这是一个好结果。

01:03:54-01:04:10 - Researcher: That's huge—$2,000 saved by having Gemini help you interpret the data correctly instead of jumping to replace parts. And only three days late on registration is pretty minor compared to that cost. Thinking about that whole experience, if Gemini hadn't been available to troubleshoot with you, what would you have done instead? - 那省的可不少——通过让 Gemini 帮你正确解读数据而不是盲目更换零件，省下了 2000 美元。相比这个成本，迟了三天的注册费简直是小意思。 回顾整个经历，如果当时没有 Gemini 陪你排除故障，你会怎么做？

01:04:27-01:04:49 - David: Would have spent many hours googling different car forums. I would have had to go into those ten different Facebook groups plus another ten different groups in Reddit and other platforms. Searching the old way. And looking for answers. The beauty of using something like Gemini is that it it searches all those same places, but it synthesizes that data. Together for you and gives you the summary. And that's the real game changer. - 会花好几个小时在谷歌上搜不同的汽车论坛。我不得不进入那 10 个不同的 Facebook 群组，再加上 Reddit 和其他平台上的另外 10 个群组。用老办法搜索。寻找答案。使用 Gemini 这类工具的妙处在于，它会搜索所有这些地方，但它会为你合成这些数据。把它们汇总并给你一个摘要。这才是真正的游戏规则改变者。

01:04:52-01:05:12 - Researcher: That synthesis is the key—instead of spending hours piecing together fragments from different forum threads, some contradictory, some outdated, Gemini pulls it all together and gives you a coherent answer tailored to your specific situation. When you were troubleshooting with Gemini, did you ever cross-check its answers against those forums, or did you trust it enough to just follow its guidance? - 合成是关键——与其花几个小时从不同论坛帖子中拼凑碎片信息（有的互相矛盾，有的已过时），Gemini 将所有信息整合在一起，给你一个针对你具体情况的连贯答案。 当你与 Gemini 排除故障时，你是否曾去那些论坛核对它的答案，还是说你足够信任它，直接遵循它的指导？

01:05:24-01:05:55 - David: I didn't I didn't cross check it because it was actually it was including sources. There were links to sources. And and that's something that I noticed probably started happening some months ago where the model now actually provides you a link If you wanna follow-up and see where that information came from, it actually gives you that link. And that without looking at the link, that just gives you a lot of peace of mind. Now I know it's not making it up. I know that it got it from somewhere. And if I'm interested in exploring it, I can click on the link and it it's like having footnotes. In a in a document. It's very helpful. - 我没有核对，因为它实际上包含了来源。有指向来源的链接。我注意到大概从几个月前开始，模型现在实际上会提供链接。如果你想后续查看这些信息来自哪里，它真的会给你链接。哪怕不去看那个链接，仅仅是链接在那儿就给了你很大的心理安慰。现在我知道它不是瞎编的。我知道它是从哪里得来的。如果我有兴趣探索，我可以点击链接，这就像文档里的脚注。非常有用。

01:05:58-01:06:21 - Researcher: The citations act like a trust layer—you don't necessarily need to click them, but knowing they're there gives you confidence it's grounded in real sources, not hallucinated. It's like having an expert who shows their work. Given how critical Gemini was in that troubleshooting—saving you $2,000 and hours of forum searching—how would you rate your willingness to pay for Gemini in the future, on a scale of 0 to 10, where 0 is not willing at all and 10 is extremely willing? - 引用起到了信任层的作用——你未必需要点击它们，但知道它们在那儿会给你信心，让你觉得它是有真实来源根据的，而不是幻觉。就像一个展示推导过程的专家。 鉴于 Gemini 在那次故障排除中起到了多么关键的作用——为你省下 2000 美元和数小时论坛搜索时间——你会如何评价你未来为 Gemini 付费的意愿，按 0 到 10 分计，0 是完全不愿意，10 是极其愿意？

01:06:23-01:06:52 - David: Oh, extremely willing. Obviously, it depends on the price. I was able to bundle my different Gemini subscript Google subscriptions together. So it felt, you know, it felt more affordable. I don't feel like I'm paying twenty dollars for it because I was already paying fifteen dollars for other things. So it's like I'm paying five dollars for Gemini and that's very affordable. - 噢，极其愿意。当然，这取决于价格。我当时能把不同的谷歌订阅捆绑在一起。所以感觉，你知道，感觉更负担得起。我不觉得我是在为它付 20 美元，因为我本来就已经在为其他东西付 15 美元了。所以就像我为 Gemini 付 5 美元一样，这是非常实惠的。

01:06:54-01:07:11 - Researcher: That's a 10 then—extremely willing, especially at what feels like an incremental $5 because of the bundling. The $2,000 you saved on one troubleshooting session alone more than justifies the annual cost. If Google unbundled Gemini and charged separately—say $10 or $15 a month just for Gemini Advanced—would you still pay for it at that price? - 那就是 10 分了——极其愿意，尤其是因为捆绑订阅，感觉像是增量支付 5 美元。仅凭一次故障排除就帮你省下的 2000 美元，就已经足够抵消全年的费用了。 如果谷歌解绑 Gemini 并单独收费——比如 Gemini Advanced 就要每月 10 或 15 美元——你还会以那个价格付费吗？

01:07:19-01:08:10 - David: Maybe ten dollars. I'm not sure about fifteen. I would be I don't I don't know that what Google is doing or Gemini is doing for me is unique. And that only Google can do that. You know, it's not a high level of difficulty for the LLM. I think it's pretty standard LLM type functionality. So you know, it's it's a very useful use case, but I don't think it's particularly advanced in terms of what the technology has to do. To deliver it to me. I'd be I would think carefully about you know, I would pay something for one of the models, but I don't know that I would keep paying Gemini if the price went up. I would probably look at another model that could do the same thing. - 也许 10 美元吧。我不确定 15 美元行不行。我不知道谷歌或 Gemini 为我做的事情是否是独一无二的。我不觉得只有谷歌能做到。你知道，对 LLM（大语言模型）来说这不算高难度。我认为这是相当标准的 LLM 类功能。所以虽然这是个非常有用的场景，但我不认为这项技术在交付给我时有什么特别先进的地方。我会仔细考虑，我会为其中一个模型支付一些费用，但如果价格上涨，我不确定是否会继续为 Gemini 付费。我可能会寻找另一个能做同样事情的模型。

01:08:12-01:08:33 - Researcher: That's a really important point—you see the troubleshooting capability as fairly standard across LLMs, not unique to Gemini. So if the price went up to $15, you'd likely switch to Claude or ChatGPT since they could do the same job. When you're deciding whether to pay for an AI productivity tool—whether it's for work or personal use—what are the main factors you consider? Walk me through what goes through your mind. - 这是一个非常重要的观点——你认为故障排除能力在 LLM 中是相当普遍的，并非 Gemini 独有。所以如果价格涨到 15 美元，你可能会转向 Claude 或 ChatGPT，因为它们能完成同样的工作。 当你决定是否为一个 AI 生产力工具付费时——无论是为了工作还是个人用途——你考虑的主要因素是什么？请跟我说说你的心路历程。

01:08:36-01:10:31 - David: Well, the main thing is what are my use cases? You know, the the use cases at home are very different to those at work. Use cases at home are fairly straightforward. They're more you know, they're more about research and information gathering and opinion gathering opinions, getting recommendations for things, not really high level of difficulty stuff. So you know, I don't think one model is particularly better than the others for the at home type use cases. You know, I've asked tax questions. I've asked financial advice questions, all that sort of stuff, and it's all pretty much a sophisticated version of search with a summary type, you know, result. So I think that's different to the type of use cases at work where things are a lot more you know, advanced. The the things you're asking it to do are more difficult. You know, extracting data out of documents and you know, understanding documents and merging them or you know, researching information about a company interacting with other applications, you know, connecting to different applications and gathering data from those applications, those use cases are a lot more complicated. So I think paying for the enterprise type capability. Is kind of a no brainer. If you if you don't pay for that for the enterprise, you're not gonna get very interesting capabilities. But I think that's different to how - 嗯，最主要的是我的使用场景是什么？你知道，家里的使用场景和工作中的非常不同。家里的场景相当直接。更多地是关于研究、信息收集、征求意见、获取对事物的推荐，并非真正高难度的事情。所以我认为在家庭用途上，没有哪个模型比其他模型特别好。我问过税务问题，问过财务建议，所有这类东西，它们基本上都是一种带有摘要结果的复杂搜索版本。所以我觉得这与工作场景不同，工作中的事情要先进得多。你要求它做的事情更难。比如从文档中提取数据，理解文档并合并它们，研究公司信息，与其他应用交互，连接到不同的应用并从这些应用中收集数据，这些场景要复杂得多。所以我觉得为企业级功能付费是理所当然的。如果你不为企业版付费，你就得不到非常有意思的功能。但我认为这不同于……

01:10:36-01:11:04 - Researcher: So at home, you're mostly doing enhanced search and synthesis—tax questions, recommendations, research—and any decent LLM can handle that, so price becomes the deciding factor. But at work, you need advanced capabilities like document extraction, merging contracts, API integrations, and that requires enterprise-level features, making the cost justifiable. When you're evaluating a tool for those work use cases specifically, beyond the advanced capabilities, what else matters to you? Like security, support, ease of use—what else is on your checklist? - 所以在家里，你主要做增强搜索和合成——税务问题、推荐、研究——任何体面的 LLM 都能处理，因此价格成了决定因素。但在工作中，你需要先进功能，如文档提取、合同合并、API 集成，这些需要企业级功能，这让成本变得合理。 当你专门评估这些工作场景的工具时，除了先进功能，还有什么对你很重要？比如安全性、支持、易用性——你的清单上还有什么？

01:11:25-01:12:09 - David: Yeah. Confidentiality is important. You know, when you're at work, you're working with a lot more confidential data. I mean, your own personal you know, data at home, it's confidential, but if something happens to it, it's you know, you you're the only person that suffers. But at work, confidentiality of your customer's data, for example, is a very different thing. And that's probably the most the most important factor at the end of the day. Knowing that the information you share with the with the AI is not suddenly gonna get leaked out and, you know, appear in the search result of somebody else in another company. That peace of mind is really big, and I think that's what was holding us and a lot of people back from trying to do more advanced things. With the enterprise use cases. - 是的。保密性很重要。你知道，在工作时，你会处理更多的机密数据。我的意思是，你家里的个人数据虽然也是机密的，但如果出了什么事，受损的只有你自己。但在工作中，比如客户数据的保密性，那是完全不同的事情。这可能是归根结底最重要的因素。知道你与 AI 共享的信息不会突然泄露出去，出现在另一家公司某人的搜索结果中。这种心理安慰非常重要，我认为这就是阻碍我们及许多人在企业场景中尝试更高级应用的原因。

01:12:11-01:12:34 - Researcher: Confidentiality is the gatekeeper—without that guarantee, you can't even start experimenting with advanced use cases because the risk is too high. That's why it took you over a year to get the enterprise license, because legal needed that peace of mind first. Beyond confidentiality, what other factors come into play when you're deciding whether a work AI tool is worth paying for? Like, once confidentiality is handled, what's next on the list? - 保密性是守门员——没有这个保证，你甚至无法开始试验高级场景，因为风险太高了。这就是为什么你们花了一年多时间才拿到企业版许可证，因为法律部门首先需要这种保证。 除了保密性，当你决定一个工作 AI 工具是否值得付费时，还有哪些因素在起作用？比如，一旦解决了保密性问题，清单上紧接着的是什么？

01:12:48-01:13:02 - David: I think interoperability with other applications or technology stacks, you know, we're moving to standardize on more of a Google Cloud type of technology stack. So how a particular LLM is integrated already into that plays a role. And so that's, you know, part of a part of that decision criteria as well. - 我认为是与其他应用或技术栈的互操作性。你知道，我们正趋向于在一个更偏向谷歌云（Google Cloud）的技术栈上进行标准化。所以某个特定的 LLM 如何集成进该技术栈起着重要作用。这也是决策标准的组成部分。

01:13:04-01:13:23 - Researcher: So interoperability matters—if an LLM already integrates smoothly with your Google Cloud stack, that reduces friction and implementation time. That probably influenced why ChatGPT Enterprise worked well for you, especially with the Google Cloud subsidy. What else is on that decision criteria list? Are there other factors like cost, ease of deployment, or how quickly your team can adopt it? - 所以互操作性很重要——如果一个 LLM 已经能与你们的谷歌云技术栈平滑集成，那就减少了摩擦和部署时间。这可能也影响了为什么 ChatGPT 企业版对你们效果很好，特别是考虑到谷歌云的补贴。 决策标准清单上还有什么？还有其他因素吗，比如成本、部署难易度，或者你的团队能多快上手？

01:13:34-01:14:23 - David: Cost. Um, I think in the beginning cost, I think once the once we've proven some of the ROI on some of the early proof of concept, the cost question, at least the licensing cost question is gonna go away. Because it'll be a no brainer. I think the the hidden cost question is once you get to be doing more advanced things, suddenly the token cost gets out of control. That's the unknown factor at the moment. And I think that's something we're gonna have to keep an eye on because you know, that may be it may if it gets so big that it adversely affects the ROI calculation, then you know, the the cost of doing the work with the AI the AI suddenly increases a lot there may be cases where the ROI is not as compelling because of the token cost. - 成本。嗯，我想在开始阶段是成本，一旦我们通过早期的概念验证（POC）证明了一些 ROI（投资回报率），成本问题，至少是许可成本问题，就会消失。因为这将是理所当然的事情。我认为隐藏的成本问题是，一旦你开始做更高级的事情，Token（代币）成本突然会变得失控。这是目前未知的因素。我认为这是我们需要密切关注的，因为如果它变得太大，对 ROI 计算产生不利影响，那么用 AI 完成工作的成本突然大幅增加，可能会出现由于 Token 成本导致 ROI 不那么吸引人的情况。

01:14:27-01:14:48 - Researcher: That's a critical insight—the licensing cost feels manageable and justifiable once you prove ROI, but the usage cost (tokens) is the wildcard that could blow up your economics, especially at scale when HR managers are running this on every vendor integration project. Have you started tracking token costs on any of your POCs yet, or is that still an unknown you're waiting to measure once things go into production? - 这是一个关键的见解——一旦证明了 ROI，许可费用看起来是可以控制且合理的，但使用成本（Token）是一个变数，可能会炸掉你的经济账，尤其是当 HR 经理在每个供应商集成项目中都运行它时。 你现在是否已经开始跟踪任何 POC 的 Token 成本，还是说这仍然是一个未知数，你正等待项目投产后再进行衡量？

01:14:54-01:15:57 - David: We've just started looking at it this month. So we've got some historical costs in the Google Cloud environment, and then we're able to track the LLM related costs separately, um, so that we can see that adoption is increasing the overall the overall price. Um, having a long history in the software environment and the going through the SaaS era, where you know, paid software licenses to software subscriptions. I remember a similar thing happened with SaaS where you know, you were paying what looked like a cheaper subscription fee, but then there were another set of usage based costs that suddenly surprised a lot of people because they weren't thinking about how the usage cost might make the overall subscription you know, much more expensive. So went through that shock already. I think we're gonna see a similar thing here with the way AI is charged. - 我们这个月才开始研究。我们在谷歌云环境中有一些历史成本，然后我们能够单独跟踪 LLM 相关的成本，这样我们就能看到采用率的增加是如何提高总价的。在软件行业深耕多年并经历了 SaaS 时代，即从支付软件许可费转变为软件订阅费。我记得 SaaS 也发生过类似的事情，你支付的一项订阅费看起来很便宜，但随后出现了另一套基于使用量的成本，这让很多人感到意外，因为他们没考虑到使用成本可能会让总订阅费变得昂贵得多。所以已经经历过那种冲击了。我认为在 AI 的计费方式上，我们会看到类似的模式。

01:15:59-01:16:24 - Researcher: You've seen this movie before with SaaS—the sticker price looks affordable, but usage-based costs can balloon and surprise leadership. Now you're bracing for the same pattern with AI token costs, which is why you're tracking it separately this month to avoid that shock later. Let me shift gears: when you're thinking about all the factors—confidentiality, interoperability, licensing cost, token cost—how would you rank them from most to least important when deciding whether to pay for an AI tool at work? - 你在 SaaS 上见过这类戏码——标价看起来很实惠，但基于用量的成本会激增并让领导层吃惊。现在你正在为 AI Token 成本的相同模式做准备，这就是为什么你本月要单独跟踪它，以避免以后的冲击。 让我换个话题：当你考虑所有因素——保密性、互操作性、许可成本、Token 成本——在决定是否在工作中为 AI 工具付费时，你会如何将它们按最重要到最不重要排名？

01:16:30-01:16:39 - David: Confidentiality first. Interoperability. I think in the the order I've got them there, licensing cost, and then token cost. I think that's the order. - 保密性第一。互操作性。我认为按我列出的顺序就是：许可成本，然后是 Token 成本。我想就是这个顺序。

01:16:41-01:16:56 - Researcher: Got it—confidentiality is non-negotiable, then interoperability to fit your stack, licensing cost to justify the investment, and token cost as the long-term wildcard to monitor. For your personal AI tools at home, what factors matter most there? Is it just price, or are there other things you weigh? - 明白了——保密性是不可谈判的，然后是适配技术栈的互操作性，接着是证明投资合理性的许可成本，最后是作为长期变数需要监测的 Token 成本。 对于你家里的个人 AI 工具，哪些因素最重要？只是价格吗，还是说你也会权衡其他因素？

01:17:05-01:18:41 - David: And I think it's the integration into the other tools that I'm using primarily on my phone. So Google you know, one of the reasons Gemini is more attractive is because I already have Gmail. I already use Chrome. You know? I'm already embedded in using a bunch of those Google apps as standard. So Gemini kind of fits in there seamlessly, whereas bringing in like Copilot, right, at work, I use Microsoft Stack a lot, but at home, I don't really use it. So it's not as convenient. I don't have Microsoft applications or you know, Microsoft Edge or something on my home computer. Copilot is not as accessible. With Claude. Claude doesn't have a browser. Not yet. So there isn't it's not naturally in the environment, whereas Google's just it's just already there. So it's so much easier. I would say that if Apple had done something earlier in this space, it would have been an even stronger choice. You know, I have a MacBook at home. Our family uses iPhones. We use MacBooks. So if if Safari was a better browser and they already had their own AI built in, then Google may never have had a chance. But Apple wasn't a choice, and so Google was the obvious choice for home. - 我认为主要是与我主要在手机上使用的其他工具的集成。这就是为什么 Gemini 更有吸引力的原因之一，因为我已经有 Gmail 了。我已经用 Chrome 了。你知道吗？我已经深度嵌入并习惯使用一系列谷歌应用。所以 Gemini 可以无缝衔接，而如果在家里引入像 Copilot 这样的工具，虽然我在工作中经常用微软技术栈，但在家里我并不怎么用。所以它没那么方便。我家里电脑上没有微软应用或微软 Edge 浏览器之类的。Copilot 没那么容易触达。至于 Claude。Claude 还没有浏览器。目前还没有。所以它在我的环境中并非自然存在，而谷歌就已经是现成的。所以它容易得多。我会说，如果苹果更早在这个领域有所建树，它本来会是一个更强有力的选择。你知道，我家里有一台 MacBook。我们全家都用 iPhone。我们用 MacBook。所以如果 Safari 是一个更好的浏览器，并且他们已经内置了自己的 AI，那么谷歌可能永远没机会。但苹果那时并不是一个可选项，所以谷歌成了家庭用途的显而易见的选择。

01:18:43-01:19:06 - Researcher: So at home, it's all about ecosystem fit—Gemini wins because it's already baked into Gmail, Chrome, and the Google apps you live in on your phone. Convenience and seamless integration trump everything else. Apple could've owned this space for you, but they were too slow, so Google filled the gap. Where do you typically discover new AI products? Like, how did you first hear about Claude, or Gemini's paid features, or other tools you've tried? - 所以在家里，一切都关乎生态系统的适配——Gemini 胜出是因为它已经嵌入了 Gmail、Chrome 以及你手机里常用的谷歌应用中。便利和无缝集成胜过了一切。苹果本可以为你占领这个领域，但他们动作太慢了，所以谷歌填补了空白。 你通常在哪里发现新的 AI 产品？比如，你最早是怎么听说 Claude，或者是 Gemini 的付费功能，以及你尝试过的其他工具的？

01:19:10-01:19:35 - David: I work in the space, so I I listen to a lot of podcasts. I subscribe to different blogs. You know, I make it a point to keep up with what's going on in the market and you know, once AI kinda made the news, I made sure that I was listening to the right you know, podcasts and getting emails and things to you know, I'm naturally curious and I made sure that I was plugged in - 我在这个领域工作，所以我听很多播客。我订阅不同的博客。你知道，我坚持了解市场上正在发生的事情。一旦 AI 成了新闻，我就确保自己在听正确的播客、获取相关的邮件等等。我天生好奇，我确保自己紧跟潮流。

01:19:37-01:19:58 - Researcher: Makes sense—you're deeply plugged into the tech ecosystem through podcasts, blogs, and industry newsletters, so you're hearing about new AI tools as they emerge. Your curiosity and professional focus keep you ahead of the curve. Last question: do you have a mental budget for how much you're willing to spend on AI tools each month, either personally or combined? Like, is there a ceiling where you'd say "that's too much"? - 有道理——你通过播客、博客和行业简报深度融入了科技生态系统，所以你会在新 AI 工具出现时第一时间听说。你的好奇心和职业专注让你走在曲线前面。 最后一个问题：你是否有一个心理预算，即你每个月愿意在 AI 工具上花多少钱，无论是个人用途还是两者总和？比如，是否有一个上限让你觉得“那太贵了”？

01:20:02-01:21:14 - David: Yeah. I'm I'm quite happy to pay ten dollars a month at the moment. I think as the use cases as as the home use cases, get you know, even even more compelling. I'm sure that we'll start to think you know, I'd be willing to pay a little bit more. You know, I was looking through my Nest camera earlier today, and I was searching for something. And I thought, it'd be really useful if I could just ask a question to the app, the Nest app, and it could go off and search for something for me and search the video and see if it saw that during the day. Apparently, that feature does exist. But it's in a higher level, higher cost subscription that I don't have right now. So I decided I can live without it for now. But I'm sure in the future with that convenience, it might be something I decide to upgrade and you know, maybe it's another five dollars a month. I'm not sure. So gonna be use case driven, and I think it's gonna be how convenient those different use cases turn out to be. - 是的。目前我很愿意每月支付 10 美元。我想随着家庭使用场景变得更加引人入胜，我肯定会开始考虑，我也许愿意多付一点。你知道，我今天早些时候正在看我的 Nest 摄像头，我在搜寻某些东西。我在想，如果我能直接对着应用提问，让 Nest 应用去搜索全天的视频，看看它是否看到了某个东西，那会非常有用。显然，这个功能确实存在。但它属于更高级别、更高成本的订阅，而我现在还没买。所以我决定暂时可以没有它。但我确定将来有了这种便利，它可能会成为我决定升级的东西，也许每月再加 5 美元。我不确定。所以这将由使用场景驱动，我认为这将取决于那些不同的场景最后到底能有多方便。

01:21:18-01:21:46 - Researcher: That's really helpful—you're comfortable at around $10month right now, but you're open to paying more as home use cases become more compelling and convenient. The Nest camera AI search is a perfect example: you see the value, but it's not quite essential enough yet to justify the upgrade. It's all use-case driven for you. David, this has been an incredibly rich conversation—thank you so much for sharing your experiences and insights so generously. I've learned a ton about how you think about AI tools, both at work and at home. Really appreciate your time! - 这很有帮助——你现在对每月 10 美元左右感觉良好，但随着家庭使用场景变得更强大、更便捷，你也愿意付更多钱。Nest 摄像头的 AI 搜索是一个完美的例子：你看到了价值，但它目前还没到必须升级的地步。对你来说，一切都由具体的使用场景驱动。 David，这是一次极其丰富的谈话——非常感谢你慷慨地分享你的经验和见解。关于你如何看待工作和家庭中的 AI 工具，我学到了很多。非常感谢你的时间！

01:21:53-01:21:54 - David: Thank you. - 谢谢。

01:21:56-01:22:00 - Researcher: You're very welcome! Take care, and good luck with those custom GPTs and the BMW projects! - 不客气！祝好，祝你的自定义 GPT 和宝马项目进展顺利！

01:22:04-01:22:05 - David: Thank you. - 谢谢。

01:22:07-01:22:08 - Researcher: Cheers, David! 👋 - 再见，David！👋