Palantir's forward deployed architect Chad Wahlquist on ontology, the death of dashboards, and the Evolve tool cutting token costs 60%
频道: TBPN
视频: https://www.tbpndigest.com/story/2026-06-04/palantirs-forward-deployed-architect-chad-wahlquist-on-ontology-the-death-of-dashboards-and-the-evolve-tool-cutting-token-costs-60
原文语言: en
统计: 共 28 轮
[0:00]
Featuring Chad Wahlquist
it. Thank you so much for taking time to come with us. Thanks for coming. Great to be with you.
Have a great rest of your time.
Thanks.
And up next, next
we have Chad Walquist. First, I'm going to tell you about Crowd Strike. Your business is AI. Their business is securing it. Crowd Strike secures AI and stops breaches. Welcome to the show. How are you doing, Chad?
Great overcoat. That's a new one. That's an Eleaniano special. It is.
Oh, yeah. He is the master.
Giving us a run for our money.
Yeah, it's fantastic. Uh, anyway, kick us off with an introduction on yourself, how you fit into Palunteer, a little bit of backstory. I'm sure we have a ton of questions to run through.
本期嘉宾：Chad Wahlquist
……这个。非常感谢你抽空来我们节目。谢谢你能来。很高兴来做客。
祝你接下来一切顺利。
谢谢。
主持人： 接下来，接下来
要请上来的是 Chad Wahlquist。不过先给大家说说 CrowdStrike：你们的业务是 AI，他们的业务是守护 AI 的安全。CrowdStrike 保护 AI，阻止安全入侵。欢迎来到节目！最近怎么样，Chad？
这件大衣真帅。这件没见过啊。这是 Eleaniano 的特制款。确实是。
哦，对。他是真正的大师。
都快把我们比下去了。
主持人： 是啊，太棒了。呃，言归正传，先给我们做个自我介绍吧：你在 Palantir 是什么角色，再讲讲你的背景故事。我们肯定有一大堆问题要一个个问。
[0:01]
Uh, first, how often do you guys do these things? Because it feels like this feels like an annual It feels like an annual event.
Yeah,
but we're getting the call every month. Karp carp karp talks about you know manipulating time working you know a quarter at palunteer is like a day or a year at an exact so that it kind of makes sense.
Yeah I'm like actually 23.
Yeah time the time warp is real. So we do these quarterly. So I'm I'm a ford deployed architect technically. Yeah.
I do what is needed and so doing the needful is kind of the palunteer way is like there's no job below me. And so no matter if I'm out on the edge with customers, I'm talking to executives, explaining the ontology, doing YouTube videos.
呃，先问一句，你们多久搞一次这个？因为感觉这……感觉像是一年一度的盛事。
对，
可我们每个月都会接到邀请。Karp……Karp 常说要“操纵时间”，你想，在 Palantir 干一个季度，就相当于在别处干一天，或者说一年，所以这么看也就说得通了。
对，我其实才 23 岁。
Chad： 是啊，时间扭曲是真的。所以我们是每季度搞一次。我呢，严格来说是一名 forward deployed architect（前置部署架构师）。嗯。
需要我干什么我就干什么。“做该做的事”，可以说就是 Palantir 的一贯作风，意思是没有哪份活儿是我不屑干的。所以我可能在一线陪着客户，可能在跟高管沟通、讲解本体（Ontology），也可能在录 YouTube 视频。
[0:02]
That's all what I'm doing. So really the goal is how do we help people decomp problems differently and apply the technology new ways.
Can AI do decomp?
Yes.
Okay. Unpack that because that feels like the secret sauce. That feels like the special thing about Palunteer is actually being able to bring someone in who understands an organization. I think a lot of people see AI tools. A lot of people see AI tools. No, a lot of people see AI tools and they and they think uh okay very defined workflow input output but now instead of just math that Python can deal with you can deal with some text and that's great but demp to me has always felt less like let's go into your HR system and understand the basic job description and like oh someone uploaded this resume versus oh Steve actually does this completely outside of that system and marketing has two two platforms for this thing and engineering has three systems for CAD files and the all the clues that have built up over decades sometimes hundreds of years for some of these organizations like you that's what was so special about the forward deployed engineer program the Palunteer model y
这些全都是我的工作。所以说到底，目标就是怎么帮大家换个方式去拆解（decomp）问题，用新的方式去应用技术。
主持人： AI 能做拆解吗？
Chad： 能。
主持人： 好，那展开讲讲，因为这感觉就是你们的秘方。感觉 Palantir 真正特别的地方，恰恰在于能派进来一个真正懂这家组织的人。我觉得很多人看 AI 工具……很多人看 AI 工具……不对，是很多人一看到 AI 工具，就会想：嗯，定义得很清楚的工作流，有输入、有输出，只不过以前只能处理 Python 能算的数学，现在还能处理一些文本了，这当然很好。但在我看来，拆解从来都不太像是“我们去翻翻你的 HR 系统，搞懂基本的岗位描述，哦，有人上传了这份简历”；而更像是发现“哦，其实 Steve 干的这件事完全在那个系统之外”“市场部为这件事用着两个平台”“工程部光管 CAD 文件就有三套系统”，还有那些积累了几十年、对有些组织来说甚至几百年的种种线索。就是，你……这正是 forward deployed engineer（前置部署工程师）项目、也就是 Palantir 模式如此特别的地方……
[0:03]
I'm surprised to hear you say I AI can do it at all it feels like the final boss
well this is where the the really the palunteer thesis is humans and AI working together
and so the way we thing about this is modeling our business process and we heard some other people talking about this of modeling my business process and the ontology um because the LMS don't necessarily have an a world view or world model of your business and your operations the ontology provides that okay
and so when we talk about demp this is really about actually now I make more data computable as well so we think about LM on the agents and I'm interacting also we use LM to make more data computable and then model that in the ontology of how things are really working and so what we're actually doing a lot of times now is is building out that world view and then running multiple agents over this actually um being combative towards each other right and so actually working against each other and having critiques and so after you do that you can also then give the human human in the loop feedback about this and iterate on this and so what we find is that's really a scaling mechanism it's like a new power tool right I think you guys were just talking about this the kind of the perspective around jobs and all this stuff it's like when you gave carpenters power tools there weren't less carpenters there were more I could do more with it right it's an empowering thing
听你说 AI 居然能做到这一点，我挺意外的，这感觉就是终极大 Boss 啊。
Chad： 嗯，这就要说到 Palantir 真正的核心主张了：人和 AI 协同工作。
我们的思路是对业务流程建模，我们也听到别人在聊这个，就是在本体里给自己的业务流程建模。因为 LLM 不一定具备关于你的业务、你的运营的世界观或者说世界模型，本体提供的正是这个。好。
所以我们说拆解，其实是说，现在我还能让更多数据变得可计算。说到 LLM，我们会想到 agent、想到跟它交互；但我们还用 LLM 让更多数据变得可计算，再把事情真实的运作方式建模进本体里。所以我们现在很多时候实际在做的，就是先把这套世界观搭起来，再让多个 agent 在上面跑，而且让它们彼此对抗，对吧，就是互相唱反调、互相挑刺。做完这一步，你还可以让人在回路（human in the loop）里对结果给出反馈，再不断迭代。我们发现，这其实是一种规模化机制，就像一件新的电动工具，对吧。你们刚才好像也在聊这个，就是关于工作岗位之类的看法：当年给木匠配上电动工具，木匠并没有变少，反而变多了，有了它我能干更多活，对吧？这是一种赋能。
[0:04]
yeah So, uh, how often like I I I I'm interested in the, uh, like the pie in the sky, Palunteer pitch, understand your entire business, run your entire business on Palunteer. And then some of the nitty-gritty where sometimes like the lowhanging fruit is like, wait, there's a like there's someone's job to just like take a form and type it into a sheet. Like, we have we've had image recognition for a long time. let's actually go and implement that and get that into a database, get that into the ontology, get that into Palunteer so then we can start building on top of it. And it feels like there might be a tension there. Obviously, both processes are speeding up, but h how do you how do you sort of like keep the project centered around the big goal while still chopping wood on all the things that actually need to happen?
主持人： 嗯。那么，呃，多久……我我我感兴趣的是，一边是 Palantir 那套宏大、有点理想化的推销说辞：理解你的整个业务，把你的整个业务都跑在 Palantir 上；另一边是一些琐碎的实操细节，有时候最容易摘的果子就是：等等，居然有人的工作就是把表单上的内容敲进电子表格？我们早就有图像识别了啊，那就真正把它落地，把数据导进数据库、导进本体、导进 Palantir，然后才能在上面继续搭东西。感觉这两者之间可能存在某种张力。当然，两边都在提速，但你们是怎么让项目始终围绕大目标，同时又脚踏实地把那些真正必须做的事一件件干完的？
[0:05]
Yeah. And I think this comes back to the forward deployed piece and like what do we deliver outcomes and and we work backwards from that rather than hey I have this data I'm going to build a data warehouse and then I'll build reports because all my data is in one place that's the that's the field of dreams and no one shows up
right and so really when we demp things and work backwards from that you know the simple things like the form filling out there's a lot of that now the one approach that we see a lot is you know enterprise software is going to force you into their box.
Sure.
Right. You go fit you go fit into this box.
Yeah. Well, then you know, okay, did I take away the special sauce which was my company because people were doing these all these kind of amalgamations. Hey, 40 ways to do a PO.
Chad： 对。我觉得这又回到了 forward deployed 这一点：我们交付的是什么？是成果。我们从成果倒推，而不是“嘿，我有这些数据，我要建个数据仓库，然后再做报表，因为我所有数据都集中到一处了”。那就是《梦幻之地》式的幻想，建好了，却根本没人来。
对吧。所以当我们拆解问题、从成果倒推时，你知道，像填表这类简单的事确实有很多。不过我们经常看到的一种做法是，企业软件会逼你钻进它们的框框里。
主持人： 没错。
Chad： 对，你得去适应，把自己塞进这个框里。
嗯。可这样一来，你就得想想：我是不是把自家公司的独门秘方给弄丢了？因为原来大家做事有各种各样的混搭方式，比如，嘿，下个采购单（PO）就有 40 种做法。
[0:06]
Yep.
Well, maybe it is okay to do 40 ways, but my software can't handle it and it's fragmented, right? And so there's there's actually a middle ground because, you know, for a long time
customization was kind of a four-letter word, right? No, no one wanted to do that. And I think that's where we think about malleable software actually. How do we help you be more different, not more similar?
Interesting. And that's so that when we decomp problems thinking about not only the the kind of the uh quantitative piece but the qualitative piece and the people and process around this how do we enable those people to do the things that made them special?
Is is software getting more malleable because I I I can look at it two ways. I can look at uh one you know obviously AI agents are incredible at coding. They can run they they can make changes very very quickly that would take you a day in just a few minutes. Y
主持人： 是。
Chad： 不过也许有 40 种做法本身没问题，问题是我的软件应付不了，而且是碎片化的，对吧？所以其实存在一个中间地带，因为你知道，很长一段时间里
“定制化”差不多是个脏字，对吧？没人愿意碰。我觉得这正是我们思考可塑软件（malleable software）的出发点：怎么帮你变得更与众不同，而不是越来越同质化？
主持人： 有意思。Chad： 所以我们拆解问题时，不光考虑定量的那部分，也考虑定性的部分，以及围绕它的人和流程：怎么让这些人继续去做那些让他们与众不同的事？
主持人： 软件真的在变得更可塑吗？因为我我我可以从两个角度看。一方面，显然 AI agent 写代码厉害得惊人，它们能跑……它们改东西非常非常快，本来要花你一天的活，几分钟就搞定了。……
[0:07]
at the same time I see you know so many screenshots of people saying I implemented this feature and the GitHub is plus a million lines of code and at a certain point like the context window is growing as fast as the code generation's growing like there's a I I'm a believer in the answer to bad slop is good slop and more slop maybe but what are you actually seeing on the malleability of software because sometimes the most malleable software in the past has been oh well there was a really incredible engineer who figured out this problem and baked it down to a 2,000line repo and you can actually just put it in your own context window so it becomes more malleable and you can use it as a building block. Yep. And that feels like that's going away and I want to make sure that we've that we're ready for when it goes away and it remains malleable.
但与此同时，我也看到大量截图，有人说“我实现了这个功能”，GitHub 上一看，代码 +100 万行。到了某个程度，context window 涨得跟代码生成量一样快，就好像……我我倒是信奉“对付烂 slop（AI 垃圾产出）的答案是好 slop”，也许还得是更多 slop。但你们在软件的可塑性上实际看到的是什么？因为过去最可塑的软件，往往是这样来的：哦，有个特别厉害的工程师把这个问题琢磨透了，把它浓缩成一个 2000 行的 repo，你直接把它塞进自己的 context window，它就变得更可塑，可以拿来当积木用。对。可感觉这种东西正在消失，我想确保等它真消失的时候，我们已经准备好了，软件依然可塑。
[0:08]
Well, I think what what's missing is the the malleable enterprise scaffolding that you and that's what we think about the ontology and foundry and the platform and then Apollo that allows us to go deploy these changes. So it gives us the right amount of structure but the right amount of freedom. So I think that's the balance we try to find is that malle malleability in the middle where we can actually scale. We can enable people to do things differently while still creating enterprise you know grade robust secure scalable software. And so it's actually a balance there about how
Chad： 嗯，我觉得缺的是那层可塑的企业级脚手架（scaffolding）。你……这正是我们打造本体、Foundry、整个平台，以及 Apollo 的思路，Apollo 让我们能把这些变更部署出去。所以它给了我们恰到好处的结构，也给了恰到好处的自由。我觉得我们努力寻找的就是这个平衡，也就是中间那层可塑性，让我们真正能规模化：既能让人用不一样的方式做事，同时依然打造出企业级的、稳健、安全、可扩展的软件。所以这里其实是一种平衡，关乎怎么
[0:09]
I can enable that engineer that you know has been doing that now they can write code much faster. they can oversee things and that enterprise scaffolding in the middle allows us to actually create the right guard rails create a safe system of work for them to go develop things in. Um and then it's also the feedback loop. So the other thing that we do with our ontology and our platforms is implicit and explicit feedback from users using it. So the udal loop that I create and really that udal loop allows our customers as they're doing workflows they're giving feedback to agents. Now can agents help them do more based on the feedback. So both explicitly saying hey that was wrong that sucked or I chose this option. Now if you do that enough agents can start to learn from that. So we actually store that in our ontology to allow it to scale. So it's really that human centric process around AI. AI is not like we shouldn't be thinking about AI from the sake of AI for AI. It's AI to enable humans to do more. Yeah,
我能去赋能那位一直在做这件事的工程师：现在他们写代码快多了，可以去统筹把关，而中间那层企业级脚手架让我们能真正设好护栏（guardrails），为他们打造一个安全的工作体系，让他们在里面放手开发。嗯，另外还有反馈回路。我们用本体和平台做的另一件事，就是收集用户在使用过程中给出的隐式和显式反馈。于是就有了我搭建的 OODA 循环，这个 OODA 循环让客户在跑工作流的同时给 agent 反馈。那么 agent 能不能根据这些反馈帮他们做更多？这里既有显式的，比如直接说“嘿，这个错了”“这个太烂了”，也有“我选了这个选项”这种。反馈攒得够多，agent 就能开始从中学习。所以我们其实把这些反馈存进本体里，好让它能规模化。说到底，这就是一套围绕 AI、以人为中心的流程。我们不该为了 AI 而搞 AI，AI 是为了让人能做更多事。对，
[0:10]
that's the frame.
Udaloo, observe, orient, decide, act, right? Uh I have a different question, but you can you can go.
Uh if you were giving if you had 30 minutes
to uh give feedback to the AI labs, what are the kind of key areas? Let's say the frontier labs, right? Uh leading models. What uh what are the kind of key areas that you would be focused on?
Yeah, I mean I think when we think about the enterprise space, you know, we
one you're like don't compete with us. No, I actually like I I think optionality is a good thing. Like I am agnostic to where you store your data, where you store, how what model you choose, what compute you use. So like we we can allow you to use any of that because the last thing that actually drives an outcome is replatforming, moving to another
这就是整个思路框架。
主持人： OODA 循环嘛，观察、调整、决策、行动，对吧？呃，我有个别的问题，不过你先问，你先。
主持人： 呃，如果你要……如果给你 30 分钟
去给 AI 实验室提反馈，你会讲哪几个关键方面？就说前沿实验室吧，对，那些领先的模型。你会重点关注哪几个关键方面？
Chad： 嗯，我觉得从企业级市场的角度来看，你知道，我们
主持人： 第一条，你肯定会说“别跟我们抢生意”。Chad： 不，其实我……我我觉得有选择权是好事。你把数据存在哪、存在什么地方、选什么模型、用什么算力，我都不设限。所以这些我们都能让你用，因为真正能推动成果的事情里，最排不上号的就是换平台、迁到另一个
[0:11]
and that goes back to the onrem culture, the secure cloud culture, ITAR compliance, like this is in the DNA of the company.
Yeah. And so how do we actually enable people where they are instead of the focus on oh if you replplatform everything to Palanteer everything will be great. like well actually you've probably been replplatforming for years. Can we enable what you have to go do these new things? So when we think about like the model companies and it's you know how do we ensure that we can give the feedback loops around you know tool usage and um you know
yeah that's the kind of that's the kind of stuff I was uh wanting to get your point of view on is like are I'm sure you're getting into the nitty-gritty
这又要回到本地部署（on-prem）文化、安全云文化、ITAR（美国《国际武器贸易条例》）合规上，这些都刻在公司的 DNA 里。
对。所以我们思考的是，怎么在客户现有的基础上去赋能他们，而不是一门心思地说“哦，只要把所有东西都迁到 Palantir 上，一切就都好了”。其实呢，你很可能这些年一直在换平台。我们能不能让你现有的东西去做这些新的事？所以说到模型公司，嗯，就是我们怎么确保能在工具调用（tool usage）之类的方面给出反馈回路，还有，嗯，你知道的
主持人： 对，这正是我想听听你看法的那类事：你们……我相信你们肯定已经在抠细节了
[0:12]
with individual models where where they're spiky where there's you know where there's shortcomings etc.
Yeah. So we we actually just launched I just put a YouTube video out last week on this new tool called Evolve. We talked about it in the kind of the halftime show where customers are using actually AI to help them understand which model. So like maybe you know the the the the meme around hey make it exist first and then make it good.
Yeah.
Most of the time I see people a building with agents they're using the latest frontier model. I just got it working and then then all of a sudden the token maxing and everyone everything else and you're like oh my gosh I just blew through my whole budget. Yep. So we built a tool called evolve that will actually go analyze the logs in production about how these models are operating, what people are doing with them, the architecture over it, um, and actually be able to swap out different models from different providers or hey actually for most of this workflow you can use this model that's older and actually without thinking and and test time compute it. It's more deterministic
具体到每个模型：它们在哪些地方能力忽高忽低（spiky）、有哪些短板，等等。
Chad： 对。所以我们其实刚刚发布了……我上周刚发了个 YouTube 视频，讲的是一个叫 Evolve 的新工具。我们在那个类似中场秀的环节里也聊过：客户其实在用 AI 帮他们搞清楚该用哪个模型。就好比那个梗说的：嘿，先做出来，再做好。
主持人： 对。
Chad： 我看到大多数人用 agent 搭东西时，用的都是最新的前沿模型。“我终于把它跑通了”，然后突然之间，token maxing（狂烧 token）之类的全来了，你就会想：天哪，我整个预算一下子全烧光了。对。所以我们做了一个叫 Evolve 的工具，它会去分析生产环境里的日志，看这些模型实际怎么运行、大家拿它们在干什么、上面搭的是什么架构，然后真的能把模型换成不同供应商的其他模型，或者告诉你：嘿，其实这个工作流的大部分环节，你都可以用这个更老的模型，而且不用开思考模式，也不用 test-time compute。它的确定性更高
[0:13]
or even cached models
cache models and then or hey if you actually just have this piece of data in the ontology then you would eliminate all this and 50% of your cost. Yep. And so you know some of the customers McCarthy talked about this at our halftime you know they they were able to in two days eliminate 60% of their token cost by rearchitecting picking a different model and in prompt tuning. So it's the combination of all those the permutations get really hard especially when it's in this probabistic models we've have tools to do this in the deterministic world
prompt tuning it's a instead of don't make mistakes it's okay to make some mistakes if the mistake is going to cost just a little bit I'm fine because don't make mistakes that's going to cost me a fortune. Well, there there were there's there was some chatter yesterday around uh uh something a model was doing to be more efficient was uh talking in in like this bad.
Chad： 或者甚至是缓存模型
缓存模型，然后，或者说：嘿，其实你只要把这块数据放进本体里，这些就全都省掉了，成本还能砍掉 50%。对。所以你看，有些客户——McCarthy 在我们的 halftime show 上讲过这个——他们两天之内就把 token 成本砍掉了 60%，靠的是重做架构、换一个模型，再加上 prompt 调优。所以是这几样手段组合起来，排列组合会变得非常难，尤其是在这种概率性模型上；在确定性的世界里，我们是有工具来做这件事的
prompt 调优就是：别再写「不许犯错」，而是写「犯点错没关系，要是这个错只花一点点钱，我能接受」，因为「不许犯错」这一句，会让我花掉一大笔钱。
主持人： 话说，昨天网上有一波讨论，说某个模型为了更高效干了件事，就是用这种……很蹩脚的方式说话。
[0:14]
Oh. Oh. Kan caveman caveman prompting. Um
the caveman prompt method actually works. What how how often are you working with a company that is having call it like a mini chat GBT moment within their enterprise and then they're just like let's not tell anyone about this because I imagine like there's all these there's clearly places where
what does that mean their product is taking off like Chad GBT or
well so they've found a way to apply AI in a way that is highly highly effective and gives them an edge
oh interesting
uh but
yeah like like the theoretical like technology transform. So, so X people are very loud, right? Figure uh they they're like, I just had this
哦，哦！原始人……原始人式提示（caveman prompting）。嗯
原始人提示法还真管用。
主持人： 你们有多经常碰到这样的公司：在企业内部经历了一次——姑且叫它「迷你 ChatGPT 时刻」吧——然后他们就想「这事咱们谁都别说出去」？因为我能想象，肯定有这么多……显然有些地方
这是什么意思？是说他们的产品像 ChatGPT 那样火起来了，还是
嗯，就是说他们找到了一种应用 AI 的方式，效果极其好，能给他们带来优势
哦，有意思
呃，不过
对，就像理论上说的那种技术变革。你看，X 上那帮人嗓门都很大，对吧？他们就会说：我刚把这个
[0:15]
done using everything.
Yeah, I just had a product work for 30 hours on this thing. They'll talk about it. But if you're a Fortune 500 and you figure out how to do something, it's not like you want to like put put your hand up and say like, I figured something out, right? Like secrets are valuable and
these advancements and kind of breakthroughs are not going to be uniform. the airline industry will never be the same and then your direct competitor copies you and you're like,
"Yeah."
And so, and so part of part of why, you know, right now the meme is token maxing. Um, and that's an obvious going to be an obvious era area of debate. People are happy to go talk about it, say, you know, CEOs might say, "Hey, let's stop doing this." Um, but there has to be all these other kind of pockets of interesting moments where we won't hear about them until they become kind of like standard operating procedures
主持人： 用上所有工具搞定了。
对，「我刚让一个产品在这件事上连续干了 30 个小时」。他们会拿出来讲。但如果你是一家财富 500 强公司，搞明白了某件事该怎么做，你可不会想举手说「我搞明白了一件事」，对吧？毕竟秘密是值钱的，而且
这些进步和突破不会均匀地发生。「航空业从此不一样了」，然后你的直接竞争对手就把你抄了，你就只能：
「是啊。」
所以，部分原因在于，现在流行的梗是 token maxing（拼命烧 token）。嗯，这显然会成为一个争论焦点，大家也乐意公开去聊，比如 CEO 可能会说「嘿，咱们别再这么干了」。嗯，但肯定还有很多零零星星的有趣时刻，我们根本听不到，直到它们变成某种标准作业流程
[0:16]
or you see it in the the the uh earnings and the economics piece, right? Yeah. Yeah. So, I Yes. Unfortunately, X is not the real world,
you know, and there there's a lot of grift and noise and, you know, podcasting, PMing and, you know, that kind of stuff that goes on,
but I I I think in the real world, yes, there is the halves and have nots. I mean, we were just talking about AIG, like when you can start to actually do the underwriting and, you know, have quotes back in hours or days instead of months on these highly complex enterprise, you know, kind of insurance agreements.
If you don't have that, how are you ever going to compete? Yeah.
And so when we think about this of the N of one, right, you know, that those are the companies that we're going after and we see where there are those moments that are not public. It's the competitive advantage. interesting category because you can imagine AIG, you know, is um, you know, working with a potential customer or renewing a policy and that customer is going and talking to all of AIG's competitors. Yep.
主持人： 或者直到你在财报和经济账上看出来，对吧？
Chad： 对，对。所以我……是的。可惜 X 并不是真实世界，你懂的，那上面有大量的忽悠和噪音，还有什么播客、PMing，这类东西层出不穷，但我觉得在真实世界里，确实已经分出了「有」和「没有」两拨。我是说，我们刚才还在聊 AIG——当你真能把核保做起来，面对这些高度复杂的企业级保险协议，几小时或几天就能给出报价，而不是等上几个月。
如果你没有这种能力，你还怎么跟人竞争？对。
所以当我们从 N of 1（独一无二）的角度去想这件事，对吧，那些正是我们要去争取的公司，我们也能看到那些不为外人所知的时刻。这就是竞争优势。
主持人： 这个类别挺有意思，因为你可以想象，AIG 正在跟一个潜在客户谈，或者在续一份保单，而那个客户会去把 AIG 所有的竞争对手都谈一遍。对。
[0:17]
And uh, if AIG is able to turn around,
you know, a quote or a policy in 24 hours and then it takes another player,
you know, two weeks because it's, you know, complicated. So many so many teams will just say like hey we you know you know especially once you have two bids you can basically say like okay this that third fourth fifth we'll kind of wait on those because we have a good option here
well it builds trust the other piece here so when you think when you see people operating that with that level of efficiency what else can you do so I see this whether I'm doing you know SAP migrations the least sexy thing you could talk about but hey if I can cut your SAP migration
主持人： 然后呃，如果 AIG 能在 24 小时内把一份报价或保单做出来，而另一家却要花两周，因为，你懂的，情况复杂。很多很多团队就会直接说：嘿，我们……尤其是一旦手里有了两份报价，基本就可以说：好，第三、第四、第五家先放一放，因为我们这边已经有个不错的选择了
Chad： 嗯，另一方面，这还能建立信任。当你看到别人以这种效率在运转，你就会想：你还能做什么？所以我在各种事情上都看到这一点，哪怕我做的是 SAP 迁移——这大概是你能聊到的最不性感的话题了——但是嘿，要是我能把你的 SAP 迁移
[0:18]
let's give it up for
yeah it's like the the least you know exciting thing on on paper but Actually, if if you're spending hundreds of millions of Yeah, you guys get it. But hundreds of millions of dollars on a migration and we can cut it in half.
Yeah,
that's a massive deal.
Uh, back on the udaloop, observe, orient, decide, act on the observation side, what is the supply and demand imbalance for dashboards? Like, and what I mean by that is is when you're working with a company, is there is there more demand for dashboards? more people asking, hey, we need a dashboard for this, we need a dashboard for that. And you have to back people off and say, I don't know if the dashboard's right for this. Like, you might just want to do an ad hoc analysis or actually go and see. Uh versus you're seeing so much opportunity that you're like, okay, we want to push dashboards out everywhere. Like what? Walk me through dashboarding right now because I've always been like sort of like, oh, the too many dashboards. You build them and then no one looks at them.
主持人： 来，掌声送给——
Chad： 对，纸面上看这大概是最没劲的事了，但实际上，如果你要花几亿美元……对，你们懂的。在一次迁移上花几亿美元，而我们能帮你砍掉一半。
对。
这可是件大事。
主持人： 呃，回到 OODA 循环——观察、判断、决策、行动——在「观察」这一环上，仪表盘的供需失衡是什么情况？我的意思是，你们跟一家公司合作的时候，是需求更多吗？更多人跑来说「嘿，这个我们要个仪表盘，那个也要个仪表盘」，然后你们得劝他们收一收，说「我不确定这件事适不适合做仪表盘，你可能只需要做个临时分析（ad hoc analysis），或者真的亲自去看一看」？还是说，你们看到的机会多到让你觉得「好，我们要把仪表盘推到每个角落」？到底是哪种？给我讲讲现在仪表盘这块的情况，因为我一直有点觉得，唉，仪表盘太多了，做出来之后根本没人看。
[0:19]
Yeah, I want to kill all dashboards.
Okay, that's my perspective. dash I mean KPIs and dashboards should be a byproduct of operational applications where I'm making decisions. So we talk about the udaloop I have to actually act for things to hit the bottom line be valuable
in the actual application
in the application. So as I need those things and it's going to inform a better decision. Yeah,
that's where I want those metrics. That should be a byproduct. Not if I go out with the goal of building a dashboard, it's going to be the field of dreams again. No one shows up.
And so yes, it should be you you're going to have to build some of those things. The other side of this also is when you think about a data warehouse like literally I won't go too deep into this technical riff but like you know Kimble and dimensional modeling was built in 96 for scaling databases and you're still modeling in the same way in 2026
Chad： 嗯，我想把仪表盘全干掉。
好吧，这是我的观点。仪表……我是说，KPI 和仪表盘应该是运营应用的副产品，是我在里面做决策的那种应用。就像我们说的 OODA 循环，我得真正采取行动，事情才能落到利润上、才能产生价值
主持人： 在实际的应用里
Chad： 在应用里。所以当我需要这些东西、而且它们能帮我做出更好的决策时，嗯，那才是我想要这些指标出现的地方。它们应该是副产品。如果我一出发就奔着建仪表盘去，那又会是《梦幻之地》（Field of Dreams）那一套，结果根本没人来。
所以没错，你确实还是得建一些这类东西。另一方面，想想数据仓库——我不打算在这个技术话题上扯太深——但你知道，Kimball 和维度建模是 1996 年为了扩展数据库而诞生的，可到了 2026 年，你还在用同样的方式建模
[0:20]
for your dashboard your Tableau whatever those things are and like that's not actually how the world works in rows and columns you need complex things to model how the world really works and that's what we think about the ontology which means I can reuse it for an operational application KPIs agents all on one single ontology which it makes it the compound effect where as I add things in I'm now compounding with each individual decision I'm design working with gets better and better and better for the next use cases I connect across my business.
Chad： 为了你的仪表盘、你的 Tableau，不管那些是什么东西。可世界实际上并不是按行和列运转的，你需要复杂的东西才能对真实世界的运作方式建模，这就是我们对本体的思考。这意味着我可以把它复用到运营应用、KPI、agent 上，全都建在同一个本体之上，由此形成复利效应：我每加进来一样东西，都是在叠加复利，我处理的每一个决策，都会让我在整个业务里接入的下一个用例变得越来越好。
[0:21]
Yeah. Is there an analogy there to just the deployment of AI tools currently? I'm I'm just reflecting on the the NoSQL boom and I don't know how strong this was. This is probably just like an online take but this idea of like why would you ever want a relational database? Why would you ever want a schema? Don't never do a migration ever again. Uh and the future looked like a win-win almost like I I think Postgress installations probably grew and so did MongoDB and other non-reational databases. uh and people use Reddus for things and they use all sorts of different tools and we built and we stood on the shoulders of giants and we got more giants and then you know that means full employment for you obviously but uh but I'm wondering like as like are you seeing uh glimmers of of the AI tools eating into different pieces of the technical stacks or is it all like yes and across the enterprises? Um I think it's yes and and in in a couple different things there is when you think about the real world it is not just rows and columns you can't describe everything with measures and attributes
主持人： 嗯。这跟眼下 AI 工具的部署有没有可类比的地方？我只是在回想当年的 NoSQL 热潮，我也不知道那股风到底有多猛，可能只是网上的一种论调，就是那种想法：你干嘛还要关系型数据库？干嘛还要 schema？以后再也别做迁移了。呃，结果未来看起来几乎是双赢——我觉得 Postgres 的安装量大概是涨了，MongoDB 和其他非关系型数据库也涨了。呃，大家也用 Redis 干各种事，用各种各样的工具，我们一边建设，一边站在巨人的肩膀上，结果巨人越来越多，这显然意味着你们永远不愁没活干。不过我想问的是，你有没有看到一些苗头，AI 工具正在蚕食技术栈里的不同环节？还是说在各家企业里，全都是「yes, and」式的叠加？
Chad： 嗯，我觉得是「yes, and」，而且体现在几个不同的地方：当你去想真实世界，它不只是行和列，你没法用度量和属性去描述一切
[0:22]
and so it's actually multimodal and so like we think about this in our ontology where you can have one semantic object that actually has a CAD file and an image a CB model and tabular stuff in one semantic thing of a plant.
Yeah.
Which means I'm starting to talk in the language of my business. So being able to have the multimodal representation where in other places oh I have to have MongoDB and I have to have a SQL database here and I have to have an S3 bucket here to put all of these different things to store them in ways well we can do that all in the ontology vectors everything else so that that's really the the goal around how do I model the real world how it actually works and make that transparent so you're not having to figure out which technology to put in a time series thing for sensors on a you know oil platform
Chad： 所以它其实是多模态的。我们在本体里就是这么考虑的：你可以有一个语义对象，里面同时装着一份 CAD 文件、一张图片、一个 CB 模型，还有表格数据，全都在「工厂」这一个语义对象里。
主持人： 嗯。
Chad： 这意味着我开始用自己业务的语言说话了。所以能做到多模态表示——换在别的地方，哎，我得有个 MongoDB，这边还得有个 SQL 数据库，那边还得有个 S3 bucket，才能把这些不同的东西分门别类地存起来——而我们可以把这些全放进本体里，向量也好，别的一切也好。所以这才是真正的目标：我怎么按真实世界实际的运作方式去建模，并且让这一切透明化，这样你就不用去琢磨，比如一个石油平台上传感器的时序数据，到底该用哪种技术来存
[0:23]
don't care right and that that's where we want to have the non- differentiated heavy lifting like truly in the platform to remove the friction about getting stuff done.
How common is it for a business with more than a hund00 million of revenue to have very little understanding of how their business actually works? Like maybe they own maybe they own maybe they know like the main thing which is like you know
we make a product and try to sell it for more than it costs to deliver. Yeah. Um but uh but but is is some element of uh how how much can chaos and mystery be reduced effectively today? Because it feels like we're entering an era like
you go back um
Chad： 无所谓，对吧？这正是我们想做的：把那些无差异化的繁重工作真正放进平台里，消除把事情做成过程中的摩擦。
主持人： 营收超过一亿美元的企业，对自己的业务实际是怎么运转的几乎没什么了解，这种情况有多常见？比如他们可能……可能只知道最核心的那件事，就是，你懂的，
我们做一个产品，然后争取以高于交付成本的价格卖出去。对。嗯，但是……如今混乱和迷雾到底能被有效减少多少？因为感觉我们正在进入这样一个时代，就像
你往回倒，嗯
[0:24]
you know 50 years and
uh
the level of like mystery in a large company would have been like is almost unconceivable today, right? because you have different time zones, different offices, you know, no email, all that stuff. And now like mystery and chaos is probably
uh reduced dramatically, but uh still there's companies that that uh maybe maybe before you start working with them, I'm curious what those look like.
主持人： 你知道，倒回 50 年，呃，那时候一家大公司内部的迷雾程度，放到今天几乎是难以想象的，对吧？因为有不同的时区、不同的办公室，你懂的，没有电子邮件，诸如此类。而现在，迷雾和混乱大概已经大幅减少了，但是呃，还是有一些公司……也许在你们开始跟它们合作之前……我很好奇那些公司是什么样子。
[0:25]
Yeah, I mean we work with a lot of a lot of different varieties of companies. Um you know, I joke that a lot of times, you know, companies make money by accident. like they don't actually know what their most profitable product is and often they're trying to sell the thing that isn't isn't actually the most profitable and actually not selling the thing that actually is profitable and it comes back to how they've modeled their data is to aggregate it up to KPIs and other metrics when you actually need to model at the finest grade how your business operates to get a true cost of goods sold for example or true cost to serve like you that's very complicated it's very complex so like we really think about how do I embrace that complexity so that I can truly understand tactically at the edge how do I do more of the things that are good and less of the bad. It's that simple. And those get peanut buttered across with KPIs and metrics. And people don't actually know how their businesses are operating. I can't tell you whether it's a hundred million dollar company or a $50 billion company how many times I see this that they don't actually understand how they're making money at a fine grain.
Chad： 是啊，我们合作过的公司类型非常多。嗯，我常开玩笑说，很多时候公司赚钱纯属意外。他们其实不知道自己最赚钱的产品是哪个，还经常卖力推销那个其实并不是最赚钱的东西，反倒没去卖真正赚钱的那个。归根结底，问题出在他们给数据建模的方式——把数据往上汇总成 KPI 和其他指标；可实际上，你需要以最细的粒度对业务的运转方式建模，才能算出比如真实的销货成本（COGS），或者真实的服务成本（cost to serve）。这件事非常复杂、非常繁琐，所以我们真正在想的是：怎么拥抱这种复杂性，让我能在一线、在战术层面真正搞明白，怎么多做好的事、少做坏的事。就这么简单。而这些东西都被 KPI 和各种指标像抹花生酱一样摊平了，大家其实并不知道自己的业务是怎么运转的。不管是一亿美元的公司还是 500 亿美元的公司，这种情况我见过多少次都数不清了：他们其实并不在细粒度上清楚自己到底是怎么赚钱的。
[0:26]
Yeah.
Uh last question. Uh, is there a world in the future where a company gets created, let's say on Stripe Atlas, and the first account they sign up for
other than that is, let's say, a Palunteer.
Yes, that's interesting.
I would love that. And so, we do have a Palanteer for builders program. We have small companies that there's people here that are two person startups, you know, that are working in their attict in Canada. I mean, like, so it it is literally um any size company. Come come work. There's a free dev tier people can come build. You actually there's actually a Shopify integration in Palunteer. You can go hook up to your Shopify and pull in Palanteer. Yeah, there are people doing this. Now, are we always great at selling it or telling the story? Sure. No.
对。
主持人： 呃，最后一个问题。将来会不会有这么一天：一家公司刚成立，比如说是在 Stripe Atlas 上注册的，而除此之外它开通的第一个账号，就是 Palantir？
Chad： 是啊，这个想法挺有意思。那我可太乐意了。我们还真有个 Palantir for Builders 计划。我们也有小公司，我们这儿就有两个人的初创团队，你知道吧，就在加拿大自家阁楼里干。我是说，真的是任何规模的公司都行，都欢迎来干活。我们有免费的开发者版（dev tier），谁都能来上手搭东西。而且 Palantir 里其实还有 Shopify 集成，你可以直接连上自己的 Shopify，把数据拉进 Palantir。对，真有人在这么干。话说回来，我们是不是一直都很会推销、很会讲这个故事？当然——不是。
[0:27]
But but there are companies doing this and I do think there's a day where it's going to be ubiquitous because I also think, you know, there's some some guys here that have, you know, they hey, my my business is dying. I was, you know, I was down 10% negative margin on on what I was selling and through using Palanteer. I they watched our YouTube videos and they built it themselves and increase to a 9 or 10% positive margin in three months.
That's great.
And so like people can go do it. I think that's the great American story is like how do we enable that and I think we'll get there. Um it might take a little time.
I love it. Well, thank you so much for taking the time. Great to see you. We will talk soon. Uh our next guest is joining in just 15 minutes. We're going
但确实有公司在这么做，而且我真心觉得总有一天这会变得无处不在。因为我也觉得……你知道，我们这儿就有几个哥们，他们当时的情况是：“嘿，我的生意快撑不下去了，我卖的东西利润率是负 10%。”后来靠着用 Palantir——他们是看了我们的 YouTube 视频自己动手搭出来的——三个月就把利润率拉到了正 9% 到 10%。
主持人： 太厉害了。
Chad： 所以说，大家是真的能自己做到的。我觉得这就是那种伟大的美国故事——我们要做的，是怎么让这种事成为可能。我觉得我们会做到的。嗯，可能还得花点时间。
主持人： 我太喜欢这个了。好，非常感谢你抽空过来，见到你很高兴，我们回头再聊。呃，我们的下一位嘉宾再过 15 分钟就到。我们要……