The Investor's Exoskeleton

All right.

好的。

Hello everyone.

大家好。

Thank you for being with me today.

感谢各位今天抽空来听我讲。

I I plan to do a short piece of curriculum here.

我本来打算在这里做一小段课程内容。

This ended up as an 82 slide deck.

结果做成了一份 82 页的幻灯片。

So I went a little crazy cuz there's so many rabbit holes to go down.

我有点收不住了，因为可以深挖的兔子洞实在太多。

To preempt a common question, yes, there will be recording sent out after and yes, we will share the PowerPoint deck.

先预答一个常见问题：是的，结束后会发回放；是的，我们也会把这份 PowerPoint 分享出去。

So if you for some reason have to drop off before we will done we are done, we will certainly not complete this all in an hour.

所以如果你出于某种原因得在结束前提前离场——我们肯定没法在一个小时里把这些全讲完。

I'm guessing maybe 75 minutes and then we'll we'll take we'll take questions at the end.

我估计大概要 75 分钟，然后最后留时间答疑。

Because there's a lot to there's a lot to get through.

因为要讲的东西真的很多。

I found myself having this similar conversation with so many people lately and I figured I would just put it all in a deck.

我最近发现自己在跟很多人反复聊着同样的内容，于是干脆把它全整理进一份幻灯片里。

And maybe that will make my conversation some of my conversations more more efficient.

也许这样能让我的对话——至少其中一部分——更高效些。

So I will go ahead and share my screen.

那我这就来分享屏幕。

And I'm really here to present to you today hypothesis, not a thesis.

今天我要向各位呈现的，是一个假说，而不是一个定论。

A hypothesis on where I think the world of AI investing is is going.

一个关于我认为 AI 投资这个世界将走向何方的假说。

So with that, for those of you who don't know me I spent 13 years as an investor.

先跟不认识我的各位说一下：我做了 13 年投资人。

So I approach all of these these questions from the investor seat not from the technologist seat.

所以我是从投资人的位置、而不是技术专家的位置来看待这些问题的。

I worked at places like Maverick, Citadel, DE Shaw, Schonfeld.

我在 Maverick、Citadel、DE Shaw、Schonfeld 这些地方工作过。

Um retired from managing money 5 years ago and I've become a professor at ASU and Fundamental our core business is Analyst Academy, a training program.

5 年前我退出了资金管理一线，成为 ASU(亚利桑那州立大学)的教授；而 Fundamental Edge 的核心业务是 Analyst Academy，一个培训项目。

As part of that program, AI has become a real interesting topic.

作为这个项目的一部分，AI 已经成了一个非常有意思的话题。

I find it endlessly fascinating.

我对它的兴趣简直没有尽头。

Sort of accidentally in my career I found myself at the intersection of fundamental investing and machine learning.

在职业生涯里，我算是误打误撞地站到了基本面投资和机器学习的交叉点上。

We don't call it machine learning anymore, we call it AI.

我们现在已经不叫它机器学习了，改叫 AI。

And so I developed a a relatively strong base of priors on the opportunities and the drawbacks from combining fundamental and machine learning signals.

于是，对于把基本面信号和机器学习信号结合起来所带来的机会与弊端，我形成了一套相对扎实的先验判断。

And so that sort of builds builds some some base of evaluating the new tools that we're evaluating in the in the process.

这也就为我们在流程中评估这些新工具打下了一定的评判基础。

And I found myself accidentally but quite entertainingly advising funds and asset managers on the adoption of AI.

而我又误打误撞、但过程相当有趣地，开始为一些基金和资产管理机构在采用 AI 上做顾问。

And we we hope to do more of that.

我们也希望今后能多做一些这类事情。

So the question of the the conversation today is how will AI change fundamental investing?

所以今天这场对话要探讨的问题是：AI 将如何改变基本面投资？

I don't think anyone knows that answer for sure.

我觉得没人能确切知道这个答案。

I'm here to make my case and the fundamental concern is that we all become context for the AI, right?

我来是要论证我的观点，而最根本的担忧在于：我们所有人都变成了 AI 的 context。

That we're feeding in management meeting notes.

也就是说，我们把管理层会议纪要喂给它。

The AI is making all the the decisions.

由 AI 来做所有决策。

We'll explore that a little bit today.

今天我们会稍微探讨一下这个问题。

But one thing I've I've observed through numerous conversations over the last 18 months on AI is that fewer asset managers are really advanced in deploying AI than one might think, right?

但过去 18 个月里，我在大量关于 AI 的对话中观察到一点：真正在部署 AI 上走得靠前的资产管理机构，比人们想象的要少。

There's certain verticals of the industry of the economy like coding where the tools are quite advanced.

在经济中的某些细分领域，比如编程，工具已经相当先进了。

They're not that advanced in institutional asset management really for two reasons.

但在机构级资产管理里，它们其实没那么先进，主要有两个原因。

Principally trust.

首要的是信任。

You know, hey, my process is reliable today.

你会想：我现在的流程是可靠的。

If you're on the investment team at a high performing asset manager, you don't really want to mess with your process.

如果你在一家业绩优异的资产管理机构的投资团队里，你其实并不想去动自己的流程。

There's concerns about hallucination, confirmation bias, low quality sources and big compliance concerns around whether it's safe to upload internal documents.

大家会担心 hallucination(幻觉)、确认偏误、低质量信源，还有围绕上传内部文件是否安全的巨大合规顾虑。

And so no institutional hedge funds I know are running internal cloud bots.

所以据我所知，没有一家机构级对冲基金在内部跑 Claude 机器人。

Autonomous research etc.

自主研究之类的东西。

Most are using Microsoft co-pilot co-work or really in the early stages of getting internal approvals for cloud cloud cloud enterprise.

大多数用的是 Microsoft Co-pilot、Co-worker，或者其实还处在为 Claude Enterprise 争取内部审批的早期阶段。

So this is still a work in progress.

所以这块仍在推进中。

I think the more fundamental concern or inhibitor to scale has been usability.

我认为更根本的顾虑、或者说规模化的阻碍，一直是可用性。

I don't know about you, I find certain elements of the LLM chatbot experience helpful.

我不知道你怎么看，我个人觉得 LLM chatbot 体验里的某些环节挺有用。

But you end up with this blank blank page problem where chatbots have been more like a better form of Google less reliable form of Google in some in some in some ways.

但你最后会遇到那种空白页难题——在某些方面，chatbot 更像是一个更好用的 Google，或者说一个不那么可靠的 Google。

It's been quite cumbersome to upload the documents that it takes to build a system around chatbots.

要围绕 chatbot 搭建一套系统，得上传的那些文件，操作起来一直相当麻烦。

So with the recent innovations in agentic capabilities, this is improving a lot as well, too.

所以随着近来 agentic 能力上的创新，这一点也在大幅改善。

I'd like to remind people, remind myself particularly with so much productivity theater out on Twitter and YouTube, hey, I can turn everything into an automation.

我想提醒大家、尤其提醒我自己：在 Twitter 和 YouTube 上有那么多生产力表演，总让人觉得“嘿，我可以把一切都变成自动化”。

The question is should you?

但问题是：你该不该这么做？

And so I'd like to start these chats with just reminding people that our objective as active managers is to beat the market, right?

所以我喜欢在这类交流一开始就提醒大家：我们作为主动管理人的目标，就是跑赢市场。

Whether that's spread in a multi-manager context or alpha in a long only context.

无论这体现为 multi-manager 语境下的 spread，还是 long only 语境下的 alpha。

That's really the objective that we're we're serving here today.

这才是我们今天真正要服务的目标。

I like to think about that objective as served by the hours I can deploy towards that goal in a given year.

我喜欢这样理解这个目标：它靠我一年里能投入到这个目标上的小时数来达成。

So roughly 3,000 hours with the objective of beating the market.

大约就是 3,000 小时，目标是跑赢市场。

So with that framing, the sort of question of does AI generate alpha becomes more clear.

有了这个框架，"AI 能不能创造 alpha"这个问题就变得更清晰了。

I think it becomes to me sort of an emphatic yes now in two buckets.

在我看来，现在这个答案是一个斩钉截铁的"能"，而且分成两块。

One in the simple sort of fundamental terms of offload alpha.

一块是简单的、基本面意义上的 offload alpha(卸载型 alpha)。

I'll give some examples of that.

我会举一些例子。

I spend less time on lower quality motion where there's not really serving comprehension or insight differentiation.

我花在低质量动作上的时间更少了——那些并没有真正服务于理解力或洞见差异化的动作。

So I can deploy those hours back in the real sort of meaty part of the investment process.

这样我就能把这些小时重新投回到投资流程里真正有分量的部分。

That I think has been true for the last number of months.

我认为这一点在过去这几个月里已经成立了。

I think what's emerging now, which is really exciting, is the the uncovering of signal alpha.

我觉得现在正在浮现、也真正令人兴奋的，是 signal alpha(信号型 alpha)的发掘。

I can cast a wider net find inflections and KPIs in a systematic way that that ultimately that ultimately can lead to signals to either add or add or reduce or buy a new position.

我可以撒下更大的网，以系统化的方式找到拐点和 KPI，而这些最终能转化为信号，去加仓、减仓或买入一个新的仓位。

So I'll walk through both of those today.

所以今天我会把这两块都讲一遍。

But really the mental model of what I'm here to present is what I call the exoskeleton.

但我今天真正要呈现的心智模型，是我所说的外骨骼(exoskeleton)。

It's sort of hypothesis that the merging of man and machine has been the ultimate ultimate target of quantamental strategies for two decades.

它算是一个假说：人与机器的融合，过去二十年一直是 quantamental 策略的终极目标。

I'm not sure that anyone has really scaled that.

我不确定是否真的有人把它规模化过。

I could write a book probably on why why it's difficult.

关于为什么它这么难，我大概能写一本书。

But the evolution of these tools I think creates an environment where we don't have to sit with a fundamental team and a quant team in the room and hash out how to execute that strategy.

但我认为，这些工具的演进创造了一种环境，让我们不必再让一支基本面团队和一支 quant 团队坐在同一个房间里，去反复琢磨怎么执行这套策略。

That we as individual fundamental investors can now wrap our process around an exo exoskeleton.

而我们作为单个的基本面投资人，现在就能把自己的流程包裹在一副外骨骼之上。

And so there's really three parts of that process.

这个流程其实分成三个部分。

It's it's observation AI isn't perfect.

第一是观察：AI 并不完美。

There's still sort of things that need to improve.

还有一些地方需要改进。

But in my opinion, it's officially now useful, right?

但在我看来，它现在正式变得有用了。

Really in the last 3 months I've seen new evolutions and outputs that I would deem personally to be institutional grade.

真的就在过去 3 个月里，我看到了一些新的演进和产出，我个人会把它们评定为机构级。

You can make your own and please do make your own conclusions on what's institutional grade.

你可以、也请务必自己去判断什么才算机构级。

But the systems that you can wrap around these processes now sort of taken many of these workflows from a red light or yellow light to a green light for me.

但在我看来，你现在能围绕这些流程搭建的系统，已经把其中很多工作流从红灯或黄灯推到了绿灯。

And so I'm I'm excited to share a few of those with you today.

所以我很兴奋，今天要跟大家分享其中几个。

We can go into a long philosophical rant about fundamental investing, the irreducible value of human in investing.

我们可以就基本面投资、以及人在投资中无法被替代的价值，来一场很长的哲学式抒发。

I'll go into a short rant today about that.

今天我会就这一点做一小段抒发。

But I feel strongly that AI doesn't replace, it augments the process.

但我强烈地认为，AI 不是取代流程，而是增强流程。

And so that's really the the next 74 slides sort of walk through a few of these things.

所以接下来的 74 页幻灯片，基本上就是把这几件事走一遍。

you part of learning new spaces is figuring out who to trust.

学习一个新领域，有一部分工作就是搞清楚该信任谁。

Um One of the most trustworthy voices in my opinion is Andre Karpathy, one of the founders of Open AI and former head of AI at Tesla and just an all around sort of clear very clear thinker.

在我看来，最值得信任的声音之一是 Andre Karpathy(Andrej Karpathy),OpenAI 的联合创始人之一、前 Tesla AI 负责人，而且整体上是个思路非常清晰的人。

Um and so I I sort of use his vibe shift as a encapsulation of the moment where in October October 2025, he went on Dwarkesh and said agents are slop, we'll be working through these issues for 10 years.

所以我把他的态度转变当成这个时刻的一个缩影：2025 年 10 月，他上了 Dwarkesh 的节目，说 agent 就是垃圾，我们还得花 10 年才能把这些问题解决掉。

There was sort of this inflection point over Christmas where people like Andre started getting going with cloud code, the coding agents.

圣诞节前后出现了一个拐点，像 Andre 这样的人开始上手用 Claude Code、用编码 agent 了。

And in sort of his encapsulation, it happened in a few few weeks.

而按他的说法，这一切就在短短几周内发生了。

The thing he's been saying now is the intelligence part suddenly feels quite a bit ahead of the rest of it.

他现在一直在说的是，智能这部分突然感觉比其余部分领先了相当一截。

There's we've gone from agents or slop to an intelligence overhang.

我们已经从「agent 是垃圾」走到了「智能过剩(intelligence overhang)」。

What does that mean?

这是什么意思？

It means that the organizational workflows, processes are now more powerful than most of us are using them for.

意思是，组织的工作流、流程如今所具备的能力，已经超过了我们大多数人实际用它们做的事。

In my experimentations, I've been able to validate that.

在我自己的实验里，我已经能验证这一点。

I'm sort of shocked by the the workflows that I can do with these tools now.

我现在用这些工具能跑出来的工作流，让我有点震惊。

And so the exciting part is this happened just in the last just in the last 3 months, right?

而令人兴奋的是，这一切就发生在过去这短短三个月里。

This is all incredibly fresh.

这全都是新得不能再新的事情。

And so in my hypothesis I sort of mark, you know, holidays 2025 as the as the entering of the new agentic era, which which matters a lot for investors.

所以在我的假说里，我把 2025 年年末假期这段时间标记为进入全新 agentic 时代的起点，而这对投资人来说非常重要。

Because chatbots are great for certain things.

因为 chatbot 在某些事情上很棒。

I've used chatbots for personal personal for personal use cases a lot.

我在很多个人场景里都大量用过 chatbot。

If I'm barbecuing steak, chicken, burgers and hot dogs, I it tells me perfectly when to time, when to flip, etc.

如果我在烤牛排、鸡肉、汉堡肉和热狗，它能完美告诉我什么时候该翻面、时间怎么把控等等。

If I need a if I need a pumpkin pie recipe, it's elite for that.

如果我需要一份南瓜派食谱，它在这方面顶尖。

But chatbots for investment process has some issues.

但把 chatbot 用在投资流程上就有一些问题。

They're cumbersome, there's no agency, they're poor poor at sourcing and retrieval natively, they hallucinate.

它们笨重、没有 agency(自主行动力)，原生的信息检索与调取能力很差，还会出现幻觉。

And before you sort of talk about the tool calling, LLMs are really natively bad at math.

而且先不说工具调用这一层，LLM 在原生的数学能力上是真的很差。

Well, that's kind of an issue in in deploying LLMs for for investment process.

而这在把 LLM 部署到投资流程时，算是个问题。

So I've I've found very specific use cases that are kind of interesting, but chatbots and chatbot wrappers haven't been that helpful for institutional investing, particularly on a zero-shot or one-shot approach.

所以我找到了一些非常具体、还挺有意思的用例，但 chatbot 以及 chatbot 套壳产品对机构级投资并没有那么大帮助，尤其是在 zero-shot 或 one-shot 的用法下。

If I go into ChatGPT and say build me an investment thesis in Jack in on Jack in the Box, it's definitively AI slop, even even even today, in my opinion.

如果我进 ChatGPT 说「帮我建一个关于 Jack in the Box 的投资论点」，在我看来那绝对是 AI slop(AI 垃圾产出)，哪怕是今天也一样。

I'll walk you through why I think that's the case.

我会带你们过一遍我为什么这么认为。

What I've sort of found and spent a lot of, you know, 2025 building ideating around is a system systems approach.

我大致找到、并且花了 2025 年很多时间去构思打磨的，是一套系统化的方法(systems approach)。

And what I found is that the systems approach is the key leverage point when taking a one-shot approach, whether that whether that's true in current agentic systems or old chatbots, that if you think about a structured process and approach around these tools, you can really get much better better outputs.

我发现，在采用 one-shot 用法时，系统化方法才是关键的杠杆点——无论这在当下的 agentic 系统里成立，还是在老式 chatbot 里成立——只要你围绕这些工具去设计一套结构化的流程和方法，你就真的能得到好得多的产出。

That included in 25 prompting and uploading the filings and transcripts and creating custom workflow documents and putting in example outputs.

在 25 年，这套做法包括写 prompt、上传各类 filing 和 transcript、制作自定义的工作流文档，并放入示例输出。

We ran an AI Academy in in in um for 150 analysts and PMs in the in uh uh September 2025.

我们在 2025 年 9 月办了一场 AI Academy，面向 150 位分析师和 PM。

Much of that now feels decayed, but that was a core part of what we taught is a systems approach to integrating with integrate integrating with AI.

其中很多内容现在感觉已经过时了，但我们当时教的核心之一，就是与 AI 整合的系统化方法。

So, if you take nothing else away from today, a systems approach is important.

所以，如果你今天只带走一点东西，那就是：系统化方法很重要。

It's a a one-shot a one-shot interaction with AI is not going to get you an institutional grade uh approach.

跟 AI 做一次 one-shot 的交互，是不会给你带来机构级的方法的。

Building your own system around these use cases is really really really important.

围绕这些用例搭建你自己的系统，真的真的非常重要。

To me, there's a few few few sort of steps of that system building that are that are important.

在我看来，构建这套系统有几个重要的步骤。

One is just building that intuition with AI.

第一个，就是建立你对 AI 的直觉。

So, when AI makes a silly mistake on math or dates or other things, pulls a number of Nvidia EPS and it's the wrong number, but it pulled it from a blog, you have some intuition around the jagged edges, and you can debug those jagged edges by saying, "Hey, don't go pull to a blog.

所以当 AI 在数学、日期或其他东西上犯了个低级错误，比如去拉 Nvidia 的 EPS，拉出来的数字是错的，而它是从某个博客上拉的，你就会对这些 jagged edges(参差短板)有点直觉，你可以这样去 debug 这些 jagged edges:"嘿，别去博客上拉。"

Make sure you're pulling this from from from Edgar." Also understanding the superpower.

确保你是从 Edgar 上拉的。"同时也要理解它的超能力所在。

So, some of it is just the native intuition, you know, building some muscle memory.

所以其中一部分就是原生的直觉，建立起某种肌肉记忆。

We encourage people to find personal use cases to build this intuition.

我们鼓励大家找到个人的使用场景来建立这种直觉。

That's a tool selection, data strategy is incredibly important.

这就是工具选择，数据策略极其重要。

The sort of uploading context, shifting now to skills and MCPs has been a big unlock for me.

从上传 context，到现在转向 skills 和 MCP，对我来说是一次很大的解锁。

It's building the workflows, prompts, and skills customized to you, really wrapping your process around this exoskeleton.

关键是打造定制化的 workflow、prompt 和 skill，真正把你的投资流程包裹在这副外骨骼(exoskeleton)之上。

Probably one of the most important points in this uh uh chart is a two-tiered two-tiered validation system.

这张图里可能最重要的一点，就是这套两层验证(two-tiered validation)系统。

I'll walk through finance benchmarks, but benchmarks LLMs are sort of 60-70% accurate on finance benchmarks, which 60% accurate on Wall Street will get you 100% fired.

我会讲一下金融基准测试，LLM 在金融基准测试上的准确率大约是 60-70%，而在华尔街，60% 的准确率会让你被彻底炒鱿鱼。

And so, building a two-tiered systematic and analog system around validation, I think is really important.

所以围绕验证建立一套两层的、系统化加人工的体系，我觉得真的很重要。

I'm now getting model updates that are 100% accurate in 15 to 20 minutes.

我现在能在 15 到 20 分钟内做出 100% 准确的模型更新。

Sort of shocking that uh I've gotten to that point, but I'm able to do I'm able to do it.

我能做到这一步，某种程度上挺震撼的，但我确实能做到。

So, I'm trying to flow with that reality.

所以我在努力顺着这个现实往前走。

And then creating a system that enhances comprehension and rigor.

然后是打造一套能增强理解力和严谨度的系统。

I hear a lot from investors that, "Hey, I used AI to get up to speed, but I don't really feel like I understand the business as well." And my rebuttal is, "Well, you're using it wrong, right?

我常听投资人说："嘿，我用 AI 快速上手了，但我并不觉得自己对这门生意的理解更深了。"而我的反驳是："那你就是用错了。

There's no rule you can't still go and read the 10-K, but deploy the tool in certain other areas in the investment process that enhance comprehension and rigor.

没有哪条规矩说你不能再去读 10-K，而是要把这个工具部署在投资流程中其他能增强理解力和严谨度的环节。

I think the ultimate goal is to get done in in 8 hours what used to take 10, but that 8-hour output is substantially more comprehensive, rigorous, and validated than before.

我觉得终极目标是把过去要花 10 小时的活儿在 8 小时内做完，但这 8 小时的产出比以前要全面、严谨、经过验证得多。

We're not trying to 10x our investment process here um in my in my in my opinion.

依我看，我们在这儿并不是要把投资流程做到 10 倍。

I don't want an analyst who 10x's idea flow and gives me sloppy consensus consensus consensus ideas.

我不想要一个把 idea flow 提升 10 倍、却给我一堆草率的、随大流的 consensus 想法的分析师。

So, there have been some real limitations of chatbots in investment research.

所以 chatbot 在投资研究里确实存在一些真实的局限。

I you know, I sort of think about the three-layer cake of investment process, and there's sort of, you know, by by nature of what institutional investors do, there's only so much that is done behind a desk, right?

我会把投资流程想成一个三层蛋糕，而按照机构投资人做事的本质，能坐在办公桌后面完成的部分其实是有限的。

A lot of the value in investment process happens in primary research.

投资流程中很多价值发生在一手研究里。

It happens by getting on the call, by getting on calls.

它发生在你去打电话、去和人通话的过程中。

It happens by getting out in the field.

它发生在你走出去、深入实地的过程中。

And the sort of more mystical part of the process happens sort of from beginning to end to end, the insight formation.

而流程中更玄妙的那部分，是从头到尾贯穿始终的——洞见的形成。

And so, a lot of insight formation investing might happen while you're slowly reading the 10-K or while you're sitting in the audience at the investor day or in the halls of a conference talking to another investor you you uh you you respect.

所以很多投资洞见的形成，可能发生在你慢慢读 10-K 的时候，或者坐在投资者日的观众席上，又或者在会议的走廊里和另一位你尊敬的投资人交谈的时候。

And so, that sort of is is a fundamental challenge of the overlay, and sort of a simple cognitive, you know, analog I I I use is I could read CliffNotes or speed read The Great Gatsby, but that's not going to sit in your soul the same way that a slow, considered reading of the book is.

所以这算是这层叠加之上的一个根本挑战，我常用的一个简单的认知类比是：我可以读 CliffNotes 摘要，或者速读《了不起的盖茨比》，但那不会像你慢慢地、用心地读完整本书那样沉淀进你的灵魂。

And so, this this might seem philosophical, but I think it's really important when it comes to the really difficult job of of driving differentiated insights.

所以这听起来可能很哲学，但我觉得当涉及形成差异化洞见这件极其困难的工作时，它真的很重要。

And across when I covered 300 stocks, there might only be a dozen or two stocks where I felt like I had real differentiated insights.

在我覆盖 300 只股票的那段时间里，可能只有十来只、二十只股票让我觉得自己真的拥有差异化的洞见。

So, the bar is very high, and we would sort of point that out to say that speed running the investment process to 10x your productivity is almost certainly going to be equivalently damaging to your investment process as speed reading The Great Great Great Gatsby is to your enjoyment of literature.

所以门槛非常高，我们想借此指出：为了把生产力提升 10 倍而把投资流程跑得飞快，几乎必然会对你的投资流程造成损害，就像速读《了不起的盖茨比》几乎必然会损害你对文学的享受一样。

Um so, that's sort of, you know, philosophical soapbox.

嗯，这算是我站在哲学高地上的一点发挥吧。

There have been certain elements like deep research I've found to be really nice.

有些特定的功能，比如 deep research，我发现真的很好用。

I really like it with the key stipulation that you need to point it towards information that you know sits in the open web.

我非常喜欢它，但有个关键前提：你得把它指向那些你知道存在于开放网络上的信息。

When you start to ask deep research things that doesn't doesn't fit in the training corpus, that's where you start to get hallucination slop or regurgitation.

当你开始拿一些不在训练语料里的东西去问 deep research，那你就会开始得到幻觉式的糊弄或复读。

Um so, for simple things like, "Hey, I want to go study, you know, what's happening in the tax rate at the 23 different states, go into the state filings, and give me a report on how to model that bull base and bear." I found that to be a nice accelerant in in an an investment process.

所以对于简单的事情，比如："嘿，我想研究一下 23 个不同州的税率情况，去查州级的申报文件，给我一份报告，告诉我怎么去做牛市、基准、熊市的建模。"我发现它在投资流程里是个很好的加速器。

But chatbots are not great, right?

但 chatbot 并不出色。

And one of the, you know, other fundamental issues of chatbots, if you really dig into the training corpus, right?

而 chatbot 另一个根本性的问题，如果你真的去深挖它的训练语料的话——

It's highly dominated by blogs and Buffett, right?

它高度被博客和 Buffett 主导。

That's that's what that's what uh these LLMs have access to.

这就是这些 LLM 能拿到的东西。

That's not the worst thing in the world, um but generally, and there's some, you know, there's some good articles.

这倒不是世界上最糟糕的事，但总的来说，里面也确实有一些好文章。

I don't want to throw shade at Seeking Alpha, Motley Fool, or investing.com, but on average, the median quality of content you'll see in the top 50 finance investing blogs that become the training corpus or the Buffett that are influenced by Buffett and his derivatives.

我不想贬低 Seeking Alpha、Motley Fool 或 investing.com，但平均而言，你在排名前 50 的金融投资博客里看到的内容——那些成为训练语料的东西——中位质量，要么是 Buffett，要么是受 Buffett 及其衍生流派影响的。

You walk through Barnes & Noble, most of the books are value investing value investing derivative books.

你走一趟 Barnes & Noble，大部分书都是价值投资、价值投资衍生的书。

The value investing can is massively overrepresented online relative to their share of actual AUM.

相对于价值投资在实际 AUM 中所占的份额，它在网上被严重过度代表了。

So, if you're to multi-manager trying to use these tools to do earnings previews, just understand there's a fundamental differentiation in the training corpus of the these tools.

所以如果你是一个 multi-manager，想用这些工具来做财报前瞻，你要明白这些工具的训练语料里存在一个根本性的偏向。

Other point is that I'm just pure pure LLM fundamentals.

另一点是纯粹从 LLM 的基本原理来说。

By sort of by nature, the hunt for alpha is an expectations gap game, right?

从本质上讲，对 alpha 的追逐是一场预期差的游戏。

Our job is to understand what's priced in, what consensus is.

我们的工作是理解什么已经被 price in、consensus 是什么。

And there's an idea is only interesting definitionally to the degree that we have a differentiated perspective.

而从定义上说，一个想法只有在我们拥有差异化视角的程度上才是有趣的。

If I if I look at a $20 stock that I think is worth 20, why am I deploying investor capital in that situation?

如果我看一只 20 美元的股票，我认为它就值 20 美元，那我为什么要在这种情况下部署投资人的资本呢？

And so, if you think about LLM mechanics, they gravitationally drift towards the most commonly expressed expressed view on a given company.

所以如果你想想 LLM 的机制，它们会像被引力牵引一样，漂向对某家公司最普遍被表达出来的观点。

They pattern match the pattern match of the modal opinion.

它们做的就是对众数观点的模式匹配。

An investor who pattern matches to the modal opinion is going to be an investor that that doesn't generate alpha over a over over a cycle.

一个对众数观点做模式匹配的投资人，将会是一个在整个周期里都无法产生 alpha 的投资人。

So, understanding that consensus bias is important is important.

所以理解 consensus 偏向这件事很重要。

And using these tools, one of the big fundamental challenges we face is the is the reality that if you've been in the seat for 10 or 15 years, a lot of what you do is tacit.

而在使用这些工具时，我们面临的一大根本挑战，是这样一个现实：如果你在这个位子上坐了 10 年、15 年，你所做的很多事情都是隐性知识。

It's almost sort of involuntary.

它几乎是某种不自觉的本能。

When I look at an idea, I start, you know, almost like playing a piano.

当我看一个想法时，我几乎是像弹钢琴一样开始的。

Uh if if you're a great great great pianist, you start to hit the keys and hit the notes without really thinking about it.

如果你是一位出色的钢琴家，你会开始按下琴键、弹出音符，却根本不用去想。

So, what you do and why you do can be intuitive and automatic.

所以你做什么、以及为什么这么做，可以是直觉性的、自动化的。

It's also why the training bur uh hurdle in this industry is tough cuz a PM has all that information in his or her head, it's hard to get it on paper for the for the new junior analyst.

这也是为什么这个行业的培训门槛很高，因为 PM 的脑子里装着所有这些信息，却很难把它落到纸面上教给新来的初级分析师。

And so, a big big hurdle in adopting these tools is making it explicit because an LLM can't read your mind, right?

所以采用这些工具的一大门槛，就是把它显性化，因为 LLM 读不了你的心思。

We're used to grunting into Google or grunting to an intern, "Hey, go build a build a model on this" without giving specific uh uh uh sort sort of structured outputs.

我们习惯了对着 Google 咕哝几句，或者对着一个实习生咕哝一句："嘿，去给这个东西建个模型"，却不给出具体的、结构化的输出要求。

Um but that's really important.

但这一点真的很重要。

It's probably the most important meta skill of working with LLMs is to be very explicit on what you do and why, and systematically deconstructing your investment process.

和 LLM 打交道最重要的元技能，可能就是把你做什么、为什么这么做说得非常清楚明确，并且系统化地拆解你的投资流程。

I'll give you some examples of that today.

今天我会给你们举几个这方面的例子。

The challenge for for the finance co-pilots building tooling in this space, too, is that investment process is highly hetero heterogeneous, right?

对于在这个领域做工具的金融 co-pilot 来说，挑战在于投资流程是高度异质的。

Whether you're a high-velocity fund focused mostly on earnings revision, sell-side narrative, or an 18-month longer-term investor, 5-year investor focusing mostly on management, right?

无论你是主要盯着盈利预期修正、卖方叙事的高换手基金，还是一个持有 18 个月、乃至 5 年、主要看管理层的长线投资人。

You know, if you look at five different investors, their their process and approach will look very different.

你看五个不同的投资人，他们的流程和方法会非常不一样。

Even in a specific fund, the way a biotech analyst at a Tiger Cub looks at names versus an Asian banks analyst will look quite different from a data and process perspective.

哪怕在同一家基金里，一个 Tiger Cub 的生物科技分析师看标的的方式，和一个亚洲银行分析师相比，从数据和流程的角度看也会相当不同。

And so, it's very hard to capture the heterogeneity of that investment process in tooling for for for investors.

所以要在给投资人做的工具里捕捉投资流程的这种异质性是非常难的。

Which is why the exoskeleton uh metaphor is so important to me in that each of those individual investors can articulate their process and build their own exoskeleton around your specific your specific process.

这也正是外骨骼(exoskeleton)这个比喻对我如此重要的原因：每一个投资人都可以把自己的流程讲清楚，围绕自己特定的流程搭建属于自己的外骨骼。

So, LLMs were a bit of a, you know, less reliable, more powerful form of Google.

所以说，LLM 曾经有点像一个不那么可靠、但更强大的 Google。

and with the sort of recent jumps in agentic capability, chatbots have grown arms.

而随着近期 agentic 能力的跃升，chatbot 长出了胳膊。

Chatbots can now run and write code.

chatbot 现在能运行和编写代码了。

They can do math.

它们能做数学。

LLMs can't do math, but Python can do math, and LLMs can reliably reliably call those agents and sub agents to do code interact with APIs and MCPs and control tools and environments.

LLM 自己做不了数学，但 Python 能做，而 LLM 可以可靠地调用这些 agent 和子 agent 去写代码、跟 API 和 MCP 交互、控制工具和环境。

So, just in the last three or four months, we've gone from passive reasoning to active agents.

所以就在过去这三四个月里，我们从被动推理走到了主动 agent。

And the active agents to me opens up a whole new slew of use cases for the investment process.

在我看来，这些主动 agent 为投资流程打开了一大批全新的用例。

Even simple things like the systematic validation, which which LLMs could not do systematic validation.

哪怕是像系统化验证这样简单的事情，而这恰恰是 LLM 过去做不到的。

I couldn't upload until three or four months until three or four weeks ago.

直到三四周之前，我还没法上传。

I could not upload an Excel file and reliably have an LLM debug that file, but you can you you you can today.

我过去没法上传一个 Excel 文件、并让 LLM 可靠地帮我调试这个文件，但今天你能做到了。

So, moving from chatbots, the you know, the new agentic work work tools, which Claude Co-worker people talk a lot about Claude Code.

所以从 chatbot，走到新的 agentic 工作工具，大家常说的比如 Claude Co-worker、Claude Code。

Claude Co-worker is just a sort of an overlay and more user-friendly way to use Claude Code.

Claude Co-worker 只是 Claude Code 之上的一层封装，是一种更好用的使用 Claude Code 的方式。

Um and Perplexity Computer, which is a very similar infrastructure to Claude Co-worker, more multimodal in nature.

还有 Perplexity Computer，它跟 Claude Co-worker 的底层架构非常相似，本质上更偏多模态。

My sort of belief is that, you know, the deep coding tools will become abstracted.

我的看法是，这些深度编码工具会被抽象掉。

They'll be very very highly user-friendly like Claude Co-worker and Perplexity Computer are.

它们会变得非常非常好用，就像 Claude Co-worker 和 Perplexity Computer 现在这样。

And we won't need to learn all the deep, you know, you won't have to spend time in a terminal.

我们不需要去学所有那些底层的东西，你不用花时间泡在终端里。

You'll just go access one of these one of these tools to get your work done.

你只要打开这些工具里的某一个，就能把活干完。

Um so, I've been sort of shocked by how powerful these tools have been to my investment process, mock investment process over the last last few months.

所以过去这几个月，这些工具对我的投资流程、模拟投资流程有多强大，让我相当震惊。

What does that mean?

这意味着什么？

It means that agentic systems can start overlay on on your systems, right?

这意味着 agentic 系统可以开始叠加到你自己的系统之上。

Before, if I were to look at a hospital company like HCA, I'd have to dig into each of these individual questions, build prompts around that, find the data, upload upload upload the data.

以前，如果我要看一家像 HCA 这样的医院公司，我得逐个钻研这些问题，围绕它们搭建 prompt，找数据，再把数据一遍遍上传。

It was very cumbersome.

这非常繁琐。

It was sort of possible in concept, but not possible in in practice.

在概念上算是可行，但在实践中并不可行。

Similar things like positioning for high velocity hedge funds, positioning is an incredibly important question, right?

类似的事情，比如高换手对冲基金的持仓结构，持仓结构是一个极其重要的问题。

To be able to pull in five or six disparate sources of data to create workflows that explain why positioning matters and how I want might want to approach differently a really crowded print long versus a really underweight print print print long will really matter.

能够拉进五六个各不相同的数据源，搭建工作流去解释为什么持仓结构重要，以及面对一个非常拥挤的多头财报、和一个仓位严重低配的多头财报，我可能想怎么区别对待，这会非常关键。

Simple signals like, "Hey, this is a highly, you know, highly shorted name into a print and in your model, you have a 2% revenue beat." Uh those are really powerful flags when you're covering 300 stocks that can give you a signal, not an answer back to back back to your investment process.

像这样简单的信号："嘿，这个标的空头很重，又要出财报了，而在你的模型里营收会超预期 2%。"当你覆盖 300 只股票时，这些是非常有力的标记，能给你一个信号——不是给你答案——反馈回你的投资流程。

These This was conceptually uh possible uh possible only in concept until recently, and I now think I'm sort of trying to build some of these in some ways or build prototypes of some of these some of these tools are now actually buildable um and uh in in real in in in reality.

这些东西直到最近还只在概念上可行，而我现在算是在尝试用某些方式把其中一些搭出来、或者做出原型，这些工具里的一部分现在确实能真正搭建出来、在现实中落地了。

It's not to say that everything is uh is is fixed by any means.

这绝不是说所有问题都已经解决了。

Data integration remains a remains a key challenge.

数据整合仍然是一个关键挑战。

Um even simple things like piping in high-quality news is is challenging.

哪怕是像接入高质量新闻这样简单的事情，也是有挑战的。

It's rather easy now.

现在其实相当容易了。

There's a number of MCPs to pipe in, you know, SEC filings and some simple market data, investor materials, and transcripts.

有一批 MCP 可以把 SEC 文件、一些简单的市场数据、投资者材料和电话会纪要接进来。

There are still real challenges around building the building the sandbox of all of the data that we consume as fundamental investors.

在搭建那个装着我们作为基本面投资人所消费的全部数据的沙盒方面，仍然存在真实的挑战。

Things like expert network transcripts, sell-side research, data and market data, fundamental data, all data, a lot of data that sits in Excel has been a challenge internal notes, etc.

像专家网络访谈纪要、卖方研究、数据和市场数据、基本面数据，所有这些数据，大量趴在 Excel 里的数据一直是个挑战，还有内部笔记等等。

So, all of these we sort of call the essential eight sort of uh uh data streams.

所以这些我们大致称之为"核心八大"数据流。

Um it's getting a little bit easier to pipe these in, but really the the the reasoning is only as good as your underlying under underlying data.

把这些接进来正变得稍微容易一点了，但说到底，推理的好坏只取决于你底层数据的好坏。

One of the things I'll call it on Claude is often still the Claude retrieval engine to Edgar breaks.

我要点名 Claude 的一件事是，Claude 到 Edgar 的检索引擎经常还是会崩。

And so, if I say build me a model, you know, with the 10Ks, it'll try Edgar.

所以如果我说用 10-K 给我搭一个模型，它会去试 Edgar。

Often it will break, and if it breaks, it goes to a log to pull the number.

它经常会崩，一崩它就转去一个日志里抓那个数字。

And surprise, surprise, that number is wrong.

然后你猜怎么着，那个数字是错的。

Is wrong is wrong quite quite often.

而且相当经常是错的。

I think you also have to carefully consider the existing quality bar.

我觉得你还得仔细掂量现有的质量基准。

A lot of people have tried to build news trackers.

很多人试过搭建新闻追踪器。

I've I've done that a little bit.

我自己也做过一点。

And then I go back and compare it to a StreetAccount tracker, which is just much better and much reliable.

然后我回头把它跟一个 StreetAccount 的追踪器一比，后者就是好得多、也可靠得多。

Simple things like getting access to Wall Street Journal, Bloomberg news, StreetAccount news.

像获取 Wall Street Journal、Bloomberg 新闻、StreetAccount 新闻的访问权限这样简单的事情。

Uh so, just cuz you can build a news scraper doesn't mean you you can.

所以，你能搭出一个新闻抓取器，并不意味着你真的能。

I've I've I've seen a lot of just inaccurate information in in news in news screeners.

我在新闻筛选器里见过大量就是不准确的信息。

So, one of the things we try to recommend people to do is build your own red light, green light, yellow light rubric and figure out what where to start.

所以我们试着建议大家做的一件事，是搭建你自己的红灯、绿灯、黄灯评估表，弄清楚从哪里入手。

What what we think is what you think is possible uh today.

先搞清楚你认为今天什么是可行的。

The usability to me has almost been one of the biggest changes in the last three months.

对我来说，可用性几乎是过去三个月里最大的变化之一。

These agentic systems, this is a screenshot from my workspace in Perplexity Computer, and I'll walk through what these things mean.

这些 agentic 系统，这是我在 Perplexity Computer 里工作区的一张截图，我会逐一讲解这些东西是什么意思。

But now I can create skills files.

但现在我可以创建 skills 文件了。

I wasted about 30, you know, I don't know, 30, 40 hours in the summer of 2020 2025 learning how to prompt by hand.

2025 年夏天我大概浪费了 30、40 个小时，靠手工去学怎么写 prompt。

I haven't written a prompt by hand because meta prompting and iterative prompting is such an easy, effective workflow.

我现在完全不用手写 prompt 了，因为 meta-prompting 和迭代式 prompting 是一套特别简单、特别有效的工作流。

Um so, I built 306 prompts based on everything we teach in Analyst Academy.

所以，我基于我们在 Analyst Academy 里教的所有内容，搭了 306 个 prompt。

Those prompts are basically now decayed in usability, but the new the new workflow or skills files.

那些 prompt 现在基本上已经不好用了，取而代之的新工作流就是 skills 文件。

And so, creating skills files that live natively in into one of these uh one of these um workspaces is incredibly incredibly um easy.

而创建原生存在于这些 workspace 里的 skills 文件，极其容易。

For example, in my modeling tool now, I can create a new modeling creator um that is built on built on how I've trained it to build a model.

举个例子，在我现在的建模工具里，我可以创建一个新的建模生成器，它是基于我训练它如何建模的方式来搭的。

I can validate that model in a different space, and then I can update that model in a different space.

我可以在另一个空间里验证那个模型，然后再在另一个空间里更新那个模型。

And so, the usability has taken a big step forward, in my opinion.

所以在我看来，可用性已经往前迈了一大步。

Even simple things like connecting a Delupa MCP with the my Fundamental Edge Model Update skill.

哪怕是很简单的事，比如把一个 Delupa MCP 和我的 Fundamental Edge Model Update skill 连起来。

I didn't think this would work, and then I got it.

我本来以为这行不通，结果我搞成了。

I tested it.

我测试了一下。

It wasn't 100% accurate, but I was able to update my Uber model, which was four quarters out of date.

它不是 100% 准确，但我确实把我那个已经落后了四个季度的 Uber 模型给更新了。

Uh not not 100% accurate.

不是 100% 准确。

There were still a few still a few issues in it, which I'll sort of talk about in the validation.

里面还有几个问题，这个我会在验证那部分讲一讲。

But validation is important cuz if you look at finance benchmarks, you're still in the 60 to 70% range.

但验证很重要，因为如果你看那些金融基准，准确率还停在 60% 到 70% 这个区间。

I'm seeing some green shoots when connected to Delupa and other things I'm hearing from other vendors that you're get now for the first time getting into the 90% range.

我看到了一些好苗头：在连上 Delupa 之后，再加上我从其他供应商那儿听到的情况，你现在头一次能进到 90% 这个区间了。

Uh but still 91% I don't know I don't know if it's fully fully there.

但即便是 91%，我也说不好它是不是真的完全到位了。

This is why agents to me have been one of the most exciting use case of agents is validation and debugging.

这就是为什么在我看来，agent 最让人兴奋的用例之一，就是验证和调试。

So, I can take this Uber model that that Perplexity Pro updated.

那么，我可以拿这个由 Perplexity Pro 更新过的 Uber 模型。

It's all very meta.

这整件事都非常 meta。

And I can say, uh "Can you check those numbers were input input it correctly?" It calls the validation the FE validations uh skill set.

然后我可以说：“你能不能检查一下那些数字是不是输入正确了？”它就会调用那套 FE 验证的 skill 集。

And it'll go through and it'll check all the numbers, right?

它会跑一遍，把所有数字都检查一遍，对吧？

It'll go through and say in this in this instance, it wasn't perfect.

它会跑一遍，然后告诉你，在这一次里它并不完美。

It was a 97% match.

匹配度是 97%。

There was a sign flipped here.

这里有一个符号弄反了。

We missed 200 million of interest income.

我们漏掉了 2 亿美元的利息收入。

That flowed through to the net income including NCI.

这一项一路传导到了含 NCI 的净利润里。

So, it's not perfect, and I didn't expect it to be perfect.

所以它并不完美，而我也没指望它完美。

But with this, I can go back into that model, and I can debug it, right?

但有了这个，我就能回到那个模型里去调试它，对吧？

So, creating these systematic validation workspaces for for for for for your for for your work.

所以，就是为你的工作创建这些系统化验证的 workspace。

The second tier of the validation system, which gets very meta, is asking these tools to create a model validation checklist for you.

验证体系的第二层——这一层非常 meta——就是让这些工具替你生成一份模型验证清单。

So, you can upload your thesis, your initial thesis, what I'm thinking, this is a gross margin story.

这样，你可以上传你的 thesis、你最初的 thesis，也就是我在想的：这是一个毛利率的故事。

This is a balance sheet story, etc.

这是一个资产负债表的故事，等等。

I really need to be accurate on the gross margin build.

我在毛利率的搭建上真的必须做到准确。

Go back in and check that.

回过头去把那块检查一遍。

You can you can build a tool like this that will spit out a checklist for you, right?

你可以搭一个这样的工具，它会替你吐出一份清单，对吧？

And so, I can go in and check off or have an intern check off, or have an Indian analyst check off gathering the documents and going through and checking the 15 or 20 mission-critical inputs, right?

然后，我可以自己进去逐项打勾，或者让一个实习生打勾，或者让一个印度分析师打勾，去收集文档、逐项过一遍，把那 15 到 20 个任务关键的输入项都核对掉。

There's sort of a myth that um models are 100% accurate.

有一种迷思，好像模型都是 100% 准确的。

I don't know if my Q3 2022 DNA uh number on my Amazon model is accurate.

我并不知道我 Amazon 模型里 2022 年三季度那个 DNA 数字准不准。

And I don't want to waste the time to go triple-check that.

而我也不想浪费时间去反复核对三遍。

That's not going to change my investment use case.

那并不会改变我的投资用例。

But if I have a gross margin-driven differentiation in my model, I need to be incredibly sure that my year-ago COGS estimates are accurate, clean, that I've that I've that I've made the proper adjustments.

但如果我模型里的差异化是由毛利率驱动的，那我就必须极其确定：我去年同期的 COGS 估算是准确的、干净的，而且我已经做了恰当的调整。

And then when management guides at 200 bips of of gross margin year-over-year, that I my model architecture is aligned with how the CFO is considering and thinking about that model architecture.

然后当管理层给出同比毛利率 200 bips 的 guidance 时，我模型的架构要跟 CFO 考虑和思考这套模型架构的方式对得上。

So, you can you can you can explain that to the system, and the system will walk you through uh checking.

所以，你可以把这些讲给这套系统听，系统就会带着你一步步做核对。

So, in that sense, it is sort of dystopian a little bit that the analyst is orchestrated by the orchestrated by the the LLM.

所以从这个意义上说，分析师被 LLM 编排调度，确实是有那么一点反乌托邦的味道。

So, if you haven't yet, I would encourage you to to uh to play with Claude Co-worker, Perplexity Computer.

所以，如果你还没试过，我会鼓励你去玩一玩 Claude Co-worker、Perplexity Computer。

If you're more coding aligned, uh technical, um you know, get your hands on Claude Code either through a VS Code or a Cursor IDE.

如果你更偏编程、更技术一点，那就通过 VS Code 或 Cursor IDE 上手 Claude Code。

Um that's a little bit harder than everyone told me it was going to be.

这件事比所有人跟我说的要难一点。

I tweeted about this, and someone said it takes about 100 hours to 100 hours to get really fluent in the IDEs.

我发过一条推，有人说要大概 100 个小时才能在这些 IDE 里真正上手。

And I said, "I don't have 100 hours, nor do the investment teams at most of my clients have 100 hours to learn just the basic tools." My hypothesis is that a lot of the complexity of the hack-together systems that are being operated in a in a terminal directly or coding IDE will be abstracted away in highly user-friendly interfaces like Claude Co-worker, Perplexity Computer, and the agentic work system that my guess, my hypothesis, I don't know for sure, will come out of OpenAI in the next few months.

我就说：“我没有 100 个小时，我大多数客户的投资团队也没有 100 个小时去学这些最基础的工具。”我的假说是，那些直接在终端或者编程 IDE 里运行的、拼凑起来的系统，其中很多复杂性都会被抽象掉，收进像 Claude Co-worker、Perplexity Computer 这样高度友好的界面里，还有接下来几个月我猜——这是我的假说，我也不完全确定——会从 OpenAI 出来的那套 agentic 工作系统。

Um so, these things will become much more intuitive, almost the way like a Repl it or a Lovable is sort of much more intuitive.

所以，这些东西会变得直观得多，几乎就像 Replit 或者 Lovable 那样直观得多。

You can one-shot things.

你可以一次性(one-shot)把事情搞定。

I'm surprised and shocked by some of the things I can one-shot in a Perplexity computer now.

我现在在 Perplexity Computer 里能一次性搞定的一些东西，让我又惊讶又震撼。

And to be very clear, I'm not sponsored by any by any any vendor any any any any vendors.

说清楚一点，我没有接受任何供应商的赞助。

So, learn these things, you know, YouTube sort of talking to someone today is like how would we figure this out if we weren't on YouTube and Twitter?

所以，去学这些东西吧，今天跟人聊天就好比说：如果没有 YouTube 和 Twitter，我们要怎么把这些搞明白？

A lot of this happens on YouTube and Twitter and then a lot of the values happens in in experimentation.

这里面很多是在 YouTube 和 Twitter 上发生的，然后很多价值是在动手实验里产生的。

So, this is a this is a you know, exoskeletons a hypothesis not a thesis cuz it's not perfect yet, right?

所以，这个外骨骼(exoskeleton)是一个假说，不是一个 thesis，因为它还不完美，对吧？

It still requires a leap of faith that, you know, some of the data challenge will improve.

它仍然需要一点信念上的跳跃——相信其中一些数据方面的挑战会得到改善。

MCPs There's argument MCPs are are a bit too brittle for certain use cases.

MCP——有一种说法是，对某些用例而言 MCP 有点太脆了。

My rebuttal back is like yes, probably, but if you have a nice validation system on top, are we now good enough?

我的回应是：是的，也许吧，但如果你在上面搭一套好的验证体系，那我们现在是不是就够用了？

There's still some technical hurdles around context rod, although a big innovation in in context was the skills files that have the sort of meta meta information that can call a skill and sort of orchestrate, you know, base essentially a series of prompts in in one one one folder.

围绕 context 还有一些技术上的障碍，不过 context 方面的一大创新就是 skills 文件——它带有那种元层面的信息，可以调用一个 skill 并做编排，本质上就是把一系列 prompt 放进一个文件夹里。

So, a lot of the challenges I had with larger prompts are now mitigated.

所以，我以前用较大的 prompt 时遇到的很多难题，现在都被缓解了。

Quantitative accuracy and compute limitations.

量化准确性，以及算力上的限制。

I do think there's a sort of reality that in the cloud ecosystem where we're accessing $2,000 of compute for $100 a month.

我确实觉得有这么一个现实：在云的生态里，我们花 100 美元一个月，就能用到价值 2,000 美元的算力。

And so, there's a question of how long that subsidization lasts.

所以就有一个问题：这种补贴能持续多久。

I'm going to try and make the most of it while it does last and hopefully the the inference curve rolls rolls down quickly.

趁它还在，我会尽量把它用足，也希望推理成本曲线能快点往下走。

So, I think the mentality of March of 9s is important.

所以我觉得，March of 9s 这种心态很重要。

From the initial DARPA self-driver to Waymo was 14 or 15 years.

从最初的 DARPA 自动驾驶到 Waymo，花了 14 或 15 年。

That first 90% is just the start just the starting gun to getting to 99.99.9.

头 90% 只是个开始，只是冲向 99.9、99.99 的发令枪响。

And I think that's a you know, a prior I've sort of considered and the quantitative use quantitative accuracies.

我觉得这是我一直在考虑的一个先验判断，也涉及量化准确性。

Um So, I think you know, the answer to me is like be very careful in terms of where you're using these tools.

所以我觉得，对我来说答案就是：在你用这些工具的地方要非常小心。

Use them for hypothesis hunch your hypothesis formation not necessarily thesis formation.

用它们来做假说的形成——你的假说形成，而不一定是 thesis 的形成。

Use cases where you're getting a signal and trading on it actively like digesting news I think are a lot more fundamentally dangerous than hey, this is my first cut on a name.

那种你拿到一个信号就主动去交易的用例，比如消化新闻，在我看来从根本上要比“嘿，这是我对一只标的的初步梳理”危险得多。

I'm going to assume this thing is 90% accurate and I'm going to go and have a thoughtful validation system system system down system downstream.

我会假设这东西有 90% 的准确率，然后在下游再去搭一套经过深思熟虑的验证体系。

Right?

对吧？

I think the other interesting debate people are having is does tech eat labor or is this sort concept of Jevons paradox where technological improvements make a resource more efficient to to use.

我觉得大家在争的另一个有意思的问题是：技术到底会吞掉劳动力，还是会出现 Jevons paradox(Jevons 悖论)那种情况——技术进步让一种资源用起来更高效，反而用得更多。

You saw this in the robo-advisor world where robo-advisors, you know, became a huge market, but actually human advisory continued to grow through that.

你在 robo-advisor(智能投顾)那个领域就见过这个现象：robo-advisor 成了一个巨大的市场，但人工顾问服务在这个过程中反而还在继续增长。

The parallel to institutional investing to me is sort of this this this concept of process mixing, right?

在我看来，对应到机构投资上，就是这种流程融合的概念。

When I started in the business in 2008, the investment process at a Tiger Cub versus a multi-manager looked very very distinct.

我 2008 年入行时，一家 Tiger Cub 和一家 multi-manager 的投资流程看上去非常非常不一样。

Fast forward 15 years and the process looked more more more more similar.

快进 15 年，这些流程看上去越来越相似了。

And so, I think that will accelerate this this sort of blending of processes that fundamental investors process will look more similar than than different.

所以我觉得，这会加速这种流程的融合，也就是说基本面投资人的流程会变得更相似，而不是更不同。

As markets become more competitive and more multi-dimensional investment process matters.

随着市场变得更具竞争性、更多维，投资流程就更重要了。

I'll give you an example.

我给你举个例子。

If I'm a high velocity idea generation fund turning my book 10 times a year, I might spend 50% of my time on management meeting earning season trying to find tone inflections, high velocity revisions, etc.

如果我是一只高换手的选股基金，一年把组合周转 10 次，我可能会花 50% 的时间在财报季的管理层会议上，去找语气上的拐点、快速的(预期)修正等等。

And that's sort of the playbook to generate alpha.

这差不多就是产生 alpha 的打法。

I don't necessarily have a week and a half to go do incredibly deep dive diligence process on the new CEO of a company that I might hold for 5 years, right?

我未必有一周半的时间，去对一家我可能持有 5 年的公司的新任 CEO 做那种极其深入的尽调流程。

Whereas a long duration fund bet the jockey investment strategy, they will do that, right?

而一只长久期基金、押注管理层(bet the jockey)的投资策略，他们就会去做这件事。

They'll go higher, you know, a consultant, etc.

他们会去雇一个顾问之类的。

to to create the New Yorker sort of profile on a CEO.

去给一位 CEO 写出那种《纽约客》式的人物特写。

Go to headquarters, spend a day with every member management, etc.

去总部、跟管理层每个成员各花上一天，等等。

It's a long duration think like an owner mindset.

这是一种长久期、像所有者一样思考的心态。

So, what's interesting interesting to me is that some of these some of these use cases if you're a long duration fund and you're not doing structured earnings previews now, the bar is pretty low in terms of how you can improve improve your process, right?

所以我觉得有意思的是，这里的一些应用场景——如果你是一只长久期基金、现在还没在做结构化的财报前瞻，那你要改进流程的门槛其实相当低。

You may miss key inflections that could inform entry and exit points.

你可能会错过一些关键拐点，而这些拐点本可以为你的进出场点提供判断依据。

But if you can take some of the process that is implemented this type of firm and create an agentric workflow around that.

但如果你能把这类公司所实施的一部分流程拿过来，围绕它搭一个 agentic 的工作流。

All of a sudden you've you've sort of deployed Jevons paradox to create a more rigorous process.

一下子，你就等于用 Jevons paradox 造出了一个更严谨的流程。

Not a speedier process, but a more rigorous process.

不是更快的流程，而是更严谨的流程。

Similarly, I would do this today.

同样地，我今天就会这么做。

If I hired a new medtech junior analyst, I would print off 15 to 20 page reports which I'll walk you through today on deep background reports on all the executives that they will be covering and inter interacting with.

如果我招了一名新的医疗科技(medtech)初级分析师，我会打印出 15 到 20 页的报告——今天我会带你们过一遍——里面是关于他将要覆盖、要打交道的所有高管的深度背景报告。

And so, step one is go read these 300 pages that are structured deep due diligence process on on on the CEO.

那么第一步，就是去读这 300 页关于这位 CEO 的结构化深度尽调。

And so, I think that's Jevons paradox in practice.

所以我觉得，这就是 Jevons paradox 的实际应用。

We won't we won't start working 15-hour weeks as investors.

我们不会因此就开始每周只工作 15 小时。

We'll just work we'll work the same or more hours, but we'll do it in a more rigorous comprehensive comprehensive way.

我们还是会工作一样多、甚至更多的时间，只不过我们会用一种更严谨、更全面的方式去做。

Why is that?

为什么会这样？

Alpha's hard Alpha's hard.

因为 alpha 很难做。

I think full displacement of fundamental investing has been a failed exercise.

我认为，想彻底取代基本面投资一直是一场失败的尝试。

There's been certain repeating patterns that that the fundamental investors have harvested over time that have become quant alphas.

确实有一些反复出现的模式，是基本面投资人多年来收割的，后来变成了 quant alpha。

Um but there are certain things that sort of an irreducible human is to fundamental investing.

但基本面投资里有某些东西，是无法被替代(irreducible)的、属于人的部分。

I'll walk through why I think that's the case.

我会讲一讲我为什么这么认为。

Um that sort of makes leads me to believe that the human elements of investing are likely to continue to be critical to to alpha generation.

这大致让我相信，投资中属于人的那些要素，很可能会继续对 alpha 的产生起到关键作用。

What are a few of those?

有哪几个呢？

I mean, investing fundamentally is a discernment business and AI wears a convincing mask of judgment, but there's no innate awareness of the world.

我的意思是，投资从根本上是一门辨别力的生意，而 AI 戴着一副以假乱真的判断力面具，但它对世界并没有内在的觉察。

They simply mimic the modal opinion in their training corpus and they fold like a like a card table when you push back.

它们只是在模仿自己训练语料里的众数观点，而你一反驳，它们就像折叠牌桌一样瞬间垮掉。

If I say is this a good business?

如果我问：这是不是一门好生意？

Is Teva good business?

Teva 是不是一门好生意？

Don't hedge.

别打太极。

Yes, it's a good business.

它会说：是的，这是一门好生意。

You know, massive global scale dermal demand and I say wrong just to clown it and it says no, it's not a good business.

全球规模巨大、皮肤科需求旺盛——然后我说“错”，纯粹是为了戏弄它，它就改口说：不，这不是一门好生意。

Operates in structure in attractive generics markets, etc.

它身处结构性有吸引力的仿制药市场，等等。

If I had an analyst that folded so quickly on pushback, that would be that would not be a use useful analyst.

如果我手下有个分析师，一被反驳就这么快垮掉，那他就不会是个有用的分析师。

So, you need you need you need real discernment in the investment process.

所以，投资流程里你需要的是真正的辨别力。

This is a game of inches.

这是一场以毫厘论胜负的游戏。

A game of small small really small edge.

一场比拼极其微小 edge 的游戏。

So, I think be careful asking LLMs for a view on things particularly without the proper context to think about think about scenarios.

所以我觉得，向 LLM 询问对某件事的看法要谨慎，尤其是在没有恰当 context 去推演各种情景的情况下。

I think there's more philosophical question of if everyone has super intelligence, no one has it.

我觉得还有一个更偏哲学的问题：如果人人都拥有超级智能，那就等于没人拥有它。

We've seen a number of priors like that in the space.

在这个领域，我们见过好几个类似的先例。

Alternative data is a prior that I lived through as a former consumer analyst.

另类数据就是我作为前消费板块分析师亲历过的一个先例。

The first four to seven years of alternative data sets being in the hands of investors when they were as more exclusive distribution was almost a license to print alpha.

另类数据集刚落到投资人手里、分发还相对独家的头四到七年，几乎就是一张印钞 alpha 的许可证。

Fast forward, you know, five or six or seven years, it became more of a knife fight where the second and third order derivative became sort of the debate.

快进到五六七年之后，它变成了一场肉搏战，争论的焦点变成了二阶、三阶导数。

Um so, markets are a complex adaptive systems where super intelligence gets adopted if if easily accessible gets adopted.

所以，市场是一个复杂自适应系统，超级智能一旦容易获取，就会被广泛采用。

And I sort of think about this is the compression of informational edge.

我倾向于把这看作信息优势的压缩。

When I started in the business, I go call 20 Wendys franchisees to try and get an edge on the same store sales print and that's compressing or the sort of fully fully compressed.

我刚入行时，会打电话给 20 家 Wendy's 加盟商，试图在同店销售的财报数据上抢到一点 edge，而这种做法正在被压缩，或者说已经被彻底压缩了。

The idea that I can know something others don't know just doesn't really exist much in the investment process any anymore.

我能知道别人不知道的东西——这种想法在如今的投资流程里已经几乎不存在了。

And so, when people say LLMs will sort of eliminate fundamental alpha, it's sort of this belief that there's some piece of information out on the open web that singularly is a source of alpha.

所以，当人们说 LLM 会消灭基本面 alpha 时，那背后其实是一种信念：认为公开网络上存在某条信息，单凭它本身就是 alpha 的来源。

And I sort of say like that's not really how it works in my my experience.

而以我的经验来说，事情其实并不是这么运作的。

The the alpha is behavioral, right?

alpha 是行为性的，对吧？

Due to market market microstructure.

这源于市场微观结构。

It's behaving the way others don't or can't behave.

它在于以别人不会、或不能采取的方式去行动。

Hey, there's two this is a should be a $30 stock going to 60, but it's at 20 cuz the next two prints are squishy.

比如，这本该是一只从 30 美元涨到 60 美元的股票，但现在只有 20 美元，因为接下来两次财报会比较疲软。

So, I want to look through those two two prints to have a better long-term view.

所以，我愿意越过这两次财报，去持有一个更好的长期看法。

That's sort of real that's the manifestation of edge in the markets today.

这才是真实的，这才是 edge 在当今市场中的具体体现。

And I think the other sort of way I think about edge is integrated perception.

我思考 edge 的另一种方式是整合感知。

It's the alpha's in your worldview.

alpha 就存在于你的世界观里。

You're connecting the dots to sort of make views on winners and losers in the industry trajectory of of industries and companies and management teams, etc.

你把各种线索串联起来，从而对行业、公司、管理层的演进轨迹中谁是赢家、谁是输家形成判断。

Both of these are both of these as as alpha pools are alive and well.

作为 alpha 池，这两者都活得好好的。

And in fact, I think the competitive set is getting more favorable favorable for the those sort of alpha pools.

事实上，我认为对这类 alpha 池来说，竞争格局正变得越来越有利。

A lot of this comes down to the simple fundamentals that stocks are priced on perception.

这在很大程度上归结为一个简单的基本面：股票的定价基于感知。

I had huge fights and debates not fights, but good spirited debates with quant teams over the years that quant systems could never get to a perfect price perfect quantitative price on stocks because price is just a function of perception.

这些年我和 quant 团队有过激烈的争论——不是吵架，而是热烈而友好的辩论——我认为 quant 系统永远无法给股票算出一个完美的定价、完美的量化价格，因为价格只是感知的函数。

If you look at the typical derivation of a 20 PE stock, only 14% of today's value comes from discounting the next 5 years of cash flows.

如果你看一只 20 倍 PE 股票的典型估值拆解，今天价值里只有 14% 来自对未来 5 年现金流的贴现。

86% lives in the lives in the future.

86% 存在于未来。

And who knows what the future looks like.

而谁也不知道未来是什么样子。

This is why we see large cap stocks often trade like penny stocks.

这就是为什么我们常看到大盘股交易得像仙股一样。

Fundamental investing is a game of hidden hidden information three-dimensional complexity in the sense that we're building a mosaic to understand the trajectory of fundamentals.

基本面投资是一场关于隐藏信息、三维复杂性的游戏，意思是我们在拼凑一幅马赛克，去理解基本面的走向。

But a lot of the game is this complex adaptive system.

但这场游戏的很大一部分是这个复杂自适应系统。

What does the market think about the future?

市场对未来是怎么想的？

What do we think about the future?

我们自己又对未来是怎么想的？

And what will the fundamental evolution of the data points do to influence that perception, right?

而基本面数据点的演变，又将如何影响那种感知？

So, this is a deeply psychological but also a deeply psychological and game to generate alpha.

所以，这是一场既深具心理色彩、又要靠它来创造 alpha 的游戏。

I think the competitive set's clearly getting easier sort of more maddening in the in the near in the near in the near term.

我认为竞争格局在短期内显然正变得更容易，也更让人抓狂。

But if fundamental investors are competing against indexers, you know, quants and pods who are constrained by duration or not investing on a fundamental base, that sort of definition means a weaker fundamental price signal.

但如果基本面投资人面对的竞争对手是指数投资者、quant，以及那些受久期约束、或并非基于基本面来投资的 pod，那么按这个定义，就意味着基本面价格信号更弱。

It means that stocks will overshoot more aggressively on the downside to upside.

这意味着股票在下行和上行时都会更激烈地过度反应。

But again, if I have a $20 stock, $30 stock going to 60 and it overshoots down to 15, if I have duration that's a mispriced asset that should ultimately be an alpha opportunity.

但同样地，如果我有一只本该从 30 美元涨到 60 美元的股票，它却过度下跌到 15 美元，只要我有足够的久期，那就是一个定价错误的资产，最终应该会成为一个 alpha 机会。

Now, that can take some time to close and easier easier said than done, Mr.

当然，这个错误可能需要一些时间才能收敛，而且说起来容易做起来难，这位……

Retired Hedge Fund PM.

……退休对冲基金 PM 先生。

Take that take that.

行，这话我认。

I take that pushback for sure.

这个反驳我确实接受。

Um but I think there's also priors in the in sort of the investment complex as well, too.

但我觉得，在整个投资体系里也同样存在一些先例。

You've seen, you know, 200 years of innovation in terms of liquidity, uh analytical abilities, whether it's the stock exchange Excel to quant models to structured accounting to Bloomberg to credit card panels.

你已经见证了 200 年来在流动性和分析能力上的创新，无论是股票交易所、Excel、quant 模型、结构化会计、Bloomberg，还是信用卡数据面板。

You know, uh a simple innovation like Excel took, you know, Ben Graham from doing his calculations on net nets with a slide rule uh to all of a sudden, if you see a net net in the market now, you don't want to buy that, right?

要知道，像 Excel 这样一个简单的创新，就让 Ben Graham 从用计算尺去算 net-net，一下子变成了——如果你现在在市场上看到一只 net-net，你反而不想买它，对吧？

Cuz it's sort of all it's sort of a sign that cash cash is burnt cash is being burnt, some broken biotech company that's likely to do something stupid.

因为这几乎是一种信号，说明现金正在被烧掉，是某家快烧完钱、很可能会干出蠢事的破败生物科技公司。

Uh so, the impact of what previously took days or weeks could now be completed reliably in in in in minutes.

所以，以前需要几天或几周才能完成的工作，现在可以在几分钟内可靠地完成。

Did that kill fundamental investing?

这杀死基本面投资了吗？

No, there was a huge boom sort of on the back of that.

没有，反而在此基础上出现了一波巨大的繁荣。

Now, in the in the near term, I love the stories from my Tiger Cub friends in the '80s and '90s when having a fax machine or using Excel well or um uh sort of the pre-Reg FD rules, etc., um you did have a big analytical edge, it's compressed, right?

当然，就短期而言，我很喜欢听我那些 Tiger Cub 朋友讲 80、90 年代的故事：那时候拥有一台传真机、或者 Excel 用得好、又或者赶在 Reg FD 规则出台之前，你确实有很大的分析 edge，但如今这种 edge 被压缩了，对吧？

So, with all this compression informational edge, surely alpha's compressed over time.

那么，在信息优势被如此压缩之后，alpha 想必也随着时间被压缩了。

Well, I think the exact inverse has happened.

然而，我认为发生的恰恰相反。

I sympathize with the argument from Cliff Asness where he talks about the less efficient market hypothesis and he uses value spreads as a proxy for that.

我认同 Cliff Asness 的观点，他谈到市场更低效假说，并用价值利差作为它的代理指标。

The eventually sort of essentially the view that if with increasing volatility in the value spreads, the market doesn't know, right?

这个观点本质上是说：随着价值利差的波动性上升，市场其实并不知道答案，对吧？

The market doesn't knows nothing about the future.

市场对未来一无所知。

Sort of a fundamental sign of of inefficiency in the markets.

这算是市场存在低效的一个基本迹象。

I think the other thing, too, is if you look at the alpha machine funds, the funds that are structured to have very low beta, sort of purely harvesting dollar alpha in the markets, those funds have gotten much bigger over last 10 years, but they've sustained 10 to 15% returns.

我认为还有一点，如果你看那些 alpha machine 基金，也就是那些结构上刻意保持极低 beta、纯粹在市场里收割美元 alpha 的基金，过去 10 年里它们规模大了很多，却依然维持着 10% 到 15% 的回报。

So, the dollar alpha from that complex is is bigger than ever.

所以，来自这一体系的美元 alpha 比以往任何时候都大。

Um so, I think that's sort of a quantitative confirmation that uh the markets aren't aren't more efficient.

因此，我认为这算是一个量化上的印证：市场并没有变得更有效。

They are less efficient, right?

它们变得更低效了，对吧？

Which is actually phenomenal news for fundamental investors who can now deploy these tools into a market that is structurally less efficient than it than it was five or 10 five or five or 10 years ago.

对基本面投资人来说，这其实是极好的消息——他们现在可以把这些工具部署到一个结构性上比五年、十年前更低效的市场中。

So, I'd say don't buy the don't buy the doomerism that fundamental investing is done.

所以我想说，别信那套说基本面投资已经完蛋了的唱衰论。

I'm sort of more I'm a little bit biased, obviously, but I, you know, the more I sort of think about the the setup, I'm more optimistic on the future of bottoms-up fundamental investing than uh than I have been in a long time.

显然我有点偏向性，但说实话，我越是琢磨眼下这个局面，就越对自下而上的基本面投资的未来感到乐观，比我很长时间以来的任何时候都更乐观。

All right, so that's the exoskeleton hypothesis.

好，这就是外骨骼(exoskeleton)假说。

Let's get a little out of philosophy and let's get a little bit more into uh practice of how we can uh we can use these tools.

咱们先从哲学里抽出来一点，多讲讲实操——怎么用这些工具。

So, four governing four governing principles before we start start that.

所以在开始之前，先讲四条统领性的原则。

Above all, we want to optimize for rigor for rigor over over speed.

首要一点，我们要为严谨度而优化，严谨优先于速度。

The sort of idea that I can build an investment thesis in 60 minutes, I think is a a really bad idea.

那种我能在60分钟里搭出一个投资论点的想法，我觉得是个非常糟糕的想法。

Markets will punish mediocre research.

市场会惩罚平庸的研究。

Look at, you know, as evidence, you know, 70 to 80% of long-only investors underperform the markets over time.

举个证据，随着时间推移，70%到80%的long-only投资人跑不赢市场。

Uh you know, you know, T+1 sell-side recommendations have no alpha.

而且卖方T+1的推荐没有任何alpha。

So, mediocre a mediocre thesis gets a sub-mediocre result due to the power law distribution of alpha of alpha in markets.

所以，由于市场里alpha呈幂律分布，一个平庸的论点得到的是低于平庸的结果。

So, at all steps, we're demanding rigor.

所以在每一步，我们都要求严谨。

We're trying to go we're trying to go deeper.

我们要做的是往更深处挖。

Now, in some of those use cases, it can be taking the the rote boring stuff off of my desk so I can go do more primary primary research.

在其中一些用例里，它可以把那些机械、无聊的活儿从我桌上拿走，让我腾出时间去做更多一手研究(primary research)。

Number two is try to build a broader net for signal identification.

第二条，是试着为信号识别撒下一张更大的网。

The idea to me is sort of democratizing the tools that I used to have at my grasp internally at some of these funds.

对我来说，这个想法有点像是把我过去在这些基金内部才能触及的工具民主化。

You having an old data team um sort of in your pocket, an investigative journalist, a quant and LP, a forensic accountant in your pocket, the sort of building more rigorous evaluation sets around various parts of the job, I think are possible.

你口袋里就揣着一个另类数据(alt data)团队、一个调查记者、一个quant、一个LP、一个法证会计，围绕工作的各个环节去搭建更严谨的评估集，我觉得这些都是可能做到的。

You know, we're compressing the mechanical parts of the job for sure with the with the important stimula- stipulation that as long as it doesn't impact comprehension.

我们确实在压缩这份工作里机械的部分，但有一个重要的前提约束——只要它不影响理解力。

If you feel like not reading the 10-K is it compressing your comprehension, read the 10-K.

如果你觉得不读10-K会压缩你的理解力，那就去读10-K。

There's no there's no reason that you can't.

没有任何理由说你不能读。

You're building your own system that gets you at the end of the process to thinking more deeply with more insights than you would would would would would would otherwise.

你是在搭建自己的一套系统，让你在流程末尾能比原本想得更深、拥有更多洞见。

So, if you're spending 3,000 hours in a year, how can you free up 600 hours to redeploy on high value tasks?

所以，如果你一年花3,000个小时，你怎么能腾出600个小时重新投到高价值的任务上？

And then it's a two-tier value validation system of system- systematic and artisanal human human verification.

然后，这是一套两层的价值验证系统——系统化验证，加上匠人式的人工核验。

So, I feel very strongly about this this this this slide.

所以，这张幻灯片我个人非常看重。

All right, so how what are a couple ideas to do that?

好，那么有哪几个思路可以做到这一点？

First, I'll give you an analog, the detailed consensus analog.

首先，我给你一个类比，详细consensus的类比。

When I started in the hedge fund world, I'd have to go collect 12 sell-side models from all the brokers.

当我刚进对冲基金这行时，我得去从所有券商那里收集12份卖方模型。

I'd have to then go in each model and pull the Goldman, you know, the Citi, the Morgan Stanley gross gross profit estimate, put it into a table, send it off to my senior analyst and do that over and over and over ahead of earnings.

然后我得钻进每个模型，把Goldman、Citi、Morgan Stanley的毛利估计值拉出来，放进一张表，发给我的资深分析师，并且在财报前一遍又一遍地这么做。

It was mind-numbing work and I didn't really learn anything about the businesses doing that.

这活儿让人麻木，而且做这些我其实没学到任何关于这些生意的东西。

Fast forward to Visible Alpha, Visible Alpha created an Excel add-in to do that where they sort of ingest the models and I can see detailed detailed consensus.

快进到Visible Alpha,Visible Alpha做了一个Excel插件来干这件事，他们会把这些模型吸进来，我就能看到详细的consensus。

So, this is sort of the mindset I'm using in thinking about the types of things to to to to to augment.

所以，这大致就是我在思考该增强哪些类型的东西时所用的心态。

Another another analogy would be the operator's knowledge, right?

另一个类比是运营者的知识，对吧？

So, the mandate to investors is yeah, the firms I worked at is to know the business like like an owner or know the business more than any other non-insider, right?

给投资人的要求——在我工作过的那些机构里——是要像所有者一样了解这门生意，或者说比任何其他非内部人士都更了解这门生意，对吧？

A lot of that is sort of the basic just talking to consultants, former executives, etc.

其中很大一部分，基本上就是去跟顾问、前高管等等聊。

And when I started that, meant 30, you know, 30 expert network calls.

我刚开始那会儿，这意味着30通专家网络电话。

Well, now I can go into Tegus, now part of AlphaSense or Third Bridge, and I can leverage transcripts to do that, right?

而现在，我可以进到Tegus——它现在归属于AlphaSense——或者Third Bridge，借助访谈记录来做这件事，对吧？

And so, does it really matter if I did those basic calls by myself?

那么，我是不是亲自打了那些基础电话，真的那么重要吗？

I think in certain cases, yes, but to compress 30 hours of calls into 3 hours of reading um is is a big up big uplift.

我觉得在某些情况下，是的，但把30小时的电话压缩成3小时的阅读，是一个很大的提升。

So, that's that's the mindset I have when I'm thinking about these things as well as signals.

所以，这就是我在思考这些事情、以及思考信号时所抱有的心态。

So, I think there's a few things I've I've toyed with.

所以，有几件事我一直在琢磨尝试。

Um you know, data dashboards, web scraping, guiding simple sort of a different analysis, and I'll I'll walk you through one that's a man- a management deep dive.

比如数据仪表盘、网页抓取、引导做一些不同的分析，我会带你过一遍其中一个——一次管理层深度剖析。

Now, natively for the first time, these tools can build dashboards.

现在，这些工具头一回能够原生地搭建仪表盘。

I built this in Perplexity.

我是在Perplexity里搭的这个。

This is for Winnebago.

这是给Winnebago做的。

I can go and say these are the key data points.

我可以直接说，这些是关键数据点。

Go go scrape and screen these.

去抓取并筛选这些。

You can also run these as scheduled tasks and cron jobs where I can go and do a web scrape of, you know, promotional intensity in a company like DraftKings and just see how that changes over time.

你也可以把这些作为定时任务和cron job来运行，我可以去抓取比如DraftKings这样一家公司的促销力度，看它随时间如何变化。

This was possible before.

这在以前也能做到。

Investors did all sorts of web scraping before, but now it's a little bit easy a little bit easier to do.

投资人以前也做各种各样的网页抓取，但现在做起来稍微容易一点了。

It's a little bit easier to ingest back into your your investment process.

把它重新导回你的投资流程里，也稍微容易一点了。

There's a number of different workflows like a guidance achiev- achievability analysis that can now be one-shot as part of evaluating management, which I'm sort of just delight- delighted by.

有若干不同的工作流，比如guidance可达成性分析，现在作为评估管理层的一部分可以one-shot完成，这一点让我相当欣喜。

A simple thing that I did, I went into met you know, top 10 medtech companies and I said, "Go back 10 years and do an analysis on which of the companies were most reliable in meeting guidance." And it built me this heat map to say, "Okay, Intuitive Surgical consistently raised, uh Zimmer at the bottom and consistently were in line or lowering." And so, you you can sort of create a rubric of how credible this CFO is when he or she gives guidance, right?

我做过一件很简单的事：我调出前 10 大 medtech 公司，然后说：“往前追溯 10 年，分析一下这些公司里哪些在兑现 guidance 上最可靠。”它给我生成了一张热力图，告诉我：“好，Intuitive Surgical 一直在上调，而 Zimmer 垫底，一直是符合预期或下调。”所以你差不多可以据此建立一套评分标准，判断这位 CFO 给出 guidance 时有多可信，对吧？

The sort of use case here is obvious.

这里的用例是很明显的。

If it's a new company I haven't I haven't uh heard about it, my analyst is pitching me a stock that this company is a beaten race story, well, I can go in and be like, "All right, if you pitch me Zimmer as a beaten race story, I'm going to be quite a bit more skeptical of that of that situation just given the base rate of guidance achievability in that company." I could do this before, of course.

如果是一家我没听说过的新公司，我的分析师向我推荐一只股票，说这家公司是个“困境反转”的故事，那我就可以进去说：“好，如果你把 Zimmer 当成困境反转的故事推给我，那考虑到这家公司兑现 guidance 的基准概率，我对这种情形会怀疑得多。”当然，这件事我以前也能做。

I've done this before many times.

我以前做过很多次。

It's highly annoying to go back into the press releases, start a year, end a year.

要回头翻那些新闻稿，从年初翻到年末，是极其烦人的。

Um you can do this a little bit in FactSet, but it's not great.

在 FactSet 里你能做一点这类分析，但做得并不好。

You don't get um um it's it's it's sort of hard to systematize.

你得不到……这件事挺难系统化的。

But I now I can do this and 19 other analyses like it very systematically in an agentic in an agentic workflow.

但现在我可以在一个 agentic 的工作流里，非常系统化地做这件事，以及另外 19 项类似的分析。

A key workflow that I've been working on is management deep dives.

我一直在打磨的一个关键工作流是管理层深度研究。

It's sort of a you know, a big hallmark of the Tiger community to go deep to go deep deep deep deep deep on a company.

对一家公司挖得极深，可以说是 Tiger 系的一大标志。

Um and I was able to sort of take all of my thoughts on what makes a good management team and put it into a deep context document.

我能够把我对“什么才算一支好的管理团队”的所有想法，都放进一份很深的 context 文档里。

What is a context document?

什么是 context 文档？

It's almost like a textbook or PDF for everything you've thought about in a company.

它几乎就像是一本教科书，或者说一份 PDF，承载你对一家公司思考过的一切。

These are really easy to create.

这些其实很容易创建。

We're guiding people through all sorts of different types types of ways to create this.

我们正在引导大家用各种不同的方式来创建它。

This can be these can be done through recording internal meetings, could be done through talking into your phone, it could be done through just writing for 10 pages on what to you makes a good management team.

这可以通过录制内部会议来完成，可以对着手机口述来完成，也可以只是花 10 页篇幅写下你认为什么才算一支好的管理团队。

That's the candidly the hardest part of building these usable tools now is getting the tacit knowledge that's in your mind out on paper, right?

坦白说，如今搭建这些可用工具最难的部分，就是把你脑子里的隐性知识落到纸面上，对吧？

From there, all I need to do is drop in this context document, which is how I evaluate management, into an agentic system.

接下来我要做的，就是把这份体现我如何评估管理层的 context 文档，丢进一个 agentic 系统里。

This is Perplexity Pro.

这是 Perplexity Pro。

And say, "Turn this into an architecture, right?

然后说：“把它变成一套架构。

Turn this into a skills architecture." You know, even in even in the system, you say, "Okay, upload this into the skills." Like the actual technical elements of turning this into a usable skills file is super easy.

把它变成一套 skills 架构。”哪怕在系统里，你也只需说：“好，把这个上传进 skills。”把它变成一个可用的 skills 文件，真正的技术环节其实超级简单。

I can turn it to a Claude, it gives me a zip file, upload it back into Claude.

我可以把它交给 Claude，它给我一个 zip 文件，再上传回 Claude。

The technical process of it takes minutes, right?

它的技术流程只需要几分钟，对吧？

It really takes minutes.

真的只需要几分钟。

The hard part, the heavy lifting, is wrapping the exoskeleton, right?

难的部分、真正吃力的活儿，是把这副外骨骼包裹上去，对吧？

Around your your investment process, right?

包裹在你的投资流程外面，对吧？

And so, rather than each of these analyses being a prompt, I can now create a system, right?

所以，这些分析不再是一个个孤立的 prompt，我现在可以创建一套系统，对吧？

Where I'm running all of these different analyses.

在这套系统里，我同时跑所有这些不同的分析。

Like before, given context window limitations, I would have had to create each of these as an each of these as an individual individual prompt, right?

而在以前，受限于 context window，我不得不把这里的每一项都做成一个单独的 prompt，对吧？

And so, um I'll give you an example.

那么，我给你举个例子。

Like if I want to go into just form four analyses, I want to look at, you know, exercise option exercise behavior as as a signal.

比如我想专门做 form 4 的分析，我想把行权、期权行使行为当作一个信号来看。

Or if I want to find situations where, you know, CEO and CFO language is, you know, diverging.

或者我想找出 CEO 和 CFO 措辞出现分歧的情形。

If I want to go do a forensic accounting example.

又或者我想做一个法务会计（财务舞弊排查）的分析。

These are things that I could have done with quite a bit of work in a chatbot.

这些事我原本在 chatbot 里花上相当多的功夫也能做到。

Um but now I can do these systematically in a in an in a in an agent.

但现在我可以在一个 agent 里系统化地完成它们。

I can go build a buyback effectiveness analysis, right?

我可以去搭建一套 buyback 有效性分析，对吧？

Um you know, uh I can go if the CFO just had a big insider buy, I can go back in minutes and say, "Has this CFO had this size of insider buy in the past?" And how did the stock react on the on the back back back back back end of that, right?

比如某位 CFO 刚做了一笔大额内部增持，我可以在几分钟内回溯并发问：“这位 CFO 过去有没有做过这种规模的内部增持？”以及在那之后股价是怎么反应的，对吧？

Um so, there's all sorts of things you can do now.

所以，现在你能做的事情五花八门。

You can scrape websites, you can go scrape LinkedIn for job posts, etc.

你可以爬取网站，可以去爬 LinkedIn 上的招聘岗位等等。

that create this structured ability to evaluate a management team, in my opinion, in a much more rigorous way than you ever could you ever could before.

在我看来，这些让你能够以一种比以往任何时候都严谨得多的方式，结构化地去评估一支管理团队。

So, comparing an output like this to a in a chatbot, um these are this is sort of if I one-shot it, it's not that good it's not that good.

所以，拿这样一份产出去和 chatbot 里的结果比，如果我 one-shot 直接出，它其实没那么好。

I covered Boston Scientific for years, so I have some context of sort of what should be in this what should be in this report.

我覆盖 Boston Scientific 有好几年了，所以对这份报告里应该有些什么，我心里是有底的。

And I'm sort of shocked by the by the quality and caliber of it.

而它的质量和水准让我有点震惊。

It's still not perfect.

它仍然算不上完美。

There are certain data things I need to pipe in and figure out MCP and some of these analyses that were uh sort of uh outlined just didn't have the data to to to run.

有些数据我得接入并搞定 MCP，而刚才勾勒出的一些分析，只是因为没有数据而跑不起来。

Uh but again, March of 9th, this will continue to this will continue to uh to improve.

但话说回来，随着 March of 9s，这一点会持续改善。

When you get into actually going to a conference, meeting with this executive, the ability to have a primer of one page of what the management has been saying, key language selection, taking in the current documents and in in sort of doing a first cut on questions, etc.

等你真的要去参加一场会议、和这位高管碰面时，能有一页纸的速览，把管理层一直在说的话、关键措辞的选择梳理出来，把当前的文件都纳进来，并且先对问题做一版初稿等等，这种能力很有价值。

There's all sorts of use cases like that for for conference preparation, um etc.

像这样的用例，在会议准备等场景里比比皆是。

And I think NLP is a really powerful over agentic overlay to to put in your systems.

我认为 NLP 是一层非常强大的、可以叠加进你系统里的 agentic 覆盖层。

The cool part of these is they scale as well, too.

这些东西酷的地方在于，它们同样可以规模化。

I'm not going to put the put the names out cuz these are all medtech companies.

我就不把名字放出来了，因为这些都是 medtech 公司。

I don't want, you know, some company sending me hate mail for saying their CEO is the worst CEO uh CEO in the industry, but if you cover medtech, you probably know.

我可不想因为说了某家公司的 CEO 是行业里最差的 CEO，就收到人家寄来的仇恨邮件，不过如果你覆盖 medtech，你大概心里有数。

You probably you could probably eyeball this one.

你大概一眼就能认出这一个是谁。

The ability to then take these workflows and not just do it on one, but then do it on all on all 10, right?

接着，你可以把这些工作流不只用在一家公司上，而是用在全部 10 家上，对吧？

And so, in the point of enhancing rigor, if I have a new analyst who's covering medtech for the first time, I'm going to do print 10 of these off, I'm going to go read them myself, validate them, make you know, fix them, etc.

所以，回到增强严谨度这一点，如果我有一位第一次覆盖 medtech 的新分析师，我会把这样的报告打印 10 份，自己先去读、去验证、去修正等等。

Um and then print those for that analyst and be like, "This is, you know, get to know these management teams, right?

然后把那些打印出来给那位分析师，跟他说：“这就是……去好好了解这些管理团队，对吧？

Here's 200 pages of reading to do um to really get the the 10-year history of of this of this executive, what you can trust and what you can't trust.

这里有 200 页要读的材料，真正把这位高管 10 年的历史吃透——哪些你可以信，哪些你不能信。

I think that's really important context for going out and conducting corporate conducting corporate access.

我认为，对于出去做 corporate access（企业调研走访）来说，这是非常重要的背景储备。

All right, so those those are a few examples and um we're building we're building the the the example set.

好，这些就是几个例子，而我们正在不断扩充这个案例集。

Um Uh a lot of this just comes to me at different different times.

这里很多想法，都是在不同的时刻冒出来找上我的。

Um I've posted a lot of this on Twitter.

这里的很多东西我都发在了 Twitter 上。

So, if you know, if you want to sort of see the new ideas there, uh but we're building a lot of things for clients, too, based on client um requests on things that can be can be done as well, too.

所以，如果你想看那里的新想法，可以去看看，不过我们也在根据客户的需求，为客户搭建很多同样能做的东西。

All right, so you're with me if you're with me that the exoskeleton hypothesis seems valid, uh that these tools are important, uh how do I start, right?

好，那么如果你认同我说的——认同外骨骼假说似乎是成立的，认同这些工具很重要——那接下来我该怎么上手呢？

So, I'll give you a few just on-ramp.

那我给你几个入门的切入点。

I have this dynamic now, which is really really interesting.

我现在观察到一个特别有意思的现象。

We go into clients and of 50 50 people, there'll be like three or four of that 50 who are so far ahead and like are way more advanced than we are in these tools.

我们去客户那边，50个人里，大概会有三四个人已经遥遥领先，在这些工具的运用上比我们还要先进得多。

And then there's like 47 who are like, "Yeah, I haven't had the time to figure out these tools yet.

然后还有大概47个人是这种状态：“对，我还没腾出时间搞明白这些工具。

Maybe I use it for pumpkin pie recipes, but I haven't adopted into my investment process." And so, what we're trying to do is sort of speak to that, you know, speak to those other 47 and give them an easy easy on-ramp to using these tools.

可能我拿它来查南瓜派食谱，但还没把它用进我的投资流程。”所以我们想做的，就是去跟那部分人对话，跟那另外47个人对话，给他们一个非常轻松的上手路径去用这些工具。

So, there's a couple of things.

所以这里有几点。

Yeah, I think one of the biggest ones we talked about is just building that muscle to articulate what you do and why, right?

我觉得我们讲过的最重要的一点，就是把“说清楚你在做什么、为什么这么做”这块肌肉练出来。

One of the most powerful examples we walked people through is figure out one thing.

我们带大家走过的最有力的例子之一，就是先搞定一件事。

Figure out one thing like, "Hey, I want to evaluate this management team." Just do a zero-shot directly into an LLM and then spend some time building the system around that.

搞定一件事，比如“嘿，我想评估一下这个管理层”，直接对着LLM做一个zero-shot，然后花点时间围绕它把系统搭起来。

Record, you know, speak into a voice recorder, build a D Burger's context document, and just see what comes out, right?

做记录，对着录音笔说话，搭一个D Burger式的context文档，然后看看会跑出什么来。

If it hits if it hits your standard, continue to improve it.

如果它达到了你的标准，就继续打磨它。

A lot of prompting is now meta-prompting or iterative prompting.

现在很多prompting其实是meta-prompting，或者说迭代式prompting。

It's the same dynamic with skills, meta-skills creation, iterative skills creation.

skill也是同样的路数——meta-skill创建、迭代式skill创建。

If I get something back on the Boston Scientific CEO and it's dinging that CEO for shareholder lawsuits on various disclosures, or something like that, sort of just an ambulance chasing thing, I can upload the skill back into an agent and say, "Hey, this is not something we should ding the tool on." So, you're refining those, you know, a skill takes three or four refinements often to get to get to get it in line with you.

如果我拿回一份关于Boston Scientific CEO的东西，里面因为各种信息披露上的股东诉讼在给这位CEO扣分，那种追着救护车揽官司式的东西，我可以把skill重新上传回agent，跟它说“嘿，这不是我们该给工具扣分的地方”，所以你是在不断精修它——一个skill往往要精修三四次才能调到跟你合拍。

You know, Claude Projects is a really uh sort of easy tool to get.

Claude Projects其实是个很容易上手的工具。

Claude Co-works great, requires a little bit more setup, but Claude Projects um is pretty phenomenal, too.

Claude Co-worker很好用，需要多花一点配置功夫，不过Claude Projects也相当出色。

And now, you know, people talk about agents, but Claude Chat and Claude Projects have agentic capabilities.

现在大家都在谈agent，但其实Claude Chat和Claude Projects就具备agentic能力。

I can go build an Excel model in Claude Claude Chat.

我可以直接在Claude Chat里搭一个Excel模型。

Uh but really just the feeling of I can do it, I can learn is a critical starting point in trying to activate that aha moment is really intoxicating.

但真正关键的起点，是那种“我能做到、我能学会”的感受，它在激活那个aha时刻的过程中特别让人上瘾。

Um I'm probably working more nights and early mornings as you sort of see the outputs like, "Man, I want to try more things." So, rather than AI freeing up more time in my week, it's actually led me to be much busier, which is there's got to be some sort of paradox uh for that.

我现在大概晚上和清早工作得更多了，因为你看到那些产出会想“天哪，我想试更多东西”，所以AI非但没给我的一周腾出更多时间，反而让我忙了很多，这里头肯定存在某种悖论。

I point people to to YouTube and Andrej Karpathy videos.

我会把大家引向YouTube和Andrej Karpathy的视频。

If you're brand new to this stuff, you know, go through his old stuff.

如果你是彻底的新手，去把他早期的东西过一遍。

You know, it's it's kind of slow time sometimes to listen about tokenization, the jagged edges, and transformers, but it is helpful to build that intuition in terms of work working working work work working with these tools.

听tokenization、jagged edges(锯齿边缘)、transformer这些，有时候确实是段慢时光，但它对建立你使用这些工具时的直觉是有帮助的。

There's some great, you know, direct direct um uh courses, you know, in OpenAI, Anthropic, etc.

OpenAI、Anthropic这些地方有一些很棒的直接的课程。

And some great resources on YouTube and Twitter, though you have to, you know, sift through a lot of low low quality.

YouTube和Twitter上也有一些很好的资源，不过你得从大量低质量内容里筛。

And then just a reminder, I sort of tell the story I I thought I was being a good dad encouraging my 12-year-old son to use AI uh for academics, and I come to learn that he was just taking a picture of his question sets and writing down the answer, which is a sort of a complete bypassing of the learning the the learning process.

再提醒一句，我常讲这个故事——我本以为自己是个好爸爸，鼓励我12岁的儿子把AI用在学业上，结果我发现他就是把题目拍张照、把答案抄下来，这等于是完全绕过了学习这个过程。

And so, you know, I sort of gave them a lecture of like, "You're not trying to get the answer.

所以我给他上了一课，大意是“你的目标不是拿到答案。

You have to actually show show your the work and create study guides and use this tool to give you a more rigorous understanding of the work since when you show up to at test time, right?

你得真正把解题过程写出来、做出学习提纲，用这个工具让你对内容有更严谨的理解，因为等你到了考试的时候——

You won't have this tool available." I think that that analogy holds in in using these tools is really being thoughtful around comprehensive comprehension bypass.

你就没这个工具可用了。”我觉得这个类比放在使用这些工具上同样成立，就是要非常审慎地对待“绕过理解(comprehension)”这件事。

You don't want to get a summary of Great Gatsby.

你不会想要一份《了不起的盖茨比》的摘要。

You want to sort of build and deepen your knowledge and and and understanding.

你想要的是去构建、去加深你自己的知识和理解。

I think in that sense, you know, there's sort of the saying in AI broadly, AI will make, you know, smart people smarter and less smart people less smart.

我觉得从这个意义上说，AI圈里有句大致的说法：AI会让聪明的人更聪明，让不那么聪明的人更不聪明。

I think that's very true in my in my journeys.

从我自己的经历看，我觉得这话非常真实。

If you have a bad process or you're lazy analyst and you want to bypass and generate AI AI slop, these tools are a bit dangerous.

如果你的流程很糟，或者你是个偷懒的分析师，想绕过流程去生成一堆AI垃圾内容，那这些工具就有点危险了。

Um but if you have a good process, you can really thoughtfully integrate these tools into your process, without a doubt, this will speed you up and it will create more rigor and more signal into into your investment process.

但如果你有一套好的流程，能很审慎地把这些工具整合进你的流程，那毫无疑问，它会让你提速，并给你的投资流程注入更多严谨度和更多信号。

And so, I think, you know, the value of a good process as, you know, finely articulated is more important than ever.

所以我觉得，一套精细表述出来的好流程，其价值比以往任何时候都更重要。

So, I think be be thoughtful about defining the skeleton of your process, whatever your process looks like.

所以我觉得，要审慎地去定义你流程的骨架，不管你的流程长什么样。

Be thoughtful about decomposing your workflows.

要审慎地去拆解你的工作流。

And one of the really easy ways to do this is a voice recording or or or or or brain dump recording in Granola or WhisperFlow.

而做这件事一个非常简单的办法，就是用语音录音，或者在Granola或WhisperFlow里把脑子里的想法一股脑录下来。

I now have this sort of folder on my on my iPhone.

我现在在iPhone上专门有这么一个文件夹。

My kids are in all sorts of sports, and so when I'm sitting in the bleachers, sometimes I'll have an idea and I'll just fire up WhisperFlow and brain dump it into a brain dump it into a a notes file and then drop that into Claude later.

我的孩子们参加各种运动，所以我坐在看台上的时候，有时冒出一个想法，就直接打开WhisperFlow把它录进一个笔记文件里，回头再把它丢进Claude。

Um and that really lowers the friction lowers the the the friction level for me.

这对我来说真的降低了摩擦、降低了阻力。

So, this brain dump dynamic sort of allows for the disambiguation that's so important in in aligning, you know, building your building your exoskeleton.

所以这种把想法一股脑倒出来的做法，能够实现那种在搭建你的外骨骼、对齐认知时至关重要的“消除歧义”。

And then take some time to learn learn about skills.

然后花点时间去了解skill。

Again, I don't think you need to learn how to build all the folders and subfolders, etc.

再说一遍，我不觉得你需要学会怎么建那些文件夹、子文件夹之类的。

cuz now these tools can do it can do it for you.

因为现在这些工具可以替你做。

But I think some basic understanding of what a skills file skills file is and then in the next seminar, this is a series, in the next seminar, we're going to go into a little bit of that and walk help help people build an up-to-speed skill, which we're thinking about just sharing if people want to see the before see the before and after.

但我觉得对skill文件是什么有一些基本理解还是有必要的，然后在下一场研讨会——这是一个系列——下一场我们会稍微深入讲一点，帮大家搭一个“快速上手(up-to-speed)”的skill，如果大家想看前后对比，我们正考虑就直接把它分享出来。

And sort of last one, especially before you use this institutionally, we have a deep and critical sort of sacred responsibility to our investors to deploy rigor.

还有最后一点，尤其是在你把它用于机构场景之前，我们对我们的投资人负有一份深切而关键、近乎神圣的责任，那就是要落实严谨。

And so a systematic validation loop is is a is a must.

所以一个系统化验证的闭环是必须的。

So really be thoughtful about both a systematic validation, which I think is one of the most under discussed but exciting elements of agentic capabilities that they can do that.

所以真的要审慎对待系统化验证，我觉得这是agentic能力里最被低估、但又最令人兴奋的一环——它们能做到这件事。

One of the tests I have a whole series of tests that I test when a new model comes out and one of it is uploading an Excel directly into chat and see if see if the model can evaluate that.

我有一整套测试，每当有新模型出来我都会跑，其中一项就是把一个Excel直接上传进对话，看看这个模型能不能对它做评估。

That wasn't possible to any sort of impressive degree until GPT 5.4 max, right?

在GPT-5.4 (max)出来之前，这件事根本达不到任何拿得出手的程度。

It really just in the last few weeks.

真的就是最近这几周的事。

Opus 4.6 still isn't quite there in my opinion.

在我看来，Opus 4.6还没完全达到那个水平。

But these tools are getting better.

但这些工具在越来越好。

They can now manage ingress and egress from Excel, which opens up a whole new realm of capabilities.

它们现在能管理Excel的ingress和egress(进出)，这打开了一整片全新的能力空间。

We've come a long way from chatbot interface.

从chatbot界面一路走来，我们已经走了很远。

When you can start to do autonomous research with Excel ingress egress and flagging systems, you're getting to systems that are capable of things that candidly are a little bit scary scary good.

当你能开始借助Excel的进出以及标记系统去做自主研究时，你就摸到了那种有能力做出一些坦白说有点“可怕地好”的事情的系统。

So that's again like March of 9s.

所以这又是那个March of 9s。

I don't know if we're there quite yet, but we're getting a lot closer.

我不知道我们是不是已经到那儿了，但我们正越来越接近。

Like we are getting a lot we're getting a lot a lot a lot closer.

我们正在越来越接近。

So we'll try and walk you through some of that.

所以我们会试着带你走一遍其中的一部分。

Today was a little bit more philosophical than some of the some of the webinars.

今天比之前有些webinar要更偏哲学一点。

This was meant to be a high-level overview of what's now possible with AI in investing meant to inspire your journey a little bit.

这一场的用意是对AI在投资中如今能做到什么做一个高层次的总览，稍微激发一下你自己的探索之旅。

I think what's important It's been very important to properly calibrate the time you deploy on AI.

我觉得很重要的一点是——恰当地校准你投入在AI上的时间一直非常重要。

I've talked to many investors who have wasted hundreds of hours over the last 3 years trying to get things to to work.

我跟很多投资人聊过，他们在过去3年里为了把东西弄好用，浪费了成百上千个小时。

I've talked to many funds who have wasted hundreds of hours and hundreds of thousands of dollars building an internal systems which don't get deployment because they get put into investors' hands and they do a simple search across 15 10-Ks and they don't find the metric that they expected to see and it sits there in the 10-K.

我跟很多基金聊过，他们浪费了成百上千个小时、几十万美元去搭内部系统，结果这些系统落不了地——因为一旦交到投资人手里，投资人在15份10-K里做个简单检索，没找到他们预期看到的那个指标，而那个指标其实就明明在10-K里。

So there's been a lot of obsolescence.

所以出现了大量的过时淘汰。

There's been a lot of abstraction where the early thing that was built now gets abstracted in the second thing.

出现了大量的抽象化——早先搭出来的东西，如今被抽象进了后来的东西里。

The hot thing used to be prompting and rag.

以前的热门是prompting和RAG。

You don't really hear much about that any any any more.

现在你已经不怎么听到有人提这个了。

So you hear about skills and agents and then I don't know in 18 months maybe it's all about intent and everything from from this 84-page slide deck is obsolete.

现在你听到的是skill和agent，然后我不知道，可能18个月后一切都在讲intent，而这份84页的幻灯片里的所有东西全都过时了。

That's probably probably the case.

很可能就是这样。

Um But this was meant to be sort of a high-level check-in of I think what's possible.

但这一场的用意，是对我认为可能的那些事做一个高层次的阶段性回顾。

In 2 weeks we'll give you an actual practical workflow of how to use Claude co-work to build an up-to-speed sort of initial primer on on on on a name.

两周后，我们会给你一套真正实操的工作流，教你怎么用Claude Co-worker给一只个股搭一份“快速上手”的初步入门材料。

We'll do a little bit of modeling with Excel and then sort of wrap up with some processes for sort of thought processes on how to take some of these agents that you build and create your own AI augmented research system that's validated in a way that's institution institutional grade.

我们会用Excel做一点建模，然后收尾时讲一些流程，一些关于如何把你搭出来的这些agent拿来、去打造你自己那套经过验证、验证方式达到机构级的AI增强型研究系统的思路。

Um So that's a sort of free webinar series.

这就是我们免费的webinar系列。

We're also working on a rebuild of our AI academy.

我们也在着手重建我们的AI academy。

This this happened in September.

这件事发生在九月。

This now I think is sort of you know feels very 2025.

现在回头看，我觉得它有一种很浓的2025年的味道。

Um As it as it was in 2025, we talked a lot about GPT projects and prompting and you know, building your prompting council etc.

就像2025年当时那样，我们大量讲了GPT projects、prompting，以及搭建你自己的prompting委员会之类的东西。

Some of that is an easy sort of next step to skills, but it's not the most updated perspective on things.

其中一些是通往skill的一个很自然的下一步，但它并不是对事物最新的视角。

So likely that that 6-month program ends end of April and likely in June we'll roll out a new program where we'll have a foundational seminar or monthly workflow labs where we'll just go more sort of deeper than we do in these sessions, you know, share some of the artifacts etc.

所以那个6个月的项目很可能在4月底结束，然后6月我们很可能推出一个新项目，里面会有一场基础研讨会，或者每月的工作流实验室，我们会比这些场次讲得更深入一些，分享一部分产出物之类的。

and then a monthly speaker Q&A sort of bringing in every month someone in the trenches building these tools so you're not just hearing from me but from other from other from other experts.

再加上每月一次的嘉宾Q&A，每个月请一位正在一线搭建这些工具的人来，这样你听到的就不只是我，还有其他专家的声音。

We're also doing quite a bit of this for for enterprises.

我们也在为企业客户做相当多这类工作。

The combination if it's interesting to you, the combination of um the combination often of what we see is a foundational seminar where we're going in live to number of firms doing a number of the of the Zoom where they customize the investment styles where we just you know, define terms, talk about where we spend stand tool selection, data prompting context and then help them with the actual workflow transformation.

如果你有兴趣的话——我们常见的组合是：一场基础研讨会，我们通过Zoom现场进入若干家公司，针对他们的投资风格做定制，我们会界定术语、聊聊工具选型上我们的立场、数据、prompting、context，然后帮他们做真正的工作流改造。

Simple things like where to bring a risk checklist in, how to build some of these tools.

一些简单的事情，比如在哪个环节引入风险检查清单、怎么搭建其中一些工具。

So if that's interesting to you, please email me info@fundamentaledge.

所以如果你有兴趣，请给我发邮件：info@fundamentaledge。

But I'm sort of more convinced and I am a sort of reflexively skeptical person.

但我现在算是更被说服了——而我是个本能上偏怀疑的人。

Someone reminded me the other day that in the early days of AI I was sort of did a webinar on Nvidia with a little bit more of a skeptical completely incorrect view on the stock.

前几天有人提醒我，在AI早期我做过一场关于Nvidia的webinar，当时对这只股票持一种偏怀疑、结果完全错误的看法。

And so I've been a little bit reflexively skeptical of AI, but I just need to follow the evidence of the workflows I see and the workflows I see are now quite quite powerful.

所以我一直对AI有点本能上的怀疑，但我只需要跟着我看到的那些工作流的证据走，而我现在看到的这些工作流已经相当相当强大了。

So I think the I think the reality that becoming AI fluent you know, the idea of deploying a swarm of agents is really important.

所以我觉得，让自己熟练驾驭AI这个现实，以及部署一整群agent(swarm of agents)这个想法，是非常重要的。

But also you blending it with the irreducible essence essence of human ability, the judgment, intuition and relationships.

但同时，你要把它和人类能力中那些不可约的本质——判断力、直觉和人际关系——融合起来。

All the value of your insights come from who you talk to every day, other investors, sell-side, you know, and so how do you how do you turn those insights those human level insights into AI scrapes and I just don't think there's a fundamental way to to do it, which sort of validates to me the the exoskeleton hypothesis.

你所有洞见的价值，都来自你每天跟谁交谈——其他投资人、卖方等等——那么你要怎么把这些洞见、这些人本层面的洞见转化成AI能抓取的东西呢？我就是不觉得存在一种根本性的办法能做到这一点，这在我看来恰恰印证了外骨骼假说。

All right.

好了。

So that was 80-something slides and I will stop my sharing and happy to take questions, comments, pushbacks, debates on on on um on on any of that.

以上就是80多张幻灯片的内容，我这就停止共享，很乐意接受关于这一切的任何提问、评论、反驳和辩论。

You can either raise your hand, although I will note that this will be will likely share this on YouTube with the YouTube link after.

你可以举手提问，不过我要说明一下，之后我们很可能会把这个视频连同 YouTube 链接一起分享出来。

So if you if you are shy, you can just um either raise your hand or just drop the question into chat and we will take the questions from from from from from chat.

所以如果你比较害羞，可以举手，也可以直接把问题打进聊天框，我们会从聊天里挑问题来回答。

And I have some time to take to take to take question on on the Q&A.

我还有一些时间可以在 Q&A 环节回答大家的问题。

All right.

好的。

Easy one first.

先来个简单的。

Will this be recorded?

这场会录像吗？

Yes, it will be recorded.

会，会录像。

Could you please compare various platforms and which one you think works the best?

你能不能对比一下各个平台，你觉得哪个用起来最好？

You know, they really have leapfrogged um and um They really have leapfrogged.

它们确实实现了跨越式的领先。

In the fall of last year I was a chat GPT user.

去年秋天我还是个 ChatGPT 用户。

You know, recently Claude has been you know, the platform the sort of hot platform just given the agentic capabilities.

最近 Claude 成了那个比较火的平台，就是因为它的 agentic 能力。

If you read the tea leaves on what open AI AI is built with their agentic systems on enterprise level and likely what's coming out of open AI in the coming months, my hunch would be that we will see something very similar using the codex system.

如果你去解读 OpenAI 在企业级用它的 agentic 系统做出了什么、以及未来几个月 OpenAI 很可能会推出什么，我的直觉是我们会看到用 codex 系统做出的非常类似的东西。

So I'd say you know, be thoughtful about about about sort of the obsolescence issue I think is not done yet.

所以我会说，要谨慎对待这个「过时淘汰」的问题，我觉得这一轮还没结束。

So I think be thoughtful about the ability to evaluate new tools new tools that that that that that that come out.

所以我觉得要认真对待评估新工具的能力，毕竟新工具会不断冒出来。

Question on Chris Murphy, why not use something like InSync for models?

Chris Murphy 的问题：为什么不用 InSync 这类工具来做模型？

I have I have in the past and I've I've used an India team to update update models.

我过去用过，我也用过一个印度团队来更新模型。

And people will use tools like Delupa and Delupa and I can analyst to update models.

人们会用 Delupa 这类工具，还有 AI analyst 来更新模型。

You know, to me I think there's a you sort of get on a soapbox about modeling, but a lot of the value of modeling is to have a chassis to do incremental analysis off of that off of that baseline.

对我来说，虽然一说到建模就容易站上讲台大谈特谈，但建模的很多价值在于有一个底盘，可以在这个基线之上做增量分析。

The simple three statements isn't always the most valuable part of the financial model.

简单的三张报表并不总是财务模型里最有价值的部分。

It's the sensitivities around what's happening.

最有价值的是围绕正在发生的事情所做的敏感性分析。

I use the example and I'm going to turn this into an agent like let's sort of use this the funny hypothetical of cos we get a press release that Costco's raising the price of a hot dog.

我举个例子，我打算把它做成一个 agent，就用这个有点搞笑的假设：Costco 发了一份新闻稿说要涨热狗的价格。

What is the sort of analysis we have to do how many hot dogs are sold, what's the what's the a sort of impact on volumes etc.

我们要做的分析是：卖出了多少根热狗、对销量的影响大概是什么，等等。

Flow that down to an EPS estimate.

把这些一路推算到 EPS 估计上。

So you know, it's not just updating models.

所以这不只是更新模型。

I sort of use as a use case, but it's the ability to to press a button in the morning press a button in the morning um you know, when there's a new acquisition or a new contract and get it sort of distill that down to an impact on the model quite quick quite quickly.

我把它当作一个用例，但关键是那种能力：早上按一个按钮，当有新的收购或新的合同时，很快把它蒸馏成对模型的影响。

I think the other thing that would be really important is ingress egress.

我觉得另一件非常重要的事是 ingress/egress（数据进出）。

Um and so having that having that model be the decision engine sort of the analytical engine of of research I think is a future and outsourcing modeling I think may be hard to hard to do that.

所以让那个模型成为决策引擎、成为研究的分析引擎，我觉得是未来的方向，而把建模外包出去，我觉得可能很难做到。

Notion of MCPs dying out.

关于 MCP 会消亡的说法。

I mean, I hear both sides of it.

我两种说法都听到过。

Um I'm using MCPs right now to pretty good effect.

我现在正在用 MCP，效果相当不错。

It's certainly better than if I go into Claude Excel right now, sometimes the retrieval mechanism to aggre breaks and so it goes and web scrapes the number.

它肯定比我现在直接进 Claude Excel 要好，有时候取数机制在做汇总时会崩，于是它就跑去网页抓那个数字。

Um I think I believe they will fix that.

我相信他们会把这个修好。

but I don't really have a great crystal ball on the tech the technical side.

但在技术层面我并没有什么很准的水晶球。

Will I share some of the skills at MD files you prepared or is that proprietary?

你会分享一些你准备的 skills 的 MD 文件吗，还是那些是专有的？

I think that's proprietary.

我觉得那些是专有的。

I'm trying to figure out I may share an up to up up up to up to speed skill in 2 weeks.

我还在琢磨，我可能会在两周内分享一个 up-to-speed 的 skill。

I'm still thinking about that.

我还在考虑这件事。

Um Um you know, uh so I don't I don't know uh about that.

所以我还不确定这个能不能分享。

A lot of this encodes, you know, the core IP of training IP I put together for for many years.

这里面很多东西编码的都是我多年积累的核心 IP、培训 IP。

So, I'm not sure I want to open source that, but I'm thinking I'm thinking about how much how much uh how much we want to uh uh we want to share.

所以我不确定我想不想把它开源，但我在思考我们到底想分享多少。

Question from Antonio.

来自 Antonio 的问题。

Hello, Antonio.

你好，Antonio。

So, do you run your processing cloud but delegate the Excel modeling to to chat GPT?

你是在云端跑你的处理流程，但把 Excel 建模交给 ChatGPT 吗？

Um I ran the the the modeling engine that does the best is Perplexity Pro because it has access to all the all the all the models.

我跑下来建模引擎里表现最好的是 Perplexity Pro，因为它能接入所有的模型。

So, I'm getting I I I got the most interesting reliable use case in Perplexity Pro uh sorry, Perplexity computer.

所以我在 Perplexity Pro——抱歉，Perplexity Computer 里得到了最有意思、最可靠的用例。

Um but Claude did a decent job in updating updating the models.

不过 Claude 在更新模型上也做得相当不错。

Um so um I still need to test out chat GPT Excel, which I believe is I believe is uh I believe is is out is out now.

所以我还需要测一下 ChatGPT Excel，我相信它现在已经上线了。

Uh uh where and when the recording will be available, we'll send it out if you're registered and we'll likely pay post it on YouTube um as well, too.

至于录像在哪里、什么时候能拿到，如果你注册了我们会发给你，我们也很可能会把它发到 YouTube 上。

What information providers have been the most useful in connecting to your agents via MCP?

哪些信息提供商在通过 MCP 接入你的 agent 时最有用？

The number one uh for me in this use case I've been using is Delupa.

在这个用例里对我来说排第一、我一直在用的是 Delupa。

Again, not sponsored by Delupa.

再说一次，不是 Delupa 赞助的。

Um but it was uh pretty easy pretty easy to connect and for the modeling use case they have these data data sheets that are quite uh quite quite um uh quite quite quite quite accurate.

但它连接起来相当容易，而且在建模这个用例里，它们有这些 data sheet，相当准确。

question from Steven.

来自 Steven 的问题。

Would you talk briefly again about validation?

你能再简单讲一下验证吗？

In short, how do you identify hallucinations such as when a a a a when a wrong number is listed?

简而言之，你怎么识别 hallucination，比如列出来的是一个错误的数字？

So, this is a really important question.

这是一个非常重要的问题。

That's why I'm sort of thinking about this two-tier validation system.

这正是我一直在思考这套 two-tier validation 系统的原因。

Number one, I want to have agents validate, right?

第一，我想让 agent 来做验证。

And I can use multiple models.

而且我可以用多个模型。

So, so the way I could do this I could have Claude build the model.

所以我可以这样做：让 Claude 建模型。

I could have Claude validate, Perplexity computer validate, I could have GPT 5.4 validate.

我可以让 Claude 验证、让 Perplexity Computer 验证，还可以让 GPT-5.4 验证。

So, three tabs open as with the validation project.

所以开三个标签页，都带着这个验证项目。

Go and validate these projects and have connections to MCPs.

去验证这些项目，并接上 MCP 的连接。

If all three give me 100% accuracy, I'm a little bit more more convinced that they're as accurate.

如果三个都给我 100% 的准确率，我就会更相信它们是准确的。

And then I can say based on the investments the trajectory the investment thesis here, this is what I really care about.

然后我可以说，基于这里的投资、走势和投资论点，这才是我真正在意的。

This is what I really need to be accurate.

这才是我真正需要准确的部分。

Create a checklist for me to go in and validate those numbers.

给我做一份 checklist，让我进去逐项验证那些数字。

I could hand it off to to an intern to go validate and source those numbers.

我可以把它交给一个实习生，去验证并溯源那些数字。

I can hand that to an Indian team to go validate and source those numbers, or I could do that on my own.

我可以把它交给一个印度团队去验证并溯源那些数字，或者我也可以自己来做。

So, that's I think to me a um not a perfect um system to catching hallucination, uh but something that so far is working out quite nicely uh quite quite quite quite quite nicely uh for me.

所以对我来说，这不是一套完美的、能抓住 hallucination 的系统，但到目前为止运行得相当不错。

All right, question from Chris.

好的，来自 Chris 的问题。

Uh Chris, what I like about Perplexity computer is I don't have to learn co-work and I'm convinced as as you are this is all extracted away.

Chris，我喜欢 Perplexity Computer 的一点是，我不用去学 co-worker，而且我和你一样确信这些底层的东西都会被抽象掉。

So, in in ways investing and learning with all evolving so fast not good use of time.

所以从某种程度上说，投资和学习都在飞快演变，花时间在这上面并不是好的时间投入。

Perplexity computer seems where this is headed swarm of agents and skills.

Perplexity Computer 看起来就是这一切的走向：swarm of agents 加上 skills。

I think that's right, Chris.

我觉得你说得对，Chris。

That's aligned with how I'm thinking about it.

这和我的思考方向是一致的。

Um even in our first uh version of the AI Academy, we spent a lot of time uh with our instructors teaching about LLMs and models and the technology and cuz the sense of that time is you need to know these things to get usability out of them.

就算在我们第一版的 AI Academy 里，我们也花了很多时间让讲师去讲 LLM、模型和技术，因为当时的感觉是你必须懂这些东西才能把它们用起来。

I think the shift to an AI to the AI accelerator is really compressing, you know, 12 hours into 3 hours of the basics.

我觉得转向 AI accelerator 是在大幅压缩，把 12 个小时的基础内容压进 3 个小时。

Um but really going into the work to the workflow sort of, you know, sharing our imagination, sharing our experimentation, sort of walking through how to use the how to use use the tool.

但更多是真正进入到工作流里，分享我们的想象、分享我们的实验，带着大家走一遍怎么用这个工具。

I I I don't want to be accused of like affiliation with Perplexity computer, but I like the the tool is really interesting.

我不想被指责说和 Perplexity Computer 有什么关联，但我确实觉得这个工具非常有意思。

Um and again, it's not perfect.

而且再说一次，它并不完美。

It still has a number of errors.

它仍然有不少错误。

But uh it's one of the first ones that just out of the box.

但它是最早一批开箱即用的工具之一。

You know, I went to use Claude Claude coding agents and you get in this approval hell where you have to approve every time uh you want to go do a web scrape, or you turn on you know, dangerously approve everything, which can go into your internal documents.

我之前去用 Claude 的编码 agent，结果掉进了那种审批地狱：每次你想去做网页抓取都得批一下，或者你打开那个「危险地全部批准」，那样它就可能碰到你的内部文档。

I'm like, I don't know if I want this tool touching my internal documents, which Perplexity computer does and it sits in a virtual environment.

我心想，我不确定我想不想让这个工具碰我的内部文档，而 Perplexity Computer 是会碰的，不过它跑在一个虚拟环境里。

Um so, I think that's where the innovation and a few people who I trust uh and are smart smarter than me sort of indicate that that multi-model approach is probably where a lot of the co-pilots go, I would think.

所以我觉得创新点就在这里，而且几个我信任、比我聪明的人也隐约表明，那种 multi-model 的路线很可能就是很多 co-pilot 的走向，我是这么想的。

Um the co-pilot, you know, there's you know, dozens of finance co-pilots that were effectively wrappers.

说到 co-pilot，市面上有几十个金融 co-pilot，本质上都是套壳。

They'll get mad if you say they're wrappers, so don't say they're wrappers.

你要是说他们是套壳，他们会生气，所以别说他们是套壳。

They're not wrappers.

它们不是套壳。

They're uh they're not wrappers.

它们不是套壳。

Um that are effectively wrappers on an LLM.

我是说，那种本质上就是套在 LLM 外面的壳。

I think it will be interesting to see many of those tools uh build around um these agentic systems with security, safety, context, etc.

我觉得会很有意思的是，看到很多这类工具围绕这些 agentic 系统去搭建，把安全性、安全防护、context 等等都做进去。

So, I'm actually like near-term much more bearish on the the finance co-pilot world, but if they can build something that is multi-model in nature that has, you know, skills embedded in the system.

所以短期内，我其实对金融 co-pilot 这个领域更看空，但前提是它们能做出一个本质上多模型、系统里嵌入了 skills 的东西。

Um it's the same sort of like I don't know where you sit on the SaaS debate, but the same sort of bull bullish part of the SaaS debate that these tools will make all the SaaS like so much more effective and user-friendly.

这跟 SaaS 那场辩论有点像，我不知道你站哪一边，但看多 SaaS 的那一派会说，这些工具会让所有 SaaS 变得高效得多、也好用得多。

I do think you'll start to see some of these co-pilots uh finance co-pilots that weren't that useful become incredibly useful in the next uh in the next few few months.

我确实觉得，你会开始看到一些原本没那么好用的 co-pilot、金融 co-pilot，在接下来这几个月里变得极其好用。

Um just activating activating some of these use cases.

就是把这些用例激活起来。

I think there's some questions on compute.

我觉得在算力上还有一些问题。

Uh how to access the compute um etc.

比如怎么拿到算力，等等。

But um that's what I'm sort of uh so, that's where we may that's where we may end up um on this I I I would think.

但这大概就是我的看法，我想这可能就是我们最终会落到的地方。

Um Question from Anonymous.

来自 Anonymous 的问题。

My firm's been lagging on AI adoption primary primary from a compliance risk management perspective to say SEC is scrutinizing AI usage in the investment process is common excuse you've heard.

我们公司在 AI 采用上一直落后，主要是出于合规和风险管理的考虑，理由是 SEC 正在审查投资流程中的 AI 使用，这是你常听到的借口吗？

Any recommendations on convincing compliance departments to be more more open to experimentation?

关于说服合规部门对做实验更开放一些，有什么建议吗？

This is a really important question.

这是个非常重要的问题。

Um many of our clients are using Microsoft co-pilot um uh because of this exact reason.

我们很多客户在用 Microsoft co-pilot，正是出于这个原因。

It's sort of been the only tool that has been approved at a number of clients.

它算是不少客户唯一获批的工具。

We get we it's sort of gone away a little bit, but we had a lot of requests to do Microsoft co-pilot trainings and our for response back is like, I have bad news for you.

这种需求现在稍微少了一点，但我们之前收到很多做 Microsoft co-pilot 培训的请求，而我们的回应是：我有个坏消息要告诉你。

Um the tool was just so bad.

这工具实在太烂了。

Um you know, recently the sort of Microsoft co-pilot co-work um integration, which is like the irony of all ironies after how much Satya has sort of dealt with with with OpenAI.

你知道，最近那个 Microsoft co-pilot 与 co-work 的集成，考虑到 Satya 之前跟 OpenAI 打了那么多交道，这真是最大的讽刺。

Um that marriage is sort of integrate with Anthro Anthropic now is sort of funny.

这段联姻现在居然转去跟 Anthropic 集成，还挺好笑的。

but um you know, I think um um uh that that is interesting that, you know, uh co-pilot with with the access co-work etc.

但你知道，我觉得 co-pilot 加上 co-work 的接入之类的，这挺有意思的。

The recommendations, I think it's just wait a little bit.

要说建议，我觉得就是再等一等。

Um the labs really want to make an enterprise push.

这些实验室真的很想在企业市场发力。

They're very smart people there.

那里都是非常聪明的人。

Um and I think they'll find ways to sort of uh convince corporate America of the compliance uh the compliance hurdle.

我觉得他们会找到办法，去说服美国大企业跨过合规这道坎。

So, I've seen just in the recent weeks more uh big name firms uh get the Claude Claude enterprise um Claude enterprise ap- approved.

所以就在最近几周，我看到越来越多知名机构拿到了 Claude enterprise 的审批。

Um so, I think that will be uh that will sort of take that Claude ecosystem wrap around enterprise-grade security of that of that as well, too.

所以我觉得，那会把整个 Claude 生态用企业级安全性包裹起来。

Um Question from Will.

来自 Will 的问题。

Does the Analyst Academy incorporate learning how to use AI?

Analyst Academy 有没有纳入如何使用 AI 的学习内容？

We haven't up until this next cohort.

到目前为止还没有，要到下一期才有。

So, in the April cohort and there's sort of debate.

也就是 4 月那一期，这里其实有一些争论。

Like there's a there's a I think a a proper debate on if I have interns if I have an intern class, should I have them do the sort of manual approach, or should I teach them AI maximalist approach?

我觉得有一个很正当的争论：如果我有实习生、有一整批实习生，我该让他们用那种手工的方式，还是教他们一套 AI 最大化的路子？

You know, we're a little bit old school, so we sort of default back to you should learn this business the artisanal way that you you uh you you earn the right to use AI in your investment process while understanding you know, using these tools today can really speed you up and drive more rigor.

你知道，我们有点老派，所以我们默认还是回到：你应该用手艺人的方式去学这门生意，你得先挣得在投资流程里使用 AI 的资格，同时也要明白，今天用这些工具确实能大幅提速、并带来更高的严谨度。

So, we're selectively integrating more AI into the Analyst Academy.

所以我们在有选择地把更多 AI 整合进 Analyst Academy。

We're rolling out other programs that have AI integration.

我们正在推出其他带 AI 整合的项目。

Um we're rolling out a modeling course, which is, you know, 8 hours of how to build a model step by step by hand.

我们正在推出一门建模课程，内容是 8 小时手把手一步一步教你搭一个模型。

Um and then also how to overlay AI into into into building that model.

然后还会教你怎么把 AI 叠加到搭建这个模型的过程里。

Um and so, I think in all of our programs, even diving deeper into the artisanal skeleton of building a process is important before you start to overlay the the AI.

所以我觉得，在我们所有项目里，在你开始叠加 AI 之前，更深入地钻研搭建流程的那副手工骨架，是很重要的。

So, more to come.

所以更多内容还在后头。

More to come on that.

这方面更多内容还在后头。

From Scott.

来自 Scott 的问题。

Hello, Scott.

你好，Scott。

How have you thought about putting together the final investment AI?

你是怎么考虑把最终的投资 AI 拼装到一起的？

I haven't.

我没有(这么做)。

Like personally, I find I find um like writing the final pitch to be something that I would want to stick with um cuz that's sort of a valuable distillation for distillation for me.

就我个人而言，我发现写最终那份 pitch 是我想留给自己亲手做的事，因为那对我来说是一次很有价值的提炼。

Often as I write my thoughts out, That's where my thoughts crystallize.

往往在我把想法写出来的过程中，我的思路才真正成形。

So, in the insight formation process personally, I wouldn't want to bypass that uh that process.

所以在洞见形成的过程里，就我个人而言，我不想绕过那个过程。

You may disagree.

你可能不同意。

You may disagree with that, but that's a that's sort of a view uh that's sort of a view that um a view view that I have.

你可能不同意这一点，但这算是我的一个看法。

How do you create the validation skills?

你是怎么创建那些验证 skills 的？

Is it derivative of the initial skill you commanded or a whole separate a whole separate whole separate script?

它是从你最初调用的那个 skill 衍生出来的，还是一整套完全独立的脚本？

Uh so, it's a derivative of the initial skill, but I went in and I I created a We have a um a section in our upcoming modeling course called the PM review, which is like when a PM looks at an analyst model, what are all the things a PM will look at to sort of check, you know, check check the architecture of the model, check the accuracy of the model, sort of the the spot check, and we built that So, the mindset is sort of like what would a PM do to check the validity of this model and then turn that into turn that into a turn that into a a a skill.

它是从最初那个 skill 衍生出来的，但我进去做了一步：我们即将推出的建模课程里有一个部分叫 PM review，就是当一个 PM 去看分析师的模型时，他会检查所有哪些东西来核对，比如核对模型的架构、核对模型的准确性，做那种抽查，我们把它做了出来，所以这个思路大概就是：一个 PM 会怎么去核对这个模型的有效性，然后把那套东西变成一个 skill。

Uh question from Michael.

来自 Michael 的问题。

Uh now that these AI agentic tools improved so much over the last 3 months, how do you see adoption curve of them changing within institutional workflows?

既然这些 AI agentic 工具在过去 3 个月里进步了这么多，你怎么看它们在机构工作流里的采用曲线会如何变化？

Does it accelerate the pace of model improvement or will compliance other bottlenecks keep the investment management industry behind early adopters?

它会加快模型改进的节奏，还是说合规等其他瓶颈会让投资管理行业落在早期采用者后面？

How do you see adoption pace differing between pods or single managers versus long onlys?

你怎么看 pod、单一 manager 与 long only 之间在采用速度上的差异？

Uh it's a it's an interesting question.

这是个有意思的问题。

I think the um adoption has been slow cuz the tool was just not very good for our use case.

我觉得采用一直很慢，因为这工具对我们的用例来说实在不太好用。

And so, you know, they think the reality of AI hype through most of 2025, you know, met the reality of chatbots not being that functionally useful for the investment process outside of a few few slices.

所以你知道，贯穿 2025 年大部分时间的 AI 炒作，撞上了现实——chatbot 在投资流程里除了少数几块之外，功能上并没那么有用。

I think that's very different now for for AI uh for sort of agentic AI systems.

我觉得现在对 AI、对 agentic AI 系统来说，情况非常不一样了。

So, my sense is that there's a real acceleration.

所以我的感觉是，确实出现了一次真正的加速。

As a as a trainer in this space, I can tell you that my inbox is quite inbox is quite busy.

作为这个领域的一名培训者，我可以告诉你，我的收件箱相当忙。

My calendar is getting quite uh my my quite busy.

我的日程也排得相当满。

So, I think firms that heretofore haven't taken it very seriously are taking it seriously, which I think is the right I think is the right call.

所以我觉得，那些至今还没太当回事的公司，正在开始认真对待，我觉得这是对的选择。

We've sort of been in the mindset of 2025.

我们一直抱着 2025 年的那种心态。

We we use the analogy of winter training.

我们用冬训来打比方。

It is it isn't spring yet, but uh you know, get do some sort of basic training uh ahead of spring.

现在还没到春天，但你知道，可以在春天之前做一些基础训练。

Um I think in the 2026, you start to you'll start to see um uh you'll start to see um um some real use cases.

我觉得到 2026 年，你会开始看到一些真正的用例。

Again, I think it'll be hard to measure because how do you measure, you know, more rigor more rigorous research?

不过我还是觉得这会很难衡量，因为你怎么去衡量更严谨的研究呢？

Um so, I think that's a that's that's a that's a challenge.

所以我觉得那是个挑战。

How do I see hedge fund analyst role changing with advancements in in AI?

随着 AI 的进步，我怎么看对冲基金分析师这个角色的变化？

This is a great question.

这是个很好的问题。

I'll tell you how I would think about it.

我来告诉你我会怎么想这件事。

Number one, as a PM or as a senior analyst, I would abstract more of the workflow the analyst workflow into my use case, right?

第一，作为 PM 或资深分析师，我会把更多分析师的工作流抽象进我自己的用例，对吧？

Simple things like the guidance achievability analysis that I might assign an analyst to do or in the past my PM assigned me the guidance uh the guidance compression analysis sort of detailed uh detailed consensus analysis of Vis Alpha does.

简单的事情，比如我可能派给分析师去做的 guidance 可达成性分析，或者过去我的 PM 派给我做的 guidance 压缩分析，也就是 Vis Alpha 做的那种详细的 consensus 分析。

I would pull more of that into my workflow.

我会把更多这类东西拉进我自己的工作流。

I would build agents around those things, systematic debugging around those things.

我会围绕这些事情搭 agent，围绕这些事情做系统化调试。

And I would steer my analyst more towards primary research, right?

然后我会把我的分析师更多地引向一手研究，对吧？

Let's go out and create, you know, creative including AI augmented.

我们出去做一些有创意的、包括 AI 增强的东西。

Let's go create, you know, AI augmented survey capabilities or hey, how can we go deeper?

我们去打造 AI 增强的调研能力，或者说，嘿，我们怎么能挖得更深？

What are the three What are the three you know, key industry trade shows in this industry?

这个行业里三个最关键的行业展会是哪三个？

I've sort of lived in two worlds as an investor, a Tiger Cub world where we were incredibly deep to a small number of companies and the multi-manager world where we were less deep on a broader subset of companies, but we're really keyed on the key inflections that drive those those stocks.

作为投资人，我算是活在过两个世界里：一个是 Tiger Cub 的世界，我们对少数几家公司钻研得极深；另一个是 multi-manager 的世界，我们对更广一批公司钻研得没那么深，但真正紧盯着驱动那些股票的关键拐点。

And so, the ability to blend that little bit to to bring more primary research into uh primary research into a multi-manager role, I think that's how I would shift the shift the uh the analyst workflow.

所以那种把两者稍微融合、把更多一手研究带进 multi-manager 角色的能力，我觉得那就是我会怎么去调整分析师工作流的方式。

After you create a context file, what's the next step?

在你创建一个 context 文件之后，下一步是什么？

Do your workflows encompass with multiple skill .md files?

你的工作流是不是由多个 skill 的 .md 文件构成的？

I'm just confused on the layering organization of a workflow.

我只是对一个工作流的分层组织方式感到困惑。

I didn't build any of this by hand.

这些东西我没有任何一样是手工搭出来的。

I built this in Perplexity computer and Claude Claude uh Claude co-work.

这个我是在 Perplexity Computer 和 Claude Co-worker 里搭出来的。

Um and really all I did is upload that context file and said, "Create some skills files for me." I uploaded those skills files, read the output, went back and made some changes to the skills file, and I was pretty impressed.

我做的其实就是上传那个 context 文件，然后说：“帮我生成一些 skills 文件。”我把那些 skills 文件传上去，读了输出，再回去对 skills 文件做了些改动，结果我挺惊艳的。

The sort of purists on the call are going to be like, "That's a really sort of inefficient way to do it." My pushback would be that didn't take me 8 hours.

在座的纯粹主义者可能会说：“这种做法其实很低效。”我的反驳是，它没花我 8 个小时。

It took me 15 15 15 minutes.

它只花了我 15 分钟。

Um and so um and so, the ability to create the skills files and even ask questions of, "Okay, how do I upload this back into upload this back into this the system now?" is quite uh quite um quite quite quite quite important.

所以，能生成 skills 文件、甚至还能追问“好，那我现在怎么把这个再传回系统里”，这一点非常非常重要。

Building your personal investing exoskeleton is it cost prohibitive or fairly inexpensive to build an exoskeleton with the Lupa Claude Perplexity pro options data subscription?

搭建你个人的投资外骨骼(exoskeleton)——用 Delupa、Claude、Perplexity Pro 加数据订阅来搭一副外骨骼，是贵得离谱还是相当便宜？

Any ballpark numbers in cost per month?

每月成本大概有个数吗？

Um The numbers can get high for sure very quickly.

这数字确实很容易迅速冲高。

Uh my token bill uh one of the pushbacks on Perplexity computer is it's not super token efficient.

我的 token 账单——对 Perplexity Computer 的一个批评是它的 token 效率不算高。

Uh so, my my token bill is getting higher.

所以我的 token 账单是在往上走的。

Um um you know, for something like uh for something like a fact set, the the the price they wanted for the MCP, I thought was um uh was not uh was not reasonable uh given I'm an existing uh uh subscriber.

像 FactSet 这种，他们对 MCP 要的价格，我觉得不太合理，毕竟我已经是他们的付费订阅用户了。

There are some new uh lower cost MCPs emerging.

现在也涌现出了一些成本更低的新 MCP。

Someone pointed me to Financial Modeling Prep that has a fundamental MCP.

有人给我推荐了 Financial Modeling Prep，它有一个基本面的 MCP。

Um um so um uh the former FinChat team has a sort of shifted to an MCP lower cost MCP.

原来 FinChat 的团队某种程度上转向了一个成本更低的 MCP。

So, I think there's an ecosystem of lower cost MCPs emerging.

所以我觉得，一个低成本 MCP 的生态正在形成。

Um But the numbers can get the numbers can get quite quite quite high.

但这数字确实可以变得相当高。

I think you can build like a pretty decent basic system for, you know, a few hundred dollars per month, but certainly an institutional grade system is going to be um uh hundreds of you know, hundreds of sorry, thousands of dollars uh per month.

我觉得每月花几百美元就能搭出一套相当不错的基础系统，但机构级(institutional grade)的系统，那肯定得每月几千美元。

We're going to explore this.

我们打算深入探究这个。

We're starting a podcast um on this as well, too.

我们也在就此启动一档播客。

And we're going to explore some of these vendors in the MCP system as well, too, to try and bring some more of that to bring some more of that to um uh to to to to light.

我们也打算把 MCP 体系里的这些供应商挖一挖，试着把更多这方面的东西讲清楚。

Probably time for uh for one or two um one or two more.

大概还有时间再回答一两个问题。

In the agentic AI era, do you think people are more empowered to build their own agent tools than relying on agents built by third-party vendors?

在 agentic AI 时代，你觉得比起依赖第三方供应商做好的 agent，人们是不是更有能力自己搭 agent 工具？

Yes.

是的。

Yes, Reting.

是的，Reting。

I I believe that that is the case.

我相信情况就是这样。

You know, there's been this sort of abstraction dynamic, this obsolescence dynamic.

一直有这么一种抽象化的动态、一种把旧东西淘汰掉的动态。

Effectively, what you can build in these tools now are your own agents, right?

实际上，你现在在这些工具里能搭出来的，就是你自己的 agent。

That's why it's so like the usability is so so exciting exciting.

这就是为什么它的可用性如此令人兴奋。

The real sort of step for for you is to identify what to build and why.

对你来说真正关键的一步，是想清楚要搭什么、为什么要搭。

And I think the tooling has has gotten um has gotten so much uh has gotten so much better.

而且我觉得，这套工具已经变得好太多了。

Um so um I think part of the training skill is teaching people to build their own agents in in internally.

所以我觉得，培训技能的一部分，就是教人们在内部自己搭 agent。

And today people are doing that with the you know, Claude code IDEs, VS code, etc.

如今人们是用 Claude Code、各种 IDE、VS Code 等等来做这件事的。

Um I my hypothesis is even that gets abstracted away, and that's why I sort of point people to Perplexity computer as a tool that abstracted that abstracted that away.

我的假说是，连这一层最终也会被抽象掉，这也是为什么我会把人们引向 Perplexity Computer，因为它就是一个把那一层抽象掉了的工具。

Um So, the ability to spin these up, spin a process up in a short amount of time, um I think is is the is the real unlock here because the way you code I don't know anything about coding, so I'm out of my depth here, but the way you code a calendar app I don't think is that different, right?

所以，能在很短时间里把这些东西、把一套流程快速跑起来，我觉得才是这里真正的解锁点，因为——我对写代码一窍不通，所以这块我说不上话——但你写一个日历 app 的方式，我觉得差别不大。

But the way you analyze a stock is really different based on what sort of stock you're looking at and what your investment style is.

但你分析一只股票的方式，会因为你看的是哪类股票、你的投资风格是什么而大不相同。

And so, this ability to create your own personalized agent agentic structure is a really important really important unlock.

所以，这种能搭建你自己个性化 agent 结构的能力，是一个非常非常重要的解锁点。

You don't want my agents because you may disagree with the weighting of my agents, right?

你不会想要我的 agent，因为你可能不认同我给这些 agent 设的权重。

People have very different philosophical, almost religious beliefs on certain parts of the workflow.

人们对工作流程里某些环节，抱着截然不同的、近乎信仰般的理念。

Uh you know, I might talk to one investor who insists his analyst build every model from scratch cuz that's how they learn the learn the business.

我可能会跟一位投资人聊，他坚持要他的分析师把每个模型都从零搭起，因为他们就是这么把生意搞懂的。

Another investor will use Can analyst models, right?

另一位投资人则会用现成的分析师模型。

And says that building a model from scratch is a waste of time.

并且说从零搭模型是浪费时间。

My job is not to judge those those opinions because if you're at this echelon of the investment community, you've been successful, like stick with it, right?

我的工作不是去评判这些观点，因为如果你已经到了投资圈的这个层级，你就是成功了，那就接着这么干好了。

Um my job is to sort of say like, "Okay, like based on your beliefs on the investment process, this is how you can build a system that can accelerate parts parts of that to institutional institutional accuracy." Question from Anonymous.

我的工作某种程度上是说：“好，基于你对投资流程的信念，你可以这样搭一套系统，把其中一部分加速到机构级的精确度。”下面这个问题来自匿名提问者。

May be time for one one or two more.

大概还有时间再回答一两个。

Uh how do you personally balance protecting the slow deliberate parts of your process against the fear of being left behind?

你个人是怎么在“保护流程里那些慢的、需要深思熟虑的环节”和“怕被时代甩下的焦虑”之间做平衡的？

This is really this is really important uh question.

这是一个非常非常重要的问题。

Number number one, you just have to understand in investing that, you know, you can't bypass rigor.

第一，你在投资里必须明白一点：严谨(rigor)是绕不过去的。

Uh if you bypass rigor and sort of buy buy a full position ahead of your real due diligence, like that might help you once or twice, but I think over time it's just the scar tissue of ensuring that you've done the rigorous work before you before you buy the buy the position.

如果你绕过严谨、在真正做完尽调之前就买满一个仓位，这也许能让你侥幸对个一两次，但我觉得长期看，那种“确保自己在建仓前把严谨的功课做足”的意识，是靠一次次教训攒出来的。

I think that's sort of one.

我觉得这算是第一点。

Number two, it's why I tell the algebra homework story of speed running an idea is tempting but dangerous.

第二，这就是为什么我讲那个代数作业的故事——赶工式地飞速把一个想法跑完，很诱人，但很危险。

Uh a lot in certain instances using these tools is about just creating more high signal reports to read.

在某些情况下，用这些工具很大程度上就是为了多生成一些高信号的报告来读。

Right?

对吧？

So, creating a more customized primer for the company based on your historical note notes around that company for an analyst.

比如，基于你多年来围绕某家公司积累的笔记，为分析师做一份更定制化的公司入门材料。

Um you know, if I have an a great RMS system I've covered HCA for 15 years, connecting all of those internal notes to create a primer for my new analyst who's covering hospitals, that's a really valuable way to give that new analyst the 15 years of context I had on on on a name.

你想，如果我有一套很好的 RMS 系统，我覆盖 HCA 已经 15 年了，把这些内部笔记全部串起来，为我那位新接手医院板块的分析师做一份入门材料，那就是把我对这只票积累了 15 年的 context 交给这位新分析师的一种极有价值的方式。

That means that doesn't mean you're speeding up, that's just means you're transferring more more more more more more rigor.

这并不意味着你在加速，只意味着你在传递更多的严谨。

Thoughts on Biverse build for a small investment team shop.

对于一个小型投资团队来说，是自建还是外购(buy vs. build)，你怎么看。

I've been building out a suite of tools internally, which has been going well.

我一直在内部搭一整套工具，进展一直不错。

Also, do you think the cost will increase decrease?

另外，你觉得成本是会涨还是会降？

$200 a month isn't cheap, but it but but it but it's worth it.

每月 200 美元不便宜，但它值这个价。

I do think to some degree, and this is just sort of passing on other views that Claude selling you 2,000 of compute for $200 or $100 a month right now.

我确实在某种程度上认为——这只是转述别人的观点——Claude 现在是在用每月 200 美元或 100 美元卖给你价值 2,000 美元的算力。

So, as sort of Uber did in the early stages, these foundation labs are likely subsidizing, you know, consumer price with their sort of VC capital.

所以，就像 Uber 在早期阶段做的那样，这些基础模型实验室很可能在用他们的 VC 资本补贴消费者价格。

That won't last forever, but also you sort of assume that the inference curve will will will will will will improve and hopefully hopefully catch up.

这不会永远持续下去，但你也可以假设推理成本曲线会不断改善，并希望它能追上来。

Biverse build is I think is a is an interesting decision.

自建还是外购(buy vs. build)，我觉得是个有意思的决定。

My sort of view is going to be easier to build, and so if that's already going well for you, I think it's fairly easy it will become easier to build a suite of tools internally even creating a a set of sort of shared skills files.

我的看法是，以后搭东西会越来越容易，所以如果这件事对你来说已经进展不错了，我觉得在内部搭一整套工具、甚至做出一套共享的 skills 文件，会变得相当容易、越来越容易。

Um a set of shared you know, context documents is something that when we talk to clients, that's what we support them on.

一套共享的 context 文档，正是我们和客户交流时给他们提供支持的地方。

Just creating the awareness of of of what to do.

就是先建立起“该做什么”的意识。

Even simple things like recording some of your your idea meetings, and after you know, 30 idea meetings recorded uh uploading those documents into an agent, and you sort of systematize the consistent pushbacks that people have on names.

哪怕是很简单的事，比如把你的一些选题会录下来，等录了 30 场之后，把这些文档上传进一个 agent，你就把大家对各只票反复提出的那些反驳系统化了。

Giving that agent to a junior analyst to mock pitch an idea before they walk into an investment committee meeting.

把那个 agent 交给一位初级分析师，让他在走进投资委员会会议之前先模拟推介一遍想法。

Simple things like that I think is where the alpha is now in the creativity.

像这样简单的事，我觉得正是如今 alpha 所在——就在这种创造力当中。

Building that building that as a tool I think is now more of a I don't want to sort of undersell how important it is cuz there still is some real data challenges internally.

把它搭成一个工具，我觉得如今更像是——我不想低估它的重要性，因为内部确实还存在一些实实在在的数据挑战。

But I think building internally has become more become easier easier easier to easier to do.

但我觉得，在内部搭建这件事已经变得容易多了。

and then last one I'll I'll give Routine the last one.

然后最后一个，我把最后一个留给 Routine。

Talking about primary research capability AlphaSense recently rolled out AI led expert interviews and the interviewees had positive experience.

说到一手研究(primary research)能力，AlphaSense 最近推出了 AI 主导的专家访谈，被访者的体验都不错。

Would this change how you firm think about primary research from the cost perspective?

从成本角度看，这会不会改变你们公司对一手研究的看法？

I think that's really interesting.

我觉得这非常有意思。

Really interesting and think in some sense like an AI cold call is is is not that interesting, but if you're paying the expert to once a month give answer three AI generated questions um and you can systematically do that every month um I even think like texting capabilities sort of an interesting thing I've heard where you can sort of do text surveys of experts in a panel and the AI system just runs that.

非常有意思，某种意义上我觉得一次 AI 打的陌生电话没那么有意思，但如果你付钱让专家每月一次回答三个 AI 生成的问题，而且你能每月系统化地这么做——我甚至觉得，我听说过一种挺有意思的做法是用短信，你可以对一个专家小组做文本问卷调查，由 AI 系统自动去跑这件事。

That's a really interesting way to build systematic close to source primary research capability.

那是一种非常有意思的方式，能系统化地建立起贴近源头的一手研究能力。

The mandate I always had in the Tiger Cub world is get as close to source as possible.

我在 Tiger Cub 那套体系里一直奉行的宗旨是：尽可能贴近源头。

Go talk to customers, go talk to clients, go talk to competitors.

去跟顾客聊，去跟客户聊，去跟竞争对手聊。

Um you know, summarizing a 10K isn't going to be a source of alpha, but getting really close to source and identifying inflections and key key drivers before they're apparent in apparent in financial statements is where a lot of a lot of alpha lies.

你想，总结一份 10-K 不会成为 alpha 的来源，但真正贴近源头、在拐点和关键驱动因素还没体现在财务报表里之前就把它们识别出来，才是大量 alpha 藏身的地方。

Um and so I think the primary research capabilities of these tools are really interesting and people are doing a lot of really innovative innovative things with that.

所以我觉得，这些工具的一手研究能力非常有意思，人们正在用它做很多很有创新性的事。

So.

就这样。

With that being said thank you for the questions.

话说到这儿，谢谢大家的提问。

I hope this was helpful.

希望这场对你们有帮助。

We will do we will do we will do three more of these.

我们还会再做三场这样的活动。

And yes, we run sort of structured, you know, academies and accelerators and stuff, but I think our mindset always is we've we try to share a decent amount of this stuff publicly to try and help drive the conversation on what's going on, and we have more of that coming with more of these webinars, a podcast a few other few other ideas as well too.

是的，我们确实办有结构化的学院、加速营之类的项目，但我觉得我们的心态一直是，尽量把这些东西相当程度地公开分享出来，来推动关于当下正在发生什么的讨论，后面我们还会有更多这类内容——更多这样的 webinar、一档播客，还有另外几个想法。

So, no no no hard sell ever from Fundamental Edge, but we'll sort of roll out program programs that can sort of help you as an individual user firm go deeper into this into this into this space.

所以，Fundamental Edge 绝不会向你硬推销，但我们会陆续推出一些项目，帮助你作为个人用户或机构在这个领域里挖得更深。

So, um uh Brett at Fundamental Edge is the is is the email or info at fundamentaledge if you want to reach out or connect, and other than that I will see see hopefully most of you uh uh in 2 weeks on the next free webinar series.

所以，如果你想联系或对接，邮箱是 Brett at Fundamental Edge，或者 info at fundamentaledge，除此之外，希望两周后大部分人能在下一期免费 webinar 系列上再见。

Thanks a lot for your time.

非常感谢大家的时间。