How the World’s Biggest Macro Hedge Funds Are Using AI | Jan Szilagyi

这项技术将极大地帮助我们以前所未有的规模从数据中提取洞见。

This was a technology that should go a long way towards helping us truly extract insights from the data on a scale that was never possible before.

你会提出一些过去会因为觉得太过离谱而直接放弃的问题。

You'll be asking questions that previously you would have dismissed because you thought this is so outlandish.

我不打算把分析师的时间浪费在这种事上。

I'm not gonna waste my analyst time on this.

但现在你不能再这样了，因为在你的咖啡还热着的时候，答案就已经出来了。

But now you can't because before your coffee is cold, you already have the answer.

该系统正在主动帮助投资者应对那些未知的未知，而投资者的许多焦虑正源于此。

The system is actively helping investors with the unknown unknowns, and that's where a lot of the anxiety investors comes from.

毫无疑问，已经有一些巨大的生产力红利出现，未来还会有更多到来。

It's undeniable that there are some huge productivity dividends that I think are already here and more coming down the line.

我的意思是，要维持这场繁荣，将需要大量的资源投入。

I mean, there is going to be a huge amount of resources that are gonna be required for this boom to basically be sustained.

前瞻指引中的任何内容均不构成购买或出售任何投资或产品的建议。

Nothing said on Forward Guidance is a recommendation to buy or sell any investments or products.

本播客仅用于信息目的，节目中任何人的观点均为其个人意见，而非财务建议，也不一定代表Blockworks的观点。

This podcast is for informational purposes only, and the views expressed by anyone on the show are solely their opinions, not financial advice, or necessarily the views of Blockworks.

我们的主持人、嘉宾以及Blockworks团队可能持有节目中讨论的公司基金或项目头寸。

Our hosts, guests, and the Blockworks team may hold positions in the company's funds or projects discussed.

一如既往，投资区块链技术涉及风险，条款和条件适用。

As always, investments in blockchain technology involve risk, terms, and conditions apply.

请自行进行研究。

Do your own research.

好了，各位。

Alright, everybody.

欢迎回到《前瞻指引》的又一期节目。

Welcome back to another episode of Forward Guidance.

今天和我一起的是一位新嘉宾，Jan Szilagyi，他是Reflexivity公司的首席执行官兼联合创始人，这家公司专注于将人工智能应用于全球各地的对冲基金。

And joining me today is a new guest for the show, Jan Szilagyi, CEO and co founder of Reflexivity, which is an AI software firm, really focused on bringing AI to different hedge funds across the world.

Jan，你的职业生涯也非常丰富，我们稍后也会聊聊你的背景。

Jan, you have a storied career as well, which we'll get into your background as well.

但首先，欢迎你来到节目，很高兴有你。

But first off, just want to say welcome to the show and great to have you.

谢谢。

Thank you.

很高兴能来这里。

Great to be here.

真的很享受这个过程。

Really enjoy it.

是的，太棒了。

Yeah, awesome.

嗯，我真的很想听听你的故事，因为我觉得这有助于更好地理解人工智能是如何被应用于全球各地的对冲基金的，尤其是在全球宏观策略中的应用，以及是什么样的经历让你走上了创立我们公司的道路。

Well, yeah, I would love to just hear a bit about you because I think it is relevant to getting to understanding better how AI is being incorporated across different hedge funds of the world, what it looks like for utilizing it across global macro and some of your own experiences that led to you towards founding us.

所以我想先请你简单介绍一下自己，你的职业生涯，以及你是如何走到今天的。

So I'd love to just start a bit of, yeah, just tell us a bit about yourself, your career, and how you got to where you are today.

当然，我很乐意分享。

Yeah, of course, happy to.

首先，再次感谢你邀请我参加这个节目。

Well, first of all, again, thanks for having me on the show.

我进入对冲基金行业其实有点偶然。

My path into hedge funds specifically was really a little bit accidental.

我知道我喜欢金融，但当我上大学最后一年时，我在耶鲁主修数学。

I knew that I liked finance, but when I was in my final year at university, so I did a degree in math Yale.

通过那段经历，我结识了斯坦利·德鲁肯米勒。

And through that ended up getting connected to Stanley Druckenmiller.

那次对话直接让我获得了与他共事的机会，真正学习如何管理全球宏观投资组合。

And that basically one conversation led to a job sitting alongside him and really learning how to manage global macro portfolios.

我发现这是一条非常有趣的职业道路，部分原因在于它让你能够提出极其前沿的问题，涉及地缘政治和经济。

And I found that to be an incredibly interesting career path in part because it allowed you to ask extremely, extremely topical questions that had to do with geopolitics, economics.

这些问题涵盖范围极广，然后你需要去思考它们可能对各种不同资产意味着什么。

They really span the gamut and then try to figure out what that might mean for a variety of different assets.

因此，你在资产类别和交易结构的选择上，感觉在某种程度上毫无限制。

So again, you felt in some way very unrestrained in so far as the asset class and the type of trade structure that you put on.

我在那里待了几年。

I spent a number of years there.

我休了两年半的时间，在哈佛攻读量化金融的博士学位，之后回来，又在斯坦身边待了几年。

I took about two and a half years off to do a PhD in quant fineness at Harvard, came back, stayed with Stan for a number of years after that.

然后我最终转去富达，与迈克·诺沃格拉茨一起管理全球宏观投资组合，而他后来进入了加密货币领域。

And then eventually moved over to manage global macro portfolios at Fortress alongside Mike Novogratz, who eventually went into the cryptocurrency land.

所有这些经历最终都汇聚成了 Reflexivity，不过我稍后再详细谈谈。

All of these experiences ultimately culminated into what is Reflexivity, but I'll talk about more about that later.

是的，完全对。

Yeah, totally.

在我们深入这个话题之前，我对你的某些经历感到很好奇，因为我觉得这很有意思——你去哈佛攻读量化金融的博士学位，而当你听德鲁肯米勒的访谈时，他总是很自豪自己只是一个真正的市场实践者，与金融学术界的观念完全相反。

Yeah, before we get into that, I am just a bit curious about some of the experience you have, because I think it is quite interesting contrast going to do a PhD in quant finance and all that at Harvard with, you know, when you listen, when I listened to Druckenmiller interviews over the world, he's always been very proud of being just a really a true proud market practitioner and sort of the opposite of the idea of academia within finance.

那段经历是什么样的？在你的经历中，结果如何？

What was that experience like and how did that net out in your experience?

我很感兴趣。

I'm curious.

我的意思是，当我跟他讨论这个问题时，很明显他认为这对我在对冲基金领域的发展并没有什么价值。

Well, I mean, think that when he and I talked about this, I think it was quite clear that he didn't think that that was gonna be value additive to my career in the hedge fund space.

我当时也意识到了这一点，但我更多是觉得这像是一个必须解决的执念。

And I think I recognized that at the time, I think I was very much thinking this is almost like an itch that I need to scratch.

回头来看，虽然当时并没有太大用处，但它确实在某种程度上帮助了我，让我开始思考如何系统化我们正在做的事情，最终促成了 Reflexivity 的诞生。

In hindsight, while it was not that useful at the time, it did prove to be more useful just in so far as then thinking about how to systematize some of the things that we were doing and ultimately how Reflexivity came to be.

但没错，当时这并不是提升职业发展的路径，尤其是在全球宏观和全球宏观对冲基金领域。

But yes, at the time, that was not a way to further your career, especially not in a global macro and a global macro hedge fund.

是的，完全说得通。

Yeah, yeah, totally makes sense.

好吧。

Alright.

嗯，就是这样。

Well, it's it's it.

让我们来聊聊 Reflexivity 到底是什么，以及你在人工智能领域都在做些什么。

Let's let's get into what Reflex is all about and and what you've been up to in the AI world.

首先，这个名字太棒了。

First off, love the name.

你知道，你参与了‘反身性’这个词的传承体系，这绝不是巧合，这个词的根源可以追溯到乔治·索罗斯，以及多年来他团队中的一些人，包括德鲁肯米勒，我相信你也与之相关。

It's, you know, it doesn't doesn't feel like a coincidence that you were part of the umbrella of that lineage of of where the word Reflexivity come from which for those that don't know is really rooted in the George Soros and then the proliferation across some of his different folks that work from over the years, including Druckenmiller, I'm sure related to you.

所以首先，你为什么选择这个名字？然后，什么是反身性？

So first off, why'd you choose that name and then what actually is Reflexivity?

正如你已经提到的，这确实不是巧合。

As you already mentioned, yes, it was not a coincidence.

我认为，对许多从事全球宏观投资的人来说，乔治·索罗斯至今仍代表着某种意义上的全球宏观投资之父，对吧？

I think conceptually for a lot of people who are in global macro, I think George Soros still represents to an extent kind of he's like the godfather of global macro investing, right?

对我们许多人来说，他之所以令人着迷，是因为他既能提出深刻的问题，又能将这些问题转化为赚钱的策略。

I think he, for many of us, made it seem just very cool because he could ask these very deep questions and yet at the same time also turn them into something that made money.

在他的案例中，赚到了非常多的钱。

In his case, quite a lot of money.

反身性这个概念对我们尤其相关，因为它真正探讨的是市场与基本面之间的双向互动：市场反映基本面，而基本面也会反过来回应市场。

The Reflexivity concept itself was relevant for us specifically because to the extent that Reflexivity really does talk about this back and forth between markets reflecting fundamentals, but then fundamentals also responding to markets.

我们构建的系统原则上是为分析师和投资组合经理设计的，使他们能够实时向公司或政府的决策者反馈，他们的决策可能将他们引向何方，从而可能改变这些决策。

The system that we have built is in principle meant for analysts and portfolio managers to be able in real time to give that feedback to whoever is running a company or a government, how and where their decisions might be taking them and therefore altering them potentially.

因此，我们对数据的理解越深入，越能判断某个决策是好是坏，这越可能成为决策者调整当前路径的信号。

So you do have the more we're able to understand based on the data, whether or not a particular decision is a good one or a bad one, the more that might then be a signal to the decision maker to potentially alter the current course.

所以，某种程度上，你是在强化这种联系，对吧？

So you're in some way, I think strengthening that link, right?

如果我们全都一无所知，就只能看着公司说：好吧，我们根本不知道它值多少钱，就像所谓的投票机与称重机的区别，对吧？

Like if we were all totally in the dark, we would basically just sort of look at the company and say like, okay, well, we have no clue what this is valued, kind of like the sort of the idea of a voting machine versus a weighing machine, right?

所以也许我们一直只是在投票。

So maybe we're just always kind of voting.

而如今，我认为借助技术，我们或许能在比以往短得多的时间尺度上成为一台称重机。

Whereas now, I think we might be able to be a weighing machine on a much, much shorter horizon than it would have been possible without technology.

这最终就是我们选择这个名字的初衷。

That ultimately was the thinking behind picking the name itself.

酷。

Cool.

所以，跟我们聊聊Reflexivity吧，我特别想听听你创办这家公司时的所见所感。

So yeah, tell us a bit about where Reflexivity and I would love for you to really weigh on what you saw when you founded it.

因为据我了解，你在ChatGPT热潮和相关炒作爆发之前就已经启动了这个项目。

Because to my understanding, you started this whole thing before the big wave of hype around ChatGPT moment and all of that.

而如今到了2026年，人人都在每天谈论AI，我们也开始看到真正可能性的萌芽。

And obviously now in 2026, everybody's talking about AI every day and we're starting to see green shoots of what's really possible.

但当初你起步的时候，这并不像现在这么明显。

But I don't think it was as obvious when you started.

我特别想听听，是什么让你决定离开全球宏观对冲基金领域，去创办这家公司？

And I would love to just hear about what did you see to make you leave the global macro hedge fund world to go start this thing?

我认为，公众层面的AI热潮确实是从ChatGPT发布开始的，但显然在此之前，人们就已经在尝试大型语言模型了。

So what I think became, well, it's definitely very clear that in some way, the kind of the public side of AI dates back to the release of ChatGPT, but obviously people had been experimenting with large language models before.

而且我认为，在ChatGPT发布之前，那些更接近这一领域的人就已经清楚地意识到，某些特别的事情正在发生。

And I think that it was even before ChatGPT clear to those who are closer that something special was happening.

但同样真实的是，当时一些技术还没有像今天这样容易获取。

What was also true, however, is that some of the technology wasn't as readily available as it is today.

所以，我们早期的很多工作也基于AI的其他一些方面，而且至今仍然如此，并不仅仅局限于大型语言模型。

So, a lot of our early work was also based on some other aspects and continues to be actually some other aspects of AI, not just large language models.

我始终持有一个非常强烈的直觉，虽然这至今尚未被完全证实，但至少现在看起来比五年前更有可能了：随着技术的发展，它将极大释放金融领域我们所拥有的所有数据中蕴含的价值。

The very strong intuition that I had continued to have, and I don't think it's necessarily been proven right yet, but at least it's looking better now than this five years ago, is that the technology as it's developing should represent a massive unlock of the value that is contained in all of the data that we have in finance.

这其中有一个非常简单的原因，那就是当你观察人类分析师如何分析某个市场事件时，他们的思维过程往往是线性的。

And there was a very simple reason for that, which is that when you notice the difference between how a human analyst will analyze a particular market event, it's quite linear.

你看到新闻标题或听到某个观点后，会一步步思考其影响、含义，并逐层分析等等。

You look at the headline or you hear of an idea and you then step by step try to think about the repercussions, what it means, analyze each step and so on.

而大型语言模型能够做到我所说的多维综合。

A large language model is capable of doing what I guess I could describe as multidimensional synthesis.

它同时处理所有这些可能性，对吧？

Like it's just doing all of these possibilities at once, right?

并不是说：‘油价上涨了。’

Not saying, Oh, oil prices are higher.

这可能会对通胀不利。

That could be bad for inflation.

这意味着利率可能会上升，或者利率上升会导致估值下降。

It means that it could be then higher rates or higher rates means multiple comes down.

相反，他们会这样想：哦，这对乙醇和汽油的供需平衡会有负面影响。

Instead, they're doing so like, oh, well, this is gonna be bad for ethanol, gasoline balance.

这对钻探公司有利。

It's good for the drillers.

这对生产商有利。

It's good for producers.

这将意味着利率上升。

It's gonna mean higher interest rates.

收益率曲线可能会变得更加陡峭，等等。

The curve could be much steeper and so on.

这是一种令人兴奋的方式来应对我们所拥有的海量数据，也是一种更快连接各点的方法。

That is an exciting way to deal with the abundance of data that we have and a much faster way to connect the dots.

因此，从概念上讲，我觉得这项技术将大大帮助我们以前所未有的规模从数据中提取洞察。

So it just felt like conceptually, this was a technology that should go a long way towards helping us truly extract insights from the data on a scale that was never possible before.

我觉得有趣的是，试图解决这个问题时，总是很有趣地观察全球宏观交易员如何以量化的方式形成观点。

What I find interesting about trying to solve for that problem is it's always so funny looking at how folks on global macro decide in a quantitative manner on a thesis.

比如，收益率曲线倒挂，对吧？

So for example, curve inversions, right?

每次只要看两年期和十年期国债利差，你知道，回顾过去几十年，人们总认为这预示着经济衰退，而其他一些研究流派则认为这种样本量太小，简直可笑。

Like every time, say the twos, tens, adverts, you know, if you look at the past few decades, there's been this idea that it leads to a recession and any sort of other school of work, you know, that small amount of sample size, which is, I don't know, kind of be laughed at.

但在全球宏观领域，这却是我们唯一拥有的工具，对吧？

But in global macro, it's all we have, right?

因为根本没有更好的替代方案。

Like there's no better alternatives.

就像，好吧，我们有10年期国债的数据，那就只能用它了。

It's like, okay, well we have an '10, so that's what we're going to run with.

而且，这感觉已经足够好，可以用来下注了。

And okay, it feels like good enough to make the bet there.

当你谈到能够整合如此大量的数据，从而得出这些新颖而有趣的洞察时。

When you talk about just the ability to synthesize so much data to derive these new interesting insights.

你觉得这能解决这个问题吗？即这么多不同的全球宏观策略都受限于如此有限的样本量？

Like what was, do you see that as a solution to this idea of just so many different global macro thesis are so constrained by just limited numbers of sample sizes?

你的客户，那些使用你们平台或人工智能的对冲基金，是否以这种方式来更精细地测试这些思路，即使样本量仍然相对有限？

Like, do you see, your clients, the hedge funds that are either using your platform or AI in general, are they using it in that way to be able to better test these seeds out in a bit more of a sophisticated way, even though the sample size is still somewhat restricted?

简短的回答是：是的。

The short answer is yes.

而且至少有两个原因，这能帮助解决你所说的样本量n相对较小的问题。

And there are at least two reasons why this can help with, I guess, as you're calling, so like the problem of n being relatively small.

首先，我给你一个具体的例子。

One is that, and I'll give you a concrete example.

假设你是一名全球宏观基金经理，如果运气好的话，你的职业生涯中至少会遇到一次或多次新兴市场危机，对吧？

Let's say that if you're a global macro manager and lucky in your career, you'll have run into at least one or several emerging markets crisis, right?

因为通常只有在这些时候，你才会看到货币和利率利差出现大幅波动。

Because that's when you'll typically see a big dislocation of currency and rate spreads and so on.

很难根据任何一个国家的历史来分析这些情况或应对措施，因为除了少数例外，这些危机通常不会频繁发生，让你觉得‘太好了’。

It is difficult to analyze that or the responses on the basis of any one country's history because, with a few exceptions, they generally don't have them regularly enough for you to be like, oh, great.

我真希望能有15个关于巴西的样本，来理解比如恶性通货膨胀之类的情况。

I'd love to have a sample of 15 for just, let's say, Brazil to understand, you know, like hyperinflation and so on.

然而，我认为反射性理论结合这项技术非常擅长的一点，是更好地定义‘相似’这个概念。

However, what I think Reflexivity can be very good at, help with this technology, is to better define the concept of similar.

这意味着，我不一定非得只研究巴西才能理解危机的内在结构。

Meaning I don't necessarily have to just study Brazil to understand the anatomy of a crisis.

我可以定义一个新兴市场国家的总体范围，然后尝试构建一个数据集，将基本面相似国家的类似经验纳入其中。

I can say, well, let's define the universe of emerging market countries and then try to create a data set that is going to effectively bring in similar experiences from countries with relatively similar fundamentals.

这又是另一种在一定程度上增加样本量N的方法。

And that's another way of then increasing the N at least somewhat.

第二个相关的原因是，如果你能更好地理解驱动这些敏感性的经济逻辑，最终你其实并不需要那么大的样本量。

The second and related one that I think also helps with a small N is if you have a better understanding of the economic logic that drives these sensitivities, you ultimately don't really need that large of a sample size.

通常需要大样本，是因为如果我们对驱动因素或内在逻辑不够清楚，就希望有足够的重复性或充分的证据，证明某种现象以某种频率发生，这样我们才敢在押注时相信它会再次发生。

A large sample size is often required because if we don't totally understand what the driver is or what the logic is, you want it to be repeatable enough or that there's enough ample evidence that it happens with such and such cadence that you trust that it'll happen again when you bet on it.

另一方面，比如我们长期研究的一个模式是：油价长期处于非常陡峭的远期升水状态，往往是原油现货价格触底的良好指标。

On the other hand, if, you know, let's say, so one pattern that we were studying for a long time is that a very steep contango in oil prices over time has been a great indicator of a bottom in cash prices of crude oil.

但即使我们只观察到这种极端情况发生过七到八次，你依然会非常有信心下注，因为你理解其中的机制，对吧？

But even though we only have probably seven or eight such occurrences of it being so steep, you would always feel very comfortable betting on that because you understand what is happening, right?

从基本面来看，生产者本质上是在把石油留在地下，他们卖出远期合约，做出将石油从现在转移到未来的决策。

Like fundamentally, producers are basically leaving oil on the ground, they're selling futures out, they're making decisions that are actually relocating oil from today to the future.

因此，今天实际上创造了一种稀缺性。

And therefore there is actually a scarcity that is created today.

如果我理解了这种价格与特定现象之间关系的驱动因素，我就没必要看到它发生57次。

If I have that kind of understanding of what drives this relationship between prices and a particular phenomenon, I don't need to see it happen 57 times.

对我而言，七次就足以让我非常有信心地下注了。

Like seven can be ample for me to be very confidently betting on it.

这很酷。

That's cool.

我真希望你能把这一逻辑应用到当下，因为现在确实很及时——我们正经历一场由伊朗引发的石油供应冲击，原油曲线也处于非常陡峭的反向市场状态。

I would love if you could actually apply that to the current day because it does seem somewhat timely that we have an oil supply shock going on right now with Iran and oil curve that is in very steep backwardation.

一出现这种情况，大家第一反应就是：好吧，我们来把这个石油冲击和2007年的情况做个比较。

Right away, the first thing everybody does is, okay, well, gonna go compare this oil shock to, you know, 2007.

我要把它和1970年代做比较。

I'm gonna go compare it to the 1970s.

我还要把它和2022年做比较。

I'm gonna go compare it to 2022 as well.

也就是说，我们最终可能需要三到五个案例。

It's like, all right, we got to end up maybe three to five.

现在我得决定我的投资组合该怎么办。

And now I have to decide what the hell I'm going do with my portfolio.

我真想听听你如何根据AI和反身性来思考当前的局势。

I would love to just hear about how you're thinking about that current situation in light of what you can do with AI and Reflexivity.

如果你现在尝试在反身性系统中测试这一点，它肯定会持续寻找其他类似供应端冲击的因素。

If you currently try and test this in Reflexivity, it's definitely going to be continually looking for other what are effectively supply side shocks.

对吧？

Right?

我觉得‘七’这个数字有点特殊，因为当时紧接着就爆发了全球金融危机。

I think 'seven was a little bit peculiar because it was basically then immediately followed by kind of a global financial crisis.

然后突然间，人们都惊呼：天哪，需求真的在消失。

And then all of a sudden people are like, oh my god, demand is actually evaporating.

对吧？

Right?

或者像新冠疫情时期，你突然觉得：再也不会有人用石油了，因为我们再也不会出门了，诸如此类。

Or like, for example, COVID when, again, you suddenly said like, no one's gonna use oil ever because we will never leave our house again and so on.

所以目前，它能够找到许多非常有价值的类比，回溯到上世纪70年代，当时出现了某些禁运，石油供应被明显地政治性限制。

So, at the moment, it's able to find actually quite a lot of very fruitful analogies going back to kinda like the 1970s and so on when you had some of these embargoes and you basically had very, very clearly politically restricted supply of oil.

当时人们在短期内根本无法确定在哪里以及如何填补那部分缺失的石油。

And people then in the short term couldn't really figure out where and how to replace some of that missing of that missing oil.

当然，现在的不同之处在于，我们已经拥有更成熟的其他能源体系。

Now, difference, of course, now is that we do have a more developed ecosystem of other types of energy sources and so on.

所以，现在的制约不像当年那么严重。

So it's not quite as constrained as it was back then.

但正如你从价格本身看到的那样，归根结底，我们显然还没有摆脱对石油的依赖，对吧？

But I mean, as you can see from the prices themselves, ultimately, we definitely have not weaned ourselves off of oil, right?

它仍然在很大程度上驱动着我们世界中的许多行为。

Like it still drives a lot of what we do in the world.

因此，我认为所有这些例子依然非常相关。

And so I would say fundamentally, all of those examples continue to be very relevant.

人工智能特别适合捕捉时代情绪，并将其与当时发生的事件、任何正在进行的谈判等进行对比。

And AI is particularly well suited to be able to also glean the sentiment of the time and compare it, like what was happening, how did they compare to any kind of negotiation that was taking place and so on.

所以，是的，我认为任何目前使用这项技术的人——如果你还没有使用，我强烈建议你尝试——都已经在理解当前局势上取得了巨大优势，并至少能帮助你规避一些潜在的负面影响。

So, yeah, I think anybody who's currently using this technology, and if you're not, I definitely recommend that you do, has had a great head start on understanding what's happening and hopefully, at a minimum protect from any fallout.

嗯。

Mhmm.

这很酷。

That's cool.

好的。

Alright.

我想稍微宏观一点来看，如果你能帮我们更清晰地描绘一下全球宏观对冲基金，或者一般来说的对冲基金，比如多策略基金，或者你们的一些客户的情况。

I wanna just zoom out a little bit and get if if you could just help us paint a a better picture or understanding of how global macro hedge funds or just hedge funds in general, you know, multi strats, I believe, or or some of your clients as well.

我真的很想听听，您能否帮我们理解这些大型基金是如何思考在投资和交易过程中运用人工智能的？

Would just love to hear if you could help us understand how are are these big funds thinking about utilizing AI in their in their investment and trading process?

无论是针对反射性理论，还是他们只是使用一般的大型语言模型之类的技术。

Whether that's specific to Reflexivity or you know, whether they're, you know, using just general LLMs and that sort of thing.

能不能给我们详细讲讲，目前最大的应用场景在哪里？

Like walk us through where the biggest use case is happening right now.

完美。

Perfect.

首先，我应该强调的是，也许为了帮助大家更好地理解，为什么金融领域的人会真正使用这项技术。

So what I should highlight first is that just maybe to zoom out a little bit to help everybody understand why is it that anybody in finance would actually be using this.

显然，每个人的第一反应是：为什么我不直接去用Claude或ChatGPT呢？

Obviously, the first thought for everybody is, well, why couldn't I just basically go to, you know, Claude or ChatGPT and so on?

为什么我不能做同样的分析？

And why couldn't I do that same analysis?

我认为，任何使用过这项技术超过几周，并尝试用它进行更复杂的量化分析的人，都会发现它给出的答案差异极大，这让你很难基于这些结果做出高风险的决策，这也很合理。

And I think anybody who's used it for longer than a couple of weeks and try to use it for any more complex quantitative analysis, then you realize that you get some a really wide discrepancy of answers that makes it difficult for you to base some high stakes decisions on this, which makes sense.

大型语言模型本质上并不是为计算而设计的。

Like, large language models ultimately were not built to calculate.

它们是为了理解大量信息而构建的。

They were built to understand a lot of things.

它们在理解方面非常出色。

They're really good at understanding it.

它们最终能够通过按需编写脚本等方式辅助计算，这确实令人惊叹，但这并非它们最初的设计目的。

It's incredibly impressive how well they are ultimately also able to help with calculation by writing scripts on demand and so on, but that is not what ultimately they were designed to do.

我们实际上创建了一种组合方案：利用大型语言模型的推理层，作为你与庞大分析库之间的中介，并结合一个专有的知识图谱，将所有分散的数据源关联起来，使推理引擎能够理解市场的第一阶、第二阶、第三阶影响等。

What we have effectively created is a combination of using the reasoning layer of these large language models to act as an interpreter between you and a vast analytical library, and then a proprietary knowledge graph that has linked all these disparate data sources to give the reasoning engine some sense of what are the first order effects, second order effects in the market, third order effects, and so on.

这意味着它能相当详细地理解，收益率曲线陡峭化可能对银行股价格产生影响等等。

Which means that it understands in quite some detail that a steepening yield curve might have implications for banking stock prices and so on.

这两者的结合——即数据与大型语言模型推理层之间的基础设施，需要解释、映射等操作——最终使我们能够帮助投资者应对那些已知的未知问题，这是我们客户的一个重要应用场景。

A combination of these two, so this infrastructure between the data and the reasoning layer of large language models that requires interpretation, mapping, and so on, ultimately means that we are able to help investors, and this is a big use case for our clients, with the known unknowns.

也就是说，那些由市场动态或投资研讨会等引发的问题。

So questions they have that might have been spurred by something happening in the market or an ideas dinner and so on.

我们希望让问题与答案之间的差距变得非常容易缩小。

And we are hoping to really just make it very easy to really shrink that gap between the question and the answer.

我们如何能最快地完成分析，从而回答你脑海中的问题，让你能够应对各种问题？

How can we as quickly as possible conduct the analysis that hopefully answers the questions that are on your mind, and you'll be able to answer all sorts of questions.

对吧？

Right?

比如，我可以缩短这个差距，你会提出一些过去因为觉得太离谱而放弃的问题。

Like, I can shrink that gap, you'll be asking questions that previously you would have dismissed because you thought this is so outlandish.

我不会浪费分析师的时间去处理这种事。

I'm not gonna waste my analyst time on this.

但现在你不能了，因为你的咖啡还没凉，答案就已经出来了。

But now you can't because it's gonna be five minutes, right, before your coffee is cold, you already have the answer.

我们能够做到的第二件事，特别是得益于这个知识图谱和认知能力，是填补盲点。

The second thing that we are able to do, particularly because of this knowledge graph and awareness, is also cover blind spots.

这意味着，该系统正在主动帮助投资者应对未知的未知。

And what this means is that the system is actively helping investors with the unknown unknowns.

而这正是投资者焦虑的来源——可能会发生什么？

And that's where a lot of the anxiety for investors comes from, which is what might be happening?

我的同行们在做什么，而我却不知道，但我本该知道？

What are my peers doing that I don't know, but I should know about?

比如，有哪些我遗漏的信息？

Like what is out there that I'm missing?

它之所以能做到这一点，是因为知识图谱有效地为它提供了处理任何市场事件、任何新闻标题的方向，并通过追踪图中的边来思考其连锁反应。

And the reason it's able to do this is the knowledge graph effectively gives it some direction of how to process any market event, any news headline, and then think about the ripple effects by just tracing out the edges of the graph.

我给你的最佳类比是，谷歌之所以主导了搜索市场，是因为它意识到，如果重新组织网络内容，从仅仅基于搜索词的频率，转变为基于页面之间的链接数量并优先考虑后者，就能获得更快、更相关的结果。

The best comparison I'll give you of this is if you consider how Google ended up dominating the search market is that it realized that if you reorganize web content, moving from just the frequency of the search word to actually the number of links between pages and prioritize that, you ended up with faster and more relevant answers.

这与此类似，因为如果你回到我最初的观点，即没错，大语言模型在多维综合方面非常出色，但它们确实需要某种映射来降低问题的维度。

This is similar because if you go back to the point that I was making initially, which is, yes, LLMs are really good at this multidimensional synthesis, but they do need some kind of a mapping to reduce the dimensionality of the problem.

比如，金融领域显然存在一个问题——几乎所有事物似乎都相互关联。

Like, finance has obviously the issue that, like, almost everything seems related.

我的意思是，这正是它令人兴奋的原因。

I mean, that's what makes it exciting.

连锁反应无处不在。

The ripple effects are everywhere.

但如果不计算每一个敏感性，你如何区分重大影响和微小影响呢？

But how do you distinguish between a big one and a small one without calculating every single sensitivity?

这时，这种映射就非常有帮助，能够有效建立某种层次感。

That's where that mapping is extremely helpful to effectively create some sense of hierarchy.

这太酷了。

That's super cool.

所以听起来，目前最大的用例之一是能够快速有效地进行验证和测试各种假设，但最终的想法仍然最需要来自真正的投资经理、首席投资官或其他相关人员。

So it sounds like one of the biggest use cases right now is the ability to quickly and effectively, you know, kick the tires and test theses out, but it's still most useful for an idea to come from the actual, you know, PM or CIO or what have you.

是这样吗？

Is that correct?

因为说实话，就我个人而言，最近我一直在极度痴迷于大语言模型，并尝试各种智能体框架。

Because you know, even in my own experience, I've been in all honesty, like very, very obsessed with LMs and trying out the different agent harnesses lately.

对我来说，就是我想去关注一下，最近几个月，我们看到制造业PMI在时隔几年后首次突破50，与此同时，制造业就业岗位也首次出现了正增长。

And for me, it's like, okay, I'm going to go, you know, I just want to, you know, in the last few months we've seen the manufacturing PMI go above 50 for the first time in a couple of years, at the same time that we saw like manufacturing payrolls go positive for the first couple of times.

我当时就想，好吧，Claude，你去查一下过去很长一段时间里，这两个指标同时突破50大关的几次情况。

I was like, okay, Claude, like go through the last few times that we've seen both of those trigger positive after a long time of being 50.

在那种情况下，接下来十二个月表现最好的资产是什么？

And what are the next twelve months best performers of assets in that regime?

结果还挺不错的。

And it was pretty good.

至少让我得到了一个想法。

Managed to at least get an idea.

但有一点是，它所使用的数据集在某些方面变得有点混乱，影响了稳健性。

But, you know, there was a bit of that robustness that in terms of the data sets that it was pulling from that got a bit messy.

但总的来说，我觉得这真的太酷了，我脑子里冒出一个随机的想法，就能去测试它，两分钟内就能得到答案。

But overall, it just seemed really cool for me that I was able to just, there's this random thought in my head and I could go test it out and I would get an answer within two minutes.

而如果没有使用LLM之类的东西，我不知道要花多长时间才能完成。

Whereas I don't even know how long that would have taken me without using an LLM or something like that.

所以，你现在发现的最大用例就是这类情况吗？

So is that the sort of thing where you find the biggest use cases right now?

是的。

Yes.

然而，重要的是要注意，当你开始深入挖掘数据的长尾部分时，Reflexivity和Claude这类系统之间的巨大差异才会真正显现出来。

However, it's important to note that where you're going to start to really see a big difference between a system like Reflexivity and Clod is when you start tapping into the kind of the longer tail of the data.

我来举个例子。

I'll give an example.

我们当时在做这个演示，正好是特朗普总统建议可口可乐用真糖替代果葡糖浆的时候。

We had, you know, we had this we were doing this demo, and this was right around the time when basically President Trump suggested that Coca Cola should replace fructose with real sugar.

显然，他们开始考虑这个方案。

And obviously, they started considering it.

但在我们做演示时，客户说：这正是我想要看到的。

But as we were doing this demo, the client said, Well, this would be a good way for me to see.

我们没必要花两天时间去整理糖市场的供需平衡数据。

We're not going to spend two days putting together supply and demand balance in the sugar market.

这并不是我们的专长。

That's not our specialty.

但看看这是否会带来任何我们需要注意的交易影响，会很有趣。

But it would be interesting to see if this has any trading repercussions that we should be aware of.

大约八分钟后，该系统基本上整合了夏威夷、佛罗里达、德克萨斯和路易斯安那这四个主要产糖州的甘蔗收成数据，并从标普全球的财务报表中识别出几家可能因这一转变而受损的公司。

In about eight minutes, the system basically pulled together data on, you know, sugar harvest across Hawaii, Florida, Texas, Louisiana, kind of the four main sugar producing states in financial statements from S and P Global and was able to identify a couple of companies that were going to be hurt by that switch.

同样，它还指出巴西是填补这一缺口的最佳选择，因为巴西地理位置接近，且拥有大量可供应的糖。

Similarly, it then identified Brazil as being the best way to fill that gap because it was close, it had quite a lot of sugar that it was able to supply and so on.

接着，它又发现了几家从巴西出口糖但似乎被低估的公司。

And then again, a couple of companies that export sugar from Brazil that appeared undervalued.

在所有这些时间内，系统实际上为你提供了两到三天的分析工作量，但更重要的是，它让你获得了那些仅靠Clod几乎无法获取的数据——除非你懂得如何梳理和关联这些信息。

In all of that time, it effectively gave you probably two, three days of analysis, but more importantly, access to the data that would be very hard unless you knew how to map it and so on for you to get just through clot.

我认为，对于某些更专业的应用场景来说，拥有可靠数据并让Reflexivity能够找到它们，这一点至关重要；而Claude可能只会基于大量关于可口可乐生产和糖市场现有的摘要，给出一些定性的推测。

And I think this is where for some of the more specialized use cases, it becomes really important that you have reliable data and that the Reflexivity is able to find it, where Claude might come back and just give you a kind of a qualitative description of what he thinks could happen based on a lot of the summaries that obviously exist in terms of coke production and the sugar market and so on.

酷。

Cool.

这也引出了一个问题：我觉得阻碍所有人全面投入这一领域的最大因素，就是大语言模型的幻觉问题。

That brings up the question too around I feel like the biggest thing holding back for everybody to be all in on this whole thing is hallucinations around LLMs.

而且，就像我刚才用PMI做案例那样，你必须要么相信LLM所做的一切都是真实、可验证且计算正确的，要么就得反过来自己去核查和测试，但这几乎就违背了使用它的初衷。

And, you know, if like using that case study of what I just did with the PMI is like, you have to either trust that everything that the LLM did was true and verifiable and calculated properly, or you have to go backwards and do and test it yourself, which almost defeats the whole purpose of using it.

所以我很好奇，在涉及巨额资金、数十亿美元的决策时，你怎么看待幻觉问题？

So I'm just curious, how do you think about hallucinations when there's very high stakes, multi billion dollar books depending on these decisions?

是的。

Yeah.

不，这是个很好的问题。

No, it's a great question.

而且这确实是我们在不断思考的问题。

And it's one that obviously we think about a lot.

我们有几种方式基本上可以完全避免幻觉。

There are several ways in which we more or less are able to avoid them entirely.

一种方式是，你从Reflexivity获得的每一个答案，最终都是以代码优先的方式输出的。

One is that every answer that you get from Reflexivity is ultimately code first output.

意思是，当你提出一个问题时，推理层的作用是理解你真正想要什么，然后接入我之前提到的解释性基础设施——即知识图谱之间的那一层，获取数据，并直接在原始数据上运行任何分析。

Meaning when you ask a question, the role of the reasoning layer is to understand what is it that you want, then dip into that interpretive infrastructure that I mentioned, the layer in between the knowledge graph and so on, procure the data, and then run any sort of analysis directly on that raw data.

如果它无法访问数据，就会返回一个错误。

If it doesn't have access to the data, it'll just return an error.

它不会凭空编造一个答案。

It doesn't then make up an answer.

它会明确告诉你：无法获取数据，因此无法执行你的分析。

It'll actually say like, hey, wasn't able to get the data, so I cannot conduct your analysis.

当然，你仍然可能遇到数据本身错误的问题。

You still run into the problem obviously of incorrect data that could happen.

但我们也会通过始终保持高度透明和完全可审计来应对这一问题。

But we try to also address that by always making it very transparent and completely auditable.

你可以看到分析的每一步，以及每个数据系列的来源。

You can see every step of the analysis and where each of the data series came from.

所以，如果你突然发现：等等，你给我展示的市盈率数据似乎来自一个我从未听说过的来源，可能完全错误——至少这会提醒你，你可以去核实。

So that if suddenly you're like, wait a second, the PE series that you're showing me here seems like it's coming from a source that I'd never heard of and that might be totally wrong, at least it alerts you to it and you can check it.

但我们尽最大努力确保你根本不需要走到这一步，因为我们的目标是已经将所有这些数据源连接并整合在一起。

But we try as hard as we can to make sure that you never have to even get to that step because the idea is that we have connected and integrated all of these data sources already.

它们通常只是投资者已经熟悉的最优质内容，对吧？

They tend to be only the most premium content that investors are already familiar with, right?

比如你使用的是IBIS和DataStream等数据。

Like you're using data from IBIS and DataStream and so on.

所以这些并不是他们不熟悉的数据源。

So it's not sources that they wouldn't be familiar with.

只是我们能够以前所未有的程度挖掘这些数据。

It's just that we're able to mine them to the extent that previously wasn't possible.

很棒。

Cool.

我想听听你对人工智能如何影响更广泛金融行业的看法。

I wanna hear about your perspective on how AI starts to impact the broader finance industry.

我最近最关心的问题是：如果我们向前看几年，每个人都能使用这些新工具，当每个人都拥有AGI大脑、随时可用时，还怎么产生超额收益和竞争优势呢？

And I feel like the biggest question I have on my mind these days is, all right, if we fast forward through the next few years and everybody has these new tools at their disposal, what happens to generating alpha and sources of edge if everybody just has, you know, AGI brains at their, you know, in the palm of their hands?

你怎么看这个问题？

How do you think about that?

在这种情况下，你在哪里获取超额收益？

Where do you generate alpha in that regime?

这是个非常有趣的问题，我只能就短期情况给你一个回答，因为我认为任何超过五年的事情，根本不可能做出任何大家愿意听的预测。

It's a very interesting question and I will give you an answer only about the shorter term of it because I think that anything beyond sort of like five years, I think it just becomes impossible to make any, any predictions anybody wanna even listen to.

我认为，就我们目前构建的技术而言，它无疑是一个赋能工具。

I think currently what I see from the technology that we have built, it's definitely an enabler.

我的意思是，它让你能够高效地测试各种不同的问题，最终使得问题本身变得更加有价值。

And what I mean by this is that it effectively allows you to test a variety of different questions very efficiently, which ultimately means that the questions become, the questions themselves become even more valuable.

如果我提出的问题非常糟糕，得到的成果也很少，系统不太可能突然说：嘿，你知道吗？

If I'm asking very bad questions that lead to very little, it's unlikely that the system will suddenly be like, hey, you know what?

你的训练框架有问题。

Your training framework sucks.

我觉得你真正应该做的是这些。

I think you should really just here's what I think you should be doing instead.

我们或许能达到那个阶段，但那更接近于真正的自主AI交易员，而不是我们目前所处的阶段。

Like, we might get to that point, but I think that is closer to really an autonomous AI trader as opposed to where we currently are.

我认为这是一个令人兴奋的时刻，最初由于一些人使用了这项技术而其他人没有使用，可能会导致市场上的差异而非缩小。

I think that's an exciting moment where I actually think there might be initially, partly because some people are using the technology and others aren't, going to create more disparity in the market rather than less.

就像你能够拥有一个并非来自……你之前提到过斯坦，对吧？

Like you will be able to have somebody who doesn't come from and you were asking about Stan before, right?

这将赋能那些本质上缺乏量化交易背景，却能对市场提出良好直觉性问题并仍能用数据验证的人。

Like this is going to empower people who in principle have very little quantitative trading, but are able to ask good intuitive questions about markets and then still test them against the data.

但它扩大了市场上可能被提出、甚至被实施的交易思路的范围。

But it kind of widens the scope of trade ideas that will be out there floated and potentially put on and so on.

现在，随着技术变得更好、更具辅助性，也就是说，它能评估你说的话，比如：‘简，我明白你认为这些是显而易见的交易，但其实你错了，我建议你考虑的是别的方向。’

Now, as the technology gets better and it becomes more assistive, in other words, you know, assess what you're saying, be like, look, Jan, I understand that, you know, you think that these are the obvious trades, but actually you're wrong, and here's what I think you should be thinking about instead.

你有没有感受到这种趋同？

Do you get a little bit of that convergence?

是的。

Yes.

但即便如此，每个人的风崄偏好、投资期限等仍然不同。

But even then, not everybody has the same risk appetite, investment horizon, and so on.

所以我认为，最终你还是会发现，人们在表达交易策略等方面仍然会存在很大差异。

So I think ultimately, you will still find that there will be a lot of variety in terms of how people express trades and so on.

一个非常不完美的类比是，当人们从通过传真机查看价格报价，转变为能够使用彭博终端的时候。

And a very imperfect analogy is probably when people switch from, you know, looking at price quotations on fax machines to then having access to like a Bloomberg terminal.

当时并没有因为每个人都拥有了彭博终端——现在确实大多数人都有了——而突然导致机会消失，因为大家都能获得更好的信息。

It wasn't suddenly that, well, because everybody has a Bloomberg terminal and it really feels like now most people do, that opportunities just disappeared because everybody had access to better information.

如果我们真的实现了智能的完全融合，那么也许这种情况是有可能发生的。

If we get to a complete melding of intelligence, then I suppose that's possible.

但我觉得这是在预测那样的世界会是什么样子。

But I think that is predicting what that world looks like.

那种感觉非常、非常不同，对吧？

It feels very, very different, right?

因为那样的话，从某种意义上讲，我们到底在交易什么呢？

Because then in some sense also like what are we actually trading?

如果我们大家本质上都想买同样的东西，那谁在卖呢？反之亦然？

Like if we're all basically wanting to buy the same thing, who's selling and vice versa?

对。

Right.

好的。

Okay.

所以听起来，这仍然需要由能够整合想法的人来使用。

So it sounds like it still needs to be utilized by somebody who's can actually put ideas together.

并不是说，在某个平行宇宙或某个时刻，普通人就会觉得，好吧。

It's not so much like, you know, in one alternative universe or at one point we get to this, say, you know, just the the average person is like, okay.

嘿，ChatGPT。

You know, hey, ChatGPT.

嘿，Claude。

Hey, Claude.

你帮我生成一个投资组合，五年内跑赢标普500，夏普比率还很高，别出错，然后就放手去做。

You know, make me a portfolio that outperforms the S and P 500 on a five year basis with a with a high sharp make no mistakes and then off to the races.

而且所有人都同时这么说。

And everybody's saying that at the same time.

然后，阿尔法就会被压缩。

And then there's alpha alpha just gets compressed.

突然间，所有人都只做多贝塔。

And then suddenly everybody's just long beta.

你真的不觉得这样的世界在未来五年左右至少已经初现端倪吗？

Like, you don't think really that world is at least on the horizon for the next five years or so?

我的意思是，说清楚一点，我认为这确实是一个非常可能的路径。

I mean, look, to be clear, I think that's a very possible path.

我只是觉得，目前金融领域最终看起来，这些关系不仅数量众多，而且极其复杂，并且不会随时间保持恒定。

I am just thinking that at the moment, finance ultimately does seem like these relationships are not only are there many of them, they're very complicated, and they don't stay constant over time.

嗯。

Mhmm.

因此，我认为，即使对于人工智能系统来说，要内部创建出可以反复应用的规则也是个挑战。

And so I think that it's a challenge even for an AI system to create internally rules that it could apply over and over again.

因为让你感到兴奋的一点——你也提到了这一点——就是：当你只有寥寥几个其他案例时，你该如何进行全球宏观交易？

Because one of the things that makes this exciting, and you highlighted this, was, well, how do you trade in global macro when you only have like two other cases?

对吧？

Right?

你可以稍微扩展一下，但从人工智能的角度来看，某些交易和投资领域仍会存在非常大的误差范围。

Like you can expand it a little bit, but there is gonna be even from an AI perspective, some margin of error that in some trades and in some areas of investing remain very, very large.

是的，也许你在做出决策时会更快，这确实有帮助。

And yes, maybe you'll be faster at arriving at a decision and that definitely helps.

但这并不意味着你在这一具体决策上会更频繁地正确。

But it doesn't mean that you'll necessarily be right more often on that specific decision.

所以我认为这是一种极其令人兴奋的技术。

So I think it's an extremely exciting technology.

我认为它绝对会让投资变得更有趣，也更有回报，因为你能够探索市场的很大一部分。

I think it's absolutely going to make investing a lot more fun and also really rewarding because you'll be able to explore such a large part of the market.

我还不清楚这些路径最终会如何发展，比如最终由一个交易员为我们所有人完成所有事情。

I'm not yet clear on how these paths are gonna then work out where eventually you just have, you know, a single trader doing it, doing everything for all of us.

有道理。

Makes sense.

这引出了一个好问题：你认为哪种交易或投资风格最适合作为AI的应用对象？

That brings up a good question around which styles of trading or investing do you think are best predisposed to AI?

因为我们之前主要讨论的是宏观策略，这很有用，因为样本量有限。

Because of course, we've been talking largely about macro, which is useful because of this idea of limited sample size.

但你想想，比如股票多空策略，你需要深入研究各种10-K文件之类的东西。

You But know, think about say, equity long short where you really focus on just like digging through different 10 ks filings and that sort of thing.

我觉得在那里AI会非常有用。

I imagine it's very useful there.

比如高频交易，它在那里有用吗？

You know, maybe high frequency trading, like, it is it useful there?

你怎么看待不同策略中哪些最适合作为AI的应用？

How do you think about the different styles in which are best predisposed?

实际上，股票策略也有道理，部分原因是我们有大量的数据可用。

You know, actually, so equities does make sense also partly because we have a lot of data available.

所以我认为这是这项技术非常有前景的应用领域。

So I think it's a very fertile ground for this technology.

另一个我认为非常有前景的领域是商品交易，因为这里有海量的微观信息。

Another area that I actually think is very promising is commodity trading, because there's an incredible amount of very micro information.

但这些信息并不总是能通过容易获取的数据源获得。

It's not always available in easy to access data sources.

因此，做得好的人常常自己也难以保持这些模型的及时更新等等。

And so people who do it well often actually themselves struggle keeping some of these models up to date and so on.

如果你能将这些框架应用于这种海量数据，比如从处理单个仓库发票开始，精确了解每个仓库每平方英尺的持仓成本，再实时对比从阿姆斯特丹运送到纽约的成本等等，我相信你能从中发现大量有趣的模式。

If you were to apply some of these frameworks on that wealth of data and, you know, going literally from processing individual warehouse invoices so that you knew what the cost of carry was for, I mean, down to basically the square foot of that warehouse, compare that to the price in real time of what it would mean to ship it from Amsterdam to New York and so on, there would be a lot of interesting patterns that I think you could uncover with, again, that wealth of data.

还有，同一商品在不同港口的价格差异等等。

Also, different prices at different ports for the same good and so on.

我的意思是，这似乎是一个特别适合用强光照射海量现有信息的领域。

Mean, it feels like an area that could definitely benefit from kind of shining, you know, a floodlight on the huge amount of information that exists.

这很酷。

That's cool.

在过去几个月里，你觉得AI代理领域是不是充满了大量热情？

What you think about over the last few months, it feels like there's been a lot of excitement around AI agents.

你如何看待这些？

How do you see those?

你如何看待AI代理在这一领域中的作用？显然，我们之前主要讨论的是数据和信息的整合，但现在还出现了一种全新的现象，即AI代理会主动去执行任务，甚至可能用于执行交易订单。

How are thinking about AI agents in landscape of, obviously, we've been, I feel like largely talking about just this data and information synthesis, but there's also this whole new thing of these AI agents going out to actually do things and potentially even, you know, could apply that to executing on orders.

我们已经讨论了很多关于投资流程的前端，但关于在执行端使用它呢？

You know, we've talked a lot about the top side of the investment process, but like what about utilizing it on the execution front?

我的观点非常乐观。

I mean, I'm very bullish on that.

我认为，在投资中确实存在许多流程，通过改进可以取得远超当前的效果。

I think that absolutely there are a lot of processes that exist in investing where you could do far better.

执行在某种程度上可能是一个自然的选择，因为到目前为止，人们有时会使用算法来执行交易。

And execution is to an extent maybe a natural one because even up to this point, people would sometimes have either like algorithms execute.

我们已经具备了使这一过程相对可靠的基础设施。

Like, we have the infrastructure already in place to kind of make that relatively reliable.

因此，我认为这有着非常光明的未来。

And so I do see that having a very, very bright future.

我认为这些智能体另一个有趣的应用领域是我们花了大量时间分析市场数据，但其实分析你自己的交易行为也会非常有意思，对吧？

The area where I think it's going to be interesting to see some of these agents also is we spend a lot of time analyzing market data, but I think actually it's going to be also interesting to analyze your own trading behaviour, right?

比如，有趣的是发现：你的想法确实不错，但你的执行能力太差了。

Like, think what would be interesting to kind of find out is, sure, your ideas are great, but you're just terrible at execution.

你总是进场太早，或者你很擅长买入，却非常不擅长卖出，诸如此类。

Like, you're always way too early or like you're really good at buying, but you're terrible at selling and so on.

就像我们之前提到的Whoop设备，对吧？

Like, it would be interesting to actually have, you know, we were talking about whoop before, right?

但举个例子，假如你有一个类似的东西，能实时监控你的交易行为，并提醒你：别在周五交易了。

But like, for example, let's say that you had something similar that would just be in real time, basically monitoring your own activity and kind of highlight, look, just stop doing trades on a Friday.

这样根本行不通。

Like, it's just not working.

每次你这么干，最后都是亏钱卖出。

Like every time when you do this, it turns out that you end up selling at a loss.

是的。

Yeah.

我当时笑是因为这完全就是我的真实经历，有时候我会看看自己那些大的主题性想法。

I was laughing there because that's just like entirely my experience where sometimes I look at my big thematic ideas.

我会想，哦，这方面我还不错。

I'm like, oh, I was kind of good on that.

但当我查看实际的交易执行时，却发现远没有表面上看起来那么严谨。

And then I look at my actual execution of the trade and it's just it's a lot sloppier than it looks on the surface.

我把所有的交易记录都导入了AI，然后问它：嘿，能给我一些有趣的洞察、相关性之类的吗？

I took all my trade logs and I put it in AI and I was like, hey, like tell me some interesting insights or like, you know, correlations or what have you.

结果它某种程度上狠狠地‘教训’了我一顿。

And it's just like, it kind of roasted me in some ways.

它说：别再交易那个资产类别了。

It was like, stop trading that one asset class.

你根本就擅长不了这个。

Like you suck at it, basically.

而且我认为，最终来说，从机器那里获得这种反馈可能比其他方式更容易。

And I mean, I think that ultimately is it's probably easier to get that feedback from a machine than it would have been otherwise.