本集简介
双语字幕
仅展示文本字幕，不包含中文音频；想边听边看，请使用 Bayt 播客 App。
大家好。
Hey, everyone.
我是科里·霍夫斯坦。
Corey Hoffstein here.
在开始之前，如果你关注资本效率和更聪明的投资组合构建，记得把日期标出来。
Before we get started, if you're into capital efficiency and smarter portfolio construction, mark your calendar.
第二届收益叠加研讨会将于2026年10月28日在芝加哥举行，主讲人是AQR的创始人兼首席投资官克里夫·阿内斯。
The second annual return stacking symposium is coming to Chicago on 10/28/2026, headlined by Cliff Asness, founder and CIO of AQR.
一整天的对话都将围绕如何叠加收益、更高效地运用资本，以及重新思考投资组合的构建方式展开。
It's a full day of conversations on how to stack return streams, put your capital to work more efficiently, and rethink the way you build portfolios.
无论你是投资顾问、资产配置者，还是只是对收益叠加感兴趣，这里都是你该来的地方。
Whether you're an advisor, allocator, or just curious about return stacking, this is the room to be in.
前往 www.returnstacked.com/symposium 了解更多信息并预订席位。
Head to www.returnstacked.com/symposium to learn more and grab your spot.
网址是 returnstacked.com/symposium。
That's returnstacked.com/symposium.
好了，我们正式开始。
Now on with the show.
好的，本。
Alright, Ben.
你准备好了吗？
Are you ready to go?
我准备好了。
I'm ready.
好。
Alright.
三、二、一。
Three, two, one.
咱们开干吧。
Let's jam.
大家好，欢迎各位。
Hello, and welcome, everyone.
我是科里·霍夫斯坦，欢迎收听《与模型调情》，这档节目将揭开量化策略背后的神秘面纱，探索那些真正驱动策略的人的因素。
I'm Corey Hofstein, and this is flirting with models, the podcast that pulls back the curtain to discover the human factor behind the quantitative strategy.
科里·霍夫斯坦是Newfound Research的联合创始人兼首席投资官。
Corey Hofstein is the cofounder and chief investment officer of Newfound Research.
由于行业监管要求，他在这档播客中不会讨论Newfound Research的任何基金。
Due to industry regulations, he will not discuss any of Newfound Research's funds on this podcast.
节目中所有嘉宾的观点均为个人意见，不代表Newfound Research的立场。
All opinions expressed by podcast participants are solely their own opinion and do not reflect the opinion of Newfound Research.
本播客仅供信息参考，不应作为投资决策的依据。
This podcast is for informational purposes only and should not be relied upon as a basis for investment decisions.
Newfound Research的客户可能持有节目中讨论的证券或资产。
Clients of Newfound Research may maintain positions and securities discussed in this podcast.
如需更多信息，请访问thinknewfound.com。
For more information, visit thinknewfound.com.
本期嘉宾是本·威灵顿，Two Sigma公司复杂特征引擎团队的负责人。
My guest this episode is Ben Wellington, head of Complex Feature Engines at Two Sigma.
在现代量化流程中，有人会说，优势可能存在于三个地方。
In a modern quant process, one might argue that edge can live in three places.
一是你能拿到的数据，二是你如何处理这些数据，三是你如何基于数据做出预测。
The data you can get your hands on, what you do with that data, and how you forecast from it.
本就卡在中间这一层——特征生成，而这也是人工智能正在最快重塑的领域。
Ben lives squarely in the middle layer, feature generation, which also happens to be the place AI is reshaping the world fastest.
所以我们把重点放在这里。
So that's where we spend our time.
我们先从什么是特征说起，为什么随着原始数据越来越普及，真正的优势越来越取决于你从数据中构建出什么。
We start with what a feature even is and why as raw data gets commoditized, the edge increasingly comes from what you build out of it.
接着我们会贯穿整个对话，聊聊大型语言模型。
Then we follow the thread running through the whole conversation, large language models.
本有一句话我一直记在心里：现在任何东西都可以变成语言，我们会探讨这句话打开了哪些可能，以及让特征创建变得如此廉价，会不会反而让优势变得人人皆有。
Ben has a line I keep coming back to that anything can be language now and we trace what that unlocks and whether making feature creation this cheap just democratizes the edge away.
最后我们会谈谈，一个今天刚入行的量化从业者，该朝着什么方向努力，以及在AI成为主导工具的行业里，哪些技能最能长期复利增长。
We close on what a quant starting out today should be building toward and which skills compound most in a career where AI is the dominant tool.
请欣赏我与本·威尔士的对话。
Please enjoy my conversation with Ben Wellington.
本，欢迎来到节目。
Ben, welcome to the show.
这次对话对我来说特别令人期待。
This is gonna be an exciting one for me to dive into.
得知能有机会和你聊天，我真的很高兴。我们之前请过不少人来聊股票和股票因子，但你做的工作在这一领域确实很新颖，深入探讨你的方向，正好能触及当今系统性量化股票投资者最前沿的问题。
And when I found out I was gonna get the opportunity to chat with you, it was great because we often, I've had a lot of people come on to talk about equities and equity factors, but I think some of the work you're doing is really novel in the conversation where to dive in here really gets to the cutting edge of what it means to be a systematic quantitative equity investor today.
非常感谢你参与这次对话。
So I really appreciate you joining me.
非常感谢你的邀请。
Thank you very much.
很高兴能来。
I'm happy to be here.
谢谢你的邀请。
Thanks for having me.
在正式开始之前，能不能先给听众们简单介绍一下你的背景？你是谁，读过什么学校，怎么一步步走到Two Sigma的？也说说你职业生涯初期是怎样的，再快速讲讲你现在在Complex Feature Engines团队的情况。
So before we dive in, maybe we can give the listeners a little bit of lay of the land and your background and who you are and would love to get a sort of sense of where you went to school and how you ended up at Two Sigma and what the beginning of your career looked like, and take us, you know, the quick run through to where you are today in the Complex Feature Engines team.
是的。
Yeah.
我在纽约大学读的博士，研究方向是自然语言处理，主要做机器翻译。
I did my PhD in natural language processing at New York University with it really a focus in machine translation.
你得回过头想想，那时候跟电脑说话、用翻译工具，这些事还特别新鲜，那时候还没Siri，也没Alexa。
So you have to kind of go back in time a little bit to think about the fact that it was still quite odd to be, say, speaking to a computer or using translation or before Siri, before Alexa, all this stuff.
当时就觉得这领域挺酷的。
And it just sounded like a cool field.
我选研究生方向的时候，其实根本不知道自己想干嘛，就翻了翻一些研究方向，觉得这个挺有意思，就选了。
So when I was looking to go to grad school, because I didn't know what to do with my life, I read some research areas and it just seemed it just seemed fun.
现在回头看，我当年选的这个方向，居然成了经济发展的核心部分之一，但说实话，那会儿纯属运气。
So dove into what, I guess, fast forward today, turned out to be a pretty central part of the growing economy, but that was mostly luck.
对吧？
Right?
我读完博士后，第一份正式工作去了纽约苏豪区的Two Sigma，那会儿公司大概一百来号人，租了个loft办公。
I did my PhD, and my first real job after my PhD was a place in Soho, New York called Two Sigma, with about a 125 or so people in sort of a loft building.
面试时见到的那些人，个个都挺有意思，特别有活力，我就决定加入了。
And when I met people on the interview, I was like, these are interesting, cool, fun people.
一进去就全身心投入了。
And so I jumped right in.
其实我刚去的时候，是在数据工程团队。
And and actually, my first role there was on the data engineering team.
那个团队负责开始收集、接入和清理我们当时正在获取的数据。
So that team was responsible for starting to collect and ingest and clean up the data that we were collecting at the time.
我2007年加入的，虽然背景是自然语言处理，但那时候更多是作为一个通用工程师在做事。
I joined in 2007, and I had my background in natural language processing, but I was more just really being an engineer more broadly at the time.
做着做着，我开始留意到Two Sigma手上收集的那些数据集，真的特别有意思。
And as I did it, I started to look around, and I noticed all these different data sets of Two Sigma that they collected that were really, really interesting.
有一次，我偶然看到一组数据，做了点分析，就特别好奇。
And at one point, I came across some data that had been collected, then I did this analysis, and I was curious.
那是一种对不同新闻媒体的政治倾向分析，看它们是偏向哪一边。
It was sort of a political analysis of different news places and whether they were leaning one way or the other.
我把这个结果发在公司内部论坛上，结果有人给我打电话，说：嘿，我很喜欢你做的这个。
And I posted this in an internal message board, and I got this call from someone at the company who said, hey, I like what you did.
也许不是所有东西都适合发给所有人，但你还能做点什么别的吗？
Maybe not the best to send everyone everything like this, but what else can you do?
后来我发现，那是Sigma那边建模团队的人，从那以后，我的工作逐渐转向了新闻领域，从新闻数据开始，后来扩展到更广泛的文本分析，一干就是十多年，专门研究文本如何以各种方式预测市场走势。
And it turned out that was somebody over in modeling at Sigma, and from there, I kind of shifted my work towards what began in the news space, news data, and eventually the textual space more broadly, and spent a good ten plus years just studying how text predicts markets in various ways.
所以，这就是我从研究生阶段，到做工程师，再到如今专注建模工作的整个过程——我们的目标就是尽可能利用所有数据，预测我们交易的标的在未来不同时间周期内的涨跌。
So that's kind of how I went from grad school to an engineering role and then eventually to a more modeling focused role where our goal is to predict using as much data as we can how the things that we trade are gonna move.
它们的价值是会上涨还是下跌？
Are they gonna go up or down in value over different time horizons?
这整个过程，是我职业生涯中很重要的一部分。
And that's a big part of the journey.
所以你现在的头衔是复杂特征引擎团队负责人，而‘特征’这个词，正好就在里面，这个词在量化投资领域可是经常被提到。
So your title today is head of complex feature engines and right in there is that word feature and that word gets thrown around a lot in the world of quantitative investing.
但据我跟很多人聊下来，每个人对这个词的理解都不太一样。
But at least anecdotally when I talk to many people, it means different things to different people.
所以我想先听听你的看法。
And so I would love to maybe just start with you.
你觉得‘特征’这个词到底是什么意思？
What does the phrase feature mean?
Two Sigma 公司是怎么使用这个词的？
How does Two Sigma really use that word?
如果你能举一些例子，从最常见的到比较冷门的，都涵盖一下，那就更好了。
And maybe if you can give us some examples that span the spectrum from the very familiar perhaps to the very alternative.
关于‘特征’这个词的定义，我们内部已经争论过很多次了。
We've had many a debate over the definition of the word feature.
这可是我们工作的核心。
It's central to what we do.
虽然大家看法不一，但我先从例子说起吧。
And I think we've landed in different areas, but let me start with the examples.
与其下定义，不如举个例子，我觉得这样更清楚。
It's easier to give an example than a definition, I think.
比如，一家公司过去一周的股价变动，就是个例子。
An example is something like the change in stock price over the last week of a company.
对吧？
Right?
这是关于我们交易对象的一个具有经济意义的事实。
It's a fact about something we trade that's economically meaningful.
至少在Two Sigma，‘经济意义’这部分很重要，但在其他行业，特征的定义可能不包含这一点。
At least at Two Sigma, the economically part is important, but you could have features in other industries where that's not part of the definition.
所以，我们其实是在问：世界上有哪些事实是值得被记录下来的？
And so really, we look at it as saying, okay, what is a fact about the world that is worth itself noting?
比如，某只股票上涨了5%，这就是关于股价的一个特征，对吧？
And so the fact that a stock went up 5% is a feature about the stock's price, right?
股价本身只是原始数据，它时涨时跌，但我们可以从中提取不同的特征，比如：它从去年以来表现如何？
The stock's price is sort of a raw piece of data, it goes up and down, but you can pull different features, well, what did it do since last year?
具体发生了什么？
How what happened?
变动了多少？
How much did it move?
波动率是多少？
What was the volatility?
这些都是你常听到的更技术性的常见特征。
Those are more technical common features that you might hear about.
但特征也可以更有趣，比如，最近有分析师上调了这只股票的评级吗？
But then features can get more interesting like, did an analyst recently upgrade the stock?
这挺有意思的，从不同角度看，它可能是关于股票的特征，也可能是关于分析师的特征。
It's kind of interesting, it's a feature about a stock, or a feature about an analyst, depending how you look at it.
再往下，会变得更具体，就我从事文本数据的经验来看，某家公司最近的新闻曝光量相比历史水平如何？
And then it can get progressively more specific, from my background in textual data, how much news coverage is there about a particular company compared to the history?
这显然是关于这家公司的特征，对吧？
That's a feature about that company, right?
我们发现它的增长在加快，但新闻报道却在减少。
We're understanding that it's growing and its news coverage are reducing.
或者他们是什么时候提交这份文件的？
Or when did they release this filing?
或者新闻稿是在星期几发布的？
Or what day of the week did the press release come out?
是在周五盘后发布的，还是更早的时候发布的？
Did it come out on a Friday after hours, or did it come on a earlier time?
这其实是一种方式，把那些源于直觉的假设提炼成具体指标——比如，周五下午发布本身就值得作为一个特征来研究。
So it's really a way to encapsulate these ideas that are usually driven by some hypothesis that's innate, like, hey, the fact that it might come out on a Friday afternoon is itself interesting enough that I'm gonna make a feature out of it.
把想法做成特征的好处是，你可以专门分析这个特征，而不是把所有新闻稿堆成一堆，扔给电脑，指望它自己找出答案。
The nice thing about making a feature is that you can then study that feature instead of just saying, I'm gonna study all press releases in a giant pile and throw it at a computer and see what comes out.
通过加入人的洞察和经验，我能逐步分辨出哪些信息才是真正重要的。
By layering human intelligence and intuition over it, I can start to pick apart what's meaningful in there.
我是在帮电脑减负：我知道周五下午特别值得关注，发布时间的长短也可能有影响，再加上我对公司运作方式的了解，所以我没必要让电脑去分析公司所有历史数据，那要求太高了。
And I'm kind of helping a computer out by saying, no, no, Friday afternoons are really interesting, as is the length maybe, as is I can kinda use my knowledge of how companies operate, so I'm not asking a computer to check everything ever about a company at all times, which is a very large ask.
通过筛选并结合人类的知识层，我们可以让这项工作变得可行——指出哪些内容有趣、哪些值得关注、哪些可能影响市场。
By filtering down and using human layers of knowledge, we can kind of make that job tractable by saying, no, here's some cool stuff, and here's what's fascinating, and here's what might move markets.
最终，这些每一点都构成了一个特征。
And in the end, each of those things is a feature.
我们相信，你不能只是把一堆原始数据扔进机器学习算法，就指望它能预测未来。
And we believe that instead of just, I mean, you can't just throw a bunch of raw data into some machine learning algorithm and tell it to predict the future.
这确实是一种方法，但我们坚信，把数据转化为特征这一中间层非常有价值，以此作为起点，才能更好地预测接下来会发生什么。
It's definitely a way to approach things, but we really believe there's value in this feature layer of converting data into features, and then using those features as our predicting starting point to understand what's gonna happen next.
一两年后，我们得重新审视那个关于周五下午的特征例子，看看它在2024到2027年的市场环境中是否还站得住脚。
In a year or two, we're gonna have to revisit that feature example about Friday afternoons and see if it survives a 20 fourseven market environment.
没错。
That's true.
随着情况变化，这一点确实很重要。
As things start to change, that's a good point.
在量化股票领域，一个重大的演变和趋势就是‘作坊式’模式，而Two Sigma则采用的是我们所说的共享平台模式。
So in the world of quant equity, one of the big evolutions and big adoptions has been the pod shop model and Two Sigma in contrast to that operates on, I guess we would call a shared platform model.
我们只有一个核心投资组合，而不是一堆相互竞争的小团队。
It's one core portfolio rather than a series of competing pods.
这种结构上的选择，如何影响你和团队每天的关注点和优化方向？你觉得这和在小团队模式下工作的研究员关注的重点有什么不同？
How does that structural choice shape what you and your team focus on and optimize for day to day And how do you think that differs from what a researcher at a pod shop would be focused on and optimizing for?
我先说明一下，我们并不是真的只有一个投资组合。
I'll first say that it's not that we have literally a single portfolio.
我们有很多产品和投资组合，但作为Two Sigma的建模人员，我们在寻找阿尔法时，并不会针对某个具体的投资组合来看问题。
We have many products and portfolios, but as modelers at Two Sigma who are trying to find alpha, we are not looking at it through the lens of any portfolio specifically.
对吧？
Right?
我们的工作是做出这些预测，但最终怎么交易这些投资组合，并不由我们决定。
Our job is to kinda make these predictions, and we don't control ultimately how these portfolios trade.
我们有专门的团队在研究，比如未来一周IBM股价会怎么走，NVIDIA又会怎样，每个人从不同的角度切入，但他们之间并不竞争，因为最终这些预测都会汇总到公司更大的投资组合里，整合起来做决策，所有人的成果都会被综合考虑。
So we have teams who are out there trying to say, okay, here's what's gonna happen to the IBM price over the next week, here's what's gonna happen to NVIDIA, And each coming from a different angle, and they're not competing against each other because in the end of the day, all of these predictions are being centralized on some of the firm's larger portfolios and being aggregated, and decisions are gonna be made based on the combination of all these individuals.
所以这其实带来了很多有意思的机会，对吧？
So that creates a lot of interesting opportunity, right?
如果各个小组彼此竞争，那确实会有理由让你不想跟其他组共享技术，或者只在本地投入资源，因为你觉得自己是在跟别的团队抢胜出的机会。
In a world where you have pods competing, there is reason why you might say, have a technology that you don't wanna share with another group, or invest truly locally because you feel like you're trying to beat out another team.
但在Two Sigma，我们想的是怎么从全局角度优化，而不是只盯着局部优化。
At Two Sigma, we're thinking, how can we globally optimize instead of locally optimize?
所以我们希望把我们做得最好的那些成果提升到一个共享平台上，让其他人也能用上。
So we want to elevate some of the best stuff we do up to a shared platform so others can use it.
比如我们之前提到过的一些特征，特别有意思，我就干脆公开出来，让其他团队去研究。
And so that might be a set of features like we discussed earlier that are so interesting, I'm gonna put it out there for other teams to study.
又或者有人发现了一个特别巧妙的机器学习算法，他们会围绕它开发一个工具，然后分享给所有人使用。
Or it might be a really clever machine learning algorithm that somebody discovered, and they're gonna build a tool around it and share it for others to use.
这样一来，我们的平台就能以一种非常惊人的方式扩展，因为有太多人从全局出发，共同让整个平台变得更好，而不是仅仅有人在使用一个现成的平台，从中提取资源。
And this really allows our platform to scale in ways that is really remarkable because there's so many people contributing kinda globally to make the overall platform better, as opposed to just a platform that exists that others kind of attach to and pull from.
我觉得在普通的小组制公司里，通常只有一个中心平台，大家主要就是从中取用。
I think in a normal pod shop, there's a centralized platform that's mostly pulled from.
团队只是在使用这个平台，而在这里，这是一个循环的过程。
It's the team utilizing the platform, and here it's cyclical.
这个平台是一个不断迭代的过程，我们的团队一直在把好的成果往上推，供大家共享，因为我们所有人都一心想着让Two Sigma做出最精准的预测。
The platform is an iterative process, and our teams are constantly pushing up into the platform for others to share because we all are aligned in making Two Sigma make the best predictions possible.
也许这么说有点简化了现实，但当我观察量化流程时，我觉得alpha主要存在于三个独立的领域里。
At the risk of perhaps oversimplifying the world, when I look at the quant process, I sort of see three main silos that alpha can really live in.
第一个是数据本身，对吧？
The first is in the data itself, right?
比如你能拿到一些别人无法获取的另类数据。
This might be getting your hands on alternative data that nobody else has.
第二个是特征生成。
The second is feature generation.
可能大家用的数据都一样，但你能把它转化成独特的观察结果，而这种转化方式是别人做不到的，却能带来深刻洞察。
It might be that people have the same data, but you're able to transform it into observations or unique observations in a way that nobody else has that can be insightful.
第三个则是预测层本身——也许大家的数据和特征都一样，但你能用一种新的方法、新的机器学习算法，从这些特征中挖掘出更多预测性信息，从而获得优势。
And the third would be sort of the forecasting layer itself that everyone maybe has the same data and the same features, but you can apply some new method, some new machine learning algorithm that extracts more information from those features in a predictive way that gives you an edge.
以你现在的视角来看，你觉得这三者中，alpha主要集中在哪个部分？过去五到十年，你的看法有变化吗？
From where you're sitting today, curious how you would weight where alpha lives in those three buckets and has your answer changed over the last five or ten years?
过去十年里，其实有点变化。
Probably over ten years, it's changed a little.
我觉得第一个方面，数据获取，确实是我们工作里很重要的一部分。
I think the first bucket, data access, I mean, it is an important part of what we do.
但要是说这十几年到二十年间，这部分没被标准化，那就太天真了。我刚入行的时候，比如做新闻领域的建模，或者跟大多数供应商打交道，他们会问：‘你们想查哪家公司的数据？’
It would be naive not to realize that some of that has been commoditized over the last fifteen to twenty years, right, that when I started doing, for example, modeling within the news space, or really across most vendors, I mean, you would reach out and they would say things like, well, which company do you want data on?
我们就说：‘所有公司。’
And we'd say things like, all of them.
他们就问：‘什么叫所有公司？’
And they'd say, what do you mean all of them?
我们就说：‘我们想追踪所有公司。’
And you'd say, we wanna track all the companies.
他们觉得这根本说不通。
I said, does that doesn't make any sense.
所以，我们确实是最早一批跟这些供应商合作、帮他们重新思考数据打包方式的人。
And so clearly, were the first people interacting with many of these vendors thinking about a different way to package their data, teaching them.
所以在那些早期，我觉得数据获取所占的比重比现在大得多，不过它依然很重要。
And so in those early days, I would say data access as a relative weight was probably larger than it is today, though it's still important.
我觉得随着越来越多的数据变得商品化，这对我们的独有数据造成了压力。
I just think that with more commoditization, it puts pressure on that proprietary data piece.
我们依然拥有大量独有数据，因为我们有很长的历史，这是别人没法比的，非常丰富。
We still have a lot of proprietary data, we have a long history that others might not have, which is very rich.
我们会从用户群体中收集自己的交易思路之类的东西。
We collect our own trading ideas from crowds and things like that.
所以我们确实有很多独有数据。
So we do have a lot of proprietary data.
我不是在低估它的重要性。
I'm not underselling it.
我只是觉得它在整体中的相对比例可能下降了。
I just think its proportional weight has probably dropped.
你提到的第二部分，就是你如何利用这些数据来构建功能和想法。
Your second part, which was what you do with it in terms of building features and ideas.
我的意思是，这真的特别关键。
I mean, I think that is just huge.
我认为，能够以独特而巧妙的方式思考如何利用数据、提出极具创意的应用点子，这本身就是巨大的优势。
I think that the ability to think about in a clever and unique way how to take data and come up with really clever ideas of how to use it is just a huge edge.
我敢这么肯定，是因为我亲眼看到我们是怎么打造出那些具有预测能力的成功模型的。
And I could say that with confidence because I see how we build successful models that are predictive.
正是那种创造性的过程，让我们忍不住说：‘这想法真绝了。’
And it's the creative process that we say, oh, that's really clever.
这推动了我们很多工作。
That drives a lot of what we do.
所以我觉得，能带来超额收益的很大一部分，正是人类的智慧——把一堆数据转化成有趣、有意义、巧妙，而且可能是别人还没想到的特征。
And so that suggests to me that a good part of what's creating this alpha is certainly the human ingenuity layer of how to convert piles of bits into features that are interesting, meaningful, clever, and maybe that you think others haven't thought of.
我给你举个例子，让你感受一下。
I'll give you just a flavor of that.
比如，我常会提到一个经典的分析师案例。
For example, I like to talk about a classic one of an analyst.
你可能会说，好吧，分析师确实会做些判断。
And you might say, okay, an analyst is making some decisions.
网上有很多论文讨论分析师的目标价是否影响市场，或者他们上调或下调评级的影响。
And there's a lot of papers out there and others about whether an analyst target price moves markets or they upgrade or downgrade.
但你能从多少个不同角度来研究分析师呢？
But how many different ways can you study an analyst?
一开始，你可能会觉得，这也没多少可研究的吧？
And at first, it's like, well, it's not that
数据确实不多。
much data.
对吧？
Right?
但当你开始想，这位分析师到底住在哪里？
And then when you start to think about, well, where does that analyst live?
他所在市场和他覆盖的公司市场之间有什么关系？
And do they live, you know, what market are they in compared to who they're covering?
他们在这个职位上待了多久？
And how long have they been in this position?
他们在分析一家公司的时候，会不会和这家公司的CEO毕业于同一所学校？
And do they go to the same school as the CEO of a company when they're covering them?
这些都是一些细小的点，但如果你能提出两百个这样的问题，创造出两百个新特征，那就会形成一份很长的清单，全是这类初步的发现。
I mean, these are small points, but in aggregate, if you can ask 200 of those questions, and you've created 200 new features, it's that long list of just kind of intro oh, yeah.
这挺有意思的。
Oh, that's cool.
我之前还真没想过这一点。
I hadn't thought of that.
这份清单看似冗长，但综合起来却能产生显著的影响。
It's a long list that in aggregate starts to make a difference.
因为现在我提到的每一点，单独来看，影响其实微乎其微，大概只有50.001%的几率吧。
Because any one of those things I'm talking about now, it's not it's like, you're right, 50.001% of the time, right?
这些行为中的信号确实非常微弱，但正是所有这些想法的累积，才真正带来了巨大的改变。
I mean, these are tiny, tiny signals in the way we act, but it's the aggregation of all that interesting ideas that I think really makes a huge difference.
这就是为什么我们在Two Sigma设立了围绕功能的团队。
And that's why we have these teams at Two Sigma that are centered around features.
这可以说是我们的核心使命。
It's like it's the core mission.
你第三个问题提到的是机器学习算法和其中的科学原理。
Your third part of your question was these machine learning algorithms and the science there.
这部分同样非常重要。
That is also a very big part of it.
我们有一个庞大的团队叫‘技术组’，专门研究这些算法本身，以及如何最大化地利用它们。
We have a large team called Techniques that really focuses on those algorithms themselves and how to make the most out of them.
所以，你有这些聚焦功能的团队，他们在思考：我能用哪些巧妙的数据？
And so you kinda have these feature focused teams where they're thinking about, alright, what are all the cool, clever data I can make?
还有技术团队，他们在想：有了这些好数据，我能用哪些厉害的算法来让它发挥作用？
And you have these techniques teams who are like, okay, given cool data, what are all the cool algorithms I can put on it to make things work?
正是这两个领域的交叉点，让我们看到了令人惊叹的成果。
And it's at the intersectionality of those two areas where we see just amazing things happening.
不同领域、不同背景的专家聚在一起做预测，这种多元的思维方式，我觉得真正提升了我们的工作效果。
It's the experts from different areas with different interests coming together to make predictions, and that diversity of approach, I think, really strengthens what we do.
我想聊聊功能层面的事，因为你刚才举了那个分析师的例子，问他们是不是和CVO上同一所学校？
I wanna talk about the feature side a bit because, you know, you gave this example of the analyst and did they go to the same school as the CVO?
他们是不是住在自己所覆盖的市场里？
Do they live in the market that they cover?
但既然你们在努力挖掘各种独特有趣的特征，我想象得到，可能会有人坐在那儿天马行空地想出一堆特征，我猜其中大部分其实没什么意义。
But given that you're trying to come up with all these unique and interesting features, I can imagine a world where well, can sit and ideate a whole bunch of features that I would presume probably are meaningless.
比如分析师有多高？
Like how tall is the analyst?
分析师的眼睛是什么颜色？
What color are the analyst's eyes?
再比如，分析师和CEO是不是同名？
You know, like do they and the CEO have the same name?
要么样本量太小，要么根本就无关紧要。
Like it's either N is too small or it's just frankly probably irrelevant.
你能给我讲讲，实际操作中，你们是怎么从这些高度同质化的数据里构思出这些特征的吗？
Can you walk me a little bit through what the process actually looks like in practice in ideating these features coming from largely commoditized data sets?
你们团队到底是怎么从人人都能拿到的数据里，挖出别人没发现的东西的？
Like how is your team really taking a piece of widely available data and pulling something out of it that the rest of the market hasn't?
我觉得坚持很重要。
I think persistence is a part of it.
你说得对，眼睛颜色大概先验概率很低，也没什么意思，对吧？
You're right that probably eye color is going to have a prior that's pretty low and probably uninteresting, right?
我不会花太多时间在这上面。
I'm not gonna spend much time on that.
事实上，我可能会直接避开这种指标，因为它只会增加噪音，对吧？
And in fact, I'd probably avoid it as an indicator because I think it would add noise, right?
所以这里面其实有人为判断的成分。
And so there is a human layer of what's possible there.
我们的目标不是随便想出什么特征都往上堆。
Our goal is not to just come up with any feature ever.
如果从人类的角度来看，比如一个分析师，人类身上其实有很多可以挖掘的特征，对吧？
Come up of a human, if we're talking about an analyst, there's a lot of features about humans, right?
但关键是，哪些特征会让大家一听就恍然大悟：‘哦，原来是这样’？
But rather, which ones do everyone's like, oh, yeah.
这些想法的来源其实五花八门，对吧？
Where these ideas come from is varied, right?
当然，有时候一些学术研究也能带来启发。
I mean, obviously, there's some academic work, which is inspirational at times.
我之前看过一篇论文，挺奇怪的，他们居然分析了分析师发的帖子，试图判断他们有没有孩子。
I actually saw a paper where they, oddly enough, looked at this academic paper actually took analysts and mapped them to posts to try to determine whether they had children or not.
然后他们把这个当成了一个指标，不过我不会告诉你这个指标是正向还是负向的，因为我有孩子，不方便说。
And then they used that as an indicator, and I won't tell you which direction it goes because I have children, so I'm not gonna tell you.
重点是，有没有孩子会不会影响一个人预测股价的能力。
But whether or not having kids affects your ability to predict stock prices.
这个例子就说明了，这类研究能走得多远。
And so that's an example that the academic did that just makes you understand how far these things can go.
我们会停下来认真思考，用多元的思维方式来应对这些问题。
And we really stop and think, and we approach it by diversity of thought.
比如，不同背景的人会带着各自的角度来解决这些难题。
So, like, people from different backgrounds come to these problems.
物理学家、数学家、计算机科学家，他们各自带着不同的视角过来。
You know, the physicist comes with the mathematician, comes with the computer scientist.
正因为背景不同，他们提出问题和解决问题的方式也各不相同。
Because they come from different backgrounds, they come up with different ways to ask these questions and solve these questions.
这正是其中很重要的一部分。
And so that's part of it.
我觉得我们还会关注现实世界中发生的事。
I think also we see things happen in the world.
比如某地发生了火车脱轨，有文章说某只股票因此下跌，你就会觉得，这挺有意思。
There's a train derailment somewhere, and there's some article written in that says, hey, this stock is down off this, and you say, that's interesting.
那我要怎么把这种现象推广开来呢？
How do I generalize that?
我该怎么把这一个发生在现实中的例子拿过来分析呢？
How do I take this one instance that's happened in the world?
每当有公司出事，我们都会想，这种情况能推广到别处吗？
And every time something happens to a company somewhere, we're kinda like, is that generalizable?
我们能不能把这个案例拿出来，想想该怎么系统地总结，然后回溯十年，看看它对所有公司、或者至少对整个行业意味着什么？
Can we take this instance and think about how it how you would write it up and play it back over a decade to all companies, or at least all companies in an industry or something like that?
所以我们一直在寻找更巧妙的方法来构建这些分析维度。
So we're kinda constantly on the lookout for just clever ways to build out these features.
一旦找到了，我们可不想把所有东西一股脑扔进一个巨型算法里，跑完就说‘搞定’了，对吧？
And once you do, what we don't wanna do is just throw everything in some giant algorithm, run it, say go when we're done, right?
我的意思是，我们本来就有自己的预判。
I mean, are we have priors.
比如，你默认火车相撞是坏事，这就是你的预判，但当你深入研究后，发现结果恰恰相反。
So if your prior is that something that a train collision is a bad thing, that's your prior, and then you go and study it and it's the opposite sign.
这显然不太对劲，对吧？
I mean, it's not great, right?
这让你觉得，哦，也许是因为人们过度反应了。
It kind of breaks with your you feel like, oh, well, maybe it's because people over.
是的。
Yeah.
事后你总能找到理由来解释任何事情。
You come out you can always post hoc rationalize anything.
但如果结果和你的预期不符，那才正是科学方法的体现。
But if it doesn't meet your priors, that's sort of the scientific method.
你提出假设，然后去验证它。
You state your hypothesis, you test it.
如果结果和你预想的相反，你不能说：‘我其实想说的是相反的方向。’
And if it doesn't go the way you said, you can't just be like, well, I meant the opposite direction.
对吧？
Right?
所以我们用科学的方式来看待这些问题，而且在可能的情况下，必须验证我们的预设。
So we're approaching things scientifically that way, and we have to verify our priors when we can.
这也是过程的一部分。
And that's part of the process.
所以我们有清晰的特征、明确的假设、规范的使用方式，最终能得到一组能用来做预测的数据。
So we have clean features, clean hypotheses, clean usage, and ultimately, we feel like that ends up with a set of data that we can use to be predictive.
我注意到一点，如果一个团队一直在不停地寻找新特征，对吧？
One of the things that strikes me here is if you have a team that is just constantly searching for features, right?
他们确实可能发现一些新颖有趣的特征，但这些特征可能和别的特征高度相关，表面上看不出来，对吧？
There's the risk that they do identify novel interesting features, but those features may have a high degree of collinearity with other features that it may not seem like they do, right?
于是你们会生成成百上千个单独看都挺有意思的特征，但实际上让组合构建的过程变得复杂多了，因为出现了多重共线性的问题。
And so you're generating hundreds, if not thousands of features that are independently interesting on their own, but actually make sort of the portfolio construction process far more complicated because you get this collinearity issue.
我想知道，当你们面对这种不断膨胀的特征库——包括那些被淘汰的和真正用上的——是怎么处理这个问题的？
I'm curious how you deal with that as you get this expanding both potentially graveyard of features as well as like actual features you're putting into implementation.
谁知道呢，也许分析师的眼睛颜色和天气之间有很强的相关性，谁知道呢？
I mean who knows, maybe the color of an analyst's eyes correlates very highly to, I don't know, the weather of who knows?
我只是在想两个完全不相关的东西，现实中它们是独立被发现的，但预测结果却高度相关。
Like I'm just I'm trying to think of two very random things that actually in practice, like they came up with independently, but are highly collinear in what they predict.
是的，我们每天的核心工作就是寻找相互独立且新颖的预测方向。
Yeah, so our job is at its core every day to come up with orthogonal and novel predictions.
要找出和过去做法有关联的东西，其实挺容易的。
It's kinda easy to come up with things that are correlated to what we've done in the past.
那些都是最容易拿到的成果，对吧？
And that's the lowest hanging fruit, right?
所以，作为阿尔法建模者，我们每天醒来都在主动寻找独立性。
And so almost by definition, we wake up every day as an alpha modeler searching for orthogonality.
共线性，就是信号的杀手。
So collinearity is the death of a signal.
对吧？
Right?
如果你发现某个东西，别人用不同的数据集也在交易，那我举个例子，比如用新闻数据交易。
If you found something that someone else is trading through a different dataset I mean, I give a class example of trading news data.
你可能会说，好吧，我要做新闻情绪交易，搞个新的情绪模型。
Now you could say, okay, I'm gonna trade news sentiment, and I'm gonna, you know, say, okay, this is my new sentiment model.
但如果这其实只是捕捉了财报发布后的新闻呢？
But what if it is just picking up post earnings stories?
所以你所谓的‘新情绪’，不过是追着财报数字跑罢了。
And so what you call new sentiment is just chasing earnings numbers.
你坐在那儿想着，我找到了一个全新的数据源，结果它不过是财报数字的粗糙翻版。
And so you're sitting there saying, I got this novel data source, and all it is is a dirty version of earnings numbers.
你可能会很兴奋，觉得自己发现了什么巧妙的东西，但等你用工具跑一遍、分析完相关性结构——这本来就是工作的一部分。
You might get excited and think you found something clever, but by the time you've run through the tools, you've explored correlation structures, which is just a part of the job.
如果你连这都做不到，那根本撑不下去。
If you can't do that, then you're not gonna make ends.
所以我们有一些明确的规则和门槛，只有当新的阿尔法信号具备足够独立性时，我们才会觉得可靠，因为我们根本不相信，把一堆和过去数据高度相关的玩意儿堆在一起能有什么用。
So we have certain rules and thresholds where we are comfortable only when new alpha signals have a certain amount of orthogonality, because we just don't believe that throwing a bunch of things in that are equally correlated to things in the past is useful.
所以没错，这其实是我们的核心所在。
So yeah, it's kind of core to what we do.
你被独立性驱动，被避免相关性驱动，目标就是找到下一个和我们以往信号关联性没那么强的新信号。
It's like you're driven by orthogonality, you're driven to avoid correlation, and your goal is to come up with the next signal that's not so correlated with our previous ones.
这其实是你起步的地方。
And so it's kind of where you start.
这甚至不只是事后才考虑的事。
It's not even just an after the fact thing.
分析师在预测收益时，得多大程度上意识到他们所针对的时间周期？又该怎么据此判断哪些特征是相关的？
How aware do the analysts have to be about the horizon over which they're trying to ultimately forecast returns and sort of back out which features would be relevant to that horizon?
我猜有些特征对高频交易特别有用，有些则适合超低频或中频场景。
I can imagine there are some features that are highly relevant to the, you know, the high frequency space or the super slow frequency space, the mid frequency space.
这些特征肯定是不一样的。
Those are gonna be different features.
他们在构思这些新特征时，心里会多大程度上考虑时间周期的问题？
How much is that in the back of their mind when they're thinking about these novel features?
这非常关键。
It's a big part of it.
对吧？
Right?
如果你的假设是研究顾客去商店的情况，那你大概会认为，比如这周很多人去了沃尔玛，这种行为是有明显迹象的。
If your hypothesis is that you're studying, I don't know, people visiting stores, you probably think that and you have a good indication that people have been visiting the stores a lot this week or whatever for going to Walmart.
但你不会觉得这种信号能在接下来五秒内就被市场定价。
You probably don't think that that's a signal that's gonna be priced in in the next five seconds.
对吧？
Right?
这种信号得慢慢渗透进去。
It's probably it's gotta make its way through.
它最终得体现在同店销售额的报告里，而那还得好几个月后。
It's got to eventually show up on some same store sales release, and it's gonna be months away.
所以不同的信号，背后都对应着不同的假设，关于它们会以怎样的路径影响市场。
And so different signals have different hypotheses of how they're gonna find their way in.
因此，建模者确实会考虑这一点，他们研究时往往本身就带着对时间周期的意识。
And so, yeah, modelers are definitely thinking about that and aware it's kind of built into what they're studying often is what horizon.
我想先聊聊这个例子，再转向更广泛的人群讨论，但这个同店销售额的数据案例特别有意思——你可能会想，当这种数据刚出来、还没人用的时候，它预测的周期会比较长；但随着越来越多的人拿到数据并快速反应，市场对它的反应速度反而会越来越快，对吧？
I wanna talk about that example before I move on to talking about just people in general, but that same store sales data example or feature is really interesting to me because you would imagine when that data maybe first comes out and there's not a lot of people using that data, the horizon over which it forecasts may be slower and then as more and more people get access to that data and act on it faster, you could actually imagine it getting the market moving on it much much faster, right?
这其实是同一个特征，但预测的时间窗口会不断缩短，直到变得太高效，反而失去了意义。
And it's sort of the same feature but the forecast horizon continues to speed up until maybe the point it becomes so efficient, it is no longer relevant.
在你看来，这种特征的生命周期是怎么演变的？
How does that play through in like the life cycle of features for you?
这其实就是特征的生命周期本身。
I mean, it is the life cycle of features.
你的目标是创造出足够新颖的东西，让它不会受到这种压力的影响。
Your goal is to create something so creative that it doesn't come under that pressure.
但这种情况并不总是能实现。
That's not always gonna happen.
总会有一些其他公司去研究这些数据，从而施加压力。
There are going to be other firms that study things, and so they're gonna put pressure on.
没错，你描述的这种现象确实存在。
And so, yeah, you do see exactly what you described.
你会看到这些信号的反应速度越来越快。
You see things speed up.
有时候甚至到信号完全失效的地步。
And sometimes to the point where, yeah, the signal's no longer viable.
我们都在尽力发挥创造力，避免陷入这种压力。
We're all trying to be as creative as possible so that we don't get that pressure.
归根结底，我们有监控系统，用来跟踪各个模型的实际表现。
And in the end of the day, we have monitoring in place that kind of tracks what's happening with a particular model.
在公司层面和投资组合层面，团队会持续观察这些情况，根据自己的分析决定对每个信号投入多少交易力度。
And at the firm level, the portfolio level, is watching that and deciding how hard to trade any signal based on their own analysis.
对吧？
Right?
作为建模人员，我们主要负责做出这些预测。
We as modelers are kind of making these predictions.
而投资组合那边——那是另一个量化团队的职责——他们会专门监控这类问题，寻找收益衰减的迹象，并据此调整权重。
The portfolio itself, those that are on the portfolio, which is a whole another quant function, are going to be monitoring for that type of thing and they're going to be setting looking for alpha decay and then appropriately setting weights to match that.
我想聊聊人这一块，因为你描述的这份工作听起来非常有创造性。但人们一提到量化或者招聘量化人才，通常只会想到硬核的技术能力：数学、统计、编程这些。
I wanna talk about the people side here because what you're talking about sounds like a very creative role and I think when people talk about quant or hiring for quant, they typically think about the hard technical skills, the math, the statistics, the coding.
你提到的很多特征都需要一定的巧思和创造性思维，在一定意义上还需要发散性思维。
A lot of the features you're talking about require a degree of ingenuity and creative thinking and you know, lateral thinking to a certain degree.
你们是怎么找到这类人的呢？
How do you go about finding those people?
你们是怎么组建团队，以及如何筛选候选人，来判断他们是否擅长发现新颖且有用的特征的？
How do you go about assembling this team and screening candidates to determine whether they're going to be good at identifying novel and useful features.
我觉得像我们这样的团队，在面试过程中寻找这类人才时，通常会进行一些开放式对话。比如我们会说，嘿，这里有一些数据。
I think for a team like mine that looks for features during an interview process, right, you have conversations that are open ended and you ask, hey, here's some data.
假设你能拿到社交媒体上的海量数据，告诉我你能预测出哪些有趣的东西。
Here's some if you had access to a fire hose from social media, tell me all the cool things you could predict.
这些问题非常开放，不是让你解数学题，而是没有标准答案的那种。
Very open ended questions that are not solve this math equation, but rather things that don't have a right or wrong answer.
我们会一起探讨，关键是看他们能想出多少巧妙的点子，以及有多少是他们独创的、别人没听过的想法。
And we walk through, and and the question is, how many clever ways and how many things that I've never heard anyone else say are they saying?
如果他们在过程中能提出新颖有趣的观点，这就是一个很好的指标。
That's a good indicator if somebody is coming up with just interesting things that are novel during that process.
我觉得这是个不错的方法。
I think it's a good thing.
说到这个，让我想起当年读研的时候，我去微软面试，记得当时有个职位是产品测试员，我确实面过这个岗位。
Mean, reminds me, look, frankly, back in when I was in graduate school, I I was looking at an interview at Microsoft, And I remember they I guess they had these product testers was one of the roles that I remember being interviewed for at some point.
他们问我：假设你面前有一台自动售货机，你的任务是测试它。
And the question to me was just like, you have a vending machine, your goal is to test it.
你会怎么测试这台自动售货机？
Like, how would you test the vending machine?
听起来可能有点傻，但其实根本没有标准答案。
Seems like a silly question, but, like, there's no right or wrong answer.
我就开始聊起来，说比如，如果你把手伸进去，它会不会夹断你的手？
And I just started talking for, you know, however long about, you know, don't if you stick your hand in, does it cut off your hand?
如果拔掉电源会怎么样？
What happens if you unplug it?
要是往里面塞张纸，会不会卡住？
Like, can you put a piece of paper in?
你能把硬币挂在绳子上吗？
Can you put a coin on a on a rope?
如果你摇一摇呢？
What if you shake it?
我能一直说下去，没完没了。
And I can go on and on and on.
你懂的。
You know?
我思考问题就是这样的。
It's kind of that's the way I think.
我总是会想，接下来会怎样？
I'm always like, oh, what's next?
接下来会怎样？
What's next?
我们想找的就是这种人——喜欢思考、解决问题，能持续不断地提出下一个问题。
So we're looking for that kind of spark of people who like to think and solve problems in sort of a continuous basis and can always find the next question in a series of questions for a particular problem.
所以我也说不准。
And so that I don't know.
那个面试问题，虽然我觉得它并不适用于类似的角色，我也从来没问过这个问题，但它那种开放的思路很有启发性，而且显然已经流传很久了——二十多年前就有人问我过。
That interview question, although I don't think it goes for anything like the same role, and I've never asked that specific question, it is inspirational in its openness and it's obviously been around a while since somebody asked it to me twenty years ago.
咱们转到人工智能吧，我觉得这会是接下来谈话的主旋律。
Let's pivot to AI, which is I think where we're gonna spend the majority of the rest of this conversation.
在我们之前的通话里，你跟我说过一句话，让我印象特别深，你原话是：自然语言处理就是把语言变成数字。
In our pre call, you said something to me that really stuck with me, which was, and I'm quoting here, NLP was about turning language into numbers.
而有了大语言模型，我可以把数字变回语言。
With LLMs, I can turn numbers into language.
现在任何东西都能变成语言了。
Anything can be language now.
你能详细说说你这句话是什么意思吗？
I'd love you to just unpack what you meant by that.
这种转变在实际应用中意味着什么？又会带来哪些以前三到五年根本做不到的新功能？
What does that shift mean practically and maybe what does that mean for the types of features that can be built today that couldn't be done three to five years ago?
对于像Two Sigma这样已经用文本数据交易了十五年以上的公司来说，能获得更多文本数据真的让人兴奋。
I think for a firm like Two Sigma that's been trading off of textual data for over fifteen years, getting more textual data is pretty exciting.
当然，我们一直在收集各种关于世界的文本信息，但能收集到的东西终究是有限的。
Well, obviously, we've been collecting all sorts of text about the world, there is a limited amount of things that you can collect out there.
对吧？
Right?
互联网有它的边界，供应商也有他们的限制。
The Internet has its bounds, vendors have their bounds.
你只能分析那些你能接触到的数据。
And so you can only analyze what you can get your hands on.
我觉得语言模型真正令人兴奋的地方在于，它们能生成文本数据。
And I think what's really exciting about LMs is that they generate textual data.
所以，什么是专有数据？也许这个概念比我们以前想的要宽泛得多，对吧？
And so this idea of what is proprietary data, maybe it's a little wider tent than we thought, right?
以前，专有数据只是我们主动收集来的，或者我们有独家访问权的数据；但现在，它也可以是我们自己生成的数据。
Before it was just data that we went out and collected or that we had unique access to, but now it's data that we can generate.
展开剩余字幕（还有 319 条）
那这具体是个什么样子呢？
And what does that look like, right?
大语言模型的每一次输出，都是一份新的文本数据。
Every output of an LLM is another textual dataset.
对于这个团队、这个公司来说，他们研究这些东西已经很久了，突然间冒出这么一项技术，就像你一直用手搓油，结果有人给你接了一根巨大的油管，哗哗地往外流。
And so for the group and the team and the company that's been studying that stuff for so long, And then here comes this technology that just just like as if you know how to use oil and you've rubbed more oil and then somebody has this giant oil hose that's coming out.
这太让人兴奋了。
It's exciting.
对吧？
Right?
这简直是一场机遇的革命。
So it's kind of this revolution of opportunity.
我之前说的‘把数字变成文字’，意思就是你可以和大语言模型聊任何话题。
And what what I meant about numbers into text was simply that you can have a conversation about anything with an LLM.
对吧？
Right?
你可以上传一个 Excel 表格，或者你家里的财务数据，然后说：帮我分析一下。
You can upload a Excel sheet or your personal finances at home, and you can say, tell me about it.
你实际上得到的，是文字信息。
And what you're getting at is text.
对吧？
Right?
只要这些文字有含义，本身就很有价值。
And so that text, if it has meaning, is interesting in itself.
作为一直研究这种文字使用方式的人，我觉得现在真是个令人兴奋的时刻，仿佛突然间涌来海量的新信息，规模远超以往任何一次。
I just find that to be a really exciting moment for those of us who have been studying the use of text this way because it feels like suddenly this onslaught of new information that is bigger than anything we've ever seen.
所以这是一次巨大的机遇。
And so it's a huge opportunity.
你在我们预通话时举了个特别有意思的例子，说是在 CEO 财报电话会上，测量他们的一些细微表情，比如眨眼频率，或者说到某些词时眼神的移动。
You gave this really interesting example during our pre call about measuring, I don't what I would call micro expressions during CEO earnings calls, you know, measuring how often they blink or where their eyes go when they say certain words.
几年前，像你提到的，要做这种项目或研究，可能得专门组一个计算机视觉团队，过程会非常复杂。但今天，借助 LLM 的工具，能快速原型化和探索这些想法，成本已经大幅降低了。
And years ago, as you pointed out, undertaking a project or research that may have been a massive, you would have had to have a dedicated computer vision team and it really would have been a very difficult process to explore something like that today, you know, with the tools that are available via an LLM to rapidly prototype and explore those ideas, the costs have collapsed quite a bit.
这改变了你们团队现在愿意探索的特征或想法吗？和以前没有大语言模型的时候相比？
How has that changed the types of features that your team is now willing to explore or ideas they're willing to try versus the pre LLM world?
是的。
Yeah.
这彻底改变了局面，变化巨大。
I mean, it's massively, massively changed things.
以前所有事情都讲究投资回报。
Everything was kind of return on investment.
你得有个先验基础。
You have a prior.
比如，CEO在这次电话会议中眨眼次数比以往多了三倍，对吧？我们其实对每位CEO都有一个基准数据。
Does the fact that the CEO is blinking three times more than they used to in this call, right, we have a baseline for every CEO.
不过这只是个假设。
This is hypothetical, by the way.
我只是说，如果你在研究这个，平时也观察过CEO们的历史表现，统计他们眨眼频率，就能算出一个z值，看看他们的眨眼次数偏离平均值有多少个标准差。
I'm just saying it like, if you're studying this and you study CEOs historically and you look at how much they're blinking and you could create, you know, a z score of, well, how many standard deviations up or down is their blinking going?
我们就假设已经有了这个数据。
Let's just say we have that.
这有意义吗？
Is that meaningful?
你心里可能觉得，这确实有点奇怪。
And like, part of you is probably saying, yeah, that's kinda weird.
他们为什么突然眨眼频率升高了？
Why are they spiking and blinking?
所以我觉得，这确实是个值得关注的信号。
And so part of me says, yes, that's an interesting signal.
我想好好研究一下。
I wanna look at that.
我打赌这里面肯定有东西。
I bet you there might be something there.
但如果有人跟你说，要花你团队六个月时间去研究，得下载数据、跑视觉分析、开发这些功能，你肯定会说，这主意不错，但还没好到值得花这么大代价。
But then if someone said to you, well, it's gonna take you six months of your team's time to study because you have to download the thing and do run vision, know, bake these features, you're gonna say, well, I mean, it's a good idea, but it's not that good idea.
我这儿一堆事要处理，与其看CEO的眨眼信号，我更想盯着财报数据，我对这个数据集的信任度更高。
I have a lot to do here, and I'd rather look at earnings numbers than CEO Blinkings, and I have higher priors on this dataset.
但如果你真能带来可观的投资回报，一旦投入成本大幅下降，整个计算方式就全变了。
But if you can really bring that ROI, the return on your investment, if that investment plummets, it really changes some of the calculation.
你之前脑子里转来转去的那些想法，比如‘真想看看这个’，‘真希望有这个功能’。
And so all these ideas you had swimming around, like, wish I could see this.
以前觉得这想法挺酷，但实现不了。
I wish that would be cool.
现在突然之间，这些都变得可行了。
Suddenly they're viable.
这真是个令人兴奋的时刻——你清单上那两千个永远没时间做的点子，现在突然能看了，心想：嘿，这些我以前想干的事，现在真能做了。
And so it's a really exciting time as well because if you have 2,000 ideas in your list of things that you're never gonna get to, and suddenly you can look at them and say, hey, I can get to these, all this cool stuff I've been wanting to do.
所以我觉得，这真正解放了思维，不再被实施难度这个限制死死捆住。
So I think it really changes, it frees the mind from having to worry so much about implementation as a constraint.
作为科学家，这就像是你在实验室里，突然间设备预算全没了，对吧？
And as a scientist, it's kinda like you're a scientist and you work in a lab and suddenly there's no budget for your equipment, right?
这就像是，当你不用再操心建造粒子对撞机的时候，一切都不一样了。
It's like, it changes things when you don't have to worry about building the particle collider.
眼前一下子冒出上千个想法。
Have a thousand in front of you.
同样，如果你不必为每个问题都投入巨额资金，就能自由地重新安排优先级，不再那么纠结于成本问题。
Similarly, if you don't have to worry about a giant investment to answer every question, it just frees you to reorder your things and not worry so much about the investment part.
那是不是说，所有东西都能靠大语言模型免费自动生成？
Does that mean that everything falls out for free out of an LLM?
不是，我可不是说这些问题全都解决了。
Like, no, I'm not suggesting that this is all solved.
但对很多项目来说，速度确实能比以前快上一百倍。
But it certainly can be a 100 times faster than it used to be for a lot of these projects.
这就带来了大量的新机会。
And so that opens up a lot of opportunity.
这里可能有人会反驳，如果大语言模型让功能开发变得便宜多了，而技术本身也不再是护城河，那大家的获取门槛就都降低了。
One of the counters here might be that if LLMs help make feature creation much cheaper and the technology in and of itself is no longer a moat, it's, you know, you have fairly democratized access.
这种优势可能会更快地被削弱，即使在那些曾经因为技术门槛而难以突破的极端情况下，这道墙也已经倒了。随着数据被迅速探索，所有人都在更快地挖掘新功能，优势消散的速度会大大加快。
It might be that edge gets eroded much more quickly that even in those harder edge cases where edge was over this wall of difficulty, that wall has come down and you might see edge erode much more quickly as all this sort of data gets explored quite rapidly and everyone converges on exploring features much more quickly.
你如何看待这种批评或风险？
How do you think about that sort of critique or risk?
确实，现在任何人都能更快地构建功能，不只是Two Sigma，只要他们想做，哪里都可以，对吧？
It's absolutely the case that people can build features faster, not just at Two Sigma, but everywhere if they want to, right?
所以，这道护城河确实变低了。
So there is the lowering of that moat.
但反过来讲，我们在大规模构建这些系统方面有着深厚的积累，早就准备好迎接这种变化了，对吧？
On the counter side, I would say we have such deep experience building these things at scale that we're shovel ready for this change, right?
这已经融入我们的基因了。
It's our DNA.
当出现十倍的提升，而你又系统性地懂得如何利用这些工具时，这真的让人无比兴奋。
And so when you have this 10x increase, and you've been set up to systematically understand how to use these things, it's just really exciting.
当你有了十倍的提升，却开始想：天啊，这么多东西，我们该怎么梳理？
When you have a 10x increase, and you're saying, oh, wow, How do we sort through this?
我们现在有海量的数据了。
We have all this data now.
确实，人们能生成大量数据。
So it is true that people can make a lot of data.
我不是在否认这一点。
Don't get me wrong.
我觉得在家用笔记本电脑就能做出来。
I think we can make it at home on our laptops.
就在那儿就能搞定。
We can make it right there.
你可以生成任意多的数据，但真正关键的还在后面。
You could generate as much as you want, but the then what is still really key.
所以，虽然创建数据集本身不是护城河，但怎么做、做什么，恰恰是护城河的核心，而经验在这其中占了很大比重。
And so while the ability to create a dataset is not the moat, the how, the what is very much part of that, and it's experience is a big part of that.
对吧？
Right?
你一直做这件事这么久，早就为我们迎来这一刻做好了准备。
The fact that you've been doing this for so long just makes us ready for that moment in time.
在我们之前的通话中，你提到过对AI工具的担忧，还说了一句特别好的话，我想用在这里。
In our pre call, you raised a concern about AI tooling specifically and and you had a great quote that I wanna use here.
你说AI的风险可能是让输出结果的多样性降低。
You described the risk as AI of AI as potentially quote, lowering the entropy of the output.
我很想听听你具体是什么意思。
And I'd love for you to expand on what you meant by that.
这个风险到底在哪？我们该怎么应对？
What is the risk and how can you fight back?
是的。
Yeah.
这确实很有意思。
This is really interesting.
对吧？
Right?
人人都说要自动化。
Everyone says automation.
把一切都自动化吧。
Let's automate everything.
听起来挺对的。
And it sounds right.
谁都想躺在沙滩上，让工具替我们干活，对吧？
Like, would love every we all wanna be on the beach with tools doing our jobs for us, right?
嘿。
Hey.
但自动化这事儿没那么简单。
But automation is a nuance thing.
如果我要搞一个纸箱自动化生产线，我希望每次产出的东西都一模一样。
So if I'm building a cardboard box automation, I want the output of that thing to be the same every time.
我希望每个纸箱都长得一样，就像在流水线上一样，这没问题。
I want my cardboard box to look like it is, I'm on an assembly line, and that's fine.
关于在投资组合中加入阿尔法收益，其实我们需要的是相反的东西，就像我们之前说的。
The thing about adding alpha to your portfolio is you want the opposite, as we talked about earlier.
我们需要的是非相关性，是独特性。
We want orthogonality, we want originality.
如果你过度自动化，最后可能会形成一个系统——不管谁用它，产出的东西都一模一样。
And so if you over automate, you could end up with a system that it doesn't matter who's using it, it creates the same thing.
如果大家都开始做同样的东西，你的投资组合就很难有竞争力了。
And if everyone starts creating the same thing, your portfolio is not gonna look nearly as good.
真正能提升投资组合价值的，反而是多样性。
Or what might strengthen the portfolio is diversity.
我们都清楚，这是投资里最经典的策略之一。
We all know that's a classic investing component.
像Two Sigma这样的地方，关键就在于成百上千个模型，每个模型都有自己的数据假设和研究员——这一点很重要——每个背后都是不同的人，正是这种差异性，让它们组合在一起时相关性足够低，从而能做出非常精准的预测。
And at a place like Two Sigma, it's the hundreds and thousands of models, each with their own data hypothesis researcher, this is key, each with their own person behind it, that leads to that orthogonality that when these things come together, the correlation is low enough that they can make really strong predictions.
所以担心的是，如果你给每个人一个一键式工具，让他们把数据丢进去、生成模型，结果所有人都按同一个按钮，指向同一个方向，产出的全是雷同的东西，虽然速度快、看起来很棒，但大家做的其实完全一样。
And so the fear is that if you give everyone a tool, a push button tool, and the tool is throw all your data in and make a model, and they all hit the button, and they all point at the same place, they're all gonna be putting in same output, and wow, I did that really fast, that's wonderful, but we're all doing the same thing.
这让我很担心，尤其是在我们这种以多样性为核心价值的行业里。
And that's terrifying to me for AI in an industry like ours where diversity is the point.
那我们该怎么应对呢？
How do we fight back?
说实话，我们把AI看作是激发原创性的工具。
Look, we see AI as an amplification tool of our originality.
回到你之前问的关于相关性的问题，我想打造的是能帮助团队成员放大自身独特性的工具。
Going back to that correlation question you asked earlier, I wanna build tools that allow our people to multiply what's unique about them.
当我用这个工具，你用这个工具，即使我们面对的是同一组数据，因为我们思考方式不同、过往经验不同、想法不同，工具就得能理解这一点。AI可以通过上下文来实现这一点，让不同的人用同一个工具，也能得出不一样的结果。
So when I use the tool and you use tool, and we point at the same dataset, because we approach it differently, because we historically do different things, we have different thoughts, we have different ideas, the tools need to understand that, and you can do that in AI via things like context, that allows them to navigate the pathways and come up with different answers even though we're using the same tool.
如果一个工具让两个背景不同的人——比如一个物理学家和一个计算机科学家——最终得出完全相同的答案，那我就觉得不对劲。
And so if a tool doesn't allow two people to come up with different answers, if those two people have different backgrounds, like if a physicist and a computer scientist try to solve the problem and they end up with the same answer, I don't like that.
这让我感到不安。
That makes me nervous.
所以我们必须清楚，这些工具的本质，应该是放大人类最擅长的东西：原创性和多样性。
So we really need to think, as these tools, as amplification of what we're good at as humans, which is sort of originality and orthogonality.
这在很大程度上反映了我们支持建模团队的AI战略。
And that's kind of maybe front to our AI strategy when it comes to empowering our modeling teams.
很多年前，我听说过一个概念叫‘理性忽视’，意思是有时候忽略某些事反而完全合理。
Years and years ago, I heard of this idea called rational inattention, which is that it can make complete sense to ignore something.
在大语言模型这件事上，我一直想着这点——能力提升的速度太快了，投入精力去对抗模型可能并不明智，毕竟再过一年半载，这些问题可能就自动解决了。
And that's been front of mind for me with LLMs where the pace of capability improvement has been so rapid and exponential spending time trying to fight against the LLM may not make sense when it could be a solved problem in twelve to eighteen months of what's capable, right?
所以，今天让你头疼或几乎做不到的事，未来可能变得轻而易举。
So, you know, what's painful or just barely possible today might be easy in the future.
所以我猜，像你们这样的团队，面对每一个想法时，都会面临一个真正的抉择。
So I would imagine a team like yours faces sort of this real choice for any given idea.
你可以现在就尝试，哪怕技术还不成熟，得对付各种坑洼；也可以等上一年半载，等技术追上来，那时做起来轻松多了，但同时，别人也更容易跟进这个想法了。
You can try an idea today and maybe fight against the technology if it's just barely possible and you know, deal with the rough edges or you know, you can wait twelve to eighteen months when the idea is much more trivial to explore, let capability catch up, but it also makes it sort of frankly easier for everyone else to explore that idea as well.
你怎么看待这个权衡？
How do you think about that trade off?
什么时候你愿意迎难而上，什么时候又会选择等等看，六个月后技术会变成什么样？
When do you sort of lean into the difficulty and when do you say, let's wait and see what's capable in six months?
首先，我觉得面对困难很重要，只有这样才真正理解问题在哪。
So I'll say, first of all, the difficulty is important to lean into just to understand what it is.
就像你手里拿着手电筒，突然照到一堵挡路的墙——那种无力感本身就是一种指引。
Like the very act of being unable to solve a problem because the technology is not there yet, it's like having a flashlight and seeing the wall that's blocking you.
如果你根本看不见那堵墙，等别人突然说‘我能炸开墙’的时候，你却连墙在哪儿都不知道，只能从头开始找。
Like if you can't see that wall, and then someone comes and they're like, can blow up a wall, and you're like, I don't know where the wall is, so let's start looking for walls.
那时候就晚了。
It's too late.
所以我觉得，与其等工具成熟，不如主动往前冲，逼自己去碰壁。等工具真出现的时候，就算你还没解决，你也清清楚楚知道该用它在哪——因为你已经连续六个月撞得头破血流，心里就一个念头：‘这事儿非得有个工具才能搞定’，结果新工具一出来，你立马明白：对，就是它！
So I think it's healthy to push the boundaries faster than the tools come, so that when those tools appear, if you haven't solved that problem, which you might have, but if you haven't, you know exactly where that tool goes because you've been darn it hitting that wall over and over again for six months, and it's like, only I could solve this, and it drops this new technology, and you're like, This is what I need.
作为一家技术前沿的公司，提前摸清技术的边界，真的特别关键。
So there's really, really it's really important as a technology kind of first company to get out ahead of what's happening to understand the limits.
另外，你刚才提到十二到十八个月。
I'll also say that you you mentioned twelve to eighteen months.
说实话，十二到十八个月，在像Alpha信号交易这么激烈的市场里，足以让你甩开所有对手了。
I mean, look, twelve to eighteen months, beating any competitor to a market like as competitive as trying to trade on alpha signals.
我的意思是，如果你是市场上唯一采取某种策略的人，十八个月可能长得像一辈子，这时候你率先发现的信号影响力会大得多。
I mean, eighteen months can be an eternity if you are the only one in the market taking certain approaches, and that means that your alpha signals are just much more impactful when you're out first.
所以我觉得，等十八个月或许能省下不少工程上的麻烦，但在这段时间里抢先构思点子，赶在市场反应过来之前行动，反而会特别有意思。
So I think waiting eighteen months is you might save yourself some engineering headache, but those eighteen months can be quite exciting to be out coming up with ideas before the market is caught up.
如果你觉得自己真的走在前面，那就趁早抓住这个优势。
And so you wanna take advantage of that while you can if you think you're out ahead.
我想聊一聊这方面背后的基础设施问题，好吗？
I wanna talk about the infrastructure side of this for a moment, right?
因为从技术选型上，大概有两种主要方向。
Because there's sort of maybe two major choices you can make.
一种是用别人数据中心里的模型，这可能会带来知识产权的顾虑，或者对供应商的版本更新、定价策略、产品路线图产生依赖；另一种是把所有东西都部署在本地，这当然会带来直接的成本，也可能限制了你能实现的功能，毕竟和供应商提供的服务相比，他们的更新速度和升级能力是不一样的。
You can use models that are living in someone else's data center, which introduces maybe some IP concerns or just general dependency concerns about a vendor's versioning or pricing or just road map decisions they make versus you could bring everything on premise and that's going to come with its own costs, both quite literally, but also maybe in the capabilities that are delivered there versus, you know, having a vendor delivered to you and the pace of change and upgrades there.
Two Sigma 是怎么权衡这些利弊的？
How is Two Sigma thinking about this set of trade offs?
归根结底，我觉得关键是多元化。
I think in the end, it's about diversification.
你不能死磕某一种做法，不然就会被它的局限困住。
You don't wanna lean into any one way of doing things because then you're held hostage by the limitations there.
那这话是什么意思呢？
So what's that mean?
这意味着你得用一种能支持扩展的方式来搭建系统。
It means you have to build in a way that will allow for growth.
别对所用的模型做太多预设。
You don't wanna make assumptions about the models that you're utilizing.
对吧？
Right?
不管你用的是供应商的模型、外部的前沿模型，还是自己内部开发的，都得确保你的系统能随时替换，得想清楚这对你来说意味着什么。
If you're going to use a model, be it something from a vendor, some external frontier model, or something you've built internally, you wanna make sure that your stack has the ability to swap that around when needed, and figure out what that mean is, what that means to you.
毕竟，你绝对不想被绑死在某一种选择上。
Because ultimately, you don't wanna ever be stuck on just one thing.
所以你说得对，这里面确实有各种权衡。
And so you're right, there are all sorts of trade offs.
我觉得你没提到但其实最重要的，是模型生命周期的掌控权。
I would say maybe the one that you didn't mention, but I think is actually the most important, is just control of the life cycle of a model.
如果你用了一个前沿模型，比如Anthropic的，结果他们说六个月后要停掉这个服务。
And what happens is if you use a frontier model, and then let's just say you're using something from Anthropic, and then they say, okay, well, turning this off in six months.
你可能会说：可我刚针对这个模型调优了这么多东西。
And you're like, but I just tuned all these models on it.
这就让你不得不赶紧迁移到下一个版本，再下一个，虽然也不是不行，但如果是外部托管的，你就完全没法掌控这种变动。
You've created a situation where you have to now upgrade your thing to the next one or the next one, and that's fine, but it is a moving target that you have no control over if it's externally hosted.
所以如果你想要主动权，自己托管模型确实能解决大问题。
So if you want control, certainly internally hosted models help a lot there.
另外还有一点要注意，就是前瞻偏差。
The other thing to be aware of is probably just look ahead bias.
如果你不清楚模型是用什么数据训练的，就很容易被它误导，以为某些东西是真实的，其实根本不是。
The fact that if you don't understand what model's trained on, you risk having it kind of fool you into thinking something that's not real.
随着模型越来越复杂，它们内部到底在想什么，会变得越来越难看懂。
These models, as they get more complex, it's gonna be harder and harder to understand what's deep in their psyche.
问题是，LLM 内部是不是本来就把 Enron 这家公司和负面印象绑在一起了？当你问起 Enron 的时候，哪怕只是回溯历史，你也会不自觉地觉得它是个坏东西。
The fact that like, does it connect the company Enron with negativity inherently in the back of the LLM and such that when you ask anything about Enron, in some point in time basis, you go back in time, you say, okay, well, what do you think?
很糟糕。
Bad.
你刚说，来聊聊这个吧。
It's like, oh, tell me about this.
糟糕，太糟糕了。
Bad, bad.
可你连问题都还没问呢，它就已经给你贴上标签了。
And you're like, oh, haven't even asked the question yet.
所以你得小心，我觉得掌控权真的很重要。
So you have to be careful, and I think control is important.
在其他条件都一样的情况下，自然还是希望完全掌握主动权。
And so all else being equal, obviously you would prefer to have full control.
但也不能否认，前沿实验室推进技术的速度确实令人惊叹。
However, it's also hard to not recognize that the Frontier Labs are pushing boundaries at a rate that is impressive.
所以你需要在掌控和前沿技术之间找到平衡，理解什么时候该用哪个，这本身就是个不断探索的过程。
And so you need to balance both that control, so you can understand what's happening, with the state of the art and figuring out when to use both is part of the journey.
你提到的这些权衡确实很到位，这本来就是把事情做对的一部分。
But the trade offs you mentioned are exactly right, and it's just part of getting it right.
随着数据更丰富、研究能力更强，我们不仅能探索更新颖的思路，还能以更低的成本和更快的速度实现。
With more data and more capability to research ideas, both that are more novel as well as potentially do it cheaper and faster.
我总觉得得问个老生常谈的问题：关于p值操纵和过拟合的风险，你怎么看？
I sort of feel like I have to ask the cliche question about the risk of p hacking and overfitting.
我想知道你们机构是怎么应对这个问题的。
I'm curious how you're dealing with this institutionally.
对吧？
Right?
你们现在是怎么让研究人员不重复发现同样的信号，或者避免过度拟合的？
How are you keeping researchers focused not on rediscovering the same signals or overfitting signals or how are you dealing with this right now?
我来谈几点看法。
I'll say a couple of things.
首先，如果你作为研究者严重过拟合，而我们在过程中又没发现，这点待会儿再说，最终你的模型表现肯定不会好。
First of all, if you as a researcher overfit terribly and somehow we don't catch it during the process, and we'll get to that in a second, then ultimately your model's not gonna do very well.
而且长远来看，这迟早会拖累你的职业生涯、你的成就，以及你正在做的事情。
And ultimately, that's gonna catch up with you in your career, in your success, in what you're doing.
对吧？
Right?
你的工作是构建能够面向未来进行交易的模型。
Your job is to create models that can trade into the future.
如果你过度拟合了过去的数据，以至于根本没法在未来的数据上表现良好，那你就不可能成功。
And if you've overfit on the past, so much so that if you can't do that, you can't overfit the future and succeed.
你只会失败。
You can and fail.
对吧？
Right?
但任何建模的人都应该在某个自然节点上自我检查。
But there is a natural point that anybody building a model should be self checking.
话虽如此，人一看见数据就容易激动，容易自己骗自己，觉得这东西太棒了。
That being said, humans can get excited and fool themselves when they see data, and they can come up with, hey, this is great.
我发现了这个！
I found this thing.
我太兴奋了。
I'm so excited.
那我们该怎么避免这种情况呢？
And so how do we check for that?
这个问题我们其实已经面对很久了。
Like, this is a problem we've been facing for a very long time.
我们用机器学习和大数据做实验，早就不是新鲜事了。
We've been doing scaled experiments with machine learning and large datasets for a long time.
所以，p值操纵这类问题，其实早就隐含在我们的做法里了。
And so questions about p hacking are kind of implicit in how we do things.
因此，我们发展出了一些工具来做检验和统计测试，也建立了先验假设。
And so we've developed tools to run tests and statistical tests, and we have priors.
所以我们已经采用了很多最佳实践，来避免过度使用p值操纵和过拟合。
And so there's just a lot of just best practices that we've employed to protect us from too much p hacking and too much overfitting.
需要注意的是，从定义上讲，每个人都会过拟合，因为未来和过去并不一样。
Now it's important to note that everybody overfits by definition because the future is not the same as the past.
如果你是用过去的数据训练模型，再去预测未来，那本质上就是过拟合。
If you've trained on the past and you're predicting the future, you overfit.
恭喜你。
Congratulations.
对吧？
Right?
从昨天到现在，世界已经变了。
The world changed since yesterday.
对吧？
Right?
所以问题不在于能不能避免过拟合。
And so it's not about not overfitting.
咱们别假装自己没过拟合。
Let's not pretend we don't overfit.
所有模型多多少少都会过拟合。
Everything is overfit to some degree.
关键在于把过拟合控制在你能接受的范围内，能衡量、能确认：尽管有噪音，未来会和过去不同，你依然保有一个值得骄傲、经得起市场波动的信号。
It's about overfitting to a degree that you're comfortable with, and that you can measure and make sure that despite the fact that there's gonna be some noise, that the future is gonna change from the past, that you still have a signal that you're proud of and that's robust to the trading environment.
你能衡量这个，对吧？
And you can measure that, right?
在Two Sigma，我们可以监控模型和投资组合，量化过拟合的程度，然后系统性地做出调整，大规模地持续这么做。
We can at Two Sigma monitor our models, monitor our portfolios, and we can measure if we're overfitting how much, how do we adjust for that, and systematically just do that at scale.
所以我觉得这本来就是职责的一部分。
And so I think it's just part of the role.
这是固有的特性。
It's inherent.
你这个问题的前提是AI，这会让情况更让人担忧吗？
Your question kind of premise on AI, does that make that more scary?
是的。
Yes.
因为一千个人做一千个实验，和三个人做一千个实验，后者要让人担心得多。
Because if you have a thousand people running a thousand experiments, it's lot more scary when you have three people running a thousand experiments.
所以我们必须确保有充分的培训，做好检查，保证模型的稳健性，别因为规模扩大就放松警惕。
So we do need to make sure the education's there, that we're doing the checks, that we're doing the robustness, that we're not letting our guard down as scale increases.
这确实带来了压力。
So it is putting pressure.
你正好说到了关键点。
You've hit the exact point.
这种压力确实加剧了那些担忧，但我们有经验，觉得能应对，也做好了心理准备——知道事情的扩展速度超出了以往，但我们知道怎么研究它，也完全能接受。
It is putting pressure on those worries, but we're experienced, we feel robust to it, and we're ready for the fact that like things are scaling faster than we're used to, but we know how to study that and we're okay with that.
我们之前花了大量时间讨论特征工程方面的问题。
We've been spending a lot of time talking about the feature engineering side of things.
在结束之前，我想问一两个关于数据方面的问题。之前通话前，你提到一个我觉得特别有意思的观点：这些数据库并不是固定不变的。
I do wanna, as we draw towards the close here, ask a question or two about the data side of things and in our pre call you made what I thought was a really interesting comment, which is these data libraries that come out aren't necessarily fixed.
数据经常会有升级，比如我刚才提到的那些。
There's often you can get upgrades to the data, I.
覆盖范围更广，数据更干净，历史跨度也更长。
Better coverage, cleaner data, a longer history.
很多人直觉上会觉得拿到新数据就该重新训练模型，但你对此持保留意见。
And I think a lot of people would intuitively rebuild their model with new data and you push back on that idea.
你说其实这并不总是最优的选择。
You said that's actually not always the optimal choice.
给我讲讲你的这个观点。
Talk me through that view.
这是个挺有意思的细微差别问题。
It's an interesting nuance question.
我是说，我有时候觉得从学术界转到像Two Sigma这样的地方，最难的部分在于，在学术界，你主要精力都放在博士论文上。
I mean, I sometimes say one of the hardest parts about moving from academia to a place like Two Sigma is that in academia, you are focused on your PhD thesis.
抱歉，我是说博士生，你得专注于你的论文，专注于某个具体的成果，你的目标就是在这单一件事上做到极致。
Sorry, you're a PhD student, you're focused on your thesis, you're focused on this result, and your goal is to be the best thing you can ever do at this one thing.
所以你把所有精力都押在这一种方法上，只要在这特定领域里稍微有点新发现，往往就被视为这段旅程的一部分。
And so you're putting all your things into this one way of doing things, and if you can just eke out a new finding in this very specific thing, that's often part of that journey.
于是你会想，我能不能把这个做得更好？
And that leads you to say, okay, can I make this better?
怎么才能把它做得更好？
How can I make this better?
就这样不断循环，直到成功为止，这本身没什么不对。
On an infinite loop until you succeed, and there's nothing wrong with that.
这其实很好。
It's good.
但如果你的工作是打造一个多元化的组合，你可能会觉得，这段旅程最后那10%的提升，却占了90%的精力。
But if your job is to make a diverse portfolio, you might believe that the last 10% of gains in that journey is 90% of the work.
那些才是真正有创新性的部分。
That's the most novel stuff.
所以当你在一个方向上不断深挖、追求改进时，回报会越来越低。
And so there's diminishing returns as you push in one area to make it improving.
所以我们真正相信的是，把时间花在同时做十件事，每件做到九成水平，可能比只专注一件事做到九成五更值得。
And so what we really believe is that that time might be better spent doing 10 things at 90% capacity than one thing at 95% capacity.
如果你追求的是多元化的成果，那就从十个不同角度去探索，而不是非得再花五年把某一件事做到极致——顺便说一句，我们任何一个项目都足以写成一篇博士论文。
And ultimately, if you're looking for a diverse portfolio, looking at it from 10 different ways, not necessarily the best you could ever do with another five years of work, because any one of our models, by the way, could be a PhD thesis.
可能三个月就能完成，也可能要花五年时间。
Like, could be a three month project or it could be a five year project.
我们都想把事情做得更好。
We all wanna make things better.
这种追求进步的本能，其实深植在每个人心里。
We're all it's built in all of us.
就是觉得，这个还能再改进一点，再改进一点。
It's like, oh, I can do this better and better and better.
而一个优秀的研究者，会这样想：嗯，明白了。
And what makes a good researcher is like, oh, okay.
等等。
Wait.
不行。
No.
不行。
No.
停一下。
Stop.
够了。
That's enough.
只要看起来足够好，就该放手上线了。
Either it looks good enough to just walk away and ship it.
这很难，因为大家都想做得更好。
That's hard for people because they wanna do better.
或者发现行不通，就该放弃，这点也很不容易。
Or this isn't working, walk away, which is also hard for people.
我觉得，这种能力才是像Two Sigma这样的地方顶尖量化人才的关键。
That skill, I think, is what makes a great quant at a place like Two Sigma.
关键是知道什么时候该放手。
It's the ability to know when to move on.
这就引出了你提到的问题——世界上的一切都在变化。
Which brings us to your question about the fact that things change in the world.
你得问问自己，当数据变好了，或者大语言模型的能力提升了，又或者出现了新算法时，该怎么办。
You have to ask yourself when data gets better, or a, when an LLM improves in its ability, or there's a new algorithm.
在理想情况下，你只要按个按钮，所有东西就能重新训练。
In a perfect world, you could hit a button and everything would get retrained.
如果你的系统架构足够完善，所有东西都以小规模方式构建，那当然，我们肯定希望一键自动升级所有内容，这比我们过去做的任何软件都更理想，没错，这很棒。
And if you had the stack, and everything was done in this sort of smaller scale thing, yeah, I mean, obviously we would love to automatically upgrade everything at a push of a button than anything we've ever done in any software, like, sure, that's great.
但在现实世界里，事情要复杂得多，对吧？
But in the real world, things are a little bit more nuanced, right?
因为你构建的模型是基于某些假设的，如果你重新开始训练，就像我之前说的博士生那样，你可能把它提升了10%，但你花的时间值不值？与其这样，不如直接去开发一个完全不同的东西。
Because you've built a model with certain assumptions, and if you go and you restart that process, see my previous comment about a PhD student, like you might be improving it and you might make it 10% better, but what is the time trade off of making your thing 10% better versus going and building something completely different?
所以我们真正看重的，是这种多元化的思路，我们把它放在首位。
And so we really believe that it's that diversity of approach that we kind of put first.
那是不是说我们从来都不更新东西？
Now does that mean we don't ever update things?
不，我们其实经常更新。
No, we do all the time.
很明显，我们一直在持续监控。
This is obviously we're constantly monitoring.
如果发现机会，我们当然会去把握，但我们会从商业角度来评估：花这些时间值得吗？新数据真能带来多大的提升？
If we see opportunity, we will chase that opportunity down, but we do it with a business lens, like what is the ROI of the time versus how much better do I really think this thing can get with this new data feed?
比如从版本三升级到版本四，我可以重新训练整个项目，但也可以想，我现有的系统已经用的是类似的数据了。
So a vendor terms from v three to v four, and I could retrain my whole project, or I could say, look, I've already got something that uses that type of data.
它表现得已经不错了。
It's doing pretty well.
为什么不继续推进手头的其他事情呢？
Why not keep going on my agenda doing something else?
当然，如果我觉得花时间回头优化它最划算，我也会回去做，但很多时候并不值得，因为还有太多领域我们根本还没触及。
And yeah, I will maybe go back and visit it if I think it's the best use of my time, but often it's not just because there's so much of the world we haven't yet captured.
所以如果你在找那种相互独立的特性，通常不会只是因为某个供应商做出来的东西稍微好了一点。
So if you're looking for orthogonality, it's rarely just gonna be in like, oh, this vendor made something slightly better.
数据和人工智能的交汇点，我们在准备会议时你提到的一个概念是‘规模化个性特征’。
The intersection of data and AI, one of the ideas you brought up on our prep call was this idea of idiosyncrasy at scale.
有些数据库里可能包含日本三家公司的消费支出信息，过去这些数据很难利用，但你提出的观点是，人工智能能让你针对单个公司做极度个性化的分析，同时又能轻松地把这种个性化分析推广到大量公司。
There are these examples of sort of data libraries where maybe you get consumer spending information about three companies in Japan that historically may have been hard to use, but maybe this idea you brought up was AI allows you to do something hyper company specific but simultaneously scale hyper company specific much more easily.
我想请你进一步展开一下这个想法，结合这种独特信息的视角，谈谈你对特征构建前沿的看法。
And I wanted to sort of get you to expand upon that thought and how you see sort of the frontier of feature creation with this idea of idiosyncratic information in mind.
对。
Yeah.
我经常用‘规模化个性特征’这个词，是为了鼓励那些原本回避公司级具体指标的人，他们总想着要更大的样本量、更多的公司、更大的机会，对吧？
I mean, I I use that term a lot, idiosyncratic scale, to inspire people who have generally maybe shied away from the very company specific measurement, because they're looking for larger sample sizes with more companies and larger opportunity, right?
你可能会觉得，投资一家公司，机会肯定比投资一百家公司要小。
And so you can believe that if you trade one company, less opportunity than you trade 100.
所以，一个只针对一家公司的特征，听起来肯定没一个覆盖一百家公司的特征那么吸引人，对吧？
And so if you have a feature with one company, it's less exciting than if you have one about 100 companies, right?
仔细想想，这其实挺自然的。
And it's kinda natural if you think about it.
但AI的独特之处在于，它能深入分析某一家公司，提炼出既针对该公司、又可能适用于其他公司的特征类别。
But the thing about AI is that it opens up its unique ability to dig in on a specific company and build categories of features that might be specific to that company, but might also be generalizable to other companies.
你可以从一家公司最细微的细节出发，再回头提炼出共性，比如：我对这家公司的销售前景怎么看，或者类似的问题。
And so you can reason your way down to the very, very specific things about a company, and then you could go back and generalize them, call that, okay, well, what's my outlook on sales, or I don't know.
而这条路径，往往对每家公司来说都是独一无二的。
And the path there might be very, very idiosyncratic to a company.
想想主动型投资者的做法，他们很大程度上依赖于对所关注对象的深入了解。
And so if you think about what a discretionary investor does, right, I mean, lot of it is just deep knowledge about some of the things that they're looking at.
但当你只看大规模的财报数据时，这些细节很难被捕捉到。
And these things are hard to pick out at scale when looking at large earnings numbers, right?
真正的机会往往藏在那些不会出现在宏观数据里的公司特定信息里。
They're usually very company specific opportunities that arise that aren't gonna show up in that space.
所以，如果你想在这样的分析中更具竞争力，同时又具备像Two Sigma那样的规模化思维，我坚信AI是关键——因为它能同时为三千家公司做分析，而对每一家做的分析都是独特的。
And so if you want to expand to be more competitive in that type of analysis, but also have this kind of scale attitude of Two Sigma, I just deep down believe that AI is a big part of that opportunity because we can use AI to do something at 3,000 companies at once while the thing it's doing for those 3,000 companies is itself unique.
所以这很有趣，既要考虑规模，又要兼顾具体性。
And so that's kind of an interesting thing to both think about scale and specificity at the same time.
因此，这正是我们正在探索的一些令人兴奋的拓展方向——利用AI更深入地挖掘那些个性化强、难以建模的公司数据部分。
And so, yeah, big part of some of the exciting expansions that we're thinking about here and how to use AI to just go deeper and deeper into the more bespoke and hard to model parts of company data.
我之前跟长期听众提过一个笑话：2009年我读研时，毕业项目是给信用违约互换定价，结果我毕业不到一年，这技能在华尔街就完全没用了。
One of sort of the, I'll say jokes long time listeners will have heard me sort of talk about before was that when I went to grad school in 2009, the capstone project of my grad school was pricing credit default swaps, which became a totally irrelevant skill on Wall Street within, you know, one year of me graduating.
对现在的特征工程师和研究人员来说，整个领域变化得太快了。
It seems like for feature engineers and researchers today, the landscape is moving very quickly.
所以当你带着团队里的初级特征工程师一起工作，他们正思考自己的职业方向，以及未来十年行业会怎么变，你觉得他们该关注些什么？
And so as you're working with junior feature engineers on your team and they're thinking about their careers right now and where the world is moving in ten years, what do you think they should be thinking about?
未来十年，哪些技能或直觉会给他们带来最大的复利增长？
What sort of skills or instincts do you think are gonna compound the most for them over the next decade?
我觉得随着AI的发展，能把自己的想法写成代码、转化成可执行格式的技术能力，其相对价值正在下降，而更重要的，是提出问题、指导AI该做什么、怎么做。
I think with the change of AI, it's that the technical skills of being able to code up your idea or put it into a particular format to execute it are obviously gonna drop in relative value to the idea to the skill of telling those things how to do or what to do.
过去，你可能会看重一个人，他特别擅长按别人的要求把东西做出来。
And so whereas you might have had somebody who is very adept at building the thing that somebody tells them to build.
我觉得这种能力正在变得没那么重要了。
I feel like that, as a skill, is growing less valuable.
因为你只要让AI去搞定，这个人能带来的额外价值就变小了。
Because if you can ask an AI to build it, you know, that person's additive value is less.
那这样一来，我们该往哪儿走呢？
And so where does that leave us?
我的意思是，归根结底，首先要学会接受AI以及它能做什么。
I mean, I think ultimately, it is a question of, first of all, getting comfortable with AI and its capabilities.
对吧？
Right?
现在还在上学的人，将来进入的会是一个AI更普及的世界。
Whoever's in school right now is going to be going into a world with more AI.
我们从学校招人时发现，每一代新人对AI都越来越熟悉。
And what you find with each generation of people that are coming out of into Two Sigma right now as we hire from schools is that they're more and more AI native.
他们越来越习惯某些东西就该是那个样子，因为他们在学校里接触的就是这样。
They're more and more expecting things to look a certain way, because that's what they have at school.
而且这变化发生得特别快。
And it's happening very quickly.
首先，你得像我说的那样，学会怎么用AI来提升自己的本事。
I think, first of all, you need to be comfortable, as I said before, how do use AI to bring your own skills forward?
不是那种只会‘做个网站’的技能，对吧？
And maybe not the skill of just quote, build a website, right?
那什么才是你独有的东西？
But what's unique to you?
你的知识和教育能带来哪些别人做不到的价值？
What does your knowledge and your education bring that you can do that others can't?
你又该怎么用AI来放大这些优势？
And how do you use AI to amplify that?
因为要想成功，你既要脱颖而出——记得我之前说过的独特性吗？——也要能规模化。
Because to succeed, you need to both stand out, see my orthogonality comment before, but also scale.
我觉得很多人把AI当工具，总觉得它会取代我，或者取代这个岗位。
And I think AI as a tool, many people think, oh, it's gonna replace me or it's gonna replace this role.
我说，不是这样的。
And I'm saying, no.
不是。
No.
不是。
No.
要让它帮你放大优势。
Let it amplify you.
别觉得它只是在替你做原来的事。
Let let like, don't look at it as, oh, it's doing the thing I did.
哇，真厉害。
Say, wow.
它能让我把原来做的事做得更好、更快，我能用自己的想法，把想干的事规模放大，而且快得多。
It can help me do what I did way better and faster, and I can use my ideas and just scale up what I want to do before, but much faster.
我觉得这才是关键：学会用下一代算法去做真正独特、聪明、有创意的事，这才是未来真正重要的能力。
I'd say maybe that's the main takeaway, that getting comfortable with how to use the next generation of algorithms to do really unique and clever and creative things is going to be the skill that's gonna carry over.
因为归根结底，未来的世界会大量依赖人类语言来互动。
Because ultimately, it's going to be a world where there's gonna be a lot of interacting through human language.
人们写代码的时间会越来越少。
The amount that your people are coding is gonna go down.
但使用提示词的人会越来越多。
The amount of people that are prompting is gonna go up.
真正决定成败的，将是你的想法。
And therefore, it's the ideas that are going to carry the day.
关键在于，你怎么最有效、最有创意地用它来达成你的目标？
It's the how do you use this in the most effective and creative way to get what you need done?
所以，我觉得这才是重点。
And so a focus on that, I think, is kinda key.
我想用一个我这个季节问过每个人的问题来结束：你现在最着迷的是什么？
And I wanna end with the same question I've asked everyone this season, which is what is something that you are currently obsessed with?
最好是跟工作无关的事情。
And ideally, is outside of work.
这可能是一个想法，也可能是一本书、一部电影，或者任何你现在挥之不去的东西。
So this could be an idea, it could be literature, art, movie, just something you can't get out of your mind right now.
你目前着迷的是什么？为什么？
What is your obsession and why?
我最近特别着迷于公共数据和政府公开的信息。
I am kind of obsessed with public data and the government's release.
我知道这跟我的工作有点关系，毕竟我是搞数据的，但我是出于个人兴趣在关注。
So I know this is kinda it's, you know, it's related to my work because I'm a data person, but, you know, done personally.
我觉得特别有意思的是，有这么多信息其实都公开着，比如停车罚单、谁被拦下超速了、餐厅的卫生检查记录等等。
And so I think it's fascinating that there's all this information out there, you know, the parker tickets that we get, the who gets know, whether or not people get stopped for speeding, health inspections at restaurants.
里面藏着大量值得挖掘的有趣内容。
And there's so much interesting stuff in there that can be done.
我特别享受探索这些可能性——如何利用现有数据，让城市和整个社会运行得更高效。
And I've really enjoyed exploring the opportunity there, how we can kind of make our cities and society run better by utilizing the data that's already there.
我觉得AI的好处在于，它让过去只有数据科学家才能接触的数据，现在普通人也能用了。
I think the nice thing about AI is that it eventually will it used to be that the peep like, only data scientists could go in and use this data.
对吧？
Right?
以前你得有那个能力才能用这些数据。
And you had to have the skill to utilize it.
但现在你不需要那些技能了，只要问问题就行，这些数据马上就能成为城市规划师、律师，还有任何人手中的有力工具。
But if you don't need that skill anymore and you can just ask questions, suddenly that data is going to be driving power for urban planners and lawyers and anyone.
它让公众机构的信息获取变得真正平等了。
It just democratizes the access to information for a public institution.
我觉得这特别令人兴奋和鼓舞，因为你不再需要什么技术门槛了，或者说很快就不需要了，你和你想了解的公共数据之间，再没有隔阂。
And so I just find that really exciting and inspiring that you don't need a purse there's no more, you know, tech requirement or there won't be soon, I guess, between you and the questions you might have on public data.
所以这正是我工作中研究和探索的领域，当然，私下里我也特别着迷。
And so that's an area that I study and explore and nerd out on related to my work, but, you know, also on a personal basis.
我觉得，越多人关注这些数据，地方政府或联邦政府就越显得效率低下，数据被撤回的速度也就越快。
I would suggest that the more and more people who get tuned into that, the more look makes, local or federal government look less efficient, the quicker that data will be pulled.
对吧？
Right?
另一个则是担忧。
Well, the other is a fear.
这些信息随着时间推移有多大的用处。
How useful that information is over time.
本，今天这场对话真是太棒了。
Well, Ben, this has been a fantastic chat.
非常感谢你抽出时间，跟听众分享这么多内容。
I really appreciate you taking the time out of your day to share all this with the listeners.
真的非常感谢。
Thank you so much.
是的，聊得特别开心，收获也很多。
Yeah, had a great time and a great conversation.
非常感谢你的时间。
Thanks so much for your time.
关于 Bayt 播客
Bayt 提供中文+原文双语音频和字幕，帮助你打破语言障碍，轻松听懂全球优质播客。