为资本市场构建人工智能系统

欢迎收听高盛交易所播客。

Welcome to Goldman Sachs Exchanges.

我是艾莉森·内森，今天和我一起的是乔治·李，他是高盛全球研究所的联席主管。

I'm Allison Nathan, and I'm here together with George Lee, who is the co head of the Goldman Sachs Global Institute.

我们将一起推出一系列节目，探讨人工智能的崛起及其对公司、投资者和经济可能带来的影响。

Together, we're hosting a series of episodes exploring the rise rise of AI and everything it could mean for companies, investors, and economies.

乔治，很高兴再次见到你。

George, great to see you again.

很高兴见到你。

Good to see you.

今天我们要深入探讨一个非常有趣的话题：在人工智能时代，真正打造产品需要什么——包括要素、挑战和方法，来构建那些首先真正有效、同时也真正重要的解决方案。

So today, we're digging into another fascinating topic, what it takes to actually build products in the AI age, the ingredients, the challenges, and the approaches behind building solutions that first and foremost actually work, but also actually matter.

我首先想听听你的看法，乔治，因为我们多次讨论过，你每天都在和公司、企业、客户聊这些话题。

And I really just wanted to first get your thoughts, George, because as we've discussed many times, you're talking to companies, firms, clients about these topics every day.

你觉得这方面实际取得了多大的进展？

How much real progress do you see being made here?

嗯，我对这次讨论非常期待，因为在我们之前的几期里，我们花了不少时间聊了AI的基础要素、基础设施层面，而现在讨论在这个环境下构建应用这个话题，我觉得非常及时也很有用，而且我们为此请到了一位很棒的嘉宾。

Well, I'm very excited for this discussion because in our prior sessions, spent a bunch of time talking about the foundational elements of AI, the infrastructural dimensions, and the topic of building applications in this environment, I think, is very timely and useful, and we have a great guest for that.

你看，我觉得在生态系统中，有很多问题围绕着这个技术栈中基础层面和应用层面之间的边界在哪里。

Look, I think there are a lot of questions in the ecosystem about what is the perimeter of the foundational dimensions of this relative to the application part of this stack.

所以，我们的嘉宾会以一种非常有效的方式帮我们理清这些问题。

So again, our guests will be able to untangle that for us in a really productive way.

好，那我们就请出今天的嘉宾吧。

All so let's bring in our guest.

他就是Chris Churchman。

It is Chris Churchman.

Chris是Marquee的负责人，这是我们公司面向机构和公司客户的数字平台，同时他也是我们公司全球银行与市场AI工作组的联席主席。

Chris is the head of Marquee, the firm's digital platform for institutional and corporate clients, and he's also co chair of the firm's Global Banking and Markets AI Working Group.

所以，Chris，你在所有这些事情上都是站在第一线的，特别是在把新的AI产品推向市场方面。

So you're on the front lines, Chris, on all of these matters in terms of bringing new AI products to the market.

欢迎来到我们的节目。

So welcome to the show.

谢谢。

Thank you.

很高兴来到这里。

Excited to be here.

太好了。

Great.

好的，Chris，我们先让听众了解一下Marquee的背景，你们的核心使命是什么，然后再聊聊你们是如何开始在其中融入一些AI能力的。

Well, Chris, let's start first by just giving our listeners a little bit of background on what Marquee is, what your core mission is, and then let's talk a little bit about how you're starting to fold in some AI capabilities there.

当然可以。

Sure.

是的，我对这次对话非常期待。

Yeah, so excited for this conversation.

感谢你们的邀请。

Thanks for having me.

Marquee是公司面向机构和企业客户的平台。

So, Marquee is the firm's platform for institutional corporate clients.

我们的目标就是帮助客户在高度不确定的情况下做出决策。

And is our help our clients make decisions under great uncertainty.

所以我们希望成为客户投资流程中不可或缺的一部分。

So, we want to become an integral part of our clients' investment process.

所以，当你考虑把生成式AI能力整合到这个平台时，你可以想象我们已经构建了什么。

And so, as you think about layering in Gen AI capabilities into the platform, you can imagine what we've built.

也许更好的描述平台的方式是我们如何整合生成式AI能力，因为你必须从要解决的问题出发，而不是仅仅说，哦，让我们在那些方面用AI，那是人们从AI开始，但让我们从投资流程开始。

And maybe it's actually a better way to describe the platform is how we've layered in Gen AI capabilities because you've gotta start with the problem to be solved, not just, oh, let's use AI on that and that and that, which is people start with AI, but let's start with the investment process.

那么，生成式AI解决方案如何成为你的助手，帮助你在投资流程的漏斗中前进呢？

And how would a Gen AI solution help you be the concierge that helps you down that investment process funnel?

所以，如果你愿意，从漏斗的顶端开始，那里有数百万篇关于不同公司、行业、经济预测等的研究文章。

So you start on the top of the funnel, if you like, which is just millions of research articles on different companies, sectors, economic forecasts, etcetera.

但还有所有的交易台评论，关于资金流向、交易想法等等。

But then you've also got all of the trading floor commentary about flows, trade ideas, etcetera.

所以，这对人们来说，找到与自己相关的内容是压倒性的。

And so that's overwhelming for people to find what's relevant to them.

然后你还有所有需要看的分析，以及来自几十家供应商的数据。

And then you've got all of the analytics that you'd need to look at, all of the data from dozens of vendors produced.

我们很幸运有一个叫做MarketView的生态系统，你可以把它想象成资本市场的Pinterest，每个组件都由各自领域的专家制作。

We've got an ecosystem, fortunately, called MarketView, which contains these you can think of it like Pinterest for capital markets where each widget is produced by an expert in their domain.

他们获取相关的供应商数据。

They take the relevant vendor data.

然后基于这些数据形成见解。

They then create an insight on that.

现在这个生态系统里有数百万条这样的见解。

And now we have millions of those in the ecosystem.

然后我们还有交易前的分析。

And then we've got pre trade analytics.

比如，好的。

Like, okay.

我要做什么交易？

What trade am I gonna do?

我可以做情景分析。

I can do scenario analysis.

我可以做回测。

I can do back testing.

所有这些我都能做。

I can do all of these things.

但从用户体验的角度来看，这非常困难。

But from a user experience point of view, it's very difficult.

我觉得我们看到的软件领域的范式转变是，我们从用户学习软件的时代，进入了软件学习用户的时代。

And I think the paradigm shift that we're seeing in software is that we've gone from a world where users learn software to where software learns users.

所以现在我们已经到了这样一个阶段：你只需要向系统表达你的意图，系统就能完全理解Marquee的所有功能，并把它们呈现给用户。因为一直以来，瓶颈就在于用自然语言表达意图——这是一个带宽很低的渠道，但实际上你想表达的意思远比你说的要多。

And so now we're at the point where you can just express your intent to the system, and the system fully understands all of the capabilities of Marquee and can deliver them to the user because that's always been the bottleneck is expressing intent in natural language as a very low bandwidth channel, but actually you mean way more than what you've actually said.

所以Marquee AI，目前只在内部使用，你问Marquee一个非常具体的问题，这就是你关心的。

And so Marquee AI, which is only available internally right now, you ask Marquee a question which is very targeted, is what you care about.

但它会根据最近发生的事情，把它拆分成多个相关的研究主题。

But then it will take what's happened recently to break it out into a number of relevant research topics.

对于每个研究主题，它会找到所有相关的研究、所有相关的交易台评论、所有相关的小组件和市场观点，也就是数据和数据分析，然后把它们整合成一个连贯的思路。

For each of those research topics, it will then find all the relevant research, all the relevant trading floor commentary, all of the relevant widgets and market views, so data and analytics, and then assemble that into a cohesive thought.

然后，如果需要进行计算，它会调出包含该数据的小组件，这些数据来自专家的可靠数据，然后运行一个计划，进行计算，比如编写并运行一些Python代码，然后给出答案，最后给你一份完整的、每个句子都有依据的文档，每个句子都基于公司里某人说过的话或你可以下达的计算。

And then if it needs to do a calculation, it will pull the widget that contains that data, reliable data from an expert, and then run a plan, do a calculation, like write and run some Python and give back the answer, and then give you a full, well grounded every single sentence in the document is grounded to something someone at the firm has said or a calculation that you can order.

这是最后最难加上去的一层。

And that is the hardest bit to layer on at the end.

然后我们可以再详细聊聊，但它的生成能力是关键，而且非常非常非常不同。

And then we can talk more about it, but the generative ability of it is key, but it's very, very, very different.

我们可以再详细聊聊。

We can talk more about it.

幻觉问题。

Hallucinations.

对。

Yeah.

我们现在就深入探讨一下，因为在你这个角色中，以及在我们最终向客户扩展这项服务时，事实性、可靠性和来源可追溯性非常重要。

Let's just double click on that now because in your role and for this service as we ultimately extend it to clients, factuality, reliability, provenance is really important.

对。

Yeah.

我知道这在你做的工作里一直是个特别难的部分。

I know that's been a really hard part of the work you've been doing.

也许可以再深入聊聊这个。

Maybe just double click on that.

这确实是特别特别难的一部分。

It's a really, really hard part of it.

而且说实话，这也是最让人沮丧的，因为通常的情况是，你看到有人搞了个很炫的演示。

And frankly, it's also the most frustrating because what typically happens is you see someone with a flashy demo.

对吧？

Right?

他们会用前沿模型，忽略所有关于事实依据的细节，然后给出一个非常自信、看似合理的答案，让人震惊。

And they'll take the frontier model, they'll forget about all of the detailed stuff around grounding, and they will come up with incredibly confident plausible answer, and it'll blow people's brains.

但问题在于，模型本身无法区分事实和推断。

But the problem with that is that the model itself cannot distinguish between a fact and an interpolation.

就是，它其实真的做不到。

Like, it just it actually can't.

所有东西都进入上下文窗口。

Everything goes into the context window.

全都经过同一个处理流程，出来之后，它真的不知道。

It all goes through the same sausage machine, comes out the other end, and it really just doesn't know.

不是它在误导你。

It's not that it's misleading you.

只是我们没注意到，因为它知道得太多了。

We just don't notice it because it knows so much.

但当你把它应用到机构环境中，当模型不了解你的实际数据时，那就是个大问题。

But when you try and apply it to institutional context, when the model doesn't know your actual data, then that's a big problem.

但这一领域的演变已经彻底改变了局面，因为早期的时候，比如，这就是为什么在AI时代构建东西如此令人沮丧。

But the evolution of this has changed things dramatically because in the early days, for example, and this is why it's so frustrating to build in the age of AI.

就拿ChatGPT来说吧。

So take ChatGPT.

对吧？

Right?

2022年11月20日。

'20 November 2022.

对吧？

Right?

我们看到这个，然后就想，好吧。

We saw that and went, okay.

嗯，模型有些事做不了。

Well, model can't do certain things.

对吧？

Right?

它不知道的话就会瞎编，而且它不会用任何工具。

It makes stuff up if it doesn't know, and it can't use any tools.

所以它不能用计算器之类的。

So it can't use calculator or anything like that.

所以你得花很多时间琢磨怎么给它这些能力。

And so you spend a lot of time thinking about how do we give it those capabilities.

然后你就想，对。

And so you go, right.

好吧。

Okay.

它上下文窗口只能处理4000个token，也就是3000个词。

It can only handle 4,000 tokens in the context window, so 3,000 words.

所以你花大量时间把内容拆成小块，嵌入到向量数据库里，基本上就是把它们变成代表含义的数字，然后全部存进向量数据库。

So you spend all of this time breaking things up into chunks, embedding them in vector databases, which is basically, like, turning them into numbers that represent their meaning, filing them all away in, like, vector databases.

结果一年后，上下文窗口变成了128000个token。

And then a year later, the context window goes to a 128,000 tokens.

你心里想，哦，那之前做的都白费了。

And you're like, oh, well, that was pointless.

然后它又变成了一百万个token。

And then it goes to a million tokens.

然后你会说，早期阶段也很有意思，因为有一篇非常著名的论文。

And then you say, well, the early stages were interesting as well because there was a very famous paper.

我觉得我们之前聊过，就是那篇《迷失在中间》，讲的是这些模型，基本上，如果你做“大海捞针”测试，就是把一些信息藏在中间某个地方，对。

I think we've talked about it, but lost in Middle, which is that these models, basically, if you do the needle in the haystack test, which is that you hide some context in the middle somewhere Mhmm.

它找不出来，因为它会优先看开头和结尾。

It can't retrieve it because it preferentially looks at the start and the end.

你会想，哦，它们跟人一样。

And you think, oh, they're like humans.

它们会注意播客的开头和结尾，中间部分就有点走神了。

They pay attention at the start of a podcast and the end of the podcast, and they sort of drift in the middle.

对吧？

Right?

这就是为什么要把内容拆成小块。

And that's why you break things up into chunks.

但有意思的是，这些模型可是基于所有人类知识训练出来的。

So but what's interesting about that is they've been trained on all human knowledge.

对吧？

Right?

所以如果你看所有人类写的文档，开头非常重要。

And so if you look at all human documents, the start is very important.

对吧？

Right?

开头，非常非常重要。

Start, very, very important.

结尾，非常非常重要。

End, very, very important.

所以它学到了这一点。

So it's learned that.

对吧？

Right?

所以谷歌把上下文长度提升到一百万token的时候，那真是了不起，他们也解决了大海捞针的问题。

And so Google, when they raised the context limits of 1,000,000 tokens, which was amazing, they solved the needle in the haystack problem as well.

所以之前那些关于把内容放在开头和结尾的折腾，其实完全没必要，因为他们能直接检索到。

So all of this finagling about put it at the start and the end, it was all completely pointless because they could retrieve it.

所以我觉得在应对幻觉问题时，问题在于人们喜欢那些听起来合理又自信的回答。

So I think as we looked at dealing with hallucinations, there's been the problem is people like plausible confident answers.

如果我告诉你，我建的系统有90%的时间是对的。

And if I told you that the system I built is right 90% time.

对吧？

Right?

10%的幻觉率，这不算什么。

10% hallucination rate, which is not.

它不会产生幻觉。

It doesn't hallucinate.

但如果我建了这样一个系统，你就得自己去判断那10个事实里哪个是编的。

But if I built you a system like that, the burden is on you to figure out which of the 10 facts is fabricated.

所以你得把所有事实都检查一遍。

So you have to check all the facts.

对。

Right.

所有的事实。

All the facts.

那你就是这么做的吗？

And is that what you do?

我是说，你实际上是怎么排查这些问题的？

I mean, how do you actually troubleshoot this in practice?

这非常痛苦，因为你要做的就是这件事。

It's very painful because what you do and this is the thing.

即使是现成的零售平台，它们也会给你来源。

Even off the shelf retail platforms, they give you sources.

但实际上，如果你尝试像我们在Market AI里做的那样，把每个内容都归因到某个来源，具体到每个来源的某一行，你用网页接地是做不到的，因为它并不是真正在用它。

But, actually, if you try and do what we've done in Market AI and say, attribute everything to a source, a specific line in each source, you can't do it with web grounding because it's not really using it.

它只是说，这是我的来源。

It's saying, here are my sources.

当我们狠狠质疑它的时候，至少它还挺诚实的，它就说，你看。

And when we challenged it hard, at least it was honest, it was like, look.

说到底，我擅长的是听起来很全面，而不是真的全面。

In the end, I'm better at sounding thorough than being thorough.

对吧？

Right?

所以你真的得搭一个框架来逼它去引用来源，因为它本质上做不到，而且它也没有动力去做。

And so you really have to build the framework to force it to ground itself because it inherently can't do it itself, and it's not rewarded to either.

有一篇著名的论文讲为什么大语言模型会幻觉，因为它们在做选择题，而且它们不会因为放弃回答而得到奖励。

There was a famous paper why LLMs hallucinate, and it's because they're doing multiple choice tests, and they're not rewarded for abstaining.

它们还不如猜一个，因为那是最优策略。

They might as well guess because that is optimal.

所以它们总是会幻觉，而且它们分不清事实和推断。

And so they're always gonna hallucinate, and they can't distinguish between a fact and an extrapolation.

你觉得我们最终会不会达到一个点，就是我们对这些模型构建的产品有完全的信心？

Do you think that we're ultimately going to get to a point though where we have complete confidence in these products built by these models?

不会。

No.

我觉得我们不会达到完全放心的程度，但这正好触及了我认为这个领域未来走向的关键。

I don't think we're gonna get to the point where we're fully confident, but it touches on where I think the puck is going with this.

所以我觉得我们已经经历了提示工程阶段，就是那种告诉模型具体该怎么做的阶段。

So I think that we've gone through the prompt engineering kind of let's tell exactly what to do.

对吧？

Right?

我们也经历了上下文工程阶段，就是想办法把一些记忆、数据、提示和操作手册拼凑在一起。

We've gone through context engineering, which is, okay, what if I can cobble together a bit of memory, a bit of data, a bit of prompts, a bit of playbooks.

我们现在正处于智能体工程阶段，也就是说，现在我们有了一个智能体循环，可以检查自己的工作并不断迭代。

We're currently in the agentic engineering phase, which is, okay, well, now we've got an agentic loop that can check its own work and and iterate.

当然，你不能把那些旧模型放进这个框架里，因为错误会累积。

And sure, you couldn't put one of those old models in this framework because the errors would compound.

对吧？

Right?

所以，它能检查自己的工作，这意味着它现在比以前任何时候都更可靠。

So the fact that it can check its own work means that it is more reliable now than it ever was before.

所以这些循环很重要。

So those loops are important.

但现在我觉得我们正在进入环境工程阶段。

But now I think what we're getting to is environment engineering.

我们怎么确保这些代理运行的环境是安全的，尽管最近的头条新闻不是这样，而且它们还能以完全授权的方式访问所有相关工具和机构知识。

How do we make sure that the environments that these agents are operating in are secure, recent headlines notwithstanding, but also that they have access to all the relevant tools and institutional knowledge in a fully entitled way.

然后我觉得下一个阶段，虽然还没有命名，但我觉得我们正在朝那个方向走，那就是任务授权工程。

And then I think the next phase, which hasn't been coined yet but I think we're going there, which is like mandate engineering.

也就是说，AI代理在谁的授权下，以谁的名义，被允许做什么。

Which is like what are AI agents allowed to do under whose authority in whose name?

我觉得这就是前沿。

And I think that's the cutting edge.

我觉得一旦有了围绕这些代理行为的保险市场或责任市场，你就会看到它们的自主性增加，但它们仍然会犯错。

I think once there's an insurance market or a liability market around what these agents do, then I think you'll see the autonomy of them increase, but they'll still make mistakes.

这就像信任任何一个人一样。

And it's like trusting any human.

机制是一样的，但我们没有像对待人类那样的追索权。

It's the same mechanism, but we don't have the recourse that we do with humans.

你刚才描述的那个发展过程，真的很有意思，而且讲得很清楚。

It's fascinating to reflect actually all that you described very eloquently that progression.

这还不到四年。

It's been less than four years.

这些AI发展的时代，都压缩在这四年里了。

Those are eras of AI development that have been packed into this four year period.

正如你所说，这些系统的可靠性置信区间已经大大扩展了。

As you say, the confidence interval in the reliability of these systems has expanded a lot.

由于技术的概率性本质，它们可能永远不会接近完美，但在这么短的时间里，各个研究领域都取得了惊人的进展。

They may never close to perfection based on the probabilistic nature of the technology, but amazing progress in different regimes of research in that short period of time.

确实如此。

Absolutely.

你提到的另一个问题是，怎么跟上这些变化？

The other question you brought up was how do you keep up?

我们会继续看到这些模型的演进。

We are going to continue to see evolution of these models.

那你打算怎么走在前面？

So, what are you doing to get ahead of that?

或者说，你能做些什么来领先一步吗？

Or can you do anything to get ahead of it?

我觉得你没法真正领先。

I don't think you can get ahead of it.

我想说两点。

I guess I'd make two points.

首先，从演示说起，你可以拿一个前沿模型接上数据，效果会让人惊叹。

One is about the demo to start with, which is that you can take a frontier model and connect it to data, and it will blow your mind.

这就是概念验证。

And that's the proof of concept.

但演示看的是它表现最好的那一天，而产品看的是它表现最差的那一天。

But a demo is judged on its best day, whereas a product is judged on its worst day.

对吧？

Right?

所有的努力都在于弥合这个差距。明白。

And all of the work is the gap between Mhmm.

就是这一点。

That.

而且这活儿既枯燥又费力，说实话还挺没人感激的。

And that is boring, hard, actually pretty thank less work.

所以如果你试图去弥补模型本身的缺陷，最后只会白忙一场，浪费自己的时间。

And so if you try and build around the deficiencies of the model, you will get burned eventually, and you'll waste your time.

举个例子，就拿上下文窗口来说吧。

So, for example, take the context window.

为了处理这个问题，大家得搭建RAG管道、做文本分块、搞向量数据库，费了老大劲，其实只是为了修修补补。

All of that work to build rag pipelines, chunking, vector databases, all the right, just to cure it.

现在你直接把整本书扔给它，对吧，或者好几本书，整个图书馆都扔给它，而不是只给它看片段。

Now you just throw out the whole book at it, right, or multiple books, the whole library at it instead of showing chunks at it.

所以，要为你发布时将会拥有的模型来构建，而不是为你刚开始开发时拥有的模型来构建。

So build for the model that you will have at launch, not for the model that you have when you're starting your development.

这就是我学到的第一课。

So that's the first lesson I've learned.

所以，假设我们构建的那些推理脚手架，一旦推理模型出现，就完全没用了。

So assume that those so all of those reasoning scaffolds we built, once reasoning models came out, totally pointless.

我们所有那些把大语言模型连接到我们的工具和数据的方式，一旦有了函数调用和MCP，就完全没用了。

All of those ways that we'd connected the LLM to our tools and our data, totally pointless once you had function calling and MCP.

所以你就白白浪费了所有的努力。

And so you just you waste all of your effort.

基本上，不要跟模型对着干。

Don't bet against the model, basically.

这就是关键。

That's the key.

另外一件事是，要构建模型学不到的东西，比如，它无法从数百万用户那里知道的东西。

And then the other thing is build what the model can't learn, like, can't know by the millions of users they have.

所以那非常具有机构特定性。

And so that's very institutional specific.

它不知道什么？

What doesn't it know?

它不知道我们的权限。

Doesn't know our entitlements.

它不知道我们的任务。

Doesn't know our mandates.

它不知道我们如何连接数据等等。

Doesn't know the way that we connect our data together, etcetera.

因为其他一切都会被模型吸收，而你会白费功夫。

Because everything else is gonna get consumed by the model and you'd have wasted your time.

这太有意思了。

It's fascinating.

正如你所说，在基础层不断变化、规则和架构都在变动的环境中构建这些系统，确实很不容易。

As you say, building these systems in an environment where the foundation layers is shifting a lot, the regimes, the architectures are shifting.

我觉得你在应对这些变化方面做得非常好，同时始终牢记一些基本原则，特别是你提到的用户意图和体验。

I think you've done a very good job of rolling with those changes, but also keeping your mind on some first principles around what you want to build and in particular, as you said, user intent and experience.

也许可以稍微倒推一下，看看你试图在系统中融入的那些用户体验基本原则是什么，而周围的环境却在不断变化。

Maybe work backwards a little bit from what are those first principles of user experience that you're trying to incorporate in the system, all while the playing field shifts around you.

是的，我认为这在一定程度上关乎如何确保信任和透明度，因为信任，我们凭直觉都知道，需要多年才能建立，却很容易失去。

Yeah, I think it's about it's partly about how do you ensure trust and transparency because trust, we all know intuitively, is something that takes years to build and is lost very quickly.

所以当你考虑这些生成式AI能力时，你可以生成各种各样的答案。

And so when you think about these Gen AI capabilities, you can generate all sorts of answers.

那么问题来了，如何生成一个人们愿意为之负责的答案？

So the question is, how do you generate an answer that someone will stand behind?

你知道，你必须非常深入地思考像高盛这样的公司到底在做什么。

You know, you have to think very deeply about what a firm like Goldman Sachs does.

做什么。

Does.

对吧？

Right?

你不仅仅是生成建议、分析或可能的情景。

You don't just generate advice or analytics or possible scenarios.

我们会为其中一个方案负责。

We stand behind one of those Mhmm.

在高度不确定的时期表明立场。

And take a view in highly uncertain time.

所以问题就变成了，我们如何构建系统，让我们的专业人才能够充分理解并支持AI生成的分析？

And so then the question becomes, how do you build in such a way that our human capital can fully understand and stand behind the analysis that was generated by AI?

这必须在设计阶段就完成。

And that has to be done in the design phase.

也就是说，必须深思熟虑地去做。

As in, it has to be done thoughtfully.

因为我认为这里有一个巨大的风险，在AI时代，我们把推理外包给这些模型，导致认知退化，使我们无法自己从基本原理出发进行推理。

Because I think there's a huge danger here that in the era of AI, we outsource our reasoning to these models, and we have cognitive atrophy that stops us being able to reason for first principles ourselves.

我觉得这确实是个很大的风险。

I think there's a real danger of that.

所以如果你回顾技术发展的历史，我们总是会制造一些外部的工具，这些工具在某些能力上比我们做得更好，然后这些技能就会退化，最后我们甚至忘了自己曾经拥有过这些能力。

So if you think about the history of technology, we have always built tools that are external to ourselves that do a capability that we had better than us, and then those skills atrophy, and then eventually we forget we even had them.

对吧？

Right?

比如说，想想方向感，在草原上跑个几百上千英里，还能找到回来的路。

So think about, like, way finding, running thousand hundreds of miles on the savannah and finding your way back.

对吧？

Right?

这种能力已经完全丧失了。

Totally lost that capability.

还有记忆。

Remembering things.

比如，吟游诗人以前能背诵整本书。

Like, Bards used to be able to recite entire books.

不用了。

No.

你不用再靠记忆去记那些东西了。

You don't need to remember anymore.

直接看书就行。

You just read it in a book.

这就是印刷术带来的改变。

That was a printing press.

对吧？

Right?

然后你就说，好吧。

Then you said, okay.

那时候我在大学里，还得去图书馆找教材来获取信息。

Well, now even when I was at university, I had to find textbooks for information.

现在呢，你根本不用出门，因为信息再也不缺了。

Now, oh, you don't even need to go anywhere or information's not scarce anymore.

所以你不必记住那些，但推理仍然很重要。

So you don't need to remember that, but reasoning is still important.

你仍然需要推理，并将其组织成论点。

You still need to reason about and structure it into an argument.

而现在我们正在把推理也外包出去。

And now we're delegating reasoning.

所以我觉得我的北极星是，我为公司打造产品时，深感责任重大，希望十年后，新加入公司的同事，他们的朋友会对他们说，天哪。

And so I think that my North Star here is I feel a great responsibility building products for the firm such that in ten years' time, a new joiner to the firm, their friends say to them, oh my goodness.

你能处理那么多推理？

You can handle that much reasoning?

这就像当战斗机飞行员和当公交车司机的区别。

It's like getting a job as a fighter pilot versus a a bus driver.

就像，哇。

It's like, wow.

就像，有很多信息向你涌来，但你一定非常聪明，非常擅长在不确定中推理，才能应付得来。

Like, there's a lot coming at you, but you must be really smart and really good at reasoning under uncertainty to be able to handle that.

所以我觉得这必须是一个主动的选择，去思考如何增强人类的推理能力，而不是把它外包出去。

And so I think that has to be a proactive choice to think about how do we empower human reasoning, not delegate it.

有意思。

Fascinating.

所以你的意思是，如果技术像历史上那样发展，我们就会失去这种能力。

So you're saying that if technology evolves as it has historically, we would lose it.

所以我们真的必须特别强调，比如——

So we actually have to really emphasize Like

在田间地头工作。

working in the fields.

对吧？

Right?

对。

Mhmm.

比如我早上去Barrie's健身。

Like I go to Barrie's in the morning.

你知道，我们得建健身房，因为我一天在办公室坐十二个小时。

You know, like we have to build gyms because I sit in an office for twelve hours a day.

对吧？

Right?

对。

Right.

所以你需要主动去做，我觉得这又回到了学徒文化，我们必须主动推动它。

So you need to actively and so I think that's comes back to the apprenticeship culture that we have to proactively push.

我觉得这不会自然而然地发生，因为初级员工学习的方式很大程度上就是这种学徒文化。

I don't think that's gonna happen by accident because a lot of the way that juniors learn is this apprenticeship culture.

所有这些都让我作为一个初级交易员，看到聊天里的价格请求，要弄清楚那是什么意思，怎么把它输入定价器，或者我要回什么价格，等等。

All of these things take me as a junior trader who is seeing that price request on a chat, working out what does that even mean, how do I put it into the pricer, or what price am I gonna put back, etcetera.

所以我们完全可以自动化这个过程。

So we can absolutely automate that.

但那样我们还能得到真正理解一切的资深交易员吗？

But then do we get the senior traders that fully understand?

当时围绕如何定价，引发了很多讨论。

There was a lot of conversations that spun off the back of how to price that.

所以你是通过实践来学习的，而且很多知识都是隐性的。

So you learn by doing, and a lot of knowledge is tacit.

这些知识从来没有被写下来过。

It was never written down.

我觉得这正是我们需要深入探讨的地方，要确保我们不会失去现在一些最优秀的人所拥有的那种隐性直觉知识，同时也要确保下一代也能掌握这些知识。

And I think that's where we really need to double click on is to make sure we don't lose that tacit and intuitive knowledge that some of our best people have today, but to ensure the next generation have it too.

那么，Chris，你对这个平台未来的期望或路线图是什么？

So, Chris, what's on your wish list going forward or your roadmap for this platform?

我们已经谈到了底层基础技术的快速演进和改进。

We've talked about the rapid evolution and improvement in fundamental technology underneath this.

随着这些技术的发展，你还想给平台增加哪些功能和特性？

As that plays out, what additional features and qualities do you want to add to the platform?

我觉得在考虑AI创新时，有两种心态很重要，需要区分开来，而大多数人用的是第一种。

So I'd say there's sort of two mindsets that I think are important to distinguish when you think about innovation AI, and most people use the the first one.

所以有一种是自动化思维，另一种是重新构想思维。

So there's an automation mindset, and then there's a reconception mindset.

目前大部分工作都是自动化思维。

And most of the work at the moment automation mindset.

那什么是自动化思维呢？

So what's an automation mindset?

就是拿一个现有流程，找出其中的认知瓶颈，比如阅读、审核这些由人完成的部分。

It's you take an existing process, you look at the cognitive bottlenecks in that process, whether it's reading, reviewing, etcetera, that are done by human.

用AI替换掉这些环节，然后得到同样的结果，只是更快。

You replace that with AI, and then you get the same answer, just faster.

对吧？

Right?

这样做风险在于会固化旧有模式。

The risk with that is you lock in legacy.

好处是能带来一些效率提升。

The benefits of that is you get some efficiency gains.

所以现在大多数人都在做这件事。

So most people are doing that right now.

然后就是重新构想的心态，这才是能带来数量级提升的地方，你要忘掉那些旧时代的遗留流程。

Then there's the reconception mindset, and this is where you get orders of magnitude improvements, where you forget about the existing processes that were fossils of a previous age.

然后你会说，好。

And you say, okay.

这里真正要做的事情是什么？

What's the real job to be done here?

对吧？

Right?

要做到这一点，你真的需要有深厚的领域专业知识。

What's the real you have to have deep domain expertise to do that.

然后你会说，好。

And then you say, okay.

假设智能是极其可扩展、灵活且丰富的。

Let's assume intelligence is incredibly scalable and elastic and abundant.

好。

K.

有哪些问题，我们以前因为人力有限或协调困难而无法回答，现在却能回答了？

What questions could we now answer that we would never be able to answer before because of limited human bandwidth or coordination, etcetera?

让我们重新构想一切可能。

And let's reconceive what's possible.

那你现在在忙些什么？能举几个例子吗？

So what are some examples of things you're working on right now?

首先是能力方面的视角。

So there's a capability lens.

在人工智能时代，我们有哪些能力可以变得更强大？

So what capabilities do we have that can be made much more powerful in the age of AI?

比如，我们有个叫“视觉结构化”的东西，对吧？那是一个直观的可视化界面，比如可以给任何衍生品定价。

For example, we have something called visual structuring, right, which is a intuitive visual interface that can price any derivative, for example.

但它本身并不定价。

But it doesn't price them.

它看的是情景，看它怎么随时间演变，什么时候支付，怎么回测。

It looks at the scenarios, how it's gonna evolve over time, when did it pay out, how does it back test.

所以想想把这个能力包装在一个AI代理里，这个代理完全理解那个工具怎么运作，以及怎么解读输出。

And so think about wrapping that capability in an AI agent that thoroughly understands how that tool works and how to interpret the output.

这样一来，你就万事俱备了，因为瓶颈不再是教用户怎么用工具、怎么解读输出。

Well, now you're away to the races because the bottleneck is not teaching a user how to use the tool and how to interpret its outputs.

你教代理这些，然后用户只需要表达意图就行了。

You teach the agent that, and then the user just has to express intent.

所以，我们怎么在AI时代，把我们公司真正拥有的、独一无二的能力发挥出来，帮助客户在不确定性下做出更好的决策。

So how do we bring to bear the real capabilities that we have as a firm, uniquely uniquely as a firm, in an era of AI to help our clients make better decisions under uncertainty.

这必须是我们的北极星。

That has to be the North Star.

有意思。

Fascinating.

好，那我们也许就到这里结束吧。

Well, maybe we just close.

这是一次非常精彩的讨论。

It's been a fantastic discussion.

我们谈了很多关于技术的局限性、它们如何演变，以及你如何应对这些挑战。

We've talked a lot about the limits of the technology, how they've evolved, how you're navigating them.

也许就提一个非常宽泛的问题。

Maybe just a very broad question.

你是否相信这种技术范式会持续扩展和进步，还是说我们面前会有一个意想不到的天花板？

Are you a believer in the continued scalability and progress of this technology paradigm or do you see some unanticipated glass ceiling ahead of us?

是的。

Yeah.

好问题。

Great question.

我相信。

I am a believer.

是的。

Yeah.

而且我觉得有两个不同的原因。

And I think for two separate reasons.

首先，关于下一个词预测，有一段时间人们都在贬低这项技术。

I think firstly, on the next token prediction, there was a period of time where people were dismissing this technology.

哦，它不过就是预测下一个词而已。

Oh, all it does is predict the next word.

结果发现，这可能是你能给一个系统的最深刻的训练方式之一。

Well, turns out that is one of the most deep training functions you could give something.

比如说，要让我准确预测“艾莉森转向乔治说”的下一个词，我得知道艾莉森是谁，乔治是谁，我们现在处于什么情境，播客里通常会说些什么，等等。

So for example, for me to correctly predict the next word of Allison turned to George and said, I need to know who Allison is, who George is, what situation we're in right now, what typically is said on podcasts, etcetera.

所以，要做好这个任务，我需要建立一个世界模型。

So I need to build a world model to be able to do that job really well.

而且我觉得令人惊讶的是，我们训练这些系统越多，它们实际上在构建合成世界模型，而很多怀疑者曾说这不会发生。

And I think that's been the surprising thing is that the more we train these things, they are actually building synthetic world models, and a lot of the doubters said that that wasn't gonna be the case.

所以我觉得随着规模扩大，第二点是，我确实被震撼到了，如果你回到第一性原理，透过噪音看本质，其实就是2017年的Transformer论文《Attention Is All You Need》，它基本上是说，你可以让上下文窗口中的每个词都关注到其他每个词，这太神奇了。

So I think as we scale it and then the second thing is, I think I've been blown away by if you go back to first principles to see through the noise, it's really the transformer paper in 2017, attention is all you need, which basically said that actually you can have every word pay attention to every other word in a context window, which is amazing.

但是后来卡普兰的缩放定律加上Chinchilla论文基本上告诉你，这些东西，如果你先做个小实验看看是否有效，然后你可以把它放大，它会在更大的规模上有效，最多能放大10个数量级。

But then the Kaplan's scaling laws plus the Chinchilla paper basically told you that these things, if you do a small experiment, then to see whether it works, you can then scale it up, and it will work at a larger scale up to 10 orders of magnitude bigger.

所以这意味着两件事。

And so what that means is two things.

首先，它证明了在基础设施上花费数万亿美元是合理的，因为你可以说那不是一场疯狂的赌注。

Firstly, it justifies spending trillions of dollars on infrastructure because you can say that that's not a wild bet.

这里有一条长期成立的规律。

There's a line here that's held a long time.

所以如果我们做10倍，我们建一个万亿美元的数据中心，我们会预期到这个结果。

So if we do 10 times, we build a trillion dollar data center, we would expect this.

所以你可以向投资者推销这个想法。

So you can pitch that to investors.

但重要的是，在实验室里，他们的首要任务高于一切，就是自动化他们自己，也就是自动化AI研发。

But importantly, in the labs, their number one priority above all else is to automate themselves, which is automate AI r and d.

那么这意味着，AI研发的瓶颈是什么？

And so what that means is, what's the bottleneck in AI r and d?

不是读论文。

It's not reading the papers.

也不是解读结果。

It's not interpreting the results.

绝对不是。

Absolutely not.

而是跑实验。

It's the running of experiments.

所以，如果AI能大规模地反复进行小规模实验，而且扩展定律成立，那就意味着实际上没有瓶颈。

And so if an AI can run actual experiments at small scale many times and scaling laws hold, then it means that there's no actual bottleneck.

这样一来，他们就能加快实验数量，并且知道这些实验会扩展。

And so they can actually expedite the number of experiments they do and know that those will scale.

所以我觉得这就是下一个范式。

And so I think that that's the next paradigm here.

在我看来，最大的缺失部分首先是自我学习。

I think the big missing piece as I look at this is, firstly, self learning.

还没到呢。

It hasn't come yet.

对。

Mhmm.

灾难性遗忘，所以所有这些正在构建的记忆架构，有可能在某个时候被模型吸收进去。

Catastrophic forgetting, and so all of these memory architectures that are being built, it's possible at some point they get enveloped in the model.

第二点是，我觉得我们还没找到合适的形态。

The second thing is I don't think we found the form factor.

我觉得我们现在处于AI的电视发明阶段，就像我们开始在电视上播放广播节目一样。

I think we're in the phase of AI where television's invented, and we start showing radio shows on TV.

我们还没有找到原生的形态。

We haven't got the native form factor.

我们现在就像是在命令行时代，还没到Windows和鼠标的时代。

It's we're on the command line versus Windows and mouse kind of era.

那么，AI的原生形态在哪里呢？

So where's the native form factor of AI?