Blackrock Aladdin Copilot: AI-Powered Investing Revolution | Agentic AI Podcast by lowtouch.ai

欢迎来到深度探讨。

Welcome to the deep dive.

今天我们要探索的，是正在彻底改变大型企业运作方式的事物。

Today we're exploring something that's really starting to shake up how big businesses work.

我们谈论的是人工智能如何远远超越聊天机器人，进入真正智能的自动化领域。

We're talking about AI moving way beyond, you know, chat bots and into really intelligent automation.

这股趋势似乎即将改变几乎所有行业，尤其是那些高度复杂的受监管行业。

It feels like it's poised to transform pretty much every industry, but maybe especially the really complex regulated ones.

当你想到谁在大规模应用人工智能时，黑石集团无疑会浮现在脑海中。

And when you think about who's doing AI at serious scale, well, BlackRock definitely comes to mind.

是的，绝对如此。

Oh, absolutely.

我的意思是，我们讨论的是一家金融巨头。

I mean, we're talking about a financial giant here.

他们管理的资产超过11万亿美元。

They manage over $11,000,000,000,000 in assets.

这个数字真是令人震惊。

That number is just staggering.

11万亿美元。

11,000,000,000,000.

是的。

Yeah.

这确实让人看清了其中的利害关系，不是吗？

It kinda puts the stakes into perspective, doesn't it?

当他们进行技术变革时，这可是大事。

When they make tech changes, it's a big deal.

没错。

Exactly.

风险极高。

The stakes are incredibly high.

他们不是仅仅浅尝辄止地接触AI，而是将其深度融入核心业务中。

They aren't just like dipping their toes in AI, they're embedding it deep within their core operations.

真正基础性的东西。

Really foundational stuff.

好的。

Okay.

那么，这就是我们这次深入探讨的目标了。

So that's our mission for this deep dive then.

我们想要层层剖析他们的系统——Aladdin，以及这个新的工具——Aladdin Copilot。

We want to peel back the layers on their own system, Aladdin, and this newer thing, Aladdin Copilot.

我们将尝试理解他们所使用的这种代理式AI如何提升效率、改善投资成果，甚至为使用它的金融专业人士带来更个性化的体验。

We'll try to understand how this agentic AI they're using boosts efficiency, helps with investment results, and maybe even makes things more personal for the finance pros using it.

当然，我们也会深入探讨背后的技术。

And yeah, we'll get into the tech behind it.

这些流程的架构应该会很有趣。

The architecture of the processes should be interesting.

我们开始吧。

Let's jump in.

要真正理解它，你得从Aladdin本身开始。

So to really get it, you have to start with Aladdin itself.

它不仅仅是软件。

It's, it's more than just software.

它是贝莱德的基础性技术系统。

It's BlackRock's foundational tech system.

把它想象成他们整个投资流程的中枢神经系统。

Think of it like the central nervous system for their entire investment process.

端到端，它将所有内容统一起来。

End to end, it unifies everything.

投资组合构建、风险分析，等等一切。

Portfolio building, risk analysis, you name it.

重要的是要明白，这不仅仅适用于贝莱德的内部团队，对吧？

And it's crucial to understand this isn't just for BlackRock's internal teams, is it?

它的影响范围更大。

It's bigger than that.

说得完全正确。

That's exactly right.

黑石无疑是最大的用户，但阿拉丁还支持各种市场，包括公开和私人市场。

BlackRock is the biggest user for sure, but Aladdin supports a whole range of markets, public, private.

它服务的对象包括机构客户和零售投资者。

It serves institutional clients, retail investors too.

此外，它为数百家客户提供强大的风险管理和分析工具。

Plus, it provides these really robust tools for risk management and analysis to literally hundreds of clients.

我认为现在它已经覆盖了70多个国家。

I think it's in over 70 countries now.

而背后开发它的团队规模非常庞大。

And the team behind it, it's huge.

阿拉丁团队大约有7000人，你猜怎么着？其中超过4000人是工程师。

Around 7,000 people in the Aladdin organization, and get this, more than 4,000 of them are engineers.

4000名工程师。

4,000 engineers.

哇。

Wow.

对。

Yeah.

只是维护和构建这个系统。

Just maintaining and building that.

这本身就是一个生态系统。

That's an ecosystem in itself.

确实如此。

It really is.

他们一直在开发一百多个前端应用程序，成千上万的专业人士每天都在使用。

They're constantly working on over a 100 different front end applications that thousands of professionals use every day.

明白了。

Okay.

所以你拥有这样一个庞大、复杂且成熟的系统。

So you have this massive, complex, established system.

你如何将人工智能融入其中？

How do you weave AI into that?

它必须不仅仅是让事情变得快一点而已。

It has to be more than just, I don't know, making things a bit quicker.

哦，远不止如此。

Oh, much more.

这就是Aladdin Copilot发挥作用的地方。

That's where Aladdin Copilot enters the picture.

你可以把它看作是一个智能层，几乎像一层连接所有这些Aladdin应用的纽带。

You can think of it as this, intelligent layer, almost like connective tissue that sits across all those Aladdin apps.

它的设计是主动的。

It's designed to be proactive.

因此它会在你提问之前就呈现相关信息。

So it surfaces information before you ask.

它优化了工作流程，让一切更顺畅。

It enhances workflows, makes things smoother.

最终帮助投资专业人士做出更快、更明智的决策。

Ultimately, helps investment pros make faster, more informed decisions.

贝莱德对AI在这里要实现的目标非常明确。

And BlackRock is super clear about what they want AI to achieve here.

他们为这次企业级AI推进设定了三个关键业务成果。

They have targeted three key business outcomes for this enterprise AI push.

三个具体成果。

Three specific outcomes.

好的。

Okay.

是哪三个？

What are they?

对。

Right.

首先，提升用户生产力。

First, boost user productivity.

简化任务，赋能系统使用者，让他们的生活更轻松， basically。

Streamline tasks, empower people using the system, make their lives easier, basically.

第二，实现他们所说的阿尔法生成。

Second, enable what they call alpha generation.

一个高大上的金融术语，但其实就是推动更好的投资回报，超越基准。

Fancy finance term, but it just means driving better investment results, beating the benchmark.

第三，提供更个性化的用户体验，根据每个用户的实际需求定制平台。

And third, provide a much more personalized user experience, tailoring the platform to what each individual user actually needs.

这充分展示了人工智能如何改变企业环境中的游戏规则。

This really shows how AI can change the game in enterprise settings.

生产力、更好的结果、个性化。

Productivity, better results, personalization.

是的，这些都很有道理。

Yeah, those make a lot of sense.

我猜，对于任何使用人工智能的公司来说，这些都是非常根本的收益。

Pretty fundamental wins for any company using AI, I'd imagine.

但这些宏大的目标如何转化为实际构建Copilot的指导原则呢？

But how do those big translate into like the guiding principles for actually building Copilot?

他们日常真正优化的是什么？

What are they really optimizing for day to day?

好问题。

Good question.

他们已经将Copilot的核心价值驱动因素分解为三项。

They've broken it down into three core value drivers for Copilot.

这些都直接关联到业务价值和效率。

Things that directly link back to business value and efficiency.

第一个是，我很喜欢这种说法：让每个用户都成为阿拉丁专家。

The first one is, and I like this framing, make every user an Aladdin expert.

通过简化操作，让每个人都能成为专家。

Make everyone an expert tell by simplifying things.

没错。

Exactly.

简化复杂的内容，让整个平台更加直观。

Simplify the complex stuff, make the whole platform more intuitive.

他们希望缩短学习曲线，你知道的，不需要几个月才能上手。

They wanna flatten that learning curve, you know, so you don't need months to get up to speed.

AI会引导你，推荐内容，并提供上下文信息。

The AI guides you, suggests things, gives you context.

因此，即使你是新手，Copilot也能帮你像用了多年一样自如操作。

So even if you're newer, Copilot helps you navigate it like you've been using it for years.

好的。

Okay.

让专业知识大众化是有道理的。

Democratizing expertise makes sense.

第二个原则是什么？

What's the second principle?

第二个是个性化平台。

Second is personalize the platform.

提供可配置、个性化的体验，释放专属于你和你角色的新效率。

Deliver experiences that are configurable, customized, unlock new efficiencies that are specific to you and your role.

想象一下，平台能自动适应你的工作方式，在你不需手动设置的情况下，为你展示最相关的内容。

Imagine the platform just kind of adapting to how you work, showing you what's most relevant without you needing manually set everything up.

这听起来非常强大。

That sounds pretty powerful.

是的。

Mhmm.

那第三个驱动力是什么？

And the third driver.

第三个是让洞察信息更加普及。

Third is democratize access to insights.

基本上，让每个使用平台的人都能更轻松地在需要时找到正确的数据和洞察。

Basically, it way easier for everyone using the platform to find the right beta, the right insights right when they need them.

再也不用翻阅无尽的菜单或费力理解复杂的查询了。

No more digging through endless menus or trying to figure out complex queries.

它让更多的人都能以数据为驱动。

It empowers more people to be data driven.

但你知道，这引出了一个问题。

But, you know, this begs the question.

对吧？

Right?

你究竟该如何构建一个足够智能的AI来处理所有这些？

How on earth do you build an AI smart enough to handle all that?

所有这些应用程序，所有这些数据？

All those apps, all that data?

是的。

Yes.

怎么做。

The how.

这就涉及到核心引擎了。

This is where we get into the engine room.

对吧？

Right?

这种支持Aladdin Copilot的智能体架构。

This agentic architecture that powers Aladdin Copilot.

听起来显然不仅仅是简单地接入一个大型AI模型。

It sounds like it's definitely not just plugging in one big AI model.

是的。

Yeah.

听起来更加复杂。

It sounds, more sophisticated.

确实如此。

It really is.

这种智能体架构专为扩展而设计，能够处理所有那些复杂的Aladdin应用以及海量的数据。

This agentic architecture, it's designed specifically to scale, to handle all those different complex Aladdin applications and the absolutely vast data sets involved.

把它想象成一个中央协调者，协调者。

Think of it like a a central coordinator, coordinator.

一个能够智能管理和调度整个专用工具或代理网络的中枢大脑。

A brain that can intelligently manage and direct a whole network of specialized tools or agents.

每个代理都是Aladdin中某个特定领域的专家。

Each one is an expert in its own little domain within Aladdin.

一个专用工具的网络。

A network of specialized tools.

是的。

Yeah.

好的。

Okay.

是的。

Yeah.

但你之前提到，有大约50到60个不同的工程团队在开发Aladdin。

But you mentioned earlier there are like 50 or 60 different engineering teams working on Aladdin.

交易、风险、合规，都是独立的领域。

Trading, risk, compliance, all separate domains.

你们如何让所有这些工具在Copilot中协同工作？

How do you get all their tools to play nicely together within Copilot?

这听起来像是集成灾难的根源。

That sounds like a recipe for an integration nightmare.

确实有可能。

It could be.

但他们有一个相当巧妙的解决方案——一个插件注册表。

But they have a pretty clever solution, a plug in registry.

这个注册表非常关键。

This registry is key.

它为所有这些专业的工程团队提供了一种标准化的方式，将他们的工具和代理接入Copilot的AI编排系统。

It basically provides a standard way for all those specialized engineering teams to onboard their tools, their agents into Copilot's AI orchestration system.

它实现了无缝协作。

It enables that seamless collaboration.

团队可以通过几种方式来实现这一点。

Teams can do this in a couple of ways.

他们可以将工具映射到现有的Aladdin API，或者对于更复杂的情况，可以构建自定义代理。

They can either map their tool to existing Aladdin APIs or for more complex stuff, they can build custom agents.

这很务实。

It's pragmatic.

它具有可扩展性。

It scales.

他们是否也在自己的围墙之外寻找解决方案？

And are they looking outside their own walls for this?

他们是否考虑过行业标准，用于这些代理之间的通信？

Are they considering like industry standards for how these agents talk to each other?

是的，他们在考虑。

They are, yeah.

最初，他们为这个代理框架构建了自己的通信协议，但现在他们正在积极研究和评估开放标准。

Initially they built their own communication protocol for this agent framework, but now they're actively looking at and evaluating open standards.

像LangChain代理协议这样的东西已经在他们的关注范围内。

Things like langchains agent protocol are on their radar.

这表明他们在考虑互操作性，而不仅仅是把自己局限在专有技术中。

Which, you know, shows they're thinking ahead about interoperability and not just locking themselves into proprietary tech.

对于企业级AI来说，这种前瞻性非常出色。

It's quite forward thinking for enterprise AI.

好吧，这让我们对架构、插件注册中心和代理有了更清晰的认识。

Okay, that gives us a clearer picture of the architecture, the plug in registry, the agents.

但让我们具体化一下。

But let's make it concrete.

为了让所有听众都能想象出这个系统实际运行的情景，能否为我们演示一下当用户提出一个简单问题时会发生什么？

For everyone listening, trying to imagine this in action, maybe walk us through what happens when a user asks a simple question.

这正是AI必须真正证明自己的地方。

This is where the AI really has to prove itself right.

当然。

Definitely.

我们以一个典型的用户查询为例。

Let's take a typical user query.

想象有人向Copilot提问，可能是打字或口头问：投资组合一中我在航空航天领域的风险是多少？

Imagine someone asks Copilot, maybe types it in or says it, what is my exposure to aerospace in portfolio one?

看起来挺简单的。

Seems simple enough.

好的，简单的问题。

Okay, simple question.

Copilot第一步会做什么？

What's the very first thing Copilot does?

第一步。

First step.

上下文感知。

Context awareness.

非常重要。

Super important.

Copilot会立即获取大量上下文信息。

Copilot immediately grabs a whole bunch of context.

用户当前在哪个应用程序中？

What application is the user in?

他们屏幕上现在显示的是什么？

What's on their screen right now?

是否有特定的投资组合或资产可见？

Are there specific portfolios or assets visible?

他们的普遍偏好是什么？

What are their general preferences?

所有这些都会被捕捉到。

All that gets captured.

对。

Right.

因为‘风险敞口’可能根据你在平台中的位置而有不同的含义。

Because exposure could mean different things depending on where you are in the platform.

是的。

Yeah.

好的。

Okay.

上下文已捕获。

Context captured.

接下来呢？

What's next?

安全性一定是个大问题。

Security must be a big concern.

非常重要。

Huge.

接下来，查询会进入他们所说的输入防护机制。

So next, the query hits what they call input guardrails.

这些实际上是用 LangChain 构建的。

These are built using lang chain, actually.

它们的作用是在问题深入之前过滤掉任何有害内容。

Their job is to filter out anything problematic before it goes deeper.

无关内容、有害语言、敏感数据，比如个人身份信息（PII）。

Irrelevant stuff, toxic language, sensitive data like PII, personally identifiable information.

必须把这些都过滤掉。

Gotta keep that out.

这是至关重要的第一道防线。

It's a critical first line of defense.

有道理。

Makes sense.

先清理输入内容。

Clean the input first.

对。

Yeah.

那么，系统是如何确定在注册表中成千上万个工具或代理中，哪个应该回答航空航天相关问题的呢？

Now how does it figure out which of those, you know, potentially thousands of tools or agents in the registry needs to answer the aerospace question?

这就是过滤和访问控制阶段。

That's the filtering and access control stage.

它会将这个庞大的工具和代理列表快速缩小，通常只剩下大约20到30个相关的。

It takes that huge list of possible tools and agents and narrows it down fast, usually down to maybe 20 or 30 relevant ones.

这种过滤会利用我们刚才提到的上下文，比如应用程序和屏幕内容，但关键还取决于用户的权限。

This filtering uses the context we just talked about, the app, the screen, but also critically the user's permissions.

你只能使用你被授权的工具。

You only get tools you're allowed to use.

啊，所以它既高效又安全。

Ah, so it's efficient and secure.

它很智能。

It's smart.

明白了。

Okay.

现在我们有了相关的工具。

Now we have the relevant tools.

这应该就是真正AI思考发生的地方。

This must be where the real AI thinking happens.

协调部分。

The orchestration part.

没错。

Exactly.

这是核心的智能协调与规划。

This is the core intelligence orchestration and planning.

Copilot 在这里使用了 GBT-4 的函数调用能力。

Copilot uses GBT four's function calling capability here.

它不仅仅选择一个工具。

It doesn't just pick one tool.

它会进行规划。

It plans.

它会确定所需的操作顺序，哪个工具应该先运行，下一个工具需要从上一个工具获取什么信息？

It figures out the sequence of actions needed, which tool should run first, what information does the next tool need from the previous one?

它会迭代地规划并执行这些步骤。

It iteratively plans and executes these steps.

所以它是动态地为回答特定问题而即时构建一个小工作流。

So it's like dynamically building a little workflow on the fly to answer the specific question.

正是如此。

Precisely.

它会从过滤后的工具集中精心选择合适的工具，并找出最佳方式组合它们的输出以获得准确的答案。

It carefully selects the right tools from that filtered set and figures out the best way to combine their outputs to get an accurate answer.

好的。

Okay.

它已经有了计划。

It's got a plan.

它已经执行了。

It's executed it.

在向用户展示答案之前，有没有最后的检查环节？

Before it shows the answer to the user, is there a final check?

你经常听说AI会编造内容。

You hear a lot about AI making things up.

幻觉。

Hallucinations.

是的。

Yep.

最后一步。

Final step.

输出防护机制，另一重安全网。

Output guardrails, another safety net.

这些检查专门设计用于发现并减少潜在的幻觉。

These checks are specifically designed to catch and reduce potential hallucinations.

他们希望确保最终答案准确可靠，在金融领域至关重要，因为一个错误的数字可能带来巨大后果。

They want to ensure the final answer is accurate, reliable, absolutely vital in finance, where a wrong number could have massive consequences.

当然。

Absolutely.

最后，用户得到了他们的答案。

And then finally, the user gets their answer.

简洁明了。

Nice and simple.

比如，你的投资组合一中对航空航天和国防领域的敞口是5%。

Something like, your exposure to aerospace and defense in portfolio one is 5%.

对用户来说无缝顺畅，但天啊，下面隐藏着一套复杂的流程在高效运转。

Seamless for the user, but wow, there's a whole sophisticated process humming away underneath.

而整个流程让我们回到了任何企业领导者思考企业级AI时最基本的问题。

And that whole process brings us right back something fundamental for any business leader thinking about enterprise AI.

信任。

Trust.

你必须能够信任这个系统。

You have to be able to trust the system.

佩德罗·维森特·瓦尔德兹，黑石公司主要的AI工程师之一，特别强调了这一点。

Pedro Vicente Valdez, he's one of BlackRock's principal AI engineers, really stressed this.

他说，以评估驱动的开发至关重要。

He said evaluation driven development is just essential.

在黑石这样的规模下，可靠性至关重要。

Critical for reliability at the kind of scale BlackRock operates at.

你不能只是构建这些系统，然后指望它们能正常工作。

You can't just build these things and hope they work.

你必须不断证明它们有效，并且在持续改进。

You have to constantly prove they work and that they're getting better.

信任至关重要，尤其是在金融领域，这是不可妥协的。

Trust is everything, especially in finance, It's non negotiable.

因此，说到这里，为什么他们目前使用的是监督式AI代理，而不是完全自主、彼此对话的AI系统？

So on that note, why are they currently using supervised AI agents, not fully autonomous ones talking to each other?

这显然是一个有意的选择。

Seems like a deliberate choice.

是的。

It is.

这是一个非常明确且我认为极具洞察力的决定，符合当前企业AI的发展阶段。

It's a very deliberate and I think insightful decision for where enterprise AI is right now.

专注于有人类监督和控制的受监管代理，而不是完全自主的代理之间直接对话，这为他们带来了几个重大优势。

Focusing on supervised agents where there's more human oversight and control in the loop rather than fully autonomous agent to agent chats gives them several big advantages.

这意味着更快的开发周期。

It means faster development cycles.

它使测试变得容易得多。

It makes testing much easier.

而且至关重要的是，它赋予了他们更多的控制权，这是至高无上的。

And crucially, it gives them more control, which is paramount.

所以这就像一个实用的过渡步骤。

So it's like a practical stepping stone.

现在获得好处的同时保持控制，未来或许再逐步走向更高程度的自主性。

Get the benefits now while maintaining control, maybe move towards more autonomy later.

没错。

Exactly.

完全自主可能是最终的愿景，也许吧，但受监管的代理是如今实现这一目标更实际、风险更低的方式。

Full autonomy might be the eventual vision, maybe, but supervised agents are the practical lower risk way to get there today.

它让他们能够快速迭代，创造价值，同时保持监管领域所必需的严格治理控制。

It lets them iterate quickly, build value, but keep those ironclad governance controls needed in regulated fields.

这对企业来说是一种明智的平衡。

It's a smart balance for an enterprise.

这很有道理，尤其是在建立内部和客户信心方面。

That makes a ton of sense, especially for building confidence internally and with clients.

那么，这种严格的测试究竟是什么样的呢？

So what does this rigorous testing actually look like?

他们是如何建立这种信心的？

How do they build that confidence?

哦，这非常全面。

Oh, it's extensive.

以系统提示为例。

Take system prompts, for example.

他们可能会内置一条核心规则，比如：绝不能提供投资建议。

They might have a core rule baked in like you must never provide investment advice.

简单的规则。

Simple rule.

对吧？

Right?

但他们需要以高度的统计置信度证明Copilot始终遵守这一规则。

But they need to prove with high statistical confidence that Copilot always follows that rule.

不容例外。

No exceptions.

他们如何测试这一点？

How do they test that?

你不能只是问它几次。

You can't just ask it a few times.

绝对不行。

No way.

是的。

Yeah.

他们生成大量专门设计用来欺骗它的合成测试数据。

They generate lots of synthetic test data specifically designed to trick it.

他们让人类领域专家审阅大量示例。

They have human subject matter experts review tons of examples.

他们构建了自动化测试流水线。

They build automated testing pipelines.

他们甚至使用其他强大的AI模型（大语言模型）作为公正的评审者，以大规模验证响应结果。

They even use other powerful AI models, LLMs, as sort of impartial judges to validate the responses at scale.

这些测试不是只运行一次，而是每天都会运行。

And these tests aren't just run once, they run daily.

每次代码变更时，它们都会作为标准CI/CD流水线（持续集成、持续部署）的一部分运行。

They run with every single code change as part of their standard CICD pipeline, continuous integration, continuous deployment.

这是一种持续的警惕。

It's constant vigilance.

哇。

Wow.

Bates 直接融入开发流程，持续验证。

Bates right into the development process, Constant validation.

那更复杂的情况呢？

What about more complex stuff?

比如不同代理之间的交互，或者在对话中保持上下文记忆。

Like how the different agents interact or remembering context over conversation.

是的。

Yep.

他们也会对此进行严格测试。

They test that rigorously too.

这被称为代理交互的端到端测试。

It's called end to end testing of agent interactions.

他们会检查所有方面。

They check everything.

AI 能否正确填写 API 参数？

Does the AI fill out API parameters correctly?

代理之间能否准确地传递信息？

Do agents pass information between each other accurately?

它如何处理跨越多个回合的来回对话？

How does it handle conversations that go back and forth over several turns?

它能否尊重应用上下文，记住你正在查看的投资组合、你的设置和屏幕上的小部件？

Does it respect the application context, remember which portfolio you're looking at, your settings, widgets on the screen?

开发者为每个插件设置特定的测试场景，提供正确答案，并持续运行这些测试以发现任何回归问题。

Developers set up specific test scenarios for each plugin, provide the correct answers, and run these tests constantly to catch any regressions.

这种细节水平。

That level of detail.

这正是决策者需要听到的，对吧？

It's what decision makers need to hear, isn't it?

尤其是当他们考虑将AI应用于自己的关键业务时。

Especially if they're considering AI for their own critical operations.

为什么这种以评估为导向的方法对企业级AI如此重要？

Why is this evaluation driven approach so vital for enterprise AI?

这其实是基础性的。

It's foundational really.

有几个关键原因。

For a few key reasons.

首先，它能防止随着新代码的加入，系统逐渐退化，保持高质量。

First, it stops the system from slowly degrading as new code gets added, keeps quality high.

其次，它能提供确凿的统计证据，证明系统确实在改进。

Second, it gives you hard statistical evidence that the system is actually improving.

这对展示价值、证明投资回报率至关重要。

That's crucial for showing value, proving ROI.

第三，如果真的出现问题，如果有回归错误发生，这种测试能帮助工程师快速发现并修复，最大限度减少错误和停机时间，而这对于任何采用AI的企业来说可能最为重要。

Third, if something does break, if a regression slips in, this testing helps engineers find and fix it super quickly, minimizes errors, minimizes downtime, and maybe most importantly for any business adopting AI.

它确保了你能可靠地向用户或客户交付可扩展、生产级别的AI，且能保证其准确性和合规性。

It's what ensures you can reliably deliver scalable, production grade AI to your users or clients knowing it's accurate and compliant.

因此，这种持续的评估循环确实是大规模构建和维持信任的关键所在。

So that continuous evaluation loop is really the secret sauce for building and maintaining trust at scale.

我觉得是的。

I think so.

佩德罗·维森特·瓦尔德斯的那句引言完美地总结了这一点。

And that quote from Pedro Vicente Valdez sums it up perfectly.

最重要的经验是，将这种端到端的评估机制嵌入到你的CI/CD流水线中，才能让你有信心持续改进系统。

The biggest lesson is that having this kind of end to end evaluation baked into your CICD pipeline is what allows you to confidently improve the system.

信心来自于实证。

Confidence comes from proof.

真是一次令人着迷的内部透视。

What a fascinating look under the hood.

我们仿佛深入探索了当今企业级AI自动化是如何真正构建的。

It really feels like we've journeyed into the heart of how serious enterprise AI automation is being built today.

是的。

Yeah.

简要总结一下，贝莱德的Aladdin Copilot就是一个典型的代理式AI案例，已经为数千名金融专业人士带来了切实影响。

To quickly recap, BlackRock's Aladdin Copilot, it's a prime example of agentic AI already making a difference for thousands of finance professionals.

它的成功真正依赖于强大的智能体架构、灵活的插件系统，以及最重要的是对以评估驱动开发的不懈承诺。

Its success really hinges on that strong agentic architecture, the flexible plug in system, and maybe above all that relentless commitment to evaluation driven development.

这赋予了它们企业级的可靠性、准确性和合规性。

That's what gives them enterprise grade reliability, accuracy, and compliance.

所以对于所有正在聆听的朋友们，无论你们是深入AI管理团队，还是在做战略决策，都要思考一下这带来的连锁影响。

So for all of you listening, whether you're deep into AI managing teams or making strategic decisions, think about the ripples from this.

黑石的这种蓝图——对精准、效率和严格验证的专注——对你们的业务意味着什么？

What could BlackRock's blueprint, this focus on precision efficiency and just hardcore validation mean for your business?

它能否在你们的行业中解锁类似的AI机遇？

Could it unlock similar AI opportunities in your industry?

这是一个强大的模式。

It's a powerful model.

是的。

Yeah.

或许最后再留一个值得深思的想法。

And maybe a final thought to chew on.

如果一个为全球金融的庞大规模和复杂性而构建的系统能达到如此高的准确性和受控自主性。

If a system built for the sheer scale and complexity of global finance can reach this level of accuracy and controlled autonomy.

是的。

Yeah.

这种具有细致评估的智能体AI模型，能为你独特的挑战带来哪些突破？

What could this agentic AI model paired with that meticulous evaluation unlock for your unique challenges?

如果你从第一天起就建立对AI的信任，会开启哪些新的可能性？

What new possibilities open up if you build that trust in AI from day one?

确实值得深思。

Definitely something to think about.

感谢您参与我们的深度探讨，我们下次再见。

Thank you for joining us on a deep dive until next time.