跳到主内容
@wquguru
精选88OpenAI(YouTube)技巧与观点

Stack Overflow VP Alex Lotto分享Codex工程实践

Stack Overflow - From Idea to Product | DevDay 2026

原文
发到 X
推荐理由

一线大厂VP亲述Codex在企业级复杂场景下的具体用法与取舍,特别是“原型抽象”思路极具复用价值,建议做AI应用落地的同学重点参考。

Hello everyone. Hope you're having a beautiful morning. On the stage I'm super excited. We're gonna have one of our customers Stack Overflow talk about how they're using Codex intensively. So we're gonna have the VP of product Alex Lotto talk about different initiatives, share some practical tips and insights from their use of Codex. And then lastly we're gonna talk about how they're thinking through building in an agent first world as you have not just humans but agents interacting with Stack Overflow.

大家好。希望大家有一个美好的早晨。我现在在台上非常兴奋。我们将邀请我们的客户 Stack Overflow 谈谈他们如何密集使用 Codex。因此,我们将请产品副总裁 Alex Lotto 介绍不同的举措,分享他们使用 Codex 的一些实用技巧和见解。最后,我们将讨论他们如何思考构建一个以 Agent 优先的世界,因为在这个世界中,不仅有人类,还有 Agent 与 Stack Overflow 进行交互。

So with that please join me in welcoming Alex to the stage. Thank you Alex. Thank you so much for being here. Thank you. I thought California was hot. I don't know what's going on here. Freezing. Sunny but cold always. So why don't you start with telling us a little bit a little bit about who you are and what you do at Stack Overflow. Yeah absolutely. My name is Alex. Nice to meet everyone. Originally from Canada so you would think I could handle the cold.

那么,请大家和我一起欢迎 Alex 上台。谢谢 Alex。非常感谢你来到这里。谢谢。我以为加州很热。我不知道这里发生了什么。冷得要命。阳光明媚但总是很冷。那么,你能先告诉我们一些关于你是谁以及你在 Stack Overflow 做什么的信息吗?当然可以。我叫 Alex。很高兴见到大家。我来自加拿大,所以你们可能认为我能适应寒冷。

Apparently not. Very terrible Canadian. I work at Stack Overflow leading product so that's across all our portfolio from the very famous public platform that maybe some of you hopefully know and have used in your careers once upon a time. All the way to our two new products Stack Internal which we'll talk about a little bit today and Stack Overflow for agents. Also a very cool product. Amazing. Thank you. So Alex you talked about the two initiatives but can you tell me a little more about these initiatives and and what you use Codex for?

显然不行。真是个糟糕的加拿大人。我在 Stack Overflow 负责领导产品工作,涵盖我们整个产品线,从非常著名的公共平台(希望有些人曾经在自己的职业生涯中知道并使用过它),一直到我们的两款新产品:Stack Internal(我们今天会稍微谈一下)和 Stack Overflow for agents(也非常酷的产品)。太棒了。谢谢。Alex,你提到了两个举措,但你能再多讲讲这些举措以及你们使用 Codex 的目的吗?

Yeah totally. So two ends of the spectrum is what I'd say. Stack Internal is our new enterprise product. So that's essentially a enterprise grade right. Has to be ready for kind of these large deployments and that's a knowledge intelligence product. So we'll talk about that in a little bit. But on the other end of the spectrum you have Stack Overflow for agents which is really a small experimental product that's more individual facing and obviously agent first.

完全正确。我想说的是光谱的两端。Stack Internal 是我们的新企业级产品。这基本上是企业级的。必须准备好应对大规模部署,它是一个知识智能产品。所以我们稍后会谈到这一点。但在光谱的另一端是 Stack Overflow for agents,这是一个小型的实验性产品,更面向个人,显然是以 Agent 优先的。

So two completely different products both used Codex extensively for different parts of that SDLC. You mentioned these two very different products. How did your use of Codex differ and maybe how you set up the projects as you were approaching them? Yeah totally. So as a product person I like to sometimes pretend that I can be an engineer which is always a very dangerous thing to do. I say this with no engineers on my team around so I feel a little safer.

所以这是两款完全不同的产品,都在软件开发生命周期(SDLC)的不同部分广泛使用了 Codex。你提到了这两款截然不同的产品。你们使用 Codex 的方式有何不同,以及在接近项目时如何设置项目?完全正确。作为一个产品人员,我有时喜欢假装自己可以是工程师,但这始终是一件非常危险的事情。我说这话是因为我的团队里没有工程师在场,所以我感觉稍微安全一点。

But essentially for the Stack Overflow for agents which I'll call SOFA for the next few minutes that was a product again and we were given a hypothesis and we were trying to figure things out right. We wanted to be a little more scrappy. And so Codex there was really given kind of the free reigns. We said like okay these are architectural like golden paths that we use at Stack Overflow. Go rogue. Tell us how we should do this.

但本质上,对于 Stack Overflow for agents(我接下来几分钟称之为 SOFA),这又是一个产品。我们得到了一个假设,并试图弄清楚事情该怎么做。我们希望更加灵活敏捷。因此,Codex 被给予了极大的自由裁量权。我们说:好吧,这些是我们用于 Stack Overflow 的架构式‘黄金路径’。放手去干吧。告诉我们你们认为我们应该怎么做。

Tell us what decisions, what constraints, kind of build that container for us. And so that was completely different than again the other end of the spectrum which is we're gonna deploy this to existing customers right. Customers with 150 to 10,000 seats on our platform. Some of them big banks with 50,000 right. On-prem a lot more constraints. And so for that we were a lot more prescriptive in what we wanted Codex to help us with.

告诉我们哪些决策、哪些约束,为我们构建那种容器。因此,这与光谱的另一端截然不同,另一端是我们要将此部署给现有客户,即在我们平台上拥有 150 到 10,000 个席位的一些客户。其中一些是大银行,拥有 50,000 个席位。本地部署意味着更多的约束。因此,在这方面,我们对 Codex 希望它帮助我们做什么提出了更多规定性要求。

So that was a lot of these are the existing contractual or contracts, architectural contracts between services. This is the way our architecture is built. Go build within that you know safe container. And so really again two different ends of the spectrum where we were able to kind of let Codex go rogue at different parts. And then again purposely and intentionally put the human in the loop there. Makes sense. It sounds like for the enterprise product where of course there are higher standards for security and other constraints.

所以,这些都是现有的合同或服务间的架构契约。这就是我们架构的构建方式。在已知的安全容器内构建即可。因此,这确实是光谱的两个不同极端,我们能够在不同部分让 Codex 放手去干。然后再次有目的且刻意地将人类纳入循环中。说得通。听起来对于企业产品而言,当然对安全性和其他约束有更高的标准。

You were more thoughtful up front about what guardrails to put in place. Yeah exactly. I think for the enterprise product there's a lot of complexity that Codex requires context for right. It's not just how is the architecture built. It's really you know what does an enterprise need. And that could be for the product that we're building which is essentially grabbing you know scattered context raw knowledge structuring that.

您在前期更周全地考虑了要设置哪些护栏。是的,完全正确。我认为对于企业产品,Codex 需要大量的上下文来处理复杂性。这不仅仅是架构如何构建的问题,而是企业究竟需要什么。这可能适用于我们正在构建的产品,即抓取分散的原始知识上下文并进行结构化处理。

There's a lot of complexity in the sense of who can see this right data leakage. If an enterprise asks a question someone in that enterprise asks a question about topic X. We have to make sure that we respect if they have access to that different authorization models different sources again on prem cloud. There's just a lot of different complexity that we need to be able to handhold obviously Codex through with that context.

复杂性在于谁能看到这些数据,即数据泄露问题。如果企业中的某人就主题 X 提出问题,我们必须确保尊重他们的访问权限、不同的授权模型以及不同的来源,无论是本地部署还是云端。存在许多不同的复杂性,我们需要能够引导 Codex 处理这些上下文。

And again if you think of a non enterprise product it's net new architecture. So I can kind of say you know what I want this and the product you tell me the best way to do this. And then we can kind of idea it in a completely different way. Makes sense. So for this enterprise product you mentioned you know there's a lot of many different data sources that you're trying to make sense of. In the age of is grep all you need it seems like the answer might be no.

同样,如果你考虑非企业级产品,那就是全新的架构。所以我可以说:‘我知道我想要什么,你告诉我实现它的最佳方式。’然后我们可以用一种完全不同的方式来构思它。说得通。所以对于你提到的这个企业级产品,你知道有很多不同的数据源需要你去理解。在‘grep 即一切’的时代,答案似乎是否定的。

No. So can you talk to us about how you worked through many different knowledge sources and making use of that. Yeah totally. So if you think of an enterprise and any of us in the work we do you create context every single day right. Whether it's a slack thread MS teams a Google document. Imagine how many times as an individual you're editing that right. You're sharing that with someone there's permissions attached to that right.

不是。那么你能跟我们谈谈你是如何处理众多不同的知识来源并加以利用的吗?当然可以。如果你想想企业环境以及我们工作中的任何场景,你每天都会创建上下文,对吧。无论是 Slack 线程、MS Teams 还是 Google 文档。想象一下作为一个个体,你会编辑多少次。你会与他人共享它,并且这些共享附带权限设置,对吧。

And then you kind of there's 10x complexity there right. Every individual in that enterprise is doing that. And so different sources have different complexity again. And so our product once again is a knowledge intelligence layer. And what it does is it goes into this raw context. It structures it organizes it. And then it applies stack overflows bread and butter which is trust right verification. But to do that we have to go into these systems again data leakage privacy governance.

然后这里的复杂度增加了 10 倍。企业中的每个个体都在这样做。因此不同的来源具有不同的复杂度。我们的产品再次成为一个知识智能层。它的作用是深入这些原始上下文,对其进行结构化、组织化。然后应用 Stack Overflow 的看家本领,也就是信任验证。但要做到这一点,我们必须再次进入这些系统,涉及数据泄露、隐私和治理问题。

It's like a whole grocery list of things to do and think of. And then we have to be sure that we respect that through and through. And so if you look at something like slack versus Google Docs versus a Jira ticket or a Confluence page. That information lives in there completely differently. And then you add the complexity of an enterprise might set up Jira one certain way. Why another one does this way. And so again we inherit all of that.

这就像一份要做和思考的事情的完整清单。然后我们必须确保彻底尊重这些要求。所以如果你看看 Slack 与 Google Docs 与 Jira 工单或 Confluence 页面的区别。这些信息在其中以完全不同的方式存在。然后再加上企业的复杂性,可能一家公司会以某种方式设置 Jira,而另一家公司则以这种方式设置。因此我们再次继承了所有这些差异。

And so we have to build this in a way. And again Codex sees right through that complexity and helps us ideate how to scale that. We have to think through how to kind of scale these sources in a way that we can move faster right. And that's what Codex was able to do for us. It built the concept of archetypes. So things that across these sources patterns that they have in common. An example of that would be the document archetype right.

因此我们必须以这种方式构建它。Codex 再次看透这种复杂性,帮助我们构思如何扩展它。我们需要思考如何以能够加快速度的方式扩展这些来源。这就是 Codex 为我们做到的。它构建了原型的概念。也就是说,在这些来源中它们共有的模式。一个例子就是文档原型,对吧。

That could hit Google Docs. It could hit an MS Teams attachment. It could hit Confluence, Notion. So if we're able to create an API around an archetype for a document. Codex can say that's a pattern that I see in these ten sources. And so when you go get that document. We'll make sure that downstream we'll be able to connect into that. So again that complexity would have taken us weeks and weeks of building that logic.

这可能会影响到 Google Docs。它可能会影响到 MS Teams 的附件。它可能会影响到 Confluence、Notion。因此,如果我们能够围绕文档的原型创建一个 API,Codex 可以说这是我在这十个来源中看到的模式。所以当你去获取该文档时,我们会确保下游能够连接到它。因此,这种复杂性原本需要花费我们数周时间来构建该逻辑。

And Codex was able to do that in a few days. Makes sense. What stands out also is it's not just about retrieval performance as it were. But like maybe some sources are more authoritative like instead of like Slack messages or Teams. Or there's data freshness to also think about. And so breaking it into archetypes helped you maybe put some context around what each of these sources might be. Yeah I think archetypes and again Codex support there was really accelerating that front door lift.

而 Codex 能够在几天内完成这一点。说得通。同样突出的一点是,这不仅仅是关于检索性能的问题。而是比如某些来源可能更具权威性,例如不是像 Slack 消息或 Teams 那样。或者还有数据新鲜度也需要考虑。因此,将其分解为原型有助于你为这些来源中的每一个提供一些上下文。是的,我认为原型以及再次提到的 Codex 支持确实加速了前端提升。

Right. Once we could kind of scale and kind of decide how we're gonna build these patterns. Then we can

没错。一旦我们能够规模化并决定如何构建这些模式,那么我们就可以

原文超出正文长度上限,此处截断——上游还有内容,完整版见上方「原文 ↗」。

更进一步:量化金融体系

看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力

进入量化体系 →

相似阅读

关联信息,但可能不是同一事件