Airbnb：从内部 AI 平台走向用户产品
原题Inside-Out AI: Rebuilding Airbnb Behind the Scenes and Across the Guest Experience
读前速览
Airbnb CTO Ahmad Al-Dahle 解释了“由内而外”的 AI 策略：先利用内部 AI 工具加速产品开发，随后将这些能力扩展到改造客端体验。
“由内而外”的 AI 策略
Airbnb 优先考虑在内部使用 AI 来加速开发周期，利用这些验证过的能力来增强面向客户的功能。
从产物转向代码
通过消除产品需求与设计文档之间的传统交接，团队现在直接与原型和代码进行工作，显著提高了拉请求的吞吐量和速度。
Everest AI 上文图
使用 LLM 和嵌入技术的自定义内部知识图允许通用型开发者快速构建复杂的专家服务（如杂货），将开发时间从几个月缩短到几周。
多模型策略
Airbnb 部署了至少 10 个定制模型，将昂贵的前沿模型用于编码等高风险任务，并使用较小的、经过训练的开源模型用于搜索等延迟敏感的任务。
异步智能体的兴起
下一个前沿是运行在容器中的智能体，由特定事件（如监控告警）触发，以自动化初始的轮班分拣和事件管理。
完整摘要
Airbnb CTO Ahmad Al-Dahle 详细介绍了了公司如何通过“由内而外”的方法向“AI 原生”组织转型。该策略涉及在内部使用 AI 来彻底改变软件工程——例如从依赖文档的工作流转向直接进行代码原型设计——这导致 60% 的代码由 AI 编写，功能交付速度提升了 80%。文章还涵盖了内部上下文图“Everest”的使用（该图用于加速杂货配送等新服务），以及一种平衡了前沿模型与基于成本、延迟和性能需求的定制开源模型的多模型策略。
目录 · 6 节
在今年1月加入 Airbnb 担任首席技术官之前，Ahmad Al-Dahle 曾任 Meta 生成式人工智能负责人，并主导了 Meta 在2023年至2025年间推出开源 Llama 模型。如今，Al-Dahle 负责将 Airbnb 打造成一家“AI 原生公司”。Prior to joining Airbnb as CTO in January, Ahmad Al-Dahle was head of generative AI at Meta and led the launch of its open source Llama models over 2023-2025. Now Al-Dahle is in charge of turning Airbnb into an “AI-native company”.
实际而言，这意味着先在内部使用 AI 加快产品开发，然后再运用这些能力变革客户体验。我们将这种做法称为“由内而外的 AI”策略，本文将深入探讨其中的细节。What that means in practice is using AI internally to speed up product development, then using those same capabilities to transform the customer experience. We’re calling this an “inside-out AI” approach and in this article we’ll dig into the details.
Al-Dahle 与 Latent Space 谈到了 Airbnb 的 AI 转型。我们首先问他，为什么他从在 Meta 构建前沿模型转而在 Airbnb 部署这些模型。Al-Dahle spoke with Latent Space about Airbnb’s AI transformation. We began by asking him why he made the shift from building frontier models at Meta to deploying them at Airbnb.
从根本上说，我喜欢追逐我认为真正艰难的前沿所在之处，”他回答道，或许这也是有意为之的双关。“我们在 Meta 时，某种程度上知道飞轮是什么样的，也明白如何一代又一代地提升模型能力。“Fundamentally, I like to chase where I believe the hard frontiers live,” he replied, pun perhaps intended. “We kind of knew [at Meta] what the flywheel looks like, we understood how to improve the model capabilities, generation on generation.”
阿尔-达赫勒认为，下一个挑战是大规模部署模型。在 Airbnb，他的目标是“推动人们改变工作方式，因为这些工具正在改变你的工作方式”，以及“将这些系统投入生产环境，并以富有创意的方式加以部署，切实提升核心用户体验”。The next challenge, Al-Dahle thinks, is deploying models at scale. At Airbnb, his goals are to “push people to work differently, because these tools are transforming how you work,” and to “take these systems and deploy them in production, in creative ways that actually add value to a core user experience.”
直接进入原型阶段，代码就是产物Jumping to prototypes, code is the artifact
自 2008 年创立以来，Airbnb 一直被视为一家在酒店业运营的科技公司。该公司于 2020 年上市，目前市值约为 930 亿美元（截至本文撰写时）。因此，要让它成为一家“AI 原生”企业，可不是件小事。From its launch in 2008, Airbnb has been considered a technology company operating in the hospitality industry. The company went public in 2020 and now has a market capitalization of approximately $93 billion (at time of writing). So it’s a big operation to try and make “AI-native.”
Al-Dahle列举了几项数据，说明爱彼迎如今已深度拥抱AI：目前其60%的代码由AI编写，发布的功能和改进比去年同期增加了近80%，而普通工程师的拉取请求处理量提升了约1.6倍。Al-Dahle reeled off a few statistics to show that Airbnb is AI-pilled now: 60% of its code is now AI-authored, it has shipped nearly 80% more features and improvements year over year, and pull-request throughput for the average engineer is up about 1.6x.
但 Airbnb 究竟是如何做到这一切的？But how exactly has Airbnb done all this?
首先，他表示，这是在改变软件工程的组织流程。传统上，流程可能是先制定产品需求，再用 Figma 进行设计，然后由工程团队实施，最后进行生产环境测试。过去，这些环节需要从一个团队移交给另一个团队。但现在，Airbnb 的产品、设计和工程团队直接开始使用原型开展工作。Firstly, he said it’s about changing the organizational process for software engineering. Traditionally, you might do product requirements, then design work in Figma, then engineering implementation, and eventually production testing. In the past, these would’ve required handoffs from one team to another. But now, Airbnb has its product, design and engineering teams move directly to working with prototypes.
缩短这段时间，实际上是许多传统软件公司必须迈出的一大步，也是它们能实现的最大节省之一，”Al-Dahle 说。“我们迈出了这一步，而这实际上为我们的产品开发流程注入了强大的助力。“Breaking that time down is actually one of the biggest savings that a lot of traditional software companies have to make the leap to do,” said Al-Dahle. “And so we made that leap and it actually ended up adding a ton of tailwind to our product development process.”
这一变化带来的一个结果是，团队直接处理代码，而不是处理产品需求文档之类的“制品”。A byproduct of this change is that teams deal directly with code, rather than dealing with “artifacts” like a product requirements document.
“我们不再过度生成各种产物，而是转而把代码和原型作为我们推理所依据的产物，”他说。“We moved away from excessive artifact generation to the code being the artifact that we reason upon, and prototypes,” he said.
AI 解决了大约一半的支持工单。AI resolves roughly half of support tickets
阿尔-达赫莱告诉我们，客户支持是他们最先引入 AI 的面向用户领域。他形容这是“最难部署的问题”，因为“出错的风险和后果都非常严重”。Al-Dahle told us that customer support was the first user-facing area that it introduced AI into. He described it as “the hardest problem to deploy,” because “the stakes and the consequences of a mistake are really high.”
Al-Dahle 表示，如今 Airbnb 约有一半的客服工单完全由 AI 解决（这与公司第二季度业绩相符，后者显示这一比例接近 45%）。Roughly half of Airbnb’s support tickets are now resolved purely by AI, said Al-Dahle (this is in line with the company’s Q2 results, which put the figure at nearly 45%).
他说，关键在于使用合成数据，在将系统投入生产环境之前进行全面测试。The key to this, he says, is using synthetic data to thoroughly test the system before taking it to production.
我们首先要做的是：构建模型、构建智能体，并在投入生产之前开始生成一批合成数据。“We start with: build a model, build the agent, and begin to generate a battery of synthetic data before you ever go into production.”
他指出，尽管智能体处理了大约一半的客户支持咨询，但对于哪些问题需要人工协助，它们会谨慎判断——例如涉及安全问题时。He notes that although agents handle about half of customer support queries, they are careful about what requires human assistance — with safety issues, for example.
因此，虽然我们解决了 50% 的工单，但对于那些我们暂时选择不让智能体解决的工单，我们其实是经过审慎考虑的。“So while we solve 50% of the tickets, we’re actually deliberate about the tickets we don’t choose to solve yet [with agents].”
Everest, Airbnb’s AI context graph
你可能还不熟悉 Airbnb 的两项服务：杂货配送和机场接送（杂货配送服务刚刚扩展到更多城市）。这两项服务都于今年早些时候推出，其中部分得益于 Airbnb 由内而外的 AI 方法。Two Airbnb Services projects you may not be familiar with yet are grocery deliveries and airport pickups (the grocery service has just been expanded to more cities). Both were launched earlier this year, thanks in part to Airbnb’s inside-out AI approach.
帮助这些产品迅速推向市场的，是一个名为 Everest 的内部组织情境图谱。Al-Dahle 表示，Everest 运用大语言模型、嵌入和基于 AI 的检索等技术来构建和查询该图谱。It was an internal organizational context graph called Everest that helped bring these products to market quickly. Al-Dahle said that Everest uses technologies such as LLMs, embeddings and AI-based retrieval to build and query the graph.
他解释说，杂货配送和机场接送服务“有点类似——都是与合作服务（外部公司）进行 API 集成，并接入我们平台的服务。”He explained that the grocery delivery and airport pickup services are “kind of similar services — API integrations with partner services [external companies] that integrated onto our platform.”
杂货服务首先构建完成，团队从中汲取的经验也应用到了 Everest 中。这让机场接送团队能够更快地完成项目。在 Airbnb 的第二季度财报中，公司表示：“杂货服务花了八到九个月开发，而机场接送服务的开发大约用了六周。”The grocery service was built first, and learnings from it were made available in Everest. That enabled the airport pickup team to do its project much more quickly. In its Q2 earnings report, Airbnb states that “groceries took eight months, nine months and airport pickups took about six weeks to develop.”
此外，使用 Everest 意味着开发人员不一定需要具备专业知识也能参与项目开发。Additionally, using Everest means its developers don’t necessarily need specialist knowledge to work on projects.
正是由于整个代码库中存在这种上下文图，”Al-Dahle 说道，“我们才能让通才处理代码中非常专业的部分。“Because of this context graph that exists across the codebase,” said Al-Dahle, “we’re able to have generalists work across very specialist parts of the code.”
这两项新服务均由首席执行官布莱恩·切斯基在 Airbnb 于 5 月发布的 2026 年夏季新品中重点介绍。Both of the new services were highlighted by CEO Brian Chesky at Airbnb’s 2026 Summer Release in May.
选择要使用的模型并自定义开放模型Choosing which models to use and customizing open models
为了让人们更了解 Airbnb 内部如何使用 AI，Al-Dahle 将其称为一家“多模型公司”，并表示公司在生产系统中部署了“许多不同的模型”。To give more of a sense of Airbnb’s internal AI usage, Al-Dahle described it as a “multi-model company” and says it deploys “a lot of different models” in its production systems.
Airbnb 混合使用前沿模型和开放模型，但大部分后训练和强化学习工作都是基于开放模型自行开展的。Al-Dahle 表示，公司至少部署了 10 个定制模型，用于生产场景。Airbnb uses a mixture of frontier and open models, but conducts most of its own post-training and reinforcement learning on open models. Al-Dahle said the company deploys at least 10 customized models for production use cases.
它还会根据成本、性能和延迟的帕累托前沿评估模型，并为每个应用选择不同的权衡方案。It also evaluates models along a Pareto frontier for cost, performance and latency, selecting a different trade-off for each application.
我们会针对每种使用场景专门进行评估，”他解释道。“比如，如果我们用模型来做搜索，就会从生产环境中抽取一组搜索查询，并据此进行衡量。如果是用于客户支持，我们也会抽取所有我们最关注的生产环境查询——包括边缘情况。“We run evals specifically for each use case,” he explained. “So if we’re using a model for search, we have a set of search queries that are sampled from production that we measure against. If we’re doing customer support, we also sample all the production queries that we care most about — including edge cases.”
“我们尽量让合适的工具做合适的事，”他补充道。例如，编程任务可以容忍更高的延迟，但出错代价高昂。因此，他们更倾向于使用能力最强的前沿模型来处理编程任务。“We try to use the right tool for the right job,” he added. For example, coding can tolerate more latency but mistakes are expensive. So they prefer to use the strongest available frontier model for coding tasks.
使用最前沿的编码模型是合理的，因为模型每产生一个缺陷，都可能让我们付出高得多的代价。“It makes sense to use the absolute most frontier coding model, because for every defect that a model produces, it could easily cost us a lot more.”
另一方面，Airbnb 的用户会大规模使用搜索功能，因此对延迟非常敏感。因此，对于这项任务，他们更倾向于使用规模较小的专用模型。Al-Dahle 声称，规模较小、经过后训练的模型有时表现优于前沿模型。On the other hand, search is used at scale by Airbnb’s users and so is very latency-sensitive. So for that job, they prefer smaller specialized models. Al-Dahle claimed that smaller, post-trained models sometimes out-perform frontier models.
针对一些特定应用场景，我们已经能够对非常小巧灵活、速度快且成本低的模型进行后训练，使其能力超越前沿水平。“For narrow use cases, we’ve been able to take really small, really nimble models that are very fast and cheap and post-train them beyond the frontier.”
异步代理：下一次重大变革Async agents: the next big shift
与其他原生 AI 公司一样，Airbnb 创建了一个内部智能体，并将其命名为 AirChat。Like other AI-native companies, Airbnb has created an internal agent — in its case called AirChat.
“我们有自己的内部智能体，称为 AirChat，它基本上包含了所有必要的 MCP 组织背景信息，”Al-Dahle 说。“We have our own internal agent that we call AirChat, which [includes] basically all the necessary MCP organizational context,” said Al-Dahle.
但 Airbnb 也开始着手实现 Al-Dahle 认为是智能体下一阶段发展的方向：在容器中运行、由事件触发的异步智能体。But Airbnb is also starting to implement what Al-Dahle sees as the next development in agents: asynchronous agents running in containers and activated by events.
我们的很多团队基本上都开始自动化值班流程了，”他解释道，“所以，如果出现问题——比如 Grafana 的阈值触发了，或者你使用的任何监控系统发出了告警——智能体就会启动，接手分诊并执行值班流程，也就是最初的值班响应。“A lot of our teams are starting to basically automate their on-calls,” he explained. “So if something trips, if a Grafana limit or whatever monitoring system that you use trips, agents spin up to then triage and run your on-call, your initial on-call.”
人工工程师可以审核代理提出的 PR，而如果代理判断警报不稳定，它也可以自行关闭事件。A human engineer can review a PR proposed by the agent, while the agent may close the incident itself if it determines that the alert was flaky.
这一流程让我们想起了 Vercel 的 AI SDK、Astro、Flue 和 tldraw 等开源项目如何利用软件工厂——由智能体组成的团队——来实施修复和开发新功能。This process reminds us of how open source projects like Vercel’s AI SDK, Astro, Flue and tldraw use software factories — teams of agents — to apply fixes and features.
Al-Dahle认为，这种方法最终将应用于其整个市场平台，例如用于监控欺诈、“信任违规”、市场质量和软件缺陷。Al-Dahle sees this approach eventually being used across its entire marketplace platform; for example to monitor fraud, “trust violations,” marketplace quality, and software defects.
我的愿景是部署大量异步代理，协助自动化和管理这个市场。“My vision is large amounts of asynchronous agents that are helping automate and manage the marketplace.”
传承软件工程的技艺Maintaining the craft of software engineering
Al-Dahle 认为，围绕成果而非功能来组建团队，是成为 AI 原生公司的关键。Al-Dahle believes that organizing teams around outcomes rather than features is key to becoming an AI-native company.
如果你的公司是按功能开发来组织的，那么在人工智能时代就会举步维艰，”他说。“如果你的公司是围绕要实现的目标、使命和成果来组织的，那么就不会有问题。“If you’re a company that organizes by feature development, you’re going to struggle in the age of AI,” he said. “If you’re a company that organizes by objectives and missions and results that you’re trying to achieve, you’re going to be fine.”
尽管如此，在向 AI 原生转型的过程中，Al-Dahle 仍有一件事感到担忧：如今 AI 能替初级工程师完成如此多的工作，他们是否还能培养自己的专业技能和判断力。That said, there’s one thing Al-Dahle worries about during this AI-native transition: whether junior engineers are developing their craft and judgement, given that AI is able to do so much work for them now.
他说，资深工程师的判断力是在多年发布系统、运行生产环境中的软件过程中积累起来的——其中也包括他们一路犯下的错误。Senior engineers, he says, have acquired their judgement through many years of shipping systems and operating software in production — including the mistakes they made along the way.
Al-Dahle 表示，解决方案的一部分是确保他们所有的工程师都能解释人工智能为他们完成的工作。Part of the solution, Al-Dahle says, is ensuring all of their engineers can explain the work AI does for them.
我正在大力推动的一点是，每位工程师都必须能够解释自己构建的东西，即使这份 PR 是由 AI 生成的，也必须能够解释。我认为，只要这一基本前提依然成立，更多初级工程师就能学会界面设计和架构这门手艺，也能认识到测试、单元测试和集成测试的重要性。所以，我们努力保持很高的专业水准。“One of the things I’m pushing very strongly is that every engineer must be able to explain, even if an AI generated the PR, must be able to explain what they built. I think if that fundamental premise remains true, then more junior engineers will learn the craft of interface design and architecture, and the importance of testing and unit testing and integration tests. So we’re trying to keep our craft really high.”