人工智能如何改变 Anthropic的工作方式

⼈⼯智能如何改变Anthropic的⼯作⽅式 2025年12⽉3⽇ How is AI changing the way we work? Ourprevious researchon AI’s economicimpacts looked at the labor market as a whole, covering a variety of different jobs.But what if we studied some of the earliest adopters ofAI technology in moredetail—namely, us? ⼈⼯智能如何改变我们的⼯作⽅式？我们此前关于AI经济影响的研究着眼于整体劳动⼒市场，涵盖了各类不同职业。但如果我们更详细地研究⼀些AI技术的早期采⽤者——也就是我们⾃⼰——会是怎样的情形？ Turning the lens inward, in August 2025 we surveyed 132 Anthropic engineersand researchers, conducted 53 in-depth qualitative interviews, and studiedinternalClaude Codeusage data tofind out howAI use is changing things atAnthropic. Wefind that AI use is radically changing the nature of work forsoftware developers, generating both hope and concern.微信公众号 404K微信公众号 404K 2025年8⽉，我们将研究视⾓转向内部，对132名Anthropic⼯程师和研究⼈员开展问卷调查，进⾏53次深度定性访谈，并分析内部Claude代码使⽤数据，以探究⼈⼯智能如何改变这家公司的⼯作模式。研究发现，AI的应⽤正在彻底改变软件开发者的⼯作本质，既带来希望也引发忧虑。 Our research reveals a workplace facing significant transformations: Engineersare getting a lot more done, becoming more “full-stack” (able to succeed at tasksbeyond their normal expertise), accelerating their learning and iteration speed, 我们的研究发现职场正经历重⼤变⾰：⼯程师们⼯作效率⼤幅提升，变得更加"全栈化"（能够完成超出专业领域的任务），加快学习与迭代速度，并着⼿处理以往被忽视的⼯作。这种能⼒范围的扩展也让⼈们开始思考其中的得失——有⼈担忧这可能导致深层技术能⼒流失，或降低对Claude输出的监督能⼒，另⼀些⼈则乐于借此拓宽思路提升格局。部分⼈感受到与AI合作增加反⽽减少了同事间协作，还有⼈担忧最终会因⾃动化取代⽽失业。 We recognize that studying AI’s impact at a company building AI meansrepresenting a privileged position—our engineers have early access to cutting-edge tools, work in a relatively stablefield, and are themselves contributing tothe AI transformation affecting other industries. Despite this, we felt it was onbalance useful to research and publish thesefindings, because what’s happeninginside Anthropic for engineers may still be an instructive harbinger of broadersocietal transformation. Ourfindings imply some challenges and considerationsthat may warrant early attention across sectors (though see the Limitationssection in theAppendixfor caveats). At the time this data was collected, ClaudeSonnet 4 and Claude Opus 4 were the most capable models available, andcapabilities have continued to advance.微信公众号 404K微信公众号 404K 我们清楚，在⼀家开发AI的公司内部研究AI影响意味着⾝处优势地位——我们的⼯程师能率先使⽤尖端⼯具，⼯作在相对稳定的领域，且⾃⾝正推动着影响其他⾏业的AI变⾰。尽管如此，我们认为研究和发表这些发现总体仍具价值，因为Anthropic⼯程师的现状可能预⽰着更⼴泛社会变⾰的先兆。研究发现暗⽰了⼀些值得各⾏业提早关注的挑战与考量（但需注意附录《局限性》部分的说明）。本次数据收集时，Claude Sonnet 4和Claude Opus 4是最先进的可⽤模型，⽽AI能⼒仍在持续演进。 More capable AI brings productivity benefits, but it also raises questions aboutmaintaining technical expertise, preserving meaningful collaboration, andpreparing for an uncertain future that may require new approaches to learning,mentorship, and career development in an AI-augmented workplace. We discusssome initial steps we’re taking to explore these questions internally in theLooking Forward section below. We also explored potential policy responses inour recent blog post onideas forAI-related economic policy. 更强⼤的⼈⼯智能带来了⽣产⼒提升，但也引发了⼀系列问题：如何维持技术专业深度、保持有效协作、以及为充满不确定性的未来做好准备——在这个AI增强的⼯作环境中，可能需要全新的学习⽅式、师徒传承机制和职业发展路径。在下⽅"展望未来"章节中，我们将分享公司内部为探索这些问题所采取的初步措施。关于潜在的政策应对⽅案，我们已在近期发布的AI相关经济政策构想博客⽂章中进⾏了探讨。 Keyfindings主要发现微信公众号 404K In this section, we briefly summarize thefindings from our survey, interviews,and Claude Code data. We provide detailedfindings, methods, and caveats in thesubsequent sections below. 本节我们简要概述来⾃问卷调查、深度访谈和克劳德代码数据的发现。后续章节将提供详细的研究结果、⽅法论和注意事项说明。 Survey data问卷调查数据 1.Anthropic engineers and researchers use Claude most often forfixing codeerrors and learning about the codebase. Debugging and code understandingare the most common uses (Figure 1). 试代码和理解代码是最常见的⽤途（图1）。 2.People report increasing Claude usage and productivity gains.Employeesself-report using Claude in 60% of their work and achieving a 50%productivity boost, a 2-3x increase from this time last year. This productivitylooks like slightly less time per task category, but considerably more outputvolume (Figure 2). 员⼯们普遍反映克劳德使⽤频率和⽣产⼒均有提升。据员⼯⾃述，他们在60%的⼯作中使⽤克劳德，⽣产⼒提⾼了50%，较去年同期增长2-3倍。这种⽣产⼒提升表现为每项任务耗时略有减少，但产出量显著增加（图2）。 3.27% of Claude-assisted work consists of tasks that wouldn't have beendone otherwise, such as scaling projects, making nice-to-have tools (e.g.interactive data dashboards), and exploratory work that wouldn't be cost-effective if done manually. 27%由Claude协助完成的⼯作属于原本不会着⼿进⾏的任务，例如规模扩展项⽬、开发锦上添花的⼯具（如交互式数据仪表盘），以及⼈⼯操作成本效益不⾼的探索性⼯作。微信公众号 404K微信公众号 404K 4.Most employees use Claude frequently while reporting they can “fullydelegate” 0-20% of their work to it.Claude is a constant collaborator butusing it generally involves active supervision and validation, especially inhigh-stakes work—versus handing offtasks requiring no verification at all. 多数员⼯频繁使⽤Claude，但⾃述仅能"完全委托"0-20%的⼯作给它。Claude是持续的⼯作伙伴，但使⽤过程通常需要主动监督和验证，尤其在⾼风险⼯作中——与完全不需要核查的任务交付形成对⽐。 Qualitative interviews定性访谈 1.Employees are developing intuitions forAI delegation. Engineers tend todelegate tasks that are easily verifiable, where they “can relatively easily sniff-check on correctness”, low-stakes (e.g. “throwaway