首页  >>  来自播客: Lenny's Podcast 更新   反馈  

Lenny's Podcast - The limiting factor—how to design an AI software factory for speed | Geoff Charles (Ramp CPO)

发布时间:   原节目
演讲者首先感谢了之前的演讲者,并分享了一个他在勒芒(Le Mans)赛车时的个人轶事。尽管他那辆500美元的赛车撞毁了,但他学到了一个重要的教训:在职业赛车中,车手很少是唯一的问题所在。车手只占影响的15%;胜利的真正决定因素是车手、赛车和团队之间的互动。这引出了核心信息:获胜不仅仅是开得更快,而是要消除驾驶周围的瓶颈。 他举了一个 F1(Formula 1)维修站的例子,其时间从20世纪50年代的67秒,大幅缩短到如今的1.8秒。这并非通过要求机械师更努力工作来实现,而是通过识别和消除瓶颈,借助专业化分工、更先进的技术和不懈的练习。F1赛车就是这一理念的典范,每年有90%的16,000个零件会更新换代。演讲者强调,公司也正处于一场类似的竞赛中,速度和执行力至关重要,尤其是在人工智能(AI)时代。对于企业而言,“时钟”从客户痛点开始,到产品解决痛点时停止。在这种背景下,AI 移除了瓶颈,但同时也转移了瓶颈,使得获胜的团队是那些能够更快识别、消除并转向下一个瓶颈的团队。 传统的产品开发生命周期——识别、定义、构建和改进——正在经历一场变革。工程师们已经实现了大部分编码工作的自动化,瓶颈已转移到产品经理(PM)身上,他们现在面临着定义、协作、协调、测试和发布等方面的更高要求。演讲者敦促 PM 们“更像工程师”,并投资于他们“自己的工厂”以加速这个循环。 Ramp,演讲者所在的公司,已在这些步骤中进行了迭代和自动化: 1. **识别(Identify):** Ramp 最初在客户痛点数据孤岛问题上举步维艰,摆脱了“吐槽渠道”,构建了一个全面的客户洞察代理(agent)。这个代理从公司所有数据源拉取数据,利用 ETL、向量搜索和聚类技术来理解产品、团队和功能。通过 Slack 代理、HTML 仪表盘,甚至“吐槽播客”都可以访问,它帮助 PM 们精确锁定要接触的客户。 2. **定义(Define):** Ramp 没有问“你想构建什么?”,而是开发了“Glass”——一个 AI 代理,它连接到 Snowflake 等内部系统,获取数据、用户研究、产品策略和代码库。Glass 充当“技术负责人”(tech lead)的角色,提供定性和定量数据以增强具体性,评估可行性,甚至构建符合产品原则和设计系统的原型。这为工程团队生成了一份“下一份合同”:包括数据支持的问题、可操作的需求和富有启发性的原型。 3. **构建(Build):** 尽管编码不再是主要的瓶颈,Ramp 还是构建了“Inspect”——一个内部编码代理,它完全配置好以理解他们的代码库。与 Slack 集成后,Inspect 可以在五秒内返回可部署的产品预览。令人印象深刻的是,Ramp 75% 的 Pull Requests (PRs) 是由 Inspect 构建的,上个月有1,000名非工程师提交了 PRs。下一个瓶颈变成了代码审查,由“ReviewBuddy”解决,它理解 Ramp 的代码库、质量检查、安全问题和提示,自动处理93%的 PRs,并将关键的7%路由给高级工程师。 4. **测试(Test):** PM 们传统上在 QA 环境中花费大量时间。Ramp 利用“Testo”实现了这一点自动化,这是一个基于浏览器的 QA 代理,它根据生产数据,以100种不同的组合启动产品。Testo 像用户一样运行,提供阻塞性反馈(bug、问题),并提供深思熟虑的设计和定性反馈。在30天内,Testo 发现了425个 bug,阻止它们到达客户。 5. **协调(Coordinate):** 随着产品交付速度加快,人类注意力成为了瓶颈。Ramp 的“Gadget”代理理解问题的意图,并连接到正式记录(Notion、Slack、Linear、工单)。它回答了85%的 PM 问题,更新路线图,提醒滞后的利益相关者,甚至可以为发布撰写帮助中心文章、博客文章和客户邮件。 6. **改进(Improve):** Ramp 意识到需要自动化反应式的小规模任务。对于许多小问题,AI 现在完全运行整个循环:包括路由给团队、匹配积压工作、去重、规划、编写代码(有人类监督)以及运行测试和 CI/CD。这个“自主循环”让人类能够专注于更大的挑战,客户或销售发现的60%的 UX 问题在24小时内得到解决。 演讲者回答了三个关键问题: * **更快且高质量地行动:** 这需要招聘理解并挑战组织以实现速度的领导者(驱动者),就像 Niki Lauda 挑战法拉利(Ferrari)赛车一样。 * **有限的资源:** 拥抱限制,就像奥迪(Audi)在勒芒(Le Mans)以燃油效率获胜一样,寻找实现世界级表现的维度。 * **PM 的未来:** 这一角色将演变为三个方向:技术型 PM(构建工厂)、品味塑造者(设定产品标准),以及总经理(GM,负责跨职能的业务成果)。 总之,演讲者重申,速度在于消除瓶颈,这些瓶颈会不断转移,产品领导者必须专注于更快地构建产品的“工厂”,而不仅仅是产品本身。他鼓励其他人“效仿我们”并分享他们的创新,强调“你将构建的最好的产品是你将发布的下一个产品”,而这始于软件工厂。

The speaker began by acknowledging previous presenters and sharing a personal anecdote about racing cars at Le Mans. Though his $500 car crashed, he learned a crucial lesson: in professional racing, the driver is rarely the sole problem. The driver accounts for only 15% of the impact; the true determinant of victory is the interaction between the driver, the car, and the team. This led to the core message: winning isn't just about driving faster, but about removing bottlenecks around the driving. An illustrative example given was the Formula 1 pit stop, which dramatically reduced from 67 seconds in the 1950s to 1.8 seconds today. This was not achieved by asking mechanics to work harder, but by identifying and removing bottlenecks through specialized functions, better technology, and relentless practice. F1 cars exemplify this philosophy, with 90% of their 16,000 parts changing annually. The speaker emphasized that companies are in a similar race, with speed and execution being paramount, especially in the era of AI. For businesses, the "clock" starts with customer pain and stops when a product solves it. AI, in this context, removes bottlenecks but also shifts them, making the winning team the one that can faster identify, remove, and move on to the next bottleneck. The traditional product development lifecycle—identify, define, build, and improve—is undergoing a transformation. While engineers have automated much of their coding, the bottleneck has shifted to product managers (PMs) who now face increased demands for defining, collaborating, coordinating, testing, and releasing. The speaker urged PMs to "be more like engineers" and invest in their "own factory" to accelerate this loop. Ramp, the speaker's company, has iterated and automated across these steps: 1. **Identify:** Initially struggling with silos of customer pain data, Ramp moved beyond a "hate channel" to build a comprehensive customer insight agent. This agent pulls data from all company sources, using ETL, vector search, and clustering to understand product, teams, and features. Accessible via Slack agents, HTML dashboards, and even a "hate podcast," it helps PMs pinpoint exactly which customers to engage. 2. **Define:** Instead of asking "what do you want to build?", Ramp developed "Glass," an AI agent that connects to internal systems like Snowflake for data, user research, product strategy, and codebase. Glass acts as a "tech lead," providing specificity with qualitative and quantitative data, assessing feasibility, and even building prototypes that align with product principles and design systems. This generates a "next contract" for engineering: data-backed problems, actionable requirements, and inspiring prototypes. 3. **Build:** While coding is no longer the primary bottleneck, Ramp built "Inspect," an in-house coding agent fully provisioned to understand their codebase. Integrated with Slack, Inspect can return a deployable product preview in under five seconds. Impressively, 75% of Ramp's Pull Requests (PRs) are built by Inspect, with 1,000 non-engineers submitting PRs in the last month. The next bottleneck became code reviews, addressed by "ReviewBuddy," which understands Ramp's codebase, quality checks, security concerns, and prompts, handling 93% of PRs automatically and routing critical 7% to senior engineers. 4. **Test:** PMs traditionally spent significant time in QA environments. Ramp automated this with "Testo," a browser-based QA agent that spins up the product in 100 different combinations based on production data. Testo runs like a user, provides blocking feedback (bugs, issues), and offers thoughtful design and qualitative feedback. In 30 days, Testo caught 425 bugs, preventing them from reaching customers. 5. **Coordinate:** As product shipping accelerated, human attention became the bottleneck. Ramp's "Gadget" agent understands the intent of questions and connects to formal records (Notion, Slack, Linear, tickets). It answers 85% of PM questions, updates roadmaps, pings late stakeholders, and can even write help center articles, blog posts, and customer emails for launches. 6. **Improve:** Ramp recognized the need to automate reactive, small-scale tasks. For many minor issues, an AI now fully runs the loop: routing to teams, matching backlogs, deduplicating, planning, writing code (with human oversight), and running through tests and CI/CD. This "autonomous loop" allows humans to focus on bigger challenges, with 60% of UX issues identified by customers or sales fixed within 24 hours. The speaker addressed three key questions: * **Moving faster with quality:** This requires hiring leaders (drivers) who understand and challenge the organization to achieve speed, much like Niki Lauda challenging Ferrari's car. * **Limited resources:** Embrace constraints, like Audi's fuel efficiency win at Le Mans, to find a dimension for world-class performance. * **Future of PMs:** The role will evolve into three tracks: technical PMs (building the factory), tastemakers (setting product standards), and GMs (owning business outcomes across functions). In conclusion, the speaker reiterated that speed is about removing bottlenecks, these bottlenecks constantly shift, and product leaders must obsess over the "factory" that builds products faster, rather than just the products themselves. He encouraged others to "copy us" and share their innovations, emphasizing that "the best product you'll ever build is the next product that you'll launch," starting in the software factory.