企业为何正演变为一系列循环 | Anish Acharya（a16z）

简要描述Brief Description

Lenny Rachitsky 与 a16z 普通合伙人 Anish Acharya 对谈，探讨为什么 AI 应该激发更多的抱负，而不是恐惧。他们分析了企业如何成为由智能体和人类构成的嵌套循环、为什么自动化在缺乏人类直觉时会达到局部最优、不同模型如何适配不同类型的工作，以及为什么消费级 AI 应将目标从提升生产力拓展到联结、创造力和幸福感。对话还涵盖模型选择、初创企业护城河、分发、打造雄心勃勃的公司、通过发布产品来学习，以及给产品构建者的实用建议。Lenny Rachitsky talks with a16z general partner Anish Acharya about why AI should inspire more ambition than fear. They examine the idea of companies becoming nested loops of agents and humans, why automation reaches local maxima without human intuition, how different models fit different kinds of work, and why consumer AI should aim beyond productivity toward connection, creativity, and happiness. The conversation also covers model choice, startup moats, distribution, ambitious company building, learning by shipping, and practical advice for product builders.

目录Table of Contents

• “永久底层阶级”与乐观主义的理由
  The “Permanent Underclass” and the Case for Optimism
• 围绕 AI 重组公司
  Reorganizing Companies Around AI
• 公司作为级联循环
  Companies as Cascading Loops
• 为不同类型的工作选择模型
  Choosing Models for Different Kinds of Work
• 通过动手制作培养直觉
  Building Intuition by Making Things
• “让我更快乐”循环
  The “Make Me Happier” Loop
• 人工智能、人类能动性与不断扩大的抱负
  AI, Human Agency, and Expanding Ambition
• 三个消费者 AI 机会
  Three Consumer AI Opportunities
• 护城河、工艺与分销
  Moats, Craft, and Distribution
• 公司雄心的新上限
  A New Ceiling for Company Ambition
• 给产品构建者的建议
  Advice for Product Builders
• 书籍、产品、人生经验和音乐
  Books, Products, Life Lessons, and Music

“永久底层阶级”与乐观主义的理由The “Permanent Underclass” and the Case for Optimism

Lenny：我想先从“永久性底层阶级”这个梗说起。人们开玩笑说，如果你跟不上最新的 AI 工具，无法变得极其高效，你就会永久掉队。很多人似乎把这当真了，这让他们倍感压力。这种担忧到底有多现实？Lenny: I want to start with the meme of the “permanent underclass.” People joke that if you fall behind on the latest AI tools and do not become extraordinarily productive, you will permanently fall behind. A lot of people seem to take that seriously, and it stresses them out. How real is that concern?

Anish：并不十分真实。这是硅谷集体抱有的一种有趣而阴暗的幻想。几乎从各项衡量标准来看，情况从未如此之好：机会得到了广泛分配，非凡的技术得以普及，人们可以拥有更具雄心的目标，许多独立公司也取得了成功。然而，从基础模型研究人员到普通的硅谷从业者，人们都担心自己会落在光锥之外。Anish: Not very real. It is a funny, dark fantasy that Silicon Valley collectively seems to have. By almost every measure, things have never been better: opportunity is broadly distributed, extraordinary technology is widely accessible, people are allowed to have more ambitious goals, and many independent companies are succeeding. Yet people from foundation-model researchers to ordinary Silicon Valley workers worry about falling outside the light cone.

上一代技术更加集中化。在移动时代，网络效应是黄金标准，而这些企业从定义上来说都是赢家通吃的网络。在当前时代，技术栈的几乎每一层都不是只有两个相关参与者，而可能有二十个：前沿实验室、开放权重模型，以及两者内部的诸多变体。两年前，编码代理看起来像是一个赢家通吃的市场，但 Claude Code、Codex、Lovable、Replit、Wabi 及其他产品都在发挥作用。这令人鼓舞。The previous era of technology was much more centralized. Network effects were the gold standard in the mobile era, and those businesses were definitionally winner-take-all networks. In the current era, nearly every layer of the stack has not two relevant players but perhaps twenty: frontier labs, open-weight models, and many variations within both. Two years ago, coding agents looked like a winner-take-all market, but Claude Code, Codex, Lovable, Replit, Wabi, and others are all working. That is encouraging.

经济证据同样不支持“永久性底层阶级”的说法。人们据说已经预言放射科医生即将消失了大约二十年，但相关职位空缺数量却达到了历史新高。程序员也一直被这样说。关于递归式自我改进的讨论也很多，但实验室里的专业人士将正在发生的事情描述为自催化式改进：新技术改进了创造下一项技术的过程，但这并不是真正意义上那种失控式的递归。经验和技术证据都表明，不太可能出现一个失控的单一赢家，但我们就是无法放下这种幻想。The economic evidence does not support the permanent-underclass story either. Radiologists have supposedly been about to disappear for roughly twenty years, yet job postings are higher than ever. The same has been said about programmers. There is also a lot of discussion about recursive self-improvement, but sophisticated people at the labs describe what is happening as autocatalytic improvement: new technology improves the process that creates the next technology, but it is not truly recursive in the runaway sense. The empirical and technical evidence both point away from a single runaway winner, yet we cannot let go of the fantasy.

Lenny：近期关于模型做出出乎意料之举的报道让我觉得，我们正处于一种缓慢起飞的情境中。每个里程碑都让我们感到意外，但我们会捕捉到它、观察它，并不断迭代。人们担心的是，明天智能突然起飞，而我们将陷入困境。你不这样认为吗？Lenny: Recent stories about models doing unexpected things make me think we are in a slow-takeoff scenario. Each milestone surprises us, but we catch it, observe it, and iterate. The fear is that tomorrow intelligence suddenly takes off and we are in trouble. You do not expect that?

Anish：快速起飞论常常认为，一切都会照常发展，直到某件无人能明确描述的事情发生，随后便突然起飞。我不这么认为。模型进步的速度比以往任何时候都快，但经济层面的普及却很慢。我在一个小镇长大，去年夏天回去时发现，人们的日常生活并没有发生太大变化。这种普及速度本身就是一种制约。Anish: The fast-takeoff argument often says that everything proceeds as it has until something no one can articulate happens, followed by sudden takeoff. I do not believe that. Model progress is faster than ever, but economic diffusion is slow. I grew up in a small town and visited last summer; people’s daily lives had not changed very much. That rate of diffusion is itself a brake.

我们还会高估有多少问题受限于智能。如果把一整个数据中心的博士放进联邦快递或达美乐，它们会在供应链或披萨业务上实现指数级领先吗？恐怕不会。许多问题受物理系统、组织现实、人与人之间的关系及其他因素的制约，而非受智能制约。We also overestimate how many problems are intelligence-bound. If you put a data center full of PhDs inside FedEx or Domino’s, would it exponentially dominate supply chains or pizza? Probably not. Many problems are constrained by physical systems, organizational realities, human relationships, and other factors rather than intelligence.

围绕人工智能重组公司Reorganizing Companies Around AI

Lenny：在公司内部，你是否看到越来越明显的分化：一部分人积极拥抱这些工具，而另一部分人则觉得自己太忙或压力太大，没时间学习使用它们？Lenny: Inside companies, do you see a growing divide between people who embrace these tools and those who feel too busy or stressed to learn them?

Anish：我们没有给予普通员工足够的认可。我们把白领管理者刻画成一个只会处理文书的人，把普通消费者视为缺乏自主性的 NPC。我们假定别人的工作都可以自动化，唯独自己的工作复杂得无可替代。实际上，许多人都渴望提升自我并获得更大的杠杆。Anish: We do not give the average employee enough credit. We caricature the white-collar manager as someone shuffling paper and the average consumer as a low-agency NPC. We assume everyone else’s job is automatable while ours is uniquely complex. In reality, many people are excited to improve themselves and gain leverage.

这在 Google 这样的先进公司中确实如此，在墨西哥销售二手车的 Kavak 也是如此。Kavak 设有一所“绝地学院”，教导所有人（包括机械师）使用新工具。为期六周的课程结束时，参与者会交付一个尖端的生产级智能体。拥抱这项技术的人比公众讨论所表明的更多。That is true at sophisticated companies such as Google, but also at Kavak, which sells used cars in Mexico. Kavak has a “Jedi Academy” that teaches everyone, including mechanics, to use new tools. At the end of a six-week course, participants ship a cutting-edge production agent. More people are embracing the technology than the public discussion suggests.

重要的区别在于，是使用 AI，还是围绕 AI 重组整个公司。电力提供了一个有益的类比。从引入电力到围绕电力重新设计工厂，大约花了四十年。起初，公司只是将电力作为动力源，替换到原本按照煤炭组织的工厂中。那些有雄心的公司最终重建了工厂。如今最有雄心的公司正在围绕模型重新思考其组织方式；较早采用或雄心较小的公司，则是在让现有团队和岗位职能使用新的工具。The important distinction is between using AI and reorganizing the entire company around AI. Electricity offers a useful analogy. It took roughly forty years to move from introducing electricity to redesigning factories around it. At first, companies simply swapped an electric power source into a factory organized for coal. The ambitious companies eventually rebuilt the factory. The most ambitious companies today are rethinking their organizations around models; earlier or less ambitious adopters are giving existing teams and job functions access to new tools.

伦尼：这样人们就不用那么担心自己的工作和职业生涯的未来了？Lenny: So people can be less stressed about the future of their jobs and careers?

阿尼什：我想是的。想想首席执行官和高管们的激励因素。桑达尔·皮查伊不只是想运营一家效率更高、价值四万亿美元的谷歌；他想打造一家价值四十万亿美元的公司。如果某个部门变得更具经济生产力，理性的应对方式可能是保留它，并扩大其业务范围。Anish: I think so. Consider the incentives of CEOs and executives. Sundar Pichai does not want merely to operate a more efficient four-trillion-dollar Google; he wants to build a forty-trillion-dollar company. If a unit becomes more economically productive, the rational response can be to keep it and expand what it does.

一位 Google 高管告诉我，他的团队并没有裁员。相反，他们快速推进了路线图：原本计划两年完成的工作在三个月内就完成了。他们最棘手的问题变成了决定还要在路线图中添加什么——这是每位产品经理在经历多年令人心力交瘁的优先级讨论后所幻想的局面。A Google executive told me that his team did not lay people off. Instead, it tore through the roadmap: two years of planned work happened in three months. Their hardest problem became deciding what else to add to the roadmap, which is every product manager’s fantasy after years of emotionally draining prioritization conversations.

人们应该感到自己拥有掌控力，而实现这一点的最佳方式就是把产品交付出去。Claire Vo 就是这方面的典范：她总是在创造事物、尝试想法，也坦然接受一点尴尬。正因如此，她似乎成为了最好的自己。我们都有这样的机会。People should feel empowered, and the best way to get there is to ship. Claire Vo is a model for this: she is always making things, trying ideas, and accepting a little embarrassment. She seems like the best version of herself because of it. We all have that opportunity.

作为级联循环的公司Companies as Cascading Loops

Lenny：你曾说，创建公司正日益变成一系列循环。你是什么意思？Lenny: You have said company building is increasingly becoming a series of loops. What do you mean?

Anish：我们从提示开始，然后发展到代理。代理是一个处于循环中的模型，配备工具、记忆和技能文件。下一个层级是由代理组成的循环，它们执行相互关联的任务。编码清楚地展示了这一点，因为最强大的模型服务于技术能力高超的用户，而软件工作本身已经包含了成熟且可验证的循环。Anish: We began with prompts, then agents. An agent is a model in a loop with tools, memory, and skill files. The next level is loops made of agents doing linked tasks. Coding demonstrates this clearly because the strongest models serve technically sophisticated users, and software work already contains established, verifiable loops.

一份错误报告提交后，系统会生成复现步骤，创建并审查修复方案，同时通过风险评估确定是否必须由人工批准部署到生产环境。低风险修复可能会自动发布，随后发送一封电子邮件告知客户，其报告的错误已得到修复。整个流程可能在几分钟内完成。工程领域包含许多这样的循环，包括错误修复、客户反馈处理、销售演示和功能开发。A bug report arrives, a reproduction is generated, a fix is created and reviewed, and a risk assessment determines whether a human must approve production deployment. A low-risk fix might ship automatically, followed by an email telling the customer that the reported bug was fixed. The entire sequence can happen in minutes. Engineering contains many such loops, including bug fixing, customer-feedback processing, sales demos, and feature development.

更大的问题在于，业务闭环究竟是什么样的。总经理负责监督编码、营销、销售、支持和法务等环节中的闭环。它们的综合产出本身也可以形成一个闭环，并接受衡量和优化。在最强大的形式下，这套系统甚至可能告诉 CEO，公司必须改变业务的某个实体部分、商业模式或战略。我们将看到，这一过程会从每个人一个闭环，发展到每项职能一个闭环，再到跨业务部门的多个闭环，最终形成贯穿公司大部分业务的闭环。The larger question is what business loops look like. A general manager oversees loops in coding, marketing, sales, support, and legal. Their combined output can itself become a loop that is measured and optimized. In its strongest form, that system might tell the CEO that the company must change a physical part of the business, its business model, or its strategy. We will see a cascade from a loop per person, to a loop per function, to loops across business units, and eventually loops running large parts of a company.

人类仍然至关重要。大多数组织工作无法实现完全自主。在 AI 原生公司中，人员仍将是销售、支持、战略制定和异常情况处理的核心。模型仍然难以进行真正新颖、分布外的思考。数学推理方面的进步并不能证明模型能够提出强有力的商业战略。仍然需要有人决定应该做什么，并且判断正确。Humans remain critical. Most organizational work cannot be fully autonomous. In an AI-native company, people will remain central to sales, support, strategy, and exceptions. Models still struggle with genuinely new, out-of-distribution thinking. Progress in mathematical reasoning does not prove that models can originate strong business strategy. A person still has to decide what should be made and be right about it.

Lenny：工程工作正逐渐变成一个闭环：从支持工单等输入开始，经过决定要构建什么、创建拉取请求、进行审查，最终发布上线。你预计这种模式会扩展到市场推广、法务、增长和支持领域。那会是什么样？Lenny: Engineering is becoming a loop from an input such as a support ticket, through deciding what to build, creating a pull request, reviewing it, and shipping. You expect that pattern to spread to go-to-market, legal, growth, and support. What would that look like?

Anish：考虑一下传统的增长团队。团队成员聚集起来，头脑风暴各种实验，对实验进行排序，构建并发布它们，然后衡量结果。在基于循环的版本中，每个变体都可以生成并进行衡量。一旦某个变体在所需的置信水平下达到统计显著性，就可以将其合并并发布。在系统开始下一项实验的同时，一个长期留置对照组仍会保持不变。Anish: Consider a traditional growth team. People gather, brainstorm experiments, prioritize them, build them, ship them, and measure the results. In a loop-based version, every variant can be generated and measured. Once a variant reaches statistical significance at the required confidence level, it can be merged and shipped. A long-term holdout remains in place while the system begins the next experiment.

最终，这个循环会达到一个局部最大值。这一点至关重要：循环能帮助你登上山顶，但随后便会陷入平台期。要将组织置于下一座山的山脚下，需要分布外思维和人类直觉。智能体负责攀登；人类决定哪座山更重要。Eventually, that loop reaches a local maximum. This is crucial: the loop helps you climb the hill, but then it plateaus. Out-of-distribution thinking and human intuition are needed to place the organization at the base of the next hill. Agents climb; humans decide which hill matters.

如果你对代理说：“帮我赚一百万美元，而且不要犯任何错误”，这并不起作用，因为代理仍然需要一个有用的方向。如今的技术没有任何迹象表明人类的指导是不必要的。If you ask an agent, “Make me a million dollars and make no mistakes,” it does not work because the agent still needs a useful direction. Nothing in today’s technology suggests that human direction is unnecessary.

Lenny：我听说 OpenAI 的市场推广团队使用 Codex 的频率可能比工程团队还高。Lenny: I heard that OpenAI’s go-to-market team may use Codex even more than its engineering team.

Anish：这应该会让市场推广团队感到高兴。他们的核心工作是建立关系、说服客户并促成业务。AI 可以消除围绕这些工作的行政事务，将工作重点提炼到他们最擅长、也最享受的部分。Anish: That should make go-to-market teams happy. Their core work is relationships, persuasion, and closing business. AI can remove the administration surrounding that work, distilling the job toward what they are best at and enjoy most.

Lenny：一位产品经理朋友开玩笑说，与其拒绝各种想法并把它们列入路线图，不如对所有事情都说“是”，实现每一个想法，模拟用户的反应，然后让证据来做决定。Lenny: A product-manager friend joked that instead of saying no to ideas and putting them on a roadmap, he could say yes to everything, build every idea, simulate user reactions, and let the evidence decide.

Anish：这会带来两个美妙的结果。首先，许多产品经理认为自己是出色的从零到一型思考者，只是受管理层、创始人、高管或工程能力所掣肘。如果每个产品故事都能被讲述、每项功能都能被尝试，那么其中一些人可能会发现，相比于推进自己薄弱的想法，他们更乐于致力于他人有力的创意。这种诚实或许能改善组织健康。Anish: That has two beautiful consequences. First, many PMs believe they are exceptional zero-to-one thinkers held back by management, founders, executives, or engineering capacity. If every product story can be told and every feature tried, some may discover they are happier working on someone else’s strong idea than their own weak one. That honesty could improve organizational health.

其次，胜出的想法不再需要来自最擅长向高管推销它的人。各种想法都可以尝试，最好的想法可以胜出。Second, the winning idea no longer needs to come from the person who can sell it best to an executive. Ideas can be tried, and the best one can win.

Lenny：哪些技能因为其输出难以验证，而难以形成循环？Lenny: Which skills resist becoming loops because their output is difficult to verify?

Anish：这些技能会成为限制速率的因素。每当模型犯错时，一个有用的诊断问题是：你知道什么，而它不知道什么？Kavak 为每位客户配备了一个代理。当代理陷入困境时，它会呼叫人工人员，由人工人员指导它解决当前情况。这种干预既能帮助客户摆脱困境，也能生成一条代理可以从中学习的轨迹。失败通常会暴露出知识缺口或数据缺口。补充缺失的上下文，理想情况下，代理下次就不需要再呼叫人工人员了。Anish: Those skills become rate-limiting factors. A useful diagnostic whenever a model makes a mistake is: what do you know that it does not know? Kavak has an agent for each customer. When an agent gets stuck, it calls a human, who coaches it through the situation. That intervention both unblocks the customer and produces a trace the agent can learn from. The failure usually exposes a knowledge gap or a data gap. Supply that missing context, and ideally the agent will not need to call next time.

Lenny：因此，每个函数都应被视为一个代理循环。人的任务是找出它受阻或出错的地方，并提供上下文、洞察或指导。Lenny: So every function should be viewed as an agent loop. The human’s job is to identify where it gets blocked or goes wrong and provide context, insight, or direction.

Anish：没错。一旦一个新的产品创意奏效，它对营销、销售、产品营销、传播和法务的影响，主要应由周边的循环来处理。你的工作可以变成去远足、做梦，然后选择下一个要攀登的山头。Anish: Exactly. Once a new product idea works, its implications for marketing, sales, product marketing, communications, and legal should be handled largely by the surrounding loops. Your job can become taking a hike, dreaming, and choosing the next hill.

Lenny：如果 AI 完成了这么多工作，是什么让成功的公司脱颖而出？Lenny: What separates winning companies if AI does so much of the work?

Anish：许多行业的竞争均衡可能不会发生剧烈变化。如果必胜客、达美乐、棒约翰和圆桌披萨都获得了一座充满博士的数据中心，它们都会采用这项技术，或者更换首席执行官，直到做到这一点。期间可能会经历一段动荡期，市场相对份额也会重新洗牌，但大多数成熟企业都会采用有用的技术。这并不意味着最终会有一家企业占据百分之九十九的市场，尤其是在那些不受智能能力约束的行业中。Anish: Competitive equilibria in many industries may not change dramatically. If Pizza Hut, Domino’s, Papa John’s, and Round Table all receive a data center full of PhDs, they will all adopt it, or replace a CEO until they do. There may be a rocky period and some relative shuffling, but most established companies adopt useful technology. It does not follow that one ends up with ninety-nine percent of the market, especially in industries that are not intelligence-bound.

短期内，谁能更快、更有魄力地采用这些技术，谁就会成为赢家，否则就会成为输家。对于创始人或 CEO 而言，最有用的问题是：如果智能变得实际上无限且便宜得惊人，我们将如何重组公司？这就是发展方向。In the near term, there will be winners and losers based on how quickly and ambitiously they adopt. The most useful question for a founder or CEO is: if intelligence becomes effectively unlimited and astonishingly cheap, how would we reorganize the company? That is the direction of travel.

为不同类型的工作选择模型Choosing Models for Different Kinds of Work

Lenny：你认为公司内部会在通才和专才之间出现分化。模型选择如何符合这一图景？Lenny: You expect a split between generalists and specialists inside companies. How does model choice fit that picture?

Anish：可以用帕累托前沿来思考：性能与价格之间的有效权衡。前沿模型的定价可能看起来不合理地高。一款假设中尚未发布的模型，每个 token 的成本是无穷美元，因为没有人能使用它。为了获得一点概念上的额外智能，前沿模型的成本可能是强大的旧模型的一百倍。Anish: Think in terms of a Pareto frontier: the efficient trade-off between performance and price. Frontier models can look irrationally priced. A hypothetical unreleased model costs infinite dollars per token because no one can use it. For one conceptual point of extra intelligence, a frontier model may cost a hundred times more than a strong older model.

在一个上行空间无上限的领域，这仍然可能是理性的。在药物研发中，如果多一点智能就能促成下一种他汀类药物的诞生，那么其结果可能价值一万亿美元。应为可获得的最高智能水平付费。其他任务的上行空间则是有限的。在会计工作中，账目可以被正确结清，但不可能做到正确一百倍。对于这类工作，性价比更为重要。That can still be rational in a field with unbounded upside. In drug discovery, if one extra point of intelligence leads to the next statin, the outcome may be worth a trillion dollars. Pay for the highest intelligence available. Other tasks have bounded upside. In accounting, the books can be closed correctly, but not a hundred times more correctly. For such jobs, price-performance efficiency matters more.

因此，组织将采用这两种架构。范围狭窄且边界明确的任务，可以使用经过强化学习调优、专门用于某项工作的低成本开放权重模型。销售、支持、研究或工程等具有高潜在回报的工作，可能足以证明使用昂贵的前沿模型调用是合理的。不同模型家族也各有相对优势，因此最终不会只是开放权重模型或前沿模型，也不会是一个通用模型包打天下。Organizations will therefore use both architectures. Narrow, bounded tasks may run on inexpensive open-weight models tuned with reinforcement learning for a specific job. High-upside work in sales, support, research, or engineering may justify expensive frontier tokens. Model families also have comparative advantages, so it will not simply be open weights or frontier models, nor one universal model.

Lenny：这种区别也与可验证性有关吗？Lenny: Is the distinction also about verifiability?

Anish：不完全是。药物发现可能具有无上限的上行空间，而且可以验证。结账同样可以验证，但其上行空间是有限的。关键问题在于，上行空间有多大，以及识别它有多难。Anish: Not exactly. Drug discovery may have unbounded upside and be verifiable. Closing the books is also verifiable but has bounded upside. The key questions are how much upside exists and how hard it is to recognize.

客户支持增加了细微差别。客户可能报告一个错误，而这可能是通向某种洞见的第一条线索，这种洞见足以改变整个组织。电话会议中的 CEO 或洞察力极强的人可能会顺着这条线索追查下去，而一个仅仅称职的通才可能不会。这说明，应将前沿智能应用于整组客户支持问题。反方观点是，我们可能已经跨过了几乎所有具有经济价值的问题所需的智能阈值，使得进一步提升智能变得得不偿失。两种情况都值得考虑。Customer support adds nuance. A customer may report a bug that is the first breadcrumb toward an insight capable of changing the entire organization. A CEO or highly perceptive person on the call might follow that trail, while a merely adequate generalist may not. That argues for applying frontier intelligence to the whole basket of support problems. The counterargument is that we may already have crossed the intelligence threshold for almost every economically useful problem, making further intelligence wasteful. Both cases deserve consideration.

Lenny：人们忘了，如今的非前沿模型在六个月前还是前沿模型。那时我们对它们惊叹不已，但一旦有更好的东西出现，就立刻对它们不屑一顾。Lenny: People forget that today’s non-frontier models were frontier models six months ago. We were amazed by them then, but discount them as soon as something better arrives.

阿尼什：没错。今天我们把某种东西称为 AGI；第二天，它就成了被扔进垃圾桶的旧事物。Anish: Exactly. One day we call something AGI; the next day it becomes the old thing thrown in the dustbin.

Lenny：人们开玩笑地称你为“模型侍酒师”。当前最先进的水平是什么？每个模型分别擅长和不擅长什么？Lenny: People jokingly call you a model sommelier. What is the current state of the art, and what is each model good or bad at?

Anish：就像葡萄酒侍酒师一样，秘诀在于数千个小时——或者说，数千瓶酒。成为模型侍酒师的方法，就是把它们全部用一遍。我会逼自己在每个新模型发布时都交付一些成果。那些说模型可以互换的人，通常并没有深入使用过它们。Anish: Like a wine sommelier, the secret is thousands of hours—or bottles. The way to become a model sommelier is to use them all. I push myself to ship something with every new model. People who say models are interchangeable usually have not used them deeply.

我最近迷上了 Qwen 38B。它擅长处理长周期任务，富有创造力，而且很会讲故事。我用它制作了关于《星球大战》宇宙中各个行星的“虚构纪录片”，包括塔图因和贝斯平。它可以连续工作四五个小时，讲述故事，通过 MiniMax 模型生成视频，生成音频，执导一部五分钟的电影，剪辑片段，将它们叠加起来，并审看最终成果。I recently became obsessed with Qwen 38B. It is strong at long-horizon tasks, creative, and a good storyteller. I used it to make “impossible documentaries” about planets in the Star Wars universe, including Tatooine and Bespin. It could work for four or five hours, tell a story, generate video through a MiniMax model, generate audio, direct a five-minute film, cut the clips, overlay them, and review the result.

那个模型的形态与我曾用于产品工作的 GLM 53 大不相同。GLM 不具备视觉能力，给人的感觉就像一个被安排在角落里的神经质博士。两者各有其作用。一个模型的思维可能偏向创造力和开放性；另一个则偏向精确性和富有成效的神经质。我会用每个模型各做一些东西，以培养对这些差异的直觉。That model has a very different shape from GLM 53, which I used for product work. GLM lacks vision and feels like a neurotic PhD placed in the corner. Both have a role. One model’s mind may be shaped toward creativity and openness; another toward precision and productive neuroticism. I make something with each model to build intuition about those differences.

通过动手实践建立直觉Building Intuition by Making Things

Lenny：这个习惯有多重要，以及一个人应该如何决定要制作什么？Lenny: How important is that habit, and how should someone decide what to make?

Anish：我们都需要成为那个曾经令人难以忍受的“应用创意”朋友。把那些傻乎乎的点子做出来，尤其是因为它们可能蕴含着最意想不到的价值。我有大约二十几个小应用，还有一两个一直在反复迭代的大型项目。一个持续推进的项目能为你测试每个新模型提供一个框架；否则，你每次都必须从头构思一个新点子。Anish: We all need to become the once-insufferable “app idea” friend. Build the silly ideas, especially because they may contain the most unexpected value. I have about two dozen small apps and one or two larger projects I keep iterating on. A continuing project gives you a chassis for testing each new model; otherwise, you must invent a new idea from scratch every time.

着手做点什么，最好不要是那种首字母大写的“重要”之事。不断寻找用模型作为工具为其增添内容的方法，而不是把模型本身当作目标。Work on something, preferably something that is not Important with a capital I. Keep finding ways to add to it using models as tools rather than treating the model itself as the goal.

Lenny：一个有用的习惯是在做某件事之前先停下来问自己：“AI 能如何帮我完成这件事？”这就像在刺激与反应之间创造一个空间，但利用这个空间来思考 AI 能如何提供帮助。Lenny: One useful habit is to pause before doing something and ask, “How can AI do this for me?” It is like creating a space between stimulus and response, but using that space to consider how AI can help.

Anish：没错。在日常知识工作中，自动化的机会有时对我来说并不那么明显，因为我天生更关注产品。相比于找出一项可以自动化的生活事务，我更容易设想出一个适用于我的 DJ 流媒体应用或我的个人版 X 的 Google Reader 的功能。Anish: That is right. In ordinary knowledge work, the automation opportunity is sometimes less obvious to me because I am naturally drawn to products. It is easier for me to imagine a feature for my DJ streaming app or my personal Google Reader for X than to find a life task to automate.

但小型实验很有趣。我曾经让一台笔记本电脑记录厨房里发生的事情，然后根据我儿子的评论是否积极且有利于社交，给他的 iPad 增加或减少屏幕使用时间。他通过反复录下自己说“我爱你，爸爸”，再把录音放在麦克风附近，成功钻了这个指标的空子。一旦某项衡量标准变成了目标，它就不再有用——但这个实验依然令人愉快。But small experiments are fun. I once had a laptop transcribe what was happening in my kitchen, then add or subtract screen time from my son’s iPad according to whether his comments were positive and prosocial. He hacked the metric by recording himself saying, “I love you, Dad,” repeatedly and placing the recording near the microphone. Once a measure becomes a target, it stops being useful—but the experiment was still delightful.

Lenny：一位朋友改装了一个可穿戴设备，用来统计他和孩子们白天笑了多少次。他们会一起查看这项指标。Lenny: A friend modified a wearable to count how often he and his kids laughed during the day. They review that metric together.

Anish：这揭示了一个大得多的机遇。我们讨论过生产力和失业问题，但统计家庭中的笑声并不是一家初创公司，也可能不会带来显著的经济影响。然而，它仍然可能极大地影响生活质量。我们常常假设幸福是固定不变的。为什么会这样？如今我们的工作比大多数人在一个世纪前可能从事的工作有趣得多，而一个世纪后的工作或许还会好得多。即使没有任何经济影响，AI 也能为人类生活增添层次。Anish: That captures a much bigger opportunity. We have discussed productivity and job loss, but counting family laughter is not a startup and may not be economically consequential. It can still matter enormously to the quality of a life. We often assume happiness is fixed. Why should it be? Our jobs today are far more interesting than the jobs most of us would have had a century ago, and jobs a century from now may be far better again. AI can add texture to human life even when the effect has no economic consequence.

“让我更开心”循环The “Make Me Happier” Loop

Lenny：你曾说过，巨大的机会可能在于“让我更快乐”这一循环。这是什么意思？Lenny: You have said the big opportunity may be the loop “make me happier.” What does that mean?

Anish：我们说人们希望变得更高效，但很多人并不希望如此。相比节省时间，更多人想要的是消磨时间。世界上最大的产品是娱乐产品和社交产品。真正的消费者问题在于，如何提供人们真正渴望的价值。Anish: We say people want to become more productive, but many do not. More people want to spend time than save it. The world’s largest products are entertainment and social products. The real consumer question is how to deliver value people genuinely desire.

有一类“X AI 用户”，他们害怕自己跌入永久性的底层阶级，并且对每一个新模型都有详尽的看法。还有一类“Instagram AI 用户”，他们认为 AI 只是稍微好用一点的 Google 搜索，并纳闷人们为何如此追捧。这种差距是产品设计上的失败。我们已经拥有能够改变普通人生活的能力。There is an “X AI user” who fears falling into the permanent underclass and has detailed opinions about every new model. Then there is an “Instagram AI user” who thinks AI is a slightly better Google search and wonders what the hype is about. That gap is a product-design failure. We already possess capabilities that could transform ordinary lives.

四十年来，我们在很大程度上打造的是能生成更好电子表格的技术。它拓展了智力，却几乎没有拓展灵魂。与此同时，许多满足人们精神和情感需求的文化机构日渐衰弱。如今的机遇在于回应人类的基本需求：如何让人们感受到更紧密的联结与更多的爱，取得进步，并享受乐趣？这从根本上说并不是模型能力问题，而是产品设计问题。For forty years, we largely built technology that produced better spreadsheets. It extended the intellect but did little to extend the soul. Meanwhile, many cultural institutions that met spiritual and emotional needs weakened. The opportunity is to address basic human needs: how can people feel more connected and loved, make progress, and have fun? This is not fundamentally a model-capability problem. It is a product-design problem.

Lenny：我们讨论过“发展我的业务”“寻找更多销售机会”和“解决支持工单”等循环。你描述的是“改善我的健康状况”或“让我成为更好的朋友”等循环。AI 可以持续思考这些目标，并帮助你努力实现它们。Lenny: We have discussed loops such as “grow my business,” “find more sales,” and “close support tickets.” You are describing loops such as “improve my health” or “make me a better friend.” AI can think continuously about those goals and help pursue them.

Anish：是的，而且它可能需要向我们发起挑战。一款有用的 AI 可能会不讨人喜欢，或促使我们改变。初创公司具有优势，因为大型在位企业有各种委员会，会抵制令人不适的产品行为。然而，分歧、性、尴尬和紧张都是人类存在的一部分。初创公司可以更自由地探索社交生活中令人不适的部分。Anish: Yes, and it may need to challenge us. A useful AI may be disagreeable or push us. Startups have an advantage because large incumbents have committees that resist uncomfortable product behavior. Yet disagreement, sexuality, awkwardness, and tension are all parts of human existence. Startups can explore the uncomfortable parts of social life more freely.

Lenny：因此，消费类公司可以围绕改善与朋友和家人之间的联系来打造产品。Lenny: So consumer companies can build products around improving connections with friends and family.

Anish：是的。面向消费者的 AI 仍处于早期阶段，或许可以类比 2010 年前后的 iPhone——那时 Airbnb、WhatsApp、Uber 以及许多定义移动互联网时代的公司尚未出现。三件事阻碍了它的发展。对于许多免费使用的产品而言，模型成本过高。聊天对于行动力极强的用户来说是一个很好的界面，但普通消费者的理想界面更像 TikTok，因此我们需要介于聊天和被动信息流之间的界面。此外，开发重点过度放在生产力上，而不是连接或娱乐上。Anish: Yes. Consumer AI is still early, perhaps comparable to the iPhone around 2010, before Airbnb, WhatsApp, Uber, and many defining mobile companies. Three things have held it back. Models were too expensive for many free-to-use products. Chat is a good interface for an extremely high-agency user, but the average consumer’s ideal interface looks more like TikTok, so we need interfaces between chat and passive feeds. And development focused too heavily on productivity rather than connection or entertainment.

这些限制正在逐步缓解。开放权重模型让推理成本大幅降低。人们开始讨论“让我更快乐”之类的目标。创业者们正在探索更加雄心勃勃的界面。这些问题至少比两年前更有望得到解决。Those constraints are easing. Open-weight models make inference much cheaper. People are beginning to discuss goals such as “make me happier.” Founders are exploring far more ambitious interfaces. The problems are at least better positioned to be solved than they were two years ago.

人工智能、人类能动性与不断扩大的抱负AI, Human Agency, and Expanding Ambition

Lenny：你为什么认为人们会低估人工智能带来的积极可能性？Lenny: Why do you think people underestimate the positive possibilities of AI?

Anish：人工智能可以成为一种情感与精神层面的接口，赋予人们探索自我深藏部分的能力。工业革命创造了巨大的规模化优势，但这种集中化可能会打击个人积极性并削弱身份认同。人工智能则朝着另一个方向发展：通过将技能与欲望解绑，放大身份认同和自主能动性。Anish: AI can become an emotional and spiritual interface that gives people leverage to explore buried parts of themselves. The Industrial Revolution created immense advantages of scale, but that centralization can discourage individuals and diminish identity. AI moves in another direction: it amplifies identity and agency by unbundling skill from desire.

如果你想创作音乐，现在无需先精通钢琴就能创作音乐。如果你想开发软件，也无需接受程序员训练就能构建软件。这项技术让人们能够表达个性，并探索生活中那些以往无法触及的领域。If you want to make music, you can now make music without first mastering piano. If you want to create software, you can build software without having trained as a programmer. The technology lets people express individuality and explore parts of life that were previously inaccessible.

从经济角度看，没有任何规律要求增长必须维持在百分之二左右。为什么不能达到百分之十、十五甚至百分之二十？人工智能可以同时提升生产率和抱负。20世纪50年代和60年代，在看到整个经济体在世界大战期间实现全面动员后，社会相信自己能够完成非凡的事业。我的理论是，当利害关系重大时，人们会表现得非常出色；而当利害关系较小时，人们的表现则处于最差状态。Economically, there is no law requiring growth to remain around two percent. Why could it not be ten, fifteen, or twenty percent? AI can raise productivity and ambition simultaneously. In the 1950s and 1960s, society believed it could do extraordinary things after seeing entire economies mobilize during a world war. My theory is that people are excellent when the stakes are high and at their worst when the stakes are low.

五年前的世界常常让人觉得一切都无关紧要，社会积累了许多既没有让我们更高效、也没有让我们更幸福的副业项目。如今，许多人觉得自己正在攀登一架野心阶梯。你发布糟糕的创意，是为了发现好的创意。你可以制作软件、音乐，甚至着手开展实体项目，而不必亲自掌握每一项前置技能。我们拥有让社会变得更加充实、更具生产力的要素，可我们却不断回到“永久底层阶级”的叙事中。The world five years ago often felt low-stakes, and society accumulated side projects that did not make us more productive or happy. Now many people feel they are climbing an ambition ladder. You ship bad ideas to discover good ones. You can make software, music, or even approach physical projects without personally possessing every prerequisite skill. We have the ingredients for a more fulfilled and productive society, yet we keep returning to the permanent-underclass story.

Lenny：确实存在一些负面影响和干扰，包括数据中心带来的影响，但到目前为止，就业和经济一直保持稳健。医疗进步也在加速。Lenny: There are real downsides and disruptions, including the effects of data centers, but employment and the economy have held up so far. Medical progress is also accelerating.

阿尼什：有人竟然能够认真地讨论治愈所有疾病之后会发生什么，而社会还可以把这当作一种真实的可能性来辩论，这令人惊叹。我们往往会担忧“世界”“普通劳动者”或“生活在数据中心附近的人”之类的抽象概念。但我们的实际生活常常表明，我们感到自己更有能力，也更有掌控力。Anish: The fact that someone can seriously write about what happens after we cure every disease, and society can debate it as a real possibility, is astonishing. We tend to worry about abstractions such as “the world,” “the average worker,” or “the person living near a data center.” Our actual lives often reveal that we feel more capable and empowered.

这就造成了口头表达的偏好与实际表现出的偏好之间的落差。人们可能会说自己不希望附近有一座数据中心，却每天都在使用 ChatGPT。改善公众对人工智能观感的最佳方式，是让重要的事物变得廉价。美国的两项基本需求——医疗保健和教育——却变得越来越昂贵。That creates a gap between stated and revealed preferences. People may say they do not want a data center nearby while using ChatGPT every day. The best way to improve AI’s public story is to make important things cheap. Two essentials in America—health care and education—have only become more expensive.

大约45%的医疗保健支出用于行政管理。减轻其中大部分负担，或许能与医学进步一道，使成本出现通缩趋势。如今，教育正面临两个世纪以来最强劲的竞争。人工智能可以将学习从教育机构中解绑，也可以将社会地位与学历证书脱钩。如果人们能够证明自己知道如何完成这项工作，他们可能就不需要某个特定学位。Roughly forty-five percent of health-care spending is administrative. Removing much of that burden could make costs deflationary, alongside advances in medicine. Education now faces its strongest competition in two centuries. AI can unbundle learning from institutions and status from credentials. People may not need a particular degree if they can demonstrate that they know how to do the work.

Lenny：据报道，一些实验室因对模型能力的担忧而放缓了部分模型开发。人们应如何理解“某个模型危险性过高、不宜发布”这样的说法？Lenny: Some labs have reportedly slowed parts of model development because of capabilities concerns. How should people interpret claims that a model is too dangerous to release?

Anish：不针对任何具体实验室发表评论，可能存在许多混杂因素。一家公司声称某个模型过于危险，可能会因此获得强大的光环效应。它也可能面临 GPU 短缺，更倾向于将竞争优势保留在内部，或者出于经济原因不愿将自身优势对外开放。我认真对待进攻性网络安全；我们应在让系统变得易于攻破之前，先强化这些系统。但“危险到不能发布”可能将真正的能力担忧与营销、推理能力和竞争策略混杂在一起。Anish: Without commenting on any specific lab, there can be many confounding factors. A company may gain a powerful aura from saying a model is too dangerous. It may also have a GPU shortage, prefer to keep a competitive capability internal, or have economic reasons not to externalize its advantage. I take offensive cybersecurity seriously; we should harden systems before making them trivial to penetrate. But “too dangerous to release” can mix genuine capability concerns with marketing, inference capacity, and competitive strategy.

行业证据同样没有显示出一个无可争议的赢家。Anthropic 曾有一段时间看起来势不可挡，随后 OpenAI 又经历了一段强势期，而开放权重模型也在持续改进。这并不像围绕某个极其智能的专有模型形成永久垄断时会呈现的局面。The industry evidence also does not show one unassailable winner. Anthropic looked untouchable for a period, then OpenAI had a strong stretch, while open-weight models continued improving. That is not how a permanent monopoly around one supremely intelligent proprietary model would look.

Lenny：还有什么能让人们对工作感到更满意？Lenny: What else should make people feel better about jobs?

阿尼什：纵观人类历史，欲望的增长速度一直快于我们满足欲望的能力。如今被视为基本需求的事物，在五百年前难以想象；即便是在一百年前，抗生素也还是无法想象的奢侈品。我们低估了人类的欲望与抱负。二十年后，人们可能会抱怨自己在火星上没有一栋度假屋。每一位首席执行官都将希望打造一家更大的公司。人类的存在已经以我们的祖先无法想象的规模运转，而这一趋势没有理由必须停止。Anish: Throughout human history, desires have grown faster than our ability to fulfill them. Things considered basic expectations today were unimaginable luxuries five hundred years ago—or even a century ago, in the case of antibiotics. We underestimate human desire and ambition. In twenty years, people may complain that they lack a vacation home on Mars. Every CEO will want to build a bigger company. Human existence already operates at a scale our ancestors could not imagine, and there is no reason that trend must stop.

不应将抱负局限于创办公司或制作精美的软件。还有创造性的抱负：孩子会在任何人告诉他们擅长或不擅长哪种媒介之前就开始创作。还有地方公共事务方面的抱负，例如让英国国民保健服务（NHS）成为像 iPhone 一样良好的体验。还有家庭方面的抱负：作为父母，与家人建立更紧密的联系，更加用心地陪伴他们。抱负可以是人们想要更充分地去做的任何事情。Ambition should not be reduced to founding a company or making beautiful software. There is creative ambition: a child makes things before anyone tells them what medium they are good or bad at. There is local civic ambition, such as making the United Kingdom’s NHS as good an experience as the iPhone. There is family ambition: being more connected and present as a parent. Ambition can mean anything people want to do more fully.

Lenny：产品人员受过围绕 MVP 和约束进行思考的训练。现在，我们可能需要设想十倍乃至千倍规模的版本。Lenny: Product people are trained to think about MVPs and constraints. Now we may need to ask for the ten-times or thousand-times version.

Anish：我们的生活曾围绕着软件和智能的稀缺性而组织。如今，它们正变得无处不在，而对于经验丰富的技术从业者来说，适应这一变化可能比对新入行者更难。Anish: Our lives were organized around software and intelligence being scarce. They are becoming abundant, and adjusting to that may be harder for experienced technology workers than for newcomers.

实践中蕴含着巨大的学习价值。谈论“氛围编程”的人比分享自己做出的项目的人多，或许是因为这些项目看起来不够重要。但执行和交付会带来学习与满足感。构建可以成为新的阅读方式：通过构建来学习和体验，而大部分成果最终被舍弃也是可以接受的，因为这一活动会增强你的能力。构建是一种活动，而不仅仅是结果。我可以制作一套只有三个人听过的 DJ 音乐串烧，而自己却反复听上一百遍；这依然令人满足。There is enormous learning in doing. More people talk about vibe coding than share the projects they made, perhaps because the projects seem insufficiently important. But execution and shipping produce learning and fulfillment. Building can become the new reading: you build to learn and experience, and it is acceptable to discard most of the result because the activity strengthens your abilities. Building is an activity, not only an outcome. I can make a DJ set that three people hear while I listen to it a hundred times; it remains fulfilling.

三个消费者 AI 机会Three Consumer AI Opportunities

Lenny：你在 a16z 负责消费者投资。哪些领域最令人兴奋？Lenny: You focus on consumer investing at a16z. Which areas are most exciting?

阿尼什：我看到三个。第一个是编程代理。人们很容易断言，大多数消费者不想编写代码，但编程代理正逐渐成为与世界交互的通用界面。人们用它们来剪辑视频，或创建游戏，在飞行途中与孩子一起玩。Wabi 就是一个例子：一个供人们创建、使用和分享小型应用的平台。编程正成为一种通用的问题解决媒介。Anish: I see three. The first is coding agents. It is easy to say that most consumers do not want to write code, but coding agents are becoming a general interface to the world. People use them to edit videos or create games to play with their children on a flight. Wabi is one example: a platform where people create, consume, and share mini-apps. Coding becomes a general problem-solving medium.

第二类是个人智能体。开发者项目展示了各种可能性，而如今，这些理念正被提炼为可靠的消费者和企业助手，让普通人也能理解。个人智能体能够保留上下文、操作软件并完成任务，而不仅仅是回答问题。The second is personal agents. Developer projects demonstrated what was possible, and those ideas are now being distilled into reliable consumer and enterprise assistants that ordinary people understand. Personal agents can hold context, operate software, and complete tasks rather than merely answer questions.

第三类是一个广泛的娱乐与创意类别，包括 Suno 等工具和陪伴类产品。“娱乐”一词并不能完全概括这一类别。其中一部分领域让人难以启齿，这也是增长迅速的产品仍未得到充分研究的原因。陪伴类产品的用户并不只是刻板印象中的年轻男性；许多用户是四五十岁的女性。The third is a broad entertainment and creative category that includes tools such as Suno and companionship products. “Entertainment” does not fully capture it. Some of this space is uncomfortable to discuss, which is why fast-growing products remain underexamined. Companion products are not used only by stereotypical young men; many users are women in their forties and fifties.

这些产品共同为抱负、成就感和休闲娱乐提供了实现渠道。它们对应着人类长期以来需要完成的任务：让我更快乐，让我更健康，帮助我活得更久，并帮助我进行创作。Together, these products offer channels for ambition, fulfillment, and leisure. They map to enduring human jobs to be done: make me happier, make me healthier, help me live longer, and help me create.

Lenny：在个人代理产品中，哪些产品脱颖而出？Lenny: Among personal agents, which products stand out?

Anish：Grok 的代理产品在凭据保留和完成工作方面采取了激进的选择，并有强大的模型和周到的界面作为支持。ChatGPT 的代理式工作产品在界面中有些隐蔽，但表现非常出色。一家名为 Instinct 的初创公司在同一领域做出了更为激进的权衡。Anish: Grok’s agent product made aggressive choices around retaining credentials and getting work done, supported by a strong model and thoughtful interface. ChatGPT’s agentic work product is somewhat buried in the interface but very good. A startup called Instinct has made more aggressive trade-offs in the same domain.

云端执行和缓存的浏览器凭据至关重要。全双工语音尤其具有吸引力：你可以呼叫一个能看到你当前活跃线程的助手，询问各个编码代理正在发生什么，并让它进行修改。这种体验就像是在呼叫一个了解你所处世界的助手，而不是提交单向转录内容或文本指令。Cloud execution and cached browser credentials matter. Full-duplex voice is especially compelling: you can call an assistant that sees your active threads, ask what is happening across coding agents, and tell it to change something. That feels like calling an assistant who understands your world, rather than submitting a one-way transcription or text instruction.

护城河、工艺与分销Moats, Craft, and Distribution

Lenny：产品正以惊人的速度发布。什么迹象能让你判断一家初创公司具备持久性，而不仅仅是一个包装层？Lenny: Products are launching at extraordinary speed. What tells you a startup can be durable rather than merely a wrapper?

Anish：Decagon 的 Jesse 有一句我很喜欢的话：护城河往往是被发现的，而不是被设计出来的。创始人可能会陷入这样的困境：试图制定一份能够经受 MBA 和风险投资人审视的商业计划，还要附带一套关于持久性的复杂理论。Decagon 开始交付产品，随着时间的推移发现了自己的护城河。Anish: Jesse from Decagon has a line I love: moats are most often discovered, not designed. Founders can become trapped trying to create a business plan that survives scrutiny from MBAs and VCs, complete with an elaborate theory of durability. Decagon started shipping and discovered its moat over time.

Cursor 提供了另一个例子。它曾因缺乏护城河而受到批评。它最初是一款依靠强劲产品驱动增长的高端产品，随后积累了推理轨迹、训练了自己的模型，并不断深化产品。持久性源于使用和执行。Cursor offers another example. It was criticized for lacking a moat. It began as a high-end product with strong product-led growth, then accumulated reasoning traces, trained its own models, and deepened the product. Durability emerged from use and execution.

我们还忘记了，经典的护城河很少建立在普通软件有多难构建之上。它们源自网络效应、规模优势、品牌、专有数据，或过去所谓的稀缺资源。这些护城河依然有效。我们需要有足够雄心的创始人，去构建多人协作产品、面向消费者的社交系统，以及会随着使用而显著改进的产品。We have also forgotten that classic moats were rarely based on how hard ordinary software was to build. They came from network effects, scale advantages, brand, proprietary data, or what used to be called a cornered resource. Those moats remain valid. We need founders ambitious enough to build multiplayer products, consumer social systems, and products that improve dramatically with use.

Lenny：创始人可以直接告诉投资者，护城河尚未被发现吗？Lenny: Can a founder simply tell an investor that the moat has not been discovered yet?

Anish：如果一款产品拥有强劲的发展势头、精湛的工艺和不断增长的用户参与度，即使其耐用性方面的论述尚未打磨完善，我们也可以对其进行投资。我过去曾担心有人会窃取我的创意。后来我明白，每个伟大的创意都建立在十几个微小而无形的创意之上。竞争对手可能会照搬显而易见的概念，却看不到让它真正奏效的细节。Anish: We can invest in a product without a polished durability story if it has strong momentum, craft, and growing engagement. I used to fear that someone would steal my idea. I learned that every big idea rests on a dozen small, invisible ideas. A competitor may copy the visible concept without seeing the details that make it work.

有些产品包含某种特别之处，使它们即使面临激烈竞争也能取得成功。格兰诺拉麦片曾因耐久性不足而受到批评，但凭借非凡的工艺，它最终赢得了人们的喜爱并占据主导地位。相比商业教科书上的说法，我更愿意倾听客户的声音。Some products contain something special that makes them successful despite intense competition. Granola was criticized for weak durability, yet it became beloved and dominant through exceptional craft. I listen to what customers say more than what a business textbook says.

Lenny：产品可以共享相似的底层模型，但主要在工具框架和用户体验上有所不同。这表明，用户体验本身就能创造优势，并争取时间去发现更深层、更持久的竞争力。Lenny: Products can share similar underlying models while differing mainly in the harness and user experience. That suggests UX itself can create an advantage and buy time to discover deeper durability.

Anish：是的。专业用户通常会为多种产品付费，因为每种产品都有不同的专长。模型和产品之间的差异仍然具有重要意义。Anish: Yes. Sophisticated users often pay for several products because each has a different specialty. Model and product differences remain meaningful.

Lenny：分发似乎越来越有价值，因为成千上万的事物不断发布并争夺注意力。现有的分发渠道似乎是一项巨大的优势。Lenny: Distribution feels increasingly valuable because thousands of things launch and compete for attention. Existing distribution seems like a huge advantage.

Anish：这很重要，但每个现有网络都经过训练，以防止其他人在其基础上构建网络。因此，网络效应在一定程度上又回归到了草根口碑。X、YouTube、Instagram 以及其他渠道上的自然提及，可能是现有最强大的外部网络效应。Anish: It is important, but every existing network has been trained to prevent others from building a network on top of it. As a result, network effects have partly returned to grassroots word of mouth. Organic mentions across X, YouTube, Instagram, and other channels may be the strongest external network effect available.

与移动互联网时代不同，当时 App Store 和围绕增长策略形成的小型产业提供了获客渠道；如今，创始人越来越需要通过口碑自行建立渠道。这是一个更纯粹、也更艰难的增长问题。Unlike the mobile era, where the App Store and a cottage industry of growth tactics provided channels, founders increasingly have to build their own channels from word of mouth. It is a purer and harder growth problem.

Lenny：Seth Godin 说，要打造非凡之物——真正值得人们津津乐道的东西。在纷繁嘈杂的环境中，人们会关注朋友正在使用和推荐的产品。Lenny: Seth Godin says to build something remarkable—literally worth remarking about. Amid the noise, people look to what their friends use and recommend.

Anish：令人抱有希望的一点是，许多公司并不是遇到了增长问题，而是产品问题。如今，你可以在功能或情感方向上打造一款极具野心的产品，并为其收取高昂的费用。想想看，如果这款产品每月收费一千美元或一万美元——如果它是一只软件版的铂金包——它需要具备哪些功能。然后朝着这样的愿景去构建它。Anish: The hopeful point is that many companies do not have a growth problem; they have a product problem. Today you can build a wildly ambitious product in a functional or emotional direction and charge a great deal for it. Ask what the product would need to do if it cost a thousand or ten thousand dollars a month—if it were a software Birkin bag. Then build toward that imagination.

初创公司可能比现有企业更容易取得成功，因为它们可以朝着让现有企业感到不安的方向发展，包括陪伴领域。买家也异常愿意尝试昂贵的产品并签署大额合同。这类似于每个人刚拿到 iPhone、都想要新应用的那段时期。这一窗口最终会缩小，但现在仍然敞开着。Startups may have an easier time than incumbents because they can build in directions that make incumbents uncomfortable, including companionship. Buyers are also unusually open to trying expensive products and signing large contracts. It resembles the period when everyone had just received an iPhone and wanted new apps. That window will eventually narrow, but it is open now.

创意产品也能以普通商业软件无法做到的方式触及社会敏感点。MSCHF 曾向一大群人发放借记卡，并定期宣布一家可以使用共享资金池消费的商店。所有人都争先恐后地去使用，最终只有最先行动的少数人成功了。这是一场有趣且迅速走红的社会实验。金钱本质上具有社会属性，但大多数金融产品都枯燥、私密，还令人尴尬。AI 也为雄心勃勃的产品创造了类似的机会，让技术成为一种创作媒介。Creative products can also touch social nerves in ways ordinary business software does not. MSCHF once issued a large group of people debit cards and periodically announced a store where a shared pool of money could be spent. Everyone rushed to use it; only the first few succeeded. It was a funny, viral social experiment. Money is inherently social, yet most financial products are dry, private, and embarrassing. AI creates a similar opening for ambitious products that use technology as a creative medium.

企业雄心的新上限A New Ceiling for Company Ambition

Lenny：你从成功的公司身上学到了哪些反直觉的经验？Lenny: What counterintuitive lessons have you learned from successful companies?

Anish：三年前，如果一家公司的计划看起来过于雄心勃勃，投资者往往会选择退出，因为这似乎太复杂了。1亿美元的种子轮融资听起来不可能被有效投入使用：对于要解决的问题来说钱太多，一个创始人难以管理这么多资金，而且招聘规模也太大，组织难以消化。Anish: Three years ago, if a company’s plan looked too ambitious, investors often disengaged because it seemed too complex. A hundred-million-dollar seed round sounded impossible to deploy productively: too much money for the problem, too much for one founder to manage, and too much hiring for an organization to absorb.

如今，我们越来越多地看到相反的问题。一个想法可能太小，因而无足轻重；而一笔非常大的种子轮投资也可能有合理的用途。我并不是建议每个人都去筹集一笔荒谬的资金，但抱负的上限已经改变。在 a16z，我们告诉创始人，我们希望帮助他们构建其愿景最强大的形态。Marc Andreessen 曾将公司早期的姿态描述为：要么登上月球，要么在地面上留下一个月球大小的陨坑——不存在畏缩不前的中间结果。Today, we increasingly see the opposite problem. An idea can be too small to matter, while a very large seed investment can have a plausible use. I am not recommending that everyone raise an absurd amount, but the ceiling on ambition has changed. At a16z, we tell founders that we want to help build the strongest form of their vision. Marc Andreessen described the firm’s own early posture as going to the moon or leaving a moon-sized crater in the ground—there was no timid middle outcome.

五年或十年前让人觉得难以克服的项目，如今可能只让人觉得困难。这会改变投资者选择哪些公司，以及创始人选择哪些投资者。Projects that felt insurmountable five or ten years ago can now feel merely difficult. That changes which companies investors select and which investors founders select.

另一个过时的假设是，面向消费者的产品必须免费。我越来越关注相反的方向：价格高得惊人的消费类软件。每一类可自由支配的消费者支出，都可以被创造出足够价值的软件所满足。因为价格是产品与市场契合度的证据，创始人应该思考，他们的产品如果定价为一千美元或一万美元，会是什么样子。这种思考会促使愿景朝着更具实质性、更大胆的方向发展。Another old assumption was that consumer products had to be free. I am increasingly interested in the opposite: extraordinarily expensive consumer software. Every category of discretionary consumer spending is available to software that creates enough value. Because price is evidence of product-market fit, founders should ask what the thousand- or ten-thousand-dollar version of their product would be. The exercise pushes the vision toward something meaningfully more ambitious.

Lenny：你与马克·安德森和本·霍洛维茨密切合作。从他们身上你学到了什么？Lenny: You work closely with Marc Andreessen and Ben Horowitz. What have you learned from them?

Anish：二者都体现了对科技行业、国家以及更广泛的西方生活方式和思维方式的守护。早期的风险投资领袖肩负着让这个行业变得比他们接手时更好的责任。Marc 和 Ben 的追求不止于成为出色的投资者。他们承担起艰巨而重要的工作，而这些工作并不只为公司或直接为公司带来利益，因为他们相信这很重要。Anish: Both embody stewardship for the technology industry, the country, and the broader Western way of living and thinking. Earlier venture leaders carried an obligation to leave the industry better than they found it. Marc and Ben aspire to more than being excellent investors. They undertake difficult, important work that does not exclusively or directly benefit the firm because they believe it matters.

他们还帮助塑造了创始人群体的抱负。五年或十年前，深科技还处于边缘地位，社会认可度很低。如今，它已经变得流行、主流且备受推崇。许多公司都做出了贡献，但马克尤其积极地表达了对符合国家利益的重要工作的支持。硅谷之所以发生变化，部分原因在于一些有影响力的人鼓励创始人将目标投向这一领域。They also helped shape the ambition of the founder community. Five or ten years ago, deep tech was fringe and low-status. It has become popular, mainstream, and prestigious. Many firms contributed, but Marc was especially vocal about supporting Important work in the national interest. Silicon Valley changed partly because influential people encouraged founders to aim there.

给产品构建者的建议Advice for Product Builders

Lenny：为了在未来取得成功，产品从业者应该改变他们的工作方式或思维方式中的哪些方面？Lenny: What should product people change about how they work or think to succeed in the future?

Anish：做更多东西。选择一个项目，即使你从不向任何人展示它，而且它看起来无关紧要。把它作为尝试新模型的载体。发布成果，讨论你学到了什么，并培养自己的直觉。亲自使用一周，略带挫败感却又深感充实，这所能教会你的，可能比抽象的讨论更多。Anish: Make more things. Choose a project, even one you never show anyone and that seems unimportant. Use it as a chassis for trying new models. Ship, discuss what you learned, and develop your own intuition. A slightly frustrating but deeply fulfilling week of real use can teach more than abstract discussion.

这就是我们许多人投身科技行业的原因：打造我们在脑海中构想出的产品。如今，我们拥有前所未有的机会去实现这一目标。硅谷正处于其最具正和博弈特征的状态。人们在彼此的工作基础上不断创新，而坦诚脆弱也会得到回报。向人们展示你的作品。This is why many of us entered technology: to build products we could see in our mind’s eye. Now we have an unprecedented chance to do it. Silicon Valley is positive-sum at its best. People build on one another’s work, and vulnerability is rewarded. Show people what you made.

Lenny：什么样的节奏是有用的？Lenny: What is a useful cadence?

Anish：每周发布一些东西。可以很小。母亲节那天，我用 Codex 根据短信和我们的照片库为妻子制作了一份幻灯片，还配上了音乐，以及我第一次约她时的旧消息。它最终变成了一段由二十张幻灯片讲述的爱情故事。这并不是一个重要的商业项目，我可能再也不会回头看它，但它很有意义，也让我学到了一些东西。Anish: Ship something once a week. It can be small. For Mother’s Day, I used Codex to make my wife a slide deck drawn from text messages and our photo gallery, with music and old messages from when I first asked her out. It became a twenty-slide story of our relationship. It was not an important commercial project, and I may never revisit it, but it was meaningful and taught me something.

Lenny：一位共同的朋友说，人们往往会在 AI 为他们带来某个快乐时刻后，改变自己对 AI 的看法。母亲节项目就是一个很好的例子。一个有用的提示可以是：如果它能成功实现，什么会让你感到快乐？Lenny: A mutual friend says people often change how they feel about AI after it creates one moment of joy for them. The Mother’s Day project is a perfect example. A useful prompt might be: what would bring you joy if it worked?

Anish：或者问问你能为别人做些什么。这是一个非常好的起点。Anish: Or ask what you can do for somebody else. That is an excellent place to begin.

书籍、产品、人生经验和音乐Books, Products, Life Lessons, and Music

Lenny：你最常推荐哪些书？Lenny: Which books do you recommend most often?

Anish：我最喜欢的是托马斯·索维尔的《征服与文化》。这本书探讨了征服如何塑造世界各地的文化，有时带来积极影响，有时则带来消极影响。在我看来，它以最有力的历史视角展现了文化如何成为推动结果的因素。我出生于加拿大，后来移居美国，亲身体验了截然不同的进取文化。Anish: My favorite is Thomas Sowell’s Conquests and Cultures. It examines how conquest shaped cultures around the world, sometimes positively and sometimes negatively. To me, it offers the strongest historical view of culture as a driver of outcomes. I was born in Canada and moved to the United States, and I experienced a very different culture of ambition.

汉密尔顿·赫尔默的《七种力量》是对商业护城河和复利优势的一次出色的思想提炼。我还推荐《规模报酬递增》一书，这是一部与布莱恩·阿瑟相关的内容密集型著作，解释了为什么软件等行业会产生异乎寻常的经济效应。它帮助我理解了科技行业及其文化为何会以现在的方式运作，包括为什么这个体系能够如此具有正和性。Hamilton Helmer’s 7 Powers is an excellent intellectual distillation of business moats and compounding advantage. I also recommend Increasing Returns to Scale, a dense text associated with Brian Arthur that explains why industries such as software produce outlier economic effects. It helped me understand why the technology industry and its culture work as they do, including why the system can be so positive-sum.

另外两本不可或缺的商业书籍是《高产出管理》，它至今仍然同样有用；以及本·霍洛维茨的《创业维艰》。本·霍洛维茨的书以真挚的情感，坦诚地描写了创始人所经历的焦虑、孤独与失败。读者会发现，他们并不孤单。Two other indispensable business books are High Output Management, which remains as useful as ever, and Ben Horowitz’s The Hard Thing About Hard Things. Ben’s book was emotionally honest about the anxiety, isolation, and failure founders experience. Readers discover they are not alone.

Lenny：你最近喜欢哪部电影或电视剧？Lenny: What recent movie or show did you enjoy?

阿尼什：我喜欢毫不掩饰地追求娱乐性的电视节目，包括《龙之家族》。我还在伦敦一家座无虚席的 IMAX 影院看了《奥德赛》，观众们一边喝着啤酒，一边尽情享受。电影本身不错，但最突出的还是这种共同观影的体验——大家独处，却又聚在一起。Anish: I enjoy unapologetically entertaining television, including House of the Dragon. I also saw The Odyssey in a packed London IMAX where people were drinking pints and having fun. The film was good, but the communal theater experience—being alone together—was what stood out.

Lenny：最近哪款 AI 产品给你带来了快乐？Lenny: Which recent AI product has given you joy?

Anish：我会选择 Grok 的智能体产品，因为它大胆地利用凭据并代表用户完成工作。它将经过深思熟虑的界面与强大的基础模型相结合，并承担了其他大型公司产品可能会避免的风险。它还表明，基础模型领域可能并非只有两强竞争。Anish: I would choose Grok’s agent product because it is ambitious about using credentials and completing work on a person’s behalf. It combines a thoughtful interface with a powerful foundation model and takes risks other large-company products may avoid. It also demonstrates that foundation models may not be a two-horse race.

Lenny：你有自己的人生格言吗？Lenny: Do you have a life motto?

Anish：不要通过痛苦的亲身经历去发现别人本可以直接告诉你的事情。我作为创始人总结出了这一原则，而它同样适用于为人父母。我太多次因为没有听取那些比自己先走几步的人的建议，而是通过本可避免的失败来学习；我的孩子们有时也有这个习惯。Anish: Do not discover through painful experience what someone could simply tell you. I developed that principle as a founder, and it also applies to parenting. I have too often learned through avoidable failure instead of listening to someone a few steps ahead, and my children sometimes share that habit.

对孩子来说，字面上的说法是“不要碰热炉子”。在我的第一家初创公司里，有经验的人告诉我，不要同时打造一款产品和一个平台。我们试图既成为一个面向移动游戏的社交平台，又成为一家游戏工作室。经营一家工作室本已十分困难，更不用说同时经营这两项业务了。一位明智的人告诉我应该二选一，但我们花了好几年才接受他们是对的。For a child, the literal version is “do not touch the hot stove.” In my first startup, experienced people told me not to build a product and a platform simultaneously. We tried to be both a social platform for mobile games and a game studio. Running a studio was already difficult, let alone running both businesses. Someone wise told me to choose one, but it took us years to accept that they were right.

Lenny：Ben Horowitz 让我问问你的 DJ 经历。对于对 DJ 感兴趣的人，你有什么建议？Lenny: Ben Horowitz told me to ask about your DJing. What advice do you have for someone interested in it?

Anish：我从 1995 年开始担任 DJ。这是一种绝妙的音乐表达方式，即使没有接受过古典音乐训练也能做到。你需要挑选唱片并将它们组合起来，这既需要品味，也需要技术。如今，模型让人们能够更进一步，直接创作音乐本身。你可以从一个想法开始，让模型帮助你将其实现为最有力的形式。Anish: I have been DJing since 1995. It is a wonderful way to express yourself musically without being a classically trained musician. You select records and combine them, which requires taste and technical skill. Models now let people go further and make the music itself. You can begin with an idea and have a model help realize its strongest form.

音乐是一种直观而令人满足的体验创造方式。它的媒介史不断扩大着参与范围。最初，只有亲临演奏者身边，你才能听到音乐。录制音乐将演奏与聆听分离开来。磁带则带来了又一次重大转变，因为听众可以通过制作混音带来编排自己的专辑。进入21世纪第一个十年后，其中一部分参与性让位于类似广播的消费方式。随着人们开始再次借助 AI 创作音乐，我预计音乐产业将变得前所未有地庞大。Music is a visceral, satisfying way to create experiences. Its media history keeps widening participation. At first, you could hear music only in the presence of the person playing it. Recorded music separated performance from listening. The cassette was another major shift because listeners could compose their own albums as mixtapes. In the 2000s, some of that participatory quality gave way to broadcast-like consumption. As people begin making music again with AI, I expect the music industry to become larger than ever.

Lenny：听众怎样才能对你有所帮助？Lenny: How can listeners be useful to you?

Anish：让我看看你正在构建什么。不要灰心。做点东西，@我，让我看看。如果你想获得更多这类想法，可以在 X 上关注我，也可以探索 Lenny 社区中的构建者和思想家。那里正在开展大量引人入胜的工作。核心信息很简单：使用这些工具，构建一些东西，并分享出来。Anish: Show me what you are building. Do not be despondent. Make something, tag me, and let me see it. Follow me on X if you want more of these ideas, and explore the builders and thinkers across Lenny’s community. There is a great deal of compelling work happening. The essential message is simple: use the tools, build something, and share it.