蒸汽、钢铁与无穷智能

Steam, Steel, and Infinite Minds

作者： Ivan Zhao， Co-founder & CEO, Notion 

每个时代都由其奇迹材料所塑造 Every era is shaped its miracle material。钢铁锻造了镀金时代 (Gilded Age)。半导体开启了 (switched on) 数字时代。如今，人工智能以无穷智能的姿态降临 (arrived as infinite minds)。历史告诉我们，掌握核心材料者，方能定义时代。

左：少年安德鲁·卡内基和他的弟弟。右：镀金时代的匹兹堡钢铁厂。

在19世纪50年代，少年安德鲁·卡内基还是个在匹兹堡泥泞街道上奔跑的电报童。当时，十之六的美国人仍是农民。不出两代人的时间，卡内基及其同辈便锻造了现代世界。马匹让位于铁路，烛光让位于电力，生铁让位于钢铁。

自那时起，工作重心从工厂转向办公室。如今，我在旧金山经营一家软件公司，为数百万知识工作者打造工具。在这个行业小镇，人人都在谈论通用人工智能（AGI），但全球二十亿伏案工作者中的大多数，尚未感受到它的存在。不久之后，知识工作将变成何等模样？当组织结构图吸纳了永不休眠的智能，又将发生什么？

早期电影常常看起来像舞台剧，单一镜头聚焦于舞台。

未来之所以难以预测，因为它总以过去的伪装示人。早期的电话交谈简短如电报。早期的电影看起来像被录制下来的戏剧。（这便是马歇尔·麦克卢汉所说的"通过后视镜驶向未来driving to the future via the rearview window"。）

当今最流行的人工智能形态，看起来像过去的谷歌搜索。引用马歇尔·麦克卢汉的话："我们总是通过后视镜驶向未来。" 

今天，我们看到这种现象体现为模仿谷歌搜索框的人工智能聊天机器人。我们正深陷于每次技术变革都会出现的、令人不适的过渡阶段。

对于接下来会发生什么，我并无全部答案。但我喜欢用一些历史隐喻来思考人工智能如何在不同尺度上运作——从个人到组织，再到整个经济。

个人：从自行车到汽车

最初的迹象，可以在知识工作的”高阶祭司" (high priests of knowledge work)——程序员身上窥见。

我的联合创始人西蒙曾是我们所说的”10倍效率程序员"，但他如今已很少亲手写代码。走过他的办公桌，你会看到他同时指挥三四个AI编程智能体工作。它们不仅打字更快，还能思考，这共同使他成为效率提升30至40倍的工程师。他会在午餐前或睡前将任务排队，让智能体在他离开时继续工作。他已化身为无穷智能的管理者 (He’s become a manager of infinite minds)。

20世纪70年代《科学美国人》一项关于运动效率的研究，启发了史蒂夫·乔布斯著名的”头脑的自行车"之比喻 (bicyle for mind)。只不过，自那以后数十年，我们一直在信息高速公路上”蹬自行车”。

20世纪80年代，史蒂夫·乔布斯将个人电脑称为”头脑的自行车"。十年后，我们铺设了互联网这条"信息高速公路" (information superhighway)。但时至今日，大部分知识工作仍依赖人力驱动。这就好比我们一直在高速公路上蹬着自行车 (pedaling bicycles on the autobahn)。

借助AI智能体，像西蒙这样的人已经从骑自行车升级为开汽车。

其他类型的知识工作者何时能开上"汽车"？必须解决两个问题。

与编码智能体相比，为何AI更难协助一般知识工作？因为知识工作更加碎片化且难以验证。

首先是语境碎片化。对于编码工作，工具和语境往往集中于一处：集成开发环境、代码仓库、终端。但一般的知识工作则分散在数十种工具中。想象一下，一个试图起草产品简报的AI智能体：它需要从Slack聊天记录、战略文档、仪表盘中的上季度指标，以及仅存在于某人脑海中的组织记忆里提取信息。如今，人类是粘合剂，通过复制粘贴和在浏览器标签页间切换，将这一切缝合起来。在语境得到整合之前，智能体将只能局限于狭窄的用例。

第二个缺失的要素是可验证性。代码拥有一个神奇的特性：你可以通过测试和错误来验证它。模型制造者利用这一点来训练AI，使其更擅长编码（例如强化学习）。但是，你如何验证一个项目管理得是否出色，或者一份战略备忘录是否优秀？我们尚未找到提升模型在一般知识工作中表现的方法。因此，人类仍需在循环中监督、引导并示范什么是"优秀"。

1865年的《红旗法》要求车辆在街上行驶时，需有一名持旗者在车前步行（该法于1896年废止）。这是一个不受欢迎的"人在回路"的例子。

今年的编程智能体告诉我们，"人在回路"并非总是可取的。这就像让某人亲自检查生产线上的每个螺栓，或者走在汽车前面清空道路（参见：1865年《红旗法》）。我们希望人类能从有影响力的制高点监督循环，而不是置身其中。一旦语境得到整合，工作变得可验证，数十亿工作者将从"蹬自行车"升级到"开汽车"，再从"开汽车"迈向"自动驾驶"。

组织：钢铁与蒸汽

公司是近现代的发明。它们随着规模扩大而效能递减，并最终触及极限。

1855年纽约和伊利铁路的组织结构图。现代公司及组织结构图是随着铁路公司演进而来的，铁路公司是首个需要协调数千人跨越远距离运作的企业。

几百年前，大多数公司只是十几人的作坊。如今，我们有了员工数十万的跨国公司。沟通基础设施（通过会议和信息连接起来的人脑）在指数级负荷下不堪重负。我们试图通过层级、流程和文档来解决这个问题。但我们一直在用人力尺度的工具来解决工业化尺度的问题，就像用木头建造摩天大楼。

两个历史隐喻展示了，当拥有新的奇迹材料时，未来的组织将如何呈现不同面貌。

钢铁的奇迹：伍尔沃斯大厦于1913年在纽约建成时，曾是世界最高建筑。

第一个是钢铁。在钢铁出现之前，19世纪的建筑高度被限制在六七层左右。铁虽然坚固但易碎且沉重；增加楼层，结构就会在自身重压下坍塌。钢铁改变了一切。它坚固却又具延展性。框架可以更轻，墙壁可以更薄，突然间建筑可以拔地而起数十层。新型建筑成为可能。

AI就是组织的"钢铁"。它有潜力在工作流程中维持语境，并在需要时呈现决策，且没有噪音干扰。人类沟通不再必须是承重墙。每周两小时的对齐会议可以变成五分钟的异步审阅。需要三层审批的高管决策可能很快在几分钟内完成。公司可以实现真正意义上的规模化扩展，而无需忍受我们曾认为不可避免的效能衰减。

一个依靠水轮提供动力的磨坊。水力强大但不可靠，且将磨坊限制在少数地点并受季节影响。

第二个故事关乎蒸汽机。在工业革命初期，早期的纺织工厂依河溪而建，依靠水轮提供动力。当蒸汽机出现时，工厂主最初只是用水轮换成了蒸汽机，其他一切照旧。生产力提升有限。

真正的突破发生在工厂主意识到他们可以完全摆脱对水的依赖之时。他们建造了更靠近工人、港口和原材料的大型工厂。并围绕蒸汽机重新设计了工厂布局（后来，电力出现时，工厂主进一步从中心动力轴分散开来，在工厂各处为不同机器安置了更小的发动机）。生产力随之爆炸式增长，第二次工业革命才真正起飞。

托马斯·阿洛姆1835年的这幅版画描绘了英国兰开夏郡的一家纺织厂。它由蒸汽机提供动力。我们仍处于"替换水轮"的阶段。

AI聊天机器人被生硬地附加到现有工具上。我们尚未重新构想，当旧有的约束消失，你的公司可以依靠在你睡觉时仍在工作的无穷智能来运行时，组织会是什么模样。

在我的公司Notion，我们一直在进行实验。除了我们的1000名员工，如今还有700多个智能体处理重复性工作。它们做会议记录、回答问题以整合团队知识。它们处理IT请求、记录客户反馈。它们帮助新员工了解福利待遇。它们编写每周状态报告，让人们不必再复制粘贴。而这仅仅是婴儿学步。真正的收益只受限于我们的想象力和惯性。

经济：从佛罗伦萨到超级都市

钢铁和蒸汽不仅改变了建筑和工厂。它们改变了城市。

佛罗伦萨与东京直到几百年前，城市还是以人的尺度建造的。你可以在四十分钟内步行穿越佛罗伦萨。生活的节奏是由一个人能走多远、声音能传多响来设定的。

接着，钢架结构使摩天大楼成为可能。蒸汽机驱动的铁路将市中心与腹地连接起来。电梯、地铁、高速公路随之而来。城市的规模和密度急剧膨胀。东京。重庆。达拉斯。

这些不仅仅是放大版的佛罗伦萨。它们是不同的生活方式。超级都市令人迷失方向、充满匿名性、更难以导航 （Megacities are disorienting, anonymous, harder to navigat）。这种"难以解读性 (illegibility)"是规模化的代价。但它们也提供了更多的机会、更多的自由。比一个文艺复兴时期人力尺度的城市所能容纳的，更多的人以更多的组合方式做着更多的事情。

我认为知识经济即将经历同样的转型。

如今，知识工作几乎占美国GDP的一半。其中大部分仍以人力尺度运作：数十人的团队、以会议和电子邮件为节奏的工作流程、超过数百人便效能下降的组织。我们一直在用石头和木头建造"佛罗伦萨"。

当AI智能体大规模上线时，我们将建造"东京"。那将是横跨数千智能体与人类的组织。是跨时区、无需等待任何人醒来、持续运行的工作流。是融入恰到好处的人类监督而合成的决策。

感觉将会不同。更快，更具杠杆效应，但最初也会更令人迷失方向。每周例会、季度规划周期、年度评估的节奏可能不再适用。新的节奏将会出现。我们会失去一些可解读性。但我们获得了规模和速度。

每一种奇迹材料都要求人们停止通过后视镜看世界，并开始想象新的世界。卡内基凝视钢铁，看见了城市的天际线。兰开夏郡的工厂主凝视蒸汽机，看见了摆脱河流束缚的工厂车间。

我们仍处于人工智能的"水轮阶段"，将聊天机器人硬塞进为人类设计的工作流程中。我们需要停止仅仅要求AI成为我们的副驾驶。我们需要去想象，当人类的组织用"钢铁"加固，当繁琐工作被委托给永不休眠的智能时，知识工作将会是何等景象。

钢铁。蒸汽。无穷智能。下一道天际线就在那里，等待我们去建造。

Steam, Steel, and Infinite Minds

Every era is shaped by its miracle material. Steel forged the Gilded Age. Semiconductors switched on the Digital Age. Now AI has arrived as infinite minds. If history teaches us anything, those who master the material define the era.

In the 1850s, Andrew Carnegie ran through muddy Pittsburgh streets as a telegraph boy. Six in ten Americans were farmers. Within two generations, Carnegie and his peers forged the modern world. Horses gave way to railroads, candlelight to electricity, iron to steel.

Since then, work shifted from factories to offices. Today I run a software company in San Francisco, building tools for millions of knowledge workers. In this industry town, everyone is talking about AGI, but most of the two billion desk workers have yet to feel it. What will knowledge work look like soon? What happens when the org chart absorbs minds that never sleep?

This future is often difficult to predict because it always disguises itself as the past. Early phone calls were concise like telegrams. Early movies looked like filmed plays. (This is what Marshall McLuhan called "driving to the future via the rearview window.")

Today, we see this as AI chatbots which mimic Google search boxes. We're now deep in that uncomfortable transition phase which happens with every new technology shift.

I don't have all the answers on what comes next. But I like to play with a few historical metaphors to think about how AI can work at different scales, from individuals to organizations to whole economies.

Individuals: from bicycles to cars

The first glimpses can be found with the high priests of knowledge work: programmers.

My co-founder Simon was what we call a 10× programmer, but he rarely writes code these days. Walk by his desk and you'll see him orchestrating three or four AI coding agents at once, and they don't just type faster, they think, which together makes him a 30-40× engineer. He queues tasks before lunch or bed, letting them work while he's away. He's become a manager of infinite minds.

In the 1980s, Steve Jobs called personal computers "bicycles for the mind." A decade later, we paved the "information superhighway" that is the internet. But today, most knowledge work is still human-powered. It's like we've been pedaling bicycles on the autobahn.

With AI agents, someone like Simon has graduated from riding a bicycle to driving a car.

When will other types of knowledge workers get cars? Two problems must be solved.

First, context fragmentation.For coding, tools and context tend to live in one place: the IDE, the repo, the terminal. But general knowledge work is scattered across dozens of tools. Imagine an AI agent trying to draft a product brief: it needs to pull from Slack threads, a strategy doc, last quarter's metrics in a dashboard, and institutional memory that lives only in someone's head. Today, humans are the glue, stitching all that together with copy-paste and switching between browser tabs. Until that context is consolidated, agents will stay stuck in narrow use-cases.

The second missing ingredient is verifiability.Code has a magical property: you can verify it with tests and errors. Model makers use this to train AI to get better at coding (e.g. reinforcement learning). But how do you verify if a project is managed well, or if a strategy memo is any good? We haven't yet found ways to improve models for general knowledge work. So humans still need to be in the loop to supervise, guide, and show what good looks like.

Programming agents this year taught us that having a "human-in-the-loop" isn't always desirable. It's like having someone personally inspect every bolt on a factory line, or walk in front of a car to clear the road (see: the Red Flag Act of 1865). We want humans to supervise the loops from a leveraged point, not be in them. Once context is consolidated and work is verifiable, billions of workers will go from pedaling to driving, and then from driving to self-driving.

Organizations: steel and steam

Companies are a recent invention. They degrade as they scale and reach their limit.

A few hundred years ago, most companies were workshops of a dozen people. Now we have multinationals with hundreds of thousands. The communication infrastructure (human brains connected by meetings and messages) buckles under exponential load. We try to solve this with hierarchy, process, and documentation. But we've been solving an industrial-scale problem with human-scale tools, like building a skyscraper with wood.

Two historical metaphors show how future organizations can look differently with new miracle materials.

The first is steel. Before steel, buildings in the 19th century had a limit of six or seven floors. Iron was strong but brittle and heavy; add more floors, and the structure collapsed under its own weight. Steel changed everything. It's strong yet malleable. Frames could be lighter, walls thinner, and suddenly buildings could rise dozens of stories. New kinds of buildings became possible.

AI is steel for organizations. It has the potential to maintain context across workflows and surface decisions when needed without the noise. Human communication no longer has to be the load-bearing wall. The weekly two-hour alignment meeting becomes a five-minute async review. The executive decision that required three levels of approval might soon happen in minutes. Companies can scale, truly scale, without the degradation we've accepted as inevitable.

The second story is about the steam engine. At the beginning of the Industrial Revolution, early textile factories sat next to rivers and streams and were powered by waterwheels. When the steam engine arrived, factory owners initially swapped waterwheels for steam engines and kept everything else the same. Productivity gains were modest.

The real breakthrough came when factory owners realized they could decouple from water entirely. They built larger mills closer to workers, ports, and raw materials. And they redesigned their factories around steam engines (Later, when electricity came online, owners further decentralized away from a central power shaft and placed smaller engines around the factory for different machines.) Productivity exploded, and the Second Industrial Revolution really took off.

We're still in the "swap out the waterwheel" phase.AI chatbots bolted onto existing tools. We haven't reimagined what organizations look like when the old constraints dissolve and your company can run on infinite minds that work while you sleep.

At my company Notion, we have been experimenting. Alongside our 1,000 employees, more than 700 agents now handle repetitive work. They take meeting notes and answer questions to synthesize tribal knowledge. They field IT requests and log customer feedback. They help new hires onboard with employee benefits. They write weekly status reports so people don't have to copy-paste. And this is just baby steps. The real gains are limited only by our imagination and inertia.

Economies: from Florence to megacities

Steel and steam didn't just change buildings and factories. They changed cities.

Until a few hundred years ago, cities were human-scaled. You could walk across Florence in forty minutes. The rhythm of life was set by how far a person could walk, how loud a voice could carry.

Then steel frames made skyscrapers possible. Steam engines powered railways that connected city centers to hinterlands. Elevators, subways, highways followed. Cities exploded in scale and density. Tokyo. Chongqing. Dallas.

These aren't just bigger versions of Florence. They're different ways of living. Megacities are disorienting, anonymous, harder to navigate. That illegibility is the price of scale. But they also offer more opportunity, more freedom. More people doing more things in more combinations than a human-scaled Renaissance city could support.

I think the knowledge economy is about to undergo the same transformation.

Today, knowledge work represents nearly half of America's GDP. Most of it still operates at human scale: teams of dozens, workflows paced by meetings and email, organizations that buckle past a few hundred people. We've built Florences with stone and wood.

When AI agents come online at scale, we'll be building Tokyos. Organizations that span thousands of agents and humans. Workflows that run continuously, across time zones, without waiting for someone to wake up. Decisions synthesized with just the right amount of human in the loop.

It will feel different. Faster, more leveraged, but also more disorienting at first. The rhythms of the weekly meeting, the quarterly planning cycle, and the annual review may stop making sense. New rhythms emerge. We lose some legibility. We gain scale and speed.

Every miracle material required people to stop seeing the world via the rearview mirror and start imagining the new one. Carnegie looked at steel and saw city skylines. Lancashire mill owners looked at steam engines and saw factory floors free from rivers.

We are still in the waterwheel phase of AI, bolting chatbots onto workflows designed for humans. We need to stop asking AI to be merely our copilots. We need to imagine what knowledge work could look like when human organizations are reinforced with steel, when busywork is delegated to minds that never sleep.

Steel. Steam. Infinite minds. 

The next skyline is there, waiting for us to build it.