Left: Young Andrew Carnegie and his brother. Right: Pittsburgh steel mills during the Gilded Age.
In the 1850s, Andrew Carnegie was a telegraph operator running through the muddy streets of Pittsburgh, when six out of ten Americans were farmers. Just two generations later, Carnegie and his colleagues forged the modern world, horses gave way to railroads, candlelight to electric lights, and iron to steel.
Since then, the work has shifted from factories to offices. Today, I run a software company in San Francisco, building tools for hundreds of thousands of knowledge workers. In this tech town, everyone talks about Artificial General Intelligence (AGI), but most of the two billion office workers haven't yet felt its presence. What will knowledge work look like in the near future? What will happen when relentless intelligence is integrated into organizational structures?
Early films were often like stage plays, with a camera filming the stage.
The future is often unpredictable because it always masquerades as the past. Early conversations were as brief as telegrams, and early films were like recorded stage plays. As Marshall McLuhan said, "We are always looking into the future through the rearview mirror."
The most prevalent form of artificial intelligence today still resembles Google Search of the past. To quote McLuhan, "We are always looking into the future through the rearview mirror." Today, we see AI chatbots mimicking the Google search box. We are caught in that uncomfortable transition period that occurs with every technological revolution.
I don't have all the answers as to what the future holds. But I like to use a few historical metaphors to think about how artificial intelligence can play a role at different levels of individuals, organizations, and even the economy as a whole.
Personal: From Bicycles to Cars
The initial signs can be seen in programmers, who are "advanced practitioners" of knowledge work.
My co-founder, Simon, was once a "ten-times programmer," but lately he rarely writes code by hand. Walking past his workstation, you'll see him simultaneously managing three or four AI programming assistants. These assistants not only type faster but also think, making him an engineer 30 to 40 times more efficient. He often schedules tasks before lunch or bedtime, allowing the AI to continue working while he's away. He has transformed into a manager of boundless intelligence.
A 1970s Scientific American study on exercise efficiency inspired Steve Jobs to coin the famous metaphor of the "bicycle of thought." However, for decades since then, we've been "riding bicycles" on the information superhighway.
In the 1980s, Steve Jobs called the personal computer a "bicycle of thought." A decade later, we paved the "information superhighway" called the internet. But today, most knowledge work still relies on human labor. It's like we've been riding bicycles on a highway all this time.
With AI assistants, people like Simon have upgraded from riding bicycles to driving cars.
When will other types of knowledge workers be able to "drive a car"? There are two questions that need to be addressed.
Why is AI-assisted knowledge work more difficult than programming assistants? Because knowledge work is more fragmented and harder to verify.
First, there's the issue of fragmented context. In programming, tools and context are often centralized in one place: an integrated development environment (IDE), a code repository, or a terminal. But general knowledge work is scattered across dozens of tools. Imagine an AI assistant trying to draft a product description: it needs to retrieve information from Slack threads, strategy documents, last quarter's data from dashboards, and organizational memories that only exist in a person's brain. Currently, humans act as the glue, piecing everything together by copying and pasting and switching between browser tabs. As long as the context isn't integrated, AI assistants will be limited to narrow uses.
The second missing element is verifiability. Code has a magical property: you can verify it through testing and error reporting. Model developers leverage this to train AI to program better using methods like reinforcement learning. But how do you verify whether a project is well-managed or whether a strategic memo is excellent? We haven't yet found a way to improve general knowledge work models. Therefore, humans still need to remain in the loop to supervise, guide, and demonstrate what is "good."
The Red Flag Act of 1865 required that a flag bearer walk in front of a car when it was driving on the street (this act was repealed in 1896).
This year's programming assistant practices have taught us that "human in the loop" isn't always ideal. It's like having someone inspect bolts one by one on a production line, or walking in front of cars to clear the way (see the Red Flag Act of 1865). We should have people overseeing the loop from a higher perspective, rather than being part of it. Once the context is integrated and the work becomes verifiable, billions of workers will move from "cycling" to "driving," and from "driving" to "autonomous driving."
Organization: Steel and Steam
Companies are a modern invention; their efficiency diminishes as they grow in size, eventually reaching their limits.
Organizational chart of the New York and Erie Railroad Company in 1855. Modern corporations and their organizational structures evolved alongside railroads, which were among the earliest enterprises requiring the coordination of thousands of people over long distances.
Centuries ago, most companies were workshops with just a dozen or so people. Today, we have multinational corporations with hundreds of thousands of employees. Communication infrastructure, relying on meetings and information-connected human brains, is overwhelmed by this exponentially increasing load. We try to solve it with hierarchies, processes, and documents, but this is akin to building a skyscraper with wood—using human-scale tools to solve industrial-scale problems.
Two historical metaphors illustrate how organizations might look differently in the future when they possess new technological resources.
A miracle of steel: The Woolworth Building in New York, completed in 1913, was once the tallest building in the world.
The first was steel. Before steel, 19th-century building heights were limited to six or seven stories. Iron, though strong, was brittle and heavy; adding more floors would cause the structure to collapse under its own weight. Steel changed everything. It was strong yet flexible, allowing for lighter frames, thinner walls, and buildings that could suddenly rise to dozens of stories, making new types of architecture possible.
AI is the "steel" of an organization. It promises to maintain contextual coherence across workflows, delivering decisions when needed without noise interference. Human communication will no longer need to act as a load-bearing wall. Two-hour weekly alignment meetings may become five-minute asynchronous reviews; senior executive decisions requiring three levels of approval may be completed in minutes. Companies can truly scale without the efficiency degradation we once considered inevitable.
A mill powered by a waterwheel. Water power is powerful but unstable and limited by location and season.
The second story is about the steam engine. In the early days of the Industrial Revolution, early textile factories were built along rivers and powered by waterwheels. After the steam engine appeared, factory owners initially only replaced waterwheels with steam engines, keeping everything else the same, resulting in limited increases in productivity.
The real breakthrough came when factory owners realized they could completely break free from the constraints of water. They built larger factories closer to workers, ports, and raw materials, and redesigned the layout around the steam engine (later, with the widespread adoption of electricity, factory owners further eliminated the central power shaft, distributing smaller engines throughout the factory to power different machines). Productivity exploded, and the Second Industrial Revolution truly began.
Thomas Allom’s 1835 engraving depicts a steam-powered textile mill in Lancashire, England.
We are still in the "replacing the waterwheel" phase. By cramming AI chatbots into workflows designed for humans, we haven't yet reimagined what organizations will look like when old constraints disappear and companies can rely on unlimited intelligence that works even while you sleep.
At my company, Notion, we've been experimenting. In addition to our 1,000 employees, we now have over 700 AI assistants handling repetitive tasks: recording meetings, answering questions to consolidate team knowledge, processing IT requests, recording customer feedback, helping new employees familiarize themselves with benefits, and writing weekly status reports to avoid manual copying and pasting… This is just the beginning. The real potential is limited only by our imagination and inertia.
Economies: From Florence to Megacities
Steel and steam have transformed not only buildings and factories, but also cities.
Until a few hundred years ago, cities were still measured on a human scale. You could walk across Florence in forty minutes, and the pace of life was determined by the distance people could walk and the range of sound.
Subsequently, steel frame structures made skyscrapers possible; steam-powered railways connected city centers with the hinterland; elevators, subways, and highways followed. The size and density of cities expanded dramatically—Tokyo, Chongqing, Dallas.
These are not merely enlarged versions of Florence; they represent entirely new lifestyles. Megacities can be disorienting, anonymous, and difficult to navigate. This "indiscernibility" is the price of scale. But they also offer greater opportunities, greater freedom, and support for more people to engage in more activities in more diverse combinations—something that Renaissance cities on a human scale could never achieve.
I believe the knowledge economy is about to undergo the same transformation.
Today, knowledge work accounts for nearly half of the US GDP, but its operation is still largely at the human scale: teams of dozens of people, workflows that rely on meetings and emails, and organizations that can hardly survive if they have more than a hundred people... We have been building "Florence" with stone and wood.
When AI assistants are deployed on a large scale, we will build "Tokyo," an organization composed of thousands of AI and humans; workflows that run continuously across time zones without waiting for someone to wake up; and decisions synthesized with just the right amount of human involvement.
It will be a different experience: faster, with greater leverage, but initially more dizzying. The rhythm of weekly meetings, quarterly planning, and annual reviews may no longer be suitable; a new rhythm will emerge. We will lose some clarity, but gain scale and speed.
Beyond the waterwheel
Every technological material demands that people stop looking at the world through the rearview mirror and start imagining a new world. Carnegie gazed at steel and saw the city skyline; the Lancashire factory owner looked at the steam engine and saw the factory workshop far from the river.
We are still in the "waterwheel stage" of AI, forcing chatbots into workflows designed for humans. We should not be content with AI acting as a co-pilot, but should imagine what knowledge work will look like when human organizations are reinforced with steel, and trivial tasks are delegated to tireless intelligence.
Steel, steam, and boundless intelligence. The next skyline lies ahead, waiting for us to build it ourselves.
