Ivan Zhao's Vision: Why AI Is the Steel of Knowledge Work

Notion CEO Ivan Zhao just published one of the most compelling frameworks for understanding AI's impact on work. His thesis: every era is defined by its miracle material — steel forged the Gilded Age, semiconductors powered the Digital Age, and now AI arrives as "infinite minds." The essay moves through three scales — individuals, organizations, and economies — using historical metaphors that cut through the hype cycle. The core argument is practical: two specific bottlenecks (context fragmentation and verifiability) must be solved before AI transforms the two billion desk workers who haven't felt it yet.

What Happened

Zhao published "Steam, Steel, and Infinite Minds," a long-form essay drawing parallels between industrial revolutions and AI adoption. The piece opens with Andrew Carnegie running through 1850s Pittsburgh as a telegraph boy, then traces how miracle materials — steel, semiconductors — reshaped entire economies within two generations.

The most revealing anecdote comes from inside Notion itself. Zhao describes his co-founder Simon Last, formerly a "10x programmer," who now rarely writes code directly. Instead, Simon orchestrates three or four AI coding agents simultaneously, effectively becoming a "30-40x engineer" who queues tasks before lunch and lets agents work while he's away. He's become, in Zhao's framing, "a manager of infinite minds."

Zhao then riffs on Steve Jobs's famous "bicycle for the mind" metaphor. His update: we've been pedaling bicycles on the autobahn for decades. AI agents represent the upgrade from bicycle to car — but only for programmers so far. The essay asks when the other two billion knowledge workers get their cars, and identifies exactly what's blocking them.

The organizational section draws on architectural history — before steel, buildings couldn't exceed five stories because load-bearing masonry hit physical limits. Steel frames unlocked skyscrapers. Zhao argues that companies face an analogous constraint: human communication infrastructure (meetings, messages, documents) buckles under exponential organizational load. AI could be the structural steel that lets organizations scale beyond current limits.

Why It Matters

Zhao's framework matters because it names the two specific problems that separate "AI works for programmers" from "AI works for everyone."

Context fragmentation is the first bottleneck. Coding agents thrive because their context lives in one place: the IDE, the repo, the terminal. But a product manager drafting a brief needs Slack threads, strategy docs, quarterly metrics from a dashboard, and institutional knowledge that exists only in someone's head. Humans currently serve as the glue — copy-pasting between browser tabs. Until that context is consolidated into something an AI agent can access, agents remain stuck in narrow use cases.

Verifiability is the second. Code has a "magical property" — you can verify it with tests and compiler errors. This is what enables reinforcement learning to improve coding models rapidly. But how do you verify if a project is managed well or a strategy memo is sound? Without clear signals of quality, models can't improve at general knowledge work through automated feedback loops.

This framing has competitive implications for every productivity tool company. Whoever solves context consolidation — pulling scattered knowledge into a unified layer — creates the platform that AI agents can actually operate on. Notion, with its all-in-one workspace model, is clearly positioning itself as that consolidation layer. Competitors like Confluence, Coda, and the broader Microsoft 365 ecosystem are racing toward the same goal.

Technical Deep-Dive

The essay's most technically interesting insight is the distinction between "human-in-the-loop" and "human-over-the-loop."

Zhao references the Red Flag Act of 1865, which required a person to walk ahead of early automobiles waving a flag. He argues that many current AI workflows replicate this pattern — humans personally inspecting every output, approving every action. This defeats the purpose of automation.

The alternative is humans supervising from a "leveraged point." In Simon's workflow, this means reviewing agent outputs in batches rather than watching each keystroke. The human sets direction, defines quality standards, and handles exceptions — but doesn't bottleneck the execution loop.

This maps to a progression Zhao outlines:

1. Bicycle — Human does the work, computer assists (current state for most knowledge workers)
2. Car — Human drives, AI agents execute (where programmers are today)
3. Self-driving — AI executes autonomously, human supervises (where programmers are heading)

The organizational parallel uses steel-frame architecture as a metaphor. Pre-steel buildings used load-bearing masonry walls — every wall carried weight, limiting height to roughly five stories. Steel frames decoupled structure from skin, enabling skyscrapers. Similarly, current organizations use human communication as load-bearing infrastructure. AI could decouple execution capacity from headcount, allowing smaller teams to operate at scales that previously required thousands.

The steam engine metaphor addresses a different dimension: energy. James Watt's engine didn't just replace muscle — it created entirely new industries (railroads, factories) that were inconceivable before cheap mechanical power. Zhao suggests AI won't just automate existing knowledge work but will enable categories of organizational activity that are currently impossible because they'd require too many human-hours.

What You Should Do

1. Audit your context fragmentation. Map where your team's knowledge actually lives — Slack, docs, dashboards, email, people's heads. The more scattered it is, the harder it is for AI agents to help. Consolidation isn't glamorous work, but it's the prerequisite.

2. Build verifiability into knowledge work. Create rubrics, checklists, and evaluation criteria for non-code outputs. If you can't define what "good" looks like for a task, neither can an AI — and you can't improve the AI's performance on it.

3. Move from in-the-loop to over-the-loop. Start with low-stakes tasks: let agents draft, then review in batches. Resist the urge to watch every step. The goal is supervision, not surveillance.

4. Watch Notion's product roadmap closely. Zhao's essay reads like a product strategy document. Expect Notion to ship deeper agent integration and cross-tool context consolidation in 2026.

Related: Today's newsletter covers more AI industry analysis. See also: What Are AI Agents for background on the agent paradigm Zhao describes.

---

Found this useful? Subscribe to AI News for daily AI briefings.