Zhao’s structural metaphor: before steel, 19th-century buildings hit a ceiling of “six or seven floors” because iron was “strong but brittle and heavy” and collapsed under its own weight. Steel — “strong yet malleable” — let frames go lighter and buildings rise dozens of stories. He argues “AI is steel for organizations”: it “maintain[s] context across workflows and surface[s] decisions when needed without the noise,” so “human communication no longer has to be the load-bearing wall.” Concretely, “the weekly two-hour alignment meeting becomes a five-minute async review,” and orgs “can scale, truly scale, without the degradation we’ve accepted as inevitable.”
Garry Tan supplies the cognitive mechanism the metaphor leaves implicit. He names the failure directly — “the business becomes too big to fit in one person’s head” — and identifies the constant behind the ceiling: “the world is basically built on things that are built around human beings, and human beings can only keep seven plus or minus two things in their head at any given time,” whereas, as the conversation puts it, “you plus an agent can keep basically three Harry Potter books in your head.” His example is an executive whose agent reads the meeting transcripts of his direct reports, letting him see two levels down and identify “what’s broken, what’s in conflict, who’s in conflict with who,” then walk into a meeting with “perfect context from the last three weeks” of a team he was never in. Tan calls this “one of the defining reasons why startups or businesses fail” — “actually like a management challenge.” The implication is that the ceiling was never really about communication style. Whose head the business has to fit in is also what makes Knowing what to build is the scarce half, so agent leverage favors veteran founders possible.
The load-bearing claim is the precise inverse of Context centralization is why coding AI works — git is a solved context repository, knowledge work has no equivalent: coding scaled because git already centralizes context; knowledge orgs degrade at scale because their context lives in meetings and human memory. Make software carry that context and the scaling limit lifts. This is the mechanism beneath AI compresses the distance between idea and execution but not between good and bad judgment at the firm level — communication overhead was a binding constraint, and removing it shifts the bottleneck upward to judgment and When products and models converge, the moat is the institution — how a company attracts people, distributes authority and compounds judgment. It also depends on durable Persistent agent memory preserves institutional knowledge that walks out the door with employees, since context can only bear load if it persists. If software carries the context, the question becomes what the people are for — and Humans move to the edge where the model touches reality — a company brain that can't touch the world is just a database gives the answer Blomfield borrows from Jack Dorsey: intelligence moves into the system, and people move to the perimeter where it makes contact with reality, because a world model that cannot touch the real world is just a database.