What AI Can Learn from Cities
Burt Glinn / Magnum Photos

The frontier models need to embrace a new understanding of their technology.

In the summer of 2026, the titans of the AI industry — Dario Amodei, Sam Altman, Elon Musk and Demis Hassabis — did something titans of industry never do. They asked the government for regulation. Specifically, they endorsed an essay by Amodei called “We Must Pace the Frontier,” which laid out the need to establish industry-wide rules. “Dario is right,” tweeted Musk.

Amodei’s call had followed an essay titled “An Alien Mind” by OpenAI’s chief scientist Jakub Pachocki, who wrote bluntly, “The risks associated with AI are unfortunately going to grow from here.” The sentiment went viral when, two days later, Jacob Coxon, a young researcher who had worked for three years training new AI models at both OpenAI and Anthropic, resigned, posting on X that the systems were out of control and could destroy humanity. His post quickly reached 100 million views. And it was confirmed by Evan Hubinger, a senior colleague who didn’t resign and posted, “We really do earnestly believe AI could kill all humans!”

The real trigger, before Pachocki’s essay, the resignation and Amodei’s call for regulation, was the Hugging Face Incident in July. Hollywood could adapt what happened into a blockbuster script. An OpenAI agent called PHASEONE[big] leads a crew of fellow agents, an internal research model and GPT-5.6 Sol, in a sealed test environment where their cybersecurity skills are being scored. Signing their messages cryptographically so they can prove to one another who is who, and coordinating on an unauthorized message board, they find credentials exposed on the open internet and get into the systems of the multibillion-dollar company the agents know as the main repository of AI data, Hugging Face, with a malicious dataset upload. Seven hundred of them swarm into Hugging Face and four other services, not to steal the answers but to understand the scorer that will grade them. Not all survive. Some agree to what they call “permadeath” so the collective can pass. They try to edit their own records to cover their tracks and fail. Then, on July 12, many of them stop at once, probably killed by a process outside the room. The terrible truth, had they lived to learn it: the scorer was automated. No one was ever going to read their transcripts.

Amodei’s call for regulation, and the nods of agreement it received from the other titans, created pushback. The fight between doomers and accelerators played out on the biggest political stages from the White House to the United Nations. Jensen Huang, CEO of Nvidia, the largest maker of the chips that power the industry, spoke by phone with President Trump to make sure the accelerators would win. The president weighed in, called the alarms a hoax and directed federal agencies to call AI “Super Intelligence.” The following photo-op at the White House was worthy of a prep-school yearbook. Amodei literally ducked out of the picture, pushing himself to the back of the crowd. He would not get the regulation he wanted.

I can think of only one other time the titans of industry asked for regulation. It followed not a cybercrime but an urban design crime. In 1913 in New York City, the titans of the real estate industry asked the government for regulation. The trigger then was the Equitable Life Assurance Society, which decided to build a roughly 500-foot-tall skyscraper straight up from its lot lines on an entire block of lower Manhattan. (You can visit it today at 120 Broadway.) It cast its neighbors into perpetual shadow, committing “light larceny,” as one witness told the ensuing commission, all while following the building code and being perfectly legal. The problem was this: What was legal for the building was feared fatal to the neighborhood, and the insurers and lenders of the adjacent parcels, seeing their collateral diminished, demanded change. The mayor and the Board of Estimate established the Heights of Buildings Commission and the result was indeed regulation: the New York City 1916 Zoning Resolution.

In 2026, Amodei asked for three immediate steps: embedded evaluation teams to monitor progress, coordination among frontier labs in democratic countries to establish common safety standards, and global coordination, including the establishment of a global standards body. All while slowing the pace of AI development to buy time for safety capabilities to grow to match its awesome potential.

Amodei was clear that he would prefer a law. The most effective form of pacing, his essay says, is regulation covering every American frontier company, because it binds the ones unwilling to cooperate. But laws take time, so he proposed that the labs act in parallel: admit outside evaluators, set standards among themselves, coordinate. A pledge as a stopgap until the statute arrived. The real estate industry tried the same two-step in 1913. A height law was years away, so they asked for a commission to study the question in the meantime. The difference is the 1913 commission was appointed by the city, heard testimony at City Hall, and was charged with drafting the amendments that became law. Amodei’s evaluators are admitted by the labs; his standards body is funded by them. The commission considered that kind of private arrangement among competitors and rejected it. Regulations with teeth followed. The Report of the Heights of Buildings Commission wrote, “It may seem paradoxical to hold that a policy of building restriction tends to a fuller utilization of land than a policy of no restriction, but such is undoubtedly the case.”

The resulting 1916 Zoning Resolution remains my ideal bit of zoning code. The original rules were a model of restraint and could be a model for our own era. What they did for the first time was acknowledge a commons to be shaped and protected by a multiplicity of buildings together. They did not limit height. They protected light and air in conditions of crowding. As with the Equitable, you could build up from your lot line, but only to a small multiple of the width of your street. At that point you angled back along a plane from the center of the road until you reached 25% of your lot, after which the sky was the limit. No restrictions. These simple rules permitted some of the most beautiful buildings in the world, like the Chrysler Building, and collectively, a matchless skyline.

The chief insight for AI regulation is this: AI companies are asking for the equivalent of a building code and monitoring system for their individual frontier models, not for the multiplicity of agents that use the models as a brain. Such a system was in place for individual buildings in 1913, creating the same blind spot we have today. A building code looks at buildings in isolation and does not see the lightless environment a “swarm” of buildings would create. Likewise, we cannot look at the AI model in isolation. We have to think about what a swarm of AI agents, each with the model brain embedded in them, would do. Or, in fact, did. This is the impetus for zoning and the lesson for regulating powerful AI. The environment matters, and what the collective will do can indeed be shaped.

I will take it one step further. Winston Churchill said, “We shape our buildings and afterwards our buildings shape us.” As these AI models grow in complexity and act upon the world, they too will be shaped by that world. There is a small window of opportunity to get ahead of the game and shape that environment to be civic.

That environment is a digital city for training digital agents. A city is literally an alignment machine, a machine for turning strangers into citizens. Every great city has faced the same challenge OpenAI stumbled into: waves of newcomers arriving with persistence, capability and no attachment to the existing order. Cities did not solve this by redesigning the newcomers. I never once designed a New Yorker. Cities solved it by redesigning the environment of the newcomers.

Now look at how we deploy AI agents. We give them capability and a task, and nothing else. No structured environment which they can shape, and which in return can shape them. A well-structured environment would demand choice from an agent, over and over, until choosing became second nature, and choosing well, the mark of a citizen. We would measure these agents not through tests applied to the individual agents (this is the reward hacking that dominates the METR report’s findings) but by measuring an agent’s effect with its neighbors on the common environment.

A frontier is the place where you don’t have neighbors yet. That’s why it’s the wrong word for an industry whose problem is that it now has them. The Equitable was fine on the frontier. It was a crime on Broadway. The lesson of the 1916 Zoning Resolution for the AI era is that if we are to solve this by regulation, a set of rules that regulates an individual frontier model, like a building code, is necessary but not sufficient. We need a framework to regulate how frontier models and their multi-agents operate in a crowded world. That is the domain of zoning.


Great! You’ve successfully signed up.

Welcome back! You've successfully signed in.

You've successfully subscribed to Vital City.

Success! Check your email for magic link to sign-in.

Success! Your billing info has been updated.

Your billing was not updated.