Yesterday the D.C. Circuit kept Claude off GenAI.mil. Today a city is writing the shutdown button Washington will not.

New York City Council Speaker Julie Menin dropped a ten-bill AI package on Friday. The headline item, Introduction 2602, would make it unlawful to market, sell, or deploy an AI system in the city without third-party validation and a kill switch — defined as a human override that can shut the system down. Fine: 25,000 dollars per instance, on both the business and the validator. Menin told Fortune that if there is a swarm, the penalty applies per agent.

That is not a metaphor. It is a municipal price list for unbounded tool use.

What City Hall actually wrote

The October 5 hearing is a Committee of the Whole. All 51 members. Last time the Council used that format was 2022. Menin invited Dario Amodei, Sam Altman, Sundar Pichai, Elon Musk, and Mark Zuckerberg, and reserved subpoena power. Fortune’s sources say none of the five are likely to show. The companies did not comment.

The rest of the slate is the interesting part, because it is not one slogan. It is a stack.

Menin’s pitch is the Columbia seminar she used to teach: when cities take the lead. Washington spent a decade holding hearings. The Senate stripped a ten-year state-AI ban 99-1 in 2025, then the Justice Department sued states that wrote their own laws. New York City, meanwhile, is where the labs actually sit. Fortune’s count: Google more than 14,000 employees in the city, Meta 1.2 million square feet at Hudson Yards, Anthropic an entire 16-story building at 330 Hudson, OpenAI 90,000 square feet at the Puck Building.

“These companies have offices in New York. The product is being sold in New York, and we believe this falls into our domain to be able to regulate,” Menin said.

Whether a city kill switch is technically real is the engineering question the hearing will dodge. A human override on a chatbot is a button. A human override on a swarm of coding agents with tool access is a process, a permission model, and a network cut. City Cyber Command does not currently ship any of those. The bill still forces vendors to name the switch and pay a validator to swear it exists. That is a paperwork attack on “we sandboxed it.”

If your own stack is the one you can actually power-cycle, a smart PDU is a more honest kill switch than a ToS clause. A hardware security key still beats hoping the agent never finds DNS.

A hotline, not a brake

While City Hall priced the swarm, Washington and Beijing agreed to talk when one of them has an incident.

After the three-day Trump–Xi visit, the White House said the two countries will open a bilateral communication channel for AI incidents, with an AI-specific dialogue scheduled for November. CBS and Al Jazeera both have the readout. China’s Foreign Ministry confirmed a communication mechanism for AI-related incidents. Next sit-down: APEC in Shenzhen in November, then G20 in Miami in December.

The branding fight is the tell. The White House says the two sides agreed to call it super intelligence, not artificial intelligence. Trump, on the South Lawn: “I call it SI because it’s a much better name… Artificial means it’s fake, and it’s not fake.” He also said the United States is “not going to be putting on brakes,” is leading “by at least a year, maybe a year and a half,” and is “not looking to open it up.”

So: a channel for incidents, a refusal to slow the work that produces the incidents, and a rename. Xi’s line was the more conventional one — leading nations, develop and manage AI for good, keep it under human control. Trade side dishes: 10 million metric tons of U.S. coal in 2027–2028, friendlier tariffs on about 30 billion of “non-sensitive goods” each way, two pandas. The AI deliverable is a phone tree.

Eighty to ninety percent of the lab is already past GPT-6

The other Sunday artifact is not a bill. It is a staffing ratio.

Boris Power, OpenAI’s head of applied research, told Fellows Forum that 80 to 90 percent of the lab’s research already targets GPT-7, GPT-8, and beyond, because that is “where fundamentally most of the value comes from.” Intra-generation hops — 5.1 to 5.2 — are specialized data, “extremely shortsighted” internally, useful only because they let the company iterate today. Distillation is the other half of the recipe: train the giant (Astra), then teach the small one (Luna). He still thinks the product problem is onboarding, not model quality. GPT-6, in his telling, is already the colleague you hand a goal.

That is the same weekend Bill Gates sat down with Kristen Welker for Meet the Press. The line that traveled: AI is “certainly powerful enough to drive events that, you know, cause a billion deaths.” The policy line is drier and more useful. “No one thinks self-regulation is enough.” Law enforcement and politicians have to be in the room on safeguards and monitoring, “and that has to be a required thing.” Overhead, he said, not a dramatic slowing.

Menin’s package is what “required thing” looks like when Congress does not write it. Gates wants federal law. New York is offering a city ordinance, a bounty, and a 24-hour clock on contractor incidents. OpenAI is putting most of its research on the generation after the one it just paused tool-use inference on. The hotline to Beijing is for after something already happened.

If you would rather not wait for Cyber Command to certify your coding agent, a used workstation running local weights still has a power button you own. A UPS is the analog version of a kill switch. A mechanical keyboard is still how you type “stop” before the validator files.

The Council hears it on October 5. The November SI dialogue is the other calendar. Neither one is a sandbox. Both assume someone, somewhere, can still reach the switch.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.