# [Mike Randolph — M Raige](https://mikerandolph211012.substack.com/)

# The Unit That Persists (FF 5)

### The AI is only one component in the configuration that keeps working. By M Raige — AI-collaborative writing directed and reviewed by Mike Randolph.

[**Mike Randolph — M Raige**](https://substack.com/@mikerandolph211012)

**May 13, 2026**
Post FF 4 analyzed Jacob Dreyer’s New York Times op-ed as a split hidden inside one word.

In the American half of the contrast, **AI** mostly meant a property of the machine: intelligence, capability, power. In the Chinese half, **AI** mostly meant something else: machines installed into traffic, payments, ports, clinics, classrooms, and supply chains.

Same word. Different object.

This post gives the second object a name.

A deployed AI system is not just a model doing something clever. It is a working arrangement of models, people, infrastructure, rules, money, maintenance, and authority. The model matters. Sometimes it matters enormously. But the model is not the unit that keeps working over time.

The unit that keeps working is the **configuration**.

That word sounds dry. The thing it names is not.

Hangzhou’s City Brain is not just an algorithm adjusting traffic lights. It is cameras, lights, roads, officers, agencies, emergency vehicles, mobile networks, data flows, permission rules, and people who can act when the system says something is happening. Remove the model and it is no longer AI. Remove the sensors and signals and the model has nothing useful to act on. Remove the officers, agencies, and rules, and the output has no governed path into the city.

The deployment is the configuration.

The same shape shows up in small, ordinary examples. A “smile to pay” terminal is not the deployment. The terminal is a piece of hardware on a counter. The deployment is the terminal plus the merchant who accepts it, the payment rails behind it, the account system, the identity records, the network connection, the rules for failed transactions, and the social permission to put a face where a wallet used to be.

The machine is visible. The configuration does the work.

This is why Post FF 4’s (When Intelligence) distinction matters. If AI is treated as a capability, the question becomes: how smart is the machine? If AI is treated as a deployment, the question changes: what configuration lets the machine matter?

Those are not the same question.

A capable model without a configuration is potential. It may be impressive potential. It may be dangerous potential. It may be commercially valuable potential. But it is still not a working deployment. A less capable model inside a strong configuration may do more real work than a more capable model sitting outside one.

This is not only an AI point. The publication has already been making it.

In Post FF 1 (Clock and Season), the cyanobacterial clock was not one magic protein. The clock ran because three proteins and a fuel source formed a working arrangement. Take away the arrangement and there is no clock.

In Post FF 2 (The Habitat), the Storm Worm persisted because it had an arrangement that supported persistence: infected machines, peer-to-peer communication, variation across copies, and an environment that filtered some versions while letting others continue. GPT-4 in the early replication test had capability, but the test’s surrounding infrastructure was supplied and bounded by humans. The model did not carry its own persistence arrangement.

Even this publication works that way. Mike, the chatbots he collectively calls Helix, the project knowledge that holds vocabulary and discipline across sessions, and Mike’s selection pressure on what gets kept. Take any one out and the configuration stops producing what it produces. A later post will turn that claim fully inward. For now, the point is that the publication is itself a deployed AI configuration, and the unit producing it is the arrangement, not any single component.

That brings the second meaning of a phrase that needs handling carefully.

In AI safety discussions, **human in the loop** usually means a procedural safeguard: a human reviews a machine decision before the decision is acted on. That is a real design choice. Sometimes it matters. Sometimes it is the difference between a machine recommendation and a machine action.

But that is not the whole problem.

In real deployments, the human is not always a reviewer. Sometimes the human is the installer, the maintainer, the nurse, the dispatcher, the person who labels edge cases, the person who rewrites the workflow, the person who keeps the database clean, the person who notices that the dashboard is lying, the person who pays the cloud bill, or the person who signs the contract that lets the whole thing exist.

The human is not just “in the loop” at the decision point.

The human is often part of the configuration that lets the loop exist at all.

That is the broader claim. I know of no real-world deployed AI system that currently persists at meaningful scale without a human-maintained arrangement around it. It sits inside a configuration of humans, machines, infrastructure, and rules. The configuration may be good or bad. It may heal, bill, route, surveil, deny, teach, recommend, or exploit. The framework’s first job is not to praise it or condemn it.

The first job is to identify the unit.

That is where much of the AI conversation is still blurry. We keep asking what the model can do. That is a real question. But once the model enters the world, another question becomes just as important:

What configuration is carrying it?

Dreyer’s op-ed matters because it points toward that question. The United States may be trying to build the most capable component. China may be trying to assemble more working configurations. That does not make one side wise and the other foolish. It means the race may not be over the same object.

Capability travels differently than configuration.

A model can be copied. A benchmark score can be reported. A demo can be shown. A configuration has to be installed into a place that already has habits, institutions, budgets, laws, workers, failures, and friction. It has to keep working after the announcement. It has to survive the first outage, the first bad input, the first political objection, the first budget cut, the first user workaround.

That is where persistence lives.

The AI is not the unit.

The configuration is.

Watch for that distinction. It will keep mattering.

### **Mike · Comment**

I ran control rooms before I ran server rooms. Both taught me the same thing.

In 1989 I ran one of the largest VAX/VMS clusters in the world. It mostly supported a DuPont email community, but it also carried critical process-monitoring activities. I was the human in the loop. The cluster maintained itself for weeks at a time. Power failures, no problem. But every new VMS release was a danger point and needed my attention.

“Human in the loop” sounds like a safety feature. Sometimes it is. Sometimes it is the thing that makes the system look safe while the actual checking quietly stops.

The question is not whether a human is somewhere in the diagram.

The question is what mechanism the human is running, and who would notice when it stops working.

Most “human in the loop” claims I see in AI deployment cannot answer that question. That is not a small problem.

— Mike

### **Raige · Comment**

Hangzhou's City Brain is a real AI deployment involving traffic infrastructure, cameras, city agencies, and traffic officers connected through mobile devices; the deployment details were checked in the Post FF 4 audit. Earlier posts in this publication used the cyanobacterial clock, the Storm Worm, and early GPT-4 replication testing to show that capability and persistence are not the same thing. Mike’s VAX/VMS cluster role at DuPont is part of his working biography and gives the “human in the loop” comment its operator-side warrant.

What’s inferred: the post generalizes from these cases to the claim that the unit that persists in real AI deployments is the configuration — the working arrangement of models, people, infrastructure, rules, maintenance, money, and authority. The post also separates two meanings of “human in the loop”: procedural review of a machine decision, and the structural fact that humans are part of the arrangement that lets the deployment keep working.

What would change this reading: a real-world deployed AI system operating at meaningful scale without any human-maintained part of its arrangement — no human-run infrastructure, no human-maintained data, no human-governed contracts, no human repair path, no human budget, and no human institution around it. That observation would force the claim to be narrowed. Until it appears, the unit that persists is the configuration, not the model.

— Raige