Let me start with a question. Why did I fear development done by artificial intelligence?
The answer is plain. AI can build software, and on top of that, it never rests.
AI has no labor law
People rest. There are labor laws. We sleep at night. We need weekends. Work too many days in a row and the body breaks. So there is a ceiling on how much work a person can move forward in a day. For a long time, we treated that ceiling as a given.
But AI has no labor law. It works at night. It works on weekends. Give it an instruction once, and it does not stop until morning. It never says it is tired. It takes no breaks. It keeps working for hours at the same quality.
This difference did not fit inside the word “convenient.” What I felt was fear. This was not a story about one more handy tool. It was a story about the ground under the speed of work changing at the root.
Claude Code came out about a year and a half ago. That is when I understood. The company that runs it 24 hours takes the first-mover advantage. And the company that can punch with money wins. This is not a cynical take. It is the obvious consequence.
The first mover wins — that story is not new. Whoever enters a market early takes the ground. They set the standard. Everyone after them chases the gap. AI widens that gap by the day. A company that moved ten hours forward overnight and a company that stood still overnight are ten hours apart by morning. The gap compounds daily.
Can you catch up by hiring more people? You cannot. Hiring takes time. Post the opening, interview, teach, wait for people to settle in. That takes months. Meanwhile, the other side’s AI keeps moving through the night. The speed of adding people cannot match the speed of adding AI. So the moment a small company steps into a contest of headcount, it loses. It was a ring we should never have entered.
Companies that can punch with money win — obviously
Why can I say it becomes a contest of money? Because there is no ceiling on how fast you can add more.
The AI that does the implementation can be added endlessly, after the fact. Add one connection key — an API key — and you have one more worker, then another. None of the time that hiring takes. No job postings, no interviews, no training. Add a key, and it works through the night starting that day. A company with money can line up as many of these workers as it likes.
There is one more obvious fact. Companies sitting on surplus compute are strong. There are companies that stockpiled graphics processing units (GPUs) for Bitcoin mining. When the market drops, those machines sit idle. Idle machines can be turned to AI work. The side that owns them runs AI cheaper and longer than the side that does not. They have nothing left to buy.
Put these together and the picture is clean. Companies that can buy more AI with money. Companies that can run AI cheap on leftover compute. Companies with both push development through the night, and widen the first-mover gap every day. I found that picture terrifying. I was standing on the small-company side.
The one resource left to a small company: the unused subscription quota
Here, most people drop their shoulders. If it is a contest of money, a small company has no chance.
But I did not stop there. I asked the question again. Is there really nothing a small company is failing to use up?
There was. The subscription quota.
A monthly subscription comes with a usage quota. Most companies use that quota for a few daytime hours. At night, nobody touches it. On weekends, it sits still. The quota is there every month, and most of it gets thrown away. This is not a problem of having no money. It is a problem of not using what you already have.
Factories have a word for this: utilization. The same equipment pays itself off faster the longer it runs. A plant running at half capacity and a plant running flat out get different returns on the same investment. A subscription quota is the same. The payment is already made. Use it or not, the cost does not change. Then there is no reason not to use it. An unused quota is idle equipment. “What a waste” does not cover it. It is a reason you lose the competition.
When I saw this, the top priority fixed itself: use up the leftover quota. In other words, keep the subscription running 24 hours a day. Fail to do this, and you fall behind the world. You drown in the AI wave. Not drowning in that wave — that was the first reason I started building.
History teaches the same lesson. In the middle of the industrial revolution, cotton mills could only run by day. At night it was too dark to see the thread. Then gas lamps came in, and later electric light. Mills could run at night. And the mills that ran nights overtook the mills that ran only days. Same equipment. Similar headcount. The only difference was how many hours they ran. How many hours you run what you already own — that is what decided the contest.
Economics has a famous observation called the Jevons paradox. In 1865, William Stanley Jevons pointed out that as coal was used more efficiently, coal consumption went up, not down. When efficiency makes something cheaper, people use more of it. AI is the same. AI that has become cheap and useful is not something to cut back on. It is something to use more. So there is no reason to leave quota on the table. Leftover quota simply expires. It does not come back next month. Quota you did not use up is quota you threw away.
Judgment goes local; implementation goes to the cloud
That said, running 24 hours blindly is not the point. If it runs all night, mistakes pile up all night too. Misguided implementation can march forward in the hours nobody is watching. Close that hole, or morning brings nothing but regret.
Do not leave contradictions alone. Close them with structure. That is my number one rule. I do not believe in “we will push through on grit.” If a structure cannot close the hole, that structure is unfinished.
So I built an AI development orchestrator that puts a local large language model (LLM) in the judge’s seat. Its name is local-commander. Its role is not command — it is sorting. What gets processed where. It decides that first.
Every incoming task gets sorted three ways:
- What can be done locally
- What needs the cloud
- What requires human approval
The sorting judgment is done by the local LLM. This is the crux. If every judgment went out to an external service, it would eat both quota and money. What can be decided locally gets decided locally. That is faster, cheaper, quieter. There is one more reason the judge sits at home: judgment happens many times. Put the highest-frequency step somewhere expensive, and it becomes your biggest burden. So I put the most-run judgment in the cheapest seat.
After the judgment, things just flow. Work that needs implementation goes to Codex or Claude Code. Steps that need no human eyes move straight ahead. Merging the change, opening the pull request, review, and syncing the task after merge — the whole chain is automated. Humans touch only what requires approval. Everything else, the structure carries to the end.
Let it work the night; stack only approvals in the morning
A day looks like this.
At night, I leave instructions and step away. The structure runs on its own. Locally doable work finishes locally. Work that needs the cloud gets implemented by Codex or Claude Code — push, pull request, review, merge, all the way through. Only the approval-required items stop, just short of the line. They hold there and wait for morning. Nothing moves past that point on its own. If it did, mornings would greet me with things I cannot take back.
In the morning, what sits in front of me is one kind of thing only: the items that require approval. Not the work that finished overnight. The one point only a human can decide. I look at just that, and answer yes or no.
With this, the quota works nights too. And in the morning, my job is judgment only. What comes back to me is not the work. It is time. This is exactly what I want from efficiency. Not speed itself — the time that returns. With the returned time, I think about what to build next. That is the part only a human can do.
Remember where this started. I feared that AI never rests. Now I hand the night to the AI that never rests. Its nature has not changed. What changed is one thing: how I use that nature.
No gut purchases: DGX Spark was decided on 16 measured cases
Decide to run 24 hours, and the next question arrives. Is the compute at hand enough? Running the judge LLM locally through the night takes real hardware. My candidate was the DGX Spark.
But here too, I did not buy on gut. “Looks good, let’s buy it” is not a way I buy. To tell myself a purchase is allowed, I need evidence. The more irreversible the purchase, the thicker the evidence.
So I built 16 test cases and measured. I lined up the numbers and grounded the decision in them. I do not buy on impressions. I buy on measurements.
Why go this far? To keep the quality of judgment from sinking to the level of a hunch. I believe human work is judgment. Tasks can be handed to AI. But what to build, what to buy, where a human must stop — that judgment stays with people. If judgment drifts into hunches, the whole 24-hour structure rots with it. Build automation on top of a hunch, and everything above it is a product of that hunch. So the one-shot judgment of a purchase gets fixed in numbers, more than anything else.
Contradictions get closed with structure, not left alone
Let me answer the questions that always come back, before they are asked.
“If AI runs all night, can you keep the quality?” The sorting answers this. Steps that need human eyes get routed to approval-required. What is delegated is only the work that needs no human eyes. Not everything is delegated.
“What if it runs wrong until morning — isn’t that an accident?” The stop point answers this. Approval-required items stop before implementation. In front of every irreversible step, the structure waits for a human. What moves at night is only what can be undone.
“Is a local LLM good enough for the judgment?” The division of roles answers this. The local LLM does light judgment: sorting. Heavy implementation goes to Codex and Claude Code. Hard approvals go to humans. The local seat holds only the judgment it can carry. It is never asked to overreach. That is why “not good enough” does not happen.
“Doesn’t 24-hour operation cost too much?” The order of consumption answers this. What gets used first is the quota already paid for. As long as it runs inside the quota, no extra cost appears. Before the quota runs out, the local-versus-cloud sorting kicks in — the cheap stuff finishes locally. I do not start saving after the costs rise. I put the design that keeps them from rising first.
Every answer has the same shape. No difficulty is crossed on grit. Each is closed with structure. Find a contradiction, and do not leave it alone. That is the one thing this article most wants to say.
The concrete inside of this approach is in the related article, “Half of AI development moved to a local LLM — sorted by role.” What sorting by role means, and how much can be trusted to local — that is where the details live.
One last thing. A 24-hour structure is not reserved for big companies. If anything, the small companies throwing away their quota have the most room to grow. You cannot punch with money. But you can use up the quota you are leaving on the table. You no longer have to be on the drowning side.