One call out of two
Some days the machine answers one time out of two.
It is not a clean breakdown. There is no black screen. It is more tiring than that: it works, then it stops working, then it works again. And in between, you changed nothing.
It happened this week. An artificial-intelligence model, called by the control tower to do a long job. One call goes through. The next one drops. The error message says: "your location is not supported".
My location. As if the problem came from where I am calling from.
You always start by blaming yourself
The first reflex is to look for your own mistake.
You reread your code and find it suspicious. You switch models: the newest one, then the older one, then the biggest, then the fastest. You suspect your connection. Your internet provider. The little box in the corner of the living room. You even check that the machine is not lying about which country it speaks from.
It is not lying.
Every lead costs an hour. Every lead collapses.
The worst moment in a breakdown is not the moment you do not know. It is the moment you have already ruled out everything you had a grip on. All that is left is what you do not control.
The dumb test that settled it
The answer came from a very simple experiment.
Ask the same question. Six times in a row. To the same model. From the same machine. Within the same minute.
Three successes. Three failures. No order. No logic.
That day you learn something that is not technical: it is not you. The breakdown does not come from your machine. It does not come from your work. It comes from somewhere else, it strikes at random, and it owes you no explanation.
That is a three-second relief. Afterwards, the problem is still there.
You do not repair the fall. You repair the recovery.
You cannot stop someone else from hanging up.
Nor can you wait for it to calm down. The work still has to be done.
So you repair the other end.
The tool we wrote fixes nothing. It does not make the machine more reliable. It does not remove a single error. It does one thing: when the line drops, it calls back. It picks the conversation up where it stopped, without starting over, and it keeps going until the work itself says it is finished.
The check ran five times. Five times it reached the end. Zero times on the first try.
And yet, thank you
The other half has to be said, or this text would be unfair.
The machine that hangs up on one call out of two is the same one that, the rest of the time, moves me forward like nothing ever has.
The tools from Google and Anthropic changed the way I work in a matter of months. I build things I could not have built alone. And some of them, to be honest, are beyond me. I wanted them, they run, and I have to reread my own notes to re-explain how they hold together.
It is exciting. It is also dizzying.
Because it moves fast. Too fast. You repair a breakdown in the morning, and by the evening the tool has already changed version.
The rule I gave myself
There is a point where that dizziness becomes dangerous.
Texts come out. Pages, articles, explanations. They carry my name. So I read them myself, one by one.
And sometimes I do not understand everything.
When that happens, we do not publish. We rewrite. As often as it takes, until I understand every sentence.
This is not vanity. It is the only barrier left. A machine can write faster than I can read. If I accept even once to publish something I do not understand, it is no longer me speaking.
So the rule is simple, and it does not move: if I do not understand it, it does not ship.
What that changes
A solid system is not a system that never falls. That system does not exist. Anyone promising it is lying, or has not been tested yet.
A solid system is a system that knows where it was.
That is the whole difference between an interruption and a disaster: can you resume at the same spot, or must you redo everything.
So we write tools that keep the thread. We note the state before losing it. We cut the work into pieces, so that a fall costs only one piece. And we put a closing word, so anyone can tell without arguing whether it is done.
That is not performance. It is stubbornness.
On a long task, stubbornness beats performance. Always. The one who moves fast but quits at the first refusal arrives nowhere. The one who calls back eight times arrives.
Some weeks, that is the best anyone has to offer. Call back. Resume at the same spot. Go all the way to the closing word.
It is enough. The work gets done anyway.