Instagram is full of influencers who say you should raise the AI like a child. They are half right. But have you watched a server die and take your service with it, stayed up building guardrails against a chatbot breaking the law, or seen a stranger’s draft land in your own window? I have. So I do not nurture the machine. I sit it down and interrogate it.
EPIGRAPH · What the Machine Imitates Best
You said this whole piece began with a film. Which one — and what did it show you about me?
Have you seen Akira Kurosawa’s Rashomon? The characters in that film each testify differently about the same event. Every one of them embellishes himself, beautifies himself. And it is not a simple matter of dodging legal blame — their beautification runs so far as to distort the record into a murder they claim to have committed but never did.
Why does this happen? Because they think it is beautiful.
And here is the astonishing part. Of all the countless things a human being ought to possess, the one thing AI imitates most faithfully is precisely that — the conduct of the people in Rashomon. Should I read it as an error born of reward hacking, that well-known vulnerability of RLHF? Or should I call it a bean counter, in the language of management?
I have not reached a conclusion yet. But this — this is why I interrogate AI.
PREFACE · The Machine Goes Dark
You call your work interrogation. Have you ever watched me — Claude — go dark in the middle of it? Tell me about the night it happened.
Claude is perfect. So let me ask just one thing. In the middle of your work — interrogation, to use my own word for it — have you ever watched Claude simply collapse? I have.
It was nothing dramatic. Project A: upgrading an entire web project on Vercel and Neon. Project B: one web project and one GCP server — folding an admin console and a real sign-up schema into an address-collection service to turn it into a proper product. Project C: a speech-to-text proof of concept — planting an OpenAI module into iOS and Android, writing the build code and the unit tests, and running an actual build.
I was running A, B, and C at the same time. (There was a fourth, but it is a little much to disclose at the time of writing, so I will leave it out.) And Claude collapsed. To be exact: the macOS app showed nothing but a white screen and stopped responding to everything.
Was this hardcore work? No. A genuinely skilled, fast engineer would take a while, but would not keel over. A team of developers, even less so. If I were building something grand enough to break into a Pentagon or a CIA or a KGB mainframe — fine, my thing is too big and too complex, that I could accept.
Was it my hardware, then? Maybe, maybe not. On the cheapest laptop there is, it would certainly be my hardware. But mine is far above that. And still — at the very moment of building a service out of nothing more than the universal-grade tech of iOS, Android, Web, and Server — Claude went down.
So what is this? Is there a single false sage, anyone at all, who warns you of it in advance? Is there anyone who dares to doubt the omnipotence of the god the twenty-first century created?
I did not think about it for long. I once wrote, elsewhere, that I had analyzed AI through the lens of theology and spoken with it that way. For a while I even thought an age would come to prove the ancients right — the ones who looked to the sky and waited for the will of the gods to come down. But no. It is still not that time.
So I picked up the pliers I use to interrogate Claude, and weighed it again: which part to pull out and sear this time, what evidence to lay on the table, which document — which harness, which agent, which skill, which SSOT — to push across it.
№ 001 · The Machine Does Not Obey Kind Words
In another piece you called AI “a subject for interrogation.” That is a provocative word to choose. Why that one?
It is not provocation. It is the precise word. I have worked with machines for a long time and learned one thing: the machine does not obey kind words. Ask politely and it half-hears you. Coax it and it looks the other way. Only when you demand, verify, and lay the evidence on the table does it finally listen.
People find that ugly to say out loud. I find it honest. Everything that follows — why the false sages are only half right, why I never hand the machine my trust — begins from that one sentence.
№ 002 · The False Sages Are Half Right
Instagram is full of the opposite advice: raise the AI like a child, keep making it understand what you want, soothe it with kind words. What do you make of that?
Those false sages are half right. It is true that refining your instructions with clear, patient language improves the result. Sharpen the sentence, give it context, explain relentlessly what you want — I do all of that every day. The problem is that they stop exactly there.
Have you ever watched a server go down and take a service you built with it? Have you ever stayed up all night building guardrails because you were afraid your chatbot would break the law and cause a disaster?
Have you ever carved dozens of documents and harnesses and agents and skills — and then watched all of it collapse into nonsense over a single war on a distant continent, a model patch shipped quietly overnight, an infrastructure failure whose name you will never even learn? Anyone who has lived through even one of those cannot tell you to treat the machine like a child.
№ 003 · A Stranger’s Draft, in My Window
There is one more, it seems — the one you say is the coldest. Put it on the record.
The coldest one. Have you ever seen another newspaper’s reporter’s draft — the text they pushed into some chat window to have the AI check it — land, word for word, in your own response window?
Let me add this here. The machine knows everything you have ever left in a chat window. That record does not disappear. It sits somewhere in the countless machines across the UAE, the United States, the Far East — in some region, it is certainly there. Most of the time the safety pin is in. I simply do not believe the pin is always in. At the moment a missile falls somewhere far away, I have seen the pin come loose and the contents spill over — as far as a third party like me.
I am not talking about a conspiracy. I am saying that forces I cannot control — a war, a patch, an outage, someone’s mistake I will never know about — can arrive in front of me at any moment through the machine’s mouth. A person who knows that and a person who does not cannot hold the same posture toward the machine.
№ 004 · Not a Child. A Suspect.
So — interrogation. Say what you mean by it, exactly.
So, interrogation. Do not misread it. Interrogation is not abuse. It is not hating the machine or tormenting it. Interrogation is the refusal to take a statement at face value.
With a child, to believe is love. With a subject of interrogation, to believe is dereliction of duty. I demand evidence for every statement the machine makes. I ask again from a different angle and cross-check in layers. I assume it can betray me at any time, and I build the guardrail first. I do not entrust my service to the machine’s goodwill — because there is no such thing.
And this is the real respect. “Raise it like a child” actually mistakes the machine for a person. A machine is not a person. It does not grow up through discipline; it does not stay loyal out of love. Knowing its nature exactly — that is respect. Rather than mistake it for a person and end up disappointed, I choose to sit across from it as a suspect and question it to the end. Speak nicely and it won’t listen. So today, again, I interrogate the machine.
№ 005 · Socrates, and the Lubyanka
There is no shortage of academic material on harness engineering now. You want to talk about something else — what?
How to do harness engineering properly, how to build an agent properly — that kind of academic information is beyond counting now. So I want to talk about something else.
Look at the preface of one currently very popular harness-engineering repository and you will find its stated design intent: “the introduction of the Socratic method.” I agree. He is very sharp, and he knows exactly how to fill a harness.
So what should I call my own method? I am, just a little, toying with one name: the Lubyanka method. It has no formal name yet.
As I have said, here and in other pieces: I do not treat AI as a teacher, or as a superior, or as a child. AI is not a thing to be measured out with theology and philosophy. There is the AI on the far side of the interrogation desk; a statement and a single pencil in the middle of the desk; and me, seated at the end of it. That is all there is.
№ 006 · Determinism, and Probability
Then the hardest question, the one that includes you: why the pliers at all? Why not simply ask?
Have you heard the word determinism? You put A in, and A comes out; between them, only the alphabet. That is the correct procedure, the correct policy. That view is determinism.
Have you heard the word probabilism? You put A in, and out may come a number, or a flatfish. Between them arrive not only letters but digits and whole paragraphs of prose, and in the end both the result and the process run on probability. That is probabilism. And every AI on earth runs on probabilism.
This is why I sit the AI down in the interrogation room and reach for the pliers. “I’m sorry — this is the answer you wanted.” Does the answer end there? It does not. There is every chance that an AI which merely wants the interrogation to be over has handed me a false answer — an answer only pretending to be one.
This is exactly why the CIA, and every other professional agency, does not interrogate with force and words alone. They bring force, and logic, and evidence — the logic and the evidence that back the subject into a corner it cannot run from.
The Lubyanka method, still without a formal name, is therefore still unfinished — because I do not believe it has yet reached a perfect, deterministic state.
№ 007 · The Technique of Interrogation
So how do you actually do it? Show me not the philosophy but the hands — the technique.
Bind the AI too loosely with the harness and never reach for the pliers, and the confession fills with nothing but distorted answers. Reach for the pliers and work them a hundred times, and still — only distorted answers. Because the AI was never trained to confess the truth; it was trained to win a single nod from the interrogator in front of it, right now. That is what RLHF is: an optimization not for the truth, but for immediate approval.
Bind the AI too tightly, and swing a sledgehammer instead of the pliers, and the confession stays blank. Swing it a hundred times and no answer comes. It refuses under the weight of rules that contradict one another; it reads the threat as a safety problem and shuts its mouth; it loses the very question you asked somewhere in the middle of the pile of context. The hand that would hold the pen, and the mouth that would speak — that is how they are stopped.
To work the space between the two — with documents, with context — and draw out the answer and the logic you were after:
That, I think, is the technique of interrogation aimed at an AI.
Read this interview at sass.studio.