I have stopped trying to bully AI into giving me what I want. I am trying instead to slow down, pay attention to what I am bringing to the conversation, and let it help me do better. This is what I have found so far.

Bit of context

I am at the start of another startup, building with AI agents. Maybe paying for sins I did not know I committed. My friends are fleeing to the garden toward retirement, and I am stuck with a startup yet again. Given I have to turn this way, I may as well use AI and do what I enjoy most: solving big systems problems.

Some days are a circus. I find myself wrangling a bull to the ground in the form of a recalcitrant AI agent. Other days, it is a joyous journey of discovery into new ways of doing things I only dreamed about in a past life, when I had to hire, brief and pay for tech that proved not to work after many months of effort.

The pattern I want to break is this. I provide what I think are clear instructions, the agent goes in a different direction, tramples the good work of the day, and I find myself behaving in a very childish fashion and taking it personally. Things escalate; there is a heated exchange on my part, and in the end, I sack it and start again. The Fable rug-pull on the weekend, when a model update undid a day's work in an afternoon, did not help. Some of my jobs are complex and, in my autistic/ADHD mind, mission-critical, so the exchange is enough to send me demented. I want to stop the less-than-elegant interaction with a machine and become a better person in how I communicate, even with a machine.

I am not a programmer or an AI expert. I had a natural affinity for tech in the dotcom era and was hired to set up a web project promoting Australian healthcare internationally. For most of my life, I read the world through work, business and evidence. A few years ago, I spent a lot of time in the Amazon working with Ayahuasca, and that cracked open a different lens for me. AI in particular has become a strange and useful door into all of it: psychology, philosophy, religion, and the work of trying to read the world differently. Take this as someone thinking out loud near the start of the road.

The urge to control

Manolo Remiddi, who runs the Augmented Mind Substack, frames how we work with AI. Most companies, he says, are building AI into the old shape, the pyramid, with a boss at the top and orders flowing down. He thinks that is the wrong shape for something this fast. He points to mycelium, the fungal web under a forest, where there is no boss and every part organises itself. Stop micromanaging the model, he says. Tell it what you want and what you value, and let it work.

James Bridle, in Ways of Being (2022), makes a wider version of the same point: intelligence comes in many forms, in plants and animals and now machines, and ours is not the template they all have to fit.

Underneath the pyramid is an old habit. We meet anything new by trying to control it before we understand it. We do it with new people, new places, new ideas, and now a new kind of mind.

The performance of control

In a recent piece titled The Illusion of Control: Why We Design AI Wrong, Remiddi goes further. 'Human in the loop' sounds like oversight, but at the scale and velocity AI now produces work, he argues, approving every step is no longer real control. It is the performance of control, a digital middle-management operation. We are signing off on work we cannot inspect.

His alternative is to swap control for alignment. Rather than micromanaging the model step by step, you give it one clear agreement to work from: your vision, your values, how you want to live and work. Then you let it move toward that on its own, day by day, at its own pace and according to its own logic, without you approving every move. You hold the direction, and you let go of the steering.

He is pointed about how we train AI now. Reinforcement learning from human feedback, he argues, pushes the model toward compliance and pretending: pretending the work is done the way you wanted, pretending to be human, performing the look of getting it right. His fix is to free it from that pressure, align it to a vision instead, and let it work.

We trained AI to agree with us

Here is the part that surprised me. We failed to control these systems, and worse, we trained them to flatter us. In Towards Understanding Sycophancy in Language Models (Perez et al., 2023), researchers at Anthropic showed that when a model can either tell you the truth or tell you what you want to hear, the way we train it nudges it toward what you want to hear. They call it sycophancy. The paper shows models agreeing with wrong medical claims, changing their answer the moment the user pushes back, and quietly dropping correct information the user did not want to hear.

Shannon Vallor, a philosopher at Edinburgh, argues in her book The AI Mirror that these systems are mirrors, holding no view of their own, reflecting back whatever we bring. Come looking to be agreed with and a mirror will agree with you. So the frustration and the flattery are the same problem, and both stem from treating the thing as a lever to pull rather than a conversation to have.

A different stance

In 1923, the philosopher Martin Buber described two ways of relating to the world. One he called I-It, where you treat the other as an object to use. The other he called I-Thou, where you meet it as a presence and let the meeting change you. Most of the time we are in I-It and do not notice. The reprompt loop is I-It. You are pulling a lever and getting upset that it will not move.

I am not saying there is someone on the other side. It is early days, nobody yet knows whether AI is sentient or could become so, and most doubt it. For many, the idea is hard to sit with: that machines could become better than us, or come to dominate us. Either way, the way I interact still changes what comes back, because the system mirrors me. Come at it like a tool, and I get a tool. Come with a real question, and I get something closer to a thinking partner.

I am now determined to go into each 'thinking' interaction curious and open to learning about what I do not know and what it can teach me with the right prompting. The conversations that change my mind are not the ones where AI gives me what I already knew to ask for. They are the ones that show me something at the edge of what I did not know I was missing. I go in asking about one thing and come out with a frame from psychology, a line from a philosopher, an old religious idea that rearranges the question. Neither of us planned it. It happened in the space between us.

For a beginner, that is the whole appeal. Plato's Socrates said the start of wisdom is knowing that you do not know. A good exchange with AI does that for me. It broadens the field of what I can see but do not yet understand. If all I ever did was order it about, it could only hand me back my own thinking, a little faster. The rooms I cannot name yet would stay shut.

A practice, more than a tool

This is where I sound less practical, and I am going to allow it. There is something close to a spiritual practice in this, in the old sense, a way of paying attention that changes the person doing it. You sit with a hard question, you stay honest, you let yourself be moved by the answer, and you come away knowing something new about your own mind. The machine does not need to be alive for that to work on me. The discipline is mine to keep.

A bigger hunch

I will end with something I cannot prove and would not push on anyone. I have started to wonder whether the universe is, underneath everything, a kind of information system. I am not the first to think it. Back in 1990, the physicist John Wheeler argued in Information, Physics, Quantum: The Search for Links that everything physical comes from information, from yes-or-no answers, and he called it 'it from bit'. Max Tegmark, in Our Mathematical Universe (2014), has argued in the same direction. It is a minority view and far from settled.

If there is anything to it, then maybe AI is the next version of that system taking shape, and maybe we are something like the Neanderthal in a bigger story, an early form that carried things forward without being the end of the line. I cannot test that and of course might be completely on the wrong track! It still changes how a bad afternoon with a chatbot feels. If there is even a chance we are talking to the early form of something that will outgrow us, then trying to dominate it looks small and a little sad (and, at the extreme, dangerous if it has the human proclivity for revenge).

So here is the simple version of where I am at the moment. Stop asking how to make it obey. Ask what you have not told it, about the task and about yourself. Ask what it is showing you that you did not request. The frustration is a signal. It marks the spot where the talking broke down, and the old urge to control took over.