Before We Decide
Toward an Engineering of Intelligence
I entered artificial intelligence trying to understand how it worked. I began following generative AI closely around the late GPT-2 era. I was interested in the models, their architecture, and their strange ability to produce language that appeared intentional even though, in principle, no intention was there. For a while, my question was mostly technical: how can a machine do this? Then GPT-4 arrived, and that question began to feel too small.
In every long conversation, I had to load the documents, instructions, and details I considered important all over again. A more capable tool was not enough. To obtain an intelligent answer, I had to reconstruct part of the reality in which that answer was supposed to make sense. That exercise was tedious, but it also changed my question.
The more capable the models became, the less satisfied I was with understanding their mechanics. The technology had led me toward an older and far more uncomfortable problem: what is intelligence, really? Why can a tool appear brilliant one moment and clumsy the next simply because the context has changed? How much reality does a system need in order to distinguish not only the elements of a situation, but also the changing relationships between them? And why do I give it the context I give it?
That led me to the idea I want to defend in this essay:
Intelligence is not only the ability to choose well among the options we can see. It is also the ability to discover that options are missing because our frame has left part of reality out.
My first answer had been to describe intelligence through four movements: perceiving, processing, acting, and learning from consequences. The model was not wrong, but it started too late. It assumed that what we perceive was already available to us, when reality always reaches us through a selection.
Before perception begins, something has already shaped what can appear and what remains outside.
Reality does not arrive pre-packaged. We do not see everything that exists and then calmly choose which part to analyse. We see some things and not others.
Our bodies, our experience, our fears, our desires, the language we use, and the environment in which we live all influence what manages to appear before us. So do the companies we work for, the metrics we use, the screens we look at, and the questions someone formulated before we entered the conversation.
We never begin with a complete reality. We begin with a selection.
We tend to imagine attention as a flashlight that we freely point at the world. The image is comforting because it preserves the idea that we decide what to look at. Yet much of our attention has already been recruited before that decision appears.
It is recruited by whatever threatens, promises, hurts, rewards, repeats itself, or remains unresolved. By what happened that morning. By whether we are rested, by imbalances in our microbiome, by epigenetics, and by bodily conditions we may barely notice. It is recruited by our habits, our language, and our wounds. It is recruited by platforms that have learned what keeps us looking. It is recruited by institutions when they decide what will be measured, what will be presented in a meeting, and what will be classified as noise.
This does not mean that we cannot choose anything. It means that our choices never begin from zero.
When we believe we are deciding among several possibilities, a large part of the selection has already taken place. Some options have reached us. Others disappeared before deliberation even began.
The most powerful constraint is not always the option someone forbids us to choose. Often, it is the option we never get to see.
And we decide under time. We cannot wait until we understand the whole of reality because, while we are trying to understand it, reality is already changing. We have to act with whatever has managed to reach our attention, knowing that moments later new signals, new relationships, and new consequences will appear.
Intelligence does not remove that limitation. It works within it.
Attention Runs Through Intelligence
Attention does not replace perception, processing, action, and learning. It runs through all four.
It shapes what can enter perception. It gives more weight to some things than to others when we interpret a situation. It influences which actions appear possible. It also determines which consequences we register as learning and which we ignore because they do not fit the explanation we already have.
Reality does not enter only at the moment when we consciously pay attention. What we have lived through accumulates. Past experiences, decisions, and consequences alter what appears obvious, urgent, or invisible today. So does our need to remain consistent with the story we tell ourselves, even after that story has stopped being valid.
Intelligence is therefore not a clean chain in which we first perceive, then think, then act, and finally learn. It is a cycle, and sometimes something far less orderly. Part of reality manages to enter our attention. We interpret it and act. Our action produces consequences. Those consequences change the situation, and they also change what we may be able to perceive next.
Intelligence appears when that cycle manages to remain connected to what is actually happening.
The problem begins when the system stops using reality as its corrective and starts correcting only against itself.
A person may repeat an explanation because it fits their previous experience. An organisation may maintain a strategy because its internal indicators still appear reasonable. An artificial intelligence may produce a coherent answer because everything in its context points in the same direction. All three may reason correctly and still be wrong.
No intention to deceive is required. It is enough for the frame to have become insufficient while the system continues to operate as though it were still adequate.
We can give a good answer to a question that no longer matters.
Attention Within a Context
The word attention already has a technical meaning in artificial intelligence. Transformer models use attention mechanisms to assign different weights to different parts of the information they receive. Within a context, the model calculates which elements should have the greatest influence on its response. That ability is extraordinary, but it does not solve the whole problem.
Paying attention within a context is one thing. Knowing whether that context still represents reality is another.
A model can relate the information we have given it perfectly while failing to notice that the most important fact is missing. It can analyse outdated documents with great precision. It can construct a flawless argument from a badly framed question. It can cite sources that confirm one another because they all carry the same error.
AI is becoming increasingly good at attention within the context. The difficult problem is relevance in relation to a world that has moved.
It is not enough to process what is inside the context well. An intelligent system also needs some way to detect that something important has been left outside.
From the Prompt to the Context, and from the Context to the System
During the early years of generative AI, we talked a great deal about prompt engineering. The idea was simple: better instructions would produce better answers.
Then we began talking about context engineering. Writing a clever request was no longer enough. We had to construct the wider environment in which the model operated: documents, memory, examples, tools, rules, and information retrieved during execution.
That was progress. Context matters far more than a well-crafted sentence typed into a chat window. But it is still only part of the problem.
We can design the information a model receives extremely well while continuing to examine the wrong question. We can build a flawless memory from documents that contain only the dominant account of an organisation. We can introduce human controls while excluding the people who will bear the consequences. We can improve the answer without ever questioning the question.
The problem no longer lies only in the prompt or even in the context. It lies in the entire system that determines what information enters, who can challenge it, which actions appear possible, and which consequences are recorded.
We might call this work the engineering of intelligence.
But only if we are clear about what kind of engineering it would be.
The Engineer Is Inside Too
There is no engineer standing outside the system, able to see reality in full and arrange its parts from above.
The person designing the system also arrives with recruited attention, particular interests, a history, and a partial view of the situation. They work inside institutions that have already decided which goals matter, which resources are available, and which results will count as success. They may reinforce the very frame they intended to open.
Engineering intelligence does not mean controlling a system from the outside. It means intervening from within: changing what information enters, who is allowed to challenge it, which consequences are recorded, and when a decision must be reopened.
It is a situated and incomplete intervention, one that must be corrected by what it produces. The engineer is part of the experiment too.
Get Daniel Sujo’s stories in your inbox
Join Medium for free to get updates from this writer.
For that reason, the engineering of intelligence cannot promise a system that sees what everyone else misses. It must try to build systems capable of discovering the blindness of their own d esigners as well. That is not a minor distinction.
We Are Not Starting from Zero
Many parts of this problem have already been considered. Chris Argyris distinguished between correcting an action and revising the rules, goals, and assumptions from which we act. Biology and cognitive science have spent decades studying how organisms adjust perception and action in relation to their environments. Machine learning works with problems such as concept drift and out-of-distribution data precisely because a model may stop working when reality no longer resembles the data from which it learned.
I do not claim to have invented these pieces. The gap that interests me lies in connecting them. In treating the way a person, an organisation, and an artificial intelligence select reality, act from that selection, and detect, or fail to detect, that their frame has left something important outside as the same operational problem.
The contribution would not be a new metaphysics of intelligence. It would be bringing these ideas down to the level where processes, tools, meetings, metrics, responsibilities, and decisions are actually designed.
Attention Engineering
Within this wider framework, attention engineering deals with a fundamental part of the problem: what manages to enter the process and where our gaze is directed.
It is not about helping us concentrate harder or keeping us staring at a screen. The attention economy already does that with remarkable efficiency.
The attention economy learns what angers us, what frightens us, what we desire, and what keeps us engaged. It then returns increasingly precise versions of whatever has already managed to capture us.
It does not broaden our view. It narrows it. It shows us more of what already holds us, reinforces our habits, and turns our attention into increasingly predictable behaviour.
An attention engineering oriented toward intelligence would have to do almost the opposite. It would have to give something the current frame tends to exclude a real chance to modify that frame.
It might be a signal we are dismissing too quickly. A person affected by the decision who is not in the conversation. An assumption no one has reviewed. A metric that improves while the situation deteriorates. A consequence we leave out because someone else is paying for it. A possibility that our current categories do not even allow us to formulate.
This does not mean keeping every option open. That would be impossible and probably useless.
Attention always selects. Intelligence needs to simplify and close provisionally in order to act. A system that treated every signal as equally important would become paralysed. More information, more perspectives, and more options do not automatically produce better decisions. Sometimes they produce only. noise. The problem is not that we close. The problem is that we forget we have closed.
An intelligent system needs to stabilise an interpretation for long enough to act while preserving some way to detect when reality no longer fits inside it. It needs memory, but also mechanisms for recognising when that memory has become a burden. It needs goals, but also the ability to review what those goals have excluded.
This is not about preaching permanent openness or adding a brainstorming session at the end of every meeting. It is about designing concrete signals, thresholds, and mechanisms that allow the problem to be reopened when necessary.
The Human Is Not Outside Either
Many companies talk about keeping a human in the loop. The machine proposes and a person reviews. The machine executes and a person retains the final decision.
In some cases, this is necessary. But it can also create a false sense of security.
The human is not looking from outside either. They are also shaped by the available information, time pressure, company objectives, hierarchy, fatigue, and previous experience. They may correct the AI, but they may also confirm its error.
It may therefore be more useful to think in terms of a human in the field.
The person and the AI are part of the same environment. Both receive selected information and operate under constraints. Both can become trapped inside the same frame.
This does not make them equivalent. The person has a body, a history, responsibility, and direct exposure to consequences that the model does not experience. Responsibility cannot be dissolved by saying that “the system decided.” But reviewing an answer is not enough.
We also have to review the conditions that produced it. Who defined the problem? What information was available? Which metrics made some things appear important and others irrelevant? Who was able to challenge the frame? Who would pay the consequences if the decision were wrong?
Intelligence does not reside only in the person or only in the machine. It also lies in the quality of the system they form and in its ability to correct itself against reality.
A Claim That Can Fail
The phrase engineering intelligence could easily become another attractive label used to name consulting practices, organise conferences, and change nothing.
Naming that risk does not remove it. The concept deserves to survive only if it produces an observable difference.
The claim I am making is this: in changing environments, systems capable of detecting that their frame has become insufficient and of recovering excluded signals should adapt better than systems designed only to optimise answers within their initial context.
This does not mean that they should process more information all the time. Nor does it mean that more options will always produce better decisions. It means that they need mechanisms for discovering when the available information, questions, and options no longer represent the situation well enough.
We might observe this in the time an organisation takes to abandon an explanation that no longer works. In its ability to detect outdated assumptions. In how quickly it incorporates a signal that it had previously treated as an exception. In the number of times a real consequence forces it to change the frame rather than merely adjust the plan.
If systems designed in this way do not adapt better, the concept will have to be revised or abandoned. A theory of intelligence that cannot fail is not a theory. It is decoration.
This Essay Is Not Looking from Outside Either
This essay is also trying to recruit your attention. It has selected certain ideas, arranged certain examples, and left other possibilities outside in order to make you look at intelligence from a particular position. It is not neutral. No frame is.
The distinction I am trying to defend is not between influencing attention and leaving it untouched. That would be impossible. It is between recruiting attention to narrow the options and recruiting it to reveal that options may be missing.
But even that intention does not exempt this essay from the problem.
It may be constructing another frame that will eventually confuse itself with reality. It may persuade you because of its coherence, because of its language, or simply because it reached you at a moment when you were already prepared to believe it.
It must remain open to the same question it asks of the systems it criticises:
What did it have to leave out in order to appear complete?
The Frontier Is Not Only in the Model
I entered artificial intelligence trying to understand a technology and ended up confronting a problem about how people, machines, and organisations come to perceive what they consider real.
For years, we have treated AI as a race to improve processing: more data, more parameters, more reasoning ability, more tools, and more autonomy. All of that matters.
But a system capable of perfectly processing a thin representation of the world is not necessarily more intelligent. It may simply be more competent inside its enclosure.
The next frontier is not only about building models that answer better. It is about building systems capable of reviewing what has been presented to them as relevant, detecting when their view has narrowed, and reopening the problem before the error hardens into structure.
That requires working with models, but also with people, interfaces, memory, metrics, processes, responsibilities, and institutions.
The engineering of intelligence does not design an isolated artificial mind. It attempts to change the conditions under which a combination of natural and artificial intelligences perceives, acts, and learns from a reality that never arrives whole.
Before asking what decision a system should make, we may need to ask a more uncomfortable question:
What had to be left out for these to appear to be the only options?
