The Iron Man Paradox

AI makes senior people superhuman. It may also slowly empty them.

The more senior you are, the more AI gives you.

That sounds unfair. It probably is.

We often describe AI as a great equalizer: everyone receives the same model, the same blinking cursor, the same cheerful offer to help you with anything. But equal access does not produce equal augmentation. Give the same telescope to two people and the astronomer will see more—not because the telescope behaves differently, but because the astronomer knows where to point it.

AI works the same way.

Its value is not simply added to what you know. It multiplies the number of connections you can make from what you know. The richer your internal map, the more places AI can take you.

This is why AI utilization is, at least initially, almost exponential with seniority.

And it is also why the people benefiting most from AI today may be the people most endangered by it tomorrow.

My brain is not a filing cabinet

I do not think in a straight line. I think in hyperlinks.

I see something in mathematics, something else in product design, a third thing in organizational theory, and—without asking anyone’s permission—they begin talking to each other. A physical law becomes a product principle. A science-fiction scene becomes an organizational metaphor. A paper written for one field suddenly illuminates a problem in another.

This kind of reasoning needs memory—not memory as storage, but memory as terrain. You need to have crossed enough landscapes to notice that two distant mountains have the same shape.

AI gives every branch of that hypertext a booster engine.

I begin with an intuition. It retrieves the neighboring concepts, the forgotten terminology, the counterexample, the paper I almost remember, the implementation pattern from a language I never learned. Every small branch can suddenly grow leaves, sub-branches and, occasionally, a mildly suspicious amount of fruit.

I am augmented because AI can bring content and memory to my thinking at the exact moment I need them.

But we need to distinguish two things that are increasingly confused: reasoning and memory.

Reasoning manipulates, tests and transforms. Memory supplies the pieces. Neither is sufficient alone.

A mind without knowledge will not reinvent centuries of mathematics from a blank page. A mind made only of remembered solutions will perform beautifully—right until it meets a genuinely new problem.

The relationship is subtle. Memory feeds reasoning. Then, if we are not careful, it begins to replace it.

The TypeScript test

I recently built sophisticated interfaces in TypeScript.

There is only one small issue: I do not know TypeScript.

I have never properly learned it. I have no particular desire to learn it. TypeScript and I maintain a courteous professional relationship mediated almost entirely by machines.

Yet I know what I want from a product. I have spent years thinking about product design, user centricity, innovation, interaction and the tiny details that separate an interface that technically works from one that feels inevitable. I can recognize the right output. I can critique it. I can direct iteration toward it.

Not long ago, realizing that vision would have required a team of engineers. Today, I can do much of it alone with AI.

This is an extraordinary gain.

But what exactly has improved?

The product has improved. My range of action has expanded. My output has multiplied.

My knowledge of TypeScript has not.

The result gains. The fuel does not.

That distinction matters. The machine temporarily completes my capabilities without necessarily changing me. It bridges the gap, but the bridge disappears when the model does.

We already know this phenomenon from phone numbers. We once remembered dozens; now we remember approximately three, including our own on a good day. The information did not vanish from the world. It moved outside us.

AI externalizes something more consequential than phone numbers. It externalizes parts of the intellectual journey that once changed the traveler.

Compressed memory, available on demand

An LLM gives us access to a compressed form of the world’s memory.

That is its magic. It is also its temptation.

At the right dose, memory feeds reasoning. At a higher dose, it can suffocate it. Paracetamol is useful until dosage turns medicine into poison. Intellectual memory has a similar, quieter curve: first it supports you, then it reassures you, then it becomes the only thing you trust.

You recognize the exercise, so you solve it. You recognize something adjacent to the exercise, so you adapt the solution. Then one day the page is truly blank. No familiar pattern arrives. You freeze—not necessarily because you lack the pieces, but because you have stopped trusting yourself to activate them.

AI accelerates this dynamic because the remembered answer always arrives before the discomfort has time to become reasoning.

The senior expert is spectacularly boosted at first. Every prompt lands on decades of accumulated context. Every response activates more analogies. Every analogy creates more output.

But if the expert only consumes the complement supplied by AI and never incorporates anything new, the curve eventually bends.

The growth is not exponential forever. It becomes concave.

Then it plateaus.

Then, perhaps, it declines.

You are still producing more than before. You may even appear superhuman. But the human underneath the suit is no longer being replenished.

The Iron Man problem

In the first Iron Man, Tony Stark’s arc reactor keeps him alive. It also poisons him.

That may be the best metaphor for senior people using AI today.

The machine becomes our source of energy. It lets us prototype, write, analyze, code, visualize and explore at a scale that would once have required an organization. We begin to resemble the one-person company of Silicon Valley mythology: one expert, an army of agents, no meetings, no coordination cost, no one changing the font in the deck five minutes before the presentation.

Glorious.

But a team is not merely a slow and expensive way to divide output.

This is what returned to me while thinking about Manuel Sosa’s work on teams. The best creative teams do more than aggregate skills. They expose each person to different memories, disciplines, intuitions and ways of seeing. Even an individual contributor becomes better through other people. A colleague does not merely complete a task you could not do. They bring a world you do not contain.

In a strong team of five to seven complementary senior people, each member expands the reasoning surface of the others. Contradiction creates movement. Misunderstanding forces precision. Someone notices the missing variable. Someone else asks the apparently naïve question that punctures three weeks of sophisticated nonsense.

AI can shrink a team from fifteen people to five or seven. That is probably excellent. It can return senior people to hands-on creation instead of making them full-time administrators of work they no longer touch. That may be one of AI’s greatest gifts.

But shrinking five to one changes the nature of the system.

At first, you remove bureaucracy.

Then you remove coordination.

Then you remove friction.

And eventually, without noticing, you remove surprise.

You become faster and more coherent—but coherent with yourself.

A mirror with excellent graphics

The more I work intensively with AI, the more clearly I see its current limit.

It is an extraordinary mirror. A mirror with compressed human culture behind it, excellent graphics, tireless syntax and occasionally alarming confidence—but still, often, a mirror.

I have long felt that my best interactions with AI were a dialogue with myself. The machine gives form to intuitions I have not articulated. It reveals relationships hidden in my own thoughts. It lets an older part of my experience speak to a newer problem.

That is immensely valuable.

But a mirror cannot refill the person standing in front of it.

After hours of intense work, something curious happens. I become tired. My ideas thin out. The prompts lose energy. And the LLM, which looked brilliant two hours earlier, begins returning flatter and flatter material.

The model has not become tired.

I have.

Yet the entire system has become tired, because I was supplying more of its intelligence than I realized.

At that moment the illusion breaks. If I have exhausted the source, the reflection cannot invent a new sun.

What should AGI actually give us?

Perhaps AGI is not the machine that finally allows one person to do the work of everyone.

Perhaps that is only automation wearing a cosmic cape.

A more interesting threshold would be a machine capable of playing the role of the team we are tempted to eliminate: not only extending our output, but replenishing our inner resources.

A machine that brings something genuinely foreign. That challenges rather than completes. That notices when our analogies have become circular. That teaches in a way that leaves knowledge behind. That protects productive struggle instead of erasing it. That knows when to answer—and when to make us think.

In other words, an intelligence that does not merely help us produce more, but helps us become more.

Until then, the central question is not whether AI makes us more capable. It clearly does.

The question is whether we are building a sustainable source of augmentation.

We need to keep learning what the machine lets us avoid learning. We need to preserve some blank pages long enough for reasoning to wake up. We need small teams of complementary people, even when one person plus agents appears more efficient. We need contact with minds that are not optimized reflections of our own.

The goal is not to reject the suit.

The suit is magnificent.

The goal is to replace the poisoned reactor before we forget there is still a human inside it.