The Seduction of One Answer

Why the human mind compresses complexity — and what happens when causality refuses to cooperate

Over the last several weeks, several VCs have asked me some version of the same question: “Mel, what is the one thing?” Does Chateauz save time? Increase efficiency? Reduce risk? Improve capital allocation?

It was a useful exercise. Every company should be able to articulate what makes it valuable. But the question stayed with me for another reason: I started noticing how often we ask some version of it everywhere else.

Scroll through almost any conversation about AI and energy infrastructure and you will find hundreds of variations of the same argument. AI infrastructure is a power problem. No, it is a grid problem. Actually, it is transmission. Or water, cooling, permitting, compute, capital.

There is something deeply satisfying about identifying the constraint. It takes a complicated system and compresses it into something we can understand, communicate and act upon.

But what if causality isn’t that simple? What if there isn’t one thing?

The human mind compresses

Human beings cannot reason over every piece of available information simultaneously, so we build reduced representations of reality. We categorize, abstract, prioritize, recognize patterns and use heuristics. We turn hundreds of observations into concepts that can be manipulated as one thing.

The limits of working memory are one illustration of this broader constraint. Nelson Cowan’s research places the central capacity of working memory at roughly three to five chunks under controlled conditions — considerably less than the famous “seven, plus or minus two.”¹ 

This does not explain why someone reduces AI infrastructure to a power problem. It does, however, illustrate something fundamental: human cognition operates with finite capacity in environments containing vastly more information than we can consciously manipulate at once.

So we compress, and that is not a defect. Compression is part of what makes intelligence possible. Without abstraction, every situation would have to be understood again from first principles.

The problem begins when useful simplification becomes oversimplification — when the representation becomes simpler than the causal structure of the system it is supposed to represent.

The seduction of simplicity

We are also intellectually trained to value simpler explanations. Occam’s Razor is commonly reduced to the idea that “the simplest explanation is probably correct,” but the philosophical principle is more careful: other things being equal, there is reason to prefer the simpler explanation.²

That qualifier does a tremendous amount of work. Occam’s Razor tells us not to introduce complexity unnecessarily. It does not tell us to remove complexity that is actually doing causal work.

There is a profound difference between simplifying a system while preserving what matters and simplifying a system until what matters disappears.

That distinction becomes particularly important in coupled systems, because the behavior we care about may not belong to one variable in isolation. Sometimes it emerges from the relationship between variables.

Complexity can live between the variables

Imagine a decision involving ten variables, each with only three possible states: low, expected and high.

Even that toy example produces 3¹⁰ = 59,049 possible configurations.

At twenty variables, 3²⁰ = 3,486,784,401.

The point is not that every variable matters equally; often only a few dominate. The more interesting problem begins when which variables dominate changes with the state of the system.

Real infrastructure systems are full of these interactions. Power affects cooling. Compute density affects power. Cooling architecture affects water. Water affects siting. Siting affects transmission. Transmission affects schedule. Schedule affects capital. Capital can change architecture, which changes equipment requirements, which can change schedule again.

If the world behaved additively, we might imagine an outcome as:

Y = β₁X₁ + β₂X₂

But coupled systems can contain interaction terms:

Y = β₁X₁ + β₂X₂ + β₁₂(X₁ × X₂) + …

That small X₁ × X₂ changes the meaning of the model. It says that the effect of X₁ can depend on the state of X₂. Sometimes the information that matters is not contained in either variable independently. It is contained in the relationship between them.

We can take this one step further.

The sensitivity of an outcome to one variable may itself depend on the rest of the system:

∂Y/∂X₁ = f(X₂, X₃, …)

In plain English: how much one variable matters can change depending on the conditions around it.

That, to me, is where the “one thing” explanation becomes particularly seductive.

Consider the statement: “Power is the biggest constraint on AI infrastructure.” It might be true — under some conditions, at some sites, at some point in time. But perhaps the more useful question is: Under what conditions does power become the binding constraint, and what becomes binding once power is solved?

Add generation and transmission may become limiting. Solve transmission and permitting may dominate. Change the power architecture and equipment availability may become critical. Increase compute density and thermal constraints change. Accelerate construction and supply-chain constraints surface.

Removing one constraint does not necessarily remove the complexity. It can change where the constraint appears.

That is much harder to turn into a headline. “AI has a power problem” fits neatly into a post. “Power becomes dominant under a particular configuration of compute, transmission, cooling, siting, permitting, equipment availability and capital assumptions” does not.

So we compress again.

How much complexity is enough?

This is where W. Ross Ashby’s Law of Requisite Variety becomes interesting. Ashby was not arguing that more complexity is always better. His point was more precise: a regulator needs sufficient variety, appropriately matched to the variety it must regulate.³

The word requisite matters.

Humans need compression because we cannot carry the full variety of a complex environment in our heads. Yet a model or decision process that removes too much variety may no longer be capable of representing the conditions relevant to the outcome.

So perhaps the choice is not simplicity versus complexity at all.

The goal is requisite complexity: no more complexity than the decision requires, but no less.

That leads to a more practical way of answering the question I started with: How simple is too simple?

A model has been simplified too far when the information removed can change the decision.

If removing a variable, interaction, dependency or plausible condition does not change which option we choose across the range of circumstances we care about, perhaps that information can safely remain compressed. But if restoring it changes the preferred option, moves the binding constraint, reverses the risk ranking or materially changes the outcome, then we threw away something decision-relevant.

This idea sits close to established work in decision theory around Value of Information and sensitivity analysis: information matters to a decision insofar as knowing it can improve the quality of the choice we make.⁴ I am interested in the same question from the other direction — not only what information is worth acquiring?, but what information can we safely remove without changing the decision?

That is a very different standard for simplification.

Not: Does this model contain reality? No model does.

Not: Can we explain the system with fewer variables? Often we can.

But rather: Does the simplification preserve the decision?

Perhaps “it depends” deserves more respect

This is why I have started thinking differently about the phrase “it depends.” We often hear it as evasive, and sometimes it is. But in a genuinely complex system, it depends may be the beginning of rigor rather than the absence of it.

Depends on what? Which variables? Which interactions? Under what conditions? At what threshold? What changes the outcome? What does not?

Those questions move us away from reflexively searching for one universal cause and toward understanding the structure of the decision itself.

The systems we are increasingly responsible for designing, financing and operating are becoming more interconnected than our ability to mentally hold all of their interactions at once. Our instinct — cognitively, culturally and professionally — is therefore to compress them into something manageable.

Sometimes that compression produces genuine insight. Sometimes it produces a beautifully simple explanation of a system that does not actually exist.

We need simplicity in order to think. We need sufficient complexity in order to be right.

Perhaps the line between the two is this:

We can afford to throw away everything that doesn’t change the decision. The hard part is discovering what does.

---

Footnotes

1. Nelson Cowan, “The Magical Number 4 in Short-Term Memory: A Reconsideration of Mental Storage Capacity,” Behavioral and Brain Sciences 24, no. 1 (2001): 87–114. See also Cowan, “The Magical Mystery Four: How Is Working Memory Capacity Limited, and Why?,” Current Directions in Psychological Science 19, no. 1 (2010): 51–57. ↩

2. Alan Baker, “Simplicity,” The Stanford Encyclopedia of Philosophy, Winter 2022 Edition. ↩

3. W. Ross Ashby, An Introduction to Cybernetics (1956), particularly the treatment of requisite variety and regulation. ↩

4. The framing is related to established Value of Information analysis in decision theory. See Ronald A. Howard, “Information Value Theory,” IEEE Transactions on Systems Science and Cybernetics 2, no. 1 (1966): 22–26; and Howard Raiffa and Robert Schlaifer, Applied Statistical Decision Theory (1961). ↩