← All writing Final · Sep 2026

More Better: Modeling and the Frame Problem

Every thought is a model, and every model leaves something out. Most failures come not from bad reasoning inside a model but from what never made it in: the frame problem. A story about my daughter's first car, clocks and clouds, and why being wrong is the light that leads to a life of discovery.

Introduction

People have a remarkable ability to build models. We look at a situation and produce a structure that represents it. Some of that structure we keep because we judge it important, and the rest we leave out, sometimes by choice and sometimes because it never crossed our minds.

Let me define terms. A model is a representational structure. Modeling is the process of creating these structures. Our senses turn light, sound, and touch into signals, and the brain represents those signals as information; it then operates on those informational structures, never on the world itself. That may seem unnecessarily technical, but it has profound implications for how we understand ourselves and our world.

Every thought is a model: a belief about a friend, a hunch about the weather, a theory of physics. They differ in scale and rigor, not in kind. The fields of study we build are models too, just larger, more deliberate, and shared:

Epistemology models knowledge itself: what we know, and how we come to know it.
Mathematics models objects through quantities and relationships.
Systems thinking models the world as stocks, flows, and feedback loops.
Forecasting models the future as probabilities and degrees of belief.
Game theory models decisions as players, values, actions, and beliefs.

What do we model? Information, because that is the only form in which reality reaches us. We never deal with the world directly; we deal with information and the models we build from it. I suspect information is also what reality is, an idea the physicist John Wheeler called “it from bit.” The Computation Conjecture argues a related, narrower claim: that computation reaches further than physicalism. But nothing in this essay depends on that stronger claim.

A model, then, compresses information. It keeps what we judge important and discards the rest. That compression is what makes models useful. It is also how they fail.

1. A Conversation About a Car

My daughter recently set out to buy her first car. I’m divorced, and she lives with her mother. She has worked hard and saved her money, and she wants to buy the car outright. She has watched relatives struggle with debt, and she doesn’t want to carry any.

I wanted to widen her options so she wouldn’t feel forced into a single choice. So I asked her to name her values. She said independence and security. From there we worked through the possibilities: what each option offered, what each one cost, and how each one served what she said she cared about.

We talked for an hour, and she eventually became upset. I was confused. I was giving her especially good guidance, after all.

Then she explained that her need was urgent. To respect her privacy I won’t share why. But the point is this: for that whole conversation I was doing excellent modeling, if I do say so myself, and I had missed an important consideration.

What bothers me most, in a useful way, is that nothing inside my model warned me. The values were right. The options were real. The reasoning held together. How good a model is on the inside tells you nothing about what it leaves out. The only signal came from outside the model, when my daughter got upset.

After I learned of the urgency our options began to shift, perhaps we take on certain risks because they may be better given the situation.

2. The Frame Problem

Philosophers, extending a problem from early artificial intelligence, call this the frame problem. In its original form (McCarthy and Hayes, 1969), it was a technical puzzle about how a logical system could represent what doesn’t change when an action occurs. Philosophers such as Daniel Dennett saw that it pointed to something much bigger: how does any reasoner decide which of the unlimited facts about a situation are relevant?

Dennett illustrated it with a robot that must retrieve its battery from a room where a bomb is set to go off. The first robot pulls out the wagon the battery sits on, but the bomb is on the wagon too. It knew the bomb was there; it just didn’t see the consequence. The next robot is built to consider the side effects of its actions, and it is still deducing that pulling the wagon won’t change the color of the room’s walls when the bomb goes off. The next one is built to ignore the irrelevant, and it sits outside the room busily filing implications as irrelevant, one after another, until the bomb goes off.

The robots can reason. What they can’t do is decide what belongs in the reasoning. Neither can we, reliably. And I believe this is where most of our failures happen. Not in the calculation. In the frame.

Statisticians have a version of this. Leave an important variable out of an analysis, and every conclusion can be biased, however carefully you calculate. They call it omitted-variable bias. Aubrey Clayton argues in Bernoulli’s Fallacy that probability is always conditional on the information you bring to it, and that methods that ignore background information produce conclusions that are confident and wrong. A model is only as good as the information it conditions on, and deciding which information to include is the frame problem again.

3. What Makes Something Relevant?

If the frame problem is about relevance, what decides relevance? I think two things: values and causality. Something is important if it can causally affect an outcome someone values.

My car conversation failed on both.

On causality: the urgency was a constraint that changed which options were even possible. It worked through timing, and timing wasn’t in my model.

On values: I asked for my daughter’s values, but the frame of the conversation (expand her options, don’t let her feel forced into one choice) came from mine. Whose values set the frame is itself a consideration, and it is one of the easiest to miss, because our own values don’t feel like assumptions. They feel like the way things are.

4. Clocks and Clouds

The frame problem would be manageable if the world outside our models were small. It isn’t.

Karl Popper, in his 1965 lecture “Of Clouds and Clocks,” contrasted systems that are regular and predictable, like clocks, with systems that are irregular and unpredictable, like clouds. Philip Tetlock picks up the image in Superforecasting:

“So is reality clocklike or cloud-like? Is the future predictable or not? These are false dichotomies…We live in a world of clocks and clouds and a vast jumble of other metaphors. Unpredictability and predictability coexist uneasily in the intricately interlocking systems that make up our bodies, our societies, and the cosmos.”

Every model is a clock we build and set down inside a cloud. The clock can be beautifully made. The cloud is still there, and it is vast. My conversation with my daughter was a small cloud. The systems Tetlock describes are enormous ones. The frame problem is the same at every scale.

5. The Rarer Failure: When Regularities Break

There is another way models fail, and it deserves mention, though I believe it is the less common one.

All models rest on regularities, which is a fancy way of saying we rely on experience. We expect gravity to work today as it did yesterday. We expect the sun to rise.

The philosopher David Hume showed that we cannot logically prove the future will resemble the past. This is now called the problem of induction. No matter how many white swans we see, we cannot justify the claim “All swans are white,” and a single black swan refutes it. Hume went further on causality: we never observe causation itself, only one event regularly following another. Our belief in cause and effect rests on the expectation that the future will be like the past. No one has solved the problem of induction. Popper and the Bayesians found ways to work around it, but not to solve it.

Sometimes regularities do break, and the result can be dramatic. Newtonian physics worked for more than two centuries before Einstein showed where it fails. But notice how this fits the frame problem. When a regularity breaks, something outside the model has changed, or was never what we assumed. A broken regularity is a consideration we didn’t know we were relying on.

Still, regularities are how sane people and societies think. The Newton-to-Einstein leap is rare. Far more often, we leave out an important consideration.

6. Every One of Us Is Wrong in Important Ways

This is not a problem for AI researchers or statisticians alone. It is the condition of every mind, in every field and every ordinary day. Every single one of us is wrong in important ways. We fail at several levels.

We miss what is right in front of us. In a well-known experiment (Simons and Chabris, 1999), people asked to count basketball passes in a video often fail to notice a person in a gorilla suit walking through the middle of the scene. Attention is itself a frame. Whatever it filters out never reaches the model.

We miss what game we are actually playing. This is the one that caught me. I was playing “help her choose well among options.” My daughter was playing “solve an urgent problem.” My modeling was excellent for the wrong game. Game theory’s first question is what game is this?, and getting it wrong makes every later calculation beside the point, however precise it is.

We misread the signals we do get. Daniel Kahneman called it WYSIATI: “what you see is all there is.” The mind builds the most coherent story it can from the information in front of it and doesn’t flag what is missing. Confirmation bias then shapes how new signals are read. When my daughter got upset, I could easily have read it as frustration with the options. It actually meant I was answering the wrong question.

Our fields have frames too. Every discipline is a trained way of deciding what counts as relevant. Economists see incentives. Doctors see pathology. Engineers see systems. That is the power of expertise, and it is also its blind spot. As the saying goes, if all you have is a hammer, everything looks like a nail. Expertise sharpens the model and narrows what can be seen from inside it.

Conclusion: Error Correction as a Way of Life

This could all sound discouraging. If every model leaves something out, and we can’t know what, why bother?

Because we have no choice. Modeling isn’t something we do; it is what we are. Every thought is a model. Refusing to model is itself a model, and usually a worse one, because no one examines it. The choice is not whether to model but whether to get better at it. I’ve written elsewhere about how to hold our models: to live above them, inhabiting them without being imprisoned by them.

This is what Popper most wanted to teach. His framework of conjecture and refutation starts from our fallibility, but it doesn’t end in a fatalistic posture. We make progress by understanding our nature more truly: we are fallible, and we can still get better. The posture that follows is to rely on regularities and make bold guesses beyond them, both held loosely, with a readiness to throw away positions that turn out to be invalid or unsound.

In practice, that means asking a few questions of any model, especially the ones we are proudest of:

What could causally change the outcome that isn’t in here?
Whose values set this frame?
What game am I actually playing, and what game are the others playing?
What would surprise me, and what would I do if it happened?

And it means treating surprise as information. Surprise is where the unmodeled breaks through. My daughter getting upset was not a failure of the conversation. It was the conversation correcting me.

We can be wrong, and that is more than okay. It is the light that leads to a beautiful life, one of discovery, surprise, and wonder.

Or, as my father would jokingly say, we can get “more better.”