Casey Barr

Mathematics, Econometrics, and Applied Topology

When Thinking Becomes Too Cheap

Lately, I’ve noticed myself trying to optimize experiences that are still novel to me.

Whether I’m getting into distance running, researching a purchase, planning a trip, or trying to learn a new skill, I increasingly find myself trying to identify the optimal path before taking the first step. I want to know which approach is best, what mistakes people usually make, and how I should sequence the work. I want the first attempt to be informed by the lessons of attempts that have not happened yet.

At first, this felt like overthinking. Maybe I was turning simple decisions into elaborate ones. Maybe I was using research as a more respectable form of procrastination.

But eventually I realized I wasn’t simply overthinking. I had internalized a new decision-making heuristic: before acting, try to find the best way to act.

Why this changed

The immediate explanation is that another round of analysis has become unusually cheap. Large language models let me brainstorm approaches, pressure-test assumptions, generate counterarguments, and get a rough map of an unfamiliar subject in minutes. They do not resolve every question, but they make it easy to keep asking questions before committing to an answer.

Historically, another round of analysis cost time, expertise, or social capital. Now it often costs a prompt.

When an input becomes cheaper, people tend to use more of it. Cheaper storage leads us to store more data; cheaper computation leads us to compute more things. Thought seems to follow the same pattern. I can keep a question open while I generate another version of the answer, another objection, or another plan.

That is often useful. It is easier to notice a missing assumption when someone—or something—can point to it immediately. It is easier to compare approaches before I become attached to one. But cheap analysis also changes the threshold at which I feel ready to act. A rough first attempt can start to feel irresponsible when a more polished plan is only one prompt away.

The convenience also changes the shape of uncertainty. Previously, uncertainty often forced a choice because resolving it required a book, a conversation, or a costly detour. Now I can label the uncertainty, ask for competing interpretations, and continue without deciding. The question remains intellectually active even when it is no longer helping me move.

Running exposes the limit

I noticed this most clearly when I decided to get into distance running.

Before I had established any running habit, I found myself thinking about the plan. Should I run three days a week or four? Treadmill or outside? Distance or time? How quickly should I increase my mileage? When should I introduce speed work?

These were reasonable questions, but they were not the bottleneck.

The important unknowns concerned my body, habits, and preferences. What pace feels sustainable? Do I enjoy running outside? Do mornings fit my schedule? How does my body respond to three easy runs in a week? What limits me first: my lungs, my legs, my attention, or my willingness to leave the house? Even a technically sound training plan would have to be adjusted around answers I did not yet have.

I could read about those questions, but I could not answer them in advance. The information did not yet exist. I had to run and find out.

I was trying to optimize away uncertainty before I had generated any experience. Running was an exploration problem disguised as an optimization problem.

That distinction applies beyond running. Some problems reward thinking before acting. In software architecture, system design, product strategy, hiring, investing, and long-term planning, careful reasoning can prevent expensive mistakes. Other problems contain information that only appears through interaction with reality. Cooking and public speaking have this property, as do most physical skills. The first attempt is not merely the execution of a plan. It is a way of measuring the problem itself.

In those cases, the plan is useful mainly because it helps you begin collecting information. Its value is not that it predicts the experience perfectly. Its value is that it makes the first experiment safe and concrete enough to teach you something.

No training plan could tell me whether a particular ache was ordinary fatigue, bad form, or a reason to stop. No schedule could tell me whether I would keep choosing a run after a long day. Those answers required a body moving through time, not a better description of the activity.

Two kinds of iteration

This is the distinction I keep coming back to: not all iteration is the same.

Avoidable iteration comes from preventable mistakes. You misunderstood the problem, failed to check an assumption, chose a poor process, or ignored information that was available beforehand. Better reasoning can reduce this kind of iteration. A conversation with an LLM can be useful here: it can expose an assumption, point out an edge case, or suggest a better starting plan before the cost of acting has accumulated.

Necessary iteration is required because the missing information can only be created through experience. You can read about running form, but you cannot know your sustainable pace until you run. You can study recipes, but you cannot fully understand how a dish tastes in your kitchen until you make it. You can prepare a speech, but you cannot know which parts lose the audience until people hear it.

The second kind is necessary because the information did not exist beforehand.

This is not a failure of planning. It is a property of the problem. LLMs are useful at reducing avoidable iteration, but they cannot eliminate necessary iteration, because there is nothing to eliminate. The missing information has to be created by an event: a run, a meal, a conversation, a presentation, or some other encounter with conditions that cannot be simulated completely from a distance.

Confusing the two creates a strange ambition: the desire to plan so well that experience becomes unnecessary. But in an exploration problem, experience is not what happens after the real work. Experience is how the work learns what it is.

An old pattern

The larger pattern is older than LLMs. Cognitive technologies change what must be done internally and move the bottleneck elsewhere.

Writing externalized parts of memory. Calculators externalized routine arithmetic. Search engines externalized information retrieval. LLMs externalize parts of preliminary reasoning: generating possibilities, organizing arguments, and exposing gaps before a plan encounters the world.

Plato’s Phaedrus contains an early version of the anxiety these tools provoke. Socrates tells a story in which the Egyptian god Theuth presents writing as a remedy for memory and wisdom. King Thamus objects that writing will produce forgetfulness, because people will rely on marks outside themselves rather than developing their own memories. Written words, he suggests, will give people the appearance of knowledge without knowledge itself.

The useful point is not whether Socrates was right. It is that a new tool changes the balance between internal effort and external support, and therefore changes what people spend their attention on. The history of cognitive technology is a history of moving bottlenecks. When memory becomes easier to externalize, retrieval matters more. When arithmetic becomes easier to automate, choosing what to calculate matters more. When preliminary reasoning becomes cheap, deciding when reasoning has done enough may matter more.

Cheap thought needs new stopping rules

Abundance creates its own distortions. When something is expensive, we ration it. When it becomes cheap, we may overuse it before learning where it helps.

Cheap reasoning can produce more options than we can evaluate, more objections than we need, and more refinement after a decision is already actionable. If another plan is only seconds away, why not ask for it? If the answer can be improved one more time, why stop now? The possibility of more analysis can make the current amount feel inadequate even when it is sufficient for the next step.

The result can look like rigor while functioning as avoidance. Not because I am necessarily afraid to act, but because the economics make further thought feel unusually responsible. It is hard to notice that analysis has stopped changing the decision when each additional pass still produces something new.

A stopping rule designed for expensive analysis may behave badly when analysis becomes abundant. The challenge is to notice when more reasoning is still reducing avoidable error and when it is only postponing the experiment that would produce the missing information.

The scarce input in running was not another training plan. It was firsthand knowledge of my body and my relationship to the activity. The next useful step was not another optimization pass. It was a run.

That will not be true everywhere. Some decisions deserve extensive analysis; some mistakes are too costly to discover through casual experimentation. But the distinction gives me a better question than “Have I researched enough?” It asks whether the next useful input can be generated by more thought or only by contact with the world.

Which problems are waiting for a better plan, and which are waiting for contact with the world?

2026 1
2025 1
2024 1
2021 2
2019 5

2026