What It Is
After three months away from work, Will could spend his time reconstructing every implementation detail or complete one task from Linear. Completing the task did not require a detailed understanding of the entire codebase. He could keep the rest in Linear and a braindump, then recover the details a particular task required. Trying to hold the whole system in working memory made starting harder without making that task more executable.
The useful level of detail depends on what you can change. A developer implementing a feature needs more detail than “build a working product,” but may gain nothing from optimizing assembly instructions. A surgeon needs tissue anatomy; a physicist working on quantum field theory needs a different account of the same physical world. Neither becomes more intelligent by applying the other's magnification to the wrong job.
Will uses resolution matching as a mental model for allocating thought: put detail where it gives you the ability to intervene. In computational terms, intelligence is adaptive compute allocation. The model has been useful in his gym routine, work reactivation and environmental changes. Its practical test is whether matching the detail of your thinking to the actions available to you helps you execute.
A closer look
Choose the useful scale
Navigation chooses a direction; execution needs the detail relevant to the intervention in front of you.
Read this diagram
Compare Where should I go?; What can I change here?.
The Core Insight: Resolution and Causal Power
A microscope reveals detail by giving up field of view. At a lower magnification, a larger pattern becomes visible, but small features disappear. More magnification is useful only when the detail changes what you can do.
The same constraint applies to reasoning. “Improve my health” leaves too much unspecified to execute, while a complete account of protein folding goes below the level at which you can consciously act. Food choices, exercise and sleep are available interventions. Understanding those behaviors deeply gives you something to change.
This model places consciousness at human scale because that is where its computation and capacity to act are suited. You cannot consciously direct protein folding, and an individual cannot personally direct civilization. You can intervene in the parts of a causal chain that your body, tools and responsibilities let you reach. Thinking above or below that level reduces the probability of execution.
Resolution Matching Examples
| Domain | Low Resolution (Macro) | Medium Resolution | High Resolution (Micro) | Where You Can Execute |
|---|---|---|---|---|
| Health | “Eat less, move more” | Macros and exercise type | Molecular biochemistry | Food choices and gym attendance |
| Programming | “Build working product” | Feature implementation | Assembly optimization | Tasks and features, for most developers |
| Learning | High-level concepts | Intermediate details | Formal proofs | Concepts that build usable intuition |
| Morning routine | “Get to work state” | Coffee, shower and desk | Neurochemical cascades | Triggers such as preset coffee and clothes |
The appropriate resolution is specific to the intervention. A food choice and an operation on tissue are both health interventions, but the people making them need different information.
Execution IS Causality
Understanding can extend beyond your ability to intervene. Execution happens where you actually change something.
An assembly expert may understand registers without shipping software. A Python developer who produces a working application has caused software to exist and users to receive its value. Measured by the change produced, the developer who ships has done more than the expert who ships nothing. The expert's additional knowledge has not provided additional causal power in that case.
A child who describes a game in English and uses AI to implement it also causes software to exist. The child works at the level of intent while the model handles implementation. The outcome is real, even though the child cannot explain the machinery at the same depth as an experienced programmer. Under this definition, that is programming: the person causes a computational outcome through the interface available to them.
The Intelligence Paradox
More intelligence can make more dimensions of a decision visible. Technical correctness becomes one consideration among ethics, long-term consequences, effects on a larger system, precedents, opportunity costs and questions about the decision process itself. A person whose cognition keeps hundreds of these considerations visible cannot simply stop seeing them. By the time the analysis feels complete, the opportunity may have passed.
A person considering fewer dimensions can decide whether an option will probably work and whether it has an obvious flaw, then act. In settings where repeated attempts matter more than a perfect first choice, the faster decision wins. The difficulty comes from resolution exceeding the decision's requirements, rather than from a lack of analytical ability.
| Cognitive Style | Dimensions Considered | Decision Speed | Execution Probability | When This Wins |
|---|---|---|---|---|
| High-resolution | 50+ factors, deep analysis | Days or weeks | Low when complexity prevents a decision | Precision matters more than speed |
| Matched-resolution | 5–10 key factors, with depth on each | Hours | High because the relevant decision becomes clear | Most contexts requiring action |
| Low-resolution | 2–3 heuristics | Minutes | High because a decision arrives quickly | Iteration speed matters most |
This is also a resource-allocation problem. Dividing 100 compute units across 50 dimensions leaves 2 units for each, which can be too little to reach a conclusion anywhere. Concentrating those resources on 5–10 relevant factors gives each 10–20 units. The second analysis covers less territory but goes far enough to support a decision.
Navigation vs Driving: The Intelligence Trap
Finding a route and traveling it require different work. Planning answers what would lead to an outcome; execution makes an intervention and encounters the result. Practice at planning does not transfer to practice at execution, so becoming a skilled planner does not train the ability to carry out the plan. A good plan can feel complete even when nothing outside the plan has changed.
Will calls these activities navigation and driving. In his effort model, navigation costs about 1–2 units, while driving costs about 4–6. Navigation ends when the causal map is satisfying. Driving ends when the destination is reached.
The proposed intelligence trap is a training imbalance. School rewards figuring out the path, and solving a problem provides a reward before anyone carries out the solution. By age 25, a person can have 20 years of navigation practice and roughly zero years of driving practice. Their skill at planning has developed while their skill at execution has atrophied. Recognizing the optimal path then makes failure to follow it especially frustrating.
The reward reinforces more planning. A person maps the whole causal graph, notices a flaw and revises it, or generates another objection to resolve. Those operations cost 1–2 units and feel productive while producing zero contact with reality. Satisfaction arrives when the path is clear. Comparing their actual behavior with an optimal execution adds resistance, but still does not make the first action happen.
The missing work has consequences beyond an unfinished task. The person has not used the intent-execution interface, paid the 4–6 unit activation cost, or received information from acting. Circuit formation requires temporal exposure; simulation does not supply those execution repetitions. In this model, intelligent people become optimized for the reward from navigation while a person with less elaborate cognition starts acting sooner.
Leveraging Intelligence for Causal Output
Analytical ability becomes useful when it addresses the conditions that keep requiring another difficult decision. Instead of repeatedly analyzing whether to work today and which task to choose, the analysis can identify an architectural change that makes work start automatically.
A forcing function changes which actions remain available. Prevention architecture changes the surroundings that cause a behavior, while triggers arrange for a context to load a routine. Designing these conditions directs navigation toward the system that produces actions. One intervention at a sufficiently influential point then replaces repeated interventions in individual actions. The surroundings continue to trigger the behavior, and the constraints continue to require it, indefinitely after the initial design.
Finding such a point does not require mapping every possible route. The useful question is where one cause has a disproportionate effect, so the analysis can concentrate on those nodes. This also gives someone who dislikes repetitive execution a reason to design a system that handles the repetition.
Execution can serve the analysis itself. A plan says that X causes Y, but only an intervention shows whether that connection holds outside the plan. Treating execution as an experiment makes failure informative and gives someone who values empiricism a reason to act. Experimentation also carries more status for that person than simply “doing stuff.” The action supplies data that planning cannot generate, allowing the next plan to improve.
AI can reduce the interval before that test. Navigation that takes 3 hours alone can take 10 minutes with AI. A timer bounds the planning period; when it ends, execution begins with the current map. An incomplete map tested against reality provides information that a perfect, untested map cannot.
AI Agents as Navigation-Driving Arbitrage
Agents let a person delegate execution while retaining responsibility for the causal model and the intended outcome. The bottleneck moves from the person's willingness to carry out each step to their ability to specify what should happen and design an agent system that works.
The person maps the causal structure, identifies leverage points and designs the agents' architecture. They must express intent precisely enough for an agent to execute. The agent handles the threshold of starting, repeats work without boredom, and gathers data through parallel actions. While a person has one body and must act sequentially, an agent swarm can drive 24/7 without willpower depletion.
The intelligent person's optimal allocation is to improve navigation and outsource all driving to agents. Precise intent and the design of interacting agent systems become the remaining navigation problems. The same planning ability that previously produced more plans than the person could execute now directs more execution.
Matching Resolution to Causal Affordances
The useful detail changes with the job. Most developers can intervene by completing a feature; implementation optimization matters when it affects that outcome. Someone changing a habit can arrange triggers and surroundings, but cannot directly rewire a neuron. A learner needs a usable pattern before every formal foundation is relevant.
Behavioral debugging also requires choosing something that can be changed. “I'm lazy” provides no specific intervention. “The work-starting routine didn't load” directs attention to the missing trigger or the competing routine. The system description becomes useful because it identifies something the person can alter.
This is an extension of pedagogical magnification from understanding to execution. Available compute limits how much detail you can process, and causal affordances limit where that detail can produce a change. Optimal foraging adds the resource question: spending 100 compute units on a level where you have zero causal power cannot repay the analysis through an intervention.
Pedagogical Implications
A student can use Python before learning assembly, calculus before real analysis, and applied reasoning before formal logic. The useful result gives them a reason to care about the machinery underneath it. Starting with prerequisites can demand depth before the student has that context.
In this model, intuition and motivation begin at the macroscopic level. A child can understand that gravity bends space and time dilates before studying tensor calculus. Describing a program's intended behavior can precede understanding memory management. The claim that a child cannot learn a subject can therefore be a failure to find a suitable resolution. At the right magnification, any concept becomes accessible.
AI acts as a resolution adapter when a human describes an outcome in English and the model produces code, syntax and implementation details. The human achieves the intended outcome without personally operating at implementation resolution. Calling this unreal programming confuses the level of the interface with the reality of the result.
Observable Patterns from N=1
Pattern 1: Gym Execution (Day 16/30)
Will concentrated on attending the gym and completing the workout instead of optimizing protein synthesis or studying muscle biochemistry. His available intervention was physical behavior; he could not consciously control protein folding. In this experiment, the activation cost fell from 6 units to 0.5 units by Day 16 of 30.
Pattern 2: Work Reactivation After 3-Month Dormancy
Completing one Linear task let Will resume work without first reconstructing every implementation detail. The task gave him a bounded part of the codebase to understand and a completion he could produce. Linear and the braindump held the wider context because working memory could not hold the full system at detailed resolution. The wiki's working-memory range is 4–7 items.
Pattern 3: Information Diet Architecture
Will changed access to information instead of repeatedly deciding whether to resist it. He kept zero apps installed, left the phone off by default and checked intentionally when needed. His comparison assigns environmental prevention a one-time setup cost of 0.5 units, against 3 units for each act of resistance. Once set up, prevention costs roughly zero per event because the trigger no longer has to be resisted.
Pattern 4: Guitar vs Lounge-Scrolling
The lounge already cued a default activity. Replacing scrolling with guitar practice changed what that state led to, so each evening no longer began with the same contest against the phone. In Will's accounting, changing the default costs about 0.5 units, compared with about 3 units for each act of resistance. The intervention acts on the transition into the routine rather than on a routine already running.
Practical Applications
Application 1: Debugging Effort Allocation
A career decision can expand to 50 factors, including market trends, ten-year projections and hypothetical opportunity costs. More factors do not help if you cannot get information about them or use the information to choose. A narrower comparison can examine five factors you can assess and act on: skills match, team quality, learning opportunity, compensation and location.
The diagnostic questions are:
- What resolution am I operating at, and how many dimensions am I considering?
- Where can I intervene?
- Am I thinking more deeply or broadly than that intervention requires?
A useful factor supports real information gathering, a meaningful assessment and an actual decision. Those requirements determine what belongs in the analysis.
Application 2: Task Execution Protocol
“Finish the project” does not specify an action. “Optimize every line of code” specifies so much work that completing something useful can disappear from view. “Complete the user authentication feature” provides both a starting point and an outcome someone can use.
1. Check resolution:
- "Finish project" = too high-level (can't execute on abstraction)
- "Optimize every line of code" = too low-level (lose the forest)
2. Match to execution resolution:
- "Complete user authentication feature" = just right
- Specific enough to start
- Broad enough to ship value
3. Execute at matched resolution:
- Don't zoom into micro-optimization (wrong resolution)
- Don't stay at "work on project" (wrong resolution)
- Stay at feature/task level (where you can ship)
Spending three weeks optimizing a function that does not need optimization fails the same test: the additional detail does not serve the required outcome. The choice of resolution depends on available compute, the interventions available and the result the task requires.
Application 3: Learning Strategy
Learning a framework can begin with what it accomplishes, followed by a simple working example. Using the example reveals the patterns and gives bugs somewhere specific to occur. Internal mechanisms become relevant when curiosity or a problem requires them. Reading the entire source before writing a line reverses that sequence and requires detailed understanding before functional capability.
High-level work can be demanding and effective. A Python developer shipping a product is working at the resolution their intervention requires. Deeper understanding is useful when it serves that intervention; for someone whose health choices are limited to eating and exercise, molecular biochemistry adds computational cost without causal benefit. Understanding what triggers overeating or enables regular exercise addresses the available controls.
Related Concepts
- Pedagogical magnification explains how to choose the detail that gives a learner useful understanding.
- Willpower supplies the finite-resource accounting used for analysis and action here.
- Optimal foraging theory examines where an expenditure of resources can repay itself.
- Agency concerns the ability to act through available interfaces.
- AI as accelerator develops the relationship between faster analysis and execution.
- Programming as causal graph construction describes how execution changes causal flow.
- State machines describes the routines and transitions used in the guitar example.
- Prevention architecture develops environmental changes that reduce repeated resistance.
- The 30x30 pattern records the behavioral repetition model behind the gym example.
- Working memory explains the capacity constraint on holding a detailed system in mind.
- The braindump provides a way to keep information available outside working memory.