The Hidden Art of Breaking Problems into Learnable Pieces

Jul 01, 2026

10 min read

2 views

72%

The Real Problem Is Not Complexity, It Is Granularity

What if the biggest reason smart interventions fail is not that the problem is too hard, but that we are trying to learn it at the wrong size?

That question sits underneath everything from language models to philanthropy. When a model learns language, it does not memorize every whole word as an indivisible unit. It gets far better leverage by splitting words into smaller parts, so that stems can recur across contexts and affixes can be learned separately. The same basic logic applies far beyond text: once you break a system into reusable parts, you stop paying full price for every new example.

This is not just a technical trick. It is a theory of change.

In complex systems, the difference between failure and progress often comes down to whether you can find the right intermediate unit of action. Too large, and every case is a one off. Too small, and you lose meaning. The art is to find a representation that is simple enough to generalize, but rich enough to matter.

That is why some interventions move the world while others barely register. The strongest ones do not merely work on individual cases. They discover a reusable structure inside a messy domain and then alter that structure directly.

Why Systems Resist Direct Learning

We often imagine that if a problem is important enough, we should attack it head on. If poverty persists, send more money. If public health is failing, educate people. If institutions are broken, pass better rules. Yet complex systems rarely yield to blunt force, because their behavior is not determined by one variable. It emerges from interactions, incentives, habits, coordination failures, and feedback loops.

That is why the most effective effort is often not the most obvious one. A system can contain many concrete instances of suffering, but the leverage may lie in a smaller number of shared mechanisms. Change the mechanism, and you change the outcomes of many cases at once.

Think of grammar in a language. If you had to learn every sentence from scratch, language would be nearly impossible. Instead, you learn a stock of reusable pieces: roots, prefixes, suffixes, and patterns. Then you can generate and understand novel phrases. Complex social systems work similarly. They are not just piles of isolated events. They are structured with recurring parts, and progress depends on finding them.

This is the deeper attraction of modularity. It allows you to separate a complicated reality into components that can be understood, improved, and recombined. In language, those components are meaningful subword units. In social change, they might be institutions, incentives, technologies, or norms.

The highest form of leverage is not solving every instance separately, but redesigning the piece of reality that keeps generating the instances.

Once you see this, many failed efforts look less like moral failures and more like poor representation choices. They were aimed at the right pain, but at the wrong level of abstraction.

The Three Leverage Points Hidden in Any Messy System

A useful way to think about intervention is to ask: where can a system be decomposed into pieces that preserve meaning while increasing reuse?

I find it helpful to separate three leverage points.

1. Recurring elements

These are the parts that appear again and again across cases. In language, stems are a recurring element. In public problems, these might be administrative bottlenecks, information asymmetries, or incentive misalignments. If you can isolate them, one improvement can have wide reach.

Example: if many education failures trace back to unreadable application forms, then improving form design is more valuable than treating each applicant as a unique case. The issue is not the individual student, but the repeated friction embedded in the process.

2. Independent components

A good decomposition lets you learn one part without entangling it completely with the others. Affixes teach something about tense, plurality, or direction independent of the stem. Likewise, some interventions target a component that can be improved without redesigning the entire system.

Example: digital payment infrastructure can improve cash transfers, benefit delivery, and fee collection simultaneously. It does not solve every social problem, but it creates a cleaner channel through which many different programs can flow.

3. Generalizable representations

The most powerful decomposition is not merely local. It gives you a model that travels. Once learned, it transfers to novel examples without needing to start over each time.

Example: a policy that reduces bureaucratic discretion may improve outcomes in welfare, licensing, and procurement. The principle is portable because it addresses a general failure mode: too much hidden friction concentrated in too few hands.

These three leverage points form a practical test. If you are designing an intervention, ask whether it identifies recurring elements, isolates independent components, and builds a generalizable representation. If it does not, it may still help, but it is probably expensive in proportion to its impact.

Novelty Is Not the Opposite of Structure

At first glance, the two ideas here may seem to point in different directions. One is about linguistic decomposition, a technical method for handling variation. The other is about novel ways to change complex systems, a search for creative social interventions. But the deeper connection is that novelty becomes possible only when structure is exposed.

This is counterintuitive. People often imagine innovation as the opposite of pattern recognition, as if original action must be unstructured improvisation. In reality, truly new moves usually come from seeing the structure everyone else missed. The breakthrough is not pure invention from nowhere. It is a better segmentation of the problem.

Consider public health. A superficial response to a disease outbreak might focus only on the disease itself. A more structural response identifies the reusable machinery of spread: testing delays, reporting bottlenecks, crowded indoor environments, weak incentives to isolate, supply chain fragility. Each of these is a part that can be independently improved. The result is not one heroic fix, but a new representation of the system.

Or consider housing. If you treat every housing shortage as a unique local tragedy, you will be overwhelmed. But if you identify repeated causes, such as zoning constraints, permitting delays, and misaligned municipal incentives, you can target the mechanisms rather than the symptoms. A single structural improvement can shift thousands of outcomes.

The same idea applies to philanthropy itself. Many charitable efforts are optimized for visible relief, because visibility is emotionally satisfying. But if the real bottleneck is structural, then impact depends on whether you can change the rules, channels, or incentives that generate repeated need. The best philanthropy is often less like handing out answers and more like improving the language in which a problem can be solved.

A Mental Model: From Casework to Compression

Here is a useful mental model: good interventions compress complexity without destroying signal.

In information terms, compression is the art of finding a shorter description that preserves what matters. In language, tokenization or subword segmentation gives the model a more compact way to represent vocabulary. It does not ignore words. It discovers structure inside them. That same kind of compression is what makes strong social interventions possible.

Suppose you are trying to reduce dropout rates. You can study each student individually, which is compassionate but hard to scale. Or you can ask what repeated patterns are producing dropout across many students: transportation barriers, unstable schedules, lack of childcare, weak advising, opaque administrative requirements. Once those patterns are named, you can intervene at the level where many cases share the same root.

This is the difference between casework and compression.

Casework treats each occurrence as mostly unique.
Compression seeks the hidden regularities that make many occurrences similar.
Casework scales by labor.
Compression scales by structure.

That distinction explains why some reforms feel modest at first but compound dramatically over time. They are not dramatic because they are flashy. They are dramatic because they alter the compressed representation of the system. They make many future actions easier, cheaper, or more effective.

A transportation system that standardizes payment, routing, or scheduling does not merely improve one ride. It lowers the cognitive and logistical cost of the entire network. A bureaucracy that simplifies forms does not merely reduce annoyance. It changes the way the institution relates to reality.

The best reforms do not just do more. They make the system more learnable.

That phrase matters. A learnable system is one where improvement does not require rediscovering the whole world every time. It has stable units, legible boundaries, and reusable interfaces. It can absorb novelty without collapsing into chaos.

Why Complex Systems Need Better Tokens

There is a subtle but important implication here: changing a complex system often requires inventing better tokens, not just better actions.

A token, in the broadest sense, is a unit that stands in for something larger. In language, tokens help a model represent words and word parts efficiently. In social systems, tokens can be forms, categories, metrics, institutions, or workflows. They are the handles by which a system understands itself.

When a system uses bad tokens, it becomes blind in predictable ways. It may classify people too coarsely. It may measure the wrong outcomes. It may force diverse realities into a single box. That leads to brittle decisions, because the system is learning from a distorted map.

For example, if a social program categorizes applicants using an oversimplified eligibility rule, it may exclude people who most need support while including those who can navigate the paperwork best. The problem is not just inefficiency. It is misrepresentation. The system has chosen the wrong unit of understanding.

This is where novel institutional design matters. A new social institution can function like a better tokenization scheme. It creates categories, procedures, and interfaces that match the underlying reality more closely. It allows the system to distinguish what was previously blurred together, and to connect what was previously fragmented.

Technological breakthroughs can play the same role. A data infrastructure, payment rail, or verification tool can convert a messy field into something that can be coordinated at scale. Political change can do it too, by rewriting the incentives that define which parts of the system are visible and actionable.

So the question is not only, what should we do? It is also, what should count as a unit in this problem?

That is a profound question, because the unit you choose shapes the solution you can even imagine.

Key Takeaways

Look for repeated mechanisms, not just repeated symptoms. If many problems have the same underlying cause, one intervention may solve a whole class of cases.
Design interventions at the right level of granularity. Too broad, and the work becomes vague. Too narrow, and it cannot generalize. Aim for units that preserve meaning while enabling reuse.
Treat institutions and processes as representations. Forms, categories, metrics, and workflows are not neutral. They define how a system sees reality, and therefore how it acts on it.
Favor changes that make a system more learnable. The best reforms reduce friction, clarify boundaries, and improve transfer across contexts.
Ask what can be compressed without losing signal. If you can compress a problem into a simpler structure that still captures the important variation, you have found leverage.

The Most Valuable Innovation Is Often a Better Decomposition

We tend to celebrate the dramatic public win: the breakthrough drug, the sweeping law, the charismatic campaign. But underneath many of the best outcomes is a quieter achievement. Someone found a way to break a problem into pieces that could be learned, shared, and improved independently.

That is what makes tokenization such an elegant idea, and why it resonates far beyond language. It is not really about words. It is about the structure of understanding itself. The world is too complex to master as a single lump. Progress begins when we discover meaningful parts.

And in the realm of social change, that means the most ambitious interventions are not always the most sprawling. Sometimes they are the ones that identify a reusable bottleneck, introduce a better institutional grammar, and let change propagate from there.

So the next time a problem feels intractable, do not ask only how to push harder. Ask a more interesting question: What is the right way to split this problem so that learning can scale?

That shift in perspective can turn hopeless complexity into something with shape. And once a problem has shape, it can be redesigned.