Why Humans Need Better Ways to Break Things Into Pieces

The hidden danger in how we slice the world

What if one of the most important differences between wisdom and catastrophe is not what we see, but how we segment what we see?

That sounds abstract until you notice a pattern. A crowd does not simply see a complex social situation and react to it as a whole. It breaks the situation into chunks. One person becomes the problem. One story becomes the explanation. One desire becomes the obvious desire. Once the world has been divided this way, people stop seeing nuance and start copying each other’s attention. Conflict intensifies, and the crowd becomes a machine.

Now consider a seemingly unrelated idea from machine learning: if you split words into stems and affixes, a model can learn more efficiently. It sees repeated pieces across many examples. It no longer has to memorize each word as an indivisible blob. A language model works better when it can generalize from reusable parts.

These two ideas meet at a surprising place. Human beings and machines both depend on tokenization, the act of breaking reality into pieces. But the quality of the pieces matters. Break things badly, and you get myths, mobs, and mistaken certainty. Break things well, and you get learning, pattern recognition, and better judgment.

The deep question is not whether we should simplify. We always simplify. The question is: what kind of simplification creates understanding, and what kind creates violence?

---

Tokenization is not just a technical trick

In language models, tokenization is the first act of intelligence. Before a model can predict anything, it has to decide what counts as a unit. Should “unhappiness” be one token, or three parts, or something in between? The answer affects how much the model can reuse what it knows. A useful token scheme does not merely compress text. It reveals structure.

That is true of people as well. We never encounter the world in raw form. We divide it into categories: friend, rival, stranger, threat, opportunity, insult, coincidence. We do this because it is cognitively efficient. A child seeing a room full of toys does not analyze each object from scratch. Attention latches onto one object because someone else has already treated it as meaningful.

This is the beginning of mimetic desire, the imitation of desire itself. We do not just want things because they are useful. We want them because they are wanted. The object becomes a token charged by social attention. In a sense, desire is a kind of language model: it predicts value from patterns of other people’s valuing.

That is why social life can escalate so quickly. Once one person’s preference becomes visible, others copy it. Then the copied desire makes the object feel even more important. Soon the object is not important in itself, but as a node in a network of rivalry. The tokens begin to feed on one another.

Human beings do not merely think in categories. They fight over the categories they create.

---

When simplification becomes scapegoating

There is a form of tokenization that helps us generalize. And there is a form that helps us misfire.

The useful kind separates a word into reusable pieces. The dangerous kind separates a complex social crisis into one guilty person. A community in distress rarely experiences its crisis as a system. It feels the friction of too many desires, too many comparisons, too much status anxiety, too much uncertainty. That diffuse instability is hard to hold in mind. So the community looks for a cleaner unit, a single explanation that can be named, pointed to, and punished.

This is the logic of scapegoating. The crowd takes an overlapping, chaotic field of conflict and converts it into a single token: the victim. The victim becomes the container for everything no one wants to understand. Resentment, fear, envy, humiliation, and confusion all get attached to one figure. Once that happens, social attention becomes startlingly coherent. Everyone agrees, not because they are right, but because they are synchronized.

That synchronization can feel like moral clarity. It often does. But it is really a failed compression algorithm. A poor tokenization has made the world seem simpler than it is.

This is why scapegoating is so seductive. It converts an overwhelming system into a legible story. It is easier to say “that person caused this” than to say “our desires have become imitative, our institutions have weakened, and our conflict is feeding on itself.” The first explanation is emotionally efficient. The second is cognitively honest.

The tragedy is that a crowd often experiences the punishment of the victim as relief. Conflict seems to vanish. People suddenly feel united. The tokenization has “worked” in the short term. But the result is counterfeit order, because the underlying dynamics have not been understood, only displaced.

---

The model and the myth share a hidden structure

Why compare a machine learning technique to a religious and social theory at all?

Because both are about representations. A model learns better when it can identify parts that repeat across contexts. A culture learns better when it can identify patterns in conflict without mistaking one person for the whole pattern. In both cases, the task is to move from surface appearances to reusable structure.

Myths are especially revealing here. A myth often preserves the memory of a social crisis, but in distorted form. The group once felt torn apart, then rallied around a victim, then experienced peace. Over time, the victim becomes sacred, monstrous, or both. The story encodes the event, but it does not explain it. It turns a crisis in interpretation into a narrative of necessity.

That is why myths can resemble truthful accounts while still being fundamentally misleading. They preserve the outline of a process without naming its mechanism. They are like a model that fits the training data while learning the wrong feature. The model outputs something plausible, but not because it understands the world. It has merely found a convenient compression.

The Bible, in this framing, is not just another myth. It is a counter tokenization. It repeatedly shifts attention away from the crowd’s accusation and toward the victim’s innocence. Instead of reinforcing the sacrifice, it exposes the machinery behind it. Where myth says, “The victim is the cause,” this alternative reading says, “The victim is where the crowd hid its own conflict.”

That is a radically different way to break up reality. It changes which parts are reusable. It teaches readers to see not only the event, but the structure that generated the event.

---

The real contest is over the units of thought

Every culture, institution, and technology competes to define what the basic units are.

A political movement may define society in terms of classes, races, tribes, or insiders and outsiders. A workplace may define value in terms of teams, metrics, or personalities. A platform may define attention in terms of clicks, likes, and shares. Each system is a tokenization scheme. Each one says, implicitly, “these are the pieces that matter.”

This is powerful because units determine what can be learned. If you tokenize a conflict incorrectly, you will train yourself on the wrong lesson. If you tokenize productivity incorrectly, you will optimize the wrong behavior. If you tokenize identity incorrectly, you will turn fluid people into rigid symbols.

Consider two managers looking at the same underperforming team. One manager tokenizes the problem as “lazy employees.” The other tokenizes it as “unclear incentives, misaligned goals, and status competition.” The first explanation is simpler. It feels cleaner. It also invites scapegoating. The second is harder, but it opens real interventions.

Or consider social media. It rewards the fastest tokenization of a person into a brand, a villain, a hero, or a punchline. Nuance is expensive. Tribal labels are cheap. The platform therefore selects for the kind of cognition that can turn complex humans into efficient symbols. Once that happens, the crowd’s mimetic energy has something easy to copy.

This is not merely a moral problem. It is an epistemic one. Bad tokens produce bad models. Bad models produce bad action. And bad action, when synchronized across a group, can become a disaster.

Most human violence begins as a classification error.

---

How to tokenize better: a practical framework

If we are always breaking reality into pieces, then the real skill is not resisting simplification. It is learning better simplification.

Here is a useful three part test for any unit of analysis:

1. Does this category repeat across contexts?
   Good tokens travel. A useful concept helps explain many cases without forcing them to be identical.

2. Does this category reveal structure, or hide it?
   If a label only names the loudest thing in the room, it may be a disguise. Better tokens point to mechanisms, not just outcomes.

3. Does this category reduce blame or concentrate it?
   If a classification makes it easier to understand a system, it probably helps. If it makes one person carry the weight of a whole collective failure, beware.

This framework applies far beyond social conflict. In science, better models tokenize phenomena at the right level of abstraction. In business, good strategy identifies the reusable patterns behind isolated failures. In personal life, mature judgment separates a person from the role they played in one conflict.

The deeper skill is to ask: what am I turning into a token, and what am I erasing by doing so?

A person who can answer that question is harder to manipulate, because they can see when a crowd is trying to hand them a prepackaged explanation. They can also see when their own mind is doing the same thing, converting discomfort into certainty too quickly.

---

Key Takeaways

• Notice your tokens. When you feel certain about a person, group, or event, ask what unit of meaning you have created. Is it a useful abstraction or a shortcut to blame?
• Prefer mechanism over label. Replace statements like “they are the problem” with questions about incentives, imitation, status, and feedback loops.
• Watch for mimetic escalation. If desire or outrage is spreading through comparison, not genuine independent judgment, the situation is becoming self-reinforcing.
• Treat consensus with suspicion when it arrives too fast. Instant agreement can signal clarity, but it can also signal that a crowd has found a convenient victim.
• Train yourself to resegment. When a situation feels impossible, try breaking it into smaller, more informative parts. The right pieces often reveal a solution hidden by the wrong whole.

---

The moral of the story: the world punishes sloppy abstraction

The same human capacity that lets us learn language, build civilization, and generalize from experience can also lead us to persecute the innocent. The difference is not whether we simplify. It is whether our simplification respects reality.

A good model breaks the world into parts so it can understand the whole. A bad crowd breaks the world into parts so it can stop understanding anything at all. One form of tokenization creates learning. The other creates a sacrificial story that feels true because it is emotionally efficient.

That is the unsettling connection between a language model and a mob. Both are systems of prediction. Both depend on the units they choose. And both can go badly wrong when they learn the easiest pattern instead of the deepest one.

The most important question, then, is not only what we want, but how we are parsing desire, conflict, and blame in the first place. Because once the world has been divided into the wrong pieces, even a brilliant mind can end up worshipping its own error.

The future belongs to people who can do the harder work: not just naming the thing in front of them, but discovering the structure that made it seem singular.