The Power of Useful Weirdness: How Embeddings and Marketing Solve the Same Problem

When similarity is not enough

What if the hardest part of building something valuable is not making it better, but making it distinct enough to be remembered?

That question sits at the center of two worlds that rarely talk to each other. In one, machine learning tries to turn messy life events into vectors, so that one medical journey can be compared to another. In the other, marketing tries to turn ordinary products into memorable choices, so that one offer does not dissolve into the blur of everything else. Both are, at heart, about finding structure in similarity without erasing difference.

That tension matters because most systems fail in one of two ways. They become too generic, and no one notices them. Or they become too eccentric, and no one trusts them. The real craft lies in a narrow band between the two: enough regularity to be understood, enough weirdness to be interesting.

The most powerful things are rarely the most average, but they are also rarely the most bizarre. They live in the space where pattern and surprise cooperate.

This is why embeddings and marketing belong in the same conversation. Both are methods for solving a translation problem: how do you make something complex legible, comparable, and actionable, without flattening it into banality?

---

The hidden similarity between medical journeys and brands

An embedding is a way of giving meaning a coordinate system. Instead of treating each item as isolated, you place it in a space where proximity matters. Two words near each other have related meanings. Two claim codes, if modeled well, may reflect similar medical journeys. A vector representation is not the thing itself, but it captures the relationships that make the thing intelligible.

That sounds technical, but the same logic drives great marketing. A brand also occupies a position in a mental space. People do not evaluate products in a vacuum. They compare them, often unconsciously, against nearby alternatives. If your product looks, sounds, and behaves like every other option, then it occupies no clear place in the customer’s mind. It is unembeddable in memory.

This is where many teams get confused. They think the goal is to remove friction by making the product universally palatable. But in practice, universal palatability often destroys structure. If everything is optimized to be slightly less annoying, slightly more average, and slightly more familiar, then nothing becomes distinctive enough to guide choice.

A good embedding is not a bland average. It is a compressed representation that preserves the useful geometry of the world. Great marketing does something similar. It compresses a product into a simple signal, but the signal must preserve the product’s identity. A luxury perfume, a rebellious sneaker, or a trusted health service all succeed when they occupy a precise place in the customer’s mental map.

The deeper shared problem is this: how do you preserve meaningful difference inside a system that rewards similarity?

---

Why the middle gets crowded and the edges get remembered

Most categories become crowded in the middle. The middle is safe. The middle is plausible. The middle is where committees land when they want to avoid risk. But the middle is also where memory goes to die.

Rory Sutherland’s idea of the right amount of weird is useful here because weirdness, when calibrated correctly, creates compressibility. A strange detail gives the brain something to latch onto. It acts like a distinctive feature in an embedding space, a coordinate that helps separate one point from another. Without it, two options collapse into one another.

Think about how people remember hotels, apps, restaurants, or even doctors. They rarely remember the one that was merely acceptable. They remember the place with the surprising lamp, the app with the odd little ritual, the restaurant with the menu item nobody else dares to serve, or the doctor who made the experience feel unusually calm and personal. These details are not just decorative. They are memory scaffolding.

The same principle applies to complex human services like healthcare. A medical journey is not a single event but a sequence of interactions: diagnosis, treatment, billing, follow up, anxiety, relief, confusion, and trust. If those journeys are represented well, patterns emerge. Similar journeys cluster together, and interventions can be targeted more intelligently. But if the representation is too coarse, every patient becomes a generic case.

That is the real lesson: distinctiveness is not the opposite of usefulness. It is often what makes usefulness possible.

When a model learns that certain claims histories are close to one another, it is discovering that some journeys share structure. When a brand learns that a certain tone, visual style, or product quirk makes it easier to recognize and talk about, it is discovering that some signals stick because they are not interchangeable. Both are forms of learning how to be similar in the right places and different in the right places.

---

The economy of attention rewards asymmetry

Attention is expensive. People do not have time to inspect every detail, compare every attribute, or appreciate every improvement. They use shortcuts. That means the world is not selected by perfect information, but by salient information.

This is where ordinary optimization often fails. Teams spend enormous effort improving features that do not change perception. They refine experiences in ways that are measurable but not memorable. In machine learning terms, they optimize the underlying signal without improving the representation that humans actually use. In marketing terms, they enhance the product without changing the choice architecture.

A new product can be objectively better and still lose if it is mentally invisible. That is why many innovations were initially dismissed. Electricity was not just a better candle. The internet was not just a better fax machine. Each needed a new frame, a new language, and often a new social ritual before people could understand what it was for.

This is where fame enters the picture. Fame is not just vanity. It is a reduction in cognitive cost. A famous thing is easier to classify, easier to recommend, easier to trust, and easier to search for. In a world of limited attention, fame functions like a precomputed embedding in the social mind. It tells people where to place you before they have done the full analysis.

But fame alone is not enough. If everyone knows you and nobody can say why you are different, fame becomes a noisy substitute for identity. The best outcome is not generic fame. It is famous weirdness, a recognizable pattern with a memorable twist.

The goal is not to be strange for its own sake. The goal is to become the kind of strange people can describe to someone else.

That is the difference between gimmick and signal. A gimmick is weird without structure. A signal is weird in a way that reinforces identity.

---

A framework: the 3 layers of memorable systems

To connect these ideas, it helps to think in three layers.

1. The underlying reality

This is the actual thing, the medical journey, the product, the service, the human experience. It has complexity, edge cases, and nuance. In healthcare, this is where claims, conditions, and outcomes live. In marketing, this is where the product’s real strengths and tradeoffs live.

2. The representation layer

This is how the underlying reality is encoded so it can be understood. Embeddings are one form of representation. Brand positioning is another. A good representation does not copy reality. It makes reality easier to navigate. It groups similar things together and separates dissimilar things enough to matter.

3. The attention layer

This is the human layer, where decisions are actually made. People are not reading your data schema. They are not inspecting your full feature matrix. They are responding to cues, stories, shortcuts, analogies, and familiarity. This layer is where weirdness, fame, and distinctiveness either help or hurt.

Most failures happen when teams confuse these layers. They optimize layer one and assume layer three will follow automatically. It rarely does. A superior product does not automatically become a superior choice. A better model does not automatically become a better intervention. You need representation and attention to align.

This is why the best systems often look a little overdesigned from the outside. They have a clear shape. They have memorable features. They create a clean mental slot. They are not trying to be everything to everyone. They are trying to become easily placeable in the minds of the people who matter.

That is not superficial. That is architecture.

---

The practical paradox: optimize for clarity, preserve one odd detail

If you want a system to spread, you usually need two things at once: clarity and a signature.

Clarity means people can quickly understand what it is, what it does, and why it matters. A healthcare model needs interpretability and actionable outputs. A product needs a comprehensible promise. Without clarity, even brilliance becomes friction.

A signature means there is one thing that breaks the pattern just enough to be memorable. It could be a visual style, a tone of voice, a ritual, a surprise feature, or a narrative frame. It should not confuse the user. It should help them remember you.

Consider a few concrete examples:

• A telehealth service that makes every interaction efficient may still blur into the rest unless it has one unmistakable ritual, like a follow up that feels unusually human.
• A consumer app can be packed with features, but if all the screens feel interchangeable, users will not form a mental map.
• A medical recommendation engine may be highly accurate, but if clinicians cannot see why it surfaced a pattern, they may not trust it enough to act.
• A brand can speak in plain language and still stand out if it owns one distinctive metaphor, visual cue, or promise.

The point is not to sprinkle eccentricity everywhere. Too much weirdness becomes noise. The point is to create one or two reliable asymmetries that make the whole system easier to remember and harder to copy.

That is also why the instruction to maintain some idiosyncrasies is so powerful. The temptation in competitive environments is to remove anything that seems unusual. But unusual things are often the anchors people use to form meaning. Strip away every peculiar edge, and you may gain polish while losing traction.

The best products, like the best embeddings, do not maximize sameness. They maximize structured separability.

---

Key Takeaways

1. Do not confuse average with accessible. If everything looks and feels generic, people cannot place you in memory or in the market.

2. Build a representation, not just a product. Whether you are modeling claims data or positioning a brand, the crucial task is to preserve the relationships that matter.

3. Use one memorable asymmetry. A distinctive ritual, tone, feature, or visual cue can dramatically improve recall without harming usability.

4. Optimize the attention layer intentionally. Better performance is not enough if no one can recognize, trust, or talk about what you built.

5. Aim for famous clarity, not bland perfection. The best systems are easy to understand and hard to confuse with anything else.

---

The real lesson: meaning lives in difference

The most interesting connection between embeddings and marketing is not technical. It is philosophical. Both reveal that meaning is not created by isolated things, but by relationships among things. A word means what it means partly because of the words around it. A product becomes compelling partly because of the alternatives around it. A medical journey becomes actionable when its structure can be compared to other journeys.

That means the goal is not to eliminate difference. The goal is to make difference legible.

In a noisy world, the winners are not always the best. They are often the ones with the best shape in the mind of the user. They are the ones that can be recognized quickly, explained easily, and remembered accurately. They are similar enough to be trusted and strange enough to stand apart.

So the next time you are designing a model, a message, or a product, ask a question that cuts across both machine learning and marketing:

What is the smallest distinctive signal that still preserves the truth?

If you can answer that, you are not just building something that works. You are building something that people, and systems, can actually find.