Secrecy
Секретность
Uniqueness Holds Depth 4 · Mine
hard to copy because knowledge is protected
Moats it grows into
- #3 Proprietary data
A data set you gathered and never published is a secret with a moat's shape: a rival can collect its own, but not yours, and the time it took you is the time it costs them. Eroding, because public data and synthetic sets close the gap.
- #4 Proprietary models
Weights and methods fitted to your data and kept unreleased: the product can be used, and even rented, without the model ever leaving your hands. Distillation from outputs is the leak to watch.
- #19 Process power
A way of working made of hundreds of small routines is a secret nobody can steal whole; an ex-employee knows their part, a visitor sees a day. The deepest of the three, and the slowest to build.
Summary
Secrecy is protecting what you know instead of what you make: the formula, the training set, the routine on the shop floor. It holds, because a rival who can see the product still cannot see how it is made. It digs three moats, Proprietary data, Proprietary models and Process power, each as deep as the secret is hard to reconstruct.
How it works
A copy needs a recipe. Most products publish theirs by existing: the feature list, the interface, the price are all visible, and a rival rebuilds them from the outside. A secret withholds the one part that cannot be inferred from the result, the data the product was fitted to, the weights, the sequence of small decisions the factory makes in order, so the copier gets the shape and not the behaviour. The edge is the gap between the visible product and the hidden inputs.
The mechanism depends on two things. The secret must be genuinely hard to reconstruct: a formula a lab can reverse in a week is a delay, not a defense, while a training set gathered over years or a process that lives in a thousand habits cannot be bought. And it must be kept: every employee who leaves, every supplier who sees the line, every model that can be distilled from its outputs is a leak, and the moat is as deep as the least careful person in it.
Why it holds — or only takes
It holds because what is hidden cannot be copied from the product; the depth depends on the kind of secret. Proprietary data is direct: a set you gathered that no one else can assemble, protected by not publishing it. Proprietary models is direct: weights and methods fitted to that data and never released, rentable from you but not rebuildable. Process power is direct and the deepest: a way of working that lives in hundreds of small routines, secret not because it is locked but because no outsider sees all of it at once. A secret reversed in a week is a delay, and maps nowhere.
Do you have it
- A rival has tried to reverse-engineer the product and shipped something that looks the same and behaves worse.
- There is a document, a data set or a line that only a handful of people are cleared to see.
- New hires take months to become useful because the way you work is not written anywhere they can read.
- You have refused to publish a benchmark, a method or a supplier name that a rival could learn from.
How to dig it
- Separate the visible product from the hidden inputs and decide which input a copier cannot infer from the result.
- Gather the data or the routine over years, and keep it where a departing employee cannot take it whole.
- Serve the model rather than shipping it: an API leaks outputs, a download leaks weights.
- Write the process down for yourselves, not for anyone else, and split it, so no one role holds all of it.
- Watch your outputs: what a rival can distil from them is no longer secret.
How it erodes
- A key employee leaves and rebuilds the routine elsewhere; a trade secret walks out in a head.
- A rival distils your model from its outputs, or public data catches up with your set.
- The product's behaviour becomes explainable: a paper, a leak, a regulator's demand for the method.
- You publish to recruit, to sell or to comply, and what is published cannot be un-known.
Solo, without capital
A solo builder can keep a secret perfectly, since there is nobody to leak, and Proprietary data is solo-reachable: a set you scraped, labelled or collected over a year is yours as long as you do not publish it. Proprietary models and Process power are not: one person cannot fund a model worth hiding or run a process big enough for its secrecy to matter. What you cannot do is stop a well-funded rival from gathering a bigger set; your secret has to be one that money does not shorten.
What the matrix says of its moats
- #3 Proprietary data Solo ✓ solo-reachable Capital ◐ medium AI → eroding Rent ~ partly
- #4 Proprietary models Solo ✗ no Capital ◉ extreme AI ↑ AI-resistant Rent ✓ rentable
- #19 Process power Solo ✗ no Capital ◐ medium AI ↑ AI-resistant Rent ✗ not for sale
In nature, in business
- Nature
- —
- Business
- magicians, Coca-Cola, trade secrets
Cases
- Coca-Cola — Never patented the formula, because a patent expires and publishes; the recipe stays a trade secret with no end date.
- WD-40 — Sells one product for decades on a formula that was never patented for the same reason, and mixes the concentrate in-house.
- Google Search — Publishes what a page should do and never how the ranking system weighs it, so rivals can see every result and copy none of the method.
Combines with
- Security — A secret is only as good as the door in front of it; access control is the practice that keeps the moat from leaking.
- Obscurity — The best-kept secret is one nobody knows exists; staying off the radar means nobody is trying to reverse it.
- Privacy — The discipline that protects customers' data is the same one that protects your own; a firm that leaks one leaks both.
Fights with
- Transparency — Trust built by showing the inner workings is bought with exactly what secrecy withholds.
- Distributed ownership — A community that owns the thing sees all of it; a moat made of what nobody outside can see cannot be shared.