The Map of Every Possible Design

Here is a picture worth an entire course. Take every possible version of some thing — every possible genome for a beetle, every possible shape for a wing, every possible layout for a keyboard, every possible strategy for a business — and spread them all out on a flat map, arranging them so that versions that are similar sit close together and wildly different ones sit far apart. That map is already strange to think about (it has as many dimensions as the thing has knobs to turn), but now do one more thing to it: at every point, build a tower whose height is how good that version is — how well it survives, reproduces, works, or wins. Tall towers are excellent designs; short ones are terrible.

What you’ve just built is a fitness landscape: a rolling terrain of peaks, valleys and ridges where altitude is quality. The geneticist Sewall Wright drew this picture in 1932, and it quietly reorganized how careful people think about improvement of every kind — in biology, yes, but also in engineering, machine learning, strategy, and your own stalled attempts to get better at something. Because once you can see improvement as movement across a terrain, one devastating fact jumps out and never lets go.

The one idea to take away

Sit with why that’s such a big deal. A process that only steps uphill has no map and no foresight — it can’t see the taller mountain, can’t plan a route, can’t tolerate getting temporarily worse to end up much better. It just feels around its own feet for the steepest way up and takes it, over and over, until every direction points down. Then it’s done — marooned on the top of its local hill, which might be a magnificent peak or a pathetic little mound, with something far better clearly “over there” and utterly out of reach. That trapped-on-a-hilltop state has a name we’ll use all course: a local optimum. It is the reason “good enough” so reliably beats “best” and then refuses to move — in genomes, in gadgets, and in you.

That cave fish is going to follow us through the course, because it’s the cleanest possible case: a genuinely improving population that is nonetheless permanently barred from a far better design, purely because the route there runs downhill. You just predicted that without knowing one fact about fish. That’s the landscape doing the work.

Watch a search climb the nearest hill — and get stuck

Reading about hill-climbing is one thing; watching a population do it is another. Below is a live fitness landscape. The horizontal axis is every possible design, laid out left to right; the height of the terrain is fitness. Press Drop a population to place one somewhere at random and let selection climb. Watch what it does: it walks uphill — always uphill — until it reaches the top of whatever hill it landed on, and then it halts.

The dashed flag marks the tallest summit on the whole landscape — the best design there is. Notice how often your climber stops somewhere lower, stranded on a local peak with the true summit flagged across a valley it cannot cross. Then try two things. Drag the ruggedness slider from smooth (one gentle hill, easy — the climber always finds the top) to rugged (a jagged range full of traps). And when your population is stuck, press Mutate / big jump to fling it to a new spot — sometimes it lands on the slope of a taller hill it could never have reached by climbing, and that’s the escape.

Two things are worth burning in while you play. First, the climber never goes downhill — that single rule is the entire engine, and it’s why it gets stuck. Second, the landscape’s shape decides its fate: on a smooth one-hill landscape, hill-climbing is flawless and always finds the best design; on a rugged, many-peaked one, the same process gets trapped almost every time. Same search, wildly different outcomes, purely because the terrain changed. Hold onto that — half of getting good results is about the shape of the problem, not the cleverness of the searcher.

Why this belongs in a latticework

Fitness landscapes earn their place because the picture is ruthlessly general — the same map governs searches that look nothing alike on the surface:

• In biology, it’s why evolution produces designs that are excellent-but-flawed rather than optimal: the eye’s backwards wiring and its blind spot, the absurd detour of the recurrent laryngeal nerve, the panda’s “thumb.” Each is a local peak that selection climbed and can’t climb back down from. Wright drew the landscape to explain exactly this.
• In machine learning, training a model is hill-climbing on a fitness landscape (there it’s called gradient descent on a loss surface — same picture, upside down). Getting stuck in a bad local optimum, and the tricks used to escape it, are daily engineering problems.
• In strategy and business, a company sits on a peak of “how we do things,” climbing it higher and higher — right until a rival on a different, taller mountain (a new technology, a new model) makes the whole first peak irrelevant. Being trapped on a local peak is what disruption feels like from the inside.
• In your own life, every “I’ve plateaued” is a local peak: your current technique is locally optimal — every small tweak makes you worse — and getting genuinely better requires a temporary step down (rebuilding your grip, your swing, your process) before you can climb a higher hill.

The thread through all of them is a single diagnostic you’ll learn to run on sight: Is this search stuck on a local peak, and if so, what would it take to cross the valley to a better one? That question — and the handful of answers to it — is the whole payoff of the course.

The map of the course

Five teaching lessons, then one exam you can’t undo. The route:

1. Mapping the Terrain — how to actually build a fitness landscape: designs as points arranged by similarity, fitness as height, and the vocabulary of peaks, valleys, ridges and neighbours. Wright’s 1932 picture, made concrete.
2. Hill-Climbing — the engine in detail: why selection is an uphill-only, local, foresight-free search, what “local” really means, and why that one rule is both its power and its prison.
3. Trapped on a Low Peak — the central consequence: local optima. Why “good enough” beats “best” and stays there, with the real biological evidence — the blind spot, the recurrent laryngeal nerve — and the idea of lock-in.
4. Ruggedness & Crossing Valleys — what makes a landscape smooth (easy) or rugged (a maze of traps), and the forces that let a search cross a valley to a better peak: mutation, drift in small populations, recombination — and the Red Queen twist where the terrain itself won’t hold still.
5. The Model Everywhere — cashing it out beyond biology: gradient descent in machine learning, engineering and design, business disruption, skill plateaus, and the master trade-off of all search — explore vs. exploit.

Then a Final Exam — graded, one question at a time, one-way: once you answer, it locks. No back button, no retries, 70% to pass.

How to use this course

One habit does most of the work: whenever you see something improving, ask what hill is it climbing, and is there a taller one it can’t reach? Treat every exercise like the cave fish — guess the outcome before you reveal the answer, because the small sting of being wrong is what makes “uphill-only search gets stuck” finally stick. And play with the landscape above until two patterns feel obvious in your gut: the climber that stops on a low peak, and the mutation-jump that occasionally saves it. A trap you’ve watched a search fall into sticks far better than one you’ve only read about.

Next up: lesson 2, where we build the landscape properly — how to turn “every possible design” into an actual terrain you can reason about, and why neighbours and height are the only two ideas you need.