At the aquarium, a shark and a dolphin swim past you. Spindle-shaped body, a fin on the back, a tail pushing from behind. Not knowing better, you would take them for the same kind of thing.
They are far apart. A shark is a fish; a dolphin's ancestors walked on four legs on land, and it still has to surface for air. Add a third: the ichthyosaur, a marine reptile from over 200 million years ago, whose fossil could almost be paired with a dolphin, and which is related to neither of the other two. Three separate origins, trimmed by the same water into the same outline.
The strange part is not that they look alike. It is that the likeness follows a rule: some things get stumbled upon again and again — eyes, this spindle shape, proteins that keep blood from freezing — while others look as though they succeeded only once. Why is the same machine so generous with one kind of answer and so stingy with another?
Close your eyes and turn your face to the sun. Red glow behind the lids. Turn away, and it darkens. Your eyes never opened, yet you know which side the light is on.
That is the whole job of the most primitive eye: a patch of skin that senses light. Between that patch and what you are using right now lie several steps, and every one of them has to pay off on the spot — nobody is storing a half-finished part for you until it becomes useful later.
A flat patch can only report bright or dark. Let it dip inward a little and the rim blocks light coming from the side, so you begin to tell which direction light comes from; deeper means sharper, and each increment is immediately useful — knowing which side the shadow is falling from is knowing which way to dodge. Close it into a near-sealed cup with only a pinhole left, and light forms an image inside: the nautilus still uses this eye today, which sees shapes, but dimly. Grow a transparent membrane over the opening to keep grit out, thicken its middle, and it starts to focus. A lens was not invented; it is a membrane slowly getting thicker.
In 1994 two researchers costed the slope out: allowing shape to change by only one percent per generation, and choosing conservative values at every turn, going from a flat light-sensitive patch to a fish-like lens eye takes roughly 364,000 generations — a few hundred thousand years for a small animal with one generation a year. That is a model, not a fossil record, but it answers the question that actually stalls people: is there enough time? Comfortably.
The count is even more telling. In 1977 someone went through the eyes of the major animal groups one by one: eyes of a basic kind have arisen independently at least 40 times in the animal kingdom, and possibly as many as 65.
"Independently" has a limit, though. An octopus eye is a camera eye like yours: one opening, one lens, one retina; their common ancestor had no such eye, and the two sides built it separately. Yet the molecule you both use to catch light belongs to the same family, and the master switch gene that starts eye development is the same one — an old part that was already around long before. What got reinvented over and over is the optical scheme; the light sensor was invented once, and has been conscripted ever since.
And building separately leaves scars: your retina is installed backwards, with the cabling laid in front of the photoreceptors and a hole punched through to let the bundle out — that hole is your blind spot; the octopus has no such trouble. One function, two implementations, each carrying the debts of its own road.
Water is about 800 times denser than air. For anything trying to be fast in it, drag overrides every other consideration, and the usable body shapes get squeezed into a narrow band: blunt in front, thick in the middle, tapering behind, driven from the rear of the body.
Saying "they became streamlined in order to swim fast" is fluent, and wrong exactly where it matters — there is no "in order to" in evolution. What actually runs in the ledger is this: each generation is born with shapes that vary a little; the ill-suited ones swim slightly slower and get eaten slightly more often, and leave slightly fewer offspring. Water needs no intention. It only needs to charge a higher price for the wrong shape. Charge it for a few million years and what is left all looks the same.
Constraint has a second source: the parts on hand. Evolution can only edit what already exists. A dolphin's tail beats up and down — its ancestors' spines flexed up and down as they ran on land; a shark's tail beats side to side. The same outline, and two perpendicular ways of rowing. History fingerprints the answer without changing the answer itself.
This has really happened, and with frightening precision. Antarctic notothenioid fish and Arctic cod each evolved almost identical antifreeze glycoproteins, whose job is to stick to newly formed ice crystals in the blood and stop them growing. The Antarctic version was remodelled from a digestive-enzyme gene; the Arctic one has nothing to do with that gene and came from somewhere else entirely. The same answer, from two unrelated starting points.
The scale can be large, too. C4 photosynthesis — a package of modifications involving leaf structure and several biochemical steps that lets a plant waste less water where it is hot and dry — has arisen independently at least 60 times in flowering plants. Bats and toothed whales each took up echolocation to move in the dark, and even the inner-ear protein that amplifies sound has been changing in the same direction in both.
But do not push this too far. The thylacine and the wolf, separated by about 160 million years, have nearly identical skulls. After the thylacine genome was sequenced in 2018, almost no signal of convergence with the wolf turned up in the protein-coding genes; abundant signs of it were found later in regulatory regions — the parts that set when, where, and how strongly a gene switches on. The outlines collided head-on, but the parts that changed were not the same parts.
There are four large islands in the Caribbean: Cuba, Hispaniola, Jamaica, Puerto Rico. Each has its own population of anole lizards, and on each island those lizards split into the same cast of roles: big-bodied ones working the crown, short-limbed short-tailed ones creeping along thin twigs, long-legged ones sprinting between trunk and ground. Four separate runs, the same division of labour, matching even in body proportions.
Why can it repeat? Two conditions hold at once. The selection pressure is narrow: walk on a thin twig with long legs and you cannot keep your footing — that is mechanics, and it does not negotiate. And the supply of variation is ample: large numbers, many generations, usable variants in the same direction turning up again and again, so raw material is never short. The better those two hold, the more alike the replays.
Sometimes even the genes repeat. After sticklebacks moved from the sea into freshwater lakes, populations everywhere repeatedly lost their bony armour and pelvic spines (fresh water is low in calcium, and the predators are a different set). In unrelated lakes, the step that loses the pelvic spine often lands on the same switch beside the same gene — and that stretch of switch happens to break unusually easily. It was not chosen; it is the cheapest thing to break, so the same "accident" replays lake after lake.
Another layer, though, does not replay. There is an experiment that has been running since 1988: twelve flasks of identical E. coli, passaged generation after generation in identical medium. The medium has always contained citrate, which this bacterium does not use when oxygen is present. Thirty thousand generations passed and not one flask had learned. At about 31,500 generations, one suddenly did. The researchers pulled that flask's frozen samples from across the years and replayed them: the trait reappeared only when restarted from the later time points, and never from the early ones, however many runs they did. Before that step, a set of changes had accumulated that showed no benefit at the time — they were the ticket of admission to the step that followed. The same environment, the same starting point: the innovation happened in one flask out of twelve, and will not replay from the wrong point in its history.
So "is evolution predictable" is too coarse a question; it has to be split. At the layer of ecological roles and body shapes, it is highly repeatable — give it a body of water, a thin twig, a freezing sea, and there are only so many answers; anyone who walks in here will collide with the same one. At the layer of one particular innovation, it often is not — that one rides on a string of prior changes that paid nothing at the time, and only once they have all landed does the rest become possible. Some read the sheer quantity of convergence as meaning a replayed Earth would look much the same; others hold that contingency overwhelms everything. The argument is unsettled — and the two sides are holding different layers.