水族馆里，鲨鱼和海豚从你面前游过去。纺锤形的身子，背上一片鳍，尾巴在后面推。不知道底细的话，你会以为它们是一类东西。
它们差得远。鲨鱼是鱼；海豚的祖先在陆地上用四条腿走过路，今天还得浮上来换气。再加第三个：两亿多年前海里的爬行动物鱼龙，化石和海豚几乎能配成一对，跟这两位都不沾边。三个出身，被同一片水修成同一个轮廓。
怪的不是它们像，是像得有规律：有些东西被反复碰出来——眼睛、这个纺锤形、防止血液结冰的蛋白；另一些看上去只成功过一次。同一台机器，为什么对一部分答案来者不拒，对另一部分一毛不拔？
闭上眼睛，把脸转向太阳。眼皮里一片红。转开，暗下去。你没睁眼，却知道光在哪边。
这就是最原始的眼睛在做的全部事情：一块能感光的皮。从这块皮到你现在用的东西，中间隔着好几步，而每一步都得当场就划算——没有谁替你保管半成品，等它以后变好用。
一块平的皮只能报告亮不亮。皮往下凹一点，边缘就挡住侧面来的光，你开始能分辨光从哪边来；凹得越深方向越准，每深一点都立刻有用——知道影子从哪边压过来，就是知道该往哪边躲。凹成一个几乎封口的坑只留小孔，光在里面就成了像：鹦鹉螺今天还用这种眼睛，看得见形状，但很暗。孔口长一层透明的膜挡沙子，膜中间厚一点就开始聚光。晶状体不是被发明出来的，是一层膜慢慢变厚。
1994 年有两个人把这条坡算了一遍：每代只允许形状改动百分之一，参数各处都往保守里取，从平的感光斑走到鱼那样带晶状体的眼睛，大约三十六万代——对一年一代的小动物就是几十万年。这是模型不是化石记录，但它回答了真正卡人的那个问题：时间够不够。够得很富余。
更说明问题的是次数。1977 年有人把各大类动物的眼睛挨个盘了一遍：眼睛这种器官在动物界至少独立起源过四十次，可能多到六十几次。
不过「独立」有个限度。章鱼的眼睛和你的一样是照相机式的：一个孔、一个晶状体、一块视网膜；共同祖先没有这种眼睛，两边各造各的。可你们用来感光的分子是同一类，启动眼睛发育的总开关基因也是同一个，那是很早就在的老零件。被反复重新发明的是光学方案；感光的零件只出现过一次，后来被反复征用。
而各造各的会留疤：你的视网膜是反着装的，接线铺在感光细胞前面，还得捅个洞把线束送出去——那个洞就是你的盲点；章鱼没这毛病。同一个功能，两个实现，各带各的旧账。
EN Close your eyes facing the sun: red glow, then dark when you turn away. That is all a primitive eye does — a patch of skin that senses light. Every step from there has to pay off on the spot, because nobody is holding a half-finished part in storage. A flat patch reports only brightness. Let it dip inward and its rim shades light from the side, so direction becomes readable; deeper means sharper, and each increment is immediately useful. Close it to a near-sealed cup with a pinhole and it forms an image — the nautilus still uses such an eye, which sees shapes but dimly. Grow a transparent membrane over the opening to keep out grit, thicken its middle, and it starts to focus. A lens is not invented; it is a membrane getting thicker. In 1994 two researchers costed this slope out: allowing only 1% change per generation and choosing conservative parameters throughout, going from flat patch to a fish-like lens eye takes about 364,000 generations — a few hundred thousand years for a small animal. That is a model, not a fossil record, but it answers the question that actually stalls people: is there enough time? Comfortably. More telling is the count. A 1977 survey of eyes across the animal phyla concluded that eyes arose independently at least 40 times, possibly 65. "Independently" has a limit, though. Octopus and human eyes are both camera eyes built separately, yet both use the same family of light-catching molecule and the same master control gene for eye development — ancient parts, repeatedly conscripted. The optics were reinvented many times; the photosensor was invented once. And building separately leaves scars: the vertebrate retina is wired backwards, with the cabling in front of the photoreceptors and a hole where it exits — your blind spot. The octopus has no such hole.
水的密度大约是空气的八百倍。在水里想快，阻力压倒一切别的考虑，可用的身体形状就被压成很窄的一段：前面钝、中间粗、后面收细，靠身体后段推。
说成「它们为了游得快而进化出流线型」很顺口，也正好错在最要紧的地方——演化没有「为了」。实际跑的账是：每代生下来的个体形状各有出入，不合适的那些游得慢一点、被吃掉的机会多一点，留下的后代就少一点。水不需要有意图，它只需要对不合适的形状收更高的费用。收上几百万年，剩下的就都长一个样。
约束还有第二个来源：手里的零件。演化只能改已有的东西。海豚的尾巴上下摆——它祖先在陆地上跑时脊柱就是上下起伏的；鲨鱼的尾巴左右摆。同一个轮廓，划水方向却是垂直的两种。历史在答案上按了指纹，却没能改掉答案本身。
这件事真发生过，而且精确得吓人。南极的南极鱼和北极的鳕鱼各自演化出几乎一样的抗冻糖蛋白，作用是贴住血液里刚形成的冰晶不让它长大。南极鱼那套是从一个消化酶的基因改来的；北极鳕鱼那套和它毫无关系，出身完全在别处。同一个答案，两个互不相干的起点。
规模也可以很大。C4 光合作用——一套让植物在高温缺水处少浪费水的改装，牵涉叶片结构和好几步生化反应——在开花植物里至少独立出现过六十次。蝙蝠和齿鲸各自摸黑用回声定位，连内耳里负责放大声音的那个蛋白都在往同一个方向改。
但别把这条推太远。袋狼和狼隔着大约一亿六千万年，头骨几乎一模一样。2018 年测出袋狼基因组之后，人们在蛋白质编码的基因里几乎找不到和狼趋同的信号；后来在调控区（管基因什么时候、在哪里、开多大的那部分）才找到大量趋同迹象。外形撞了个满怀，改的却不是同一批零件。
EN Water is roughly 800 times denser than air, so for anything fast, drag overrides everything else — which squeezes the usable body shapes into one narrow band. Saying they "became streamlined in order to swim fast" is fluent and wrong where it matters: there is no "in order to" in the ledger. What actually runs is this: shapes vary, poorly shaped individuals swim slower, catch less and get eaten more, and leave fewer offspring. Water needs no intention; it only needs to charge a higher price for the wrong shape. Constraint has a second source: available parts. Evolution edits what is already there. A dolphin's tail beats up and down because its land-running ancestors flexed their spines that way; a shark's beats side to side. Same water, same outline, perpendicular strokes — history fingerprints the answer without changing it. The convergence can be startlingly precise. Antarctic notothenioid fish and Arctic cod independently evolved nearly identical antifreeze glycoproteins that bind tiny ice crystals in the blood and stop them growing. The Antarctic version was remodelled from a digestive-enzyme gene; the Arctic one has no relation to that gene at all. Same answer, unrelated starting points. It also repeats at scale: C4 photosynthesis — a package of leaf-anatomy and biochemical changes that wastes less water in hot, bright, dry places — has arisen independently at least 60 times in flowering plants. Bats and toothed whales each evolved echolocation, and even the inner-ear protein that amplifies sound has been changing in the same direction in both. But do not push the rule too far. The thylacine and the wolf, separated by about 160 million years, have near-identical skulls. After the thylacine genome was sequenced in 2018, almost no convergence showed up in protein-coding genes; the signal turned up later in regulatory regions — the parts that set when, where and how strongly genes switch on. The outline collided head-on; the parts changed did not. Convergence reaching down to the molecule tends to happen where the job is narrow and the solutions few.
加勒比海上有四个大岛：古巴、伊斯帕尼奥拉、牙买加、波多黎各。每个岛上都住着一群安乐蜥，而每群都各自分化出同一套角色：树冠上活动的大个子、细枝上爬的短腿短尾、树干到地面之间冲刺的长腿……四次分头进行，同一套分工，连体型比例都对得上。
为什么能重复？两个条件同时成立。一是选择压很窄：细枝上走，腿长了就站不稳，这是力学，不商量。二是变异供应足：数量大、世代多，同一个方向上的可用变异反复出现，不缺原料。这两条满足得越好，重播越像。
有时连基因都重复。三刺鱼从海里进到淡水湖之后，各地反复丢掉骨甲和腹刺（淡水里钙少、天敌也换了一批）。互不相干的湖里，丢腹刺这一步常常落在同一个基因的同一段开关上——而那段开关恰好特别容易断。不是它被挑中，是它最容易坏，于是同一个「意外」在不同的湖里一次次重演。
可另一层就重播不出来。有个从 1988 年跑到今天的实验：十二瓶完全一样的大肠杆菌，在完全一样的培养液里一代代传下去。液里始终有柠檬酸，而这种细菌在有氧条件下不会用它。三万代过去，一瓶都没学会。到大约三万一千五百代，一瓶忽然会了。研究者取出这一瓶历年的冷冻样本重播：只有从比较晚的时间点重来才可能再次出现，从早期重来怎么跑都不出现。这一步前面先攒下了一些当时看不出好处的变化，它们是后来那步的入场券。同一个环境、同一个起点，创新只发生在十二分之一，换个历史点还重播不出来。
所以「演化可不可预测」问得太粗，得拆开问。生态角色和外形那一层高度可重复——给定一片水、一根细枝、一个结冰的海，答案就那么几个，谁走到这儿都会撞上。某一次具体创新那一层往往不可重复——它依赖一串当时毫无用处的前置变化，碰齐了才有后面。有人把趋同之多读成「重播地球还会得到相似的世界」，也有人认为偶然性压倒一切。这个还在争论——而两边握着的其实是不同的那一层。
EN Four islands in the Caribbean — Cuba, Hispaniola, Jamaica, Puerto Rico — each carry anole lizards, and on each island the lizards independently split into the same cast of roles: crown-dwelling giants, short-limbed twig specialists, long-legged sprinters of trunk and ground. Four islands, four separate runs, the same division of labour down to body proportions. Two conditions make repetition possible. The selection pressure is narrow: walking on a thin twig with long legs does not work, and mechanics does not negotiate. And variation is abundant: large populations and many generations keep supplying usable variants in the same direction. The better both hold, the more alike the replays. Sometimes even the genes repeat. Sticklebacks entering freshwater lakes have repeatedly lost their bony armour and pelvic spines, and in unrelated lakes the pelvic loss often lands on the same switch beside the same gene — a stretch of sequence that happens to break unusually easily. It was not chosen; it is the cheapest thing to break, so the same "accident" recurs. Another layer does not replay at all. Twelve identical flasks of E. coli have been passaged since 1988 in identical medium containing citrate, which the bacterium does not use when oxygen is present. For 30,000 generations none of the twelve learned to. At about 31,500 generations, one did. Replaying frozen samples from that lineage, the trait reappeared only when restarted from later time points, never from early ones — meaning earlier, apparently useless changes had accumulated as the ticket of admission. Same environment, same start: one innovation, one flask in twelve, and unreplayable from the wrong point in history. So "is evolution predictable" is too coarse a question. At the level of ecological roles and body shapes, highly repeatable. At the level of a particular innovation, often not — it rides on a string of prior changes that paid nothing at the time. Some read the sheer amount of convergence as meaning a replayed Earth would look much the same; others hold that contingency dominates. The argument is unsettled — and the two sides are largely holding different layers.