春天，天还没亮，窗外那只鸟就开始叫。一天接一天，几乎踩着点。
你问一句「它为什么这时候叫」。有人会答「白天变长了，它身体里的激素跟着变」。也有人会答「叫得早的公鸟更容易占住一块地盘」。这两个答案听上去在抢同一个位置——其实谁也没挡谁的路。
行为比骨头难办得多。骨头会变成化石，摆在桌上谁都能量；行为落地就消失，还特别容易被我们顺手安上意图。而日常语言偏偏只给了我们一个「为什么」。
回到窗外那只鸟。四种答案，全都成立。
一，此刻它身体里在发生什么。白天变长，光信号进到脑子里，一串激素跟着变；管唱歌的那几块脑区在繁殖季会胀大一圈，过了季又缩回去——这件事最早是在金丝雀身上量出来的。
二，这一只这辈子怎么长成这样的。它小时候听过谁唱，在哪一段时间听的，后来又练了多久。
三，带着这种叫法的个体，留下的后代是多还是少。这一格最容易说歪。别说「它为了求偶才唱」——账不是这么记的。真实的记法是：每年出生的公鸟，在开口的时间和力气上各有出入，其中某些个体占住了更好的地盘、配上了对、把幼鸟养活了，于是它们那一套在下一代里占的比例高一点。没有谁想干什么，只有谁留下的后代多。
四，这套行为在祖先那边什么时候出现的。它的近亲有没有类似的叫法，哪一支上第一次冒出来。
1963 年，研究动物行为的尼可·廷贝亨把这四问写成一张清单，从此成了这一行的基本功。清单的用处不在于它有四格，而在于它逼你先说清楚：你现在答的是哪一格。
说句诚实的：天亮前那一阵为什么特别密集，到今天还有好几种解释在竞争——光线太暗不好找虫子、清晨空气传声传得远、这个时候还有力气唱恰好说明昨晚过得不错。第三格是四格里最难落实的一格，因为「说得通」太便宜了。
EN A bird starts singing outside your window before dawn, day after day, almost on schedule. Ask why, and you may be told the days are getting longer and its hormones are shifting; or that males who start earlier hold better territories. Those answers sound like rivals for the same slot. They are not. Four kinds of answer all hold at once. First, what is happening in its body right now: longer days, a cascade of hormones, and song-control regions of the brain that swell during the breeding season and shrink afterwards — first measured in canaries. Second, how this individual came to sing this way: who it heard as a chick, during which window, and how long it practised. Third, whether individuals who sing like this leave more offspring. This is the slot that goes wrong most easily. Not "it sings in order to attract a mate" — the ledger does not work that way. What actually happens is that males vary in when and how hard they sing, some of them hold better territories, pair up and fledge young, and their variants make up a slightly larger share of the next generation. Nobody wants anything; some simply leave more descendants. Fourth, when this behaviour appeared in the lineage, and whether close relatives do anything like it. Niko Tinbergen set these out as a checklist in 1963, and it became the basic discipline of the field. Its value is not the four boxes but the pressure it puts on you to say which box you are answering. One honest note: why the pre-dawn chorus is so dense is still contested — too dark to hunt insects, sound carries further in cool morning air, having energy to sing then advertises a good night. The third box is the hardest to pin down, because a plausible story is cheap.
白冠麻雀是北美常见的小鸟。把一只刚孵出的雄鸟单独关起来，一辈子不让它听见任何同类的歌，长大后它照样会唱。唱出来的东西含糊、简陋，可还是听得出是麻雀那一路，不会被当成别的鸟。所以「什么都没学到就唱不出来」是错的：它身上本来就带着一张很粗的草图。
关键的一步在后面。如果在它出生后大约第十天到第五十天之间，放同类的歌给它听，长大后它能唱得很像；错过这段时间再放，没用。而且它挑食——这段时间里放别的物种的录音，它基本不收。它带着的不只是草图，还有一把筛子，知道该往心里记哪一类声音。
更麻烦的还在后面。听过了也不一定行。1965 年有人做了这样一件事：让小鸟正常听过老师的歌，然后在它自己开口之前弄聋它。结果它唱不出来。它需要听见自己的声音，一遍一遍地跟记忆里那段比对着调。
那么请回答：这只鸟的歌，几成是天生的？模板是自带的，内容是听来的，调校靠的是听自己。这个问题问不出东西，就像问一块饼干的形状里面粉占了几成。1953 年有人把这层意思写成一篇很凶的批评文章，从此「天生对学来」这个二分在动物行为学里就没法再天真地用了。
还有一件事连窗口的形象也一并拆了。上面那个「第十到第五十天」，是用录音得出来的。1984 年有人换成活的老师——一只真鸟，关在旁边，看得见、能互动。结果五十天之后还学得会，甚至学会了另一个物种的歌。窗口不是一扇焊死的门，更像一条成本曲线：过了那一段，同样的输入得配上更强的社交信号才推得动。
EN The white-crowned sparrow is an ordinary North American bird. Raise a male alone so that he never hears another of his kind, and he still sings — a blurred, impoverished song, yet recognisably a sparrow's and not another species'. So "learns nothing, sings nothing" is wrong: he carries a rough sketch already. The decisive step comes next. Play him recordings of his own species between roughly day 10 and day 50 after hatching and he will sing well as an adult; play them outside that window and it does not take. And he is picky — recordings of other species during that window are largely ignored. He carries not just a sketch but a filter that knows which sounds to keep. Then it gets harder. Hearing is not enough. In 1965 someone let chicks hear a normal tutor and then deafened them before they first sang. They could not produce the song: they need to hear their own voice and tune it against the remembered model. Curiously, deafening after the song has crystallised changes little — by then the ears no longer matter. So: what fraction of this song is innate? The template is built in, the content is heard, the tuning runs on self-listening; remove any one and it fails. The question yields nothing, like asking what share of a biscuit's shape is flour. A fierce 1953 critique made the innate-versus-learned split unusable in its naive form. One more finding dissolves even the window. That day-10-to-50 figure came from tape tutors. In 1984 researchers used a live tutor — a real bird nearby, visible and interactive — and the birds learned past day 50, including the song of an alien species. The window is not a welded door but a cost curve: past that stretch, the same input needs a stronger social signal behind it.
1966 年有个设计得很漂亮的实验。给老鼠喝水，水分两种：一种是甜的；另一种一舔就触发闪灯和咔哒声，姑且叫「又亮又响的水」。喝完的后果也分两种：一半被弄得恶心，一半挨一次轻微的电击。
结果不是「都学会了」，而是交叉的。喝甜水被弄恶心的，从此躲甜味；喝甜水挨电击的，照喝不误。喝「又亮又响的水」挨电击的，从此躲那个灯和响声；喝「又亮又响的水」被弄恶心的，照喝不误。
同一只老鼠，同样是两件事凑一块，学不学得会取决于配的是哪两件。而且味道和恶心这一对，隔上几十分钟还能连上，一次就够。这在当时主流的学习理论里是不该发生的——那套理论认为什么和什么配都差不多，要紧的是配了多少次、隔了多久。
你身上也有这个。吃过一次坏掉的东西吐了，那个味道你能躲很多年；可你在某首歌响着的时候被吓了一跳，一般不会从此怕那首歌。
这不是谁替老鼠做了设计。能这样读：祖先里「什么容易和什么连上」的各种配置都出现过，其中把味道和肚子难受连得特别快的那些个体，在吃错东西这件事上少死了一点，于是这套配置在后代里占的比例慢慢高了上去。
印刻是同一件事的另一种长相。灰雁破壳后的头几个小时里，会认定看见的第一个大体合适、会动的东西并跟着走，通常是母亲，实验条件下也可能是人。这不是「学得特别快」，是「在一段很窄的时间里，只对很窄的一类输入敏感」。
把这两件摆在一起，「这是本能还是学来的」那个二分就彻底塌了。本能和学习不是两个抽屉。被留下来的从来不是某个具体动作，而是一整套「什么时候、对什么东西、学起来有多容易」的偏向。行为看着灵活，底下的偏食是硬的。
EN A neatly built 1966 experiment gave rats two kinds of water: one sweetened, the other rigged so that licking triggered a flashing lamp and a clicking relay — call it bright-noisy water. The consequence also came in two kinds: half the animals were made nauseous afterwards, half got a mild shock. The outcome was not "everyone learned"; it crossed over. Sweet water plus nausea, and the rat avoids the taste; sweet water plus shock, and it drinks on. Bright-noisy water plus shock, and it avoids the lamp and the click; bright-noisy water plus nausea, and it drinks on. Same animal, same pairing of two events — whether it can be learned depends on which two. And taste-with-nausea links up across tens of minutes, from a single trial. Under the dominant learning theory of the day that should not have happened: that theory held that any cue pairs with any consequence about equally, with repetitions and delay doing the work. You have this too. Throw up once after eating something spoiled and you can avoid that taste for years; get startled while a song is playing and you do not usually end up fearing the song. Nobody designed this for the rat. Read it this way: ancestors varied in which cues linked easily to which consequences, and those whose taste linked fast to gut illness died slightly less often from eating the wrong thing, so that configuration grew as a share of later generations. Imprinting is the same thing wearing another face. In its first hours out of the shell, a greylag gosling fixes on the first roughly suitable moving object and follows it — usually its mother, sometimes, in experiments, a person. That is not fast learning; it is narrow sensitivity inside a narrow stretch of time. Put the two together and the false binary collapses. Instinct and learning are not two drawers. What gets passed on is never a particular action but a whole set of biases about when, toward what, and how easily learning happens. Behaviour looks flexible; the fussiness underneath is hard.