Can a Machine Think? 機器會思考嗎？

🔊The idea behind the test was three centuries old. “Reason,” wrote Thomas Hobbes in 1651, “is nothing but Reckoning (that is, Adding and Substracting)” of the consequences of names. If thinking is calculation, then anything that calculates thinks, and what it is made of is irrelevant. Leibniz, who built one of the first machines that could multiply, dreamed of a notation in which every dispute could be settled as accountants settle theirs: “Let us calculate.” The modern version of the idea is called functionalism. A mental state, it says, is defined by what it does, by its causes and effects within a system, in the way that a program is defined by its steps and can run on machines of quite different construction. On this view the mind stands to the brain as software stands to hardware, and a suitably programmed computer would not merely imitate thinking. It would think.

這個測驗背後的想法，已經有三個世紀之久。霍布斯在 1651 年寫道：「理性不過就是計算（也就是加與減）」，計算的是名稱之間的推論。如果思考就是計算，那麼任何會計算的東西都在思考，而它是什麼材料做的並不相干。萊布尼茲造出了最早能做乘法的機器之一；他夢想有一套記號，讓每一場爭論都可以像會計師解決帳目那樣解決：「我們來算一算吧。」這個想法的現代版本叫作功能論。它說，一個心理狀態是由它「做什麼」來界定的，也就是由它在一個系統之內的原因與結果來界定；就像一個程式是由它的步驟界定的，可以在構造完全不同的機器上執行。照這個看法，心靈之於腦，就像軟體之於硬體；而一台程式寫得恰當的電腦，就不只是在模仿思考。它是在思考。

🔊The opposite intuition is as old. The Daoist book Liezi tells how an artificer named Yan Shi presented King Mu of Zhou with a performer he had made. It walked, bowed, and looked about, “truly a human being.” When its chin was touched it sang in tune, and when its hand was lifted it danced in time. The king took it for a real person, until at the end of the show it winked at his concubines. He was furious and would have had the maker executed on the spot. Yan Shi at once took the figure apart, and it proved to be an assembly of leather, wood, glue, and lacquer, colored white, black, red, and blue. Every organ was there. The king removed the heart, and the mouth could not speak. He removed the liver, and the eyes could not see. Delighted, he asked: “Can human skill really equal the work of the Maker of things?” The story records the two reactions people still have. While the performance lasts we see a person, and when the case is opened we see parts.

相反的直覺一樣古老。道家的《列子》記載，一位名叫偃師的工匠，把自己造的一個倡者獻給周穆王。它行走、俯仰、顧盼，「信人也」。碰它的下巴，它就合著音律唱歌；抬它的手，它就應著節拍起舞。穆王把它當成真人，直到表演快結束時，它對王的侍妾拋了個眼色。王大怒，當場就要殺掉偃師。偃師立刻把那個人形拆開；原來它是皮革、木頭、膠、漆拼合而成的，塗著白、黑、紅、青的顏色。每一個器官都齊全。王拿掉它的心，口就不能說話；拿掉它的肝，眼就不能看。王高興了，問道：「人之巧乃可與造化者同功乎？」這個故事記下了人們至今仍有的兩種反應。表演進行的時候，我們看見一個人；外殼打開的時候，我們看見零件。

🔊Leibniz gave the second reaction the form of an argument, and he gave it against his own machines. Suppose, he wrote in 1714, that there were a machine “so constructed as to think, feel, and have perception.” Imagine it enlarged until one could walk into it “as into a mill.” Inside “we should, on examining its interior, find only parts which work one upon another, and never anything by which to explain a perception.” A century later Ada Lovelace, describing the Analytical Engine that Charles Babbage had designed, added a different limit. The engine, she wrote, “has no pretensions whatever to originate anything. It can do whatever we know how to order it to perform.” Leibniz’s objection concerns experience and Lovelace’s concerns originality. They have been confused with each other ever since.

萊布尼茲把第二種反應寫成了論證，而且他是對著自己造的機器提出的。他在 1714 年寫道：假設有一台機器，「被造得能夠思考、感受、擁有知覺」。想像它被放大，大到人可以走進去，「像走進一座磨坊」。在裡面，「我們檢視它的內部，只會發現一個推動一個的零件，永遠找不到任何可以用來解釋一個知覺的東西」。一個世紀之後，勒芙蕾絲在描述巴貝奇設計的分析機時，加上了另一種限制。她寫道，這台機器「絲毫沒有『創始』任何事物的自命。凡是我們知道怎麼命令它去做的事，它都能做」。萊布尼茲的反駁關乎經驗，勒芙蕾絲的反駁關乎原創。此後這兩者一直被混在一起。

🔊In 1980 the American philosopher John Searle turned the mill into a room and put himself inside it. He imagined a program that passes the Turing test in Chinese, a language of which he knew not a word. He is locked in a room with baskets of Chinese characters and a rule book in English. Slips of paper covered in characters are pushed under the door. He looks up their shapes in the book, which tells him which characters to push back. To the people outside, the replies are those of a fluent speaker. Searle has done everything the computer does: he has run the program. And he understands nothing. He does not know whether the slips are about the weather or about his mother. His conclusion is that carrying out a program, however good, is not sufficient for understanding. A program is defined by its syntax, the rules for moving symbols according to their shapes. Understanding requires semantics, knowing what the symbols mean, and no amount of shuffling will supply it.

1980 年，美國哲學家瑟爾把那座磨坊變成了一個房間，並且把自己放了進去。他設想有一個程式，用中文通過了圖靈測驗；而中文他一個字也不懂。他被鎖在一個房間裡，裡面有一籃一籃的中文字，以及一本用英文寫的規則書。寫滿了字的紙條從門縫底下塞進來。他在書裡查那些字的形狀，書告訴他該把哪些字塞回去。對外面的人來說，那些回覆是一個流利的說話者寫的。瑟爾做了電腦所做的一切：他執行了那個程式。而他什麼也不懂。他不知道那些紙條談的是天氣，還是他的母親。他的結論是：執行一個程式，不管程式多好，並不足以構成理解。程式是由它的「語法」界定的，也就是依照符號的形狀搬動符號的規則。理解需要「語意」，也就是知道那些符號是什麼意思；而不管怎麼搬來搬去，都搬不出語意來。

🔊Searle chose Chinese because he could not read it. A reader who can should substitute a language that is closed to him, and the exercise below does so.

瑟爾選中文，是因為他看不懂。看得懂中文的讀者，應該換成一種對自己關閉的語言；下面的練習就是這樣做的。

🔊The paper drew more replies than almost any other in the field. The systems reply grants that the man does not understand and says that he is only one component: the understanding belongs to the whole system of man, book, and baskets, as it belongs to a brain and to no single neuron. Searle answered that he could memorize the book and do the work in his head, and he would still understand nothing. The robot reply says that the symbols mean nothing because they are connected to nothing: give the system cameras and limbs, so that its word for “tea” is caused by tea, and meaning will follow. Searle answered that the camera’s signals would reach him as more symbols. The brain simulator reply imagines a program that copies, neuron by neuron, the brain of a Chinese speaker. Searle answered that a simulation of a storm leaves everyone dry. Each exchange ends in a clash of intuitions about the same point, and no decisive move has been found by either side.

這篇論文招來的回應，幾乎比這個領域裡任何別的論文都多。「系統回應」承認那個人不懂，並且說他只是一個零件：理解屬於由人、書、籃子組成的整個系統，正如理解屬於一個腦，而不屬於任何單一的神經元。瑟爾回答：他可以把整本書背下來，在腦子裡做這件工作，而他仍然什麼也不懂。「機器人回應」說，那些符號沒有意義，是因為它們什麼也沒連上：給這個系統裝上攝影機與肢體，讓它那個代表「茶」的詞是由茶引起的，意義就會隨之而來。瑟爾回答：攝影機的訊號到了他這裡，只是更多的符號。「腦模擬回應」設想一個程式，一個神經元一個神經元地複製一位中文使用者的腦。瑟爾回答：模擬一場暴風雨，誰也不會淋濕。每一回合都以對同一點的直覺衝突收場，而雙方都還沒有找到決定性的一著。

🔊For seventy years the debate concerned imaginary machines. It now concerns real ones. Systems trained on enormous quantities of human writing carry on conversations that most people cannot tell from a person’s, and they were built by neither of the methods the philosophers had in mind. Nobody wrote their rule book. They adjust billions of numerical settings until they predict text well, and what goes on inside them is only partly understood by the people who make them. One camp sees the Chinese room come true: a system that has learned which strings of words follow which, with no access to the world the words are about. Another replies that to predict what a physicist will say next one must model a good deal of physics, and that understanding may come in degrees and consist in such abilities. Lovelace’s limit is equally unsettled, since these systems produce things nobody ordered them to produce.

七十年來，這場辯論談的是想像中的機器。現在談的是真實的機器。用巨量的人類文字訓練出來的系統，能進行多數人分辨不出是不是真人的對話；而它們並不是用哲學家心裡想的那兩種方法造出來的。沒有人替它們寫規則書。它們調整數十億個數值設定，直到能把文字預測得很好；而它們內部發生的事，連製造它們的人也只懂一部分。有一派看到的是中文房間成真了：一個系統學會了哪些字串跟在哪些字串後面，卻接觸不到這些字所談的那個世界。另一派回答：要預測一位物理學家接下來會說什麼，就必須對不少物理學建立模型；而且，理解也許有程度之分，也許就是由這類能力構成的。勒芙蕾絲的那條界線同樣懸而未決，因為這些系統會產出沒有人命令它們產出的東西。

🔊Turing’s second prediction, at least, has come true. People speak of machines thinking, and few bother to contradict them. Whether the usage records a discovery or only a habit is the question he set aside in 1950. It helps to keep three things apart. A machine may match us at tasks, and that can be measured. It may or may not understand, and here the room is still the best test of what we mean by the word. And it may or may not feel, which was the subject of the last lesson. “We can only see a short distance ahead,” Turing ended his paper, “but we can see plenty there that needs to be done.”

至少，圖靈的第二個預測成真了。人們談論機器在思考，而很少有人費事去反駁。這個用法記錄的是一項發現，還是只是一種習慣？這正是他在 1950 年擱到一邊的那個問題。把三件事分開會有幫助。一台機器在各種任務上可能跟我們並駕齊驅，這是量得出來的。它可能懂、也可能不懂；在這一點上，那個房間仍然是檢驗「我們說『懂』是什麼意思」的最好辦法。它可能有感受、也可能沒有；那是上一課的主題。圖靈在論文的結尾寫道：「我們只能看到前方不遠的地方，但我們看得到那裡有許多需要做的事。」