达特茅斯暑期研究项目

1955 年 8 月 31 日的提案很短，也很昂扬：两个月，十个人，下一年夏天，在新罕布什尔州汉诺威的达特茅斯学院。预算列到铁路票、房租、秘书费和复印费，总计 13,500 美元。目标却没有同样收敛——让机器使用语言、形成抽象与概念、解决当时留给人的问题，并改进自己。

四位发起人来自不同位置。约翰·麦卡锡在达特茅斯教数学；马文·明斯基研究神经网络与脑模型；纳撒尼尔·罗切斯特参与过 IBM 701 的设计；克劳德·香农已经建立信息论，并把逻辑带进开关电路。提案抬头列出他们当时的机构身份，也标明这是一份寻求暑期资助的研究计划，而不是已经完成的学科章程。提案没有假装这些路线已经汇成一套理论。它逐项列出自动计算机、语言如何用于计算机、神经元网络、计算规模的理论、自我改进、抽象、随机性与创造性，像把散落各处的未解题临时装订在一起。

装订线就是一个新名称：Artificial Intelligence。

这个词最初承担的不是答案，而是召集。提案的共同猜想写得很清楚：学习的各个方面以及智能的其他特征，原则上都能得到足够精确的描述，从而由机器模拟。语气里的乐观不难看见：作者们相信，一群经过挑选的科学家共同工作一个夏天，就能在一个或多个问题上取得显著进展。但提案正文也留下了比口号更清醒的句子。它承认当时计算机的速度和存储可能不够，却把主要障碍指向人们尚不会充分利用已有机器；它讨论穷举为何低效，也承认需要一套衡量计算复杂度的理论——在复杂性理论尚未成形的年代，这是一种先见的不满。

后来被称为“达特茅斯会议”的 1956 年夏天，并不是十位奠基者连续两个月坐在同一张桌旁。参与者分批来访，小组与个人工作交错，也没有在散会时交出统一的智能理论。纽厄尔与西蒙带来的 Logic Theorist 成果，其主要工作在会前已经展开；提案里预想的“十人共度两个月”与实际到访节奏并不重合。若把这一切压缩成一次庄严的诞生仪式，反而看不清它真正完成了什么。

它完成的是行政与想象的基础设施。一个名称让控制论、符号推理、语言研究、神经模型与自动机理论之间产生可见的邻接；以后可以围绕它设实验室、申请经费、办会议，也可以反对它、宣布它失败，再在多年后重新解释它。1956 年夏天没有交出统一理论，却交出了可引用的议程标题——对资助机构和后来的院系建制来说，标题往往比未完成的理论更先起作用。麦卡锡后来发明 Lisp，创办斯坦福人工智能实验室；明斯基在 MIT 推动 AI 研究；香农的信息论与开关理论继续为计算提供底层语言。这些后续各有自己的时间线，并不都从汉诺威的同一间屋子同时出发。学科并非在那个夏天被造好。更准确地说，一群尚未达成一致的人，获得了一个可以长期争夺的共同名词。

提案文本本身值得当作史料细读，而不是只摘一句名言。它把“自动计算机”“编程语言能否表达启发式”“神经网络”“计算规模”“自我改进”“抽象与感官信息”“随机性与创造性”并列，说明 1955 年的作者们并不共享单一技术路径，却愿意共享一张问题清单。洛克菲勒基金会等资助方看到的，是一份预算清楚、时间短、人员有限的暑期项目；历史留下的，却是比预算大得多的名称遗产。

提案预估了两个月。名称此后存活了数十年，边界仍在移动：有时被等同于符号搜索，有时被等同于神经网络，有时被媒体缩成聊天机器人，有时被政策文件写成需要监管的通用技术。青年式的时间表几乎立刻失败；临时的装订却持续有效——那些仍拒绝待在单一学科里的问题，至今还在同一个词下面被争论。命名没有结束争论，它只是给争论一个可以反复返回的地址。

The 1955 proposal itemized train travel, rent, secretarial help, and duplicating costs. The total request was $13,500. Against that modest ledger stood an agenda with almost no modesty at all: language, abstraction, concepts, self-improvement, problems then reserved for human beings. Ten researchers, two months, the following summer in Hanover, New Hampshire, at Dartmouth College. The document is dated 31 August 1955.

John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon did not pretend to possess a common theory. The proposal’s letterhead lists their institutional affiliations of the time and marks the document as a request for summer research support, not a finished constitution for a discipline. Their own work approached the subject from mathematics, neural models, computer engineering, information theory, and switching circuits. The proposal likewise reads less like a manifesto for one method than a bundle of difficult headings: automatic computers, how language might be used with them, neuron nets, the size of a calculation, self-improvement, abstraction, randomness, creativity. The questions were stitched together before their answers were.

The stitch was a new phrase: artificial intelligence.

The new phrase assembled an agenda before anyone could explain the whole of it. The authors grounded the project in a shared conjecture—that every aspect of learning, and other features of intelligence, could in principle be described precisely enough for a machine to simulate it. Their confidence is unmistakable. A carefully selected group, they wrote, could make significant advances on one or more problems during a summer. Yet the document also contains a more sober technical instinct. Existing machines might be too slow or too small, but the authors suspected that ignorance about how to use them was a larger obstacle. Exhaustive search already looked inadequate. A theory of computational complexity was still missing—and the proposal already treated its absence as a problem worth naming.

The summer later remembered as “the Dartmouth conference” was not one unbroken two-month council of ten founders. Participants arrived at different times. Individual and small-group work overlapped. Logic Theorist had already been advancing in the hands of Newell and Simon before the gathering; the proposal’s picture of ten people sharing two continuous months did not match the actual visiting schedule. No unified account of intelligence was carried out of the building when the summer ended.

Calling the event a birth ceremony makes for an orderly origin story, but it understates the peculiar power of naming. The summer of 1956 did not deliver a unified theory; it did deliver a citable agenda title—and for funders and later departmental structures, titles often act before unfinished theories do. The phrase gave neighboring problems in cybernetics, symbolic reasoning, language, automata, and neural modeling an address. Laboratories could be founded around it; grants could invoke it; conferences could dispute its boundaries. McCarthy later developed Lisp and founded the Stanford AI Laboratory; Minsky helped organize AI research at MIT; Shannon’s information theory and switching work continued to supply a lower language for computation. Those careers have their own timelines; they did not all depart from one Hanover room at once. A field need not agree on its object before institutions begin to form around the disagreement.

The proposal text itself is worth reading as a source rather than as a single quoted slogan. It places side by side automatic computers, whether programming languages can express heuristics, neural nets, the size of calculation, self-improvement, abstraction from sensory information, randomness, and creativity. The 1955 authors did not share one technical path; they were willing to share one problem list. Funders saw a summer project with a clear budget, a short horizon, and a limited cast. History kept a name larger than the budget.

The proposal allowed two months. The name has now survived for decades without acquiring a stable frontier—sometimes identified with symbolic search, sometimes with neural nets, sometimes reduced in the press to chatbots, sometimes written into policy as a general technology to regulate. Its youthful schedule failed almost immediately, while its temporary binding held: questions that still refuse to stay in one discipline continue to be argued over under the same name. Naming did not end the argument. It gave the argument an address to which participants could keep returning—sometimes to defend the field, sometimes to declare it exhausted, sometimes only to rebrand the same problems under a safer phrase.