黎明前的黑暗：AI的艰难探索 | Before Dawn: The Turbulent Path of AI

黎明前的黑暗：AI的艰难探索

Before Dawn: The Turbulent Path of AI

那是1997年的深秋，瑞士日内瓦。国际神经信息处理系统大会（NeurIPS）的会场里，一篇名为《长短期记忆》的论文正在被宣读。台下的听众中，有人窃窃私语，有人若有所思，更多的人则将信将疑。这篇由Sepp Hochreiter和Jürgen Schmidhuber合著的论文，提出了一种革命性的架构——长短期记忆网络（Long Short-Term Memory, LSTM）。在当时，没有人会想到，这个看似小众的研究成果，将在接下来的二十年间统治整个自然语言处理领域，成为机器翻译、语音识别、文本生成的核心引擎。更没有人会想到，当这个时代最终落幕时，它将被一个名为"Transformer"的架构所取代，而那场革命的种子，早在LSTM诞生的同一年就已埋下。

It was the autumn of 1997 in Geneva, Switzerland. Inside the conference hall of the Neural Information Processing Systems (NeurIPS) conference, a paper titled "Long Short-Term Memory" was being presented. Among the audience, some whispered to each other, others pondered deeply, and still more remained skeptical. This paper, co-authored by Sepp Hochreiter and Jürgen Schmidhuber, proposed a revolutionary architecture—the Long Short-Term Memory network (LSTM). At that time, no one could have predicted that this seemingly niche research would dominate the entire field of natural language processing for the next two decades, becoming the core engine for machine translation, speech recognition, and text generation. And certainly, no one anticipated that when this era eventually drew to a close, it would be supplanted by an architecture called the "Transformer"—a revolution whose seeds had already been planted in the very year LSTM was born.

当我们今天回望这段历史时，会发现1997年是一个奇妙的年份。就在同一年，一位名叫Kenneth Church的语言学家在一篇被广泛引用的论文中发出了著名的警告："统计火起来，语义学out了"（"Statisticalickos" / "Semantics is out"）。这句话既是对那个时代学术风向的精准概括，也是对即将到来的数据驱动范式的预言。然而，预言者自己恐怕也没有料到，这场统计革命的浪潮会如此汹涌，以至在二十多年后，当人们终于有能力将"统计"与"语义"重新融合时，整个领域已经沧海桑田。

When we look back at this period of history today, we find that 1997 was a remarkable year. In that same year, linguist Kenneth Church issued a famous warning in a widely cited paper: "Statisticalickos" / "Semantics is out." This phrase was an accurate characterization of the academic climate of that era and a prophecy of the data-driven paradigm to come. Yet the prophet himself likely never imagined how surging the wave of this statistical revolution would become—so much so that more than twenty years later, when researchers finally gained the ability to re-integrate "statistics" with "semantics," the entire field had undergone a transformation of monumental proportions.

人工智能的诞生：从符号到智慧的梦想

The Birth of AI: From Symbols to the Dream of Intelligence

要理解为什么LSTM和RNN会成为那个时代的主角，我们必须回到1956年的夏天。那个闷热的七月，在美国新罕布什尔州汉诺威镇的达特茅斯学院，八位年轻学者聚在一起，签署了一份后来被载入史册的提案。这份提案只有寥寥数页，却宣告了一门全新学科的诞生：人工智能（Artificial Intelligence）。

To understand why LSTM and RNN became the protagonists of that era, we must return to the summer of 1956. That sweltering July, eight young scholars gathered at Dartmouth College in Hanover, New Hampshire, and signed a proposal that would later be inscribed in history. This proposal, just a few pages long, proclaimed the birth of an entirely new discipline: Artificial Intelligence.

提案的起草者是约翰·麦卡锡（John McCarthy），一位年仅29岁的助理教授。他给这门新学科取了一个富有远见的名字——"人工智能"。在提案中，他写道：

The proposal's primary author was John McCarthy, a mere 29-year-old assistant professor. He gave this new discipline a visionary name—"Artificial Intelligence." In the proposal, he wrote:

> "我们将尝试找出如何让机器使用语言形成抽象概念，解决人类尚未解决的各类问题，并自我完善。我们相信，智能的某个或多个方面可以如此精确地描述，以至于可以用机器来模拟它。" > > "We propose that a 2-month, 10-man study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College in Hanover, New Hampshire. The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it."

这段话充满了那个时代特有的乐观主义精神。在麦卡锡和他的同行们看来，智能不过是一套可以形式化的规则系统，而人类的任务就是找出这些规则，然后用逻辑和推理将它们编码到计算机中。这种信念催生了人工智能的第一个范式——符号主义人工智能（Symbolic AI）。

This passage overflows with the optimism characteristic of that era. In the view of McCarthy and his contemporaries, intelligence was merely a system of formalizable rules, and humanity's task was simply to discover these rules and encode them into computers through logic and reasoning. This belief gave birth to the first paradigm of artificial intelligence—Symbolic AI.

符号主义的核心假设是：认知是一种符号操作。思想就像计算机中的数据，而推理就像对数据进行操作的程序。这个假设建立在坚实的基础上：数理逻辑已经相当成熟，形式系统已经被证明具有强大的表达能力，而香农的信息论则提供了处理符号的数学框架。更重要的是，这种观点与人类对自身智能的直观理解高度吻合——我们确实用语言思考，用概念推理，用逻辑论证。

The core assumption of symbolism was that cognition is symbolic manipulation. Thoughts are like data in a computer, and reasoning is like programs operating on that data. This assumption rested on solid foundations: mathematical logic was already quite mature, formal systems had proven capable of powerful expression, and Shannon's information theory provided a mathematical framework for processing symbols. More importantly, this view aligned closely with humanity's intuitive understanding of its own intelligence—we do think in language, reason through concepts, and argue through logic.

在达特茅斯会议之后的十年里，符号主义人工智能取得了令人振奋的进展。艾伦·纽维尔（Allen Newell）和赫伯特·西蒙（Herbert Simon）开发了"通用问题求解器"（General Problem Solver, GPS），这是第一个尝试模拟人类通用推理能力的程序。虽然GPS很快被发现能力有限，但它证明了复杂的问题解决可以通过明确的符号操作来实现。另一位先驱者司马贺（Herbert A. Simon）甚至大胆预测：计算机将在十年内成为国际象棋世界冠军。

During the decade following the Dartmouth conference, symbolic AI achieved exciting progress. Allen Newell and Herbert Simon developed the General Problem Solver (GPS), the first program attempting to simulate general human reasoning capabilities. Although GPS was soon found to have limited abilities, it demonstrated that complex problem-solving could be achieved through explicit symbolic manipulation. Another pioneer, Herbert A. Simon, boldly predicted that computers would become world chess champions within ten years.

更引人注目的是早期的机器翻译研究。1954年，IBM和乔治城大学联合展示了第一个机器翻译系统。尽管这个系统只能处理简单的俄语到英语的翻译，而且需要大量人工干预，但它激起了人们对全自动高质量翻译的憧憬。1966年，美国政府发布的ALPAC报告（自动语言处理咨询委员会报告）却给这股热情泼了一盆冷水。报告指出，机器翻译不仅速度慢、成本高，而且质量远不如人工翻译。这个结论直接导致了美国政府削减对机器翻译研究的资助，第一次人工智能寒冬就此降临。

Even more striking was early machine translation research. In 1954, IBM and Georgetown University jointly demonstrated the first machine translation system. Although this system could only handle simple Russian-to-English translation and required extensive human intervention, it sparked dreams of fully automatic high-quality translation. In 1966, the U.S. government released the ALPAC Report (Automatic Language Processing Advisory Committee Report), which poured cold water on this enthusiasm. The report concluded that machine translation was not only slow and costly but also far inferior to human translation. This conclusion directly led to cuts in U.S. government funding for machine translation research, and the first AI winter arrived.

专家系统：符号主义的辉煌与局限

Expert Systems: The Glory and Limitations of Symbolism

尽管遭遇了挫折，符号主义人工智能并没有就此沉沦。20世纪70年代末到80年代初，一个新的方向异军突起——专家系统（Expert Systems）。这种系统试图将人类专家的知识编码成规则，然后利用这些规则来解决特定领域的问题。

Despite setbacks, symbolic AI did not falter. In the late 1970s to early 1980s, a new direction emerged—Expert Systems. These systems attempted to encode human expert knowledge into rules and then use these rules to solve problems in specific domains.

MYCIN是专家系统最著名的先驱之一。这个由斯坦福大学在70年代开发的系统，用于辅助诊断细菌感染和推荐抗生素治疗。MYCIN的知识库包含约450条规则，能够根据患者的症状、化验结果和病史，推断出可能的致病菌，并给出相应的治疗建议。在一项著名的实验中，MYCIN的表现被与十位人类传染病专家进行比较。结果令人惊讶：MYCIN的表现与人类专家相当，甚至在某些方面更胜一筹。

MYCIN was one of the most famous pioneers of expert systems. Developed at Stanford University in the 1970s, this system assisted in diagnosing bacterial infections and recommending antibiotic treatments. MYCIN's knowledge base contained approximately 450 rules, capable of inferring likely pathogens and providing appropriate treatment recommendations based on patient symptoms, lab results, and medical history. In a famous experiment, MYCIN's performance was compared with ten human infectious disease specialists. The results were surprising: MYCIN performed comparably to—and in some aspects surpassed—human experts.

这个成功案例点燃了整个领域的热情。1980年，卡内基梅隆大学为数字设备公司（DEC）开发了XCON系统，这是一个用于配置计算机订单的专家系统。据估计，XCON每年为公司节省了数千万美元。IB