深蓝击败国际象棋世界冠军

第一局：人。第二局：机器。第三至第五局：和。第六局：机器。总比分 3.5 比 2.5。

1997 年 5 月纽约的六局重赛，可以压缩成这样一行。最后一局落在 5 月 11 日。它使 IBM 深蓝（Deep Blue）成为第一个在标准比赛时限下、以整场比赛击败在位国际象棋世界冠军的计算机系统。加里·卡斯帕罗夫一年前曾在费城以 4 比 2 战胜深蓝的前身；重赛带着明确的旧账，IBM 团队有一年时间重新设计。六局采用标准时限，结果计入公开人机对抗史；它不是快棋表演，也不是让一方使用互联网助手的表演赛。赛后 IBM 将系统退役，比分被固定为可引用的历史结果，而不是持续迭代的在线棋力榜。

IBM 团队扩充了残局数据库，改进局面评价函数，邀请更多特级大师参与调校棋类知识。机器以 32 个处理器并行工作，配合专用棋类芯片，每秒可评估约两亿个局面——这是厂商与技术论文中反复给出的数量级（厂商自报的吞吐，与真实对局中的有效搜索深度仍需结合剪枝与评价函数理解）。这个数字很大，却不能自行下棋。搜索必须知道哪些分支值得深入，评价函数必须把王的安全、子力、空间和位置等因素变成可比较的值，开局库与残局库则把长期积累的棋类知识带进计算。深蓝的力量由速度、专用芯片、搜索方法、数据库和人的棋艺共同组成。赛后发表在 Artificial Intelligence 等刊物上的技术文章，把这些组件拆开说明，而不是只留下“电脑赢了”一句。

“蛮力”因此是个不完整的描述。没有两亿次评估，系统达不到同样深度；没有人设计的评价与剪枝，更多计算也会浪费在庞大的博弈树里。机器不像人那样看棋，也不会盲目遍历全部合法未来。它把一种高度结构化的任务拆到适合大规模计算的程度：状态空间虽大，规则明确，胜负可判定，评价可调校。

六局的顺序扩大了比赛的象征性。卡斯帕罗夫先胜，深蓝追平，随后三和，最后一局决定整场结果。公众看到的是一项被视为高度智力的活动出现清楚的负分；IBM 展示的则是一项计算与领域工程。两种叙事在同一张比分表上重叠，却并不等价。赛后关于是否应允许调参、是否应公布日志的争论，也说明“正式比赛条件”本身仍可被继续定义——时限、是否允许赛间修改、日志是否公开，都会改变公众读到的故事。

深蓝不会把棋盘上的能力迁移到语言、医学或日常判断。它不通过这六局获得通用学习能力，也没有因为赢棋而理解胜负对人的意义。比赛划出的边界恰好有两面：复杂的智力任务可以被专用系统攻克；攻克一个任务仍不足以说明机器拥有一种不受任务约束的智能。深蓝退役后，计算机在国际象棋上的统治地位继续加深，但那是棋类与搜索的故事线，不是通用智能的证书。

第六局之后，比分不再增长。那台专门为国际象棋建造的机器留下了一份可拆解的胜利：每秒评估多少局面，搜索剪掉哪些分支，人类棋谱和评估函数怎样进入计算，能力又在哪里随棋盘一起结束。后来每一次“机器超过人”的宣告，都可以先放回这样一张明细表里，看清楚赢的是哪一项任务、在什么规则与硬件之下。明细表填不满的地方，才是口号开始生长的地方。

Game 1: human. Game 2: machine. Games 3–5: draws. Game 6: machine. Final score 3½–2½.

The May 1997 rematch in New York compresses into that line. Game six fell on 11 May. It made IBM’s Deep Blue the first computer system to defeat a sitting world chess champion in a match under standard time controls. A year earlier in Philadelphia, Garry Kasparov had beaten Deep Blue’s predecessor 4–2; the rematch carried that ledger, and IBM’s team had a year to redesign. Six games at standard time entered the public record of human–machine play—not a blitz exhibition, not a show where one side used internet help. Afterward IBM retired the system. The score froze as a citable historical result rather than a living online rating board.

The team enlarged endgame databases, refined the evaluation function, and brought more grandmasters into tuning chess knowledge. The machine ran on 32 processors with specialized chess chips, evaluating on the order of two hundred million positions per second—the magnitude repeated in vendor and technical papers (vendor-reported throughput still needs pruning and evaluation to become effective search depth). The number is large and still cannot play chess alone. Search must know which branches deserve depth; evaluation must turn king safety, material, space, and position into comparable values; opening and endgame books bring accumulated chess knowledge into the computation. Deep Blue’s strength was speed, chips, search method, databases, and human chess craft together. Technical articles later published in venues such as Artificial Intelligence take those components apart rather than leave only “the computer won.”

“Brute force” is therefore an incomplete description. Without two hundred million evaluations the same depth is unreachable; without designed evaluation and pruning, more compute still wastes itself in a vast game tree. The machine does not see the board as a person does, and it does not blindly traverse every legal future. It decomposes a highly structured task to the scale large computation can use: a large state space, clear rules, decidable outcomes, tunable evaluation.

The order of the six games enlarged the match’s symbolism. Kasparov won first; Deep Blue equalized; three draws followed; game six decided the match. The public saw a clear deficit in an activity treated as high intelligence; IBM showed computation and domain engineering. Two narratives share one scoreboard without being equivalent. Post-match arguments over mid-match tuning and log publication also show that “official match conditions” remain open to definition—time controls, whether changes between games are allowed, whether logs are public, all change the story the public can read.

Deep Blue does not transfer board strength to language, medicine, or everyday judgment. It does not acquire general learning from these six games, and it does not understand what winning means to people. The boundary the match draws has two faces: a complex intellectual task can fall to a specialized system; winning one task still does not show a form of intelligence free of task bounds. After retirement, computer dominance in chess only deepened—that is the storyline of chess and search, not a certificate of general intelligence.

After game six the score stopped growing. The machine built for chess left a disassemblable victory: how many positions per second, which branches search cut, how human books and evaluation entered the calculation, and where capability ended with the board. Every later announcement that “machines surpassed people” can first be returned to such a ledger—which task, under which rules and hardware. Where the ledger cannot be filled, slogans begin to grow.