视频库 / NO.123ASK THE BIG QUESTIONS
视频库 / NO.123
字幕 字幕位置
--:--
点击播放,这里会跟随视频显示当前句的中英字幕。

The Creativity Code - Marcus du Sautoy

节目发布 2019-06-03 · Oxford Mathematics
马库斯·杜·索托伊
本期追问 · 点击跳到视频对应位置
8:07 AlphaGo 的第 37 手,真能算作机器的创造性行为吗?19:25 知道一幅画出自机器后,先前的感动就作废了吗?15:16 程序员说不清算法怎么得出结果时,作品还归功于人吗?59:52 AI 是人类创造力的威胁,还是把人推出定式的工具?
归入 Ⅲ·11 艺术需要作者吗? →
EDITED TRANSCRIPT · 依据现场录音编译整理,可划线生成便签
编者按:本文整理自牛津大学数学教授、西蒙尼公众科学理解教授、皇家学会院士马库斯·杜·索托伊(Marcus du Sautoy)的一场公开演讲,主题围绕他的著作《创造力密码》(The Creativity Code)展开:从 AlphaGo 的第 37 手谈起,逐一检验人工智能在绘画、诗歌、小说与音乐上的表现,并在现场邀请听众用投票卡做了一系列「人还是机器」的判别测试。全文依据现场录音编译整理,仅删去口语枝节与现场调度的技术性对话,论证、例子与转折悉数保留。

数学家会被电脑取代吗

在座有多少位数学家?请举手让我看看。好,那么各位数学家大概都有和我一样的经历:去参加聚会,别人问你是做什么的,我现在已经开始撒谎了,说自己是国际间谍之类的。因为大多数时候,我得到的反应就那么几种。一种是对方立刻逃到聚会的另一头,把我晾在原地;不过临走前,他们总要告诉我自己当年会考数学考了几分,我一直搞不懂这是什么心理。要是他们肯留下来,往往会说这么一句:「拜托,你这行现在早该被电脑取代了吧?」

我想这主要是因为人们以为,我在办公室里干的事就是把长除法算到小数点后很多位。假如真是那样,我的电脑早就让我失业了。但你们知道,正如刚才主持人所说,我们现在多少都感到有些威胁:人工智能步步逼近,看起来无所不能,做着各种有趣的事。既然电脑本来就是逻辑和数学的产物,我这份工作难道不该第一个受到威胁吗?

九十年代「深蓝」(Deep Blue)击败卡斯帕罗夫的时候,我就经常听到这种话。人们习惯把下棋和做数学相提并论:棋子有一套合乎逻辑的走法,你追求一个终局,证明的终点是「证毕」(QED),赢下棋局就像赢下证明。所以九十年代很多人对我说:下一个就轮到你了。

围棋:数学家的最后护盾

可我从来没有真的被国际象棋吓到。我们数学家一直握着另一种游戏当护身符,用来抵挡「电脑能做数学」的说法。那就是古老的围棋,起源于中国,在十九乘十九的棋盘上落黑白子,你要抢在对方围住你之前围住对方的地盘。这种游戏的复杂度极高,远远超过国际象棋,需要大量的模式识别,而这恰恰是数学的核心:发现隐藏的模式。下围棋,尤其是做数学,要想知道自己往哪儿走,需要很多直觉和创造力,你往往说不清为什么要走这一步。你在这个世界里泡得久了,慢慢就培养出一种感觉。

计算机科学的课堂上历来有个传统说法:国际象棋可以自动化,因为我们能理解每一步的逻辑后果,能顺着可能性的树往下走;围棋则是电脑永远下不了的游戏。事实也是如此,凡是试图把人类下围棋的方式编成程序的尝试,全都失败了,那些算法连业余棋手都赢不了。所以我感觉相当安全:计算机科学家说电脑下不了围棋,那它当然也下不了数学这盘更复杂的棋。

几年前我受到了不小的震动,这个故事你们大概都知道。伦敦的一个团队宣布,他们的算法不但能在高水平上对弈,而且能在最高水平上对弈。这就是 DeepMind。戴密斯·哈萨比斯(Demis Hassabis)当年在剑桥读书时,被告知了那条老规矩:电脑下不了围棋。这对戴密斯来说等于是红布挑逗公牛。他后来创办了这家公司,设计出一套自认为能与顶尖棋手抗衡的算法,然后向韩国九段棋手李世石发起挑战。李世石对这个算法完全不屑一顾,认为它根本不可能接近自己的水平,放话要五比零把它碾碎。

比赛五局。我坐在 YouTube 前面一局不落地追看,因为我意识到自己的饭碗恐怕真的危险了。我眼看着李世石一局比一局沮丧:第一局输了,第二局输了,第三局又输,三局过后整场比赛已经输掉。他赢了第四局,那一局至今被他视为自己一生中最伟大的一局,因为他终究在一局里战胜了这个算法。第五局他又输了,最终一比四。

自下而上:会学习的代码

这两年间发生了什么变化?变的是写代码的方式。一种新式代码登场了,能做过去的代码做不到的事,这也是我们牛津非常关注的方向:深度学习,或者叫机器学习。

过去的代码是自上而下写出来的。你必须事先知道程序要如何运行,把游戏规则告诉它,你得理解整个场景,机器只负责执行。它当然会下国际象棋,因为你叫它执行这件事,它能算得更深,分析更多局面,可归根结底还是人类在告诉程序做什么。现在不同了,代码是自下而上写出来的,它像孩子一样学习。过去的情形好比父母的 DNA 生出一个孩子,但孩子永远拴在父母的 DNA 上,不会学习新东西。现在突然有了一种代码,能在遇到新环境时自我调整、变异、重设参数,仿佛有一层元代码(meta code)在指挥代码:如果做错了,就这样改。DeepMind 正是用这种方法训练 AlphaGo 的:通过对弈、通过失败来学习。

当然,他们先拿一些简单的游戏练手,从雅达利(Atari)游戏开始,那正是我小时候沉迷的东西。我最喜欢的一款叫「打砖块」(Breakout):一个小球上下弹跳,你操控一块挡板把砖块打掉得分,蓝砖一分,越往上分数越高。机器得到的只有屏幕上的像素和分数,没人告诉它要去接球。它一开始只是随机移动挡板,每次碰到球就看见分数上涨,于是重设参数,转而优先往球飞来的方向移动。

我和朋友玩这个游戏时,发现了一个绝妙的窍门,得意得不得了:在左边凿出一条隧道,把球送上去,之后你就什么都不用干了,球会自己在天花板和砖块之间来回弹。让我目瞪口呆的是,DeepMind 学会玩这个游戏时,居然学出了同一个窍门。不只是人类,电脑也懒,不想来回挪挡板,所以把球送上去了事。这实在惊人:六百局之后,它仅凭随机移动挡板、观察哪些动作能让分数涨得更快,就自己学会了这个窍门。

它下围棋也是这个路子。先把互联网上大量的人类对局拿来,学习人类是怎么赢、怎么输的,这是它最初的学习材料。然后它开始制造合成数据:自己和自己下,不同版本的自己互相对弈。输了一局,它就分析是哪些步导致自己没能达到高水平,然后重设参数。这样过了一阵,它就到了可以挑战李世石的高度。

第 37 手与 AI 的创造性

乍一看你会说,好吧,电脑变得很强,能把局面算得很深,如此而已。但我认为这几局棋里发生了更了不起的事。

第一局很有意思。李世石想,既然它是靠学习人类棋谱练出来的,那么打败它的最好办法也许是下得不像人。于是他下了一盘非常搅局的棋。可 AlphaGo 足够聪明,从容应对了他的招数,那些招数结果反倒成了弱手,他输了,因为他没有下自己的正常棋路。所以第二局李世石决定下自己熟悉的、标准的高水平棋。

围棋老师很早就会教你几条策略,其中一条是开局要沿着棋盘边缘走,落在第一到第四线上,因为开局阶段是边角与内部的争夺,如果太早往棋盘中间落子,会被视为弱手,因为那时你还没有确立要紧的地盘。第 36 手下完,李世石觉得需要抽根烟,上了酒店顶楼。AlphaGo 不需要抽烟来提神,它坐在那儿想了一会儿,然后让人类代理落子。这场比赛考的不是机器人技术,纯粹考的是思维,所以棋盘旁还是有个人替它落子,至今对 AI 来说,捡起一颗棋子放到棋盘上仍是件挺难的事。它让人类把一颗子放在了第五线上。我已经在把这个 AI 拟人化了。AlphaGo 执黑,它把这颗黑子放在第五线。我记得 YouTube 上所有的解说员都倒吸一口气,惊呼:AlphaGo 犯了大错,开局绝不该走这种棋,李世石要赢了,比分要一比一了。所有人都很笃定。

李世石抽完烟回到座位,看了看 AlphaGo 的这一手。你们真该去看那段录像,非常好笑,他完全不敢相信这个 AI 居然下出这么蠢的一步。但他比旁人更多疑:它为什么走这一步?直到今天他还在问,它为什么走这一步?

国际象棋和围棋有个很不一样的地方:国际象棋越下越简单,因为棋子不断被吃掉;围棋越下越复杂,因为棋子不断被添上去。随着棋局展开,越来越多的子落在棋盘上,地盘从右下角一路往上延伸。结果第 37 手的那颗黑子,正好决定了那片地盘归 AlphaGo 而不归李世石。那是极富灵感的一手,第 37 手把子放在那儿、打破人类对这盘棋的固有理解的这个决定,为 AlphaGo 赢下了第二局。我为此非常兴奋,因为我相信,这正是我们应当称为人工智能的创造性行为的例子。

Boden 三要素与局部最大值

过去几年我在皇家学会的一个委员会里工作,研究机器学习在未来十年会给社会带来什么影响。戴密斯也在这个委员会里,还有一位哲学家玛格丽特·博登(Margaret Boden)。我和她谈过很多关于创造力的问题。她一直很关心她所谓的「铁皮罐头」(电脑)能做什么,以及它们能否有创造力。她给创造力下了一个很好的工作定义。我不确定这是最好的定义,关于创造力究竟指什么,哲学上有无数争论,但今晚拿它当工作定义足够好用。

创造力首先要求「新」(new)。电脑很容易做出新东西,而且新不新可以客观判断。但还有另外两个性质:它得带有「惊讶」(surprise),这就主观一些了;它还得有「价值」(value),这同样相当主观。所以电脑若想具备人类认可的创造力,就得学会理解我们觉得什么令人惊讶、什么有价值。

游戏的妙处在于,这些性质可以很快得到裁决。惊讶?有,所有解说员都惊呼它犯了错误。价值?有,这一手为 AlphaGo 赢下了一局。于是我感兴趣的是:如果它能在游戏这样封闭的环境里有创造力,那它还能在哪里有创造力?能在数学里吗?我和戴密斯坐在一起时开过一个玩笑,我们俩刚一起当选皇家学会院士,我说:你能不能让 AlphaGo 也当上院士?我书里的一部分故事就源于此:戴密斯说,我们已经在干了。DeepMind 确实在研究如何做出一个有创造力的 AI 数学家。

这种新型 AI 让我兴奋的地方在于,以 AlphaGo 为例,它不仅棋下得好,还教会了我们一种全新的下法。我们原以为在既有规则下,比如那些关于开局落在第几线的讲究,人类的棋艺已经登顶,存在一种最优下法。AlphaGo 在这几局里向我们展示:我们自以为站在顶峰,其实只是站在数学家所说的「局部最大值」(local maximum)上。那不过是斯诺登山,而远处还有一座珠穆朗玛峰,有一种高得多的下法。AlphaGo 靠冒险试探,走下了这道适应谷地,找到了更好的路。现在 AlphaGo 教会了我们新的策略,帮助我们把棋下到高得多的水平。

所以这本《创造力密码》要走的旅程就是:这种通过与周围数字世界互动来学习的 AI,除了在游戏里,能不能在别的领域也有创造力?

从 Lovelace 告诫到 Lovelace 测试

最早构想代码这个概念的人之一,就已经在设想代码可以做艺术性的事。我们每年庆祝「埃达·洛夫莱斯日」(Ada Lovelace Day)。洛夫莱斯的母亲喜欢让她接触各种科学思想,带她去看巴贝奇(Babbage)的分析机。她一看到这台机器,就意识到它能做的远不止长除法和乘法,还能做更令人兴奋的事,于是她开始写代码,让分析机做些有趣的事情。这就是我们纪念她的原因:她为一篇关于分析机的论文所写的注记,被视为「用代码让机器做有趣的事」这一想法的源头。

她当时就在琢磨,这台机器也许能做比科学计算有趣得多的事。她写道,这台机器或许能谱写出复杂精妙、任意规模的科学音乐作品。她已经想到了音乐,而音乐和数学的联系本来就很深,都是关于模式的,让机器把模式跑出来,也许就能造出音乐。AI 能否作曲这个挑战,我们稍后再谈。

但她也留下了一句告诫:要提防对分析机能力产生夸大的想象,它绝无「原创」任何东西的能力,只能做我们命令它做的事。我想这正是我们过去一贯的看法:自上而下的编程,是人在告诉电脑做什么,所以如果电脑显得有创造力,那是因为人有创造力,人把它编进了一套规则,电脑只是在执行。

但现在情况变了。代码在与我们的艺术数据、音乐互动时开始改变、变异,变成了连原始编写者都不太清楚它如何运作的东西。机器学习产出的程序,也许已经与原始编写者脱钩。于是挑战来了:这种正在脱离原始代码的新型 AI,能不能在代码和编写者之间拉开距离?

你们大概都听说过图灵测试(Turing test):电脑在线上冒充人类与你交谈,你能否被说服对面是个人?图灵提出的这个挑战很有分量:机器能否处理自然语言并实时应答。而针对艺术领域,现在有人提出了一个新挑战,叫「洛夫莱斯测试」(Lovelace test)。测试内容是:机器能否原创一件有创造性的艺术作品,并且这个过程可以重复。也就是说,它不能是硬件的某个故障,代码必须知道自己在做什么,不能是塞进去的随机噪音,随机到代码无法重现自己的产物。但真正的关卡在这里:写下这段代码的程序员,在代码学习、变异之后,无法解释算法是如何得出它的输出的。

这就是我想和大家一起探究的问题:过去几年里,AI 在理解我们所谓的艺术和创造力、并产出自己的版本这件事上,做得有多好?这也是这本书的主线。我想我们已经相当乐意让 AI 开车,甚至当我们的医生,虽然后者让人有点紧张:假如它在第二局围棋里能下出第 37 手,那它建议你吃一片看起来极其危险的药时,你吃不吃?那是失误,还是能救你一命的绝妙洞见?在某些领域我们已经放心了,但我们始终认为有一样东西是人类独有的:我们的创造力,我们的艺术产出,那是人之为人的意义所在,我们用音乐、绘画、诗歌和小说来表达它。如果 AI 能接近这个被视为人类独有的东西,那就是一个非常激动人心的时刻。

视觉艺术:伦勃朗与 GAN

那么它做得有多好?大约六年前,我为 BBC 的《地平线》(Horizon)栏目做过一期关于 AI 的节目,那是图灵诞辰纪念。当时我对 AI 的状况相当失望,有一道坎它似乎怎么也迈不过去:视觉识别。我们的大脑非常擅长接收海量信息,此刻我眼前有各种颜色、各种面孔,我却能把它们整合成一个故事:一群听众来听我的演讲。视觉当时是 AI 的一大障碍,而机器学习恰恰攻克了它。

今晚我们也用了一点机器学习来做几个小实验。这里有一台摄像头。原理是这样的:机器学习先看一些猫和狗的图片,要求区分它们,起初会出错,但它不断提出更多问题,逐渐越答越对。各位面前有投票卡,我会出几道题,算是 AI 艺术版的图灵测试或洛夫莱斯测试,你们要判断哪些是人做的,哪些是没有灵魂的 AI 做的。

我们是这样训练这个 AI 的:你们举起卡片时,我们要它能识别蓝面(代表人类)和红面(代表机器人),又不能被那边一件红夹克或一件蓝衬衫搞混。所以我们随机拍了一大批照片,把要识别的目标放进去,做成大约六百张训练图像,每张都标注了「四个红色机器人、三张人脸」之类的信息。它就在这六百张图上逐渐学会了区分,现在给它一张图,它能又快又准地数出多少红、多少蓝。这就是机器学习的现场演示。

我从视觉艺术开始讲,因为视觉正是 AI 创造力最成功的领域。你们大概认得出这位画家:伦勃朗。荷兰有个团队决定试一试,能否让 AI 的视觉识别理解伦勃朗极具个人特色的画风。伦勃朗画过很多肖像,数据量不小,大约三百幅,虽然比不上我们训练 AI 常用的规模。伦勃朗对光的运用非常特别,AI 要学会如何在肖像上布光,还要学会那个年代特定的服饰风格。屏幕上的两幅画,一幅是伦勃朗的真迹,另一幅是 AI 在学习伦勃朗风格之后产出的「新伦勃朗」。你们的任务是:分辨哪一幅是真的。

先测试一下卡片是否好用。请全体举蓝面,可以看到系统开始识别你们了,有点不稳定,抱歉,你们只能相信自己的票是算数的。现在翻到红面,上方的柱条显示比例,百分之百红。再翻回蓝面,柱条又倒向蓝色。好,可以用了。

为了不带偏向,我抛硬币决定问左边这幅。你认为左边这幅是 AI 画的,请举红面;认为是人画的,请举蓝面。相当多人举红,也有人举蓝,红略占上风,大概六比三。来看答案:你们已经很厉害了,作为一个整体,你们投对了,左边那幅确实是 AI 画的伦勃朗。

有意思的是,这个团队做的不只是二维图像。看过伦勃朗原作的人都知道,他的颜料用法很特别,有很强的立体感,所以他们连颜料的堆叠高度都分析了,最后用 3D 打印做出成品。当他们请伦勃朗专家来评审时,这位专家对整个项目态度非常傲慢、不屑一顾,但他能找出的唯一毛病是:颜料的用法比这幅肖像的画风早了二十年。团队觉得挺满意,如果这就是全部的错处,那已经不错了。

你也许会问,再来一幅伦勃朗有什么意义?我们已经有那么多美妙的伦勃朗了,还要新的干什么?我最喜欢的艺术评论家、《卫报》的乔纳森·琼斯(Jonathan Jones)肯定这么想。我喜欢读他的文章,因为他对一切和 AI 有关的东西都全盘否定。关于这个伦勃朗项目他写道:这是对人性中一切创造性的可怕、无品味、麻木不仁、没有灵魂的践踏,是把技术用在了永远不该用的地方。坦白说,一个艺术评论家穿那种衬衫,我不太信得过他的品味。不过在某种程度上他有道理:再造一幅伦勃朗有什么意义?

我认为确实有意义,因为这种 AI 的妙处在于,它开始识别出我们人类在数据里漏掉的东西。伦勃朗这里倒还没有,我没看到关于伦勃朗的新见解,但比如杰克逊·波洛克(Jackson Pollock),对他作品的算法分析揭示出,波洛克泼洒颜料时在创造一种非常特殊的数学形状,我们可以分析并衡量这些画的分形维度。AI 给了我们新的洞见。书里我还讲了 Netflix 算法的故事:它只拿到我们对电影的喜恶和影片编号,对影片内容一无所知,却仅凭喜恶数据就把影片聚成了相似类型,你一看,这些全是喜剧,那些全是惊悚片。但它时不时会把一些影片聚在一起,仅仅因为我们的喜恶体现出对这些片子的某种共同感受,而我们并没有给这种类型起过名字。就好像 AI 透过我们的喜恶,发现了一种值得取新名字的分组方式,看到了影片之间一种我们从未命名的结构。所以回望过去是有意义的。但最激动人心的还是向前看:能不能让 AI 做出新东西,打破常规,做些令人兴奋的新事情?

下一道题:这八幅画里,四幅是人画的,四幅是 AI 画的,你们要判断哪边是哪边。同样问左边,左边四幅是人还是 AI?请投票。这简直像脱欧公投一样,你们这次远没有那么确定,指针慢慢偏向红色,还在变,看来你们认为左边那组是 AI。来看答案:其实左边四幅是人画的。这道题你们觉得难了一些。

有意思的是,在我看来,AI 画的那四幅复杂度反而高得多。这些画曾在巴塞尔艺术博览会展出,我记得是 2016 年,当时没有告诉任何人有 AI 参与,只请观众谈感受。结果人们对 AI 作品的情感投入,比对人类作品的投入还要多。而当你告诉某个人,其实那是电脑画的,他们会非常沮丧:天哪,我和它产生了亲密的共鸣,可那里面根本没有情感世界。这种先体验、再得知出自 AI 的反应非常有趣,我想在座的大多数人也一样,觉得自己被骗了。可假如我告诉你们,刚才主持人讲的笑话其实全是 AI 编的,你们也都笑了,那这会让你们的笑作废吗?我不觉得。

我认为不必对此太感威胁,原因在于这种 AI 是在我们的情感世界上学习,以此决定下一步。它并没有与我们脱节,它是有情感世界的,只不过那是我们情感世界经过新滤镜之后的映照。

这个项目尤其有趣的是这些画的生成方式。右边四幅画不是由一个算法生成的,而是两个算法几乎在互相竞争中做出来的,它们把创作变成了一场博弈,这叫生成对抗网络(Generative Adversarial Network,GAN)。第一个算法负责创作,它先学习过去所有的艺术作品,成了一位艺术史学家,通过对图像做机器学习并被告知每幅属于哪种风格,学会了给艺术分类,懂得什么是立体主义、什么是点彩派。然后它的任务是创作一件不属于任何已知风格的东西,真正打破模子,做出用它学到的参数无法归类的作品。但同时它还得做出人类会认作艺术的东西,它从过去一千五百年的艺术中已经学到了我们认为什么算艺术,所以知道自己能把边界推到多远。第二个算法是判别器(discriminator),它的任务是说:不行,你这还困在立体主义里;或者说:你走得太远了,这根本不是艺术。正是两者之间的较量,最终产出了这些图像。

创造的双人舞:生成与筛选

这个想法令人兴奋的地方在于:算法往往可以大批量地生产东西,难的是挑出哪些有意思。这里有第二个算法在做挑选和判别,而我认为这非常接近人类创作的实际方式。保罗·克利(Paul Klee)谈创作时说过,在创造行为的最初,创造的冲动刚刚发生,紧接着就会出现一股反向的运动,一种接纳性的运动,这意味着创作者在控制自己产出的东西。你做了点什么,然后审视它:不行,我不确定,扔掉,再来。法国诗人保罗·瓦莱里(Paul Valéry)也谈过这两种心智:任何发明都需要两个人,一个负责组合,另一个负责挑选。作为数学家我深有体会。我在世界各地有合作者,我们就扮演这两种角色:和德国的合作者一起时,我是那个疯狂的创造者,他是判别者,负责把东西打回去;而和中东的合作者一起时,他是疯狂的创造者,我是判别者。正是靠这种组合,我们才能一起取得进展。

文字:诗歌成功,小说跑偏

视觉世界说完了,文字呢?刚才我们已经听到了文本预测会产生怎样奇怪的效果。有意思的是,AI 在文字上遇到了相当大的困难,可这偏偏是 AI 最早想做的事情之一。图灵离开布莱切利庄园之后,去曼彻斯特实现他的一些构想,那里有「曼彻斯特通用计算机」。团队成员某天发现实验室里到处出现信件,是情书,署名曼彻斯特通用计算机,大家都莫名其妙。直到一位同事承认,是他给电脑写了个程序,用一个模板,再用图灵刚为这台电脑做的随机数发生器往模板里随机填入情意绵绵的词语。看几封之后你就能看出模板来了,所以不算高明。

不过诗歌是 AI 相当成功的一个领域。部分原因是诗歌也是一种漂亮的封闭形式,不要求太大规模的结构。另一个原因是,各位作为观众,在面对一件艺术品时会投入大量自己的创造力,对吧?艺术家会留出空间,让你去填补、去解读那些含混之处。诗歌有一种谜语般的特质,读的时候你会带入特别多自己的创造力。

现在我给你们三首诗,请判断是 AI 写的还是人写的。我来念第一首,大意是:我的伙伴,凡俗之躯,将我摇篮般的心埋葬;温暖的心跳与冰冷的心跳为伴;我比你更早还是你先在我们的力量前失败,躺在步枪的废墟里,那曾是一个艺术的世界。是机器人还是人?红代表机器人,蓝代表人类。压倒性地投红,少数人认为是人。先不揭晓答案。

第二首风格截然不同,念起来都费劲,大意是:杀虫剂的小护士们,实时的反应,被驱逐的成包的排他性的心痛,无力感,被杂志开除的、不体面的、雇员化的、内化的暴力,冻结的无助。这是 AI 还是人?你们在猜我是不是在耍你们,是不是双重诈唬,因为这显然像代码。有一点偏向人类,但仍有不少人拿不准。好,你们最后选了人。

第三首,大意是:想象此刻黑烟醒来,飞过这些年,飞向另一天;纠缠的祈祷的念头,水的另一边,我看见它已经在这里;她面孔的序列,一连串分享过的老朋友,走过了他们的梦。机器人还是人?你们几乎一致投了机器人。等一等,那位说得对,是的,你们全押了机器人。

揭晓答案。第一首,你们非常确信是 AI,可那是杰拉德·曼利·霍普金斯(Gerard Manley Hopkins)的诗,他要在坟墓里翻身了。我选霍普金斯,是因为我从来没读懂过他的任何一首诗,你们没有识破。第二首,你们判断「太像代码了,不可能是 AI」,猜对了,那是一位年轻的澳大利亚诗人 Mez Breeze 的作品,她非常关注计算机代码自身的诗意和节奏与它对我们可能表达的意义之间的交互,专注于这个古怪的交界地带。既然前两首都是人写的,剩下的第三首,那唯一一首读起来通顺有意义的,恰恰是 AI 写的。它出自雷·库兹韦尔(Ray Kurzweil)创造的「控制论诗人」(Cybernetic Poet),这是一个机器学习程序,他把叶芝、济慈、艾略特的诗喂进去,然后让程序创作出叶芝与艾略特的某种融合体。库兹韦尔正是鼓吹「奇点」(singularity)的人之一,奇点指的是电脑可能比人类更聪明的那一刻。

所以诗歌做得还不错。真正让 AI 吃力的是长篇写作。它能生成相当有趣的小段文本,但写小说仍远远超出它的能力。不过也有人尝试过。有一个叫 Botnik 的团队,都是《哈利·波特》的铁杆书迷,他们对只有七卷感到失望,想要第八卷,想知道后来发生了什么。于是他们让机器学习读遍 J. K. 罗琳的全部作品,学习她关心的主题和写作风格,再用算法创作第八卷。我特别喜欢这本书的书名,叫《哈利·波特与一幅看起来像一大堆灰烬的肖像》,光凭这个书名我就想读。

开头还不错,第一句大意是:魔法,这是哈利·波特认为非常好的东西。很好,它已经领会到这套书讲的是魔法。接着是「皮革般的雨幕抽打着哈利的幽灵,他穿过场地走向城堡」,「皮革般的雨幕」我觉得是个很美的意象。但从此往后就渐渐失控了:罗恩站在那里跳着一种狂乱的踢踏舞,看见哈利,立刻开始吃赫敏的家人;罗恩的衬衫和罗恩本人一样糟糕。所以 AI 很擅长局部生成有意义的片段,却没有长程结构的把握。

音乐:巴赫与 AI 的拼接实验

音乐呢?洛夫莱斯给我们留下了这个挑战:AI 能不能作曲。音乐里布满了模式。收音机里响起一段音乐,你往往很快就能听出作曲家是谁,因为他们有各自的风格,各自特有的声音世界。AI 能学会吗?能做出水准足够高的仿作,或者做出新东西吗?

AI 总是从巴赫开始,因为巴赫的作品里有大量算法在运作。比如《音乐的奉献》,那是巴赫为一位公爵写的一组小曲,形同谜题:给你一个小小的算法,你得把它展开,才能看出音乐的真面目。所以巴赫确实是 AI 的好起点。几周前的 Google 涂鸦你们玩过吗?那是庆祝巴赫生日的,你输入一行旋律,它会为你配上其余三个声部。这背后的机器学习叫 Magenta,它学习了巴赫写的所有众赞歌(chorale)。众赞歌是很好的材料,它们不常转调,相当规矩乏味,但能让机器学到大量关于和声进行的知识。填和声有点像填数独,学会规则就行。我往里输了一段《音乐的奉献》的主题,那是个相当难的挑战,因为这段主题几乎听不出调性。放出来之后,说实话相当糟糕。但涂鸦有个很好的设计:你可以说「这太烂了」,程序就会学习,承认这种配和声不对;你觉得好,它就重设参数,说好的,我多来点这种和声。你在给反馈的同时就在训练算法。

我和作曲家艾米莉·霍华德(Emily Howard)在皇家北方音乐学院设立了一个中心,叫 PRiSM,意为「科学与音乐的实践与研究」,我们关注的正是科学和数学世界与作曲世界之间的交互。我们做过各种项目,比如合作写过一首把数学证明表现出来的弦乐四重奏。现在我还带了一位作曲家博士生,叫罗伯特·莱德劳(Robert Laidlow),他的四重奏上周刚在威格摩尔音乐厅首演。他本来今晚要来,可惜病了,很遗憾。他一直和我以及本校的几位同事合作,用目前最好的软件做了一首 AI 与巴赫的混合作品,这就是我们要为大家演奏的。

系里有出色的音乐家,Kobe 会上台演奏。刚才我们做的都是直截了当的「AI 还是人」测试,这次的挑战稍有不同:这首曲子有时是 AI,有时是巴赫,我不告诉你们它在两者之间切换了多少次。你们要做的是找出接缝:能不能听出「天哪,这难听,这是 AI 了」,或者「对,这是巴赫」。曲子大约四分钟,我们演奏两遍,第一遍记录你们的判断,第二遍回放给你们看。因为现在只剩一块屏幕了,我想了个非常「模拟」的即兴办法,一会儿你们就明白。

请先全体举蓝面,我们从蓝色开始。一旦你觉得音乐进入了不是巴赫的部分,就翻到红面;觉得回到了巴赫,就翻回蓝面。就这么简单。为了同步,我来倒数:四、三、二、一。

(演奏。)

我猜有几个瞬间巴赫要在坟墓里翻身了,你们待会儿会看到。有意思的是,当音乐是巴赫的时候,你们的判断非常自信,能感觉到这是对的;不是巴赫的时候,你们就变得犹疑不定,而有那么几处突然涌起一片红,那些是露馅的时刻。

让我说明一下我们做了什么。我们取了巴赫的《第四英国组曲》,把其中若干段落挖掉,请 AI 把空缺填上。我们用的机器学习其实相当简单,叫 Clara,由 OpenAI 开发。OpenAI 这个团队致力于让 AI 向全世界开放,埃隆·马斯克参与了创立。这个软件由克里斯汀·佩恩(Christine Payne)在 OpenAI 项目中编写,后来发展成了 MuseNet,我们还会继续用它改进这首曲子。这个软件的工作方式是预测性的:看到目前为止发生了什么,然后决定接下来是什么。有些 AI 也会很聪明地反向工作,知道曲子要去哪里,从而做出预测。但这个软件没有长期记忆,这是作曲软件的一大难题:如何让它记住自己过去做过什么,并在后续决策中加以利用。

演出前我和 Kobe 聊过,他说自己能非常清楚地感觉到 AI 在哪儿,因为 AI 没有身体,它毫不在意写出极其别扭的指法,而巴赫写的东西在手指下总是顺畅妥帖。我想这是 AI 普遍面临的挑战之一:它没有「具身」(embodied)。你确实能感觉到这一点,我第一次听 Kobe 弹时的感受就是:哇,这真是黏糊糊的一团。

现在回放。Kobe 要再弹一遍,AI 的部分对他来说很痛苦,巴赫的部分则很过瘾。屏幕上会显示你们刚才的投票,我用手势告诉你们哪些段落是 AI、哪些不是。我承认自己稍稍做了手脚:我让你们从蓝面开始,其实是想看看你们要多久才能发现,开头根本不是巴赫。你们花了相当长的时间才察觉。倒数:四、三、二、一。

(回放。)这段你们很快就抓住了,这是巴赫,能听出它有明确的方向感。(继续。)这里又回到巴赫。(继续。)现在是 AI 了。(演奏结束。)

非常感谢 Kobe。有一点我对作曲家罗伯特要求非常严格:AI 的部分必须纯粹是 AI,不允许他有任何机会去改进它。我之所以坚持,是因为你去看很多项目,包括我书里写到的许多,尽管标榜有 AI 参与,仔细一看就会发现有大量的人工干预。「全是 AI 做的,人类完全没插手」当然是更好听的故事,所以我们把规矩定死:AI 的部分只能是 AI。有意思的是,我们只请它填最后一个和弦,它却漏掉了一个关键的音,那个音本来能让全曲得到解决。但你们也看到了,这首曲子相当能以假乱真,露馅的时刻不多,要挑出哪些是 AI 其实很难。你们对哪些是巴赫判断得很准,只是有几个可怕的瞬间你们以为是 AI,其实是巴赫,这可怜的巴赫真要在坟墓里翻身了。

结论:工具而非威胁

这一切又有什么意义呢?我认为对创作艺术家来说意义是明确的:这不是竞争,而是协作。这是一种新工具,能推动我们人类自身的创造力。在绘画领域我们已经看到了这一点,而在音乐里,我见过的最有意思的故事是一个学会了演奏爵士乐的 AI,叫「爵士续奏者」(Continuator),由弗朗索瓦·帕谢(François Pachet)和他的团队在巴黎的索尼实验室研制。他们让 AI 以非常接近爵士乐手的方式学习:在某个音符序列之后,下一步各种走向的概率是多少。然后他们办了一场音乐会,观众很难分辨什么时候是人在演奏、什么时候是 AI 在演奏。真正触动我的,是那位用自己的演奏训练了这个 AI 的爵士乐手贝尔纳·吕巴(Bernard Lubat)听到 AI 回奏给他时的反应。他说:这个系统向我展示了那些我本可以发展出来、却要花上好多年才能发展出来的想法,它领先我好几年,然而它演奏的每一个音,都毫无疑问是我。

这正是令人兴奋的地方。我们人类往往表现得非常像机器:思维陷在固定的轨道里,一遍又一遍地重复同样的想法,在创造性工作里尤其如此。我知道在我自己的数学研究里,我总是反复尝试同样的路数,有时需要有什么东西把我从惯常的思路里推出去,让我看到自己的可能性世界其实丰富得多。就像贝尔纳·吕巴,他站在房间里,聚光灯打在他身上,却没有意识到自己的声音世界里还有那么多可以玩的东西。

所以对我来说,AI 走向创造力这件事,令人兴奋的地方不在于威胁,而在于机遇:它可能推动我们不再像机器那样行事,重新成为更有创造力的人。谢谢大家。

排版 + 横图 + 来源,粘贴即成稿
章节 · 点击跳转视频
0:02 数学家会被电脑取代吗 ▶ 正在看
2:15 围棋:数学家的最后护盾 ▶ 正在看
5:18 自下而上:会学习的代码 ▶ 正在看
8:07 第 37 手与 AI 的创造性 ▶ 正在看
12:13 Boden 三要素与局部最大值 ▶ 正在看
15:16 从 Lovelace 告诫到 Lovelace 测试 ▶ 正在看
19:25 视觉艺术:伦勃朗与 GAN ▶ 正在看
33:40 创造的双人舞:生成与筛选 ▶ 正在看
34:35 文字:诗歌成功,小说跑偏 ▶ 正在看
42:28 音乐:巴赫与 AI 的拼接实验 ▶ 正在看
59:52 结论:工具而非威胁 ▶ 正在看
本期小问 · 档案清单
8:07 AlphaGo 的第 37 手,真能算作机器的创造性行为吗? ▶ 正在看
19:25 知道一幅画出自机器后,先前的感动就作废了吗? ▶ 正在看
15:16 程序员说不清算法怎么得出结果时,作品还归功于人吗? ▶ 正在看
59:52 AI 是人类创造力的威胁,还是把人推出定式的工具? ▶ 正在看
本期讲者
马库斯·杜·索托伊牛津大学数学教授、西蒙尼公众科学理解教授,皇家学会院士。著有《创造力密码》(The Creativity Code,2019),本讲即围绕该书展开。
01数学家会被电脑取代吗
0:02
[Music]
[音乐]
便签笔记
0:15
he's becoming quite a hard act to follow these days actually I think he's going to put me out of a job with all his lineup of jokes how many mathematicians do we have in the room just to get me a sense of who I'm talking to okay great so the mathematicians in the room you probably have the similar experience that I do when I go to parties and you get this question about what you do I don't pass you all fake it I'm starting to fake it actually and say I'm an international spy or something but because most of the time I get a kind of stock set of reactions one of them is that they just flee to the other end of the party and I'm abandoned but at you always before they go they always tell me what they've got in their GCSE or o-level which is I'm not really quite sure what that's about but some but if they do stick around one of the things they often say is he's trying to work out the tech here which is yeah doing a talk about tech is always a disaster actually using the tech but so one of the questions I often get or statements
他现在真是让人很难接上话了,我觉得他讲了这么一连串笑话,快要抢走我的饭碗了。在座有多少位数学家?我想先了解一下我面对的是什么样的听众,好,很好,那么在座的数学家们,你们大概和我有过类似的经历:我去参加聚会时,总会被问到你是做什么的。别告诉我你们都不装一下。我现在也开始装了,会说自己是国际间谍之类的,因为大多数时候我得到的反应总是那几种。其中一种是他们直接逃到聚会的另一头,把我一个人晾在那儿,不过在走之前,他们总会告诉我他们GCSE或O-level考试考了多少分,我也不太清楚这是要说明什么。但如果他们真的留下来了,他们常说的一句话是——他正在琢磨这设备怎么用——是啊,做一场关于科技的演讲,真要用起科技来总是场灾难。总之我常被问到的问题或者说法之一是:哦,得了吧,你这工作现在肯定早就被电脑取代了吧。我
便签笔记
1:16
oh come on surely you must have been put out of a job by a computer by now I think it's mostly because people think that what I do in my office up here is long division to lots of decimal places and if that were true certainly my computer would have put me out of a job but you know we're all as Alan said we're all a little bit threatened at the moment about this advancing AI that it seems to be very powerful doing lots of interesting things and surely aren't computers all about kind of logic and mathematics so wouldn't my job be one of the first to be threatened I got this a lot during the nineties actually when deep deep blue beat Kasparov because often people used to compare the idea of playing a game of chess to doing mathematics there are certain logical Meuse you can make with the pieces there's a kind of end game that you're after at the end of the proof the QED you a you sort of win the game or you win the proof and so a lot of people said to me during the 90s well come on you must be next
觉得这多半是因为大家以为我在楼上办公室里做的事,就是把长除法算到小数点后好多位。如果真是那样,我的电脑当然早就把我的饭碗抢走了。不过正如Alan说的,我们现在都多多少少感到一点威胁,这个不断进步的人工智能似乎非常强大,能做很多有意思的事情。而电脑不就是关于逻辑和数学的吗?那我的工作岂不是最先受到威胁的之一?其实在九十年代我就经常被这么问,当时深蓝战胜了卡斯帕罗夫,因为人们常常把下一盘棋和做数学相提并论:棋子有一定的逻辑走法,最后有一个你追求的残局,就像证明结尾的QED,你赢下这盘棋,或者说完成了这个证明。所以九十年代很多人对我说,得了吧,下一个就轮到你了。但我从没觉得国际象棋特别有威胁性,而且
便签笔记
02围棋:数学家的最后护盾
2:15
but I never felt felt particularly threatened by chess and actually there was always another game that we mathematicians always use as our kind of protective shield against the idea that computers could do our subject because there's another game and that's the ancient game of Go this Chinese game played on a 19 by 19 grid where you put black and white stones down you try to engulf the other person's territory before they engulf yours and this is a grain which has a high degree of complexity much more complex actually than chess there's a lot of pattern recognition that needs to go on to be able to play this game and that's something that really mathematics is about it's about spotting patterns underlying patterns but quite often when you're playing go and especially when you're doing mathematics to be able to quite know where you're going requires a lot of intuition a lot of creativity not quite sure why you're making moves you spend a lot of time in this world and you build up kind of kind of a feel for
我们数学家一直还有另一种游戏,用来抵挡“电脑能做我们这门学科”的说法,因为还有一种游戏,那就是古老的围棋,这种中国的棋类在19乘19的棋盘上下,你放下黑白棋子,试图在对方围住你的地盘之前先围住他的地盘。这种游戏复杂度极高,实际上比国际象棋复杂得多。要会下这种棋,需要大量的模式识别,而这正是数学的核心所在——发现模式,发现底层的规律。但很多时候,当你在下围棋,尤其是在做数学的时候,要想清楚知道自己要往哪儿走,需要大量的直觉、大量的创造力,你也不太清楚自己为什么这么走。你在这个世界里待很久,慢慢培养出一种下围棋或者下象棋的感觉。传统上
便签笔记
3:13
playing go or or doing chess and very traditionally in computer science lectures they would always say yeah chess is something we can automate because we can understand the kind of logical implications of playing particular moves we can follow the the tree of possibilities through go was always said to be a game that no computer would have ever be able to play and certainly any attempt of somebody trying to encode the way that a human plays this game always failed the attempts to try and encode playing go in some sort of algorithm wouldn't even beat an amateur at this game so I felt pretty safe because computer science say you can't play go so you certainly won't yet we will be able to play the complex game of mathematics so I got a little bit of a shock a couple of years ago and you're probably aware of this story when a team in London declared that they had got an algorithm that they believed could compete at not just a high level but the highest level this is deepmind in london and actually was dennis has
在计算机科学的课堂上,人们总会说,国际象棋是我们可以自动化的,因为我们能理解下某一步棋在逻辑上的推论,我们可以沿着可能性的树状结构推演下去。而围棋一直被认为是电脑永远不可能会下的游戏。当然,任何试图把人类下这种棋的方式编码进程序的尝试都失败了,那些想把围棋下法写成某种算法的努力,甚至连在这个游戏上赢过业余棋手而已,所以我当时觉得挺安全的,因为计算机科学界都说你没法下围棋,那你当然更不可能去玩数学这么复杂的游戏了。所以几年前我受到了一点冲击,你们大概也知道这个故事——伦敦的一支团队宣布,他们做出了一个算法,他们相信这个算法不仅能在很高的水平上竞争,而且能达到最高水平。这就是伦敦的 DeepMind,其实是 Demis
便签笔记
4:18
abyss who went to see elect he went to Cambridge to do his study and he was told this a old adage that you can't complete a program a computer to play go and this was like a red rag to Dennis and so he went away and set up this company and they devised this algorithm that they thought could play the best and they challenged Lisa Dahl a korean grandmaster and Lisa Dahl was totally dismissive of this algorithm that wouldn't be able to get anywhere near the level that he could play at he said I'm going to demolish this thing 5-nil they were gonna play over five games but he got a little bit of a shock i sat and watched these games obsessively on YouTube because I realized that my life was probably under threat and as I watched I saw Lisa dole get more and more depressed throughout the games he lost the first game he lost the second he lost the third he'd lost the match already after three he won the fourth game and he now regards that as the greatest game that he's ever played that he was able to beat this algorithm in
哈萨比斯,他去剑桥读书,在那儿别人告诉他一句老话:你不可能编出一个能下围棋的程序。这话对 Demis 来说就像红布之于公牛,于是他跑去创办了这家公司,他们设计出了这个他们认为能下到最强水平的算法,然后他们向韩国的顶尖棋手李世石发起了挑战。李世石当时完全看不起这个算法,觉得它根本达不到自己的水平。他说我要 5 比 0 横扫这玩意儿——他们要下五盘棋。但他受到了一点冲击。我当时就坐在那儿,在 YouTube 上着了魔一样地看这些比赛,因为我意识到我的饭碗大概也危险了。看着看着,我看到李世石一盘比一盘更沮丧。他输了第一盘,输了第二盘,输了第三盘,三盘之后这场比赛他就已经输定了。他赢下了第四盘,而他现在认为那是他这辈子下过的最伟大的一盘棋,因为他能在其中一盘里击败这个算法。第五盘他又输了,所以他 1 比 4 输掉了比赛。那么在过去
便签笔记
03自下而上:会学习的代码
5:18
one game and he lost the fifth game so he lost 4-1 what had changed in the last two years the style of coding has changed and we've got new sort of code on the block which is able to do things the code in the past couldn't do and it's something that we're very interested here in Oxford this idea of deep learning or machine learning so code in the past used to be written in a very top-down manner you really have to know what the thing what you was you how the thing was going to behave you told it the rules of how it was going to play you had to understand the setting and the machine just implemented that yeah sure it could play chess cuz it was told to implement this thing it could go deeper it could analyze more situations than a human but the human was still telling the program what to do what has changed is that the code is now written in a very bottom-up manner we've got a sort of code which is learning very much like a child learns in the past it was like a the parents DNA would give birth
两年里到底发生了什么变化?编程的方式变了,我们有了一种新的代码,它能做到过去的代码做不到的事情。这也是我们牛津这边非常感兴趣的东西——深度学习或者说机器学习。过去的代码是以一种非常自上而下的方式写的,你必须真的知道这个东西会怎么运作,你要把规则告诉它,你必须理解整个场景,而机器只是执行你告诉它的东西。是的,它能下国际象棋,因为你让它去执行这套东西;它能算得更深,能分析比人类更多的局面,但归根结底还是人类在告诉程序该怎么做。而现在变化的是代码变成了以一种非常自下而上的方式写出来的。我们有了一种代码,它的学习方式非常像小孩子学习。过去就好比父母的 DNA 生出一个孩子,但这个孩子始终
便签笔记
6:17
to a child but the child would be still attached to the DNA of the parent it wouldn't learn and I think knew but suddenly we've got code that can adapt and change and mutate and re parameterize itself as it encounters a new environment there's almost like a meta code which is telling the code how to change if it gets something wrong and change and mutate and this is what they use to actual trained alphago to to learn how to play this game by playing games and failing back they started with some other games to start with some some simpler games so in fact they started with Atari games claims I used to be obsessed with actually when I was a kid and one of the ones I really loved was this one called breakout and where you have a little ball which ping pongs up and down you've got a paddle and you've got to knock these out new score points the Bloor just one point up to read to the higher points the machine was only given the pixels on the screen and the score it had to learn how to play this game
被拴在父母的 DNA 上,它不会学习、不会产生新东西。可突然之间,我们有了能够适应、能够改变、能够变异、能够在遇到新环境时重新调整自身参数的代码。几乎可以说有一层元代码在告诉这段代码如果做错了该怎么改变、怎么变异。他们正是用这种方法训练 AlphaGo,让它学会下这个棋——通过不断对弈、不断失败来学。他们一开始其实是从别的游戏、从一些更简单的游戏入手的,事实上他们是从雅达利的游戏开始的。这些游戏我小时候也曾经很着迷,其中我特别喜欢的一个叫《打砖块》(Breakout):有一个小球上上下下地弹来弹去,你有一块挡板,你得把这些砖块打掉来得分,蓝色的只有一分,越往上到红色分数越高。机器得到的只有屏幕上的像素和分数,它必须自己学会怎么玩这个游戏。没人告诉它
便签笔记
7:12
wasn't told anything about the fact that you had to hit this ball but it was randomly moving the paddle and every time it hit the ball it saw the score go up so it Reap parameterised itself a turn and said i'm going to prioritize moving towards where the ball came now when I was with my mate and we played this we were very pleased when we found a fantastic hack because you can create a little tunnel on the left-hand side and then if you get the ball to go up there then you don't have to do any work at all because the ball just bounces backwards and forwards what I was absolutely staggered when you see how this deep mind actually you learnt to play the Atari game it learnt the same hack it is you know not just humans but the computer is also lazy doesn't want to move this thing around and so shoots it back up again this is extraordinary after 600 games it had learned just by randomly moving the paddle and seeing that what the moves were that made that the score go up fast it had learnt how
必须要接住这个球,它只是随机地移动挡板,而每次碰到球,它就看到分数上涨。于是它重新调整自身参数,说:我要优先把挡板朝球来的方向移动。当年我和我哥们儿玩这个游戏的时候,我们发现了一个绝妙的技巧,特别得意——你可以在左边打通一条小隧道,然后只要让球从那儿钻上去,你就根本什么都不用做了,因为球会自己在上面来回反弹。让我彻底震惊的是,当你看到 DeepMind 是怎么学会玩这个雅达利游戏的,你会发现它学会了同一个技巧。你知道,不光人类懒,计算机也懒,它也不想一直挪来挪去,所以它把球又打回上面去。这太不可思议了。600 局之后,它仅仅通过随机移动挡板、观察哪些动作能让分数涨得最快,就学会了
便签笔记
04第 37 手与 AI 的创造性
8:07
to do this hack so this is how it then went on to play the game of go so what it did was to first take all the human games that are on the Internet a lot of games encoded on the Internet it learned how humans lost and won the games and that was its first material it learnt on then it started to create synthetic data it started to play itself different versions of itself would play a game and then if it lost a game it would understand which were the moves that meant that our that it actually was failing to play at a high level and it would reap ramit rise itself and so after a while it got to this high level quite amazing than it did suchly it could challenge lease at all now at first sight you say ok so the community's just got very good it can analyze it's gained very deeply but I think that there was something more amazing that happened during these games the first game interestingly Lisa Dahl decided that if it had done learning on how humans play it maybe the best way to beat this computer was to play rather
这个技巧。接着它就用同样的方式去下围棋。它做的第一件事,是把所有人类在互联网上留下的棋谱都拿过来——网上有大量的棋谱记录——它学人类是怎么输、怎么赢的,这是它学习的第一批素材。然后它开始生成合成数据,开始自己跟自己下,不同版本的自己互相下一盘棋,如果输了,它就去搞清楚哪些手棋导致它没能下出高水平,然后重新调整自身参数。这样过了一阵子,它就达到了这个高度,相当惊人,于是它就可以去挑战李世石了。乍一看你会说,好吧,这台机器只是变得很厉害,能把局面分析得很深。但我认为在这几盘棋里发生了更了不起的事情。有意思的是,第一盘棋里李世石判断,既然它是学习人类怎么下棋的,那么打败这台电脑最好的办法也许就是下得不像人类。所以他下了一盘非常搅局的棋,但 AlphaGo
便签笔记
9:12
unlike a human so he actually played a very disruptive game but at the alphago was smart enough to actually just cope with the moves that he was making and they may turn out to be quite weak moves and he lost the game because he was not playing his standard game so in the second game least had all decided to play a much more standard a high level game that he knew very well very early on in the game your go master teaches you a few things strategies one is that you should play on the kind of age of the board so you're encouraged to pray on the first second third and fourth rows in because there's a kind of competition early on for the edges and the kind of internal part of the board and if you play too far into the middle of the board early on it's considered a weak move because you're not really establishing important territory at that point so Lisa doll on the 36 move of this game decided that he needed a cigarette break and he went up to the top of the hotel had a cigarette alphago didn't need to
足够聪明,完全应付得了他下出的这些手,而这些手最后反而显得相当弱,结果他输了这盘棋,因为他没有下自己的正常棋。所以第二盘棋,李世石决定下一盘他非常熟悉的、更标准的高水平棋。在开局阶段,围棋老师会教你一些策略,其中一条是你应该往棋盘的边角一带下,也就是鼓励你下在第一线、第二线、第三线和第四线,因为开局阶段是在争夺边角以及棋盘内部的势力,如果你开局就下得太靠棋盘中间,那会被认为是弱手,因为在那个阶段你并没有真正确立重要的地盘。所以在这盘棋的第 36 手,李世石决定去抽根烟,他上到酒店顶楼抽了根烟。AlphaGo 不需要靠抽烟来提神,
便签笔记
10:10
smoke in order to get stimulation so it's that there and it's not for a while and then it asked the human player because this wasn't an exercise in robotics this was an exercise in just pure thought so there was still human this is actually still quite difficult for an AI to actually pick a stone up and place it on the board but it told the human player to place a stone on the fifth row in so I've circled it he was playing he or Bernie already anthropomorphize the AI the AI is playing black and it put this stone or in on the fifth row in which to put a little white circle on all the commentators I remember this on YouTube they all gassed and went whoa Wow alphago has made a huge mistake never play that sort of move early on in the game this is a Lisa door will be in it'll be one one after this and they were all very complacent and Lisa it all came down after his cigarette break looked at what alphago and you can see you should watch this bag is so funny because it just Lisa don't like cannot
它就坐在那儿,想了好一会儿,然后让人类棋手代它落子——因为这不是一场机器人学的比拼,而是一场纯粹思维的比拼,所以还是有人在场,对 AI 来说要拿起一颗棋子放到棋盘上其实还挺难的。它告诉人类棋手把棋子下在第五线上,我在图上把它圈出来了。它执黑——我这已经在把 AI 拟人化了——AI 执黑,把这颗棋子下在了第五线上,就是我画了白色小圈的那个位置。我记得 YouTube 上所有的解说员都倒吸一口气,哇地叫起来:AlphaGo 下了一步大错棋,开局绝不能下这种手;这下李世石要扳回来了,之后就是一比一了。他们都非常轻松笃定。李世石抽完烟下来,看到 AlphaGo 下的这手,你能看到——你真该去看看这段,太逗了——李世石简直不敢相信这个 AI 居然下了这么一手,
便签笔记
11:11
believe what this AI is just suggested what a stupid mood but he's he's a bit more suspicious things you know what that why is he done that move I still said it why did it do that move um it turned out that as the game built up and there's something rather different about chess and go because chess gets simpler as the game goes on because pieces get taken off but go gets more and more complex because more and more pieces get put on as the game built up and more and more pieces were put on the board and territory was building up from the bottom right hand corner and it turned out that alphago's move at move 37 that black stone meant that it was it that won that territory rather than Lisa doll it was an incredibly inspired move it won alphago the second match that decision on move 37 to put the stone there and break the tradition of how humans thought we should play the game and for me this was really exciting because I believe that this is an example of what we should call a creative act by artificial
多蠢的一手棋啊。但他心里又有点疑虑:你知道,它为什么要下这一手?我到现在也还在想,它为什么下这一手。结果随着棋局展开——国际象棋和围棋有个很不一样的地方:国际象棋越下越简单,因为棋子会被吃掉;而围棋越下越复杂,因为棋子越来越多。随着棋局展开,棋盘上的子越来越多,右下角一带的地盘慢慢形成,结果发现,AlphaGo 第 37 手下的那颗黑子,正好意味着那块地盘归了它,而不是李世石。那是一手极具灵感的棋,正是第 37 手在那里落子、打破了人类关于这盘棋该怎么下的传统,为 AlphaGo 赢下了第二局。对我来说这特别令人兴奋,因为我相信这就是一个可以称之为人工智能创造性行为的例子。过去几年我在英国皇家学会的一个委员会里待过一段时间,
便签笔记
05Boden 三要素与局部最大值
12:13
intelligence I spent some time on a committee at the Royal Society over the last few years we've been looking at the impact that machine learning is having on society over the next kind of ten years Dennis was on the committee and there was also a philosopher Margaret Bowden and I talked to her quite a bit about the idea of creativity she's been very interested in what she calls these tin cans can do computers and the idea of whether they could be creative and she had a very nice working definition of what we should call creative I'm not sure it's the best one and we can argue there's lots of philosophical debate over the idea of what we mean by creativity but I think this is going to be quite useful working definition as we go forward tonight so creativity is something which should be new well computers can make new things quite easily and we can objectively judge whether something is new but you have two other qualities it should also have an element of surprise now that's a little bit more subjective and also
我们一直在研究机器学习在未来十年左右对社会的影响。Demis 也在这个委员会里,还有一位哲学家 Margaret Boden,我跟她聊了很多关于创造力的想法。她一直很感兴趣的是她所说的这些“铁皮罐头”——也就是计算机——能做什么,以及它们是否可能具有创造力。她给出了一个很好用的、关于什么算创造性的工作定义。我不敢说这是最好的定义,我们当然可以争论,关于创造力到底指什么有大量的哲学争议,但我觉得今晚我们往下讲的时候,这会是一个相当好用的工作定义。首先,创造性的东西应该是新的——计算机造出新东西是很容易的,而且我们可以客观地判断某样东西是不是新的。但还有另外两个特质:它还应该带有某种意外性,这就比较主观了;以及价值,这也相当主观。所以一台
便签笔记
13:12
value that's also quite subjective so a computer if it's going to be creative as something we as humans believe is creative it's going to learn to to know what we think is surprising and has value what I think so nice about games it's of course you can judge this kind of these qualities quite quickly surprise yes the commentators all went or it's made a mistake value yes this move one alphago the game and so what I've been interested in is to look at if it can be creative in this very close environment of a game where else can it be creative can it be creative in mathematics actually I sat next to demos and I I joke to him we just become both FRS is and I said well could you get alphago to become an FRS and part of the story of my book is about demo said yeah we're already on the case so they're already a deep mind looking at big making a creative a I mathematician but so the the journey that what I think is exciting about this kind of new AI is that they're disappearing is that in the case of
计算机如果要做出我们人类认为具有创造性的东西,它就得学会知道我们认为什么是意外的、什么是有价值的。我觉得游戏的好处在于,你可以相当快地判断这些特质:意外性?有的,解说员们都惊呼说它下错了;价值?有的,这一手为 AlphaGo 赢下了比赛。所以我感兴趣的是,既然它能在游戏这种封闭的环境里表现出创造力,那它还能在哪里有创造力?它能在数学上有创造力吗?其实我当时就坐在 Demis 旁边,我们俩刚好都刚当选皇家学会院士,我就开玩笑说:你能不能让 AlphaGo 也当上皇家学会院士?我这本书里有一部分故事就是关于这个的——Demis 说:是啊,我们已经在做这件事了。所以 DeepMind 已经在研究怎么造出一个有创造力的 AI 数学家了。而这段旅程里我觉得这种新型 AI 最激动人心的地方在于,就 AlphaGo 而言,它不仅把这个游戏下到了
便签笔记
14:12
alphago it not only played the game at a high level but it taught us how to play the game in a new way we thought we'd reached a kind of peak of playing with these kind of rules that we had about playing and the ones of for throw-in and you know there was a kind of optimal way to play the game what alphago has shown us in these games is that although we thought we were at the peak of a performance in playing this game actually this was only what we mathematicians call at a local maximum actually this was like Snowdonia and there was actually a much higher amount in an Everest a new way to play the game that alphago had experimented by taking risks gone down this kind of adaptive valley and found a much better way to play an hour ago is Nell taught us a new way to play this game new strategies that are helping us to play the game at a much higher level and so the journey of this book which is called the creativity code is to look at well this AI that disappearing that seems to be able to to learn about through its
很高的水平,它还教会了我们一种全新的下法。我们原本以为,凭着我们已有的那些下棋规则,我们已经到达了某种巅峰,比如该下在第几线之类的讲究,我们以为已经有了一套最优的下法。而AlphaGo 在这几盘棋里向我们展示的是:虽然我们以为自己已经站在这个游戏的表现巅峰上,但其实这只是我们数学家所说的“局部最大值”。其实这只是斯诺登峰,而在别处还有一座高得多的珠穆朗玛峰——一种全新的下法。AlphaGo 通过冒险去试探,走下了这个适应度的山谷,找到了一种好得多的下法。所以 AlphaGo 教会了我们一种新的下法,一些新的策略,帮助我们把这个游戏下到高得多的水平。所以这本书——书名叫《创造力密码》(The Creativity Code)——的整个旅程,就是去看看:这个正在出现的 AI,它似乎能通过与周围
便签笔记
06从 Lovelace 告诫到 Lovelace 测试
15:16
interaction with the kind of digital world around it could it perhaps be creative in other realms not just creative in a game the fact one of the first people to think of the idea of code was already suggesting that Koh might be able to do things of an artistic nature so we celebrate Ada Lovelace day Ada Lovelace every year Ada Lovelace was taken by her mother to see Babbage's analytical engine her mother used to like to expose her to lots of different ideas scientific ideas and when she saw this machine she already began to realize that this could do more than just the long division or the multiplication that it could do something a little bit more exciting and she started to write down code to make the analytic engine do interesting things and that's why we sort of celebrate Ada Lovelace the notes that she wrote for a paper about the analytic engine we regard as the first idea of code to make machines do interesting things already then she was thinking about the fact that this could do maybe things which are a little bit
数字世界的互动来学习,那它有没有可能在别的领域也具有创造力,而不只是在游戏里?事实上,最早想到“代码”这个概念的人之一,当时就已经在设想代码或许能做一些具有艺术性质的事情。所以我们每年都会纪念阿达·洛芙莱斯日。阿达·洛芙莱斯的母亲曾带她去看巴贝奇的分析机,她母亲喜欢让她接触各种各样的思想、各种科学思想。当她看到这台机器时,她就已经开始意识到,这东西能做的不只是长除法或者乘法,它可以做一些更让人兴奋的事情。于是她开始写下代码,让分析机去做一些有意思的事情,这就是我们纪念阿达·洛芙莱斯的原因。她为一篇关于分析机的论文所写的注记,被我们视为“让机器做有趣的事情的代码”这一想法的起点。那时候她就已经在想,这东西也许能做一些比单纯的
便签笔记
16:21
more interesting than just sort of scientific calculations then she wrote the engine might compose elaborate and scientific pieces of music of any degree of complexity or extent so she's already thinking about music a place of course which has quite a lot of connection with mathematics the idea of patterns getting the machine to to kind of run out patterns perhaps it could Bank music and we'll come to that a little bit later on the challenge of whether a I can write music but she offered a word of caution when she wrote this she said it is desirable to guard against the possibility of exaggerated ideas that might arise as to the powers of the analytic engine it has no pretensions whatever to originate anything it can do whatever we order it to perform and I think that's what we always felt in the past is that this kind of top-down coding well it's the human that's telling the computer what to do so if the computer is being creative that's because the human has been created if it is encoded that in just a set of rules
科学计算更有意思的事情。她写道:这台机器也许能创作出任何复杂程度或任何长度的、精巧而科学的乐曲。所以她那时就已经在想音乐了——音乐当然和数学有相当多的关联,都是关于模式的想法,让机器去生成模式,也许它就能创作音乐。关于 AI 能不能写音乐这个挑战,我们稍后会讲到。但她也给出了一句告诫,她写道:我们有必要防止人们对分析机的能力产生夸大的想法。它丝毫不能自命有创造任何东西的能力,它只能做我们命令它去执行的事情。我想这也正是我们过去一直以来的感觉:这种自上而下的编程,说到底是人在告诉计算机该做什么,所以如果计算机表现出创造力,那是因为人有创造力,是人把创造力编成了一套规则,
便签笔记
17:21
that the computer is just implementing but I think something has changed now the code is beginning to change mutate as it interacts with say our new artistic data music that is starting to become code that the original coder doesn't quite know how it's performing so this machine learning is producing programs which now perhaps disconnect itself from the original coder so here's the challenge can this new ie AI that's appearing which seems to be moving on from the original code written by the coder kind of put some distance between the code and the coder so there's you've probably heard of the Turing test kind of computer part itself off in an interaction online as I landed you know what would you be convinced that that was a human talking or is it just an AI computer so touring put this down as quite a good big challenge can it process natural language and respond real-time so here's a new challenge that has been offered connected to the artistic realm is called the Lovelace test so the test is can a machine
而计算机只是在执行。但我觉得现在有些东西变了,代码开始会改变、会变异,比如当它与我们新的艺术数据、与音乐互动时,它开始变成一种连最初的程序员都不太清楚它是怎么运作的代码。所以机器学习正在产生这样一些程序,它们或许已经和最初的编写者脱钩了。于是就有了这个挑战:这种正在出现的新 AI,它似乎已经超出了程序员最初写下的代码,它能不能在代码和编写者之间拉开一些距离?你们大概听说过图灵测试:计算机能不能在网上的交流中把自己伪装起来,让你确信跟你说话的是个人,而不是一个 AI?图灵把这作为一个相当好的重大挑战:它能不能处理自然语言并实时回应?而现在有人提出了一个与艺术领域相关的新挑战,叫做洛芙莱斯测试。这个测试是:机器能不能原创出某种有创造性的艺术作品,而且这个过程是可重复的——所以它
便签笔记
18:28
originate a creative work of art some sort of the process is repeatable so it shouldn't just be some sort of glitch in the hardware somehow the code should know what it's doing um it shouldn't be some sort of randomness which is put in there such that the code wouldn't be able to reproduce what it's done but here's the challenge yet the programmer the person who wrote the code that has now learns a mutated is unable to explain actually how the algorithm produced its output so this is the challenge that I want to explore with you how good has AI been in the last couple of years and it's really the story of the book in understanding what we regard as art and creativity and being able to produce its own version of that and I think you know I think we're quite happy that our AI is going to be driving our cars or maybe even be our doctors although a little nervous of that you know if it suggests move 37 in the game to of go and suggest you take a pill that looks incredibly dangerous do you do that is it mistake or is it
不能只是硬件出了个故障;代码在某种意义上应该知道自己在做什么,它不应该是里面加入了某种随机性,使得这段代码无法复现它已经做过的事情,但是问题就在这里——程序员,也就是写这段代码的那个人,如今代码已经自我学习、发生了变异,他反而说不清楚算法究竟是怎么产生出这个结果的。所以这就是我想和大家一起探讨的挑战:过去这几年里 AI 到底做得有多好,这其实也正是这本书的主线——AI 如何理解我们所谓的艺术与创造力,并且能够产出它自己版本的作品。我想,大家大概挺乐意让 AI 来开我们的车,甚至当我们的医生,虽然对后者会有点紧张。你想想,如果它在围棋里下出第 37 手那样的棋,然后建议你吃一颗看起来极其危险的药,你吃不吃?那是个失误,还是
便签笔记
07视觉艺术:伦勃朗与 GAN
19:25
incredibly insightful move that's gonna save your life so I think we are quite happy in certain realms but I think the one thing that we regard as uniquely human is our own creativity our artistic output that's what it means to be human we express it in music in art in poetry and novels so if AI can get close to doing something that reward isn't uniquely human I think this is a very exciting moment so so how good is it well I think I made a program for the BBC a horizon of about six years ago it was a during anniversary about AI and six years ago I was pretty disappointed in the state of AI that point and there was one hurdle that seemed AI seem to be finding really difficult to achieve and that was vision recognition our brain is very good at taking in a huge onslaught of information you know lots of different colors here people faces but I'm able to integrate this into a single story about seeing an audience it's come to the talk I'm giving so vision was one of the great hurdles for AI at the time
一步极具洞察力、能救你命的妙招?所以我觉得在某些领域我们是挺乐意接受的,但我认为有一件事是我们视为人类独有的,那就是我们的创造力、我们的艺术产出,这才是身为人的意义所在。我们通过音乐、绘画、诗歌和小说来表达它。所以如果 AI 能够接近做到这件事,那说明这件事并非人类独有——我觉得这是一个非常令人兴奋的时刻。那么它到底做得有多好呢?大概六年前我为 BBC 的《地平线》做过一个节目,那是关于 AI 的一期周年特辑。六年前我对当时 AI 的状况相当失望,当时有一道坎,AI 看起来非常难以跨过,那就是视觉识别。我们的大脑非常擅长接收海量涌来的信息——你看这里有各种各样的颜色、有人、有面孔——但我能把这些整合成一个完整的画面:我看到了一群观众,他们来听我的讲座。所以在当时,视觉是 AI 面临的一大难关,
便签笔记
20:32
and this is one of the things that it's been able to do with this idea of machine learning and so actually we're going to do we've used a bit of machine learning to do and little few experiments during this lecture so we have a camera here so the point is that a machine learning that it gets shown some images of cats and dogs and it has to distinguish them and it gets it wrong to start with but it starts to ask more and more questions which helps it to get it more and more right so we've got two we're going to be doing some tests here so you'll see you've got some some cards in front of you you're going to be doing I'm gonna give you some challenges some AI are art kind of Turing test Lovelace tests and you're gonna have to decide what you think is made by a human and what do you think is made doesn't have a soul and you know that's that's an AI so in order to do this right so we're going to get you up on your display one and we want the HDMI so that's interesting so display - ah okay that's interesting
而这恰恰是 AI 借助机器学习这个思路已经做到的事情之一。所以我们接下来其实也用了一点机器学习,在今天这场讲座里做几个小实验。我们这里有一台摄像机。基本思路是,机器学习就是给它看一些猫和狗的图片,让它去区分,一开始它会分错,但它会不断提出越来越多的问题,从而让它越做越准。所以我们有两……我们接下来要在现场做几个测试。你们会看到面前有一些卡片,你们要做的是——我会给你们出几道题,算是给 AI 艺术做的一种图灵测试、洛芙莱斯测试,你们要判断哪个是人做的,哪个是没有灵魂的东西做的,也就是 AI 做的。所以为了做这件事——好,我们现在要把画面切到你们的一号显示器,我们需要 HDMI,嗯,这就有意思了……二号显示器——啊,好,有意思,它在那边是能用的。没错,不过一旦你做 AI 和技术演示,那就是
便签笔记
21:39
it worked in there so exactly but when you're doing AI and tech it's always asking for trouble so okay well I do you think you can sort that out it's not a disaster but okay so um this is how we trained our AI so what we did was we want when you throw up your cards we want the AI to be able to recognize what you're putting up that it's a blue which will be for human and a red for robot but he doesn't want to get confused by a red jacket over there or a blue shirt and so we had to train this AI on pictures so what we did was - he just took a random load of pictures and then we put the things that we were trying to get the AI to recognize and so there'll be a training image so we had about 600 training images and basically it would be told there are four red robots and three human faces and gradually it learns over those six hundred faces to be able to distinguish these and such that now when we show it an image it can count quite quickly and effectively how many robots and how many humans there are so we're going to get
自找麻烦。好吧,你觉得能搞定吗?不算灾难,好的。那么,我们就是这样训练我们这个 AI 的。我们想做的是:当你们举起卡片时,我们希望 AI 能识别出你们举的是什么——蓝色代表人类,红色代表机器人。但我们不希望它被那边的一件红夹克,或者一件蓝衬衫搞混。所以我们必须用图片来训练这个 AI。我们的做法是,先随便拿来一大堆图片,然后把我们想让 AI 识别的东西放进去,这就成了训练图像。我们大概用了600 张训练图像,基本上就是告诉它:这里有四个红色机器人和三张人脸。慢慢地,通过这六百张图,它学会了区分它们。于是现在给它看一张图,它就能相当快速、有效地数出里面有多少机器人、多少人类。所以我们要请你们来试试。这真的就是机器
便签笔记
22:52
to you so so this is really machine learning in a kind of action so what I'm going to do is to offer you some challenges so here you are this is um okay so you you've have to improvise okay so I'm gonna say that's fine cuz I'll just swap over between the HDMI and that'll come bring you up so that's why okay so here you are so these this is a project that was done so I'm going to start with visual art because vision has been the place where our AI creativity AI is being very successful so you can probably recognize the artist here this is Rembrandt and there was a team in Hollands that decided to see whether they could get their AI vision recognition to sort of understand Rembrandt's very particular style of painting Rembrandt did quite a lot of portraits so it was quite a bit of data not as much data as we use to actually train our AI and but you know the in the region of 300 portraits and of course one of the things Rembrandt has is a very special use of light so it was going to learn how to put light on a
学习的实际应用。那么我要做的就是给你们出几道题。好,你们看这是……好吧,你得临场发挥一下。好,我说没关系,因为我可以在 HDMI 之间来回切换,那样就能把画面调出来了。所以,好,你们看,这是一个已经完成的项目。我打算从视觉艺术开始,因为视觉正是 AI 创造力表现得非常成功的领域。你们大概能认出这里的画家,这是伦勃朗。荷兰有一个团队想看看他们能不能让 AI 视觉识别在某种程度上理解伦勃朗那种非常独特的绘画风格。伦勃朗画过很多肖像画,所以数据量还算可以——虽然没有我们训练自己这个 AI 用的数据多,但大概在 300 幅肖像左右。当然,伦勃朗的一大特点就是他对光的独特运用,所以 AI 要学会怎样给一幅
便签笔记
24:06
portrait a very particular style of address at that particular time so one of these images is a Rembrandt the other one is the product of the artificial intelligence learning on Rembrandt style and producing a new Rembrandt so the challenge for you is can you tell which of these is the real Rembrandt now I'm going to just do a little experiment to first of all to see whether our cards are working so let's switch over to show you so here you are so I'd like you to have all put up your blue cards so let's see whether and you can see it's starting to pick out some of you a little bit edgy I'm sorry you'll just have to feel like your vote does count and okay now turn them over to red let's see and you'll see the bar at the top is recording so we'll be able to test the proportion there so that's 100% red now turn back to blue and you'll see the bar shoot back to the to the blue side okay so good so it seems to be working so now not to prejudice the thing so I'm going to let's go back I just show you the pictures again so
肖像布光,还有那个年代非常特定的着装风格。这两张图里,一张是伦勃朗的真迹,另一张是人工智能学习伦勃朗风格后创作出的一幅新的“伦勃朗”。所以给你们的挑战是:你们能分辨出哪一张是真正的伦勃朗吗?我先做个小实验,首先看看我们的卡片能不能正常工作。我们把画面切过来,你们看。请大家都举起你们的蓝色卡片,我们看看效果。你们能看到它开始识别出你们中的一些人了。有点让人紧张,抱歉,你们只好接受“我这一票是算数的”这种感觉了。好,现在把卡片翻成红色,我们看看。你们会看到顶上的那根条形图在记录,这样我们就能测出比例。现在是 100% 红色。再翻回蓝色,你们会看到那根条形图一下子弹回蓝色那边。好,很好,看来它是能用的。那么现在,为了不影响你们的判断——我们回到刚才那页,我再把两幅画给你们看一遍,然后我会问你们其中一幅。
便签笔记
25:14
I'm you know ask you about one of these so let me flip my coin so okay so I'm going to ask you about the painting on the Left I want you to boat where do you think the painting on the Left is done by a AI or is it done by Rembrandt so if you think the painting on the Left is by AI I want you to show your red faces to me and if you think it was a human then I want you to show your blue faces and okay so here we're starting to get so quite a lot of you voting red and we're seeing although some of you are also voting blue so it is edging over more to red I'm saying probably about 60 30 okay so good let's see whether how good you we're at this one so we'll go back to here so which one was the AI is in fact so you are pretty good already so you can feel good about yourselves so as an audience at least you voted correctly so yes the one on the left is in fact the the AI Rembrandt um now it's interesting that not only did they do this as a 2d image but you know if you've seen a Rembrandt it has a very his use of paint
我抛个硬币决定。好,我要问你们左边这幅画。我希望你们投票:你们觉得左边这幅画是 AI 画的,还是伦勃朗画的?如果你觉得左边这幅是AI 画的,就把红色那面朝我举起来;如果你觉得是人画的,就把蓝色那面举起来。好,好,现在开始出结果了,相当多人投了红色,不过我们也看到有些人投了蓝色。所以它在往红色那边倾斜,我估计大概是六比三吧。好,很好,我们来看看你们这题答得怎么样。我们切回这边,看看到底哪一幅是 AI 画的——事实上,你们已经相当厉害了,可以为自己感到高兴。至少作为一群观众,你们投对了。是的,左边那幅确实是AI 版伦勃朗。有意思的是,他们不只是做成一张二维图像。你们要是看过真的伦勃朗,就知道他对颜料的运用非常特别,很有立体感。他们甚至还去
便签笔记
26:27
is very special it's very kind of 3d effect and they even went to the extent of analyzing the kind of height of the paint and so they 3d printed interestingly and when they asked Rembrandt expert to come along and review their their result and of course the Rembrandt Expo it was incredibly snooty and dismissive about the whole project but the only thing that he could find to criticize it about was well the use of paint is 20 years earlier than the style of the actual portrait so the team felt actually quite good that they'd managed to just you know that if that was all that was wrong with it okay so but you might say well what's the point about another Rembrandt we've got wonderful Rembrandt's why do we need any more Rembrandt so certainly my favorite art critic Jonathan Jones in The Guardian I love reading Jonathan Jones because he's always totally dismissive about anything to do with a I this is what he wrote about this Rembrandt project what a horrible tasteless insensitive and soulless travesty of all
分析了颜料的厚度高度,然后把它 3D 打印了出来。有意思的是,当他们请一位伦勃朗专家过来评鉴他们的成果时,那位伦勃朗专家当然对整件事表现得极其傲慢、不屑一顾。项目,但他唯一能挑出的毛病就是——颜料的用法比这幅肖像本身的风格早了 20 年,也就是说和实际画风对不上,所以团队其实感觉挺好的,因为他们做到了,你知道,如果问题就只有这一点,那已经很不错了。好,不过你可能会说,再多一幅伦勃朗有什么意义呢?我们已经有那么多精彩的伦勃朗了,为什么还需要更多?当然,我最喜欢的艺术评论家、《卫报》的乔纳森·琼斯——我特别爱读乔纳森·琼斯,因为他对任何跟 AI 沾边的东西都是彻底不屑一顾——他是这么评价这个伦勃朗项目的:这是一场可怕的、毫无品味的、麻木不仁、没有灵魂的闹剧,糟蹋了人性中一切有创造力的东西,
便签笔记
27:26
that is creative in human nature when technology is used for things it never should be used for but frankly anyone who wears out sort of shirt as an art critic I really not quite sure I trust very much in their their critique but but to a certain extent he has a point you know what is the point about creating other Rembrandt well I do think there is a point because the wonderful thing about this ai is it starting to recognize things that we as humans have missed in the data so not so much in Rembrandt I haven't seen anything new in size in Rembrandt but for example something like Jackson Pollock a kind of algorithmic analysis of Jackson Pollock has revealed the Pollock is doing something very special when he splatters paint around that he's creating a very special mathematical shape that we can actually analyze and kind of judge a kind of the fractal the greenness of this a dimension of these paintings so a is giving us new insight there's a wonderful story I tell in the book about the Netflix algorithm that just took our
这就是把技术用在它根本不该用的地方。不过说实话,一个身为艺术评论家却穿成那样衬衫的人,我真不太确定该有多信任他的评论。但在某种程度上他说得也有道理,你知道,再造一个伦勃朗到底意义何在?我确实认为是有意义的,因为这个 AI 最了不起的地方在于,它开始识别出我们人类在数据中忽略掉的东西。在伦勃朗身上倒不太明显,我还没在伦勃朗身上看到什么新洞见,但比如说杰克逊·波洛克,用算法去分析杰克逊·波洛克,就揭示出波洛克在泼洒颜料时其实在做一件非常特别的事,他创造出了一种非常特殊的数学形状,我们真的可以去分析它,去评判这些画作的分形特性、它的分形维度。所以 AI 确实给了我们新的洞见。我在书里讲过一个很棒的故事,关于 Netflix 的算法:它只拿到我们对电影的喜欢和不喜欢,以及
便签笔记
28:24
likes and dislikes of films and the numbers of the films didn't know anything about the films but just from our likes and dislikes is able to clump them together into films of a similar sort of genre so you can see oh yeah look these are all comedy films these are all thrillers but every now and again it would clump films together because on our likes and dislikes which is kind of expressing our common feelings for a film in ways that we didn't really have a aim for that genre it was almost as if the AI had spotted through our likes and dislikes that there was a kind of way of clumping films together that deserved a new name it had spotted a new sort of structure in the films that we hadn't kind of named so I think there is a point about looking backwards but I think the most exciting thing is looking forward can we get the AI to do new things to to break them all to do exciting new things so here's your next challenge four of these paintings are done by a human four of these paintings are done by an AI you have to now jerk
电影的编号,它对电影本身一无所知,但仅凭我们的好恶,它就能把它们聚成风格类似的一类,所以你能看出来,哦对,这些都是喜剧片,这些都是惊悚片。但时不时地,它会根据我们的好恶把一些电影聚在一起,而这其实是在以某种方式表达我们对一部电影的共同感受,只是我们从来没有为那种类型起过名字。就好像AI 通过我们的好恶发现了:存在一种把电影归类的方式,值得给它起个新名字。它在电影中发现了一种我们从未命名过的新结构。所以我认为回望过去是有意义的,但我觉得最令人兴奋的还是向前看:我们能不能让 AI 去做全新的事情,去打破常规,去做激动人心的新东西?那么这是给你们的下一个挑战:这里有四幅画是人画的,四幅画是 AI 画的,你们现在要判断哪边是哪边。我们还是按
便签笔记
29:23
judge which is which so we'll do the same since I've flipped we will do the left-hand one again so the left-hand one is the one you're going to be voting on do you think the four paintings on the left are by the human or by the AI so let's turn you turn it over to you so your chance to vote those four paintings on the left are they okay so now all look there's a brexit vote going on there yeah so you seem to be still going for the lip it's much less convinced by that but there's it's just edging a little bit over to the oh yeah still changing all right but I think that's you think that one's the AI one okay let's go back and see what I think it's an awesome supporter over there okay so which one was which no in fact those were for blue so a few people yes yes I knew that yes so in fact so you find that one a bit more difficult it's interesting because in some ways I would say that the four which are produced by AI have much greater complexity to them and these were actually shown at Basel Art Fair a
同样的方式来,既然我已经调换过了,我们还是投左边那组。也就是说,左边这四幅是你们要投票的对象,你们觉得左边这四幅画是人画的还是 AI 画的?好,交给你们,这是你们投票的机会:左边那四幅画到底是……好,看啊,这简直是在搞脱欧公投了。是啊,你们好像还是更倾向于左边,不过这次没那么有把握,票数还在一点点往那边挪,哦,还在变。好吧,我想你们认为那组是 AI 画的。好,我们回过头看看结果。那边有个热情的支持者啊。好,那么哪组是哪组呢?其实那些是蓝色的那组,所以有一些人——对对,我就知道——所以实际上,你们觉得这一轮更难一些。这很有意思,因为在某种意义上我会说,AI 生成的那四幅画复杂度要高得多。这些画其实几年前在巴塞尔艺术展上展出过,
便签笔记
30:39
couple of years ago I think this is 2016 these were and nobody was told there was any AI involved in this they were just asked to give their feedback on the paintings and the feedback on the AI one so they were people were much more emotionally engaged with the AI ones than they were with the human ones and then of course you see when you then tell somebody well in fact that was created by a computer really upsets people I think you know oh my gosh I had intimate but that's you know there's no emotional world going on inside there I think this is really interesting the reaction one has to to experiencing something then finding out it's done by AI and I think most of you but I mean I also feel like I've been cheated somehow um but if I tell you a joke you know Ellen's jokes were all actually made by an AI and you laughed at them all but then if I tell you now no they actually just ran do the AI joke app does that invalidate your laughter I don't think it does but why I don't think we should get too threatened by
我想是 2016 年吧。当时没有人被告知这里面有 AI 参与,人们只是被请去对这些画作发表感受。而对 AI 那组画的反馈是,人们在情感上比对人类创作的那组投入得多得多。然后当然,你会看到,当你告诉别人:其实那是电脑创作的,人们真的会很不高兴,你知道,会说天哪,我刚刚还那么投入,可是那里面根本没有情感世界。我觉得这非常有意思:人先是体验到某种东西,然后发现它是 AI 做的,那种反应。我想你们大多数人——我的意思是,我自己也会觉得像是被骗了一样。嗯,但如果我给你讲个笑话,你知道,比如说艾伦的那些笑话其实都是 AI 编的,而你们都笑了,那如果我现在告诉你,其实他只是打开了那个 AI 笑话应用,这会让你的笑声变得不作数吗?我不这么认为。而我觉得我们不该被这件事吓到,原因在于:这个 AI 是在
便签笔记
31:40
this is because this AI is learning on our emotional world to produce its next step so it's not disconnected it has got an emotional word it's representing our emotional world but in a sort of new filter what's interesting about this project I think especially is the way that these paintings were created because these four paintings on the right were not created by one algorithm but two algorithms almost working in competition against each other they almost made it into a game something called a creative or generative sometimes adversarial Network so the first algorithm was tasked with creating the art and what it did was to learn on all of the art of the past and it learnt it became a kind of art historian it learned how to classify art in particular styles it understood when something was Cuba start or pointillist art by doing a machine learning process or on the images and being told what was in which particular style then it was tasked with creating something that didn't fit into any of those styles so
我们的情感世界上学习,从而产生它的下一步。所以它并不是与我们脱节的,它确实有一个情感世界,它在呈现我们的情感世界,只不过经过了一层新的滤镜。我觉得这个项目特别有意思的地方在于这些画是怎么被创作出来的,因为右边这四幅画不是由一个算法生成的,而是两个算法几乎在互相竞争中完成的。他们几乎把它变成了一场博弈,叫做创造性网络,或者生成式网络,有时也叫对抗网络。第一个算法的任务是创作艺术品,它的做法是学习过去所有的艺术作品,它学会了——它变成了某种艺术史学家,学会了如何把艺术按特定风格分类。它能识别出某件作品是立体主义还是点彩派,做法是对这些图像进行机器学习,并被告知哪幅属于哪种风格。然后它被要求创作出不属于上述任何一种风格的东西,所以它真的是在努力打破窠臼,
便签笔记
32:43
it was really trying to break the mold it had to make something that couldn't be classified given the parameters that it had learned but it was also tasked with creating something that we as humans would recognize as art so it had already learned from the all all the art last 1500 years what we regarded as art and so it knew kind of an upper limit of where how much it could push the idea the second algorithm was tasked it was the discriminator algorithm was tasked with either saying look that I think that's still stuck you're still stuck in Cuba start there or else was saying you've gone wait fahren that is an art at all and it was the competition between these two that ultimately led to these images now that's what's exciting with this particular idea is because very often algorithms can just churn out loads of things but the challenge is choosing which ones are interesting so we had a second algorithm which was doing some choosing and discriminating and this is very close I think to actually how
它必须做出一件在它已学到的参数下无法被归类的作品。但与此同时它还被要求创作出我们人类仍能认可为艺术的东西,所以它已经从过去一千五百年的全部艺术中学到了我们眼中什么算是艺术,因此它知道自己能把这个概念推到多远的上限。第二个算法的任务——也就是判别器算法——它的任务要么是说:我觉得这还是老一套,你还困在立体主义里;要么是说:你走得太远了,这根本算不上艺术。正是这两者之间的竞争,最终产生了这些图像。这个想法之所以让人兴奋,是因为算法往往可以批量生产出一大堆东西,但真正的挑战在于挑出哪些是有意思的。所以我们有了第二个算法,专门负责挑选和甄别。我觉得这其实非常接近
便签笔记
08创造的双人舞:生成与筛选
33:40
humans work creatively his Paul clay talking about the act of creation already at the very beginning of the productive act shortly after the initial motion to create occurs a thirst counter motion the initial movement of receptivity this means that creator controls where the water is produced so far is good there's always that you know you do something and then you just is that I'm not sure I'm gonna throw it away I'll do something again here's Paul valéry a French poet talking about the idea of these two kind of mind sets um it takes two to invent anything the one makes up combinations the other one chooses and I certainly find that as a mathematician that I have collaborators around the world where we kind of play these two roles so I have a collaborator in Germany where I'm the kind of mad creator and he's the discriminator kind of knocking things down whilst a collaborator here in the Middle East that I have he's the kind of mad creator and I'm the discriminator in that case and it's by doing that kind of
人类的创作方式。这是保罗·克利谈创作行为的一段话:在创作行为的最开始,在最初的创作冲动出现后不久,就会产生第一个反向动作,也就是接受性的最初运动。这意味着,创作者要判断自己迄今为止所做的东西好不好。总有那么一下,你知道,你做出点什么,然后你会想:我不确定我要不要把它扔掉。接下来我再讲点别的。这位是法国诗人保罗·瓦莱里(Paul Valéry),他谈到过这两种思维方式的想法发明任何东西都需要两个人:一个负责组合,另一个负责选择。作为一名数学家,我确实深有体会我在世界各地都有合作者,我们大致就扮演着这两种角色。比如我有一位合作者在德国,我是那个疯狂的创造者,而他是鉴别者,负责把想法一个个推翻;而我在中东的另一位合作者他是那个疯狂的创造者,我则是鉴别者。正是通过这样的组合,我们才能一起
便签笔记
09文字:诗歌成功,小说跑偏
34:35
combination that we actually make progress together okay so we done the visual world what about the written word and the written word we already heard how text a kind of prediction produces some strange effects the written word interestingly a is having quite a lot of difficulty with but then again it's one of the first things that a I was actually interested in in trying to do this is the Manchester Universal computer after touring left Bletchley Park he went up to try and realize some of his ideas in Manchester and the team there were rather perplexed when letters started appearing around the lab which were kind of love letters and written by the Manchester Universal computer and they were sort of perplexed by this until one of the team admitted that in fact he'd written a a program for the computer which was a template and he was using a Rand number generator that touring had just created for the computer which was randomly filling in the template with words amorous words so after a while you'd
取得进展。好,视觉领域我们讲完了,那文字呢?关于文字,我们刚才已经听到文本这种预测方式会产生一些奇怪的效果。有意思的是,文字恰恰是 AI 相当吃力的领域,不过话说回来,这也是 AI 最早想尝试去做的事情之一这是曼彻斯特通用计算机。图灵离开布莱切利园之后去了曼彻斯特,想把他的一些想法付诸实现,那里的团队相当困惑,因为实验室里开始出现一些信件,是那种情书署名是曼彻斯特通用计算机写的。大家一头雾水,直到团队中有人承认其实是他给计算机写了一个程序,就是一个模板,然后用图灵刚刚为这台计算机写的随机数生成器,随机往模板里填入词语,那些甜言蜜语的词。所以过一阵子你就会
便签笔记
35:38
spot the template so not very good but poetry is somewhere where AI has been quite successful I think partly because it's again nice closed form it's not asking sort of too much large scale structure also I think that ah actually you know as an audience you bring a lot of your own creativity when you work look at a piece of art don't you I mean yeah I think that's the point an artist leaves room for your own will to fill things with this and ambiguity to things and so you know poetry has a slight nomic quality that I think especially is something that you bring a lot of your creativity to when you read it um so here are your challenges now I've got some poems for you I want you to vote whether you think these poems are precious guys already on oh god poetry now so whether these poems are by AI or are they by humans so but or not okay so here's your first poem so I'll read you the poem and then we'll go to see what you think about it and I won't read it also mortal my mates burying my rocker
看出模板来,所以效果并不好。不过诗歌是 AI 相当成功的一个领域,我想部分原因是它同样是一种漂亮的封闭形式,不要求太宏大的结构。另外我也觉得,其实作为受众,你在欣赏一件艺术品时,自己也带入了大量的创造力,对吧?是的,我觉得这正是关键:艺术家会留出空间,让你用自己的意志去填补,给作品留有模糊性所以你知道,诗歌带有一点谜语般的特质,我觉得尤其是读诗的时候,你会带入很多自己的创造力。好,那么下面是给大家的挑战。我准备了几首诗,想请大家投票判断这些诗是……大家已经开始了?天哪,现在轮到诗歌了。判断这些诗是 AI 写的还是人类写的。好,这是第一首诗。我先把这首诗读给大家听,然后我们看看大家怎么想。诗是这样的:“凡人啊,我的伙伴,埋葬我摇晃的心,暖搏与冷搏相伴”
便签笔记
36:44
heart warm beats with cold beats company shall i earlier or you fail at our force and lie the ruins of rifles once a world of art okay no stop yeah so do you think that is sport or not so now you're just voting row at red for robots blue for okay I think okay so it's a massive boat for red they're a few thinking it's human and blue okay so that was your first challenge so let's go back to your next challenge I won't reveal them yet okay I'll reveal I've got three poems for you okay so here's your next one this is quite different even to read there are small nurses of pesticide reaction of real time of packs of displaced exclusionary heart hurt of powerlessness of magazine fired non dignified as head fatty employed internalized violence a frozen helpless this is off-white Cholula did okay um so that is AI or do you think that's human okay so let's over to you to vote am i messing with you or you know it was good okay so you think yeah that's so you that's quite a little yeah a little bow
“我该早些,还是你败于我们的力量,躺在曾经那个艺术世界的枪械废墟之中。”好,停。那么你们觉得这是不是机器写的?现在开始投票,红色代表机器人,蓝色代表人类。好,我看到一大片红色,也有少数人觉得是人写的,举了蓝色。好,这是第一道题,我们接着看下一道题,答案我先不揭晓。我一共准备了三首诗。好,这是下一首这一首很不一样,连读起来都不一样:“细小的农药护士,实时的反应,被迫迁徙的排他心群,伤痛之心,”“无力感,弹匣射出,无尊严如脂肪之首,被内化的暴力,一种冻结的无助。”这是……好吧,那么这是 AI 写的,还是你们觉得是人写的?好,交给大家投票。我是在耍你们吗?还是说它其实挺不错?好,你们觉得……嗯,这次有一些人投了人类,不过还是有人
便签笔记
38:04
for human there but still some people thinking okay is this a double bluff because that's clearly code okay so you I think you've gone for human there okay so right this is working okay so um here's your next challenge ports or not imagine now the dark smoke awakened to fly all these years to another day notions of tangled pries the other side of water I see it is already here sequences of her face series shared and old friends passed their dreams but or not okay so over to you you think they're all bought yeah okay so you're going for bought on that one so so that's right so you think they're all being they're all a I that's a sort of thing I would do isn't it you know okay so let's see so you've gone for the for the bots on that one Sagan oh there you into human yes that's right they sorry you're absolutely right yeah thank you very keen that one up yes so let's go back let me give you the answers then so the first one you or pretty convinced that was AI paroled Jarrell Manley Hopkins will be turning in his grave
在想:这会不会是双重迷惑?因为这明显像是代码。好,我看你们这次选了人类好,这个办法行得通。好,那么下一道题:机器还是人?“此刻想象,黑烟苏醒,飞越这许多年,飞向另一天;纠缠的祈求之念,在水的另一边,我看见它已经在这里;她面容的序列,共享的系列,还有老友们逝去的梦。”机器还是人?好,交给大家。你们觉得都是机器?好,那你们这次选的是机器。没错,你们觉得这几首全是 AI 写的——这确实是我会干的事,对吧。好,我们来看看。你们这次投给了机器人……哦,你们那边投了人类?是的,没错,你们完全正确,谢谢,眼光很毒。好,我们回过头来我来公布答案。第一首,你们相当确信是 AI 写的——其实作者是杰拉德·曼利·霍普金斯(Gerard Manley Hopkins)他要在坟墓里翻身了。你怎么看?我选霍普金斯,是因为我从来
便签笔记
39:18
what do you think honorable I chose Gerald Manny Hopkins because I've never understood any poem that Gerald Manny Hopkins has ever written so didn't sniff that one out second one yes you did sniff out that okay that's that's too much like code it can't be an AI it's actually a young Australian poet called miss breeze and she's very interested in this kind of interplay between computer code having its own kind of poetry and rhythm to it yet having meaning that might have something to say to us and so she's very interested in this kind of weird interface between the two actually we've had two humans so that leaves the last one is the human did I really mess with you no the last one you sniffed out actually it was the only one that made any sense at all was in fact the only one crazy by an AI this is actually Ray Kurzweil created something called the cybernetic poet and this is a machine learning process where he took poems of Yeats Keats Eliot and then the the bot was kind of tasked with creating something was a kind of fusion
没读懂过霍普金斯写的任何一首诗,所以这一首大家没识破。第二首,是的,你们识破了。你们说这太像代码了,不可能是 AI——它其实出自一位年轻的澳大利亚诗人叫 Miss Breeze,她非常关注这样一种互动:计算机代码本身也具有诗意和节奏,同时又承载着某种可能对我们言说的意义,所以她非常着迷于这两者之间奇特的交界地带。这样我们已经有两首是人写的了,那最后一首也是人写的?我真是在耍你们吗?不,最后一首你们识破了——其实唯一读起来还算通顺的那一首,恰恰是唯一由 AI 写的。这其实是雷·库兹韦尔(Ray Kurzweil)做的他创造了一个叫“控制论诗人”(Cybernetic Poet)的东西,这是一个机器学习过程,他输入了叶芝、济慈、艾略特的诗然后让这个程序去创作,比如生成叶芝和艾略特的某种融合体。当然
便签笔记
40:21
of Yeats and Eliot for example so of course Rico as well is one of the people who's talking about the idea of the singularity the moment when computers might actually be more intelligent than humans and so that moment the singularity and so poetry energy is not doing too bad it's a kind of longer scale writing that AI is having real difficulty at it so it can generate quite interesting sort of text right generation or sort of small scale but it's kind of the idea of writing a novel is still way beyond it although there have been some attempts to write novels so there was a very interesting case by a team called Bach botnik they decided their big Harry Potter fans and they decided they were very disappointed there are only seven volumes of Harry Potter and they wanted an eighth they want to know what happens next um so what they did was they got the machine learning to take all of JK Rowling's writing learnt kind of ideas that she's interested in her style of writing and decided they would create an
库兹韦尔也是那批谈论“奇点”概念的人之一——奇点就是计算机可能真正比人类更聪明的那一刻。所以在诗歌上 AI 表现得还不赖,真正让它非常吃力的是更长篇幅的写作。它可以生成相当有意思的文本,小规模的片段没问题,但是要写一部小说,对它来说还差得很远,尽管确实有人尝试过让 AI 写小说有一个非常有意思的例子,是一个叫 Botnik 的团队。他们都是哈利·波特的铁杆粉丝他们觉得很失望:哈利·波特只有七部,他们想要第八部他们想知道后面发生了什么。于是他们用机器学习读入了 J.K. 罗琳的全部作品,学习她感兴趣的题材和她的写作风格,然后决定做一个算法来生成第八部。所以
便签笔记
41:21
algorithm to create an eighth book so here is the beginning of this eighth book I actually love the title this is called Harry Potter and the portrait of what looked like a large part of Ashe I'd read that I'd read that so that it starts off pretty well magic it was something that Harry Potter thought was very good so good it's already picked up that these books are about magic you know no bad leathery sheets of rain lashed at Harry's ghost as he wore leather he sheets of rain I think that's a beautiful image leathery she's the rain as he walked across the grounds towards the castle after this it began to lose the plot a little bit um ron was standing there and doing a kind of frenzy tap-dance he saw Harry and immediately began to eat Hermione's family very much good Ron's Ron shirt was just as bad it's Ron himself so AI is having it's very good a kind of local generation of things which have some sort of meaning but it doesn't have many very good sense of a long term structure okay what about music
这就是第八部的开头。我特别喜欢这个书名,它叫《哈利·波特与看起来像一大堆灰烬的肖像》这书我真想读一读。开头写得还挺不错的:“魔法,这是哈利·波特认为非常好的东西。”很好,它已经学到这些书是讲魔法的了接下来就不太行了:“皮革般的雨幕抽打着哈利的幽灵……”皮革般的雨幕,我觉得这是个很美的意象——皮革般的雨幕,当他穿过场地走向城堡时。再往后就开始有点跑偏了:“罗恩站在那里,跳着一种狂乱的踢踏舞。他看见哈利,立刻开始吃赫敏的家人。”很好。罗恩的衬衫也一样糟,那就是罗恩本人。所以 AI 很擅长局部地生成一些具有某种意义的内容,但它对长程结构把握得不太好。好,那音乐呢?可爱的 Seder 给我们出了这个难题:AI 能不能创作音乐?音乐里充满了大量的模式
便签笔记
10音乐:巴赫与 AI 的拼接实验
42:28
lovely Seder gave us this challenge of whether a I could produce music and music has full of lots of patterns a composer when you hear a piece of music on the radio you can probably very quickly pick out what the composer is because they have particular styles particularly the sort of sound world that they have can the AI learn that and be able to produce something at a particularly good level to to replicate or do something new so AI always starts on Bach Bach is where AI always us because Bach has a lot of algorithms at work if you do look at something like the musical offerings Bach wrote the musical offerings these little pieces he wrote for Duke Ferdinand as as I've kind of puzzled to solved there was a little algorithm you had to show which you had to expand and and see what the music actually meant so so actually Bach is a very good place actually for AI to start and something you may remember a few weeks ago and the Google Doodle did you have a go on the Google doodle it was celebrating box and birthday and this
当你在广播里听到一段音乐时,你多半能很快听出作曲家是谁因为他们各有独特的风格,尤其是他们特有的音响世界。AI 能不能学到这一点并在相当高的水准上复制它,或者创造出新的东西?AI 总是从巴赫开始。AI 总是从巴赫入手,因为巴赫的音乐里有大量算法在起作用。如果你看看《音乐的奉献》,巴赫写的《音乐的奉献》,那些是他为费迪南公爵写的小曲子,就像一道道待解的谜题:里面有一个小算法,你得把它展开,才能看出这段音乐到底是什么意思。所以巴赫其实是 AI 非常好的起点。你们可能还记得几周前的 Google Doodle,你们玩过那个涂鸦吗?那是为了纪念巴赫的生日
便签笔记
43:28
Google doodle you could put in a line of music and then it would harmonize the other three voices so the machine learning it to take this is based on something called magenta it had taken all the corrals that bark had written Corral's are very good because they kind of they don't change a key very often they're very sort of boring and but it can learn a lot about the way that a crawl and a coronis Phinney in the harmony is a bit like filling in a Sudoku you've got just learn the rules so this would actually the little Google doodle would fill in with your tune so I actually put in the a bit of the musical offerings so this is my attempt on it because the musical offering is actually quite a difficult challenge it doesn't seem to have any key to it at all and when I played it back it was really rubbish actually but what was very nice about the Google doodle is you could then say I thought that is rubbish and the thing would learn and actually say okay that harmonizing didn't work and so you were training the
在那个涂鸦里,你可以输入一段旋律,它会把另外三个声部的和声配出来。这个机器学习模型——它基于一个叫 Magenta 的项目——学习了巴赫写过的所有众赞歌众赞歌很合适,因为它们不太频繁转调,比较“无聊”,但是从中可以学到很多众赞歌配和声的方式,配和声有点像填数独,你只要学会规则就行所以这个小小的 Google 涂鸦会为你的旋律配上和声。我当时输入的是《音乐的奉献》的一小段,这是我的尝试,因为《音乐的奉献》其实是个相当难的挑战,它似乎根本没有明确的调性。播放出来的结果其实相当糟糕。不过 Google 涂鸦有个很棒的地方就是你可以告诉它:我觉得这个很烂。然后它会学习,说好吧,这个和声不行。所以你给出反馈的同时就在训练这个算法;如果你觉得好,它就会
便签笔记
44:26
algorithm as you gave your feedback and if you thought it was good it would kind of we privatize it so okay I'll do more of that sort of harmony so very nice so what we've done and I've got a actually set up a Center in the Royal Northern College of Music with a composer Emily Howard it's a called prism which stands to practice and research in science and music and what we're here any interested in is this kind of interplay between the scientific and mathematical world and the world of composition so we've done various projects we wrote a string quartet representing mathematical proofs and together but I've now got a PhD student who's a composer which is really exciting Rob laid low he actually had his quartet premiered last week at the Wigmore Hall he was going to be here tonight unfortunately he's ill which is a real shame but he's been working alongside me and various other people and she here in the university as well and to produce a piece using some of the best software we have around at the moment to create a
强化这一点,说好,那我多写点这样的和声。非常不错。所以我们做的事情是——我其实和作曲家 Emily Howard 一起,在皇家北方音乐学院成立了一个中心,叫 PRiSM,全称是“科学与音乐中的实践与研究”。我们感兴趣的正是这种互动:科学与数学的世界,和作曲的世界之间的互动。我们做过很多项目,比如写过一首弦乐四重奏来表现数学证明。现在我还带了一位学作曲的博士生,这非常令人兴奋,他叫 Rob Laidlow,上周他的四重奏刚在威格摩尔音乐厅首演。他本来今晚要来的很遗憾他病了,真的很可惜。但他一直和我以及其他一些人合作,还有大学里的同事一起,用我们目前能拿到的最好的软件创作了一首曲子,做出一种
便签笔记
45:28
kind of hybrid AI bark piece and and this is what we want to play you so the challenge here is so we've got a fantastic musicians in the department here and Kobe is going to come up and play this piece for you which is so in there we've just been doing very straight AI human tests this is a slightly different challenge you've now got a piece which sometimes is AI and sometimes is bark and I'm not going to tell you how many times it moves between one and the other and what I want you to do is to see whether you can see where the joins are can you tell oh gosh no that's horrible that's gone AI or oh yeah that's bark so we're going to run this twice the piece is about four minutes long and so we're going to record your thoughts on this and then hopefully if we can do this I'm gonna play this back to you oh gosh because now we don't have two screens now it's gonna be quite okay no III know what I can do okay yeah I've thought of an improvised way of doing this I'm going to do it very analog
AI 与巴赫的混合作品。这就是我们想放给大家听的。所以这里的挑战是——我们系里有非常棒的音乐家Kobe 待会儿会上台为大家演奏这首曲子。刚才我们做的都是很直接的“AI 还是人类”的测试,这次的挑战略有不同:这首曲子有时是 AI 写的有时是巴赫写的,我不会告诉你们它在两者之间切换了多少次。我想请你们做的,是看看能不能听出接缝在哪里。你能不能察觉到——天哪,不,这太难听了,这肯定切到 AI 了或者,对,这是巴赫。我们会放两遍,这首曲子大约四分钟长,我们会记录你们的判断,然后如果可行的话,我再回放给你们看。糟糕,因为我们现在没有两块屏幕,这就有点……好吧,不,我知道该怎么办了。好,我想到了一个临时的办法,我要用非常“模拟”的方式来做,待会儿你们就会看到我的解决方案。好,那么也许我们先
便签笔记
46:31
so you'll see what my solution in a minute okay so so perhaps we can give a big round and pause to Kobe who's going to come and play this piece we have a page-turner as well and so I just to set this thing off can you all show your blue faces so we're going to just start you off all on blue so the idea is as soon as you think that the music has gone into something which is not by bark you turn to the red face and if you think it's gone back to bark you move to the blue vase okay it's very simple okay so now in order to sync these things I'm going to have to count this down so we go hopefully we can show these two things again so I've got you ready ready combing okay so I'll go 4 3 2 1 [Music] [Applause] [Music]
用热烈的掌声欢迎 Kobe 上台演奏这首曲子,我们还有一位翻谱员。为了开始这个环节,请大家都举起蓝色的牌子,我们就从全蓝开始。规则是:一旦你觉得音乐进入了不是巴赫写的部分,你就翻到红色那一面;如果你觉得又回到巴赫了,就翻回蓝色。好,很简单。现在为了让这两件事同步,我得倒数一下。希望我们能把这两样东西一起显示出来。好,大家准备好了吗?Kobe 准备好了吗?好,我数 4、3、2、1。[音乐][掌声][音乐]
便签笔记
51:26
I was raising I think there are moments when bar will be turning in his grave as you'll see but it's inching because I think when it was barks there was much more confidence in your knowledge of that he could feel it was right but when it wasn't it was really kind of edgy and there were moments when his sunny surge read which they were give away moments so so we're going to replay this so you can actually see what what's your answers were this was actually so what we did was we took one of the English sweets so this is the fourth English sweet and what we did was to take these bits out of the piece of music and then ask the AI to fill in the gaps and the machine learning was it's interesting because we actually use quite a simple piece of machine learning it's called clora it's developed now it's developed by open AI this team which is trying to make AI very open to the world something that Elon Musk has helped set up and so this is actually buy a piece of software written by Christine Payne as part of the open AI
我刚才在想,有些时刻巴赫真的要在坟墓里翻身了,你们等下就会看到。但这很有意思,因为我觉得当它是巴赫的时候,你会更有把握,能感觉到这是对的;而当它不是的时候,就会让人觉得很别扭。有些地方那种突兀的走向就是明显的破绽。所以我们要再放一遍,让你们看到自己的答案是怎样的。我们其实是这样做的:我们取了一首英国组曲——这是第四号英国组曲——我们把这段音乐中的一些片段挖掉,然后让 AI 来填补空缺。这个机器学习很有意思,因为我们用的其实是一个相当简单的机器学习模型,叫 Clara,现在由 OpenAI 开发,这个团队致力于让 AI 对全世界开放,埃隆·马斯克参与创办了它。所以这其实是 Christine Payne 作为 OpenAI 项目的一部分写的一款软件,而且还在不断发展,我们会
便签笔记
52:41
project and it's growing so we're going to keep on working on this to to make it even better with something called muse net but what's interesting is that this is kind of predictive in the sense of it here's what's happened up to date and then makes a decision about what will be next so some of the AI is working very cleverly in working backwards as well so knowing where the piece is going it can make some prediction but this piece of software does not have a long-term memory and this is one of the challenges to create a piece of music software that can actually know about what it's done in the past and exploit that in its decisions as it goes forward and one of the things I talked about with Cobie beforehand he could feel the AI very clearly because the AI is not embodied so it has no trouble with giving really awkward fingerings whilst Bach was writing something that would fit very nicely under the fingers um so I think is one of the challenges of AI very generally is the idea of it not being
继续用一个叫 MuseNet 的东西把它做得更好。有意思的是,这个模型本质上是预测式的:它看到到目前为止发生了什么,然后决定接下来会是什么。有些 AI 做得很聪明,还会反向工作,也就是知道曲子要往哪里去,从而做出一些预测。但这款软件没有长期记忆,而这正是挑战之一:如何做出一款音乐软件,能真正知道自己之前写过什么,并在后续的决策中加以利用。我事先和 Kobe 聊过一件事:他能非常清楚地感觉到 AI 的部分因为 AI 没有身体,所以它会毫无顾忌地写出非常别扭的指法,而巴赫则是……写出来的东西非常贴合手指的运动,嗯,所以我觉得 AI 面临的挑战之一,非常普遍地说,就是它没有身体、没有具身性,你真的能感觉到这一点,我是说我第一次
便签笔记
53:44
embodied and you could really feel that that it just did I mean when I first playthrough it was just like wow this is just really gooey um okay so let's play that back so what I was hoping to do is have things on one screen and the other but I think what I would do so we're going to play oh yeah exactly so let's say we'll get the thing up here so we're going to get poorer Kobey's got to play this again the painful the bits the bark is great fun and what I will do analog wise and so what's going to do this on the screen but I will just show you as we're going along which bits are AI and which are not and so you'll be able to see what you voted and I'll tell you what the answers are okay all right okay so as you I sort of slightly bias things by asking you to put all up the human face I wanted to see how long it would take until you actually spotted that the opening wasn't bark at all it was in fact so it took quite a long time so we're going to start so we're gonna count this down so we get the sinking right hopefully so
从头弹一遍的时候,就觉得,哇,这也太黏腻了,嗯,好,那我们回放一下,我本来想做的是在一个屏幕上放一样东西,另一个屏幕放另一样,不过我想我会这么做,我们来播放——哦对,没错,那我们就这么说吧,我们把东西放到这上面来,所以我们要请可怜的 Kobe 再弹一遍,那些折磨人的段落,巴赫的部分非常好玩,我在现场会做的是,这些会显示在屏幕上,但我会在我们进行的过程中直接告诉你们哪些段落是 AI 写的,哪些不是,这样你们就能看到自己投了什么票,我也会告诉你们正确答案是什么,好,好的,好,其实呢,我刚才稍微误导了大家一下,因为我先让你们为“人类”举手,我是想看看要过多久你们才会发现,开头那一段根本不是巴赫,实际上是——所以你们花了相当长的时间才反应过来,那我们开始吧,我们来倒数,这样时间点能对上,希望如此,好
便签笔记
54:44
there you all are ready so uh four three two one [Music] you pricky you picked that up quite quickly that that was bark I think you can hear that it's got some sort of direction to it [Music]
大家都准备好了吗,那么,四、三、二、一 [音乐] 你们这次挺快就听出来了,那段是巴赫,我想你们能听得出来,它有某种方向感 [音乐]
便签笔记
56:22
robach [Music]
巴赫 [音乐]
便签笔记
57:03
oh now you can be the Cabal [Music] [Applause]
哦,现在轮到你们了 [音乐] [掌声]
便签笔记
57:48
[Music] [Applause] [Music] and thank you very much to Kobe now one of the things I was very strict about with Rob the composer was that it had to be AI and it was not allowed any chance to try and improve the AI so we were very strict about because very often when you look at projects and the many in the book although it's saying it's an involvement of AI you can see there's a lot of human inputs it makes a much better story if you just says the AI and the human isn't involved at all so it's very strict with this that the portions that where AI had to be only AI and interestingly we asked it just to fill in the last chord and it missed out one crucial note which made the whole thing resolved so but you can see from that that actually it was pretty convincing and there were few given way moments but actually quite hard to pick out which was the AI I think it were pretty good on what was bark although there are a few horrific moments when you thought it was AI and bark will be poor turning in
[音乐] [掌声][音乐]非常感谢 Kobe,我在和作曲家 Rob 合作时有一点要求得特别严格,那就是它必须是纯 AI,绝对不允许任何人为地去改进 AI 的输出,我们在这一点上非常严格,因为很多时候你去看那些项目,包括书里提到的很多例子,虽然说是“AI 参与创作”,但你能看出里面有大量的人类输入,如果直接说这是 AI 做的、人类完全没参与,故事会好听得多,所以我们对这一点非常严格,AI 负责的部分就必须是纯 AI 的,有意思的是,我们只让它补上最后一个和弦,结果它漏掉了一个关键音,而正是那个音让整段音乐得以收束,所以,不过从这里你们也能看出它其实相当有说服力,虽然有几个露馅的瞬间,但要挑出哪一段是 AI其实相当难,我觉得你们在判断哪些是巴赫时表现得挺好的,虽然有几个可怕的时刻,你们以为那是 AI,而那其实是巴赫,可怜的老巴赫要气得在坟里翻身了,嗯,我想再说一次,这么做的意义何在呢?我认为
便签笔记
11结论:工具而非威胁
59:52
his grave so um and I think that again what's the point of this well I think there is a point for creative artists and I think this is not about competition this is about collaboration this is a new tool to push our own human creativity we've always already seen that in the realm of art but in music one of the most interesting stories I saw was the idea of an AI that have been trained to play jazz and it's called the jazz continue ater something constructed in sunny labs in Paris by Francois fouché and his team and they got the AI to learn in a very way to a jazz musician learning what the probability is of the kind of next move after a certain sequence of notes um and they then did a concert where people found it very difficult to tell when it was the human playing when it was the AI but what struck me was the jazz musician who at this a I've been trained on his response to hearing the AI play back to him and he said this is Bernhard lumad the system shows me ideas I could have developed but that would have taken me
对创作型艺术家来说是有意义的,我觉得这不是关于竞争,而是关于协作,这是一个能推动我们自身人类创造力的新工具,我们在美术领域已经看到了这一点,而在音乐领域,我看到的最有意思的故事之一,是一个被训练来演奏爵士乐的 AI,它叫“爵士延续器”(Jazz Continuator),是由巴黎索尼实验室的 François Pachet 和他的团队做出来的,他们让这个 AI 用非常接近爵士乐手学习的方式,去学习在某个音符序列之后下一步走向的概率是多少,嗯,然后他们办了一场音乐会,观众很难分辨什么时候是人在演奏、什么时候是 AI 在演奏,但真正打动我的,是那位爵士乐手——AI 正是用他的演奏训练出来的——在听到 AI 把他的东西弹回来时的反应,他叫 Bernard Lubat,他说这个系统展示给我一些我本来也可能发展出来的想法,但那需要我花很多年才能真正发展出来,它比我
便签笔记
60:54
years to actually develop it is years ahead of me yet everything it plays is unquestionably me another thing this is what's exciting because I think that we as humans often end up behaving very much like machines we get stuck in our ways of thinking we just perform the same ideas over and over again especially in creativity I know that in my own mathematics I try the same things over and over again and sometimes I need something to push me out of the way I've been thinking and see that my world of possibilities is much richer because like Bernard lubob was in a room the spotlight was on him he didn't realize there was so much more to play with his in within his sound world so the exciting thing for me is that this movement into an AI that might be creative it's not about a threat it's about an opportunity it's about the fact that this thing could push us to be behaving less like machines and actually become more creative again as humans thank you [Applause]
领先了好多年,然而它弹的每一样东西又毫无疑问就是我,还有一点,这正是让人兴奋的地方,因为我觉得我们人类往往最后表现得非常像机器,我们会困在自己的思维定式里,一遍又一遍地重复同样的想法,在创作中尤其如此,我知道在我自己的数学研究里,我也会一次又一次地尝试同样的东西,有时候我需要某种东西把我从既有的思路里推出来,让我看到我的可能性世界要丰富得多,就像 Bernard Lubat 一样,他待在一个房间里,聚光灯打在他身上,他没有意识到在他自己的声音世界里,还有那么多可以玩的东西,所以对我来说,令人兴奋的是,这场朝向“可能具有创造力的 AI”的变革,它不是一种威胁,而是一种机会,关键在于这个东西能推动我们不再那么像机器,而是作为人类重新变得更有创造力,谢谢大家 [掌声]
便签笔记
62:07
you
谢谢
便签笔记
视频总结 · 一句话概括与核心要点

一句话概括

牛津数学家 Marcus du Sautoy 以 AlphaGo 的"第37手"为切入点,论证自下而上的机器学习已能产生符合"新颖、惊奇、有价值"定义的创造性行为,并通过绘画、诗歌、音乐的现场"AI 还是人类"投票测试,说明 AI 在视觉艺术和短篇诗歌上已相当逼真、在长篇叙事上仍然失败,最终主张 AI 是推动人类跳出思维定势的协作工具而非威胁。

核心要点

  • 围棋曾是数学家的"保护盾",AlphaGo 打碎了它:讲者在 1990 年代深蓝击败卡斯帕罗夫时就被问"你是不是下一个",但他以围棋(19×19 棋盘、依赖模式识别与直觉、传统算法连业余棋手都赢不了)自我安慰;DeepMind 的 Demis Hassabis 正是因为在剑桥被告知"围棋不可编程"而反向发力,最终 AlphaGo 以 4:1 击败李世石,李世石唯一赢下的第四局被他视为"生涯最伟大的一局"。
  • 变化的本质是从"自上而下"到"自下而上"的编码方式:过去程序员必须事先告知规则,机器只是执行(因此创造力归于人类);如今代码像孩子一样在与环境交互中自我重参数化、变异。例证:DeepMind 只给 Atari《Breakout》的像素和分数,随机移动球拍,600 局后自己学会了讲者小时候发现的"左侧打隧道让球自动反弹"的偷懒技巧。
  • 第37手是 AI 创造性行为的标志案例:围棋传统教学要求开局在第 1–4 线布子,第 5 线被视为弱手;第二局李世石抽烟回来看到 AlphaGo 下在第五线,所有 YouTube 解说都惊呼"大错"。但围棋与国际象棋相反——棋子越来越多、局面越来越复杂——随着右下角领地扩展,正是这颗子决定了归属,为 AlphaGo 赢下该局。
  • 创造力的工作定义来自哲学家 Margaret Boden:新颖 + 惊奇 + 有价值:新颖可客观判断,惊奇和价值主观,因此 AI 若要"有创造力"必须学会人类认为什么是惊奇和有价值的。棋类游戏是理想测试场(解说惊呼=惊奇,赢棋=价值);讲者进一步推演:人类以为的"最优打法"其实只是局部极大值(Snowdonia),AlphaGo 通过冒险穿越"适应性谷底"找到了更高的 Everest,并反过来教人类新策略。
  • Lovelace 测试取代 Turing 测试作为艺术创造力标准:Ada Lovelace 早在 19 世纪就预言分析机可"创作任何复杂度的音乐",但同时警告机器"无法原创,只能执行命令"。Lovelace 测试要求:机器产出可重复(非硬件故障或随机噪声)的艺术作品,且编写代码的程序员本人无法解释算法如何得出该输出——这正是机器学习让"代码与编码者脱钩"后才可能满足的条件。
  • 视觉艺术是 AI 最成功的领域,现场投票也证实真假难辨:荷兰团队用约 300 幅伦勃朗肖像训练 AI 生成"新伦勃朗",甚至分析颜料高度做 3D 打印;伦勃朗专家唯一能挑的毛病是"用漆风格比肖像早 20 年"。第二轮测试中观众把 AI 画作误判为人类,这批作品 2016 年在巴塞尔艺术展匿名展出时,观众对 AI 画作的情感投入反而高于人类作品。
  • GAN 的"生成器 vs 判别器"结构与人类创作心理高度同构:生成算法先学习 1500 年艺术史成为"艺术史学家",再被要求创作不属于任何已知流派、但仍被人类认作艺术的作品;判别算法则负责判定"你还困在立体主义里"或"这已经不是艺术了"。讲者引用 Paul Klee 和 Paul Valéry("发明需要两个人:一个组合,一个选择"),并以自己与德国、中东合作者互换"疯狂创造者/挑剔判别者"角色为例。
  • 文字领域两极分化:短诗成功,长篇叙事失败:三首诗测试中观众把 Gerard Manley Hopkins 的诗当成 AI、把澳大利亚诗人 Mez Breeze 的代码风格诗判为人类、而唯一"读得通"的一首恰是 Ray Kurzweil 的"控制论诗人"融合叶芝与艾略特的产物。反例是 Botnik 用 J.K. 罗琳全集训练生成的《哈利·波特与看起来像一大堆灰烬的肖像》:开头尚可,随后罗恩"疯狂踢踏舞并开始吃赫敏的家人"——AI 擅长局部生成,缺乏长程结构。
  • 音乐实验揭示 AI 的两个结构性短板:无长期记忆、无身体:讲者与皇家北方音乐学院 PRiSM 中心的博士生作曲家 Rob Laidlow 合作,用 OpenAI 的 Clara 模型填补巴赫第四英国组曲被挖空的段落,严格禁止人工修饰。钢琴家 Kobe 演奏时观众对巴赫段落投票自信、AI 段落"边缘化",且开头就是 AI 却过了很久才被察觉;AI 甚至在最后一个和弦漏掉了解决所需的关键音。钢琴家能"感觉到" AI 是因为它给出极别扭的指法——巴赫的写法则天然贴合手指。

结论与值得注意的细节

  • 核心立场:AI 创造力是"协作而非竞争"。最有力的证据来自巴黎 Sony 实验室 François Pachet 的 Jazz Continuator:爵士乐手 Bernard Lubat 听到以自己风格训练的 AI 演奏后说,"它展示给我的想法我本可以发展出来,但要花好几年……它领先我好几年,然而它演奏的每一个音都毫无疑问是我。"
  • 反转式警示:讲者认为人类本身常常"像机器一样"陷入固定套路(包括他自己做数学时反复尝试同样的方法),AI 的价值恰恰在于把人类推出思维定势——"让我们不再像机器,重新变得更有创造力"。
  • AI 的"回望"价值:算法分析揭示 Jackson Pollock 泼溅画具有可量化的分形维度;Netflix 推荐算法仅凭用户好恶就聚类出人类尚未命名的新"类型"——AI 能在数据中看到人类错过的结构。
  • 对被 AI 作品"欺骗"的情绪反应:讲者承认得知作品出自 AI 时会有"被骗"感,但反驳说笑话若由 AI 生成并不会让你的笑声失效——因为 AI 是在人类情感世界的数据上学习,它是人类情感的"新滤镜",并未与之断连。
  • DeepMind 已在着手打造"创造性 AI 数学家":讲者问 Hassabis 能否让 AlphaGo 当选皇家学会院士,对方回答"我们已经在做了"。
  • 技术现场翻车的自嘲:演讲一开始 HDMI 切换失灵,讲者调侃"做科技演讲还用科技总是灾难",投票用的卡片识别系统本身也是用约 600 张训练图像做的机器学习分类器(红=机器人、蓝=人类)。
核心句型 · 8
1. be a hard act to follow
“He's becoming quite a hard act to follow these days”
形容前一位表现太好、后来者难以超越。演讲开场承接主持人时常用,也可用于评价离任者。
2. like a red rag to (a bull / sb)
“This was like a red rag to Dennis”
「像红布之于公牛」,指某事反而激起对方的斗志或怒气。可省略 bull 直接接人名。
3. what has changed is that …
“What has changed is that the code is now written in a very bottom-up manner”
以 wh- 名词性从句做主语,突出「变化的关键点」。适合对比过去与现在时引出核心论点。
4. not only … but (also) …
“It not only played the game at a high level but it taught us how to play the game in a new way”
递进结构,第二部分是重点。演讲中用来把「做得好」推进到「教会我们新东西」。
5. what I think is so nice about X is (that) …
“What I think so nice about games it's of course you can judge this kind of these qualities quite quickly”
用 what 从句先聚焦评价对象,再给理由。口语中引出个人观点的常见句式。
6. it takes two to …
“It takes two to invent anything the one makes up combinations the other one chooses”
「…需要两方」,源自 it takes two to tango。后接 the one … the other … 分述两种角色。
7. this is not about X, this is about Y
“This is not about competition this is about collaboration”
否定—肯定的并列,用于重新框定问题。结尾总结时用来纠正听众可能的误解。
8. X will be turning in his grave
“Gerald Manley Hopkins will be turning in his grave”
「某人会气得在坟里翻身」,幽默地表示某事会令已故名人不满。可用于文学、音乐评论。
词汇精讲 · 90 · 按出现顺序
a hard act to follow idiom 0:15
难以超越的前任表演者/前一位讲者
stock /stɑːk/ adj. 0:15
老套的、千篇一律的(stock reactions)
flee /fliː/ v. 0:15
逃离
long division n. phr. 1:16
长除法
end game n. 1:16
残局;最终目标
protective shield n. phr. 2:15
护盾,此处比喻用来抵挡质疑的说辞
engulf /ɪnˈɡʌlf/ v. 2:15
吞没、包围
intuition /ˌɪntuˈɪʃən/ n. 2:15
直觉
encode /ɪnˈkoʊd/ v. 3:13
编码,把…写成程序
adage /ˈædɪdʒ/ n. 4:18
格言、老话
red rag n. phr. 4:18
激怒人的东西(源自 a red rag to a bull)
dismissive /dɪsˈmɪsɪv/ adj. 4:18
轻蔑的、不屑一顾的
demolish /dɪˈmɑːlɪʃ/ v. 4:18
摧毁;(比赛中)彻底击败
obsessively /əbˈsesɪvli/ adv. 4:18
着魔般地
top-down adj. 5:18
自上而下的(由规则驱动)
bottom-up adj. 5:18
自下而上的(由数据驱动)
mutate /ˈmjuːteɪt/ v. 6:17
变异、突变
paddle /ˈpædl/ n. 6:17
(游戏中的)挡板;球拍
hack /hæk/ n. 7:12
取巧的技巧、妙招
staggered /ˈstæɡərd/ adj. 7:12
震惊的
synthetic data n. phr. 8:07
合成数据(机器自我生成的训练数据)
disruptive /dɪsˈrʌptɪv/ adj. 9:12
搅局的、打乱常规的
establishing /ɪˈstæblɪʃɪŋ/ v. 9:12
确立(地盘、地位)
anthropomorphize /ˌænθrəpəˈmɔːrfaɪz/ v. 10:10
拟人化
complacent /kəmˈpleɪsnt/ adj. 10:10
自满的、掉以轻心的
inspired /ɪnˈspaɪərd/ adj. 11:11
极具灵感的、神来之笔的
working definition n. phr. 12:13
工作定义(暂时采用、便于讨论的定义)
on the case idiom 13:12
正在处理此事
local maximum n. phr. 14:12
局部最大值(数学/优化术语)
taken by phr. 15:16
被…带去(be taken by sb to)
elaborate /ɪˈlæbərət/ adj. 16:21
精巧复杂的
guard against phr. v. 16:21
防范
pretensions /prɪˈtenʃənz/ n. 16:21
自命、妄称(have no pretensions to)
originate /əˈrɪdʒɪneɪt/ v. 16:21
原创、发起
disconnect /ˌdɪskəˈnekt/ v. 17:21
切断联系、脱钩
glitch /ɡlɪtʃ/ n. 18:28
小故障
repeatable /rɪˈpiːtəbl/ adj. 18:28
可重复的
hurdle /ˈhɜːrdl/ n. 19:25
障碍、难关
onslaught /ˈɑːnslɔːt/ n. 19:25
猛攻;(信息的)狂潮
distinguish /dɪˈstɪŋɡwɪʃ/ v. 20:32
区分
asking for trouble idiom 21:39
自找麻烦
improvise /ˈɪmprəvaɪz/ v. 22:52
临场发挥、即兴应对
prejudice /ˈpredʒədɪs/ v. 24:06
使有偏见、影响(判断)
edgy /ˈedʒi/ adj. 24:06
紧张不安的
edging over phr. v. 25:14
逐渐倾向、缓慢移向
snooty /ˈsnuːti/ adj. 26:27
傲慢势利的
travesty /ˈtrævəsti/ n. 26:27
拙劣的模仿、歪曲
soulless /ˈsoʊlləs/ adj. 26:27
没有灵魂的
splatters /ˈsplætərz/ v. 27:26
泼溅
fractal /ˈfræktl/ n./adj. 27:26
分形(的)
clump /klʌmp/ v. 28:24
聚成一团、归成一类
genre /ˈʒɑːnrə/ n. 28:24
体裁、类型
emotionally engaged phr. 30:39
情感投入的
invalidate /ɪnˈvælɪdeɪt/ v. 30:39
使无效
adversarial /ˌædvərˈseriəl/ adj. 31:40
对抗性的(adversarial network 对抗网络)
pointillist /ˈpwæntəlɪst/ adj. 31:40
点彩派的
break the mold idiom 32:43
打破常规
discriminator /dɪˈskrɪmɪneɪtər/ n. 32:43
判别器(GAN 中的判别网络)
churn out phr. v. 32:43
批量产出
receptivity /ˌriːsepˈtɪvɪti/ n. 33:40
接受性、感受力
knocking things down phr. 33:40
推翻、否决想法
perplexed /pərˈplekst/ adj. 34:35
困惑的
amorous /ˈæmərəs/ adj. 34:35
示爱的、多情的
closed form n. phr. 35:38
封闭形式(结构固定、篇幅有限)
ambiguity /ˌæmbɪˈɡjuːəti/ n. 35:38
模糊性、多义性
double bluff n. 38:04
双重虚张声势、反向欺骗
turning in his grave idiom 38:04
气得在坟里翻身
sniff out phr. v. 39:18
嗅出、识破
interplay /ˈɪntərpleɪ/ n. 39:18
相互作用
fusion /ˈfjuːʒən/ n. 39:18
融合
singularity /ˌsɪŋɡjəˈlærəti/ n. 40:21
奇点(机器智能超越人类的时刻)
lashed /læʃt/ v. 41:21
抽打
lose the plot idiom 41:21
失去条理、跑偏
frenzy /ˈfrenzi/ n. 41:21
狂乱
replicate /ˈreplɪkeɪt/ v. 42:28
复制、再现
harmonize /ˈhɑːrmənaɪz/ v. 43:28
配和声
rubbish /ˈrʌbɪʃ/ n./adj. 43:28
(英)垃圾;很烂
premiered /prɪˈmɪrd/ v. 44:26
首演
hybrid /ˈhaɪbrɪd/ adj. 45:28
混合的
joins /dʒɔɪnz/ n. 45:28
接缝、拼接处
give away phr. v. 51:26
露馅、泄露(give-away moments)
predictive /prɪˈdɪktɪv/ adj. 52:41
预测性的
embodied /ɪmˈbɑːdid/ adj. 52:41
具身的、有身体的
fingerings /ˈfɪŋɡərɪŋz/ n. 52:41
指法
gooey /ˈɡuːi/ adj. 53:44
黏糊糊的;(比喻)腻的
resolved /rɪˈzɑːlvd/ v. 57:48
(音乐)解决、收束到稳定和弦
convincing /kənˈvɪnsɪŋ/ adj. 57:48
有说服力的、逼真的
collaboration /kəˌlæbəˈreɪʃən/ n. 59:52
协作
unquestionably /ʌnˈkwestʃənəbli/ adv. 60:54
无可置疑地
stuck in our ways idiom 60:54
固守成规、困在老习惯里
理解自测 · 11 题
1. AlphaGo 对李世石的五番棋最终比分是多少?李世石赛前如何预测?

最终 AlphaGo 以 4:1 获胜。李世石赛前完全不把算法放在眼里,声称要 5:0 横扫。实际过程中他连输三局、提前失去比赛,第四局赢下一盘(他自认为是平生最伟大的一局),第五局再度落败。讲者在第二章提到自己在 YouTube 上着魔般地追看这几局,因为意识到「我的饭碗也危险了」。

2. 讲者用哪个雅达利游戏说明「自下而上」的学习?机器学到了什么?

《打砖块》(Breakout)。机器只得到屏幕像素和分数,没人告诉它要接球;它随机移动挡板,发现碰到球分数会涨,于是调整参数优先朝球移动。约 600 局后,它自行发现了讲者小时候和朋友引以为豪的「左侧挖隧道让球在上方来回反弹」的技巧。讲者以此说明新代码能通过失败反馈自我重参数化,而非执行人写的规则。

3. Margaret Boden 的创造力工作定义包含哪三个要素?讲者认为哪些是主观的?

三要素是:新(new)、意外(surprising)、有价值(valuable)。讲者指出「新」可以客观判定,计算机很容易造出新东西;而「意外」和「价值」则相当主观,AI 若要被人类认可为有创造力,必须学会人类认为什么是意外、什么有价值。他在第五章还指出,游戏是检验这三点的理想场景:解说员的惊呼证明了意外,赢下比赛证明了价值。

4. 讲者为什么把 AlphaGo 的第 37 手称作「创造性行为」,而不只是「算得深」?

因为这一手同时满足 Boden 的三要素。它「新」:打破了人类开局只下一至四线的传统;它「意外」:所有解说员都以为是大错棋,李世石也难以置信;它「有价值」:随着棋局发展,正是这颗黑子赢得了右下角地盘并赢下第二局。讲者进一步论证(第 15 段),人类围棋理论只是「局部最大值」,AlphaGo 通过冒险走下山谷找到了更高的峰,因此它不只是复制人类而是教会了人类新下法。

5. 「Lovelace 测试」与图灵测试有什么不同?它为什么针对现代机器学习特别贴切?

图灵测试考察机器能否在对话中让人误以为是人,核心是自然语言与实时应答。Lovelace 测试针对创造性,要求三点:作品原创、过程可重复(不是硬件故障或纯随机)、而且编写代码的程序员无法解释算法如何得出结果。第三条正对应第 18 段讲者的核心观察:代码在与数据互动中变异,已与原编写者「拉开距离」,这与 Lovelace 本人「机器只能执行命令」的告诫形成直接对照。

6. 讲者如何回应「再造一幅伦勃朗有什么意义」的质疑?

他承认评论家 Jonathan Jones 的批评有一定道理,但认为 AI 的价值在于「识别出人类在数据中忽略的东西」。他举了两个例子:对波洛克滴画的算法分析揭示其分形结构,可以量化分形维度;Netflix 算法仅凭用户好恶就把电影聚成人类尚未命名的新类别。这是「回望过去」的意义;他随后指出更激动人心的是「向前看」——让 AI 打破常规做新东西,由此引入 GAN 画作实验。

7. 讲者为什么说 AI 画作并非「与人类情感世界脱节」?这一论证对「被骗」的感受意味着什么?

他在第 31 段指出,AI 是在人类的情感世界(过去 1500 年的艺术数据)上学习,产出是人类情感经过「新滤镜」的呈现,因此并非没有情感基础。他用笑话做类比:若得知笑话是 AI 编的,先前的笑声并不因此作废。推理链是:作品引发的情感反应是真实的,而其来源本身就承载着人类情感,因此「作者是机器」不足以取消作品的价值——这直接回应了巴塞尔观众得知真相后的不适感。

8. 从三首诗的实验结果,讲者推出了什么关于读者判断的结论?

结果出人意料:观众认定是 AI 的第一首出自霍普金斯,像代码的第二首出自诗人 Mez Breeze,而唯一读得通顺的第三首恰恰是库兹韦尔「控制论诗人」的作品。讲者由此说明两点:一是诗歌是封闭形式,不需要长程结构,AI 容易胜任;二是读者在读诗时投入了大量自己的创造力去填补意义,所以「读得通」并不能证明作者是人。这与随后哈利·波特续写「局部通顺、整体跑偏」的例子形成对照。

9. 演奏者 Kobe 为什么能清楚感觉到哪些段落是 AI 写的?这揭示了 AI 的什么局限?

因为 AI 没有身体(not embodied),不知道手指的物理约束,会写出极其别扭的指法;而巴赫写的东西天然贴合手指运动。讲者在第 47 段还指出另一个局限:Clara 模型只做前向预测、没有长期记忆,无法利用自己之前写过的内容。这两点解释了为何 AI 在补最后一个和弦时漏掉了关键解决音——它缺乏对整体结构和身体经验的把握。

10. 如果有人反驳「AlphaGo 只是在海量搜索后碰巧下出第 37 手,谈不上创造」,讲者会如何回应?

讲者可能承认「碰巧」的表面描述,但会指出 Lovelace 测试已排除纯随机:AlphaGo 的过程可重复,它自估该手人类下出的概率仅万分之一却仍选择它,说明是基于学到的价值判断而非偶然。他还会援引第 15 段的论证:人类几百年的围棋理论只是局部最大值,AlphaGo 通过自我对弈探索了人类不敢冒的风险,找到了全局更优解并改变了人类的下法。若「碰巧」能系统性地教会人类新策略,那么区分「碰巧」与「创造」的界限本身就值得怀疑。

11. 讲者「AI 是把人推出思维定式的工具」这一结论,放到数学研究的情境中还成立吗?

讲者本人认为成立,并明确以自己为例:他在数学研究中会一遍遍尝试同样的思路,需要外力把他推出定式。他与 Hassabis 的对话表明 DeepMind 已在研究「有创造力的 AI 数学家」。但迁移到数学有额外条件:数学证明的「价值」比围棋胜负更难客观判定,而讲者强调 AI 必须先学会人类的价值标准。因此该结论在数学中成立的前提,是能像围棋那样给出可快速验证的价值信号(例如证明是否成立);一旦有了这个信号,Lubat 式的「它领先我多年、但仍然是我」的协作模式便可复现。

精读便签
下载便签 手机:长按图片也可保存
← 上一期 · NO.122Stuart Russell: Provably Beneficial Artificial Intelligence
订阅苏菲周报 每周一封:本周入库的精读、一个值得带走的问题、一条苏菲按。免费,随时退订。
免费 · 每周一封 · 一键退订
苏菲拉底 THE SOPHIE LAB · ASK THE BIG QUESTIONS · THINK DEEPLY · SEE THE WORLD DIFFERENTLY 内容仅供学习 · thesophielab.com