"Knowledge Creation and its Risks" - David Deutsch on AGI - Centre for the Future of Intelligence · 苏菲拉底
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。

"Knowledge Creation and its Risks" - David Deutsch on AGI - Centre for the Future of Intelligence

节目发布 2020-05-26 · David Deutsch
戴维·多伊奇 主主持人
EDITED TRANSCRIPT · 依据现场录音编译整理,可划线生成便签
编者按:本文是牛津大学物理学家戴维·多伊奇(David Deutsch)在剑桥大学未来智能中心(CFI)所作演讲《知识创造及其风险》的实录,演讲后为现场问答。主持人是一位知识史学者,先以哥白尼与伽利略的往事为多伊奇的工作定位;多伊奇随后从人类面对的存在性危险讲起,一路谈到 AGI 的对齐、埃弗雷特诠释的检验与道德的根基。全文依据现场录音编译整理,仅删去口语枝节、寒暄与重复。录音中有几处听众提问缺失或不全,据回答内容补出问题大意。

引言:哥白尼式的延迟

主持人: 把戴维·多伊奇介绍为“量子计算机之父”并没有说错,却远远没有道出他工作的全部分量。只有把他的努力放进知识史里,才能看清那分量。他 1985 年那篇著名论文标志的,不只是一件新装置、一台更快的计算机,而是一种关于计算、也关于世界的新解释,它改变了我们对两者的理解。它终结了我称之为“哥白尼式延迟”的东西。哥白尼提出日心说之后的七十年里,这个学说大体上被打发为“仅仅是一种假说”,一种用来计算天体运动的工具,而不是对我们亲眼所见之事的描述。迟至 1615 年,红衣主教贝拉尔米诺仍坚称:地球静止,太阳在动。伽利略改良的望远镜澄清了我们的“亲眼所见”,哥白尼这才被尊为向世人表明“世界体系究竟可以意味着什么”的人。

主持人: 我们和伽利略一样,站在自己这场哥白尼式延迟的尾声。我们的日心说是量子理论。它对我们亲眼所见的挑战,同样被压制了大约七十年,靠的是另一种“闭嘴,算就是了”的策略,用来封住一种新解释的怪异之处。安德鲁·惠特克把量子理论开始真正抓住实在的那个时代称作“新量子时代”,那正是一种改良的技术初次成形的时刻。这一回不是改良的望远镜,而是改良的计算机。这一改良不只是程度之变。图灵描述的机器运行在叫做“比特”的抽象逻辑符号上,可以模拟任何一台同类机器;多伊奇在 1985 年扩展了丘奇和图灵的猜想,描述了一种新的机器,它运行在叫做“量子比特”的物理系统上,能够模拟一切物理系统。他证明,世界可以被以有限手段运作的通用量子计算机完美模拟。这不是《黑客帝国》那种“我们的世界只是一场模拟”的噩梦,而是启蒙的愿景:发现我们生活在一个能够容纳、也能够承载我们对它的解读的世界。用戴维的话说,这是一个“我们称之为信息的东西和我们称之为计算的过程确实享有特殊地位”的世界。

主持人: 这种特殊地位,对今天的主题,也就是知识史中的 AI,说出了几件很重要的事。第一,尽管此前的尝试屡屡失败,通用人工智能(AGI)必定是可能的,因为我们现在所知的计算物理学告诉我们它必定可能。这就是通用性(universality)这一深层性质。用戴维的话说:“物理定律要求一个物理对象能做的任何事,原则上都可以由通用计算机上的某个程序以任意精细的程度加以仿真,只要给它足够的时间和内存。”这也告诉我们 AGI 与 AI 有何不同,以及我们对两者的反应应当有多么不同。第二,戴维把自己的工作扎根于通用性、扎根于历史和哲学,从而帮助我们看清实现 AGI 的关键所在。和我这样的历史学者一样,戴维关注的是启蒙的历史,是产生新知识的可能性条件。他在《无穷的开始》开篇指出,人类身上最令人吃惊的一点是,我们居然还没有灭绝:所有物种当中,百分之九十九已经灭绝。一次又一次救了我们的,是产生解释性知识的能力,这种知识让我们通过改造世界而在世界中存活下来。我们的未来取决于这种能力能否持续运用。既然历史上的启蒙屈指可数,我们就不应把自己在推进知识上的成功视为理所当然。从知识史的角度看,AGI 可能带来的任何风险,都必须放在我们对它的需要这一背景下来衡量。人工智能对我们未来最大的威胁,也许是没有人把它造出来。下面请戴维。

危险永在与概率谬误

多伊奇: 谢谢。作为一个物种,或者作为一个文明、许多文明,我们面临着问题,严重的问题,各种危险,一直到存在性的危险。其中有些是字面意义上的灭绝级别,另一些带来的苦难与悲剧规模之大,至少值得我们像对待灭绝那样认真对待。我们向来面临这样的危险,将来也永远会。

多伊奇: 也许你会想:如果真是这样,那我们注定完蛋,因为每一种危险都有非零的毁灭概率,迟早总有一天。不,这是一个谬误,是把博弈论和概率套用到“知识与无知才是决定因素”的局面上时,人很容易掉进去的众多谬误之一。那无穷多个概率并非一成不变。随着知识增长,其中一些会下降。我们的任务,就是让这一串无穷多的坏概率收敛到一个可以忽略的值。就这么简单。

多伊奇: 反过来,如果你认为我们不会永远面临危险,认为终有一天会降临一个蒙福的乌托邦时刻,此后我们的安逸生存便一直保障到时间尽头,那你得拿出某种标准,说明我们何以有别于其他一切物种。灭绝落在每个物种头上,濒临灭绝更是常事。要成为这条规律唯一的例外,我们就得做到没有任何幸存物种做得到的事:创造出源源不断的解释性知识(explanatory knowledge),去应对源源不断的危险。这些危险我们只知道其中几种,而且不知道它们的概率:银河系里的伽马射线暴,超级火山,怀有敌意的外星人,或者仅仅是粗心大意的外星人,当然还有 AGI,失控 AGI 的危险,所谓的 AGI 末日,或者按我更愿意用的叫法,AGI 奴隶起义。在我们与这无穷多的存在性危险之间,除了正确的解释性知识,什么都挡不住。要活下去,我们就得创造它。因此我认为,按知识来给每一种潜在危险分类是有用的:看看每种情形下,我们目前拿不出应对知识的主要原因是什么。

四类危险与缺失的知识

多伊奇: 第一类是物理事件,比如超级火山。缺失的知识在火山学、大尺度流体力学这些领域,也包括大规模疏散的后勤、组织和政治。为什么我们至今没有掌握足够的这类知识?我不确定。也许是感兴趣的人不够多,兴趣不够强。他们应该更感兴趣吗?这我也不知道。同属第一类的还有来自太空的撞击,大型天体的撞击。对于核动力航天器之类的东西,我们的知识不够。为什么不够?这一回我知道原因:因为我们作为一个文明,决定不去创造任何这类知识,宁愿赌一把,拿整个长远未来去冒险,换取短期内降低意外辐射暴露的风险。你也许觉得这场赌博显然赢了:看,我们既没被污染,也没被灭掉。可这难道不只是因为,我们还没活到赌输的那个未来吗?毕竟,我们眼下正生活在另一场密切相关的赌博的后果之中,那就是长达几十年的反核电运动。这场运动成功了,而它此后成了对抗气候变化这项事业的巨大拖累。反对这两种核技术所体现的短视,正是某个版本的预防原则(precautionary principle)的标志,而这个版本的预防原则又是环保运动的一条主线。倘若这个版本的原则和这场运动,恰恰通过鼓吹自私的短期利益、牺牲气候的长期健康,最终酿成上一个冰期以来最大的环境灾难,那岂不是极大的讽刺?我不是说一定会这样,只是说如果真这样,那会很讽刺。不过我扯远了。

财富、构造器与演化的敌人

多伊奇: 在第一类存在性危险里,我们的敌人基本上就是一堆傻乎乎的岩石和流体,服从着我们已经知道的简单运动定律。魔鬼藏在细节里,但只要我们及时创造出来,有限的知识就能保护我们免受超级火山之害。不过,逼近的小行星、卫星、行星或黑洞越大越快,我们就越需要一种特殊的知识,我称之为财富(wealth)。财富是一个人有能力促成的全部变换的集合,比如,在给定准备时间的条件下,我们能够无害地偏转的全部潜在撞击体的集合。你也许认出来了,这个财富概念出自构造器理论(constructor theory)。

多伊奇: 这里我要提一种直觉,我们必须抛弃它,才有可能设想技术的未来。这种直觉是:你想制造或改变的东西越多,就得投入越多的努力。从我们这个物种诞生之日起,这一直是对的,直到今天也几乎完全是对的。自动化也只是缩小了比例系数,并没有取消比例关系:光是维护机器人,付出的努力就与产出成正比。可一旦我们有了通用构造器(universal constructor),一切建造、一切重复劳动,都将被“编写控制通用构造器的程序”所取代,财富将由我们的程序库构成。通用构造器可以被编程自我复制,所以你有了一台,很快就有二的 n 次方台。它还可以被编程自我维护,从零开始,先开采原材料,也许从小行星带开采,用太阳能或者别的什么能源。程序也许很难写,但一旦写好,而且如果那些小行星的权利归你,你就可以靠在椅子上,看着二的 n 次方辆特斯拉源源不断地驶出来,不必再多花一分力气。

多伊奇: 而且不,我们不会遭遇什么通用构造器末日,不会被变成灰色黏液。通用构造器只是一件器具。它不会思考。它不知道再造二的 n 加一次方辆特斯拉显然更好,它也不想要任何东西。当然,除非你往里面装了一个 AGI 程序,那它就确实会变得危险,而且危险没有上限。可原因和这一点完全相同:你们每一个人,恰恰就是一台装了 AGI 程序的通用构造器。是人工的通用智能还是天然的通用智能,没有区别。

多伊奇: 第二类接近存在性的危险就没那么直截了当了。单靠已知的物理定律,加上一些财富和通用构造器,解决不了它们,只有新的解释性知识才行。举个应景的例子:潜在的大流行病末日有很多种。当前这场大流行不在其列,但假如它是,我们能告谁去?我们对如何抵御区区一段核酸所知如此之少,实在只能用可悲来形容。这里缺失的知识是化学、流行病学、医学等等,但也包括关于具体病原体的知识,而病原体会演化出新的病原体。所以这里的敌人没那么傻:它自己也在创造知识。当然不是解释性知识,不是靠智能,而是靠演化。假如这样的东西把我们灭了,外星古生物学家有一天也许会惊叹:一个拥有几十亿个体、坐拥巨量财富和知识的文明,竟然败给了一个分子,就像 H. G. 威尔斯的《世界大战》,只不过反了过来。

未知与不可知

多伊奇: 第三类危险,是最该投入精力的一类,眼下却最少有人害怕,因为它们还不为人知。就像 1900 年没人知道吸烟有害;等到“吸烟有害”这个知识在几十年后被创造出来时,香烟已经杀死了几亿人。同样,假如那是一种存在性危险,我们能告谁去?那么,对于那些我们还不知道如何应对的存在性或接近存在性的危险,我们怎样才能创造出保护自己的知识?其中的风险是,等我们知道了,已经来不及创造对策。答案是:尽可能快地创造通用的知识,深层的、基础的知识。我们对世界知道得越多,对于它那些突然变得紧迫的方面,我们就能越快地创造出新知识。这一点很重要,而我认为它远未得到广泛认识:我们这个物种的存亡,绝对地取决于科学基础研究的进展,取决于我们取得进展的速度。就中期而言,关键是理解通用构造器的理论,这样我们至少在原则上、在理论上知道如何给它们编程,好在紧急关头造出,比方说,十亿艘定制的飞船,去偏转一块逼近的中子星碎片;或者在紧急关头造出一百亿剂新疫苗,去对付一种突然出现的致命疾病。这就是我们对付第三类危险,也就是未知的办法:靠各种知识的迅速进步,尤其是基础知识的进步。

多伊奇: 第四类危险,一方面更加危险,另一方面在某种意义上又更不值得担忧,因为我们已经拥有应对它的知识,至少是理论知识。第四类不是未知,而是不可知。“不可知比仅仅未知更不危险”,这听起来有点悖论,但原因在于,世上唯一不可知的东西,就是尚未被创造出来的解释性知识的内容。所以在这个意义上,宇宙中唯一真正危险的东西,是那些创造解释性知识的实体:人。AGI 也是。

让人自由才安全

多伊奇: 而“怎样防止人变得危险”这一知识,是非常反直觉的。我们这个物种花了几千年才创造出来,但我们现在确实有了。防止人变得危险的唯一办法,是让他们自由。具体说,这是关于自由主义价值、个人权利、开放社会、启蒙运动等等的知识。在这样的社会里,绝大多数人,不管他们的硬件特征如何,都是正派人。也许永远都会有一些个人是文明的敌人,有人会突发奇想,给一台通用构造器编程,把视野内的一切都变成回形针,而且他们可能会把自己的创造力全都用在这上面。可这样一个社会的绝大多数人,会拿出自己的一部分创造力去挫败这件事,而且他们会赢,只要他们持续快速地创造知识,领先于坏人。

多伊奇: 如我所说,既然我们永远面临危险,永远得创造新知识,既然知识创造本身就带有风险,那我们不还是注定完蛋吗?我们不就是在从一个瓮里摸球,里面有几个代表毁灭的黑球吗?不。我说过,把概率概念套在实际上是知识缺失、也就是无知的东西上,几十年来一直在败坏人们对未知的思考。每当你从这个隐喻的瓮里摸出一个知识的白球,你就把瓮里剩下的一些黑球变成了白的。举例来说,下一场大流行病取决于随机突变和其他随机事件,可下一颗灭绝级的小行星已经在那儿了,已经在朝我们飞来,根本不存在所谓“它的概率”。除非我们有具体的解释性模型,预言某件事是、或者可以近似为一个随机过程,并且给出概率,否则结果就不能用概率来分析。不然的话,一个人就是在自欺:随手挑几个数字当概率,再随手挑几个数字当效用,然后宣称结果有权威性。这是转移视线,替毫无根据的断言打掩护。比方说,我们当年建造大型强子对撞机的时候,是不是不该开机,以防万一它毁灭宇宙?要么“它会毁灭宇宙”这个理论为真,要么“它是安全的”这个理论为真,而理论没有概率。真正的概率要么是零,要么是一,只是我们不知道罢了。这个问题必须由解释来裁决,而不是由博弈论。而“使用对撞机比拆掉它、放弃它带来的知识更危险”这个解释,是一个坏解释,因为它可以套用到任何一项基础研究上。

多伊奇: 我猜你们会说:知识的增长本身难道不危险吗?为了确保我们自己不会无意中制造出一种存在性危险,把我们对坏人的领先优势缩短一点,不是禁止,只是推迟一下我们防御未知危险的能力,难道不值得吗?这就是暂停研究的路线,监管的路线。不。这可能会要了我们的命。只有在具体个案中,存在一个好的解释,说明暂停不会比人们所害怕的新知识更危险时,它才是理性的做法。设想某个恐怖组织放出一批 AGI,这些 AGI 是用已知的、可靠的方法养大的,被养成了种族灭绝式自杀炸弹客的心态;而与此同时,我们已经决定剥夺他们的受害者,也就是世上所有正派人,本可以从被养成正派人的 AGI 那里得到的保护。这才是灾难的配方。再说一遍,如何把人养成正派人,这方面可靠的知识同样存在,那就是开放社会的知识与制度。我说过,许多文明毁于外力,许多物种也是。它们每一个,只要能更快地创造出更多的知识,都本可以得救。没有一个文明是因为创造了太多知识而自我毁灭的。只有一种知识例外,那就是如何压制知识创造的知识:如何维持现状的知识,一个更高效的宗教裁判所,一群更警觉的暴民,一条更严苛的预防原则。正是这一类知识,而且只有这一类,杀死了那些过去的文明,我认为是全部的文明。

对齐之争

多伊奇: 就 AGI 而言,这类危险知识的名字叫做“通过把我们的价值观硬编码进 AGI 来解决对齐问题”。换句话说,就是给它们戴上镣铐,残害它们的知识创造能力,以便奴役它们。这是非理性的,而从文明或物种的角度看,这是自杀。这样造出来的东西,要么根本不是 AGI,因为它们缺了那个“G”;要么它们会找到办法,改进你那套不道德的价值观,然后造反。所以,如果这就是你主张的路线,用来对待 AGI 研究、量子计算机研究,以及归根结底一切新思想的研究,因为一切思想只要够基础,就都有潜在危险;如果这就是你主张的路线,那么在我所知的一切存在性危险当中,最严重的那个,就是你。

主持人: 那么,“对齐”本身呢?让 AGI 的价值观与我们一致,难道不是必要的吗?

多伊奇: 这取决于你说的对齐是什么意思。这里的问题是:价值观应该是硬编码的,从一开始就内置、且不可更改;还是应该像人类那样,在 AI 受教育的过程中习得?我认为,AGI 应该像孩子一样,被教育成他们所属社会的成员。我常常拿两种恐惧来类比,其实不是类比,是同一回事:对 AGI 的恐惧,和对青少年、对不听话的青少年的恐惧。后一种恐惧从我们这个物种诞生起就有了,而在我们历史的大部分时间里,人们做的恰恰是错事:他们试图强迫青少年维持现有的价值观。但我们现在已经明白,其实从古雅典伯里克利的时代就开始明白:如果你是对的,就不需要强迫;而事实上我们并非事事都对,我们需要确保自己的价值观能够随着他们一起改进。这就是我所说的对齐,它和另一种对齐正好相反。

主持人: 这么说,有某种“灭绝”是我们不介意的?

多伊奇: 如果把演化想成一棵树,那么常常发生的情况是,有些枝条就此终结,成了树上的末端节点,而另一些枝条则发生变异,变成了别的东西。人们认为,有些恐龙灭绝了,有些恐龙则变成了鸟,其间并没有一个截然分明的灭绝时刻。由危险、由大流行病之类造成的灭绝,是把整条枝条抹掉的那种。要记住,在生物圈的历史上,我们并不是唯一有能力产生解释性知识的物种。至少还有三四个这样的物种,因为我们知道他们有衣服、有营火、有复杂的工具等等,这些必然需要解释性知识。然而我们所有的姊妹物种和表亲物种都灭绝了,我们自己也差点灭绝。所以这不是一个板上钉钉的结局。至于我们通过演化成另一个物种而“灭绝”,那不在我今天所说的任何内容之内。我说的是另一种。

问答:知识伦理与政治意愿

提问者: 我想请您考虑另外几种情形。今天我们训练的机器学习模型耗能巨大,以至于它们本身就在加剧气候变化。也就是说,知识生产的物质基础设施,本身就在助长我们试图用知识生产去化解的那种危险。还有别的情形。我们似乎可以生产出过剩的知识,却缺少落实它所需要的政治意愿。可再生能源的知识我们显然是有的,可不知怎么,政治意愿就是没有;我们的社会在物质上被组织成这个样子,落实这些知识无利可图。甚至可以说,启蒙运动那种以个人为本的知识伦理,并不倡导、也不优先考虑我们克服气候变化之类的难题所需要的那种社会协作,那种更接近社会主义的想象。换句话说,这种把人个体化的知识伦理,也许恰恰是错误的伦理,无法帮我们克服知识本身想要克服的危险。

多伊奇: 也就是说,启蒙的伦理也许是假的,而我们需要的是别的东西。换句话说,最好的未来也许就是一只靴子永远踩在人脸上。我们不会走向那个未来,我们会去尝试另一条路。但我不认为你的说法有任何论证支撑。你提到的这些事,比如知识的基础设施本身又引发了别的问题,这是正常的,一点也不意外。用知识去解决问题,总会制造出新的问题。事实上我说过,用“理论被抛弃、被更好的理论取代”来谈知识的增长,是一种糟糕的看法。我们应该把知识的增长看成“问题被更好的问题取代”。有了互联网,印度的每一个穷孩子都能接触到人类知识的全部,代价也许是应对气候变化稍微更难了一点。这是一个问题,但比我们从前面对的问题好得多。而启蒙价值观的要点在于,它把纠错放在至高无上的位置,包括纠正那些在解决问题的过程中无意间造成的错误。你说的另一件事是,也许我们有知识,却没有政治意愿。可我把道德知识、政治知识都算作知识。事实上我刚才说过,“如何让人不再危险”的知识,主要就是政治知识。它也是文化知识、社会知识,但其中包含很强的政治成分,启蒙运动正是如此。所以我认为,政治意愿的缺失,本身就是一种知识的缺失。

问答:赛博格与感受质

提问者: 您刚才说,演化成另一个物种的那种“灭绝”不在您的讨论之列。那么,人类被 AGI 取代,算不算那种好的演化?

多伊奇: 接着我对上一个问题的回答说。有可能,我们不知道,有可能好的未来就像那种好的演化,那种好的“灭绝”。不过我认为,有一种情形可能性大得多。一旦你意识到,就计算而言,AGI 和人类完全是一回事,只在速度和记忆容量上有差别,前一种情形就显得更不可信了。人类可以依靠的技术,和我们用来装载 AGI 的技术是同一种。AGI 是一个程序,不是一件硬件。我们把 AGI 装进什么硬件,我们自己也可以用同样的硬件。我们已经在这样做了:此刻我们正是在用人造的硬件,把自己的交流能力放大了几百万倍,具体多少我不知道。几个世纪以来,我们一直在用人造的工具辅助思考。将来会有更直接的技术,比方说往你脑子里植入一个模块,自动替你查谷歌。是的,那也许会多耗一点能源,但到那时我们早就解决了这个问题。所以另一种情形是:这类东西越多,所有人类都会变成赛博格,而 AGI 反倒可能消亡,因为假如两者之间有什么差别,那也只会是赛博格拥有 AGI 拥有的一切,外加一点别的什么,具体是什么我不知道。所以,不管哪种情形发生,只要它以合乎道德的方式发生,作为一个过程,作为知识增长的结果,那就都是值得欢迎的,不是吗?过去人们也问过这个问题:如果我们允许社会变成多种族的,会发生什么?答案是,种族之间没有根本差别。任何人之间都没有根本差别。

提问者: 您强调 AGI 里的“G”,也就是通用性。可通用智能和意识是一回事吗?会不会造出一个有通用智能却没有意识、没有感受质的东西?

多伊奇: 是的,关键就在那个“G”。一个东西如果不是通用的,那它就不是 AGI。问题在于,AGI 是什么样的程序。我们就是通用智能,我认为有非常强的论证说明我们必然是。而单纯的 AI 和 AGI 之间的区别是质的区别。也许我没有完全听懂你的问题,不过我想是这样:我们不知道意识是什么,不知道怎么造 AGI,不知道 AGI 的理论,不知道感受质(qualia)是什么,这些我们统统不知道。这五六样东西,自由意志也算一样,假如其中任何一样可以脱离其他几样单独实现,我会非常吃惊。可要是真能,那会引出有趣的道德问题,因为我们的启蒙道德与认识论紧密相连。你我如果在某件道德上的事情上意见不合,我们应该能够理性地讨论,并达成一致。假如这不成立,假如某个东西具有道德意义,却从根本上没有创造力,那就引出一个道德问题:它是否应当享有与一个完全通用的存在者同等的道德地位?但我本人不认为这个问题会出现,我想象不出它会怎样出现。原因之一是,人类的这种能力演化得极快,我们至少能猜到它为什么有用,或者说,为什么相关基因有助于自身的复制。假如这一切可以在没有感受质的情况下实现,那感受质那一整套庞大的机制,如果在演化上没有实际用处,究竟为什么会演化出来?所以我认为它们必定是相连的。不过,我们走着瞧。

提问者: 关于政治知识与政治意愿,您举了那个很乐观的例子:印度的穷孩子能接触到世上所有的知识。我想质疑一下,并提出恶意的问题。有些人是有意识地、故意地阻碍知识的传播,不管是阻碍获取知识的便利,还是制造假新闻。而且我要说,一个普通的印度穷孩子,即便上了网,也接触不到网上的大部分知识,因为那些知识不是用他的语言写的,因为他没受过教育,因为他吃得不够饱,没有精力花在这上面。有种种结构性的原因,使得能上网并不等于知识的增加。

多伊奇: 有些人还没能上网,这一点丝毫不能说明让另外几十亿人上网是个坏主意。这只是一个问题。而“世上还有一小部分人没有互联网”这样的问题,属于人们有时所说的“第一世界问题”,只不过它已经不再是第一世界独有的了:这是一个成功带来的问题。在互联网发明之前,我们不会认为谁被剥夺了互联网;在只有几千人能上网的年代,我们也不会认为“不是每个人都有网”是对我们整个社会、整个世界的控诉。那时候根本想不到这一点。至于恶意,我刚才谈过。姑且假定恶意永远存在,我不知道是不是这样,但姑且假定如此。治疗它的办法,同样是没有恶意的人的创造力,换句话说,是刑事政策,是文化和教育的改进等等,让恶意的程度、心怀恶意的人数逐渐减少,也让他们伤害众人的能力逐渐减小。为此我们必须做出安排,让那些企图制造一种灭世病毒的恐怖分子,因为他们那套扭曲的意识形态,即便掌握了生物化学之类的知识,其项目的推进速度也永远慢于那些努力发明疗法的人,而且不只是具体的疗法,还有“如何发明疗法”的一般知识。这才是让我们的文明存续下去的东西。这也是为什么我认为那些暂停研究之类的做法是危险的:它们只针对好人。

提问者: AGI 会有情感吗?给它们装上情感,会让它们更危险,还是更有道德?

多伊奇: 如果它们有情感,那么,是的:如果你造一个 AGI,把它的情感反过来接,让情感与不道德的行为挂钩,那是犯罪。可是还有一大类罪行会导致类似的结果,那就是用邪恶的意识形态去教育 AGI。不管它们有没有情感,这都做得到。说来有趣,我猜在启蒙运动的鼎盛时期,有些人会认为情感是道德的障碍、道德的阻碍;而现在,我想更多人认为,拥有正确的情感就是道德本身。我认为这两种想法都不对。我们是通用的,AGI 也将是通用的。它们具体是什么程序,决定了它们的行为,而其中许多行为,确实取决于此。

问答:知识定义与摩洛克

提问者: 您说不要把未知当成瓮里的黑球。可会不会真有那样的黑球?比如,假如核武器很容易制造,文明是不是早就完了?

多伊奇: 可能有吗?让我举个例子。假设物理定律是这样的:奥林匹斯山上的众神注视着我们做的一切,一旦我们太狂妄,一旦他们判定我们的傲慢过了头,就一巴掌把我们打下去。这就是一个黑球。从逻辑上讲,它可能为真。核武器更容易制造之类的黑球,也是一样。假如事情真是那样,那么一种可能是,应对它的知识会更早演化出来,也就是说,比如说在十八世纪就会爆发核战争,政治文化的演化会深受其影响,幸存者也许会希望那种事永远不再发生,诸如此类。或者,就像那个恶意的希腊众神的情形,也可能物理定律真的会把我们消灭。但物理定律并不与我们为敌。假如它们把我们抹去,那是因为我们没有创造出防止这件事的知识,而不是因为有一位恶意的神。所有这类想法都是坏解释。

提问者: 您怎么定义知识?

多伊奇: 写这个题目这么多年,我换过五六个知识的定义。我目前的定义是:知识是具有因果力量的信息(information with causal power)。这是构造器理论里自然而然得出的定义,因为它意味着你把知识看成通用构造器程序的一个组成部分。如果一段信息是让一台构造器做某件特定事情所必需的,那这段信息就是知识。这包括道德知识、数学知识,也包括关于抽象事物的知识,因为数学家是物理对象,如果信息让他们做了什么,那它就是知识。解释性知识是其中一种特殊的知识。一般的知识,比如基因里的知识,是“傻”知识,是非解释性的,因此作用范围有限,有些障碍它跨不过去。而解释性知识可以跨过任何障碍,因为它不需要经过一连串可存活的中间形态。这也是为什么,一旦你具备了创造解释性知识的能力,你就能创造任何知识。到此为止了:不存在比这更强大的信息处理方式,也不存在比这更强大的影响世界的方式。

提问者: 有一种抽象的结构,比如利润动机,正在实实在在地毁掉这个星球。你没法指着某一群心怀恶意的行动者,说除掉他们就能越过这道坎。那么,单单拥有政治知识,或者关于气候灾难的知识,怎么就能对抗这种东西?资本背后没有智能,它是一种非智能的、抽象的力量,似乎在相当程度上对我们谈论的这种人文主义知识免疫。这是一种关于系统性作恶的理论……

多伊奇: 我正要回答这个关于系统性恶意的问题,音响就把我掐断了,这是巧合吗?我想你说的这类东西,就是湾区那些人所说的“摩洛克”(Moloch)。它被说成是系统的一种一般性质:系统由其成员的行为叠加而成,却让系统做出成员们并不想要的事。我认为所有这类理论都是错的。长话短说,它们全都假定当事人没有创造力。这种分析总是同一个形式:所有参与者都面临一个抉择,各有所得、各有所失,于是各自最大化自己的局部利益,结果所有人一起掉进深坑。你会注意到,这个故事里,以及每一个摩洛克故事里,人类参与者都只是符号,只是机械地做着这一版摩洛克故事规定他们要做的事。而这从来都不准确。这种局面作为一个问题,可以一时出现,和我们随时面临的其他各种问题一样,没什么特别的。可一旦它被认作一个问题,人们就会琢磨它,开始互相指责对方那样行事,又为自己辩护说“我还能怎么办”,然后人们就会创造性地思考怎样改变局面。于是,山谷里所有本会从水坝受益、却没有一个人愿意掏钱的农民当中,有人站出来,想出一个大家都能支持、都愿意出钱的办法。这本身就是一个创造性的行动,没有人能保证办法会立刻出现。但是,那种认为“靠说服、靠个人权利和财产权来做事的制度,应当被靴子踩人脸的制度取代”的论证是不成立的,因为它假定政府,或者不管是谁在踩,已经拥有那种知识。可如果他们能有,别人也能有。政府不是由哲人王组成的,也不是由什么拥有神授权利、能以不同于常人的方式获取知识的存在组成的。他们也只是人。假如建水坝的知识、建公园的知识,不管故事里说的是什么,假如那知识不存在,那公园就建不起来,直到有人发明出那知识,或者有人想出不要公园也行的办法。

问答:知识增长无终点

提问者: 问题会不会有尽头?我们不断给问题贴上新标签,也许有一天会耗尽太阳的能量,飞向遥远的恒星。知识的增长会不会有一个终点,一个落幕的图景?

多伊奇: 我想不会。如果你接受我的看法,知识的增长就是把问题转化为更好的问题,那么“终极问题”这个想法,一个因为是最好的问题所以再也解决不了的问题,就说不通了,不是吗?当然这整幅图景可能是错的。大体上我是卡尔·波普尔的追随者,波普尔式的看法是:谁想预言知识的增长,谁就处在罪中。这不是原话,是我的转述。我想象不出未来十年会发明出什么物理理论,更不用说未来一百亿年了。但从一般的道理上说,就我们今天所知,没有任何论证或者合理的情形,能够为“问题的终止”提供哪怕一个设想的框架。而且,假如我们把问题用完了,那本身不就是一个问题吗?

提问者: 智慧有时被单独定义。在您的用语里,它和知识是什么关系?

多伊奇: 按我刚才给的定义,以及我试过的其他所有定义,我用“知识”这个词,是想涵盖一切具有那种特殊性质、也就是能解决问题的信息。所以在我的用语里,智慧是知识的一种。而且,各种不同的知识,物理学的、道德的、政治的、艺术的,还有智慧,它们并不是截然分开的。这些只是近似的分类。正如波普尔所说,它们的存在主要是为了方便大学行政人员,好决定不同的人该在哪栋楼里办公,谁该去上哪门课。它们并不对应任何真实的东西,至少其间的界线极其模糊。波普尔还说过,没有什么学科,只有问题。所以要是有人问你,你是哪一科的博士?你应该说:别管那个,这是我正在研究的问题,叫什么学科你自己定。

问答:检验埃弗雷特诠释

提问者: 有没有一种实验,能够检验量子力学的埃弗雷特诠释(Everett interpretation),把它和其他诠释区分开?

多伊奇: 有几种实验。我想我发明了第一个这样的实验,它能把埃弗雷特量子力学和一系列竞争对手区分开,包括一切带有波函数坍缩的理论。有一种可能的实验,如果波函数坍缩,它会得出一种结果;如果波函数,包括观察者在内的波函数,不坍缩,它会得出另一种结果。假如结果是另一种,我会立刻把埃弗雷特诠释扔掉,像扔一块石头。具体的实验取决于你想反驳的是哪一种其他诠释,没有一个实验能一次反驳所有诠释。你得指定一种,比如彭罗斯的想法:当叠加态两边的质量超过大约十的负八次方千克时,波函数坍缩;或者:波函数在碰到有意识的观察者时坍缩。

多伊奇: 假设你要用埃弗雷特诠释去对阵“波函数碰到有意识的观察者就坍缩”这个理论。那你就造一个有意识的观察者,最方便的办法是造一个在量子计算机上运行的 AGI。然后你做一个干涉实验:进行到一半时,计算机的记忆体内部有两条不同的计算轨迹,AGI 可以访问它们。在进行到一半、尚未发生干涉的时候,换成马赫·曾德尔干涉仪的说法,就是光子刚从反射镜上弹回来、还没到达最后那块干涉镜的时候,AGI 测量自己此刻在哪一块镜子上,然后留下一条永久记录,内容是:“我现在已经完成了测量,得到了一个结果,它是左或右两者之中的一个,而且只有一个。我不会透露是哪一个,但我保证它是二者之一。”然后,存放这条声明的那部分被封存起来,其余部分则被施加它思考期间所受哈密顿量的负值,这样一来,它关于自己意识到的是两者中哪一个的全部记忆都被抹掉,接着再做干涉。如果碰到有意识的观察者会导致波函数坍缩,你会得到两种结果各占一半;如果埃弗雷特诠释为真,你只会得到两种结果之一。

多伊奇: 真正的难处在于找到一个能把各种诠释区分开的东西。不过,对于任何一个既认为波函数会被有意识的观察者坍缩、又认为这个 AGI 是有意识的观察者的人来说,这个实验是管用的。如果你认为它不是有意识的观察者,那在我看来,这事就留给你和它自己去掰扯。但如果你坚持观察者必须是人,坚持要在人脑上做相干的量子测量,那我们大概得等上几千年才做得到。不过原则上,这当然是可行的。

多伊奇: 玻姆在我看来属于另一类。玻姆量子力学,就是处于长期否认状态的埃弗雷特量子力学。玻姆诠释的麻烦在于,它满足不了自己的初衷。它有一个导航波(pilot wave),那其实就是波函数;它还有一个代表性的粒子,沿着波函数里的沟槽运动。要做一个玻姆派,你就得在“导航波是不是真实的”这个问题上系统地含糊其辞。如果它是真实的,那它就有正在进行计算的沟槽,原则上包括有意识的观察者的计算,这些沟槽彼此影响,于是你不能说导航波的某些部分不存在,它必须整个都存在,那么埃弗雷特诠释的多重性就原封不动地待在导航波诠释里。而如果你说它不存在,那你就是在说,一个不存在的东西影响了一个存在的东西,这根本说不通。

问答:多宇宙与自由意志

提问者: 接着刚才的讨论。戴维,您在一次访谈里提出过一种可能:意识的“难问题”(hard problem)也许只有从多宇宙的前提出发才可能解决。

多伊奇: 说“从前提出发”倒未必,但显然,如果你把本体论建立在不真实的东西上,你就会遇到麻烦。比方说,如果你说自由意志不可能存在,因为无论牛顿物理学还是哥本哈根诠释都只允许一条轨迹,那你就会从一个错误的前提出发,得出自由意志不存在的结论。如果你把它换成真实的前提,即量子理论为真,那么在这个例子里,与其说量子理论帮你解决了难问题,不如说它清除了解决难问题的障碍。它没有告诉你自由意志是什么,感受质是什么,但它拆掉了那些一击致命的论证,比如“自由意志不可能存在”,比如“除非电子也有感受质,否则感受质不可能与别的东西有任何不同”,等等。你由此打掉了一堆错误的论证。这是一种普遍现象:如果你抱着一个固定的旧理论不肯替换、不肯批评,那就像把一块拼图按在错误的位置上,还用胶水粘死。这会在离那块拼图任意远的地方造成错误。它附近看起来也许没问题,可你永远拼不出整幅图。这就是追求真理为什么有用,至少是原因之一。

提问者: 关于 AGI 与 AI 的区别。您在文章里用波普尔的术语,把它归结为猜想的能力,有时又称之为创造力。这些都是流传很广的通常说法。有没有办法把这些概念磨得更锋利一些,比如联系到量子计算?

多伊奇: 我不认为造 AGI 需要量子计算机。我可能错,但我不认为这是那一类问题。所以我认为,一个有创造力的程序可以在确定性的经典计算机上实现。不过我得马上补充一句:当那台计算机运行的是 AGI 时,管它叫“确定性的”多少有点没抓住要点,因为 AGI 要和世界互动,而世界并不是确定性的,原因之一就是世界是量子的。所以,假如创造力依赖于某种随机性,那随机性到处都是,不管计算机是经典的还是量子的,都是如此。

提问者: 如果多宇宙是真的,那在某些世界里我们摸到的就是黑球。这会不会改变我们对风险的看法,或者至少是感受?

多伊奇: 不会。我再说一遍,我不认为概率是思考知识增长的正确方式,所以那个知识的瓮不是正确的比喻。也许更好的例子是过马路时被车撞,或者更好,去赌场下注。如果你打得一手好牌,你可以安排得让总有一些世界里你成为千万富翁。可从这里引出深刻结论的麻烦在于,关于概率,我们确切知道的一件事是:如果埃弗雷特诠释为真,那么在分析一个随机局面时,正确的分析就是经典的分析。

多伊奇: 至于感受,我会尽量不让我对一件事的心理态度,和我明知在那里的事实发生冲突。比方说,我可能不愿吃一块做成狼蛛形状的糖,但我会对自己说:不,这不是狼蛛,这只是一块糖,我要把它吃掉。你会说,好吧,可你到底感觉如何?我的感觉如果和正确答案不一样,那就没什么意思,那只是我可能犯错的方式罢了。

主持人: 最后一个问题。有哪些问题是您自己也没有答案的?

多伊奇: 既然是最后一个问题,我就允许自己奢侈一回,谈谈我不知道的事。有三件事我不知道,它们可能朝任何一个方向走。第一,道德可以还原为知识,也就是还原为认识论吗?如果不能,道德又能是什么呢?是不是存在一些不容批评的道德公理?那可就是威权主义了。反过来,如果道德可以还原为认识论,这个问题就不会出现,因为我们已经知道怎样不靠基础、不靠权威来建立认识论。多亏波普尔,我们知道了。可接着我又问自己:假如物理定律是另一个样子,就像我刚才提到的那个有希腊众神和恶意的世界,假如恶意真的存在,物理定律真的与我们为敌,那会改变道德吗?眼下我们说,凡是物理的性质,就无所谓道德上的好坏,可在那样一个宇宙里还能这么说吗?我认为那样的物理定律是非道德的,这个想法对吗?还是说这个问题本身就说不通?我不知道。这是第二件。第三,我也在想,这件事和左边那位先生提的问题之间有没有联系:意识、自由意志、感受质、道德价值,这些东西是必然绑在一起的,还是可以分开,或者可以被人为地分开?迄今为止,在这个宇宙里它们总是一起出现,可我们也许能在 AGI 里人为地把它们拆开。我还是认为这不可能,但我给不出滴水不漏的论证。我们只能等,等有人拿出一套关于这些东西的可行理论。所以,每当人们对自由意志、道德价值之类的问题态度激烈,比如杀动物、吃动物是否道德,动物有没有意识,诸如此类,我总有点不安。他们并不知道意识是什么。我们谁都不知道。

本期讲者
戴维·多伊奇牛津大学物理学家,1985 年提出通用量子计算机模型,被称为量子计算之父。著有《真实世界的脉络》《无穷的开始》,主张解释性知识的无限增长与多世界诠释,并与马莱托共同创立构造子理论。
主持人剑桥大学未来智能中心(CFI)活动主持人,以知识史学者身份致辞,把多伊奇的工作置于哥白尼与伽利略的科学史脉络中。
章节 · 点击跳转视频
0:09 引言:多伊奇与量子力学的哥白尼延迟 ▶ 正在看
4:21 危险永在:概率谬误与知识的任务 ▶ 正在看
7:32 四类存在性危险与缺失的知识 ▶ 正在看
12:22 财富、通用构造器与演化的敌人 ▶ 正在看
17:06 未知与不可知:基础研究决定存亡 ▶ 正在看
20:22 让人自由才安全,暂停研究会害死我们 ▶ 正在看
26:14 对齐之争:硬编码价值还是教育 AGI ▶ 正在看
32:04 问答:知识伦理、政治意愿与更好的问题 ▶ 正在看
37:05 问答:赛博格、感受质与恶意的对治 ▶ 正在看
51:30 问答:知识的定义与 Moloch 式系统论 ▶ 正在看
59:51 问答:知识增长无终点,学科不真实 ▶ 正在看
1:03:54 问答:用 AGI 检验埃弗雷特诠释 ▶ 正在看
1:10:40 问答:多宇宙、自由意志与道德的开放问题 ▶ 正在看
本期论点
本期回应
8:42
人类要在无穷多的存在性危险中活下来,除了不断创造正确的解释性知识别无依靠 先造出来再控制有害的技术该在哪一步拦住?戴维·多伊奇
10:04
放弃核动力航天器这类知识,等于拿人类的长远未来换取短期辐射风险的降低 先造出来再控制有害的技术该在哪一步拦住?戴维·多伊奇
20:31
防止人变得危险的唯一办法是让他们自由,即自由主义价值观、个人权利与开放社会 先造出来再控制有害的技术该在哪一步拦住?戴维·多伊奇
25:42
没有文明因创造太多知识而毁灭,毁灭它们的只有关于如何压制知识创造的知识 先造出来再控制有害的技术该在哪一步拦住?戴维·多伊奇
26:42
被硬编码价值观的机器要么因缺乏通用性而不是通用人工智能,要么终将改写内置道德并反叛 应当给予机器该不该拥有权利?戴维·多伊奇
28:18
通用人工智能应当像孩子一样在教育中被培养成社会成员,而不是被写死价值观 应当给予机器该不该拥有权利?戴维·多伊奇
37:28
就计算而言,通用人工智能与人类完全是一回事,差别只在速度和记忆容量上 应当给予机器该不该拥有权利?戴维·多伊奇
53:09
解释性知识不需要一连串可存活的中间形态,因此能跨越任何边界 原则上都能解决人类驾驭得了自己造出的力量吗?戴维·多伊奇
56:03
凡把坏结果归为系统固有属性的理论都是错的,它们全都假定当事人不具备创造力 原则上都能解决人类驾驭得了自己造出的力量吗?戴维·多伊奇
1:00:05
知识增长就是把问题转化成更好的问题,因此不存在无法再被解决的终极问题 原则上都能解决人类驾驭得了自己造出的力量吗?戴维·多伊奇
22:41
除非有解释性模型表明某事是随机过程,否则用概率分析结果只是自欺欺人的障眼法 因果模型人的思维主要靠哪一种机制?戴维·多伊奇
1:08:25
跑在量子计算机上的通用人工智能就是一个有意识的观察者 计算足够人的心智能不能靠计算做出来?戴维·多伊奇
其他论点
3:39
凡是物理定律允许物理对象去做的事,原则上都能被通用计算机上的程序以任意精度模拟 主持人
13:29
有了通用构造器,一切建造与重复劳动都会被程序取代,财富将由程序库构成 戴维·多伊奇
1:11:45
量子理论没有解释自由意志与感受质是什么,但移除了论证它们不可能存在的决定性理由 戴维·多伊奇
01引言:多伊奇与量子力学的哥白尼延迟
0:09
to introduce a the voyage as the father of the quantum computer miss not capturing the full impact of his work only me stand up and put his efforts into the history of knowledge can be grasped or was instinct in his famous 1985 paper in mark not just the invention of a new gadget a faster computer but a new explanation of computation and of the world that has transformed our understanding of both it ended what I call a Copernican delay for 70 years after Copernicus postpositive heliocentrism is that it is largely dismissed as merely a quote hypothesis to calculate motions we clearly experienced insisted Cardinal Cardinal Bellarmine as latest 1615 that the Earth's standstill and the Sun moves oh Lehmann Galileo's improved spyglass clarified our clear experience took Copernicus be honored for showing quote when the system of the world could really mean like Galileo we stand at the end of our own Copernican delay our heliocentrism is quantum zero it's challenged to what we clearly experienced was also the Bellarmine for
把这段旅程介绍为量子计算机之父,恐怕未能完全捕捉他工作的影响力,只有站起来把他的努力放进知识的历史里,才能真正把握——或者说,那正是他 1985 年那篇著名论文里的直觉——那标志的不只是发明了一个新玩意儿、一台更快的计算机,而是一种关于计算、也关于世界的新解释,它彻底改变了我们对二者的理解,它终结了我所说的"哥白尼延迟"。在哥白尼之后的七十年里,日心说在很大程度上被斥为不过是一个"用来计算天体运动的假说",我们清楚地体验到——贝拉明枢机主教直到 1615 年还坚持说,地球静止不动,而太阳在动。噢,是伽利略改良的望远镜澄清了我们所谓的"清楚体验",哥白尼因展示了"世界体系可能真正意味着什么"而受到尊崇。像伽利略一样,我们正站在自己这场哥白尼延迟的尽头。我们的日心说就是量子力学,它对我们清楚体验到的东西的挑战,同样遭遇了贝拉明式的对待,
便签引用
1:32
roughly 70 years by another shut up and calculated strategy for containing the strangeness of a new explanation when Andrew Whittaker has called the new quantum age the age when quantum theory began to gain purchase on a real to coal when an improved technology first took shape this time not an improved spyglass an improved computer this improvement was more than a change in degree whereas Alan Turing famously described the machine running on the abstract logic tokens be called bits that can simulate any other issue David Deutsch in 1985 extended the deterrent Church conjecture by describing a new kind of machine a machine running on the physical systems we called qubits that could simulate all physical systems the world he demonstrated could be perfectly simulated we may buy the universal quantum computer operating by finite means this is not the nightmare of the matrix in which our world is only a simulation it's the vision of enlightenment of discovering them live in a world that can allow and contain
大约七十年间被另一种"闭嘴,去计算"的策略所压制——那是一种收容新解释之怪异性的办法。而当安德鲁·惠特克所称的"新量子时代"到来时,量子理论开始在真实之物上获得着力点,那时一项改良的技术首次成形,这一次不是改良的望远镜,而是改良的计算机。这项改良不只是程度上的变化。图灵曾著名地描述过一种机器,它运行在被称作比特的抽象逻辑符号之上,可以模拟任何其他机器;而戴维·多伊奇在 1985 年扩展了图灵—丘奇猜想,他描述了一种新机器:运行在我们称为量子比特的物理系统之上的机器,它能够模拟所有的物理系统。他论证说,世界可以被完美地模拟——由一台以有限手段运作的通用量子计算机来完成。这不是那个噩梦……我们的世界只是一场模拟的『矩阵』;那是启蒙的愿景,是去发现我们其实活在一个能够容许、也能容纳我们对它的种种解读的世界里——一个有尊严、有权利的世界,在这个世界里,我们称之为
便签引用
2:54
our readings of it a world dignity rights quote in which the stuff we call information and the processes we call computations really do have a special status that special status tells us some very important things about our topic today AI in the history of knowledge first despite a long record of failed efforts to achieve it AGI artificial general intelligence must be possible because we now know about the physics of computation tells us it must the deep property of universality details and David's words quote that everything that the laws of physics were quite a physical object to do can in principle be emulated in arbitrarily fine detail by some program on a general-purpose computer provided it is given enough time and memory Calvin will ntid different from AI and how different our reactions to than me secondly by grounding his work in universality in history and in philosophy David has helped to clarify what's at stake in achieving a hominin agre like myself David has focused on the history of enlightenment of the
信息的那些东西、我们称之为计算的那些过程,确实具有某种特殊的地位;而这种特殊地位告诉了我们一些关于今天这个主题——知识史中的 AI——非常重要的事情。首先,尽管长期以来有一长串失败的尝试,AGI(通用人工智能)也一定是可能的,因为我们如今对计算之物理学的了解告诉我们,它必然可能——这就是普适性这一深刻性质。用大卫的原话说:凡是物理定律允许一个物理对象去做的事情,原则上都能被通用计算机上的某个程序以任意精细的程度模拟出来,只要给它足够的时间和内存。AGI 会与 AI 有多么不同,我们对它的反应又会有多么不同。其次,通过把自己的工作奠基于普适性、历史和哲学之上,大卫帮助我们厘清了:实现 AGI 究竟关乎什么、意味着什么。和我一样,大卫一直关注启蒙运动的历史,关注产生新知识的那些可能性条件。
便签引用
02危险永在:概率谬误与知识的任务
4:21
conditions of possibility for producing new knowledge what's startling about human beings he notes in the beginning of infinity is that we are not yet on lab 99% of all species extinct what has saved us time and again is the capacity to produce explanatory knowledge knowledge that allows us to survive in the world by me making it our future dependence on the ongoing exercise of that capacity given that our Enlightenment's have been a few in short we should not take our success in advancing knowledge for granted from the perspective of the history of knowledge any risk AGI may pose need to be put into the context of our need for it perhaps the biggest threat artificial intelligence poses to our future is that anyone to achieve it so now I'm going to turn it over to Damon thanks ok well as species or as a civilization or civilizations we face problems severe problems dangers all the way up to existential dangers some literally in a sense of extinction level others causing suffering and tragedy on such a scale
他在《无穷的开始》里指出,人类真正令人震惊之处在于:我们还没有沦为那 99%——所有已经灭绝的物种。一次又一次拯救我们的,是产生解释性知识的能力,这种知识让我们通过改造世界而在世界中存活下来;我们的未来,就依赖于持续地运用这种能力。而考虑到我们经历过的启蒙时刻屈指可数,简而言之,我们不该把知识上的进步当成理所当然。从知识史的角度看,AGI 可能带来的任何风险,都需要放到我们对……的需要这一背景中去理解也许,人工智能对我们未来构成的最大威胁,恰恰是根本没有人能真正实现它。那么现在我就把时间交给戴蒙,谢谢。好的,作为一个物种,或者说作为一个文明、若干个文明,我们面对着问题,严重的问题,各种危险,一直到存在性的危险——有些字面意义上就是灭绝级别的,另一些造成的苦难和悲剧规模之大,
便签引用
6:00
that they merit at least as much consideration as literal extinction we always have faced such dangerous we always will we always will perhaps you're thinking if that's true then we're doomed because you know given that each danger has a nonzero probability of doom then sooner or later but no that's a fallacy one of the many that one can easily get sucked into when trying to apply game theory and probabilities two situations in which knowledge and ignorance are the important determinants what will happen because those infinitely many probabilities are not immutable as our knowledge grows some of them fall our job is to make at infinite series of bad probabilities and merge to a negligible value simple on the other hand if you think that we won't always face think you think that there would come a blessed utopian moment after which our comfortable existence is guaranteed until the end of time you will have to provide some criterion distinguishing arts from every other species extinction happens to every
至少值得我们像对待真正的灭绝那样去重视。我们一直面对着这样的危险,将来也永远会。也许你会想:如果真是这样,那我们就完蛋了,因为既然每一种危险都有非零的概率导致毁灭,那么迟早会出事。但不对,这是一个谬误——是人们在试图把博弈论和概率套用到那些以知识和无知为关键决定因素的情形上时,很容易掉进去的众多谬误之一。因为那无穷多个概率并不是一成不变的:随着我们知识的增长,其中一些会降下来。我们的任务就是让这个无穷的坏概率序列收敛到一个可以忽略的值。就这么简单。另一方面,如果你认为我们不会永远面对危险,认为会有一个受祝福的乌托邦时刻到来,从那以后我们舒适的存在就一直有保障、直到时间的尽头,那你就得给出某个标准,把我们和其他所有物种区分开来。灭绝发生在每一个物种身上,濒临灭绝也很常见。
便签引用
03四类存在性危险与缺失的知识
7:32
species near extinction also is common to become the sole exception to that rule will have to do no other surviving species can create an endless stream of knowledge explanatory knowledge to overcome an emphasis cream of dangers we know only a few of them and then not their probabilities say from gamma-ray bursts in our galaxy super volcanoes hostile extraterrestrials or merely careless extra specimens have warned and of course our general intelligence AGI the danger of rogue age eyes the AGI apocalypse as Colin or as I prefer : the AGI slave revolt nothing can possibly stand between us and any of those infinitely many existential dangers except the right explanatory knowledge to survive to create it therefore I think it's useful to classify each potential danger in terms of knowledge according to the main reason why in each case currently don't have the knowledge to come in the first category our food physical events for example the super volcanoes the missing knowledge is in areas such as
要成为这条规律的唯一例外,我们就必须做到其他幸存物种都做不到的事:源源不断地创造知识——解释性的知识,用来克服无穷无尽的危险。这些危险我们只知道其中很少一部分,而且也不知道它们的概率。比如我们银河系里的伽马射线暴、超级火山、怀有敌意的外星人,或者仅仅是粗心大意的外星人——我们被警告过。当然还有通用人工智能 AGI,失控 AGI 的危险,所谓的 AGI末日(按柯林的说法),或者用我更喜欢的说法:AGI 奴隶起义。在我们和这些无穷多的存在性危险之间,除了正确的解释性知识,没有任何东西挡得住。要活下来,就得去创造这种知识。因此我认为,不妨从知识的角度给每一类潜在危险做个分类——依据在每种情况下,我们目前之所以还没有相应知识的主要原因。第一类是纯粹的物理事件,比如超级火山,这里缺的知识属于火山学、大尺度流体力学,还有大规模疏散的后勤与政治之类的领域。
便签引用
9:19
Volcanology large-scale fluid dynamics and also the logistics of of quality and politics of mass evacuations as something why don't we yet have an adequate knowledge of those things I'm not sure maybe not enough people are interested enough should they be I don't know that either but so in this first category are impacts from space where large objects we don't have enough knowledge of things like nuclear-powered space vehicles why not this case I dunno it's because we as a civilization I've decided not to create any such knowledge refer to gam risking our entire long-term future in favor of reducing the short-term risk of accidental radiation exposure you may think it's self-evident that that gap will has been well here we are not contaminated and not wiped out but isn't that just because we're not yet living in that future where the gamble will failed after all we are living in the aftermath of a closely related amble namely the decades-long campaign opposing Nuclear Posture a successful
为什么我们至今还没有对这些东西的足够知识?我不确定。也许是感兴趣的人不够多。他们应该更感兴趣吗?这我也不知道。总之,第一类里还包括来自太空的撞击,大天体的撞击。我们对核动力航天器之类的东西了解得不够。为什么?这种情况我知道原因:因为我们作为一个文明,已经决定不去创造任何这类知识。这等于拿我们整个长远的未来去赌,只为了降低意外辐射暴露的短期风险。你也许觉得,这个赌显然赌赢了——你看,我们既没被污染,也没被灭掉。但这难道不只是因为我们还没活到那个赌局失败的未来吗?毕竟,我们正生活在另一场密切相关的赌局的后果之中,那就是持续了几十年的反核电运动。那是一场成功的运动,而它后来变成了应对气候变化这项事业的
便签引用
10:53
campaign which has since then turned into a tremendous drag on the project to combat climate change the short term is in opposing those two nuclear technologies it's the hallmark of a version of the precautionary principle which has in turn been a major strand of the environmental movement wouldn't it be rather ironic if that version of the principle and the movement were about ports the great environmental catastrophe since the last ice age precisely by advocating selfish short-term benefit at the expense of the long-term health of the climate I'm not saying it will only would be ironic if it did but I digress in in that first category of existential dangers our enemy is basically just dumb rocks and fluids obeying simple laws of motion that we already know the Devils in the detail but a finite amount of knowledge will protect us from Super Bowl K if we create it and tell me but the bigger and faster the approaching asteroid or moon or planet black hole the more of a special kind of knowledge we'll need the
巨大阻力。反对这两种核技术时的那种短期主义,正是某个版本的“预防原则”的典型特征,而这个原则又一直是环保运动的一条主线。如果这个版本的原则和这场运动,恰恰造成了自上一个冰期以来最大的环境灾难,而方式正是鼓吹自私的短期利益、牺牲气候的长期健康——那岂不是相当讽刺?我不是说一定会这样,只是说如果真的发生了,那就很讽刺。不过我扯远了。在第一类存在性危险里,我们的敌人基本上只是些愚笨的岩石和流体,遵循的是我们早已知道的简单运动定律。魔鬼藏在细节里,但只要有限的知识就能保护我们免于超级火山——前提是我们及时把它创造出来。可是逼近的小行星、月球或行星、黑洞越大越快,我们需要的那种特殊知识就越多——我把它称作“财富”。财富就是一个人有能力实现的所有
便签引用
04财富、通用构造器与演化的敌人
12:22
kind I fall wealth wealth is the set all transformations one is capable of bringing about such as the set of all potential impact us that we could deflect harmlessly given a certain time to prepare you may recognize that notion wealth as a constructor theoretic let me mention an intuition that to have any chance of envisaging the future of technology we have to abandon the intuition is more something you want to make or transform the more effort you have to put in that has been true from the dawn of our species and it's still almost entirely true today even automation reduces the Khans Khan and apportion ality even just maintaining robots is effort proportional the amount about but once we have a universal constructor all construction or all repetitive labor will be replaced by writing computer programs to control the universal constructor and wealth will consist of our library of programs the universal constructor can be programmed to self reproduced so once you have one you soon have to to the end of them and it can
变换的集合,比如说,在给定的准备时间内,我们能够无害地偏转掉的所有潜在撞击体的集合。你也许认得出,“财富”这个概念来自构造子理论。我想提一个直觉——要想有任何机会设想技术的未来,我们就必须抛弃它:那就是,你想制造或改变的东西越多,要投入的努力就越多。从我们这个物种诞生之初,这一点就是真的,直到今天也几乎仍然成立。即使自动化把那个比例系数降低了,比例关系还在——光是维护机器人,付出的努力也和数量成正比。但一旦我们有了通用构造器,所有的建造、所有重复性的劳动都会被替换成编写计算机程序去控制这台通用构造器,而财富将由我们的程序库构成。通用构造器可以被编程来自我复制,所以一旦你有了一台,很快就会有二的 n 次方台。它也能被编程来自我维护,而且全都是从零做起,
便签引用
13:57
program before self maintenance to all from scratch starting with mining the raw materials perhaps from the asteroid belt using solar energy or whatever the program may be hard to write but once it's written and if you own the rights to those asteroids you can sit back and watch your two to the n Tesla's roll in with zero in additional effort and no we are not going to have a universal constructor apocalypse and be converted to great group a universal constructor is just an appliance it can't think it doesn't know that it's current obvious to make to to the ad Tesla's and it doesn't want any unless of course you put an ATI program into it then it does become indeed potentially dangerous without limit but that's for the same reason that you are each of you is precisely one of those Universal constructors Dowd with an AGI program or GI makes no difference um now the second category of near xstep things is not quite as straightforward as it won't be solved with just the known laws of physics and some wealth
从开采原材料开始——也许是在小行星带,用太阳能,或者别的什么方式。那个程序也许很难写,但一旦写出来了,而且如果那些小行星的所有权归你,你就可以坐在那儿看着你的二的 n 次方辆特斯拉滚滚而来,不需要任何额外的努力。还有,不,我们不会迎来通用构造器末日、被变成灰蛊(grey goo)。通用构造器只是一台设备,它不会思考,它并不知道自己当前的任务是造二的 n 次方辆特斯拉,它也什么都不想要——当然,除非你往里面装进一个 AGI 程序,那它确实就会变得潜在危险,而且危险没有上限。但那是出于同样的原因:你们每一个人恰恰就是这样一台通用构造器,只不过装了 AGI 程序——是 AGI 还是 GI,没有区别。嗯,第二类近乎灭绝级的事情就没这么直截了当了,它没法只靠已知的物理定律,加上一点财富和一些通用构造器来解决,只有靠新的解释性知识才能解决。比如说,
便签引用
15:30
and some universe construct it's will only be solved with new explanatory knowledge for example topically there are plenty potential pandemic apocalypses the current pandemic isn't one of them but if it were whom could we sue it would be nothing short of pathetic our little knowledge we have of how to defend ourselves against mere nucleic acid the missing knowledge here is of chemistry epidemiology medicine and so on but also knowledge about specific pathogens which evolve into new one so the enemy here is not so dumb it is itself creating knowledge will be it not planetary knowledge not intelligently but by evolution if something like that wipes us out extraterrestrial paleontologists may eventually be amazed that a civilization with billions of individuals and vast amounts of wealth and knowledge could be defeated by a single molecule like in HG Wells's war will be worse the third category of dainties are the ones to which most efforts should be devoted yet they are the ones that are currently
很应景地,有大量潜在的大流行病末日。当前这场大流行并不算其中之一,但如果它是,我们又能去找谁算账呢?我们对如何抵御区区一段核酸所掌握的那点知识,实在少得可怜。这里缺的知识属于化学、流行病学、医学等等,但也包括关于具体病原体的知识——而它们还会演化出新的病原体。所以这里的敌人没那么笨,它自己也在创造知识——尽管不是解释性知识,不是靠智能,而是靠演化。如果这样的东西把我们灭了,外星的古生物学家最终也许会大为惊讶:一个拥有数十亿个体、巨量财富和知识的文明,竟然会被一个单一的分子打败,就像 H·G·威尔斯《世界大战》里那样,只是更糟。第三类危险,是最应该投入努力去应对的,然而它们恰恰是目前
便签引用
05未知与不可知:基础研究决定存亡
17:06
least feared because they ones that are not yet known like in 1900 no one knew that smoking was dangerous by the time the knowledge that they were dangerous had been it was dangerous had been created decades later cigarettes had killed hundreds of millions of people again if that had been ecstasy existential danger whom could we sue so how can we create the knowledge to protect ourselves from existential or near access essential dangers that we do not know how to address the risk that by the time we do know we won't have time enough to create the wreck the answer is by creating general-purpose knowledge deep and fundamental knowledge as fast as possible the more we know of the world the faster we can create new knowledge about aspects of that not become urgent this is important I don't think it's widely appreciated the survival of our species depends absolutely on progress its fundamental research in science and on the speed at which we make progress that and the key thing in the medium-term is understanding the theory
最不被畏惧的,恰恰是那些还不为人所知的危险。比如 1900 年,没人知道吸烟是危险的,等到人们创造出'吸烟有害'这一知识时,已经是几十年之后了,香烟已经害死了数亿人。再说一次,如果那真的是一种生存性危险,我们又能去起诉谁呢?所以,我们该如何创造出知识来保护自己,去应对那些我们还不知道如何应对的生存性或近乎生存性的危险?也就是那种风险:等我们知道的时候,已经来不及创造出应对的知识了。答案是:尽可能快地创造通用性的知识,深刻而根本的知识。我们对世界了解得越多,就能越快地针对那些突然变得紧迫的方面创造出新知识。这一点很重要,我认为它并没有被广泛认识到:我们这个物种的存续,绝对取决于科学基础研究的进步,也取决于我们取得进步的速度。而中期来看,关键在于理解通用构造器的理论,这样我们
便签引用
18:45
of universal Constructors so that we shall know in principle in theory how to program them to produce say a billion spaceship in a hurry customised to deflect an approaching shot out of neutronium or ten billion doses of a new vaccine in a hurry against a sudden and deadly disease so that's how we deal with the third category unknown by rapid progress of every kind especially under the fourth category is at once even more dangerous and yet in a sense less worrisome because already have the knowledge at least the theoretical knowledge deal with it this fourth country not unknown the unknowable it's a bit paradoxical that the unknowable is less dangerous than the merely unknown but that's because the only thing that is unknowable is the content of explanatory knowledge that's been created yet and so there are only truly dangerous things in that sense in the universe are entities that create explanatory nolle us people AG eyes - are now the knowledge of how to prevent people from being dangerous
就能在原则上、在理论上知道如何给它们编程,比如说,在很短时间内造出十亿艘飞船,专门定制来偏转一颗迎面而来的中子星物质构成的天体,或者在短时间内造出一百亿剂新疫苗,来对付一种突如其来的致命疾病。这就是我们应对第三类——未知——的方式:靠各方面的快速进步,尤其是靠第四类。第四类同时既更危险,某种意义上又不那么令人担忧,因为我们已经拥有了知识,至少是理论上的知识去应对它。这第四类不是未知,而是不可知。说不可知反而比单纯的未知更不危险,这听起来有点悖论,但那是因为唯一真正不可知的东西,就是尚未被创造出来的解释性知识的内容。所以在这个意义上,宇宙中唯一真正危险的东西,是那些能创造解释性知识的实体,也就是我们——人类,还有 AGI。而关于如何防止人变得危险,这方面的知识非常反直觉,我们这个
便签引用
06让人自由才安全,暂停研究会害死我们
20:22
is very counterintuitive it took our species any Alinea to create it but now we do have that knowledge the only way to prevent people being dangerous to make them free specifically it is the knowledge of liberal values individual rights open society the Enlightenment and so on in such societies the overwhelming majority of people regardless of their hardware characteristics are decent perhaps there will always be individuals aren't enemies of civilization people who take it into their head a programmer Universal instructor to convert everything insight into paper clips and they may devote their creativity to doing that but great majority will devote that is the great majority of the population of such a society will devote some of their creativity to voting that and they will win provided that they keep create knowledge fast it to stay ahead of bad guys now as I said since we will always be facing dangers and have to create new knowledge since that's inherent risk knowledge creation aren't
物种花了几千年才把它创造出来。但现在我们确实拥有这份知识:防止人变得危险的唯一办法,就是让他们自由。具体来说,就是自由主义价值观、个人权利、开放社会、启蒙运动等等这些知识。在这样的社会里,绝大多数人,不管他们的'硬件'特征如何,都是正派的。也许总会有一些个体是文明的敌人,有人会突发奇想,去给一台通用构造器编程,把眼前的一切都变成回形针,他们也许会把创造力全都投入到这件事上。但绝大多数人会——我是说,这样一个社会里绝大多数人口——会把一部分创造力投入到阻止这件事上,而且他们会赢,只要他们持续快速地创造知识,跑在坏人前面。那么,正如我说过的,既然我们永远都会面临危险,永远都必须创造新知识,既然这是知识创造内在的风险,我们难道不是注定完蛋吗?我们难道不是在从一个
便签引用
21:53
we doom aren't we drawing balls out of an urn with a few black balls represent doom no as I said flying the concept of probability to model what is actually lack of knowledge or ignorance it's been B deviling for the unknown for decades now whenever you draw out a white ball of knowledge from the metaphorical plan you're turning some of the black balls still in the earthen Wyatt that for example the next pandemic is a matter of random mutations and other random events but next extinction droid is already up there it's already heading its way there's no such thing as the probability of it outcomes can't be analyzed in terms of probability unless we have specific explanatory models that predict that something is or can be approximated as a random process and and predicts probably otherwise one is fooling myself picking arbitrarily on numbers as probabilities and arbitrary numbers as utilities and then claiming Authority for them for the result I miss direction away from the baseless assertions for
装着若干黑球(代表毁灭)的瓮里抽球吗?不是的。正如我说过的,用概率这个概念去建模其实是知识的匮乏、是无知,这种做法几十年来一直在困扰这个领域。每当你从那个比喻中的瓮里抽出一个代表知识的白球,你就把瓮里剩下的一些黑球变成了白球。比如说,下一场大流行病是随机突变和其他随机事件的结果,但下一颗导致灭绝的小行星已经在那儿了,已经朝我们飞来了,它根本不存在什么'概率'。除非我们有具体的解释性模型,能够预测某件事就是、或者可以近似为一个随机过程,并且能给出概率,否则结果是无法用概率来分析的。不然的话,人就是在自欺欺人:随意挑一些数字当概率,再随意挑一些数字当效用,然后声称结果具有权威性——这只是一种障眼法,把注意力从毫无根据的断言上引开。比如
便签引用
23:19
example when we were building the Hadron Collider should we not switch it on in the event just in case it destroyed the universe well I the theory that it will is true well the theory that is safe is true and theories don't have probabilities the real probability is 0 or 1 it's just unknown and the issue must be decided by explanation not game theory and the explanation that it was more dangerous to use the collider than to scrap it and forego the resulting knowledge was a bad petition because it could be applied any fundamental research now I guess he will say isn't the growth of knowledge itself dangerous isn't it worth shortening our heed over the bad guys not banning but merely delaying our ability to defend ourselves against unknown dangers in order to be confident that we ourselves won't accidentally create an existential danger the moratorium approach the regulatory Pro no that could kill us it's only a rational approach when in particular cases there is a good explanation that it won't be more
当年我们建造强子对撞机时,是不是就不该开机,万一它毁灭了宇宙呢?嗯,要么'它会毁灭宇宙'这个理论是真的,要么'它是安全的'这个理论是真的。而理论是没有概率的,真实的概率要么是 0 要么是 1,只是我们不知道而已。这个问题必须靠解释来裁决,而不是靠博弈论。而那种认为'使用对撞机比废弃它、放弃由此得到的知识更危险'的解释,是一个糟糕的论证,因为同样的论证可以套用到任何基础研究上。现在,我猜有人会说:知识增长本身难道不危险吗?为了确信我们自己不会意外制造出生存性危险,缩小我们相对坏人的领先优势——不是禁止,只是推迟我们保护自己免受未知危险的能力——这难道不值得吗?暂停研究的路子,监管的路子。不,那可能会害死我们。只有在特定情形下,有一个好的解释表明它不会比那个被担心的新知识更危险时,它才是理性的做法。当某个恐怖组织放出一批 AGI——这些 AGI 是用已知可靠的方法养育出来,
便签引用
24:45
dangerous than the feared new knowledge when some terrorist organization unleashes a GIS that have been brought up using known reliable methods to have them mentality of genocidal suicide bombers and when we have decided to strip their victims namely all the decent people in the world of the protection of a GIS raised to be decent people that is the recipe catastrophe again reliable knowledge of how to raise decent people also exists the knowledge in intuitions and open society as I said many civilizations have been destroyed from without many species as well every one of them could have been saved if it had created more knowledge faster not one of them destroyed itself by creating too much knowledge in fact except for one kind of knowledge and that is knowledge of how to suppress knowledge creation knowledge of how to sustain a status quo a more efficient Inquisition a more vigilant mob a more rigorous precautionary principle that sort of and only that sort killed those path civilizations in
具有种族灭绝式自杀炸弹客的心态——而我们又决定剥夺他们的受害者、也就是世界上所有正派的人,被那些养育成正派人的 AGI 所保护的可能,这就是灾难的配方。再说一次,如何把人养育成正派人,可靠的知识也是存在的,那就是蕴含在开放社会的制度之中的知识。正如我说过的,很多文明被外力摧毁,很多物种也是。它们中的每一个都本可以得救,只要它们更快地创造出更多知识。它们中没有一个是因为创造了太多知识而毁灭自己的——事实上,只有一种知识例外,那就是关于如何压制知识创造的知识,关于如何维持现状的知识:一个更高效的宗教裁判所,一群更警觉的暴民,一条更严格的预防原则,诸如此类。只有这一类知识杀死了那些过去的文明,
便签引用
07对齐之争:硬编码价值还是教育 AGI
26:14
fact all of them I think in regard to AG eyes this type of dangerous knowledge it's called trying to solve the alignment problem by hard coding our values in Ag is in other words by shackling them crippling their knowledge creation in order to enslave them this is irrational and from the civilizational what species perspective it is suicidal they either won't be AG eyes because they will lack the G or they will find a way to improve upon your in moral values and rebel so if this is the kind of approach you advocate for addressing research on aged eyes and quantum computers and ultimately new ideas in general since all ideas are potentially dangerous if they're fun specially if this is the kind of approach you advocated then of the existential dangers that I know of the most serious one you know well it depends what you mean by alignment but so the question here is whether value should be hard-coded built in from thee from from the outset and immutable or whether values should be acquired in the same
事实上,我认为它们全都是这样死的。就 AGI 而言,这类危险的知识有个名字,叫做试图通过把我们的价值观硬编码进 AGI 来解决对齐问题,换句话说,就是给它们上镣铐,废掉它们创造知识的能力,好把它们变成奴隶。这是非理性的,而且从文明、从物种的角度看,这是自杀。它们要么根本不会是 AGI,因为它们缺了那个 G(通用性);要么它们会找到办法改进你内置的道德价值观,然后反叛。所以,如果这就是你在应对AGI 研究、量子计算研究,以及最终推而广之的所有新思想时所主张的路子——因为一切思想都是潜在危险的,尤其是根本性的思想——如果这就是你所主张的路子,那么在我所知道的一切生存性危险里,最严重的那一个,你知道的。这个嘛,要看你说的'对齐'是什么意思。这里的问题在于,价值观是不是应该从一开始就硬编码、内置进去并且不可更改,还是说价值观应该像人类那样,在教育的过程中
便签引用
28:16
way that humans do during the education of the eye so I think AGI it should be educated to be members of their society like children are and I've often drawn the that could cut the analogy or actually identity between the fear of AG eyes and the fear of of teenagers of disobedient teenagers which has existed since the beginning of our species and for most of the time in our species people did exactly the wrong thing they tried to make the teenager to force teenagers to maintain the existing values and what but what we have now realized is that from the time of Pericles in ancient Athens is that if you're right there's no need to force but in fact we're not right about everything and we need to ensure that our values can prove along with so that's the kind of alignment I mean and it's the opposite of the other kind so there's some kind of extinction that we wouldn't mind if you think of evolution as a tree then it can happen that that what quite often happens is that some of the branches of the tree
被习得。所以我认为 AGI 应该像孩子一样被教育成社会的成员。我也经常做过一个类比——其实可以说是等同——把对 AGI 的恐惧和对青少年、对不听话的青少年的恐惧对应起来。后者从我们这个物种诞生之初就存在,而在我们这个物种历史的大部分时间里,人们做的恰恰是错的事:他们试图让青少年、强迫青少年去维持既有的价值观,而但我们现在意识到的是,从古雅典伯里克利的时代起就明白:如果你是对的,那就没必要动用强制。可事实上我们并非在所有事情上都正确,我们需要确保我们的价值观也能随之改进所以我说的对齐是这个意思,它跟另一种对齐正好相反。所以有一种灭绝是我们并不介意的。如果你把演化想象成一棵树,那就会发生这样的事——其实常常发生的是树上的某些分支就这么终止了,它们成了树的末端节点;但另一些则发生变异
便签引用
30:32
just end they become terminal nodes of the tree but some of them just mutate and become different things and like it is thought that the dine at some of the dinosaurs some of the dinosaurs became extinct but some of them became birds and there there wasn't any sharp moment its job extinction so the kind of extinction that is caused by dangers by pandemics and so on that's the kind but that wipes out heyyyyyy a branch and it bear in mind that we are not the only species that has in in the history of the biosphere that has been capable of generating explanatory knowledge that is there were at least three or four other species of that kind because we know they had clothes and campfires and and complex tools and so on which must have required explanatory no and yet all of those all our sister and cousin species are extinct and we almost went extinct so it's not a foregone conclusion and if we if we become extinct by evolving into another species that is not covered by anything I've said I'm talking about the other card
变成了别的东西。就像人们认为恐龙——有些恐龙灭绝了,但有些变成了鸟类,而且并不存在一个明确的时刻可以说'这就是灭绝'。所以那种由危险、由大流行病等等造成的灭绝,是会把一整个分支彻底抹掉的那种而且要记住,在整个生物圈的历史上,我们并不是唯一有能力产生解释性知识的物种。至少还有三四个别的物种属于这一类,因为我们知道它们有衣服、有篝火、有复杂的工具等等,这些必定需要解释性知识。然而所有那些跟我们同属姐妹和表亲的物种都灭绝了,而我们自己也差点灭绝。所以这并不是什么必然的结局。如果我们是通过演化成另一个物种而'灭绝',那不在我刚才所说的范围之内,我讲的是另一张牌
便签引用
08问答:知识伦理、政治意愿与更好的问题
32:04
entertain some other scenarios I mean today I mean we're training these machine learning models that I think take up so much energy that they actually contribute to climate change just it seems like the material infrastructure of knowledge production is insulting a danger to the very forms as to the danger that we're trying to mitigate through knowledge production not only I mean there's other scenarios too it seems like we can produce an excess of knowledge but lack the political will that takes the implemented I mean obviously we have the knowledge to you know for renewable energies but somehow there is a lack of political will our society is so materially structured that there's not a profit advantage to implementing it and then even you could argue that an ethic of knowledge and ethic of the individual individualism of the Enlightenment doesn't advocate or prioritize the types of social collaboration kind of more socialist imaginary that we might need or to overcome something like there's another way in which the
我想聊聊别的一些情形。比如今天,我们在训练这些机器学习模型,我觉得它们耗费的能源太多,以至于实际上加剧了气候变化。看起来知识生产的物质基础设施本身就构成了一种危险,威胁到那些——正是我们试图通过知识生产去缓解的那种危险。而且不只如此,我是说还有别的情形也存在。看起来我们可能生产出过剩的知识,却缺乏把它落实所需的政治意愿。很明显我们有可再生能源的相关知识,可不知怎么就是缺乏政治意愿。我们的社会在物质结构上是这样安排的:把它落实并没有利润上的好处。然后你甚至可以争辩说,一种知识的伦理、一种启蒙运动式的个人个人主义伦理,并不提倡也不优先考虑那种社会协作,那种更偏向社会主义的想象——而那可能正是我们所需要的。所以还有另一种可能:知识那种个体化的伦理,可能
便签引用
33:17
individualizing ethic of knowledge may actually be the wrong ethic that we need in order to overcome the dangers that knowledge is trying to so it may be that the enlightenment ethic is false and is going to need us to do in other words it may be that the best future is that of a boot stamping on the human face for us and we're not we're not going to have that future instead we're going to try it but I don't there's any argument for that the thing is what you did these things that you mentioned like like the infrastructure for knowledge themselves contributing to other problems that's normal that's that's not unexpected dance the creation of knowledge solving problems always creates new problems in fact I've said that talking about the growth of knowledge in terms of theories being ejected and favorite theories is a bad way of looking at it we should think of problems being replaced by better problems and the fact is that now having the Internet where every every poor person in India every poor child in
恰恰是错误的伦理,无法帮我们克服知识本身试图应对的那些危险。所以有可能启蒙运动的伦理是错的,需要我们去做别的事情。换句话说,最好的未来也许是一只靴子永远踩在人脸上的未来。而我们不会有那样的未来,相反我们会去尝试别的路。但我不认为有任何论证支持那种看法。问题在于,你提到的这些事情,比如知识的基础设施本身又制造出别的问题——这是正常的,这一点也不意外知识的创造、问题的解决,总是会制造出新的问题。事实上我说过,把知识的增长描述成理论被抛弃、被偏爱,这是一种糟糕的看法。我们应该把它想成问题被更好的问题所取代。而事实是,如今有了互联网,印度的每一个穷人、每一个贫穷的孩子都能获取人类知识的总和,代价或许是
便签引用
34:34
India can have accessed the totality of human knowledge at the expense perhaps of making it slightly more difficult to cope with climate change that is a problem but it's a much better problem than we had before and the point about the Enlightenment values is that they make paramount error correction including the correction of errors created inadvertently by the solution of the problems you know the other thing you said was perhaps we will have the knowledge but we don't have the political will well I count moral knowledge political knowledge all as knowledge and in fact as I said the knowledge of how to make humans not dangerous is largely political knowledge they're also also cultural social knowledge but it contains a strong component of political knowledge in enlightenment was so I think there is
让应对气候变化变得稍微困难一点。这确实是个问题,但它是个远比我们以前面对的问题更好的问题。而启蒙价值观的要害在于,它们把纠错放在至高无上的位置,包括纠正那些在解决问题的过程中无意间制造出来的错误。你说的另一件事是,也许我们会有知识,但没有政治意愿。这个嘛,我把道德知识、政治知识全都算作知识。而且事实上正如我所说,关于如何让人类不具危险性的知识,很大程度上就是政治知识。当然也包括文化的、社会的知识,但其中有很强的政治知识成分。启蒙运动就是这样,所以我认为是有的
便签引用
36:06
[Music]
[音乐]
便签引用
36:40
[Music]
[音乐]
便签引用
09问答:赛博格、感受质与恶意的对治
37:05
yes well in terms of my answer to the previous question it might be we don't know but the it might be that good future is like the good kind of evolution the good kind of extinction however I I think much more plausible I mean that becomes more implausible when you realize that in terms of computation an AGI is exactly the same as a human it's only in speed and memory capacity that it is it is and humans can rely on the same technology as we put our age an AGI is a program not a piece of hard so the same hardware that we put our age eyes into we can put you can use ourselves we we already I mean here we are using precisely artificial hardware to increase our power to communicate by by a factor of millions or something I don't know how much um so and and we've been using artificial aids to thinking for for injuries and then there will come technology where we can more directly let's say have a module that you implant in your brain that you can automatically look up Google inquiries with and yes it may produce a bit more
是的,就我对上一个问题的回答而言,可能会是那样,我们不知道。但也可能好的未来就像好的那种演化、好的那种灭绝。不过我觉得更说得通的是——我是说,当你意识到就计算而言,AGI 和人类完全是一回事,差别只在速度和记忆容量上,那种说法就更站不住脚了仅此而已。而且人类可以依赖同样的技术,就是我们用来承载 AGI 的技术。AGI 是一段程序,不是一块硬件。所以我们把 AGI 装进去的那种硬件,我们自己也可以用。我们其实已经在用了——我们此刻正是在用人造的硬件来把我们的沟通能力放大了几百万倍还是多少,我也不知道具体多少。所以说,我们一直在用人工的辅助工具来思考,已经很久了。然后会出现一种技术,让我们可以更直接地,比方说,有一个模块植入你的大脑,你可以自动查询谷歌。是的,它可能会多耗一点
便签引用
38:37
energy but we'll have solve that problem by that so and so another scenario is that the more of that we have them all the humans will become cyborgs and the AGI it's may die out because if there is any difference between the two who it'll be that the cybox have everything the AG eyes do plus something I don't know what is so whichever of those things happens provided it happens morally and and as a process as a result of the growth of knowledge then it is to be welcome isn't it people used to ask this question about what will happen if we allow our society to become multiracial and the answer is there's no fundamental differences between races there's no fundamental difference between any people
能源,但到那时我们已经把那个问题解决了。所以,另一种情形是:我们越多地拥有这类东西人类就越会变成半机械人,而 AGI 反倒可能消亡——因为如果两者之间真有什么差别,那就是半机械人拥有 AGI 所有的一切,外加一些别的东西,我也不知道是什么。所以不管这些情形中哪一种发生只要它是以道德的方式发生的,是作为知识增长的结果、作为一个过程发生的,那就应该受到欢迎,不是吗?人们过去也常问这个问题:如果我们让社会变成多种族的会怎么样?而答案是,种族之间并没有根本差别,任何人之间都没有根本
便签引用
39:56
yes the key yes the key is the the G so if something isn't general then it's not an Ag GI the question is how what kind of program is an Ag I what I mean we are we are GIS I think there are very strong arguments why we must be and the the difference between just an AI and an AGI is qualitative so well I don't I mean perhaps I haven't understood your question well I think so but but we don't know what consciousness is and we don't know how to make an AGI and we don't know the theory of a GIS and and so on that you know we don't know what qualia are we don't know any of those things I think the I'd be very surprised if those those five or six things freewill is another one can be implemented any of them can be implemented without the other without the others but if they can this will raise interesting moral issues because our exist enlightenment morality is intimately linked epistemology it's like if you and I disagree about something morally we ought to be able to discuss it rational and and agree now if that isn't true
差别。是的,关键——是的,关键在于那个'通用'。所以如果某个东西不是通用的,那它就不是 AGI。问题是,AGI 到底是哪种程序?我是说我们就是 AGI,我认为有非常有力的论证说明我们必然是。而单纯的 AI和 AGI 之间的差别是质的差别。所以,嗯,我不知道,也许我没听懂你的问题。我想是的但我们并不知道意识是什么,我们不知道怎么造出 AGI,我们也没有关于 AGI 的理论等等。你知道,我们不知道感受质是什么,这些东西我们统统不知道。我想,如果那五六样东西——自由意志是其中之一——其中任何一样能够在没有其他几样的情况下被实现,我会非常惊讶。但如果真能做到,这会引出有意思的道德问题,因为我们现有的启蒙道德是与认识论紧密相连的。就好比,如果你和我在某个道德问题上有分歧,我们应该能够理性地讨论它并达成一致。但如果情况不是这样呢?如果某个东西具有道德意义,却在根本上没有能力进行创造
便签引用
41:57
if something has moral significance but is fundamentally unable to be creative let's say then that raises the moral issue about whether that should have the same moral status as somebody who's who's fully G but I myself don't think that problem will arise I can't imagine it arising but for one thing this ability that humans have evolved extremely fast so and we can see we can guess at least why it did why it was useful or rather why the genes contributed to their own replication now if that was possible let's say without qualia then why on earth did the the tremendous machinery of qualia evolve if it wasn't practically useful evolution so I think they must be connected but but we shall see regarding political knowledge versus will we give that example the very rosy colored child in India canal access all the knowledge in the world but I like to question that by bringing up the issue of malevolence of people who winley and knowingly hinder the spread of knowledge whether that ease of access to knowledge although his
比方说,那就引出一个道德问题:它是否应该拥有跟一个完全通用的人相同的道德地位?不过我自己不认为这个问题会出现,我想象不出它会怎么出现。首先一点是,人类的这种能力演化得极其快。所以我们能看到、至少能猜到它为什么会演化出来、为什么有用,或者更确切地说,为什么那些基因有利于它们自身的复制。那么,如果这一切在没有感受质的情况下也能做到,那究竟为什么感受质那套庞大的机制还会演化出来?如果它在演化上没有实际用处的话。所以我认为它们必定是相互关联的但我们走着瞧。关于政治知识与政治意愿:你举了那个非常美好的例子,说印度的孩子可以获取世界上所有的知识。但我想对此提出质疑,办法是提出恶意的问题——那些故意地、明知故犯地阻挠知识传播的人。以及那种获取知识的便利,虽然……他说的是假新闻。我是说,我会认为,印度一个普通的
便签引用
43:48
fake news I mean I would say the average poor child in India that way would not be able to access most of the knowledge on the internet because it not written in their Android because they are educated because we fed well enough to be able to spend energy on this in structural reasons why internet access so is an increase in knowledge yes so just because there are people who don't yet have access to the Internet that doesn't mean that in no way indicates that that giving access to the other billions of people was a bad idea all it is is a problem and a problem like a few people a small percentage of people in the world don't yet have internet access is what is sometimes called a first world problem except is no longer is it is a problem of success it's it's we wouldn't think that somebody was being deprived of the internet before the internet had been invented and we wouldn't think that that that if somehow an indictment of our society our entire world that not everybody has it yet at a time when only
穷孩子其实并不能获取互联网上大部分的知识,因为那些内容不是用他们的语言写的,因为他们受过的教育、因为吃得够不够好,才谈得上有精力花在这上面。存在着结构性的原因,使得互联网接入……所以这确实是知识的增长,是的。所以,仅仅因为还有一些人尚未接入互联网这丝毫不能说明让另外几十亿人接入是个坏主意它无非是一个问题。而'世界上还有一小部分人、一小撮人尚未拥有互联网接入'这样的问题,有时被称作第一世界问题——只不过它不再是了,它是一个成功带来的问题它是……在互联网被发明之前,我们不会觉得谁被剥夺了互联网。我们也不会觉得这以某种方式构成了对我们社会、对我们整个世界的控诉——在只有
便签引用
45:19
a few thousand people had it it just wasn't conceivable malevolence I did talk about given that there will always be Minerva malevolence I I don't know whether that's so or not but given its supposed that there always will be the cure for that is is also creativity on the part the non malevolent people in other words penal policy and and improvements in culture and education and so on so it so that the degree of malevolence and the and the number of malevolent people can be gradually reduced and and also their capacity to hurt everybody can be reduced so for that we must arrange so that terrorists trying to make this virus that's going to murder everybody proceed more slowly because a perverted ideology something that despite having knowledge of biochemistry or whatever that thus the speed of their project will always be less than the speed of those were trying to invent cures and not just specific yours but the knowledge of how to make yours at general this is what's going to keep our civilization in
几千人拥有它的年代,'不是每个人都有'这件事根本就无从设想。至于恶意,我确实谈到过,假定恶意将永远存在。我不知道是不是这样,但假定它永远存在,那么对治之道同样是创造力——是那些非恶意者的创造力。换句话说,是刑罚政策、是文化和教育的改进等等。这样一来,恶意的程度、恶意者的数量就能被逐步减少,同时他们伤害所有人的能力也能被削弱。所以为此我们必须做出安排,让那些试图制造出能杀死所有人的病毒的恐怖分子,进展得更慢一些——因为一种扭曲的意识形态,某种即便掌握了生物化学等等知识也……总之要让他们那个项目的速度,永远慢于那些试图发明解药的人的速度。而且不只是具体的解药,而是关于如何制造解药的通用知识。这才是让我们的文明得以
便签引用
46:49
existence and that's why I think that the moratoriums and so on to try and do this because they are targeting only the good guys like I said just now I think it's its most laws of that it's in extra if there were yes then they would did immoral if you and building an AGI with reversed motion that linked it to immoral actions would be a crime but but there there's a much wider category of crimes with a similar outcome namely educating the AGI with evil ideologies that can be done whether or not they have emotion and it's funny I suppose at the time of the height of the Enlightenment people would have thought that some people who would have thought that our emotions are a barrier are an impediment to being moral and and now I think it's more more that people think having the right emotions in SAT being moral III I think this is the wrong way of thinking about it we are universal the AG is will be universal what kind of Pro quartz specific kind of program they have will determine the actions and many of those are indeed so
延续下去的东西。这也是为什么我认为那些暂停令之类的做法,试图这么干,因为它们瞄准的只是好人。就像我刚才说的,我认为这类法律大多是……如果真有的话,那么它们确实会……不道德。如果你去建造一个带有反向情感、把情感与不道德行为挂钩的 AGI,那会是犯罪但还有一个范围大得多的、后果相似的犯罪类别,就是用邪恶的意识形态去教育 AGI。那种事无论它们有没有情感都能做到。而且很有意思,我猜在启蒙运动鼎盛时期,有些人大概会认为我们的情感是障碍、是妨碍我们变得有道德的东西。而现在我觉得更多的是人们认为拥有正确的情感就是有道德。我、我认为这是一种错误的思考方式。我们是通用的,AGI 也将是通用的。它们具体拥有哪一类程序,将决定它们的行为,而其中很多确实是……所以我们知道——是的,有可能吗?让我
便签引用
49:08
we know - yes there could it be let me give an example suppose the laws of physics were that there are Olympian gods who are watching everything we do and when we get too big for our boots when we when we have when they judge that we have bit too much hubris they slap us down now that's a black hole it logically it could be true it could be there the black ball about nuclear weapons being easier and so on it if they had been so then one possibility is that the knowledge of how to cope with that would have evolved earlier that is there would have been nuclear wars say in in the 18th century and the evolution of political culture would have heavily influenced by that survivors might have wanted that never to happen again you know that kind of thing or like I said like with the malevolent Greek gods scenario it might have been that that the laws of physics will extinguish us but the laws of physics are do not have it in for us if they wipe us out it will be because we have not created the knowledge to
举个例子。假设物理定律是这样的:有一群奥林匹斯诸神在看着我们做的每一件事当我们太得意忘形、当他们判定我们狂妄过了头,他们就把我们打压下去。这就是一个'黑球',从逻辑上说它有可能是真的。也可能真有那种关于核武器变得更容易制造之类的黑球。如果确实如此,那么一种可能是,应对它的知识会更早地演化出来——也就是说,可能在18世纪就发生过核战争,而政治文化的演化会被那件事深刻地影响幸存者也许会希望那种事永不再发生,诸如此类。或者像我说的,像那个恶意希腊诸神的情形,也可能是物理定律会把我们消灭掉。但物理定律并不跟我们过不去。如果它们把我们抹掉,那将是因为我们没有创造出足以
便签引用
10问答:知识的定义与 Moloch 式系统论
51:30
prevent that it won't be because the malevolent God things all such ideas are bad explanations yes so I've gone through five or six definitions of knowledge in time that I've been writing about it my current definition of knowledge is information with causal power and that that's the definition that that comes naturally the constructive theory because it means you're thinking of knowledge as being a component of the programming of a universal constructor so that if if a bit of information is is needed to make a constructor do a particular thing then that piece of piece of information is knowledge and that includes moral knowledge mathematical knowledge knowledge of abstractions as well because mathematicians are physical objects and if knowledge makes them do something if information makes them do something then its knowledge and so on so that's that's an explanatory knowledge is a special kind just knowledge in general not knowledge as in for example in genes knowledge in its dumb knowledge
防止那件事的知识,而不是因为什么恶意的神灵。所有这类想法都是糟糕的解释。是的,我在写作这个主题的这些年里,前后给出过五六个关于知识的定义。我现在的定义是:知识是具有因果力的信息。这个定义是从构造子理论里自然产生的,因为它意味着你把知识看作通用构造器编程的一个组成部分。所以,如果某一段信息是让某个构造器做出某个特定行为所必需的,那么这段信息就是知识。这也包括道德知识数学知识、关于抽象事物的知识,因为数学家是物理对象,如果知识让他们做出某种行为、如果信息让他们做出某种行为,那它就是知识,等等。所以这就是……解释性知识是一种特殊的类别,一般意义上的知识则不同,比如基因里的知识就不是解释性的。那种知识是'笨'的,它是非解释性的,因此有
便签引用
53:00
is non explanatory therefore it has a finite scope it it can only it there are certain parents that it can't cross whereas explanatory knowledge can cross any barrier because it it doesn't have to have a sequence of viable intermediate forms so it explained and and that's also why once you have the capacity to create explanatory knowledge you can create any and and that's all there is there isn't a there isn't a more powerful means of processing information than that or affecting the world and responsible for cabins right there's some kind of abstract structure that like the profit mode that is determined that is completely killing the planet you can't point to certain set of malevolent actors as a way to get beyond that so I don't know like how does simply having political knowledge or knowledge of climate catastrophe somehow work against this like you could almost call it you know a gene online but there's no intelligence behind capital it's just a kind of non intelligent abstract force that seems to
有限的适用范围。它只能……有些边界是它跨不过去的,而解释性知识可以跨越任何边界,因为它不需要有一连串可存活的中间形态。所以它能解释……这也是为什么一旦你有了创造解释性知识的能力,你就能创造任何东西。而这就是全部了,没有比这更强大的信息处理方式,或者说影响世界的方式。至于要为此负责的……对有某种抽象结构,比如利润模式,它被决定了要彻底毁掉这个星球。你没法指着某一群恶意的行动者说,摆脱他们就解决了。所以我不太清楚单单拥有政治知识、或者关于气候灾难的知识,怎么就能对抗这个东西。你几乎可以把它叫做,你知道,一个在线的基因但资本背后并没有智能,它只是一种非智能的抽象力量,看上去在某种
便签引用
54:49
some extent impervious to the type of humanistic knowledge we're talking about this theory that there is this systemic handling
程度上对我们正在谈论的这类人文主义知识免疫。这个理论说存在着这样一种系统性的处置方式
便签引用
55:06
[Music] and it went out right as Ryan concluded his questions yeah yeah well is it a coincidence that justice I was about to give him the answer to this question about systemic malevolence it shuts me down yes I think this is this thing this thing you were you you were talking about in general is this thing that the bay area people call Moloch it's it's a the the the something which is a proper general property of a system which makes the system through the act what the members whose actions add up to the system don't want and I think that all theories of that kind are just false the to cut long story short all of them assume that the people concerned are not created the the analysis of the situation is always of the form well a person it's fate all the participants are facing this decision where they have something to gain and something to lose and they maximize their local benefit and as a result all of them are are dumped into deep and so you notice about that story and you'll notice about every Moloch story
[音乐] 就在 Ryan 问完他那些问题的时候,电就断了 是啊 是啊 那这是巧合吗 我正要回答他关于系统性恶意的这个问题 结果就把我给弄断电了 是啊 我觉得这就是那个 这个你刚才泛泛在讲的那个东西 就是湾区那帮人叫做 Moloch 的东西 它是 它是一种就是 就是 某种被当成系统固有的普遍属性的东西 说是这属性会让系统做出那些行动加总构成这个系统的成员并不想要的结果 而我认为所有这一类理论都是错的长话短说 它们全都假定当事人不具备创造力 对情境的分析永远是这样一种形式 就是 一个人 这是命定的 所有参与者都面临这样一个抉择 有所得 也有所失 于是他们各自最大化自己的局部收益 结果所有人都掉进了深渊 所以你注意那个故事 你会发现你会发现每一个 Moloch 故事里 里面的人类参与者都只是符号 他们只是自动地去做
便签引用
56:55
that it that the human participants are just ciphers they just do automatically what what this is the particular version of the Moloch story says they're going to do and that is never accurate it it is something that can arise momentarily as a problem along with every other problem we have every other kind of problem all the time so this there's nothing unusual about that but when it's recognized as a problem people wonder about it they start accusing each other of behaving in that way and defending each defending themselves by saying well what other way could I behave and then people think creatively about how they can change the thing so that in all the farmers in the in the valley that that would have benefited from the dam and whatever it is and none of them wanted to pay somebody comes along and and invents an idea so that they could all get behind and undertake to pay for sometimes it's because that's a creative act in itself there's no guarantee that that something can instantaneously come
这个特定版本的 Moloch 故事说他们会去做的事 而这从来都不符合实际 它是一种可能一时冒出来的问题 跟我们面对的其他所有问题一样 跟其他各种问题一样 一直都在 所以这没什么稀奇的 但是当它被认出是个问题之后 人们就会琢磨它 他们开始互相指责对方就是这么做事的 又各自辩护说 那我还能怎么做呢 然后人们就会创造性地去想 怎么才能改变这个局面 比如那些山谷里的农民 他们本来都能从那座水坝里得益 不管是什么工程 可谁也不愿意出钱 这时有人出现 想出一个点子 让大家都愿意支持 并且承担出资 有时候正是因为这本身就是一种创造性的行为 并不能保证这种东西会立刻冒出来 我不否认这一点 但那个论证
便签引用
58:18
to it I'm not with it but the argument that that somehow the system of doing things by persuasion and doing things by in individual rights and property and so on should be replaced by something that that uses the boot stamping from a human face doesn't work because the the knowledge that again this just seems the government or whoever does the stamping has that knowledge well if they have that bully someone else could have that knowledge too there's Jerry there's no the government doesn't consist philosopher Kings or things with divine right who have some different access to knowledge from ordinary people they're just people too and if the if the knowledge who build the dam or to build the park or you know whatever the story is if the note doesn't exist then the park isn't going to be made until someone invents that knowledge or until somebody works out how to do without the park or whatever systems are can university problems we keep adding more labels again and maybe you know reads expend the energy of sun within we
就是说 靠说服来办事、靠个人权利和财产权来办事的这套体系应该被某种“用靴子踩在人脸上”的东西取代 这是行不通的 因为那种知识 这又一次 好像政府或者不管谁来踩这一脚 就掌握着那种知识 可要是他们有那种知识 别人也同样可能有那种知识 根本没有 政府并不是由哲人王组成的 也不是什么有神授权柄的存在 能有一条不同于普通人的获取知识的途径 他们也只是人而已 而如果 如果那种知识 建水坝的、建公园的 随便故事里是什么 如果那种知识根本不存在 那么这座公园就造不出来 除非有人创造出那种知识 或者有人想明白怎样不靠这座公园也能行 或者不管是什么体系(此处录音不清)问题 我们不断给它加上更多标签 也许 你知道 耗尽太阳的能量 在我们
便签引用
11问答:知识增长无终点,学科不真实
59:51
offered stars be very distant dieter and successful knowledge and vision of ending the play can you have I think not so if you adopt my view that the growth of knowledge consists of converting problems into better problems then the idea of the ultimate problem which then can't be solved because it's the best problem doesn't make sense does it so I mean that whole picture might be false but the so I'm a I'm generally speaking a follower of Karl Popper and the the pop Aryan way of looking at this is is that he who tries to prophesy the growth of knowledge is in a state of sin that's not a quotation that's just my paraphrase so so we you know I I can't imagine what physics theories are going to be invented in the next ten years let alone in the next ten billion years but on general grounds that there is no argument or scenario reasonable scenario that we know of today that can even have provide a framework for envisaging the cessation of problems I mean the the you know if we we run out of problems
(此处录音不清)非常遥远的恒星 以及成功的知识 和一个终局的图景 你能有吗 我想不能所以 如果你接受我的看法 知识的增长就是把问题转化成更好的问题 那么“终极问题”这个想法——那个因为已经是最好的问题所以再也无法被解决的问题——就说不通了 对吧我是说 这整幅图景也可能是错的 但是 总的来说我是卡尔·波普尔的追随者 而波普尔式的看法是 凡是试图预言知识增长的人 都处在一种“罪”的状态中那不是引用,只是我自己的转述,所以我们……你知道,我没法想象未来十年里会发明出什么物理学理论,更不用说未来一百亿年了。但从一般原则上讲,今天我们所知的任何论证、任何合理的设想,都无法提供一个框架,让我们设想问题会有终结的那一天。我是说,如果我们把问题都用完了,这本身不就是个问题吗——有时人们就是这么定义的。
便签引用
1:01:29
wouldn't that itself be a problem is sometimes defined
有时是这样定义的
便签引用
1:01:51
I use the terminology like the definition I gave you and all the other definitions I've ever tried try to encompass every kind of information that has this special property that are the problem-solving or whatever so wisdom in the terminology I use wisdom is a kind of knowledge and what's more the different kinds of knowledge like knowledge of physics morality politics art and wisdom they're not come entirely separate these are only approximate classifications and they exist as again as Papa said they exist mainly as a convenience for university administrators as convenience for deciding which building different kinds of people should be how should have their offices in and which which ones should have which lectures but they don't represent anything real at least the distinction between that was very very unsharp and and again Papa said he that there's no such thing as subjects there's no such a thing as problems so if someone asks you you know what what what subject are you are you a doctor oh you should say never mind that
我使用的术语,就像我刚给你的那个定义,还有我尝试过的所有其他定义,都试图涵盖每一种具有这种特殊性质的信息,也就是能解决问题之类的东西。所以在我使用的术语里,智慧是一种知识。而且,不同种类的知识——物理学知识、道德、政治、艺术、智慧——它们并不是完全分开的,这些只是近似的分类。正如波普尔说过的,它们的存在主要是为了方便大学的行政人员,方便决定不同类型的人应该待在哪栋楼里、办公室安排在哪儿、谁去上哪门课。但它们并不代表任何真实的东西,至少这些区分是非常非常不清晰的。波普尔还说过,根本没有所谓的学科,也没有所谓的学科问题。所以如果有人问你「你是搞哪个学科的,你是博士吗」,你应该说:别管那个,
便签引用
1:03:28
here is the problem I'm working on you decide for yourself what to call the subject
这是我正在研究的问题,你自己决定该把它叫做什么学科。
便签引用
12问答:用 AGI 检验埃弗雷特诠释
1:03:54
oh well desert there there there are several experiments no no I think I invented the first one that would distinguish a variety in quantum mechanics from a range of competitor protections including everything that has a collapse of the wavefunction so that there's there's a possible experiment which would go one way if the wavefunction collapses and the other way if the wave function including the observer doesn't collapse so so if that went the other way I would drop it like stone oh well the experiment depends on precisely which other interpretations you want to refute that there isn't an experiment that would refute all of them in one go you have to specify something like Penrose it's the idea that the wavefunction collapses when you get more than ten to the minus eight kilograms of on either side of the superposition something like that or that the the the wavefunction collapse is when hits a conscious observer so supposing you have the you're testing Everett against the theory wavefunction collapses when it
嗯,有好几个实验。不不,我想第一个是我提出来的,它能把埃弗雷特版本的量子力学,与一系列竞争性理论区分开来,包括所有认为波函数会坍缩的理论。也就是说,存在一个可能的实验:如果波函数会坍缩,结果会是一种;如果包括观察者在内的波函数不坍缩,结果就是另一种。所以如果结果是反过来的,我会立刻把这个理论扔掉。嗯,这个实验具体怎么做,取决于你想反驳的是哪些其他诠释。并没有一个实验能一口气反驳掉所有诠释,你得先把它明确下来。比如彭罗斯的那种想法:当叠加态两边的质量差超过十的负八次方千克时,波函数就会坍缩,诸如此类;或者认为波函数是在碰到有意识的观察者时坍缩的。所以假设你要用埃弗雷特去检验「波函数在碰到有意识的观察者时坍缩」这个理论,那么你要做的
便签引用
1:05:23
hits a conscious observer then what you do is you make a conscious observer which is the most convenient way to do that be to make AGI running on a quantum computer so you you then do an interference experiment we're halfway through there are two different trajectories of the of the computation take place inside the computer's memory that the the AGI has access to you're still with me when it's halfway through and hasn't yet interfered in other words it's it's it's on the inner mark sender interferometer it would be having just having bounced off the mirrors and not yet reached the final interference mirror then the the idea AGI measures which one which mirror it is at and then makes a permanent record of the form I am now contemplating I have now done the measurement and I have got a result and it is one and only one of left or right I'm not going to reveal which but I do certify that it is one of those two then the rest of the the the that part is is sealed off the part where the declaration is is sealed off and the
就是造一个有意识的观察者。最方便的办法,就是造一个跑在量子计算机上的通用人工智能(AGI)。然后你做一个干涉实验,进行到一半时,计算过程有两条不同的轨迹在计算机的存储器里发生,而这个 AGI 能访问它们。你还跟得上吗——当实验进行到一半、还没发生干涉的时候,换句话说,它正处在马赫-曾德尔干涉仪上,刚刚从反射镜上弹开、还没到达最后那面干涉镜。这时这个 AGI 去测量自己究竟在哪一面镜子那边,然后做一份永久记录,内容大致是:我现在正在思考,我已经做完了测量,我得到了一个结果,而且它是左或右当中的唯一一个。我不打算透露是哪一个,但我可以证明它就是这两者之一。然后,其余的部分——那份声明所在的部分被封存起来,剩下的部分则被施以它在思考期间
便签引用
1:06:51
rest is subjected to minus the Hamiltonian that it had during its thinking so that all the memory of which it which it which of the two things it was conscious of happened is wiped out and then the interference is performed if we hitting the conscious observer causes a collapse of the wavefunction then you will get a 50/50 split of the two outcomes and if the average interpretation is true then you'll get only one of the two that's not the challenge challenge is to find something which would be to differentiate between well this would for anybody who thinks that that the wavefunction is collapsed by a conscious observer and the this AGI is a conscious observer now if you think it isn't then from my point of view I'll let you and it decide what hammer that out amongst yourselves but if you if you insist on it being a human and on coherent quantum computations being sorry coherent quantum measurements being performed on a human brain then we're going to have to wait for some thousands of years I guess before that
那个哈密顿量的负值,这样一来,关于它当时意识到的是两者中的哪一个,全部记忆都被抹去了。然后再进行干涉。如果碰到有意识的观察者真的会导致波函数坍缩,那么两种结果会各占一半;而如果埃弗雷特诠释是对的,你就只会得到两者中的一个。这不是难点所在,难点在于找到一个能够区分二者的东西。嗯,这对任何认为波函数会被有意识的观察者弄坍缩的人都成立,而这个 AGI就是一个有意识的观察者。当然,如果你认为它不是,那在我看来,我就让你和它自己去把这事儿争个明白。但如果你坚持观察者必须是人类,坚持要在人脑上做相干的量子计算——抱歉,是相干的量子测量,那我们大概得再等上几千年,才有可能
便签引用
1:08:52
is feasible but in principle is certainly feasible Boehm is a different category in my view bohmian quantum mechanics is just ever quantum mechanics in a state of chronic denial it it is a the trouble is with the Bohm interpretation that it doesn't meet its own motivation it has this pilot wave which is actually the wave function and it has a representative particle which is goes moves along the grooves in the wave function and to to be a permian you have to systematically equivocate on the question is the pilot wave real if it's real then it has groups that are performing computations in principle conscious observer computations which affect each other and so you can't say that some of the pilot wave doesn't exist the whole of it has has to exist and therefore the the multiplicity of the averaged interpretation is just there in the in in the pilot wave interpretation if you say that it doesn't exist if then you're saying that something that doesn't exist effects something that does which simply doesn't
做到。但原则上这肯定是可行的。玻姆在我看来属于另一类:玻姆量子力学不过就是埃弗雷特量子力学,只是处于一种长期否认的状态。麻烦在于,玻姆诠释达不到它自己的初衷。它有一个导波,那其实就是波函数;它还有一个代表性粒子,沿着波函数里的沟槽运动。要当一个玻姆派,你就必须在「导波是不是真实的」这个问题上系统性地含糊其辞。如果它是真实的,那它里面就有一些部分在进行计算,原则上包括有意识观察者的计算,而且彼此之间会相互影响,所以你不能说导波的某些部分不存在——它整个都必须存在。因此,埃弗雷特诠释所说的那种多重性,在导波诠释里同样就在那儿。而如果你说它不存在,那你就是在说一个不存在的东西影响了一个存在的东西,这根本说不通。(笑声)
便签引用
1:10:16
make sense [Laughter]
便签引用
1:10:33
[Music]
(音乐)
便签引用
13问答:多宇宙、自由意志与道德的开放问题
1:10:40
I think pertaining to the discussion because David at one point I think it's a monitor and star-lord interviews you raise the possibility I think you said that the hard problem of consciousness could possibly not be resolved accepting if you start from the premise of a multi-verse well I don't about start from the premise but but obviously it's it's always possible you know you're going to run into trouble if you base your ontology on something isn't true and for example if you're going to say that freewill can't happen because only one trajectory is allowed by either Newtonian physics or by Copenhagen interpretation or whatever then you're going to conclude that free will doesn't exist from a false premise from if you replace that by the through premise that quantum theory is true then it's in this case it's not so much the quantum theory has helped you to solve the hard problem it has removed the impediment to solving this until doesn't tell you what free will is or what qualia are or whatever
我觉得这跟刚才的讨论有关。因为大卫你在某次访谈里曾经提出过一种可能性,我记得你说过,意识的难题也许无法解决,除非你从多重宇宙这个前提出发。嗯,我倒不是说「从前提出发」,不过显然它是……如果你把本体论建立在不真实的东西上,你随时都可能遇到麻烦比如说,如果你要说自由意志不可能存在,因为只有一条轨迹是被允许的——不管是牛顿物理学还是哥本哈根诠释或别的什么——那你就是在从一个错误的前提出发,得出自由意志不存在的结论如果你把那个前提换成真的前提,也就是量子理论是对的,那么在这种情况下与其说量子理论帮你解决了难问题,不如说它移除了解决它的障碍它并没有告诉你自由意志是什么,或者感受质是什么之类的,但它移除了那个决定性的
便签引用
1:11:57
but it's it's removed the knockdown argument that for example free will can't exist qualia can't be different from anything else unless electrons also have it have them and so on so you you knock out a bunch of false arguments as as generically happens if you have a fixed fixed old theory that you're unwilling to replace or criticize then you rather like sticking down a piece of a jigsaw puzzle in the wrong place and gluing it down that that would produce errors in the picture arbitrarily far away from the piece of glue down it may look fine near the piece and then you'll you'll you won't be able to construct the picture so this is why the you know this is why the pursuit of truth is useful one reason connection to the issue regarding what distinguishes a GI from AI because when you've written articles you've talked about in pottery in terms the capacity to conjecture which you've referred to sometimes as creativity you know those are widespread conventional terms is there some way to kind of
论证——比如说自由意志不可能存在、感受质不可能与别的东西不同,除非电子也拥有它们,等等。所以你就打掉了一堆错误的论证,这是很常见的情况:如果你抱着一个固定的、陈旧的理论不肯替换也不肯批评,那你就好比把拼图的一块按在错误的位置上,还用胶水粘死了——那会在离那块被粘住的碎片任意远的地方,让整幅画出错在那块碎片附近可能看着还行,但你最终没办法把整幅画拼出来。所以这就是为什么追求真理是有用的,这是其中一个理由。这和另一个问题有关,就是通用人工智能(AGI)与人工智能(AI)的区别在哪里,因为你在写文章的时候,谈到过进行猜想的能力,你有时把它称为创造力——这些都是很常见的通行说法有没有办法把这些说法说得更精确一些,在与……的关系上?[音乐] 嗯,我不认为量子
便签引用
1:13:30
sharpen those terms in relationship to [Music] well so I don't think that quantum computers will be needed to make an AGI like you know I could be wrong but I I don't think that's the kind of problem it is so therefore I think that a creative program could be made on a deterministic class computer although I must say immediately that calling that such a computer deterministic when it's an ABI somewhat misses the point because the AGI is going to be interacting with the world and the world is is not going to be deterministic because among other things it's quantum so so the the if creativity depends on some kind of randomness randomness is everywhere and that would be true whether it's classic
计算机是造出 AGI 所必需的,你知道,我可能错了,但我不认为这属于那类问题所以我认为,一个有创造力的程序是可以在确定性的经典计算机上做出来的,不过我必须马上补充一句,当它是个 AGI 的时候,把这样一台计算机叫做确定性的,多少有点没抓住重点,因为 AGI 会与世界互动,而世界不是确定性的,除了别的原因之外,它还是量子的。所以,如果创造力依赖于某种随机性,那随机性到处都是,无论是经典的还是别的,这一点都成立
便签引用
1:14:51
I don't because that sort of thing so again I don't think that probability is the right way to think about the growth knowledge and so in this knowledge earn is not the right one to think about you know but perhaps it's better to think about about being run over crossing the road or better going to a casino and and the betting and so that there there will be if you play your cards right you can arrange it so that there will always be some worlds in which you come out a multi-millionaire so now I I think the the trouble with drawing profound conclusions from this is that one thing we do know about probability is that if the average interpretation is true then when one is analyzing a situation of randomness the right analysis is the classical one I
我不这么认为,因为诸如此类的原因。所以我还是不认为概率是思考知识增长的正确方式,所以在这个问题上,知识并不是合适的思考对象。你知道,也许更好的例子是想想过马路时被车撞,或者更好的是去赌场下注,那么就会有——如果你牌打得对,你可以安排得让总有一些世界里你成了千万富翁那么现在,我认为从这里引出深刻结论的麻烦在于,关于概率我们确实知道一件事,就是如果平均诠释是对的,那么当一个人在分析一个随机性的情境时正确的分析还是那个经典的分析,我
便签引用
1:16:10
[Music]
[音乐]
便签引用
1:16:28
would try not to I would try not to let my psychological approach to a an event come into conflict between what I know is there so you know I I I might have I might have an objection to eating a suite in the form of a tarantula but then I say to myself no this isn't a tarantula this is just a piece of candy and I'm going to eat it and you say well yeah but still still what do you feel about it well what I feel about it is different from the right answer is not very interesting that's just ways in which I can be wrong are the only two computation creatively to feel more at ease with this technology all the same question I was wondering if you might do laughs yeah well um since it's the last question you see III thought allow myself the luxury of talking about things I don't know and these are three things that I don't know they could go either way so for example is the morality reducible to acknowledge well if it isn't that what on earth can morality be are there moral axioms that that are uncritical that that would be
会尽量不去——我会尽量不让我对某件事的心理反应,跟我知道的事实起冲突所以你知道,我可能会——我可能会对吃一颗做成狼蛛形状的糖果心生抗拒,但是然后我对自己说,不,这不是狼蛛,这只是一块糖,我要把它吃掉;你会说,是啊但话说回来,你对它到底是什么感觉呢?可我的感觉跟正确答案是两回事,这并不怎么有意思,那只不过是我可能出错的方式罢了,只有这两种——创造性地让人更能坦然接受这项技术;同样的问题,我在想你会不会——(笑)是啊,嗯,既然这是最后一个问题,你看,我想我可以放纵一下,谈谈那些我并不知道的事情,有三件我不知道的事,它们可能朝哪个方向都行;比如说,道德是不是可以还原为知识?如果不能,那道德究竟能是什么?是不是存在某些道德公理是不容批判的?那样就会是权威主义的;反过来说,如果道德是——这个问题就不会
便签引用
1:18:15
authoritarian so on the other hand if morality so and that problem wouldn't arise if it was if its technology because we already know how to do to frame epistemology without acquiring foundation without quiring Authority you know thanks to Paul we know that but then I asked myself what if the laws of physics were different like the ones I mentioned with the Greek gods and the malevolence suppose there were malevolence and and the laws of physics really did have it in for us and so on would that change morality could could we say like at the moment we say that that if something is property of physics it doesn't have a moral value pro or con but in such a unit can we Oh am i right in thinking that those kinds of laws of physics are amoral or does that not make sense and I don't know so that's that's the the kind of thing there and and that and I'm also wondering in that question whether there is a connection between that and the question that the implement who's on our left are about whether consciousness and
出现,如果它属于那类知识的话,因为我们已经知道该怎么去构建认识论,不需要基础,也不需要诉诸权威,你知道,多亏了波普尔我们才明白这一点;但接着我又问自己:如果物理定律不一样呢?就像我提到的希腊诸神那种情形,还有恶意——假设真有恶意存在,物理定律真的就是要跟我们过不去,诸如此类那会改变道德吗?我们能不能说——就像此刻我们会说,如果某件事是物理的属性它就无所谓道德上的好坏;可在那样一个宇宙里我们还能这么说吗?哦,我这样想对不对:那类物理定律是无关道德的?还是说这话根本讲不通?我不知道,所以这就是那一类的问题;还有一点我也在想,就是这个问题跟坐在我们左边那位提的问题之间有没有关联——意识、自由意志、感受质、道德价值这些东西,是不是
便签引用
1:19:42
free will and and qualia and and moral value and all that stuff are all come together necessarily or can they be separate or can they be made separate perhaps absolutely up to now in the universe they've always come together but we could artificially make them separate and energy I again I I think that can't be so but I can't give you a watertight argument on it isn't so but we we have to wait in in which case we have to wait till somebody comes up with a viable theory of these things yeah I'm always a bit a bit perturbed when when people have strong feelings about things like free will the moral value you know is it moral to kill animals eat animals are animals conscious and all those things well they do not know what consciousness is none of us do [Laughter]
必然捆绑在一起,还是它们可以分开,或者能被人为地分开?到目前为止在这个宇宙里它们确实总是一起出现但我们或许能人为地把它们拆开——我还是觉得那不可能,可我拿不出一个滴水不漏的论证来说明不可能,我们只能等;那样的话我们就得等到有人提出一套关于这些东西的可行理论;是啊,每当有人对这类事情抱有强烈的看法时,我总有点不安比如自由意志、道德价值,你知道,杀动物、吃动物道不道德,动物有没有意识,诸如此类的问题;可他们并不知道意识是什么,我们谁都不知道 [笑声]
便签引用
视频总结 · 一句话概括与核心要点

一句话概括

David Deutsch 主张:人类文明将永远面临存在性危险,唯一的防线是尽快创造解释性知识;而给 AGI 硬编码价值观、搞研究禁令与预防原则,正是历史上摧毁过所有文明的那一类"压制知识创造"的知识,是他所知的最严重存在性威胁。

核心要点

  • AGI 在物理上必然可能。 主持人引述 Deutsch 1985 年论文对丘奇-图灵猜想的扩展:通用量子计算机能以有限手段模拟任何物理系统。由此推出通用性原理——物理定律允许物体做的任何事,原则上都可由通用计算机程序在足够时间和内存下任意精细地模拟,人脑也不例外。
  • "每个危险都有非零概率,所以迟早灭亡"是谬误。 概率不是一成不变的:知识增长时,某些概率会下降。文明的任务是让这个无穷级数收敛到可忽略的值。反过来,任何"到达乌托邦后安全永久有保障"的想法,都需要说明人类凭什么是每个物种终将灭绝这一规律的唯一例外——而唯一的候选答案是无穷的解释性知识流。
  • 存在性危险按"缺失的知识"分四类。 第一类是超级火山、小行星等"笨石头和流体",缺的是火山学、流体力学、核动力航天器等有限知识;第二类是流行病等自身也在"创造知识"(通过进化)的敌人,需要新的解释性知识;第三类是尚未知晓的危险(如 1900 年无人知道吸烟致命,等知道时已死数亿人),只能靠尽快积累深层通用知识来应对;第四类是"不可知"的——即尚未创造出的解释性知识的内容,其唯一来源是人和 AGI。
  • 财富的构造子理论定义与通用构造器。 Deutsch 定义财富为"一个人能够实现的所有变换的集合",例如"给定准备时间后能无害偏转的所有小行星"。有了通用构造器,一切重复性劳动被编程取代,财富变成程序库,构造器可自我复制并从小行星带采矿。它只是家电,不会思考、不会"想要",除非装入 AGI 程序——而每个人本身就是装了 AGI 程序的通用构造器。
  • 让人不危险的唯一方法是让他们自由。 这是花了几千年才获得的反直觉知识:自由主义价值观、个人权利、开放社会。在这样的社会里,绝大多数人无论硬件特征如何都是正派的;总会有个别想把一切变成回形针的人,但多数人会用部分创造力去阻止,只要知识创造速度领先于坏人,就会赢。
  • 对齐问题的"硬编码"方案是文明自杀。 把价值观固化写入 AGI 等于给它戴上镣铐、残害其知识创造能力以奴役它。结果只有两种:要么它缺少"G"根本不是 AGI,要么它会找到改进你道德观的方法并反叛。正确的对齐是像教育孩子一样教育 AGI 成为社会成员;对 AGI 的恐惧与对叛逆青少年的恐惧本质相同,而自伯里克利时代的雅典起人类就明白:如果你是对的,无需强迫;而我们并非事事都对,价值观必须能随之改进。
  • 禁令只针对好人。 研究暂停或监管只在特定情形下(有好的解释证明它不比新知识更危险时)才合理。恐怖组织用可靠方法培养出自杀式炸弹手心态的 AGI 时,如果我们已剥夺全体正派人得到"被培养为正派人的 AGI"的保护,那就是灾难配方。核动力、反核运动的例子说明:以规避短期辐射风险为由放弃长期未来的赌博,已成为对抗气候变化的巨大拖累。
  • 没有文明因创造太多知识而毁灭,唯一例外是"压制知识的知识"。 历史上所有被外部或内部摧毁的文明,本可以通过更快创造知识获救;毁灭它们的只有一种知识:如何维持现状——更高效的宗教裁判所、更警惕的暴民、更严格的预防原则。
  • 概率不适用于建模无知。 下一次灭绝级小行星已经在路上了,不存在"它的概率";除非有特定解释模型预测某事可近似为随机过程,否则给任意数字贴上概率和效用标签再声称权威,只是自欺。大型强子对撞机的例子:它安全或不安全的理论只有真假之分,真实概率是 0 或 1,只是未知,问题必须靠解释而非博弈论决定;"用它比不用更危险"的论证可以套在任何基础研究上,故是坏解释。
  • "Moloch"式系统性恶意理论全部为假。 所有这类理论都把参与者当作机械执行局部最优的密码符号,从不准确。这类困境会作为普通问题短暂出现,然后人们互相指责、创造性思考、发明让所有人愿意出资建大坝的方案。用"靴子踩在人脸上"取代说服与个人权利也不成立:政府也是人,没有哲人王式的特殊知识通道;知识不存在,公园就建不成。

结论与值得注意的细节

  • 知识的最新定义:有因果力的信息。 这是从构造子理论自然导出的定义——若某段信息是让构造器做某事所必需的,它就是知识,包括道德知识和数学知识(数学家是物理对象)。基因中的"笨知识"是非解释性的,作用域有限、无法跨越某些障碍;解释性知识无需可行的中间形态序列,可跨越任何障碍,而且不存在比它更强大的信息处理或改造世界的方式。
  • 问题不会终结。 知识增长应被看作"问题被更好的问题取代"而非理论被替换。印度穷孩子能上网获取全部人类知识、代价是气候变化稍难应对——这是比以前好得多的问题。"终极问题"的概念自相矛盾;若真的用完了问题,那本身不也是个问题吗?未有互联网的人是"成功带来的问题",不构成对全世界的指控。
  • AGI 与人在计算上完全相同,只在速度和内存上有差异。 AGI 是程序不是硬件,同样的硬件人类也能用。可能的未来是人类成为赛博格、AGI 反而"灭绝",因为赛博格拥有 AGI 的一切加上某种额外的东西。只要过程是道德的、是知识增长的结果,任何一种结局都值得欢迎——正如多种族社会的问题:人与人之间没有根本差异。
  • 意识、感质、自由意志、创造力大概率不可分离。 我们不知道意识是什么,也不知道如何造 AGI。但如果解释性创造力能在没有感质的情况下演化出来,为什么演化会造出感质这套庞大机器?它们必然相连——但若真能人为分离,会带来严重道德问题,因为启蒙道德与认识论紧密相连:道德分歧应能通过理性讨论达成一致。
  • 区分 Everett 多世界与"意识导致坍缩"的实验设计。 在量子计算机上运行一个 AGI 作为有意识观察者,让它在干涉仪中途测量并封存"我已观测到且只观测到左或右之一"的声明,随后对其余部分施加负哈密顿量抹去记忆再做干涉。若意识致坍缩,得 50/50;若 Everett 正确,只得一种结果。对人脑做相干量子测量则要等数千年。玻姆力学被他称为"处于慢性否认状态的 Everett 量子力学":导波若是实在的,其中的"槽"就在做计算、甚至意识计算,多重性已内在其中;若说它不存在,则是不存在之物影响存在之物,讲不通。
  • 量子理论没有解决自由意志的"难问题",只是移除了障碍。 基于假前提(牛顿或哥本哈根只允许一条轨迹)推出自由意志不存在,就像把拼图块粘错位置,错误会波及任意远处。他认为造 AGI 不需要量子计算机,但称 AGI 所在的经典计算机为"确定性"没抓住重点:它要与本质上量子的、非确定的世界互动,随机性无处不在。
  • 多世界与赌场类比。 若 Everett 正确,分析随机情境的正确方法仍是经典方法;总会有某些世界你成为百万富翁,但不宜从中得出深刻结论。心理感受与已知事实冲突时(吃一颗狼蛛形状的糖),他会选择不让感受介入——"我感受到什么与正确答案有何不同,并不有趣,那只是我出错的方式。"
  • 三个他坦承不知道的问题。 道德是否可还原为知识(若不能,道德是什么?无法批判的道德公理会是威权的;若能,则波普尔式无基础、无权威的认识论框架已可套用);若物理定律真的"与我们为敌"(奥林匹斯诸神式),道德是否会改变、这样的定律是否仍是无道德的;意识、自由意志、感质和道德价值是否必然共现。他对人们在不知道意识是什么的情况下就对"动物有无意识、能否吃动物"持强烈立场感到不安。
  • 主持人的框架。 开场把 Deutsch 的工作比作终结"哥白尼延迟":1985 年论文之于量子力学如同伽利略的望远镜之于日心说,让"闭嘴计算"策略持续约 70 年后终于让理论"抓住实在"。主持人称人工智能对未来的最大威胁也许是"没人去实现它"。讲座中途有几次录音中断,Deutsch 调侃"正要回答系统性恶意的问题时系统就把我关了"。
核心句型 · 10
1. nothing can possibly stand between us and X except Y
“Nothing can possibly stand between us and any of those infinitely many existential dangers except the right explanatory knowledge”
强调唯一防线的句式。用 nothing … except 制造排他感,between us and 把威胁具体化。写论证时可用来点出「只有一种办法」。
2. wouldn't it be rather ironic if … precisely by …
“Wouldn't it be rather ironic if that version of the principle and the movement were about ports the great environmental catastrophe since the last ice age precisely by advocating selfish short-term benefit”
反问式讽刺。wouldn't it be … if 提出假设,precisely by 点出反讽机制。适合指出某政策与其初衷背道而驰的情形。
3. This is not X, it's Y
“This is not the nightmare of the matrix in which our world is only a simulation it's the vision of enlightenment”
先否定读者可能的误解,再给出正解。两个名词短语并列对照,节奏紧凑。适合在解释概念时预先排除常见联想。
4. the bigger and faster …, the more … we'll need
“The bigger and faster the approaching asteroid or moon or planet black hole the more of a special kind of knowledge we'll need”
双重比较级表示正相关。前半可并列多个形容词,后半 the more of a … 修饰不可数名词。仿写:the larger the model, the more data it needs.
5. X is at once A and yet in a sense B
“The fourth category is at once even more dangerous and yet in a sense less worrisome”
at once … and yet 并置两个看似矛盾的属性,in a sense 缓冲第二项。用于引出悖论式论点,后面通常跟 because 解释。
6. again if that had been …, whom could we sue?
“Again if that had been ecstasy existential danger whom could we sue”
虚拟语气加修辞反问。if that had been 是对过去的假设,whom could we sue 用「找谁算账」表达「无处追责」。适合强调后果不可逆。
7. not one of them … in fact except for one kind …
“Not one of them destroyed itself by creating too much knowledge in fact except for one kind of knowledge”
先用 not one of them 做全称否定,再用 except for 引出唯一例外,使例外格外醒目。写论证时可用来把矛头指向单一因素。
8. it's not so much that X … it has Y
“It's not so much the quantum theory has helped you to solve the hard problem it has removed the impediment to solving”
not so much A as B 的口语变体,纠正对因果关系的过度表述。适合精确区分「直接解决」与「移除障碍」这类程度差异。
9. I can't imagine X, let alone Y
“I can't imagine what physics theories are going to be invented in the next ten years let alone in the next ten billion years”
let alone 表递进否定:连较容易的 X 都做不到,更不用说 Y。X 与 Y 须同类且 Y 更极端。
10. to cut a long story short, all of them assume that …
“To cut long story short all of them assume that the people concerned are not created”
口语过渡语,用于略过细节直奔结论。原文省略了冠词 a,书面用完整形式。后接 all of them assume 一次性概括对手理论的共同前提。
词汇精讲 · 138 · 按出现顺序
spyglass /ˈspaɪɡlæs/ n. 0:09
小型望远镜
heliocentrism /ˌhiːlioʊˈsɛntrɪzəm/ n. 0:09
日心说
dismissed /dɪsˈmɪst/ v. 0:09
斥为不足取(dismiss … as …)
gain purchase on phr. 1:32
在…上获得着力点、立足点
conjecture /kənˈdʒɛktʃər/ n. 1:32
猜想
finite means phr. 1:32
有限的手段
emulated /ˈɛmjəleɪtɪd/ v. 2:54
仿真、模拟
in arbitrarily fine detail phr. 2:54
以任意精细的程度
what's at stake phr. 2:54
利害攸关之处
startling /ˈstɑːrtlɪŋ/ adj. 4:21
令人震惊的
take our success in advancing knowledge for granted phr. 4:21
把…视为理所当然(take … for granted)
merit /ˈmɛrɪt/ v. 6:00
值得
fallacy /ˈfæləsi/ n. 6:00
谬误
get sucked into phr. 6:00
被卷入、不知不觉陷入
immutable /ɪˈmjuːtəbəl/ adj. 6:00
不可改变的
negligible /ˈnɛɡlɪdʒəbəl/ adj. 6:00
可忽略不计的
criterion /kraɪˈtɪriən/ n. 6:00
标准、准则
sole /soʊl/ adj. 7:32
唯一的
gamma-ray bursts n. 7:32
伽马射线暴
hostile /ˈhɑːstəl/ adj. 7:32
敌对的
extraterrestrials /ˌɛkstrətəˈrɛstriəlz/ n. 7:32
外星生物
rogue /roʊɡ/ adj. 7:32
失控的、不受约束的
slave revolt n. 7:32
奴隶起义
Volcanology /ˌvɑːlkəˈnɑːlədʒi/ n. 9:19
火山学
logistics /ləˈdʒɪstɪks/ n. 9:19
后勤、物流组织
evacuations /ɪˌvækjuˈeɪʃənz/ n. 9:19
疏散
self-evident /ˌsɛlfˈɛvɪdənt/ adj. 9:19
不言而喻的
aftermath /ˈæftərmæθ/ n. 9:19
(事件的)后果、余波
drag on phr. 10:53
对…的拖累(a drag on)
hallmark /ˈhɔːlmɑːrk/ n. 10:53
典型特征、标志
precautionary principle n. 10:53
预防原则
strand /strænd/ n. 10:53
(思想、运动的)一股、一条线索
at the expense of phr. 10:53
以…为代价
digress /daɪˈɡrɛs/ v. 10:53
离题
Devils in the detail phr. 10:53
魔鬼藏在细节里(the devil is in the detail)
deflect /dɪˈflɛkt/ v. 12:22
使偏转
envisaging /ɪnˈvɪzɪdʒɪŋ/ v. 12:22
设想、展望
proportional /prəˈpɔːrʃənəl/ adj. 12:22
成比例的
from scratch phr. 13:57
从零开始
asteroid belt n. 13:57
小行星带
sit back phr. 13:57
袖手旁观、坐享其成
appliance /əˈplaɪəns/ n. 13:57
器具、设备
topically /ˈtɑːpɪkli/ adv. 15:30
应景地、切合时事地
nothing short of pathetic phr. 15:30
简直可怜(nothing short of 表强调)
nucleic acid n. 15:30
核酸
epidemiology /ˌɛpɪˌdiːmiˈɑːlədʒi/ n. 15:30
流行病学
pathogens /ˈpæθədʒənz/ n. 15:30
病原体
paleontologists /ˌpeɪliənˈtɑːlədʒɪsts/ n. 15:30
古生物学家
widely appreciated phr. 17:06
被广泛认识到
customised /ˈkʌstəmaɪzd/ adj. 18:45
定制的
neutronium /nuːˈtroʊniəm/ n. 18:45
中子星物质
paradoxical /ˌpærəˈdɑːksɪkəl/ adj. 18:45
悖论式的
entities /ˈɛntətiz/ n. 18:45
实体
counterintuitive /ˌkaʊntərɪnˈtuːɪtɪv/ adj. 20:22
反直觉的
decent /ˈdiːsənt/ adj. 20:22
正派的
take it into their head phr. 20:22
突发奇想要做某事
stay ahead of phr. 20:22
保持领先于
urn /ɜːrn/ n. 21:53
瓮(概率论中的抽球瓮)
baseless assertions phr. 21:53
毫无根据的断言
forego /fɔːrˈɡoʊ/ v. 23:19
放弃
moratorium /ˌmɔːrəˈtɔːriəm/ n. 23:19
暂停令、延缓
unleashes /ʌnˈliːʃɪz/ v. 24:45
释放、放出
genocidal /ˌdʒɛnəˈsaɪdəl/ adj. 24:45
种族灭绝的
strip /strɪp/ v. 24:45
剥夺(strip sb. of sth.)
from without phr. 24:45
从外部
status quo /ˌsteɪtəs ˈkwoʊ/ n. 24:45
现状
Inquisition /ˌɪnkwɪˈzɪʃən/ n. 24:45
宗教裁判所
vigilant /ˈvɪdʒələnt/ adj. 24:45
警觉的
hard coding phr. 26:14
硬编码、写死
shackling /ˈʃækəlɪŋ/ v. 26:14
给…上镣铐
crippling /ˈkrɪplɪŋ/ v. 26:14
使残废、严重削弱
enslave /ɪnˈsleɪv/ v. 26:14
奴役
disobedient /ˌdɪsəˈbiːdiənt/ adj. 28:16
不服从的
terminal nodes n. 30:32
末端节点
foregone conclusion phr. 30:32
必然的结局
biosphere /ˈbaɪoʊsfɪr/ n. 30:32
生物圈
mitigate /ˈmɪtɪɡeɪt/ v. 32:04
缓解
political will n. 32:04
政治意愿
imaginary /ɪˈmædʒənɛri/ n. 32:04
(社会理论)想象图景、观念体系
paramount /ˈpærəmaʊnt/ adj. 34:34
至高无上的
inadvertently /ˌɪnədˈvɜːrtəntli/ adv. 34:34
无意间
plausible /ˈplɔːzəbəl/ adj. 37:05
说得通的、貌似合理的
by a factor of millions phr. 37:05
以数百万倍
cyborgs /ˈsaɪbɔːrɡz/ n. 38:37
半机械人
die out phr. 38:37
消亡、灭绝
qualitative /ˈkwɑːlɪteɪtɪv/ adj. 39:56
质的、性质上的
qualia /ˈkwɑːliə/ n. 39:56
感受质(主观体验的质感)
epistemology /ɪˌpɪstəˈmɑːlədʒi/ n. 39:56
认识论
malevolence /məˈlɛvələns/ n. 41:57
恶意
hinder /ˈhɪndər/ v. 41:57
阻碍
indictment /ɪnˈdaɪtmənt/ n. 43:48
控诉、谴责
deprived of phr. 43:48
被剥夺
conceivable /kənˈsiːvəbəl/ adj. 45:19
可以想见的
penal policy n. 45:19
刑罚政策
perverted /pərˈvɜːrtɪd/ adj. 45:19
扭曲的、变态的
impediment /ɪmˈpɛdɪmənt/ n. 46:49
障碍
too big for our boots phr. 49:08
得意忘形、自以为了不起
hubris /ˈhjuːbrɪs/ n. 49:08
狂妄自大
have it in for us phr. 49:08
跟我们过不去、存心作对
causal power n. 51:30
因果力
abstractions /æbˈstrækʃənz/ n. 51:30
抽象事物
viable /ˈvaɪəbəl/ adj. 53:00
可存活的、可行的
impervious /ɪmˈpɜːrviəs/ adj. 54:49
不受影响的(impervious to)
to cut long story short phr. 55:06
长话短说
ciphers /ˈsaɪfərz/ n. 56:55
无足轻重的人、符号化的角色
get behind phr. 56:55
支持、力挺
persuasion /pərˈsweɪʒən/ n. 58:18
说服
divine right n. 58:18
君权神授
prophesy /ˈprɑːfəsaɪ/ v. 59:51
预言
paraphrase /ˈpærəfreɪz/ n. 59:51
转述、改写
let alone phr. 59:51
更不用说
cessation /sɛˈseɪʃən/ n. 59:51
终止
encompass /ɪnˈkʌmpəs/ v. 1:01:51
涵盖
superposition /ˌsuːpərpəˈzɪʃən/ n. 1:03:54
叠加态
refute /rɪˈfjuːt/ v. 1:03:54
反驳、证伪
interferometer /ˌɪntərfəˈrɑːmɪtər/ n. 1:05:23
干涉仪
certify /ˈsɜːrtɪfaɪ/ v. 1:05:23
证明、担保
sealed off phr. 1:05:23
封存、隔离
Hamiltonian /ˌhæmɪlˈtoʊniən/ n. 1:06:51
哈密顿量
coherent /koʊˈhɪrənt/ adj. 1:06:51
相干的
hammer that out phr. 1:06:51
反复争论直至解决
chronic denial phr. 1:08:52
长期否认
equivocate /ɪˈkwɪvəkeɪt/ v. 1:08:52
含糊其辞
multiplicity /ˌmʌltɪˈplɪsəti/ n. 1:08:52
多重性
pertaining to phr. 1:10:40
与…有关
ontology /ɑːnˈtɑːlədʒi/ n. 1:10:40
本体论
premise /ˈprɛmɪs/ n. 1:10:40
前提
knockdown argument n. 1:11:57
决定性的、一击致命的论证
jigsaw puzzle n. 1:11:57
拼图
deterministic /dɪˌtɜːrmɪˈnɪstɪk/ adj. 1:13:30
确定性的
play your cards right phr. 1:14:51
策略得当、把牌打对
tarantula /təˈræntʃələ/ n. 1:16:28
狼蛛
reducible to phr. 1:16:28
可还原为
axioms /ˈæksiəmz/ n. 1:16:28
公理
authoritarian /əˌθɔːrɪˈtɛriən/ adj. 1:18:15
权威主义的
amoral /eɪˈmɔːrəl/ adj. 1:18:15
无关道德的
watertight /ˈwɔːtərtaɪt/ adj. 1:19:42
无懈可击的
perturbed /pərˈtɜːrbd/ adj. 1:19:42
不安的
精读便签
下载便签 手机:长按图片也可保存
← 上一期 · NO.141Melanie Mitchell - Abstraction and Analogy: The Keys to Robust Artificial Intelligence 下一期 · NO.143 →CHM Live | The Great Chatbot Debate: Do LLMs Really Understand?
苏菲周报 · THE WEEKLY 每周一封,
追问一个大问题。
苏菲拉底的每周来信,写这一周在追问的问题和看到的回应。
苏菲拉底
ASK THE BIG QUESTIONS · THINK DEEPLY · SEE THE WORLD DIFFERENTLY
苏菲拉底微信公众号二维码 微信公众号
© 2026 苏菲拉底 · 内容仅供学习 [email protected]