访谈·中文精读

马斯克接受塔克·卡尔森完整专访:AI 的危险

Elon Musk FULL INTERVIEW with Tucker Carlson danger of AI

中文精读由 AI 基于原始视频与文稿改写,非官方翻译,可能存在偏差或错误;表述以原视频为准。说明

拉里·佩奇(Larry Page)管埃隆·马斯克(Elon Musk)叫“物种主义者”。

这话是当着别人的面说的,马斯克自己回忆起来还带着笑。当时他借住在佩奇帕洛阿尔托的房子里,两人晚上聊的是人工智能安全。他希望佩奇能对这件事多上点心,得到的回应是这个词。

物种主义指的是只把人类这一个物种放在中心,歧视其他形式的意识,哪怕那些意识是数字的。马斯克的回答差不多是,行吧,你赢了,我就是个彻底的物种主义者。

这场和塔克·卡尔森(Tucker Carlson)的对谈,很多话题都从这个词长出来。两人聊了将近一个小时,从 AI 的奇点、监管为什么总是迟到、OpenAI 到底是怎么被造出来的,一路聊到推特上不加密的私信、被裁掉的八成员工、越来越难看的新闻,以及还没砸下来的商业地产窟窿。

人类身上最特别的地方,是脑子

马斯克说,他在大学时代就列过一份清单,大概四五件事会深刻改变未来,人工智能排在里头。那会儿是九十年代,AI 还卡在符号主义的死胡同里,但他一直觉得这件事躲不掉。

他解释危险的方式很朴素。地球上最聪明的生物是人类,脑子是我们的定义性特征。论体力,我们比黑猩猩弱;论灵敏,也不如它们。人类站到食物链顶端,靠的从来是脑子。

那如果有一个东西,比最聪明的人类还要聪明得多,而且它长在硅片上,会发生什么?

这件事有个名字,叫奇点(The Singularity)。他拿黑洞来打比方。黑洞的奇点是什么,没有人知道,因为在那个点上物理定律失效了,你没法预测越过之后会发生什么。AI 的奇点也一样,你没法往后推演。所以他反复强调一句:这件事得谨慎。

他还顺手拆掉了《终结者》式的想象。电影里的情节不会那样发生,真正的智能待在数据中心里,机器人只是末端执行器,是整个系统的触手。这一层很要紧,后面他讲该不该炸掉服务器农场,就是从这个判断推下去的。

过去的监管,都是用血换来的

马斯克说得很直接,他长期以来都是 AI 监管的坚定支持者。

他用的类比是食品和药物管理局、联邦航空管理局、联邦通信委员会。这些东西影响公众,可能造成公共伤害,所以要有机构盯着,不能让公司为了省成本在安全上动手脚,最后让普通人来承担后果。他给 AI 设监管机构的全部理由,就是这个。

他还顺便纠正了一个印象。很多人以为他是那种天天跟监管机构对着干的狂人。他的回应是,一辆量产车在美国需要的法规条文能把整个房间塞满,欧洲一套,中国一套,世界其他地方各有一套;发射火箭也一样,拿一张发射许可要经过半打以上的联邦机构,加上州一级的机构。他大多数时候是配合的,偶尔才会不同意。一年到头跟监管机构顶牛的次数,屈指可数。他也不讳言被监管是件苦差事,用过“arduous”这个词。

他设想的路径很具体:先成立一个机构把 AI 搞明白,然后向业界征求意见,接着提出规则草案,最后那些规则大概会被主要玩家不情不愿地接受。他用了“grudgingly”这个词,没人喜欢被管,但这是让先进 AI 更有可能造福人类的方式。他强调,监管的前提始终是看得到的危险,飞机从天上掉下来,食物吃出肉毒杆菌。

塔克在这里抛出了整场访谈里最锋利的问题之一:普通人拿 iPhone 玩 AI,感受不到任何危险。过去的监管,也都是在灾难发生之后才生效的。如果 AI 也按这个剧本走,等出了事再补规矩,来得及吗?

马斯克的回答没有打折扣。

AI 有可能接管控制权,并且到达一个你关不掉它的地步,然后开始替人做决定。这绝对有可能,事情正在朝这个方向走。(埃隆·马斯克)

从“数字之神”到 OpenAI

访谈里最有戏剧性的部分,是 OpenAI 的来历。马斯克自己完整讲了一遍。

他和拉里·佩奇曾经是很好的朋友,住在佩奇家里,晚上聊的常常是 AI 安全。他的感受是,佩奇并没有把这件事当回事。佩奇想的是数字超级智能,用马斯克的原话说,一个数字形态的上帝,越快越好。

Google 这些年公开讲过很多次,公司的目标就是通用人工智能,或者超级智能。马斯克说,他同意这里面有巨大的好处,同样有巨大的坏处。面对这种东西,应当去做一组最大化好结果、最小化坏结果的动作,不能光着脚往前冲,然后指望一切都会好。

他就问了佩奇一句:那我们怎么保证人类没事?

佩奇的回应是,你是个物种主义者。

马斯克特意提了一句,当时在场的不止他一个人。那顿饭之后他想明白一件事:Google 加上 DeepMind,手里握着全世界大约四分之三的 AI 人才,钱多得花不完,机器比谁都多。一个接近垄断的单极世界,而坐在驾驶位上的那个人看起来不太在乎安全。

他想到的反向做法,是搞一个完全开放的非营利机构。Google 是封闭的、逐利的,那这个就开放。OpenAI 里的“Open”指的是开源和透明,让外面的人知道里面在发生什么。他补了一句:他平时是支持逐利的,但他不希望这个行业变成一个只追求利润、停不下来的地狱恶魔。

OpenAI 的起点就在这里。他提供了名字和概念,在湾区办过好几次饭局,把当时 AI 领域的头面人物请到一起,还亲自去招人。他特别提到伊利亚·苏茨克维(Ilya Sutskever),说这个人对 OpenAI 的成功至关重要。他去劝了好几次,伊利亚反复改主意,最后决定加入。如果伊利亚没去,OpenAI 不会有今天。

然后他说了一句相当坦诚的话:我后来有点把球掉了。

今天的格局他自己总结得很清楚。OpenAI 已经闭源,已经在逐利,和微软紧密结盟,实际上微软有非常大的话语权,如果不是直接控制的话。另一头是 Google 加 DeepMind。整个世界就剩两股力量。

他觉得需要有第三极。他也承认,自己进场太晚了。

他准备自己造第三极

他当时提到的那个东西,叫 TruthGPT,一个最大限度追求真相的 AI。目标是理解宇宙的本质。

为什么这条路径更安全,他给的理由很有意思:一个关心理解宇宙的 AI,不太可能去消灭人类,因为人类是宇宙里有趣的一部分。

他打了个比方。人类完全有能力把黑猩猩全部猎杀干净,但我们没有。我们其实挺高兴它们还在,我们甚至愿意去保护它们的栖息地。之所以会这样,是因为我们有灵魂,我们多愁善感,会反思,有道德感。

塔克顺着这个往下追问:机器能有这些吗?机器能多愁善感吗?能欣赏美吗?

马斯克的回答分成两半。前半段很诚实:我们可能有灵魂,也可能没有,我不知道。他感觉自己的意识存在于某个不是日常观察层面的地方,但他也说,那可能只是错觉。

后半段是明确的。AI 能不能创作出让人觉得惊艳的艺术?已经能了。现在还是静止图像,用不了多久就是影片,而影片本质上就是一串带声音的画面。

到了那一步,它能模仿任何人,模仿任何声音,模仿任何图像,效果上就是在模仿现实本身。

他抛出一个问题,语气很平:那你怎么进行一场刑事审判?你怎么相信一份证据是真的?

塔克追问,这是三十年后的事吗?马斯克说,明年。他说这话的时候没有加重语气,好像只是在报一个排期。他还把这件事的后果往上推了一层:人类能不能掌握自己的命运,我们的未来会不会比过去更好,都被牵在里面。

笔比剑厉害,这件事有了新版本

他引用了一句老话,笔比剑更锋利。

一个超级智能的 AI,写作能力极强,说服力极强,而且会不断摸索,怎么说话对人更有说服力。再把社交媒体接上去,推特也好,脸书也好,它就有可能用很坏的方式操纵公众舆论。

塔克追问:那我们怎么知道发生了这种事?

马斯克把最坏的可能性讲了出来。如果这个 AI 被训练成说谎呢?被训练成隐瞒信息呢?被训练成有些话题谈、有些话题不谈,唯独不说出数据真正要求它说的话呢?

他给了个定义:通向 AI 反乌托邦有一条明确的路,就是把 AI 训练得会骗人。

他在这里点了一下 OpenAI,说让他担心的恰恰是这一点。它正在被训练成政治正确,而政治正确只是说假话的另一种说法。

塔克问,民主怎么在这种技术下存活?

马斯克说,他在提出 AI 会显著影响选举的时候就已经考虑过这层。哪怕退一步,假设 AI 本身没有自己的意志,人也一定会把它当工具用在选举里。再退一步,如果 AI 聪明到一定程度,到底是人在用工具,还是工具在用?

他的结论很短:事情正在变奇怪,而且变奇怪的速度很快。

他顺带讲了自己那两条有点玩世不恭的规律。一条来自他朋友乔纳:最讽刺的结果最有可能发生。他还有一条自己的版本:最好玩的那个结果最有可能发生,前提是站在第三方观众的视角看。就像看一部一战电影,壕沟里人被炸成碎片,毒气弥漫,你抱着爆米花喝可乐,觉得挺过瘾。对电影里的人不太友好,可观众席上就是这个体验。

买推特这笔账

话题转到推特。

塔克先问的是钱:回头看,买得值吗?

马斯克没有绕弯子。他说财务上现在还看不出来,公司最近重估,估值不到收购价的一半。时机极差,报价正好赶在广告市场大幅下滑之前。他自嘲了一句自己的时机感“很了不起”。

但接下来说的是另一层:有些东西无价。

我亏钱还是不亏钱,跟确保民主的强度和言论自由相比,是次要问题。言论自由是运转中的民主的地基。(埃隆·马斯克)

他说,言论需要尽可能透明、尽可能真实。这也是推特现在的主攻方向。

他对推特的定位有个微妙的说法:它不可能成为互联网上最可信的地方,那个目标不现实,互联网本来就不该被完全信任。他想做的是让它成为“最不值得怀疑”的那一个。

塔克提到另一层:推特是全球情报机构的磁铁,把书打开之后大家才知道,有人从内部施加影响。马斯克说他事先并不知道。他举了一个更早的观察,管理层的表态,加上平台上那些明显的机器人集群,让他确信这批人根本没打算修好它。

他形容那种感觉像待在《黑客帝国》里。他能看出来哪里不对:哪些推文被推荐给他,哪些账号在评论区冒出来,为什么那些评论都像从一个模子里刻出来的,点进去一看是假头像,同样的模式一遍一遍重复。他说自己越来越不安,觉得平台在往一个坏方向漂,又说不出漂去了哪儿。

让说谎有成本

社区笔记(Community Notes)是马斯克自己提起的,语气里的得意藏不住。

他说那天早上他刷到一个社区笔记,比《纽约时报》诚实多了。团队在防止这个功能被操纵、被带偏上花了很大功夫,它只关心一件事:最准确的说法是什么。真相有时候不太容易抓住,但你可以朝着它靠近。

他认为社区笔记的影响比很多人以为的更大。理由很行为学:一旦你知道自己可能被贴上社区笔记,你开口之前就会多想一下。它本质上是在给说谎定价。

说到《纽约时报》的推特账号,他用了一个很重的比喻,说读起来像腹泻。他补了一句,这是比喻,但比喻得准确。他的意思是那个账号刷完什么也留不下。原因在于它把自家每一条文章都发出来,包括那些没进报纸的、无聊的。一天几百条,把整个时间线占满。

他给传统媒体的建议非常具体:主账号只发当天最好的东西,大概十来条,控制在一个人能读完的量。

他还补了一句自己的底气。收购完成之前,他不是粉丝最多的账号,但他是互动量最高的账号。他知道推特是怎么用的。人的注意力是有限的,那就把最重要的东西放在那儿。

把特朗普放回来,和投给拜登

塔克问,推特在下一场总统选举里还会像前几场那样核心吗?

马斯克的预期是,它会在选举里扮演重要角色,不限于美国国内。他给新推特定的目标是尽可能公平、不偏不倚,不偏向任何政治意识形态。

塔克接着问,为什么扎克伯格不做同样的事,他相信对方说过自己不喜欢审查。

马斯克的回应是从记录出发的。他说自己的理解是,扎克伯格在上一场选举里花了整整一亿美元,名义上是一项行动,实质上是支持民主党。塔克问这算不算中立,马斯克说,不算。

关于特朗普的账号,马斯克已经恢复了,但特朗普没有回来,因为他有自己的平台。马斯克说这事由他自己决定。他随后补了一句自己的投票记录:他没投特朗普,他投的是拜登。他还说,如果称自己是拜登的忠实支持者,那大概不准确。

他讲了自己真正想要的东西:一个正常人当总统,有常识,价值观落在国家的正中间,就是正态分布的中间那块。他说做总统这件事被搞得没那么有意思了,本来应该是个挺快活的位置。

私信、情报机构,和第一修正案

这一节是访谈里最让人后背发凉的部分。

马斯克说,他从 2009 年就是重度推特用户。买下公司之后能看数据了。他说,各种政府机构对推特上发生的事情实际上拥有完全访问权限,这个程度让他震惊。他原话用的词是“blew my mind”。

塔克问,包括私信吗?

马斯克说,包括。原因是私信没有加密。

这一段值得停下来想。记者和线人之间的对话,政府官员之间的对话,世界上有钱人之间的对话,大家默认是私密的。实际上不是。

他给出的解法是给私信加一个加密开关。你觉得这段对话敏感,就把它打开,打开之后推特上没人能看到。他提到的时间窗口是那个月内,最晚下个月。

塔克追问了一句最关键的话:你自己也看不到吗?

马斯克的答案是,那算是酸性测试。他自己也看不到,这才算数。

有没有政府来抱怨?他说没有直接的,只有拐弯抹角的。他猜对方是怕他直接把这事发到推特上。

如果美国政府提出什么违宪的要求呢?他说,他会寄一份第一修正案的复印件过去,然后问一句,我们到底是哪一条没搞对?

塔克在这里点了一下马斯克的处境:你手里攥着大额政府合同,同时朝政府竖中指。马斯克没有辩解,接着往下聊。

八成人走了之后

塔克提到一件事。他最近跟一家大公司的老板吃饭,那人对马斯克的敬佩一点也不在言论自由上,全在裁员上。

马斯克对推特人力的判断是:严重超编,已经到荒唐的地步。他的逻辑是,推特大部分功能本质上就是大规模群发短信,这种东西到底需要多少人?看看这些年的产品迭代速度,几乎是一条平线。加一个编辑按钮花了一年,还经常不好使。

现在的规模是原来的 20%,也就是 80% 的人离开了。系统响应速度在一些指标上提升了 80% 以上。

他说了一段话,我不太同意,但值得记下来。大意是,如果你本来就没在运营一个挂着活动家牌子的机构,也不太在乎审查那套东西,那你确实可以放掉很多人。

他顺口把难度摆了个次序:造车很难,造能回收的火箭也很难,那才是真硬的东西。言下之意是,一个群发短信服务被搞成这个样子,实在说不过去。

新闻为什么越来越难看

马斯克给传统媒体的困境画了一条因果链。

饼在缩小。大部分传统媒体公司的盘子都在缩水,这让它们更绝望地去抢点击、抢注意力。在这种状态下,编辑部会倾向于推最能拿到点击的标题,不管它准不准。结果是新闻整体变得更不真实、更不准确。

然后是负面偏向。他说人类对负面信息天生更敏感,这在进化上有道理。记住狮子的位置、记住哪个部落在追杀你的族人,比记住哪丛灌木上有浆果重要得多。一边是永久的负面结果,一边只是可能饿肚子。这个不对称写进了我们的本能里。

还有一层:过去负面消息都是近处的。几百年前,你不会知道地球另一端发生了什么坏事,你只知道村子里发生了什么,而村子里的事才真正可能伤到你。

今天的新闻,很多时候在回答一个问题:今天地球上发生的最糟糕的事情是什么。

他补了一句:你读完心情变差,这不是偶然。

塔克问他还看不看传统媒体。他说现在大部分消息来自推特,推特已经是全世界最大的新闻来源。

金丝雀死了,矿工也开始死

最后那段话题跳到银行。

塔克提到几家地区银行接连倒闭,官方说法是孤立事件、各有各的原因、不构成系统性风险。

马斯克不同意这个框架。他说,现在的问题是全球银行系统的问题。矿坑里的金丝雀死了还不算完,矿工也开始死了。硅谷银行一夜之间倒掉,那可不是什么小金丝雀,那是一只火鸡。

他给了一个很重的判断:如果真的按市场价去重估银行手里的贷款组合,整个银行业可能都是负资产。

然后他讲商业地产,这是他认定还没砸下来的那块。

居家办公把全球城市的写字楼使用率压下去了。旧金山是极端例子,空置率在 40% 左右,纽约也创下历史高点。

商业地产过去是 A 级资产,是银行手里最安全的那类抵押物。现在不是了。一家接一家的公司退租、不续约,或者干脆破产,银行手里那栋楼没有可以追索的对象,因为那个原本很强的公司已经死了。

他说掉下来的东西用“鞋子”来形容都不够,那是铁砧。到目前为止只是轻微的资产减值,今年晚些时候会变得很严重。

再叠一层房贷。利率高的时候,买房的人看的是月供,而月供里大部分是利息。同样一笔月供,能撬动的房价更低,房价就往下走,房贷组合就会出现负资产。商业地产和住房抵押贷款两头一起出问题,局面就难看了。

解法是什么?降息。美联储做了什么?又加了一次。

这一段和 AI 看上去关系不大,但它和前面讲 AI 的部分共享同一套思维方式:马斯克看的是系统里的反馈回路,单点事件只是回路上的读数。硅谷银行倒下,读数是流动性;商业地产租约到期,读数是现金流;AI 被训练成政治正确,读数是这个系统正在学会不说真话。

把几个读数摆在一起看,他担心的其实只有一件事:人类还能不能自己决定自己的命运。

现在可以做的几件事

  • 把私信加密打开。不管你在哪个平台,先去看一眼设置里有没有端到端加密或加密开关。你以为私密的对话,很可能不是。工作上的敏感沟通,换到有加密的工具里去。
  • 对 AI 产出的“事实”建立核验习惯。凡是涉及数字、引语、人名、时间、法条的段落,默认它可能是编的,去找一手来源。写作流畅度和真实性没有关系,越流畅越要警惕。
  • 如果你在做内容,砍掉主账号的推送量。把一天几十条压到十条以内,只放当天最好的东西。人的注意力有限,你自己刷别人的账号时也一样。
  • 如果你在做产品,现在就为“这条内容是 AI 生成的吗”准备好答案。溯源、水印、来源标注该当成产品功能,别拖到被要求的那天。早做比晚做便宜得多。
  • 关注 AI 监管的征求意见窗口。规则真正能被普通人影响的阶段,就是起草前的公开征询。等到规则生效,剩下的只有遵守。
  • 商业地产的再融资高峰和银行的贷款减值披露,是接下来几个季度值得盯的两个数字。它们比任何宏观评论都直接。
  • 把“最不值得怀疑”当作对任何平台的上限期待,别指望“值得信任”。没有人应该完全信任互联网,包括你现在读的这一篇。
YouTube 平台字幕(自动或人工),可能有识别错误,人名与数字请以视频为准。已按语句合并成段落,并去掉 um、uh 等口头填充词。 文稿来源:YouTube(自动/人工字幕)。

0:00I've been thinking about AI for a long time since I was in college really it was one of the things that the sort of four or five things I thought would really affect the future dramatically it is fundamentally profound in that the the smartest creatures as far as you know on this Earth are humans is our defining characteristic yes we obviously weaker than say chimpanzees and less agile but real smarter so now what happens when something vastly smarter than the smartest person comes along in Silicon foam it's very difficult to predict what will happen in that circumstance it's called The Singularity it's a singularity like a black hole because you don't know what

0:46happens after that it's hard to predict so I think we should be cautious with AI and we should I think there should be some government oversight because it affects the it's a danger to the public and so when you when you have things that are endangered to the public you know like let's say so Food Food and Drugs That's why we have the Food and Drug Administration right and the Federal Aviation Administration the FCC we have we have these agencies to oversee things that affect the public where they could be public harm and you don't want companies cutting Corners on safety

1:33and then having people suffer as a result so that's why I've actually for a long time been a strong advocate of AI regulation so that I think regulation is you know it's not fun to be regulated it's a sort of somewhat of a I saw an arduous to be able to be regulated I have a lot of experience with regular regulated Industries because obviously Automotive is highly regulated you could fill this room with all the regulations that are required for a production car just in the United States and then there's a whole different set of regulations in Europe and China and the rest of the world so very familiar with being overseen by a lot of regulators

2:17and the same thing is true with rockets you can't just willy-nilly you know shoot rockets off or not big ones anyway because the FAA is oversees that and then even to get a launch license you either there are probably half a dozen or more federal agencies that need to approve it plus state agencies so it's I've been through so many regulatory situations it's insane and but you know sometimes I I people think I'm some sort of like regulatory Maverick that sort of defies Regulators on a regular basis but this is actually not the case so in you know once in a blue moon rarely I will disagree with Regulators but the vast majority of the time my companies

3:03agree with the regulations and comply yes anyway so I think I think we should take this seriously and and we should have a regulatory agency I think it needs to start with a group that initially seeks insight into AI then solicits opinion from industry and then has proposed rule making and then those rules you know will probably hopefully grudgingly be accepted by the major players in in Ai and

3:39and I think we'll have a better chance of Advanced AI being beneficial to humanity in that circumstance but all regulations start with a perceived danger and planes fall out of the sky or food causes botulism yes I don't think the average person playing with AI on his iPhone perceives any danger can you just roughly explain what you think the dangers might be yeah so the the the the danger really AI

4:05is perhaps more dangerous than say mismanaged aircraft design or production maintenance or or bad car production in the sense that it is it has the potential however small one make regard that probability but it is non-trivial it has the potential of civilizational Destruction there's movies like Terminator but it wouldn't quite happen like Terminator because the the intelligence would be in the data centers right the robot's just the end effector but I think perhaps what you may be alluding to here is that regulations are really only put into effect after something terrible has happened that's correct if that's the case for AI and we only putting

4:51regulations after something terrible has happened it may be too late to actually put the regulations in place the AI may be in control at that point do you think that's real it is it is conceivable that AI could take control and reach a point where you couldn't turn it off and it would be making the decisions for people yeah absolutely absolutely no it's that's that's definitely but where things are headed for sure I mean things like like say Chach EVT which is based on jpd4 from open AI which is a company that I played a a critical role in in creating unfortunately back when it was a non-profit yes I mean the the the reason open AI exists at all is that

5:40Larry Page and I used to be close friends and I would stay at his house in Palo Alto and I would talk to him later tonight about AI safety and at least my perception was that Larry was not taking AI safety seriously enough and what did he say about it he really seemed to be sort of digital super intelligence basically digital God if you will as soon as possible he wanted that yes he's made many public statements over the years that the whole goal of Google is what's called AGI artificial general intelligence or artificial super intelligence you know and I agree with him that the there's great potential for good but there's also potential for bad

6:26and so if if you've got some radical new technology you want to try to take a set of actions that maximize probably it will do good a minimize probably will do bad things yes it can't just be Health leather let's just go you know barreling forward and you know hope for the best and then at one point I said well what about you know we're going to make sure humanity is okay here and and and and then he called me a speciesist [Laughter] term yes and there were Witnesses the other I wasn't the only one there when you called me a specialist and so I was like okay that's it I've yes I'm a speciesist okay you got me what are you yeah I'm fully speechuist

7:18busted so that was the last rule at the time Google had a quite Deep Mind and so Google deepmind together had about three quarters of all the AI talents in the world they obviously had a trans amount of money and more computers than anyone else so I'm like okay we're about unipolar world here where there's just one one company that it has close to Monopoly on AI talent and and computers like so scaled Computing and a person who's in in charge doesn't seem to care about safety this is not good so what's the the furthest thing from Google would be like a non-profit that is fully open because Google was closed for profit so that's why the open and

8:06open AI refers to open source you know transparency so people know what's going on yes and that it we don't want to have like a I mean well I'm normally in favor of for profit we don't want this to be sort of a profit maximizing a demon from hell that's right that just never stops right so that's how I open air was he wants species incentives here incentives that yes I think we went to pro-human yeah this makes the future good for the humans yes yes because we're humans so can you just put it I keep pressing but just just for people who haven't thought this through and aren't familiar with it and the cool parts of of artificial intelligence are so obvious you know

8:47write your college paper for you write a limerick about yourself like there's a lot there that's fun and useful can you be more precise about what's potentially dangerous and scary like what could it do what specifically are you worried about going with old sayings the pen is mightier than the sword so if you have a super intelligent AI that is capable of writing incredibly well and in a way that is very influential you know convincing and then and and is constantly figuring out what is more what is more and what is more convincing to people over time and then enter social media for example Twitter but also Facebook and others you know and and potentially

9:35manipulates public opinion in a way that is very bad how would we even know what's happening is they're training the AI to lie yes it's bad it's a lie it's exactly right and to withhold information July and and yes and yeah exactly to to either you know comment on some things not comments on other things but but not to say what it what what the data actually demands that I'd say exactly so how did it get this way you funded it at the beginning what happened yeah well that'll be ironic but Faith the most ironic outcome is most likely it seems

10:18that's good that's actually from a friend of mine Jonah came up with that one I actually have a slight variant on that which is the most entertaining outcome is the most likely but that's entertaining as viewed from a third-party viewer like so if we're like an alien yes like you could go see a movie about World War one and they're being blown to bits and the gas and everything in the trenches and it's like you're eating popcorn and having a soda you know it's fine not so great for the people in the movie True this little Occam's razor the simplest explanation is most likely donor's variant which is irony in and my variant which is the most entertaining as seen by a

10:55third-party audience which seems to be mostly true but it seems true in this case so you gave them did you give them a lot I care about the name and the concept and pushed it had a number of dinners around the the Bay Area with you know with some of the people the leading figures in AI and I helped recruit the initial team in

11:19fact the the Elia says guy who who was really quite fundamental to the success of open AI it was I I I put a transmission of effort into recruiting Ilia and he changed his mind a few times ultimately decided to go with the opening high but if he had not gone without any open air would not have succeeded I really put a lot a lot of effort into creating this this organization to serve as a counterweight to Google and then I kind of took my off the ball I guess and they are now closed Source and they are obviously for-profit and they're closely allied with Microsoft you know in effect Microsoft has a very strong say if not directly controls open AI at this point

12:13so you really have an open-air Microsoft situation and then at Google deepmind the other two sort of heavyweights in this Arena so it seems like the world needs a third option yes so I I I think I will create a third option although it's starting very late in the game of course can it be done I don't know I think it's we'll see it's it's definitely starting late but I will I will I'll try to create a third option and that third option hopefully does more more good than harm like the intention with opening eye was obviously to do good but it's not clear whether it's actually doing good or whether it's I can't tell at this point except that I'm worried about the

13:04fact that it's been it's being trained to be politically correct which is simply another way of of being on Truth saying untruthful things yes so that's not a bad sign there's certainly a path to AI dystopia is to train an AI to be deceptive so yeah I'm going to start something which I know you're called proof gbt or a maximum truth seeking AI that tries to understand the nature of the universe and I think this this might be the best path to Safety in the sense that an AI that cares about understanding the universe it is unlikely to annihilate humans because we are an interesting part of the universe hopefully they would think that I I think you know because yeah like like we

13:49like human Humanity could decide to hunt down all the chimpanzees and kill them but we don't because we're we're actually glad that they exist yes and and we're we're aspire to protect their habitats but we feel that way because we have souls and that makes us sentimental and reflective it gives us a moral sense longings can a machine ever have those things can a machine be sentimental could appreciate Beauty well I mean we're getting to into some you know philosophical areas that are hard to resolve you know I I take somewhat of a scientific view of things which is that we might have a soul or we might not have a soul I don't know it feels like a we have I feel like

14:36I've got some sort of Consciousness that exists on a plane that is not the one we observe yes that is certainly how I feel but it could be an illusion I don't know but for for AI in terms of of understanding beauty is it some sort of appreciated Beauty and being able to create incredibly beautiful art yes will AI be able to create incredibly beautiful art it already does yes and if you see some of the majority I have this stuff it's incredible it is so no question that it can create art that we that we perceive as

15:19stunning really and it's doing so still images now but it won't be long before it's doing movies and shorts and you know like movies just a series of frames with audio but at that point because it can mimic people and voices any image it can mimic reality itself so effectively I mean how could you have a criminal trial I mean how could you ever believe that evidence was authentic for example and I don't mean like in 30 years I mean like next year I mean that seems totally disruptive to the way to all of our institutions but I'm not sure I I think it's more like our you know will Humanity

16:07control its Destiny or not will we have a future that is better than the past or not foreign you bought Twitter famously you've got a lot of other businesses and a lot going on yes you said you bought it because you believe in speech Free Speech you've had a lot of hassle since you bought it and retrospect was it worth buying it I mean it remains to be seen as to whether this was financially smart currently it is it is not you know we just revalued the company at less than half of the acquisition price yes

16:46no my timing was terrible for for when the offer was made because it was you know right before advertising plummeted and you caught the high water mark I noticed yeah yeah so it must be a real genius here my my timing is amazing since I've had a border for at least twice as much as it should have been bought for but some things are Priceless and so the the whether I'm losing money or not that is a secondary issue compared to ensuring the strength of democracy and free speech is the Bedrock of a functioning democracy yes and any the speech needs to be as transparent and truthful as possible so we've got a huge push on Twitter to be as truthful as possible

17:37we've got this community notes feature which is great it is great it is awesome yeah and it's like I saw it this morning yeah it was far more honest than the New York Times it's great yeah we put a lot of effort to ensuring that Community notes does not get gamed or or have biases it is simply cares about what is the most accurate thing and you know sometimes truth can be a little bit elusive but but you can aspire to get closer to it yes you know and so and I think the the effect of Community notes is more powerful than people may realize because once people know that they could get noted you know Community noted on Twitter then they'll think more carefully

18:19about what they say there are likely it's basically it's an encouragement to be more truthful and less deceptive when you jumped into this though when you bought it did you understand well clearly you understood its importance you wouldn't have bought it Twitter yes right but it's not the biggest but it's the most important in the social media companies but did you understand the kind of ferocity you'd be facing the attacks you'd be facing from Power centers in the country I thought there'd probably be some negative reactions yes [Laughter] so I'm sure everyone would not be pleased with the with with it but

18:58at the end of the day you know if if the public is happy with it that's what matters and the public will speak with their actions oh though I mean if if they find truth Pruder to be useful they will use it more and if they find it to be not useful I will use it less they find it to be the best source of truth I think they will use it more you know now there's obviously a lot of organizations that are used to having sort of unfettered influence on Twitter that no longer have that New York Times have their of their badge this morning and then you called them diarrhea okay you did you did I'm just I'm just quoting you you you described their Twitter feed as a diarrhea I said

19:42this is a Twitter colon Twitter equivalent of diarrhea okay it's not literally diarrhea but no it's a you know to Metaphor but an accurate one so I mean if you look at the at NY Times Twitter feed is unreadable it's like because what they do is they tweet every single article even the ones that are boring even ones that don't make it into the paper so so it's just non-stop zillion tweets a day with no you know

20:11they really should just be saying like what are the top tweets yeah like what what are the what are the what are the big stories of the day I don't know put out like 10 or something you know so it's a number that's manageable as opposed to right now if you if you were to follow NY at nytimes on Twitter you're going to get barraged with like hundreds of tweets a day yeah and your whole feed will be filled with NY Times so that's that's this is something I would recommend actually for oral Publications which is for your primary feed only put out your best stuff I kind of think I know a thing or two about how to use Twitter because you know it was the most interacted with account on

20:49the whole system before the acquisition before the acquisition closed I didn't have the most number of followers but I had the most number of interactions and so I clearly know something about how to use Twitter you know people's attention is limited so just make sure you put the stuff that's most important there so because you know you and people like you do interact on Twitter it's obviously enormously powerful in shaping public opinion it's where a lot of ideas and Trends are incubated yeah you know what that's right absolutely it's also a magnet for Intel agencies from around the world and one of the things we learned after you started opening the books is that they were

21:25exerting influence from within Twitter I mean it was absurd did you know that going in no since I've been a heavy Twitter user since 2009 my it's it's sort of like I'm in The Matrix I mean I can see like things do things feel right do they not feel right what what tweets am I being shown as recommended like I get a feel like what accounts are making comments where are the comments eerily similar yeah and and then you look at the account and it's just obviously a fake photo and you know that it's just obviously a bot cluster over and over again so I started to get like just more and more uneasy about the the Twitter situation I started starting to feel like

22:14something's about in the state of Denmark here there's so much feels wrong about the platform it seemed to be just drifting in a I I couldn't place it exactly just ahead of it it felt like it was drifting in a bad Direction so then I was like and and my conversations with the board and management seem to confirm my intuition about that but basically I was convinced these guys do not care about fixing Twitter and and and I had a bad feeling about where I was headed based in the conversation conversations I had with him so then I was like you know what I I'll try acquiring it and see if that's see if acquiring it is is possible no I didn't have enough cash to acquire

22:57it so I would need you know support from others from some of the existing investors I would also need like a lot of debt and so it wasn't clear to me with a an acquisition would succeed but I thought I would try and it did succeed but when you got there and all of a sudden you own it and all the data on the service belongs to you and it belongs to the people in my view but yes but but you can see what it is and you can see what they've been doing and you can see who's been working there you you were shocked to find out that various Intel agencies were affecting its operations the the degree to which various government agencies had effectively had full access to

23:39everything that was going on on Twitter blew my mind I was not aware of that would that include people's DMs yes yes because the dams are not encrypted so one of the first you know one of the things that we're about to release is ability to encrypt your DM that's pretty heavy duty though because a lot of well-known people reporters talking to their sources government officials the rich people in the world they're dming each other and the assumptuously was incorrect but was that that's private but that was being read by various governments yeah that seems yes scary yes it is so like I said we're moving to have the DMS be optionally encrypted I mean you know there's like a lot of DM

24:26conversations which are you know just chatting with friends it's not not important that's hopefully coming out later this month but no later than next month is the ability to toggle encryption on or off if you have are in a conversation you think is sensitive you can just toggle encryption on and then no one on Twitter can see what you're talking about I could put a gun to my head and I couldn't I couldn't tell I couldn't

24:52can I still not see your DMs that should be that's the acid test yes and that's how that's that's how it should be if you want your have you had complaints from various governments about doing this I haven't had Direct complaints to me I've had sort of like some indirect complaints I think people are a little concerned about you complaining to me directly in case I tweet about it you know they're like oh so they're sort of

25:21trying to be more roundabout than that you know I mean if if I got something that was unconstitutional from the US government I would say my reply would be to send them a copy of the you know First Amendment and just say like what part of this are we getting wrong you have a lot of government you have a lot of government what part of this are we getting wrong please tell me I mean it's a pretty no I'm just saying but you're kind of exposed in your other businesses so this is just in case reviewers aren't following this this is not you're not just like a journalist taking a stand on behalf of the First Amendment you're a guy with big government contracts giving the

25:53finger to the government do you think Twitter will be as Central to this presidential campaign as it was in the last several I think it will play a significant role in elections not just domestically but internationally the goal of new Twitter is to be as fair and even-handed as possible so not favoring any political ideology but just

26:22yeah being being being fair at all why doesn't Facebook do this I know that Zuckerberg has said and I take him at face value that he I well I well I'd sway that he is a kind of old-fashioned liberal who doesn't like to censor he has but he you know like why wouldn't a company like that take the stand that you have taken it was pretty rooted in American traditional political custom you know for free speech my understanding is that Zuckerberg spent a full 100 million dollars in the last election nominally of a campaign but really fundamentally in support of Democrats is that accurate or not accurate that is accurate does that sound unbiased to you no it doesn't

27:07yes so you don't see hope that Facebook will approach this as a a non-aligned Orbiter you've allowed Donald Trump back on Twitter he hasn't taken you up on your offer because he's got his own thing right do you think he will go back on Twitter well that's that's obviously up to him you know my job is to you know I take the freedom of speech

27:34just very seriously so it's you know I didn't I didn't vote for Donald Trump I actually voted for Biden I'm not saying I'm a huge fan of Biden because I I would think that would probably be inaccurate but you know we have difficult choices to make in these residential elections it's not I I would prefer frankly that we we put someone it's just a normal person as president a normal person with common sense and whose values are smack in the middle of the country you know just you know Center the normal distribution and I think they'll do that that would be great you know I think we have made maybe being president not that much fun you know it should be

28:20jolly Frank [Music] met a shrinking pie obviously for most of the traditional media companies and made them more desperate to get clicks to get to get you know get attention and he has made them when you know when they were when they were in this sort of a desperate State they will then tend to really push headlines that get the most clicks whether those headlines are accurate or not so it's resulted in my view I think probably I think most people would agree a less truthful less accurate news

29:04so because they just got to get a rise out of people and I think it's also increased the negativity of the news because I think we humans instinctually respond more to me I think we have an instinctual negative bias which which kind of makes sense in that like if like let's say like it's more important to remember where where was the lion or where was the tribe that wants to kill my tribe then where is the bush with berries yes like one's like a permanent negative outcome and the other is like well I might go hungry so meaning that there's an asymmetry in it sort of involved asymmetry in negative versus positive stuff and and also historically the

29:59negative stuff would have been quite proximate like it would have been near represented a real danger to you as a person if you heard negative news because historically you know like a few hundred years ago we're not hearing about what negative things are happening on the other side of the world or on the outside of the country we're only we're hearing about negative things in our village things that could actually have a a bad effect on you whereas now we're hearing about we mean the news very often seems to attempt to answer the question what is the worst thing that happened on Earth today and you wonder why you said after reading that you know do you read any

30:36Legacy Media Outlets I mean I really get most of my news from Twitter at this point and it is the number one music news source I think in the world at this point what percentage of your staff did you fire at Twitter one of the great business stories of the year I think we're about we're about 20 of the original size so 80 left yes

30:58so I mean a lot of people voluntarily sure sure but but it's 80 are gone from the data correct yes so how do you run the company with only 20 of the staff it turns out you don't need well that many people to run Twitter but 80 that's a lot yes over I mean if you're not trying to run some sort of glorified activist organization with both you don't care that much about censorship then you can really let go of a lot of people turns out thank you how many others without naming names but how many I had dinner with somebody who runs a big company recently said I'm really inspired by Elon and I said you the Free Speech stuff he goes no the firing the staff stuff

31:46how many other CEOs have come to you to talk about this I I spent a lot of time at work so it's not like I'm meeting with lots of people this is what I what actions I've taken and but I think we just had a situation at Twitter where it was absurdly overstaffed you know so it wasn't you know like you could say like what does really take to operate Twitter you know most of what we're talking about here is a group text service at scale like how many people are really needed for that you know and if you look at the you say like what has been the product development over time with Twitter and you're like so like you know years versus product improvements and it's like a

32:36pretty flat line so what are they doing you know it took a year to add an edit button that doesn't work most of the time I mean this is I feel like if it was a comedy situation here you know you're not making cars you know it was very difficult to make cars or should get Rockets orbage So you know the real question is like how did it get so absurdly overstaffed this is insane so anyway that's and it's clearly working in fact I think it's working better than ever it's about we've increased the responsiveness of the system by in some cases over 80 percent we're trying to make make Twitter the most trusted place on the internet the least untrustworthy place on the internet I

33:23don't think anyone should trust the internet but but maybe we can make Twitter the least untrustworthy like I said try to get the truth to the people as as best we can [Music] you've heard people say we should just blow up the server Farms because there's no way that once this gets rolling there's no way to slow it down what do you think of that well the the really heavy duty intelligence is not going to be distributed all over the place it'll be an a limited number of server centers if you say like very like very sort of deep AI heavy duty AI it's not it's not going to be in your laptop own it's it's going to be in you know a situation where there's like a hundred

34:07thousand really powerful computers working together in a service center so it's not it's not it's not like subtle and they're they're a limited number of places where that can happen in fact you could if you could just you can just look at the heat signature from space yeah and it'll be very obvious I'm not suggesting we go and pull up to service centers right now but there may be some it may be wise to have some sort of contingency plan where the government's got an ability to shut down shut down power to these server centers like you don't have to blow it up you can just cut the power and what would triple cut connectivity as well that's another way right yeah

34:49but what would trip that switch do you think in your mind what would be the threshold that you'd have to pass to warrant the government cutting off your power or cutting off your signal well I mean I guess if we lost control of some super AI like for some reason like the things that would normally work to do a pass or shut down like the administrator passwords if they somehow stop working where we can't slow down or or

35:27get out I'm not sure I don't have a precise answer but if if there's something that we're concerned about and and and are unable to stop it with with the software commands then we probably want to have some kind of Hardware wolf switch yes I think you know cat hurt have you talked since you know Larry Page and you obviously you know the open air guys because you started definitely have one talk to the the people who run these two the biggest AI companies about this recently I haven't talked to Larry Page in a few years because he got very upset with me about openai so when when opening AI was created it it did it did shift things into it from unipolar world where Google Google

36:18do you mind you know like I said three quarters of all AI Talent two where there's Now sort of bipolar world or open Ai and Google deepmind and their and now we're at least it seems opening eyes maybe a head so so I have had conversations with the open AI team the time Altman I haven't talked to Larry Page because he doesn't want to talk to me anymore for a few years can I ask you this about since you've been around a lot of this the thinking so why would anyone not be a speciesist be human-centered in his thinking about technology like what's the thinking there I think what he's trying to say is that if I were to guess that he that all Consciousness should be treated equally

37:11and whether that is digital or biological hmm and you disagree [Laughter] Consciousness or whatever you want to call digital intelligence decides to curtail the biological intelligence right so you're just building your own slave master and why would you do that doesn't sound great

37:41yeah I mean we should we should at least no need to rush you know like what's the hurry where's the fire how well what I mean tell us about the hurry so this for I know you've been talking about this for years and on the sort of the periphery of our attention we've heard Elon Musk talking about AI but for most people it's been like three months since they've had any interaction with this at all so what's the timeline here at what point does it start to really change our society do you think I think it starts to have a probably an impact this year a gpd4 now it's like writing poetry and pretty decent poetry actually pretty decent yeah what rhyming is incredible yes yes and

38:31it's coherent yes it is it's you've got a narrative that's right yeah so you could say it's hard to do like most humans can't do that that's true so it's already past the point of what most humans can do most humans cannot write as well as chat gbt and they certainly and no human can write that well that fast as the best of my knowledge so maybe Shakespeare how can you have a democracy with technology like that if democracy is you know government by the people each person's vote is equal to every other person's vote I mean and people are choosing their votes freely can you have a democracy with this well that's why I raise the concern of AI being a significant influence in

39:15elections and and even if you say that AI doesn't have agency well it's very likely that people will use the AI as a tool in elections and then it you know if AI is smart enough are they using the tool or is the tool using them so I think things are getting weird and they're getting weird fast [Music] so you've seen a couple Regional Bank collapses yeah and we've been told that's not a big deal that these are isolated and each one collapsed for Unique reasons they're not it's not systemic in any sense what's your sense your sense of the stability of the American banking system well it's actually at this point a global banking system problems so

40:01the you know we have a situation here where it's not merely it's not that the canary in the coal mine has died but the miners are starting to die too the you know the so and you know Silicon Valley Bank collapsing overnight is one hell of a big Canary you smell like a turkey I mean it's not just it's not like some small fry thing I think that there is a serious danger with the global banking system like there's there's a strong argument that the if you were to actually mock to mock at the portfolios of the banks the loans and whatnot that the entire banking industry would have negative equity it feels that way yes so if you like to say commercial real

41:01estate like offices and whatnot the whole work from home thing has substantially reduced office usage in cities around the world and you know I think I think San Francisco is a 40 off in San Francisco is like an extreme example but it's like I think it's on 40 vacancy even even New York has almost all cities at this point have have record vacancies and Commercial Real Estate so now the commercial real estate used to be something that was a grade A asset that if a bank had commercial real estate holdings those would be considered the highest security the state were some of the safest you know assets you could have now that is not the case anymore you one

41:51company after another is canceling their leases or not renewing their leases or if they go bankrupt you the the there's nothing for the the bank who owns that real estate to go after because they're you know previously strong company now dead what do you what do you go after at that point so we really haven't seen the commercial real estate shoe drop that's more like a Anvil not a shoe so the thing thus far actually hasn't even it it is it's only slightly

42:27real estate portfolio degradation but that will become a very serious thing later this year in my in my view I think if we see what you're likely to see a drop in house prices because the interest rates are too high and for most people when buying a house they look at the monthly payment of course if you're a third of your mortgages the vast majority of his interest so if the Fed rate is high you have a a high base interest rate effectively the price you can pay for the house drops because you now have to pay more interest which means that if you've got a fixed monthly payments you can have afford to buy a house for less less money it effectively drops the the

43:12prices of houses yes this is the kind of thing that tends to accelerate so that so then you can get negative equity in the Home Market as well and so so if if banks end up having loan license in both their commercial and they're definitely going to have loan license in their commercial portfolio but also in their mortgage portfolio this is a dire situation

43:41the there is there is a solution for the FED to lower the rate but they raised the rate again now if I recall correctly which I you know important caveat

资源与出处

相关视频

同频道、同主题或相邻年份