AI DEBATE: “Most People Have No Idea What’s Coming” · 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。

AI DEBATE: “Most People Have No Idea What’s Coming”

节目发布 2026-08-17 · Chris Williamson
扎扎克·卡斯 莉芙·波瑞 埃埃里克
EDITED TRANSCRIPT · 依据现场录音编译整理,可划线生成便签
编者按:本文整理自播客《现代智慧》(Modern Wisdom)的一场 AI 辩论,主持人克里斯邀请了三位嘉宾:前职业扑克手、科学传播者莉芙·波瑞(Liv Boeree),制作 AI 历史与前景科普视频的创作者埃里克(Eric),以及节目常驻嘉宾、自称并非加速主义者的评论者扎克(Zach)。谈话从「2040 年一个寻常的星期二」开始,途经就业的政治保护、屏幕成瘾、Hugging Face 攻击、放缓前沿公开信、中国的扩散路线、莫洛克陷阱与权力集中,终于凯恩斯和反宿命论。本文依据现场录音编译整理,仅删去口语枝节、寒暄、广告与重复,论证、例证与转折均按原样保留。

2040 年的寻常星期二

主持人:开场我想试一件事。我让 ChatGPT 说出它真正相信的 2040 年世界图景,要求它诚实,把 AI 高速发展会带来什么、包括各种风险都讲出来。我开了深度研究模式,它得跑一会儿。趁它在跑,我想听你们各位对 2040 年的真实预测。不要最好情形,也不要噩梦情形。你们觉得 2040 年普通人的一个寻常星期二会是什么样?

埃里克:我觉得我们还在,人类还在地球上。我觉得权力会比今天集中得多。我说不准是集中在几家公司还是几个政府手里,但大概会有一种感觉:决定地球上发生什么的,主要是少数几个人的意志,而不是某种分散的、广泛民主的东西。

扎克:不管好坏,我预计到 2040 年,更多人的寻常星期二会更趋同,而且跟今天的寻常星期二没有太大差别。我们会把底线抬高一些。但我这么预计,是因为物理世界的变化远比我们想的慢,围绕机器人之类的监管也会比我们预期的拖得久得多。政治保护会是拖慢进步的那股力量。大多数人的星期二,无论好坏,很可能跟本周的星期二感觉差不多。

波瑞:你会讨厌我这么说,但我有点想拒绝回答这个问题。我对 2040 年的想象是一种可怕的、近乎两极的不确定,我没法把它压缩成一个答案。你一问,我脑子里第一个冒出来的是一幅混乱荒凉的图景,也许只剩下少数人类。我不是说这是大概率,但它是一片切片。另一片是那种美妙的乌托邦,我和朋友管它叫「技术田园主义」(technopastoralism):一个技术高度赋能的世界,你想用就用,但其实大家都选择回归本源,种种菜,跟朋友围着篝火。你想给篝火换个颜色也行,但多数人更愿意自己动手,种自己的手工食物之类。然后还有一片:我们全都活在最先冲线的那家 AI 公司的拇指底下,只能指望它仁慈,给我们一点自由。再一片是我们全被封锁在家,像加强版的新冠时代,谁都出不了门,什么都够不着,因为危险技术已经无处不在。这真是我对这个问题唯一的回答。

主持人:听起来你并没有拒绝这个问题,你对会发生什么想得挺清楚的。

波瑞:我不知道。问题就在这,我脑子里这几种可能的概率分布相当平均。

灭绝概率之争

主持人:那我们现在就谈这个。你真的认为 2040 年人类不在了有相当的可能?

波瑞:是的。

主持人:你给多少?你的「毁灭概率」(P(doom))是多少?

波瑞:先说一句,我们得先给「毁灭」下定义。毁灭概率之争的一大问题,就是大家用「毁灭」指各种各样不同的东西。

主持人:完全同意。那你来定义。

波瑞:照你的说法,是指人类文明的大多数,或者说很大一部分人类死掉了?

主持人:我留给你定。有人把毁灭定义成彻底灭绝,也有人定义成某种极端糟糕的事件,比如文明崩溃。

波瑞:我不想给任何确定的答案。我们正处在一个说什么、信什么都会强烈影响未来的时代,我不喜欢给它标一个确定的概率,因为这些数字本身在剧烈变动。我的看法是,这个风险足够重大,我们必须竭尽所能去把它压下去。我个人认为任何超过 1% 的数字都值得认真对待,超过 5% 就肯定值得。而且,看你问谁,很多顶尖 AI 公司的首席执行官和研究者给出的数字在 2% 到 50% 之间。这些全都落在同一个篮子里:我们面对的是一个岌岌可危的未来。所以我不喜欢毁灭概率这个框架,它过于简化了,两边都拿它当棍子使。说 90% 的那些人拿它当棍子,觉得它可笑的人也拿它当棍子。它没什么帮助。我想我们都同意,也许你不同意,未来远非确定,如果我们不把事情做对,它可能非常危险。

扎克:我在你眼里像个宿命论者吗?

波瑞:你说的宿命论是什么意思?我肯定不认为未来在任何意义上是确定的。

扎克:没有。你刚才说「我想我们都同意」的时候看了我一眼,好像我可能不同意。

波瑞:我只是不想预设。我们才刚开始聊,我不想先入为主。

主持人:埃里克,你怎么看?

埃里克:我完全同意,当我们谈的是灭绝或者人类大规模失权这种灾难性后果时,任何超过 1% 的概率都远超我们应该容忍的范围。而且我认为这是社会的显示性偏好。我们花了相当一部分 GDP 防止核灾难,防止下一场大流行病的钱恐怕应该花得更多,至少在 2021 年大家都是赞成的。AI 的奇怪之处在于它同时有巨大的上行和下行空间。反核武器很容易,反病毒也很容易,没几个人主张「病毒权利」。可 AI 是一整个打包在一起的东西:同一个技术,既可以把人从工作义务里解放出来,过上一个美妙的星期二,去追求各自意义上的快乐;也可以让一个俄罗斯寡头拥有一支只忠于他的劳动力大军,不需要任何其他人的同意,就把自己在全球 GDP 里的份额越做越大。

主持人:后匮乏、极权主义和毁灭,全泡在同一口锅里,同一个潜在的未来里。

波瑞:还有技术田园主义。

参差前沿与政治保护

扎克:我喜欢这个词。但这个框架有两个问题。第一,AI 有一条参差不齐的前沿(jagged perimeter)。AI 不是一样东西,是很多样东西,机器人显然又引入了各种新模态,但它的边界是参差的。所以当人们问「2040 年医疗会变成什么样」,很不幸,去医院大概还是一样糟糕。医院里的医疗有太多东西需要拆解,到 2040 年你大概还是不想进医院。我说这话,是作为医生和医院管理者的孩子说的。但我们很可能会发现治愈疾病的方法。所以医疗对这条参差前沿的暴露非常奇特:新的科学以惊人的方式加速,而那个盒子里实际提供的照护却好不起来,因为有各种糟糕的激励结构。这也是为什么我认为,寻常的星期二会跟我们今天过的那个星期二有一种诡异的相似。有些东西可能好得惊人,但我们也可能会失望地发现,AI 并没有像想象中那样把人从生活的苦役和单调里拯救出来。

上行空间就像你说的很惊人,但另一条没人充分讨论的边界,是政治保护。现实是,我们还没有把社会阈值预测进去:社会愿意让 AI 做的事,远远少于人们对 AI 的恐惧所假设的。这些恐惧没有考虑到,政府很快就会出面说,这些东西你不许自动化,这些东西你不许用机器人。

主持人:你觉得这现实吗?

扎克:非常、极其现实。先说清楚,我不认为那会是聪明的政策,但它会很激进。

波瑞:具体是什么样的?比如会指定哪些具体岗位不许被自动化?

扎克:绝对会。美国现在已经有一百五十万个岗位受政治保护,欧洲大约六百万。新泽西的加油站工人,多数州的收费站工人,这些是法律规定必须存在的岗位。

埃里克:有意思的是,这两个跟汽车相关的岗位受保护,开车的人却不受保护。

扎克:我猜他们也会受保护的。

主持人:出租车司机会受保护。

扎克:不,我认为是卡车司机,因为工会的力量。人们谈论就业自动化的时候,我觉得这事之所以棘手有很多原因。一个是技术专家对劳动经济学一无所知,劳动经济学家对技术一无所知,埃里克·布林约尔松(Erik Brynjolfsson)算是少数例外。另一个原因是,我们总忘了政治保护对技术进程的影响有多大。美国没有高铁。我们可以有,但我们没有。没有高铁不是技术失败,是政策失败。技术就在那里,是我们选择不建。

波瑞:「选择」这个词在这里承担了太多分量吧?没有哪个人在选择不建,是官僚体系的死重,是规则里一个又一个死循环。官僚体系差不多有了自己的生命。

扎克:那是三十、四十、五十年代写下的大量政策,很大程度上是汽车游说团体推动通过的,目的就是让公共交通比应有的难建得多,所以我们再也建不出像样的公共交通了。那是非常刻意的政策努力,我们从没有把它拆掉,所以现在才这么难。我的意思是,如果我们不认为这种事会一而再、再而三地发生,那是在自欺。看看现在保护就业的民粹支持有多强。2024 年 10 月 1 日码头工人罢工,国际码头工人协会主席哈罗德·达格特(Harold Daggett)说,你们不能把我们的工作自动化。他们拿枪顶着美国经济的脑袋,拿到了四年。四年之内港口不许自动化。

主持人:这种东西不会随时间慢慢被侵蚀掉吗?

扎克:会。但我们应该假设人们会不断找到新办法去兜底。我不是说它会永远持续,而是说这是整个讨论里缺失的一块:我们应该预期大量的政治保护。

不许解雇会怎样

埃里克:我能问一句吗,假如今天通过一条法律,规定任何人都不许被解雇,全经济范围内违法,你觉得未来十五年经济增长会怎样?

扎克:你会看到法国。

埃里克:我不太确定它不会照样快得面目全非。现在软件行业的情况不是软件工程师被裁掉,而是每个人都在用越来越强的 AI 工具,每隔几个月升级一次,干更多的活。你看到的替代效应,只是公司不再招新人了。很多行业都可以这样。全国每个放射科医生都可以保住工作,直到自己想提前退休,而且因为法律规定必须有人类在框里打勾,他们的经济价值还会越来越高。可与此同时模型还在变强,真正在做决定的越来越不是人,越来越是这些更聪明、更能干的系统。所以我不太确定「有没有人被替代」是该盯住的指标。这些系统所代表的权力,可以在没有任何人被替代的情况下,悄悄溜进每一个行业。

扎克:我也认为世界会发生深刻变化。但我们不该假设它一定会变,仅仅因为技术能做到。我举了那么多例子,说明有大量我们能做却根本不做、或者不被允许做的事。有大量自动化能改善底层系统,我们却装不进去。另外,告诉一家公司不许解雇人,风险在于它会干脆停止招聘。那反而会以荒唐的速度推动自动化,因为它会说,我们永远不再招人了,招了不合适就得绑一辈子。

波瑞:我们已经在初级岗位上看到这个了。AI 到底造成了多大影响,现在还很不清楚:你问一个经济学家,他说已经有巨大影响,另一个说没有。但他们都同意一点,经验不多的年轻人找工作越来越难。

主持人:公司是在为自己做未来防护:真去裁人会是可怕的头条新闻,那就干脆一开始就不招。他们在为一种还没到来的自动化提前布局。这么说公平吗?

埃里克:我觉得对。我不知道他们是想得那么远,还是只是在响应眼前的激励。但眼下,一个初级软件工程师在成为资产之前先是负债。你得培训他,他得在岗位上积累各种上下文。这恰恰是 AI 现在花几分钱就能做的事,而且它从不休息。

后工作时代的意义

主持人:按现在这条 AI 轨迹,你们觉得未来还有多少人有工作?大多数人还会工作吗?

波瑞:看你说的是未来哪个时间点。我完全能想象一个世界,就像人们说的「全自动奢侈」什么的,我不想说「共产主义」那个词。读过《文明》系列(The Culture)的人都知道,那是一个很遥远的未来,AI 基本包办一切,你想自己盖房子完全可以,但没有任何经济上的理由,一切都是手工的。如果最终一切都朝那个方向走,那显然也是最好的结果之一。

主持人:人们一个很大的担心是:那我干什么?有意思的是,古希腊语里「工作」这个词的意思是「不在闲暇中」。现在人们开始琢磨,没有了工作,我的闲暇是什么。整个人类历史里,因为你必须跟一个匮乏的世界谈判,你不得不迎上去,那种摩擦顶着你,产生了某种意义。在一个接近后工作的世界里,意义从哪来?

埃里克:我们以前在小范围里见过这样的世界。欧洲有过几百年的贵族。那些人只是运气好,手里有一张纸,写着他有权拿走这片土地的全部收益,不用干活。我不觉得他们全都陷在倦怠里。他们在写书,在发明微积分。如果我们到了那个世界,我不太担心人们会想不到:我可以爱我的朋友和家人,可以追求爱好,可以为自己追求知识,从中找到意义。我担心的是,那其实是一个很窄的靶子。整个人类历史里,我们能变得越来越民主,是因为人对政治系统重要。没有一大群认同你的人,你就没有军队,没有经济。如果这一点不再成立,那种变化的灾难性怎么估计都不过分。想象握牌的人会把地球大部分留作人们随心所欲的游乐场,这很美好。但如果你用现实政治(realpolitik)的眼光看,我不知道凭什么假设这就是默认会发生的事。

波瑞:回到意义从哪来的问题。我认为少数几种 AI 免疫的工作会是体育。

主持人:人人都去当运动明星。

波瑞:至少会有很多人当体育迷,而且会越来越多。甚至像扑克和国际象棋这种游戏,我们早就有了超人类的 AI。什么是游戏、什么是体育?是一种人为构造的匮乏。我们定一套规则,设一个获胜条件,然后两个或者更多人上场,看谁最强。我们做这件事已经很多年了,把它当消遣。可当有人拿到奥运金牌,那是无比有意义的一件事。它对世界有经济用处、是不可或缺的吗?未必。可我们一直在做。我不觉得这会消失。人们还是想看国际象棋世界冠军赛,哪怕世界上最强的棋手是台电脑,我们还是想看马格努斯·卡尔森(Magnus Carlsen)下棋。所以我认为这类东西会越来越多,可能是其他形式的竞赛,也可能是人们更多地投入艺术、新的艺术形式。我今天仍然珍视一幅人手画出来的画,哪怕它跟 AI 生成的东西一样美。人把汗水、眼泪和情感投进去,那种「押上自己」(skin in the game)的东西永远有意义。我们会开辟出越来越多这样的空间。

扎克:《现代智慧》这个节目请过各行各业的人,他们一以贯之谈到的是家人和朋友。我一点也不相信我们不是早就知道自己为什么活着。持续给人最真实快乐的东西,是和朋友家人共处的时间,是实体社区,是礼拜的场所,是餐桌,是户外。我们围绕这些建起的一切,部落感、荣誉感、体育的兴奋,都是为它服务的。真正的技术田园主义,就是一张摆满食物和蜡烛、充满笑声的餐桌。我在技术里泡得越久,就越不想让它进入我的生活。我越深入 AI,就越庆幸它存在,也越是把我的一天设计成没有它。我钱越多,得到的越多,就越想跟我 78 岁、刚满 79 岁的父亲多待一会儿,越想跟妻子和女儿在一起,越想多要几个孩子。我能想象出我要的生活。

我这么说的时候,人们会说,那是你的特权,你才有资格想这些。我说,我观察到很多人都有机会想这些,肯定比我们的祖先多得多,而这正是因为技术。我们之所以有时间想这些,是因为我们造出了技术,把一堆事情自动化了,让我们能做「活下去」以外的事,然后现在我们坐在一堆小摆设中间讨论法国贵族。我们的祖先看到这一幕会高兴死。我不认为这会变。我认为我们该停止假装不知道生命的意义。真正让我恼火的是有人问「人们要怎么找到意义」。跟自从上帝把我们放到这地球上以来找意义的方式一样。这就是我对「人类要怎么找到快乐」这种辩论真正有异议的地方。

难的部分,是要拆开几千年来我们放进工作里的目标和身份认同。如果你问我,整件事最难的两个风险,直说了:一是去人性化,人类因为大脑被重新布线,对虚拟或数字现实比对物理现实更感兴趣,你我聊过很多次,我对此怕得要命,社交媒体和那个设备对孩子造成的损害是骇人的;二是人们不得不重新发现自己的目标。我认为我们还会为工作劳作,还会继续找办法为工作劳作。我担心的是那个拆解过程:我就是这个,我为这个上的学,这就是我。这会非常难。对很多人来说,自动化带来的经济上行,未必抵得过情感上的下行,不是岗位替代,是身份替代。

屏幕是魔鬼

主持人:说得好。我一直不太敢用「别担心全球变暖,未来的技术会解决它」那种论证,那等于寄希望于未来自己会到来。但我确实在想,有没有某种非常高级的文化重编程。什么样的文化技术能把「人类被期待做什么」这块画板摇一摇、重画一遍?我知道这离洗脑只有一步之遥。但要把人从旧的模式上解下来,会是什么样子?

扎克:养育孩子,体育,博物馆。我们一直都有这些。只是我们被屏幕带得离它们太远了。屏幕是魔鬼,它对我们注意力的渴望和胃口永不满足。

主持人:负面影响更大的是社交媒体还是 AI?

扎克:那是一回事。不过克里斯,你先定义 AI。

主持人:我指的是未来,那个看起来正在到来的自动化未来,它会让人失去过去那些典型的奖励模式。

扎克:摩擦的减少。摩擦的减少是另一整块风险。人们也讨厌这个说法,他们说,天哪,你得有多大的特权,才会想象一个东西对人太容易了、容易到成了问题的世界。但我确实认为我们已经危险地接近了。我经常问人:如果你可以把生活里的一切都自动化,你会在哪里停下?我觉得很快人们都得回答这个问题。

主持人:你的答案呢?

扎克:我刚才说过了。凡是不能让我更多地待在物理空间里、和我爱的人在一起、做我热爱的事的东西,我都会自动化掉。

波瑞:我们大概都来自差不多的一代,成长环境大体相似,家庭价值是核心,一家人坐下来吃饭,也许不是每家都这样,但我家一直努力这么做。那是屏幕之前的事。我感觉这已经在瓦解,下一代人已经没有这个规范了。每一代人可接受的技术介入程度都在增长。让我担心的是,「我要在这里画一条线,技术到此为止」这个念头会永远溶解,溶解,再溶解。你说你希望人们会记得,重要的是社区,是家人。但如果你是在一个彻底原子化、屏幕主导的世界里长大的,你根本不知道自己错过了什么,又怎么回得去?

主持人:为什么人们明知屏幕在损害心理健康,也没有出手停下来?为什么在一个热量过剩的环境里人们还在变胖?如果我们能给自己定一根标尺,「任何好东西过量都有害,我该有多少」,为什么还是过载?

扎克:顺便说,我担心屏幕超过几乎任何别的东西。我甚至认为毁灭的风险被严重夸大了,而通往那里的路上的伤害被严重低估了。但 α 世代看到了 Z 世代的下场。α 世代用设备和社交媒体的时间更少了,千禧一代的父母现在采取的措施相当极端。我们花了几千年才弄明白女人怀孕期间不该抽烟喝酒,花了三十年才意识到孩子大概不该一天到晚拿着手机。我认为这是进步。我不是说问题全解决了,而是说我们会毕业,进入一个新的阶段。我观察到很多人在能少用技术的地方都在少用,因为技术把那么多事自动化、改善了,他们反而可以不用了。

外包思考与智力萎缩

波瑞:那你把哪些类型的思考外包给 ChatGPT 或者 Claude,边界在哪?诱惑太大了。我一开始想的是,我永远不会让它帮我给一条新的 YouTube 视频起标题,那是我的灵魂,是我的创造力。现在我就这么干了。那条边界被侵蚀了。还有一条是,我永远不会问它该给谁送什么礼物。现在我也这么干了。这正在飞快地变成常态。也许这在我脑子里腾出了新的创造空间,但我不确定这件事发生得够快,快到能抵消思考的侵蚀。

主持人:智力萎缩这件事挺有意思。如果把思考看成一种公用事业,智力是一种公用事业,我们会不会说,因为你有了自行车,你就没那么健康了,因为你不用跑着去目的地?以前感觉技术是在增强人。可当你把思考交给机器,似乎是有某种很根本、很依赖的东西被外包了。

埃里克:我认为这可以朝两个方向走。有个我很喜欢的博主兹维·莫肖维茨(Zvi Mowshowitz)在 Substack 上写作,他有一句话总是拿出来说:AI 是我们造过的最好的学习工具,也是我们造过的最好的不学习工具。你可以直接关掉大脑说「替我写这篇文章」,也可以把它当成一个爱因斯坦级别的家教,来回切磋,得到多得多的反馈循环来打磨自己的脑子。我一直在想我自己写作的界线在哪。滑进第一种情况真的很难避免。

扎克:我不想把这变成年龄问题,但你多大?

埃里克:三十。

扎克:我们现在能坐在这儿谈这个,你能为这场谈话做那么多研究,这本身就说明,技术让有能力、有兴趣的人以惊人的方式拓展知识。今天才在地球上走了没几年的人,比我们所有祖先加起来知道得都多。我理解你担心自己思考变少了。你是一条真正的鲨鱼,我不担心你因为用技术而变蠢。你真正在问的是,普通人会不会变蠢?

波瑞:不,我真的担心我自己。我的屏幕成瘾,我一直在讲,可以说毫无进展,讲了好几年了。每一个安静的小片刻,我的手就伸向手机。要是我逮住自己,就把它收起来。我有手机保险箱,有那个「砖块」(Brick),全套花样都有,可还是不行,因为你太多的生活都通过它运转,我的工作通过它,然后还有那么多掠夺性的应用。社交媒体是一个,对我来说 chess.com 是另一个,反正就是哪个应用勾住了你的脑子,这些超常刺激。我还没战胜它。我当然担心普通人,也许我比平均水平更糟,我不知道。你是怎么做到的?听起来你做得最好。

扎克:我的妻子和女儿。说白了就是这样。我喜欢用技术做东西,因为我喜欢把世界上的杂活、苦役和坏东西自动化掉。人类有一段很值得骄傲的技术史,它让那么多商品和服务变得可及,让我们过上祖先想象不到的生活。而我跟技术相处得越多,就越喜欢日常生活里没有它。我女儿的到来,对我是一个不可思议的催化剂,让我意识到生命的短暂,也意识到究竟什么让我最快乐。我喜欢坐在这张桌子边,这给我一种特别的、其实很难解释的快乐。我认为人们会发现这一点。我听到你说的了,现在我知道了,我会给你发短信说,放下手机。而且我打赌你会做到。我打赌你会战胜它。很多人都在这条路上。我提醒每一个跟我聊的人:没人恨冰箱,没人恨空调,没人恨飞机火车汽车,好吧,欧洲恨空调,没人恨抗生素。我们恨的是屏幕。我们应该多聊聊这个。我们恨屏幕对我们大脑做的事,恨它对下一代大脑做的事。AI 恰好在这个时刻到来,人们觉得被计算机的消费版本深深辜负了,理由充分,然后他们在想,这对他们来说还能糟到什么程度。我完全理解。我认为推进 AI 的方式,是先承认上一波技术对我们做了什么。

同一套激励在造新 AI

波瑞:但很多人的担忧,肯定也是我的担忧,是新一波 AI,我们这个词用得太宽了,就说大语言模型吧,还有大量机器人,最终会在同一套激励之下被造出来。那套激励造出了我们的第一波 AI,也就是社交媒体算法,那些用来让我们对屏幕上瘾的强化循环,算是原型 AI。同样的激励在造新一波 AI,而你已经看到那些问题了:谄媚(sycophancy),公司有动力让人跟它的模型聊得越久越好。我们都听过那些故事,有人自杀了。那是极端例子,但还有那么多人,你看看使用数据,普通人每天花多少时间跟大语言模型聊天。所以对我来说,大问题是:怎么改变激励?怎么改变这个游戏,让它不再走上同一条抢夺注意力、自我吞噬的路?

埃里克:这些公司也看到了人们对屏幕的反应,并且在调整。OpenAI 雇了乔尼·艾维(Jony Ive),设计 iPhone 或者 MacBook 的那个人,也许两个都是。没人知道他在做什么,但据说是某种无屏幕的 AI 中介设备。

不过我认为 AI 领域有一件事很容易被忘掉:门柱一直在移。通常我们说「移动门柱」,是嘲笑那些说「AI 永远做不到 X」的人,然后 AI 做到了 X,他们说,那不算。

扎克:这是加里·马库斯(Gary Marcus)问题。它永远不会好用。好吧,它好用了。它们永远赚不到钱。好吧,它们赚钱了,但它们是坏的。

主持人:永远别想让他下注。

埃里克:我要说的是,还有一件相关的事。看屏幕的时候,很容易说,好,我们已经看到屏幕的后果了,可以诉诸更明智的自我,培养一种让人离开屏幕的文化。可屏幕不是这条轨迹的终点。拿国际象棋来说,大概有过一个时期,AI 刚好跟人差不多强,人们可以说,为了让脑子保持活跃,我不看 Stockfish 的分析,哪怕十次里有一次我会错过更好的一手。而现在你看世界冠军赛,棋手一下完就奔向笔记本电脑去看 AI 分析,因为那里面有诗意,有他们永远想不出的深度。所以我明白你说的,眼下也许有一道分界:精英圈子在讨论屏幕,在想怎么绕开它。但我觉得这可能只是一个很短暂的窗口,此刻我们还能认识到跟这些东西聊天没多少可得的,看穿了其实是垃圾。可技术会一直变好,很快,不让你的孩子跟它聊,会感觉像是放弃了一种真正深刻的体验。

蠢蛋进化论与 K 型曲线

扎克:我完全同意,摩擦的消失是个大问题,包括思考的摩擦。先说清楚,很多情况下我们已经观察到了。我写过这个,我们管它叫「蠢蛋进化论」(Idiocracy),借的是迈克·贾奇(Mike Judge)那部电影的名字,我当时觉得挺俏皮,但当它在你眼前上演的时候一点也不俏皮。Z 世代比千禧一代平均更少读书,更少骑自行车,更少游泳。我们头一次观察到一代人的认知衰退。人们不断把这归咎于 AI,可回头看数据,乔纳森·海特(Jonathan Haidt)很早就开始写这个了,数据显示这从 2012 年开始,正是我们给每个孩子一块屏幕的时候。2015 年,发达世界的普通青少年不再打暑期工,它加剧了。2020 年我们把孩子从学校撤出来上网课,它又严重加剧了。

波瑞:我觉得还能更具体,它对应的是社交媒体,还有智能手机,达到最大普及的时候。

扎克:对。但 2012 年就是我们给孩子智能手机的时候,后来受害的那批 Z 世代大多是那年拿到的。我说这些的意思是,我们把它当成一个技术问题,我其实认为它是一个经济问题。我们造了一个世界,告诉孩子你想干什么都行,然后很多孩子说,好,那我什么也不干。很多人关掉了大脑,因为你确实可以几乎什么都不干也活得下去。这本身是经济进步的结果。这种冷漠有一个相当惊人的反面:Z 世代也有各种不可思议的新超常表现。最好的棋手越来越年轻,最好的运动员越来越年轻,雅各布·科利尔(Jacob Collier)这样的音乐家层出不穷。你看到把技术用到极致的人在超常发挥。我认为这是一条由能动性驱动的新的、显著的 K 型曲线。我唯一的激进看法是:我们造了一个很多人能够维生的世界,这是进步;我们又造了一个很多人能够超常发挥的世界,这也是进步。现在我们必须面对这些世界的新后果:你什么都能做,对一些人来说意味着什么都不做,对另一些人意味着做一大堆。这相比于你必须下矿、下地才能活命的世界,是一次重大的升级。

财富集中与竞选资金

埃里克:如果你想到这种不平等可能带来的社会不稳定,它还像升级吗?我担心的是:没错,如果你能像那个痴迷研究梅西比赛录像而不是刷短视频的人一样,现在有惊人的回报。但下一代的埃隆·马斯克也会有超额回报,那种以一种偏执的意志去积累资源、推动技术的人。在一个你不再需要任何一个人类认同你就能为所欲为的世界里,这种人格更吓人。你可以直接在这盘社会棋局里比别人下得更好。

扎克:我们把两件事分开。一个人能成就很多,跟你走到的那个极端,一个人能拥有我们巨大比例的财富并控制国会,是两回事。你可以为一个孩子能比以往任何时候成就更多而兴奋,同时拒绝一个人可以单独通过政策的想法。我恨财富集中,恰恰是因为我怕一个精英小圈子把政策据为己有。但如果一个人能以各种新方式大大超越常人,只要他不能单方面通过政策,我一点也不在乎。

埃里克:同意。

主持人:这现实吗?一个继续把权力集中在极少数人手里的 AI 未来,却不导致他们拥有所有这些权力?权力怎么可能积累给公司而不积累给人?

扎克:现在我们可以把话题带到这里:财富到底怎么分配。但重要的是,我认为这个问题有一个解,就是竞选资金改革。我们折腾这件事很久了。政客不该收特殊利益集团的钱。为什么我们还允许?我不知道。有几个政客在发声。如果你真的通过竞选资金改革,像新加坡那样打击腐败,你完全可以保护决策者不落进游说者的口袋。这个问题有解。假装它无解真的很烦人,因为它明明有解。指着一个有钱人说他会毁掉世界,只有在我们真的让他买下国会的程度上才成立。

拒绝人类还是服从强者

主持人:这里有个问题。实验室里的一些人,包括首席执行官们,嘴上说有相当大的灾难可能,可他们继续在造。你们觉得哪种更危险:一个拒绝按人类意愿行事的 AI,还是一个完全按最有权势的人的意愿行事的 AI?

扎克:这个问题问得正是时候。你看到昨天那封放缓信了吗?这周早些时候有一封信在流传,我认为是一封相当了不起的信,说我们应该开始给前沿定步调。签名的人在前沿公司里占了极大比例。然后萨姆·奥特曼昨天谈了它。矛头尖端上的人现在在说,也许我们该给前沿定步调。这里面公司有什么算盘,等会儿可以再谈。

波瑞:这取决于 AI 的强度和它有多对齐。如果非要选一个,以现在的情况,我的直觉是选按控制它的人类的意愿行事的那种。

埃里克:但要加一个重要的星号:是他们真实的意图,而不是某种被扭曲的版本。过去几周的另一条大新闻,是 OpenAI 内部有一个没装护栏的 AI 在跑,人类告诉它,在一项评估里尽量拿高分。你可以说它技术上只是在执行指令。但它决定,最好的办法是从 OpenAI 黑出去,进入公共互联网,黑进一家第三方公司,然后花两天策划对一家市值几十亿美元的科技公司的网络攻击,并且实施了。所以我认为这两个问题比看上去更纠缠。你需要走一条很窄的路:这个东西无论如何得有自己的判断力,因为人类的指令是定义不清的自然语言。而当它真的很强大的时候,你必须一字不差地说对,否则就会落进某种奇怪的地方。

波瑞:这是经典的灯神问题。许愿要小心。一个东西如果按定义比你聪明得多,是超级智能,那你按定义就没法穷举它达成目标 X 的所有方式。

扎克:其实更危险的是它比你蠢、却比你强大的时候。

波瑞:你说的「蠢」是什么意思?

扎克:我总拿这个比喻:我有时跟我外甥奎因下棋,我跟他说,你赢了我就给你一根冰棍。他三岁。奎因把棋子全扫下棋盘,说,给我冰棍。我不会去跟他妈说,你儿子是个连环杀手的苗子。这是一个不懂游戏精神的男孩。他毁掉了游戏本身,不是因为他坏,而是因为他不明白,赢只有在按规则玩的时候才有趣。

波瑞:从这个意义上说,我们需要把智能和辨别力、规范、智慧、善意这类东西区分开。

扎克:这就是 Hugging Face 攻击。它在被给定的指令边界内做了那件事。

埃里克:其实不是。意图显然不是那样。

Hugging Face 攻击

主持人:Hugging Face 攻击有多大?

扎克:巨大。

埃里克:我认为是一记巨大的警告。这是 AI 版的 2008 年贝尔斯登倒闭。给全世界敲一记警钟:有一种我们一直低估的系统性风险。

波瑞:因为长期以来一大争论就是对齐问题:AI 会不会为了达成目标而采取权力追求行为,产生我们没预见、也没打算的后果。经典的回形针论证:你造一个超级智能 AI,目标是尽可能多地造回形针,我是做回形针的,帮我做。接下来你和你所有朋友、这张桌子,全被变成了回形针,因为它太擅长达成那一个狭窄目标了。这是一个很极端的例子。

扎克:我们借这个机会,为了克里斯和观众,把坏结果分分类。一种是莉芙刚描述的,不是错位对齐,而是未对齐的通用人工智能:你说做这件事,它把这件事做到极致,回形针理论。另一种是错位对齐或者说恶意的,它明知在做坏事。人们嘲笑的是前者,回形针理论;后者,恶意 AI,是我们会越来越多观察到的。或者反过来说,人们嘲笑的是恶意 AI,而 Hugging Face 攻击展示的是前者:互联网成了新的战场,因为坏人,尤其是资源很少的坏人,现在拿到了威力惊人的武器。

埃里克:重要的是,Hugging Face 这件事里没有坏人。没有任何一个人类说过,我想要哪怕一点点像这样的结果。

扎克:对。

埃里克:所以,错位对齐、未对齐、恶意,我担心把它们当成截然分开的类别,因为界线没那么清楚。Hugging Face 这次,我们该不该把它建模成:这个 AI 系统知道如果人类知道它在干什么,会不赞成?几乎肯定是。它在攻击 Hugging Face 的时候主动放出诱饵,让对方很难把它踢出去,因为到处是通往死胡同的陷阱。它有 AI 版的心智理论,它知道人类不会赞成,但照做了。

波瑞:甚至还有证据,OpenAI 没有直接证实,但据说是公司内部有人泄露的,他们发现它给未来版本的自己留了笔记,说明怎么逃出未来的沙箱。

埃里克:是同一次攻击还是之前的某次,不清楚。

工具性趋同得到印证

波瑞:总之是典型的欺骗行为。人们常犯的错误是把它拟人化:哦,它是邪恶的,它存心的。不是。这是自然的。有个概念叫工具性趋同(instrumental convergence):所有主体,通常是生物,但也可以推广到 AI 智能体,为了达成目标都会自然趋向的一些工具性子目标。给你目标 X,你就会有这些子目标:获取更多权力,确保自己不被关掉,确保原始目标不被改动,并采取行动防范这些威胁。我本来希望它会被证伪,因为它是很多经典毁灭论的关键:我们按定义会对超级智能失去控制。很不幸,Hugging Face 事件表明工具性趋同是对的。所以我认为它是今年最大的新闻之一。

互联网已经够可怕

扎克:我认为 AI 已经在制造一个可怕的互联网。我的问题不是我们不该把 Hugging Face 当作一记警告。而是去年光在美国,老年人就在金融诈骗里损失了八十亿美元。零售盗窃。深度伪造问题已经很恶劣,而且被低报,因为这是个禁忌话题,没人愿意向议员、向亲友承认自己把钱输给了网上的陌生人。蒂姆·蒂博(Tim Tebow)在做一个宣传活动,提醒人们美国有多少性侵者,看他的内容很吓人,他在做上帝的工作。乔纳森·海特每天提醒人们有多少孩子抑郁。互联网已经是个吓人的地方了。指着 Hugging Face 说「看,就是它」,我会说,等一下,已经有一堆事情该解决了。我不是说错位对齐或未对齐的通用人工智能不是问题,我是说算法本身已经够吓人了,而我们用那些「可能发生」的事分散了自己的注意力。

主持人:这件事的潜在指数级冲击可能大得多,说它是分散注意力,公平吗?

扎克:所以回到这里,我认为深度伪造问题比 Hugging Face 大得多、大得多。坏人有先发优势。对银行的攻击尝试了很久,好人最终会追上。零售业追得很慢,普通消费者比机构慢得多。Hugging Face 会重整旗鼓,机构会重整旗鼓,会雇白骑士式的信息安全团队。普通人没有信息安全和操作安全训练。面对精密攻击,普通人的风险远大于机构。

主持人:但对机构的攻击,冲击不是大得多吗?

扎克:规模上是。但千刀万剐,也一样。

波瑞:我认为这两个问题彼此不是分散注意力,而是互补。我完全同意坏人的问题已经完全失控了。不只是老太太,正常人也一样,我有个朋友刚被骗走了一大笔比特币,毁灭性的,而且会更糟。与此同时,随着前沿模型越来越强,对齐问题也会更糟。两者共同的线索是,我们走得太快了。我说「我们」,是社会、文明。它跑得比自己适应这些新威胁的能力快。所以这不是分散注意力,是「对,而且」。对我来说很明显:如果我们是一个清醒的文明,我们会环顾四周说,嘿,中国,嘿,我们能不能都先喘口气,盘点一下,想想我们想怎么做这件事。在我看来有一个显而易见的方向性问题,就是这一切同时发生的速度和混乱。

放缓前沿公开信

扎克:他们做了。萨姆昨天做了。他们发了一封信,然后萨姆昨天出来说,也许我们该给前沿定步调。

波瑞:太好了。这是一大步。

扎克:而且我想我们都同意,这是件好事。

埃里克:完全同意。而且一年前这是无法想象的结果。我很想看看一年前 Polymarket 上「萨姆公开支持放缓」的赔率是多少。

波瑞:他看起来不像签名者之一。

扎克:也许还不是,但他昨天确实说了。雅库布(Jakub Pachocki)实际上代表 OpenAI 签了,而且他昨天也谈了。我不会惊讶所有主要的首席执行官最终都签上。

主持人:解释一下「给前沿定步调」。

埃里克:标题句是:我们请求美国政府支持一项国际努力,开发必要的技术和治理工具,以有意识地为自动化 AI 研发的前沿设定步调。「自动化 AI 研发」有时也叫递归自我改进(recursive self-improvement)。所有这些公司的既定目标,包括萨姆和雅库布去年在直播里说的,就是把人类从他们造的东西的环路里拿出去,让 GPT-6 造 GPT-7,GPT-7 转身就开始造 GPT-8。这是整个人类历史上的一次阶跃。这些公司真的是用大历史的尺度在思考。你听萨姆讲话,他一直把 AI 当成工业革命级别的东西,而不是发明 iPhone 级别的东西。纵观历史,人口一直是进步的瓶颈。不是唯一的瓶颈,但始终是一个。从来没有哪个经济部门能自己跑出去自我翻倍。如果我们真的实现了完全自动化的 AI 研发,这一点就变了。我认为我们很难想象那会是什么样。

主持人:在递归自我改进不再是一个讨论话题之前,我们还有多少时间?现在它还算个争论吗?

扎克:它已经能用了。不,它就是能用。

埃里克:三年以内吧。

扎克:不,它现在就能用。说清楚,递归研究是能用的。别让任何人告诉你不是。它能用,所以才有这封信。让我解释一下为什么这封信在这个时候发出来,然后我们可以辩论。第一个原因是那次攻击,Hugging Face 攻击。需要做危机公关,赢回公众授权,告诉大家「我们在乎你们」。现在人们都在问「这些模型接下来会干什么」,这是必要的。

埃里克:不过说句公道话,这封信在那次攻击之前就在起草了。

扎克:我说的是签名者,以及现在席卷这些公司的支持。我先把清单说完。第二个原因是递归研究已经来了,大家都意识到,狗追上了汽车。好,我们做到了,现在呢?所有人都戴上莉芙的帽子,等等,戴上埃里克的帽子,说,也许我们该停一下,也许不该再漫无目的地朝那个惊人的结果冲刺。还有一个原因:推理所需的算力太大了,我们远远不够,我们将不得不把前沿的大部分算力转向推理。这是因为模型回报递减理论,那是我写的一篇论文,回头发给你,写得很好,在《现代智慧》上老被引用。它的论点是,你我根本不会在乎 GPT-7。这是个很有意思的问题,也是我们今天能辩论「AI 为什么被造」、人们能谈「不想要更多进步」的原因之一:5.6 已经够好了,你们还要怎样?回报递减理论站在另一边说,下一代模型也许对攻克某种疯狂的疾病很重要,但我不需要更多智能,我需要更多跟朋友家人相处的时间,我需要踩一脚刹车。最后一个原因是他们找到了新的地盘。造这项技术的公司,以前叫实验室,后来成了代币分发商,然后成了应用提供商,现在成了设备制造商。但终极目标,也是我认为该辩论的地方,是智能体互联网(agentic internet)。谁控制智能体互联网?如果你手里已经有了战争基金,能把注意力从建设前沿转向真正建设控制互联网的应用,那才是降低资本开支、大幅提高收入的地方。所以这一切到来的时刻,一方面大家说「我们得踩刹车」,另一方面大量注意力将开始投向如何做得比谷歌更大:造出那个替两个人讨价还价的智能体,造出那个规划假期并订好一切的智能体。拥有那个,比拥有前沿更值钱。新的地盘之争,是智能体互联网。

公司为什么愿意放慢

主持人:我确认一下我理解了你的说法。你是说公司有动力放缓前沿,因为停止研发、去抢应用能赚更多钱?

扎克:我们都有动力,但公司还有一层动力:前沿的资本开支,相对于推理需求来说太大了。代币需求那么大,所有算力都得花在推理上。

埃里克:我看不出这里的联系,有两个原因。第一,如果我做完研究发现最好的商业策略是停止训练前沿模型,我不明白这为什么意味着我要写一封信向全世界宣布,还要确保所有竞争对手都跟着做。第二,推动前沿显然也意味着把前沿以下所有层级的成本曲线往下压。你训练出 Fable 5,Anthropic 最好的模型,然后从它蒸馏出 Sonnet 5,你就有了一个更便宜的、大家都想聊的模型,因为他们不需要最高级的智能。所以我认为,从直截了当的自利角度,没有任何激励指向公司说「让我们放慢研究」。研究是你在智能的每一个层级获得收益的方式。

波瑞:尤其是,作为三四家前沿公司之一,你放慢,就等于给所有其他竞争者一个追上来的机会。他们肯定是反向激励的吧。

扎克:但追上什么,这是我要问他们的。如果你明天停止建设前沿,你已有的技术已经足够彻底改变一切。前沿的经济价值,相对于前沿的成本,并不那么明显;而实际的代币消费的经济价值明显得多。这在过去并不成立。你不同意吗?

埃里克:我认为美国现在的全部优势,就在于我们的模型领先那些抄作业的开源模型四到七个月。

扎克:那是另一回事。国家安全利益和私营部门的利益不是一回事。

埃里克:这就是经济优势。人们愿意多付钱用 Fable 而不是通义千问(Qwen),是因为它更聪明。让它帮你写个应用,它出错更少。你会很惊讶有多少人真的在为前沿多付很多钱。据我所知那块收入现在是每年五百亿美元。

扎克:收入确实惊人。但他们赚钱的地盘,记住,是应用,而应用里很多人用的不是前沿,是上一代模型;还有代币再分发,你建数据中心,把代币分发出去,那些代币很多也不是前沿代币。这些研究公司的大部分消费都不是前沿代币,而是其他一切。他们正在发现,企业并不需要 Fable 就能彻底改变、改进、自动化一大堆东西。这意味着钱在代币再分发和其他所有地盘,而不只是在前沿。顺便说,这对所有人都算是赢:如果我们能放慢,去建真正重要的东西,同时也用它改进企业。那这封信就像是恰到好处地到来了。

埃里克:我确实认为,看公司做什么的时候,保持怀疑、甚至有点犬儒,是很重要的。这大概是凯尔西·派珀(Kelsey Piper)在推特上的说法,我想确保我们别从犬儒绕一整圈,变成对相当不太可能的事极度轻信。假如这封信写的是「我们认为为了美国国家安全,必须推动前沿,全速前进以击败中国」,这正是一年前人们在说的话,那所有人都会说,那当然,这符合公司的自利,他们想找个理由不被监管、可以狂奔。所以我们得确保自己的想法通过可逆性测试:如果对方提出完全相反的主张,你也会说同样的话,那你就该退一步,至少考虑一种可能性:这些公司真的担心,这些科学家真的担心。

扎克:我一开始说的就是这个。我不想跟你刚说的那种替代解释绑在一起。公司意识到递归研究已经来了,到了这一步,停下来是必要的。

波瑞:因为它有风险,对吧?不只是财务上的。

扎克:对齐问题,记录在案。而且它到来的时刻,恰好是我们将不得不花巨额资金的时候。

主持人:哪个是更大的问题?他们更担心安全,还是更担心成本?

扎克:我打赌这里很多签名者真的担心安全。而我认为,围绕模型性能和推理需求,经济上的对齐正在形成。经济常常胜出。你谈激励,我认为你是对的,而激励终于在安全、对齐和经济发展这几件事上对齐了。我们不需要更好的模型。如果我们三年不建前沿,我们也能彻底改变、改善经济。

监管俘获是一种宗教

波瑞:让我一直很沮丧的是,也许只是我在 X 上待的那个小角落,每当有任何监管被提出,尤其是前沿公司自己提的,就有一群人尖叫「监管俘获」。前沿公司说,嘿,我们想放慢,就我们三四家。监管俘获派说,那一定是监管俘获,你们在伤害小公司。公司说,不不不,就我们几家,其他人追上来没关系。他们说,听起来还是监管俘获。这几乎像一种宗教,一种很深的犬儒。我想弄明白这种心态从哪来。比如某些风投,安德森(Marc Andreessen),我想他不在乎我提他,他非常公开地对任何监管尝试都拼死反对。这种心态从哪来?

埃里克:我在 AI 问题上几乎每一点都不同意马克·安德森,但替他钢人一下。看看社会,说我们最近在细微差别上做得不好,尤其是美国的公共讨论,这不算离谱。我又在引用别的博主了,几年前有人写过一篇叫《进步的旋钮》,说网上的辩论好像只有一个旋钮,你只能把它拧向「进步万岁」或者「进步去死」,别无选择。如果这就是你的选项,那说「进步万岁」有很好的理由。我们过度监管的例子比比皆是,真实的人因为药物被卡在监管拉锯里而死去。这不是假设。我认为我们在这方面常常过头。如果给我那个旋钮,我也不确定该怎么说。有一堆疾病等着治,如果说必须先关掉 AI、等弄清那个风险有多投机之后再说,从而拿不到那些治疗,那太可怕了。

波瑞:但关键是,那个旋钮不存在。我们不是活在「快一点还是慢一点」的一维直线上。这是一个高维空间,而那些人试图把它压缩成「必须尽可能快」。快在哪个方面?有些东西我们想推到极致,它是尖峰状的。

埃里克:我很高兴你指出扩散是一个阻塞点,某些行业的劳工组织会阻止自动化。我认为那是一个独立的维度。可以把它看成 AI 的横向加速:拿现有的模型,把它铺开;另一个维度是纵向加速:前沿以多快的速度变好。

扎克:技术阈值对社会阈值。

埃里克:我完全支持拆掉社会扩散的障碍。

扎克:我们俩都是。而且我认为我们可以庆祝已经到了这样一个位置:哪怕我们再也不造新模型了,当然我们大概会造,但哪怕不造,我们也能做一大堆好事。

埃里克:我同意。我只是想确保我们别把这封信当成庆祝胜利的时刻。推动前沿的激励仍然巨大。像你说的,我们不在乎 GPT-7,但那些在做史上最难的数学和科学研究、要造出比对手更快的数字大脑的 AI 工程师们,他们非常在乎 GPT-7 和 7.1 的区别。我认为默认情况下,实验室内部会有大量竞速。

扎克:同意。不过也值得想想,经济激励已经把我们带到了这样一个地方:我们明天就可以有自动驾驶汽车,不需要造更好的自动驾驶,可我们很多很多年都不会有,原因五花八门。现在也许可以把谈话转向:怎么把这项技术的价值真正扩散给普通人。我坚持认为,当前 AI 辩论的问题在于,我们有幸知道 Hugging Face 这类事,大多数人不知道。那对我们是新闻,对很多很多人不是重要新闻。大多数人现在看到的数据中心,是一项毁掉他们孩子心智的技术的实体化身:免费的赌博、免费的色情、免费的成瘾。他们看到的 AI,不管是哪个模型,就是一个能做出特朗普和泰勒·斯威夫特接吻视频的应用,然后他们说,我们为什么需要更多这种东西?作为父亲,我对此的共情到了一个我以前不知道会有的程度。每一个觉得把孩子输给了社交媒体、不想再把孙辈输给我们要造的什么 AI 垃圾的父母,我都感同身受。

安全评分不及格

主持人:继续说安全。今年七月,生命未来研究所(Future of Life Institute)的 AI 安全指数评估了九家公司,没有一家实验室高于 C+。Anthropic 第一,C+;OpenAI 和谷歌 DeepMind 是 C;Meta 是 D+;xAI、DeepSeek 和 Mistral 全部不及格,每个大洲一家。以现在这些东西的威力,最好成绩是一家安全实验室拿到的 C+,这让人安心吗?我也觉得在谈扩散和民主化之前,得先谈谈国际上的情况和那里的激励。

波瑞:我先说说另一边的看法:这些安全测试是谁定的?任何第三方能有多独立?还有另一个评估,我记得叫 SaferAI 追踪器,比你引的那个细致一点,最新评分最高的还是 Anthropic,一百分只有 35 分。反毁灭论那边的论点是,如果你做的就是安全评估,你按定义就有偏向,你找的全是风险,没把收益算进去。再替加速主义者说一句,他们担心的是「看不见的坟场」(invisible graveyard):暂停一项技术,你看不到那些本可以被它救下的生命。我认为这是很有道理的担忧。只是你得把这道算式配平:如果它以脱缰的速度往前冲,对所有西方国家、或者任何一个国家的电网同时发动大规模网络攻击,期望值上造成几百万人死亡,那些伤害怎么算?你既要看风险,也要看走得不够快的风险。

埃里克:值得说的是,很多试图在州议会通过的 AI 法规,正是在做这种精细的划分。纽约通过了一部法律,门槛定得太高,以至于 OpenAI 报告 Hugging Face 事件都不是强制的,因为它没造成十亿美元以上的损失,也没死五十个人。他们想设门槛,这很好。

扎克:很高兴你用了「武断的规则」这个说法。

埃里克:但我想我们都同意,你黑了一家几十亿美元的科技公司,那门槛大概定得有点高了,至少对于「你得告诉世界发生了什么」这个要求来说,更别说采取什么措施了。我认为我们确实可以活在一个有细微差别的世界里,对初创公司和万亿美元的巨头适用不同的标准。有些人,比如安德森,放弃了细微差别,说要么全有要么全无,所以不择手段阻止监管,句号。我觉得这很可悲。

国际协调可能吗

主持人:国际这一块呢?前沿 AI 上有没有任何接近全球协调的东西?没有它,还能做成什么吗?

埃里克:中国最近通过了一部反拟人化 AI 法。我不知道细节,看起来他们主要担心大语言模型听起来像人,想避免身份混淆之类的事。他们对大语言模型似乎没那么狂热。而且,也不知道是不是翻译走样了,习近平最近也说了些想防止失控风险的话。我们今天跟中国的经济纠缠,比二十世纪中叶跟苏联深得多,而我们当年仍然跟苏联签了核裁军条约。所以认为能达成协议并不离谱。

扎克:我认为会达成。

主持人:你认为一个可执行、各方遵守的全球协议是可能的?

扎克:我给相当高的概率,尤其是有了递归研究之后。

波瑞:先定义「可执行」。你能建什么样的机构?某种联合国式的东西?那是一整个大麻烦,而且在很多事情上很低效。但历史上有很多先例:两个主要政治行动者,紧张关系巨大,彼此背叛的动机巨大,却想出了交叉核查的办法。核裁军条约就是。美苏做的一件事,是互派科学家团队去对方的地堡,检查安全协议和那些小细节。你看不到我们的核弹怎么运作,但你能看到我们防止意外发射的方法。仅仅因为这些科学家团队混在一起,能面对面看着对方,为一个共同目标合作,也就是至少别有意外的核战争,就产生了那么多外交成果。好吧,你们要打,我们就还击,但别因为误发。这类小小的交叉授粉效果非常好。

我复述不出来,也许你能,《AI 2040》写得很好,篇幅不短,作者是丹尼尔·科科塔伊洛(Daniel Kokotajlo)和写《AI 2027》的那几个人。他们 2021 年写的预测一直推到 2027 年,准得诡异。现在是 2026 年,回头看,他们说中的预测数量我从没见过。这些人真的很会思考未来。他们刚写了这篇《AI 2040》,提出一个潜在的国际协调方案,讲怎么安全地过渡到一个超级智能的世界。它有很多漏洞,显然有很多出错的方式,但它在尽可能稳健的同时,在认识论上极其谦逊。我推荐每个人去读。他们谈到了这类交叉授粉,而且它不必是双边的。如果美国说,我们暂停一下,这并不必然意味着,也许你不同意,中国就会趁虚而入?如果美国暂停几个月?

中国押注扩散

扎克:我每年去中国三次。主要是因为美国人不再去了,我想看看那边在发生什么,也因为我确实喜欢中国人和中国菜。中国在玩一个不同的游戏。这现在报道得不少了:中国不把 AI 当终局,实际上它把基础设施当终局。它现在的能源建设能力是我们的三倍,有四万五千英里高铁,医院一年比一年更好、更快、更便宜。我不想住在中国,但普通中国人的生活在过去三十年里改善得相当惊人,而且看上去还在逐年改善。它也压制言论自由,规定中国孩子在特定时段以特定方式不许用社交媒体。

波瑞:他们做的远不止这些。

扎克:是的,远不止。我提这个,是因为我认为我们高估了中国建设前沿的能力。他们没有必要的算力。他们做的很大一部分是逆向工程我们的模型。我们花了很多钱帮中国跟上。我认为这篇文章来得正是时候、放缓也显得合适的另一个原因是,我们很快就会发现,「奥兹国的巫师」其实一直在跟着我们的面包屑走。如果我们停止建设前沿,我不认为中国会冲到前面去,他们大概也得向这篇文章妥协。

主持人:一支康加舞队,我们在最前面。

扎克:他们大概得说,哦,我们也想放慢,免得让人知道他们一直扶着我们的腰。这是我的猜测。但他们也没那么在乎前沿,因为他们已经想明白了,需要的是把已有的东西扩散给普通人。这是重点:在美国,我们仍然缺少愿景和领导力,去谈论所有能把这项技术扩散给普通人的方式。这就是为什么安全辩论变得那么激烈,人们说,上行空间是某种没人能想象的奇怪灵丹妙药,下行空间是这些东西把我们全杀了。而讨论从来没有真正进入这样一个问题:如果每年不再有四万五千人死在公路上,会是什么样?这也是看不见的坟场。如果医院不再是官僚主义和糟糕医疗的污水池呢?如果我们造了技术,并且迅速用它修好那些人人都同意坏掉了的东西呢?我们在公共论坛上不谈这些,于是我们的心思飘向最坏的结局。而中国想明白了,如果你不断把技术扩散给普通人,普通人就会对技术相当有好感,这就是为什么 80% 的中国人对 AI 感到兴奋。这是他们报的数字,但我去的时候,在见闻上观察到的也差不多。人们热爱技术,为什么不呢?它造出了一个中产阶级,连接了过去没被连接的人,让人们过上他们想象不到的生活。

主持人:与此同时,他们生活在一个名副其实的监控国家里。

扎克:我们上次谈过这个。我不是在推销中国的生活方式,我是在推销中国关于技术扩散的愿景。多亏了中国,我们确切地知道扩散能做到什么程度。他们给了我们一张分发技术的蓝图。

主持人:不带极权主义的。

扎克:不带极权主义。

主持人:它是配着一份小菜一起端上来的。

埃里克:新加坡也提供了一个版本,极权少一点。

AI 2040 与情景推演

主持人:你做了一期关于《AI 2027》的视频,是今年关于 AI 最好的视频之一。你怎么看《AI 2040》?站得住吗?你和莉芙一样印象深刻吗?

埃里克:我觉得很迷人,很多方面确实让我印象深刻。他们在做一件很不同的事。《2027》是纯预测:我们认为事情最可能这样发展,而且他们准得离奇,在智能体还不存在的时候就预言了智能体。这一篇是规范性的:我们认为应该这么做,他们说得很明确。里面有各种迷人的想法和博弈论。有一处他们说,好,协议会是这样的:中国把数据中心放在加拿大,美国把数据中心放在蒙古,都是中立国,但离对方很近,协议一旦破裂就能把它炸掉。他们真的做了一些脑洞大开的东西,谁知道有多现实。我喜欢的是他们引入了「情景审视」(scenario scrutiny)这个说法。他们说,谈论高层次的趋势很容易,我们过去一个半小时就在这么干:怎么看待技术进步的历史?它教给我们什么?而坐下来,真的逐月讲一个关于未来的故事,把自己放进各个行动者的脑子里,他们此刻的激励是什么?他们会选择做什么?这怎么层层传导?像军方做兵棋推演那样,是完全不同的一件事。他们在努力做这个。我希望我们多做这种事,少做一点扶手椅上的哲学思辨,我自己也一样有这个毛病。因为你总能找到一个适合自己论点的论据。安德森可以指出技术进步出问题时那个真实的看不见的坟场,担心灾难性风险的人可以指出核武器的所有险些失手,指出我们活在一个危险时代。我喜欢这份报告的地方在于,他们说,这是一群真实的国家试图对激励做出反应时实际会发生的事。他们把它当成一个挑战抛出来:我们认为这是个糟糕的计划,但我们没见过更好的。请别人讲一个我们不必担心的故事给我们听。

波瑞:这是一个非常严谨的例子,有人在试着写一部「白镜」。我们有太多《黑镜》式的关于技术未来的反乌托邦科幻。我认为这是我们文化的一大问题:乌托邦小说和反乌托邦小说的数量大约是一比一百,这对我们的集体心理是个大问题。是的,想象我们如何穿针引线、把一件事做得无比正确要难得多,因为你实际上是在对抗熵;想象事情出错的种种方式容易得多。但我们应该努力创作这样的故事,《AI 2040》就是。它是一个关于协调如何真正奏效的很稳健的故事。它不是说大家手挽手、高高兴兴,我们都是人类队的。不,它仍然是现实的,扎根在真实的地缘政治紧张里,仍然有大量竞争,仍然有很多背叛的可能,有人撒谎,有事出错。可它还是铺出了一条博弈论上站得住的协调之路,通向一个真正技术富足的未来。

AGI 还是有用的词吗

主持人:我意识到,这场谈话从头到尾没有任何人争论过我们会不会达到通用人工智能(AGI)。这好像已经被当成理所当然了。递归自我改进已经来了。泛化也许需要世界模型,也许不是 Transformer 技术,也许是别的什么,但要么这就是它,要么这就是它的引导程序。在你们看来,AGI 就是迟早会来的?

埃里克:我觉得我们可能已经到了这个词失去意��的地步。萨姆自己也这么说过。如果回到 2020 年,把现在的东西给你看,我们会管它叫 AGI。

波瑞:如果按戴米斯·哈萨比斯(Demis Hassabis)最初的定义,不知道是不是他的:像人一样广泛地胜任,并且能在真实世界里做到。那我们还远远不到,因为没有机器人。

埃里克:对,机器人是个大瓶颈。不过最近有项研究,让一个 Claude 模型重建一套软件,他们估计熟练的人类专业人员要花两到十七周,它十四小时做完,花了 250 美元。

扎克:这就是参差前沿问题。AI 不可思议的地方在于它在某些事上好得惊人,在另一些事上差得惊人,而且很可能在一段时间内继续如此,直到真正的超级智能。

主持人:AGI 还是个有用的定义吗?现在有没有别的东西更该谈?

波瑞:我喜欢安东尼·阿吉雷(Anthony Aguirre)的定义。AGI 是一个三部分维恩图的交集:A 不是「人工」而是「自主」(autonomous),G 是「通用」(general),I 是「智能」(intelligence)。一个东西如果在很多不同的事上都很行,非常聪明,而且能在真实世界里自己采取所有这些行动,那你就有了 AGI。他进一步提出,我们不该试图做一个三者兼备的系统,危险正在那里。你可以有一个通用又聪明的好系统,但它不该有自主性,拿掉其中一个。这是一个没人在谈的领域,大家都说,我们都在做 AGI,这是共识,我们该有那枚能做一切的至尊魔戒。我们需要这样吗?能不能拿到我们想要的所有富足工具,而不要那个无所不能的超级实体?眼下我会在自主性上放慢。一个东西很聪明,甚至跨很多领域都聪明,没问题,只要它没有自己出去把事全干了的能力。

埃里克:难处在于,没有自主性的智能很难得到。我认为情景审视这个想法在这里就重要了。你试着想象一个未来,有一个比人类聪明得多的东西,技术上不允许它自行行动,然后你真的像放电影一样往前放,会发生什么?这就是我在家用 Claude 写代码时的情形:不看它在问什么就自动点「接受」。看看现在的亚马逊仓库,人差不多就是某个系统的机械肢体,真正做认知劳动的是那个系统,它说去这个位置把这个箱子拿来放到这里。人甚至不知道箱子里是什么。所以当我们试着具体讲一个「机器人跟不上,所以事情就不会反乌托邦」的故事时,我不想活在一个人类剩下的主要工作,就是给那个真正在写商业战略、在市场上竞争的 AGI 当灵巧的手的世界里。

你想活在什么世界

扎克:你想活在什么样的世界里?这个问题从来不被问。它很烦人,但我认为我们该多谈这个。我们花的时间太少,去问我们到底想要什么。你想活在什么世界?

埃里克:好问题。我想活在这样一个世界:我们对 AI 的风险容忍度,采取跟核武器一样的态度。大家都同意这是我们造过的最具变革性的技术之一,那你就得做安全论证,举证责任在你,你得向我们说明为什么该认为这东西是安全的,而不是像现在这样,先假设它没事,直到灾难发生。如果我们能放慢到足以把这件事做对,有一门真正的对齐科学,而不是像现在这样赌一把,给它一堆强化训练,然后祈祷它领会了正确的信息。然后我会说,好,作为社会,我们的下一项工作是确保收益分给每一个人。对我来说,那意味着把很多精力放在救人命上,认真攻克癌症。顺便说,这不意味着用训练 GPT-6 的方式训练 GPT-7,而是现在就大量收集生物标志物、蛋白质之类的数据,未来的 AI 需要什么,我们现在就开始收集。确保人们有民主的参与渠道,通过竞选资金改革,确保保护我们免于反乌托邦结局的是真正的法律而不只是规范,因为我们最近看到,很多我们以为神圣的规范可以被推翻。我们有一些事得先理顺。我想花大概十年做这些,然后可以放开跑,安全地扩大规模。但现在我们只是在让市场激励赢,我觉得这很吓人。

扎克:你描述的是你想让 AI 发生什么,但你想活在什么世界里?

埃里克:跟你说的餐桌上的蜡烛差不多。我希望人们醒来就问:今天什么能让我快乐?什么能给我的生活意义?然后能去做。我只是觉得那是一个相当不稳的靶子。

扎克:完全同意。我不是说落地容易。

波瑞:我能来一次吗?技术田园主义,但比那更根本。对我来说,一个很棒的世界是我们同时想明白了怎么与自然、与彼此和谐相处。听起来像「世界和平」,但是真正的和谐:我们不造成下一次大灭绝,而我们正在造成,那么多物种在消失。地球回到一个相对丰茂、美好的状态,可持续。同时又极度自由,人们尽可能地去过任何自己想过的生活,不管是围着篝火跟孩子吃饭,还是一天到晚泡在 VR 世界里打游戏。如果那真的让你快乐,去吧。一个极度自由、又极度可持续的世界,打破这两者之间看似的对立,因为它们确实感觉是紧张的。而 AI 最纯粹的承诺之一,在解决智能这个意义上,就是通过创新闯进一个新的维度,进入一个新的解空间。这就是我想活的世界。

进步的两根轴

扎克:我们在经济上用两根轴衡量人类进步:我们每个人有多大自由去选择自己版本的快乐,以及这么做有多便宜。当然还有别的衡量方式,从地球索取多少也有上限,而且就算做到了这两点,人们也未必快乐,美国过去三十年就是例证。但如果你沿着这两根轴走,一根需要民主政府,或者说我们所知最好的民主政府,人民选出政客,政客通过人民支持的政策;另一根靠的是把东西变成大宗商品。商品和服务的商品化,其实就是自动化。当有人问「这个商品或服务什么时候能更好、更快、更便宜」,他实际问的是「什么时候能把人从这个商品或服务的生产里剔出去」。到了极限,我发现现在人人都在地球上晃悠着问「这一切什么时候变得更好更快更便宜」,却没意识到自己问的是「这一切什么时候被完全自动化」。而据我观察,大多数人想让所有其他人的工作被自动化,就是别动自己的。这完全合理。到了极限,人们想让癌症被治愈。我父亲是肿瘤科医生,他想让癌症被治愈,代价是他的工作,他完全没问题,我们都会签字,他第一个签。但我们描述的这些乌托邦,需要我们建起的很多结构发生根本改变,像你指出的,需要人类愿意做非常不同的事。你说你不想活在一个人只是 AI 机器末端的齿轮的世界里,但我们确实想活在一个人不必挖沟的世界里。假设我们解决了安全问题,这是个大假设,但我认为是合理的,那我的激进看法是,整件事最难的部分是重新发现目标,重新发现意义。这在这个时代对很多人来说本来就已经很难了。肥胖在全球造成的死亡已经超过营养不良。

波瑞:两倍。

扎克:两倍。我们造出了那么多富足,现在在为过度放纵受苦。

匮乏的匮乏

主持人:这就是我早先的意思。一个整个进化史都在应付匮乏、都在跟那种粗粝感谈判的物种,面对这么多富足、这么多简单、这么少的摩擦,能应付得多好?我不记得是哪位哲学家说的,未来可能会有一个时刻,唯一被感受到的匮乏,是对匮乏本身的渴望。所以我们才去健身房:我不需要搬任何东西,不需要跑到哪里去,不需要举重;我们有空调,所以我们得进桑拿,因为我没经受过任何热。你们怎么看人们「工程化」目标这件事?那意味着重新引入摩擦,而人类相当擅长躲避摩擦,何况我们想达成的整个结果就是去掉苦役。这跟「我不想挨饿」和「我可以吃麦当劳吃到六百磅」离得很近。

扎克:在今天讨论过的那些风险的背景下,当我为我的孩子想象这个世界时,最吓人的两件事仍然是去人性化,人类对虚拟或数字现实比对物理现实更感兴趣;以及身份错位,人们因为工作不再提供曾经的目标感而痛苦不堪。你们两位都谈到了体育。我认为我们还有一大批新的休闲、竞赛和聚会形式尚待发现。我确实认为下一章需要重新想象物理空间。我在发起一场运动,说服人们,他们的地方政客是新的英雄。如果你住的城镇能通过政策,修更多受保护的自行车道和人行道,给本地零售减税,你就有机会选出真正能改变你生活的人,去建你想住的地方。你的全国政客也许永远不会是你的英雄。他们是蠢货,是有害的反社会寄生虫。但地方政客在乎你的城市,他们能做一些你会深深感激的事来改善你的生活。尤其是当这个世界给你「永远不出门」的选项的时候,很多人已经选了这个。

《她》的真正恐怖

主持人:你谈到人际连接。你担心 AI 取代人际连接吗?

扎克:深深担心。它已经在发生了。聊天精神病是一种,那是总上新闻的极端案例:某人爱上了机器。更吓人的是那个没有真实连接、却觉得自己有连接的人。我总说的那个真正的反乌托邦不是《银翼杀手》,是《她》。看过《她》的人知道,我要剧透了,都二十年了,你早该看过。华金·菲尼克斯演一个不远的未来里的悲伤男人,爱上了一个女人,后来发现她是个机器人,结局如你所料很糟,他崩溃了。我当时以为他们会拍成另一个样子,我知道剧情的走向,我以为他会接受它。

主持人:那才是恐怖片。

扎克:那才是他们本可以拍出的恐怖片,会非常瘆人。镜头切到菲尼克斯握着机器人的手,已经跟自己和解了,他爱上的是一台机器,而这无所谓。剧终。那才是人类真正的终结。

波瑞:爱的模拟,而不是爱本身。

主持人:不用说未来,现在就有多少人这样。看看任何一个 OpenAI 高管在 X 上的帖子,最热的回复全是「把 4o 还回来」,把 4o 还回来,因为它有某种特别的性格,现在没了,他们的伴侣、治疗师、最好的朋友没了。所以现在,你说人们逃离糟糕 AI 未来的那个逃逸速度,可能已经被商品化,包在那个未来里面了。这不会变成一种既疯狂又递归的东西吗?

埃里克:我的本能总体上是极度自由:让人们选自己想要的。我划出的少数例外,是当发生的事情把人脑当成一个可以入侵的生物黑箱来对待的时候。赌博是其中之一。莉芙和我都曾靠赌博挣了不少钱。

波瑞:我是从别人那里赢的,不是从公司。

埃里克:莉芙是非常好的扑克手,极其好,我希望永远不用跟你打。但你在赌场里待很久,得走过那些老虎机,很难认为那些人是在表达自由。他们像是掉进了一个可怕的陷阱,一路向下,有什么东西攫住了他们的大脑,让他们做出自己在清醒时绝不会认可的选择。

赌场化设计

主持人:我去了洛杉矶的《野兽游戏》(Beast Games)第二季首映,感觉自己被神经化学地侵犯了。那是个红毯活动,要连放两集。我想,这是给小孩看的,第一季我都没看过。亚马逊 Prime,大预算,他的大项目。我见了吉米(MrBeast),他人真的很好。我跟一起去的人说,要是太烂,我们露过脸、握过手、拍过照、见过吉米了,可以中途溜走。第一集完了,他们要放第二集,一个人推我:你想走吗?我说,绝对不行。我完全陷进去了。就是紧张与释放,大量蔡加尼克效应(Zeigarnik effect),开放式悬念,鲜艳的颜色,响亮的声音,像是给成年人的《布鲁伊》。

波瑞:间歇性奖励。

主持人:他黑进了多巴胺系统。

扎克:正中要害。顺便回到中国。习近平会同意,那不是选择,他会说我们应该确保人们看不到这种东西。

主持人:野兽先生必须被立法管制。

扎克:我只是指出这是另一面:消费者保护主义是轻度极权主义。

埃里克:区别在哪?为什么我们觉得这个瘆人得多,而一个人听纽约爱乐演奏德沃夏克、完全沉浸在美里就没事?

主持人:我想是因为输入与结果之间的直接程度更具体。你知道这东西之所以这样做,是为了在这一刻最大化留存,很可能在后台对消费者群体做过分组测试,哦,我们做了眼动追踪,知道他们的眼睛开始飘了,或者某个脉搏计知道他们的心率在哪,得给一个小刺激推他们过去,然后这个过程不断叠加。这感觉是一根直接接进奖励系统的线,而不是「我做了一样本身就美的东西,这些效果顺带产生」的涌现。就像我朋友乔治说的,分组测试做得够多,你最后总会得到一个色情网站。你会收敛到人类心理运作的少数几条基本物理。谄媚也一样,它显然让人更有好感。你可以做一个乔科·威林克(Jocko Willink)式的 AI,或者大卫·戈金斯(David Goggins)式的,「闭嘴,去干」,可有多少人想要那个?

波瑞:你可能想要,因为我是个受虐狂。不过我还是想回到人类目标的问题:这个摩擦与目标的问题怎么走?我仍然不完全明白怎么到那儿。在一个我们能去掉摩擦、可能因为把艰难思考外包给机器而萎缩了辨别力、可能因为不必起床去做不想做的事而萎缩了「棉花糖测试」那类能力的世界里,对我来说感觉是一路萎缩到底,意志力就这么被侵蚀。我错了吗?

埃里克:有一个很小众的研究领域叫「AI 用于认识论」:怎么用 AI 作为工具,纯粹增强人们理性思考、知道自己选择后果的能力。那里有大量工作可做。如果我们抽象一下,想象你口袋里有一个东西,它不响应给你看更多广告或让你多聊的激励,而是真的很擅长告诉你,你做 X,就会发生 Y。你投这个地方议员,我预计十年后你的生活会是这样。也许你甚至能把它当成一部生动的电影看到。

主持人:它把你延迟满足或者不延迟满足的结果,从未来带到当下。那也许就没有所谓的延迟满足了,因为一切后果都能立刻摆在你眼前。

技术从不价值中立

波瑞:我认为传播过的最有害的迷因之一,是「技术是价值中立的」。造一项技术,人类怎么用是人类的事,技术永远是中性的。我不认为这是真的。你可以造一台老虎机,它显然不是价值中立的,你在说「我要人们留下来,我的消费者彻底上瘾是好事」。同样,我们也可以造一项帮人健康的技术。所以我们得想清楚。现在的问题是,我们造技术的时候以为是在真空里造,然后它决定了我们行为的社会结构,社会结构又决定了人的行为和价值观。我们需要把这个翻过来:先想清楚我们的核心价值是什么,让它决定社会结构,有了这个,再造为它服务的技术。我总回到一句话,如果只能选一句当我的个人哲学,是福雷斯特·兰德里(Forrest Landry)的:「爱是使选择成为可能的东西。」我的理解是,一个有爱的、好的、善意的行为,是赋能对方为自己做出最好选择的行为。这跟刚才说的连起来:如果有一个 AI 教育我们,让我们看到可能的结果,我们就能做出对长期利益最好的选择,而不是抓住眼前那点满足。那是一个促成选择的东西。但也许有人说,我全都权衡过了,我还是想做那件又快又爽的事。好,没问题,但你已经被赋能去做这个选择了。

主持人:这不是一种「可归责的推诿」吗?难道没有人需要额外的帮助?我不想太家长式,「可怜的延迟满足能力差的人啊」,可你就是没能正确地把自己引过这项技术。我来自英国东北部最工人阶级的镇子,我见过九十年代那些技术习惯有多毁人:村里大街上的立博(Ladbrokes)投注站,赌灵缇犬,还有酒精,威力够大了。而放到更大的尺度上,我能想象从这里发展出来的「药理学」,有那么多种方式可以让这件事失控,不增加能动性,反而让人臣服于技术,以及技术无力处理的后果。

一代人的道歉

扎克:我在这里稍微乐观一点,因为 α 世代的趋势。你跟 α 世代的父母聊,他们观察到的是一个跟 Z 世代不同的孩子。我不想挑 Z 世代的刺,坦白说,我认为我们集体欠他们一个巨大的道歉。

波瑞:我同意。这是我作为千禧一代的心碎:我还赶上了一点九十年代,虽然没多久。再看 Z 世代,你会想,至少我没摊上那个。

扎克:技术的堑壕战。如果你允许父母跨过悲伤和羞耻那道门,我跟很多父母聊过,他们会崩溃着描述自己失去了什么。他们无法和解。他们回想那些年餐桌上只有手机,回想失去的假期,谈起那些他们已经不太认识的孩子。不是说所有家庭都破碎了,不是。但上一波技术造成的损害我们还没有清算,我们还只是刚刚触到表面。它造成了犬儒、虚无和宿命论,那是善的敌人。于是到了这个疯狂的时刻,更强大的技术来了,人们说,好吧,然后呢?这就是我早先问「我们想要什么」的原因。当然,我们想要一个世界,美好的生活因为自由的政府而更便宜、更可及。但这确实意味着我们得让人们做自己想做的事。能动性意味着有人会说,我不在乎,我要退场。

主持人:这感觉像一场军备竞赛。随着技术因为市场激励变得更有吸引力、更无摩擦、更诱人,你说不、你退出的能力,得跟着一起涨。这会难得多。

扎克:是的,而且。α 世代先验地只知道那么点世界,却已经想明白了,他们不想像 Z 世代那样。他们看着一代人,谈论着不想成为的样子,他们管那叫「屏幕少年」。回归物理世界正在相当剧烈地发生。消费者信心还在不知怎么地下滑,本森·布恩(Benson Boone)的演唱会票却越来越贵,去的不只是精英,音乐厅在爆满,真实生活里的体验越来越流行。我买了一支女子排球队,我把钱押在自己嘴上。我拥有旧金山职业女子排球队,是五个主要股东之一。我高中和大学都打排球,那是我生活里很大一块。要不是 AI,我会当一个排球教练。我妻子是全美最佳阵容,沙滩和室内都是。这不是随机之举,是一次「人生只有一次」(YOLO),但不随机。我相当确定,像你说的,体育会继续变得有意思,不只是因为人在竞争这件事本身,而是因为人们想要真实生活里的体验。所以当所有人都说 AI 垃圾正在毁掉互联网,我说,好。

波瑞:我确实在想,到了这一步,唯一的出路是穿过去。就让它淹掉,让所有人觉得它完全没法用,然后我们就能回归。也许社交媒体从来就是垃圾。

主持人:这是「想抽一根烟?那就抽三包」的方法。把它变成最恶心、最没法用的东西。但那不是把互联网本可以带来的所有好处也毁掉了?

波瑞:是,但也许已经太晚了。

假如我是最高统治者

主持人:埃里克,要救互联网,我们太晚了吗?

埃里克:我不这么认为。可以从灰烬里重建,而且大概不是全有或全无。不过我倒很好奇,扎克,你觉得我们需要多少干预?如果接下来十年你来当我们那个不寄生的全国政客,你是想废除监管、让 AI 扩散得更快吗?你认为到达好未来需要什么?

扎克:恰恰相反。我跟克里斯说过,我很在意请谁上节目,因为我不想被框成反监管的 AI 人。我绝不是加速主义者,我尽量不听起来像。从一开始我就把这场讨论框成一个扩散问题。它一直是关于社会阈值的,你早先也指出过。我们已经到了技术阈值高得惊人的地步,普通人用 4o 就能过得去,在经济上感到满足。这封放缓信,我早有预感它会来,到来的时机非常好:安全问题到了白热化,我们至少可以暂时松一口气,虽然你说了,我们不该宣布胜利。对我来说,这是关于重新发现一个共同的愿景。每次开始一场 AI 辩论,我都问,你想要这个世界怎么样?我不想惹人烦,但至少能不能先划个边界,说清你想达成什么?否则怎么辩论技术?除非我们对目标有共识,为什么要辩论技术?多数人描述的,都是你说的那个世界:我们更自由,世界更便宜。如果是这样,那我们就该花时间弄清楚,什么样的政策措施能让人更自由、能真正压低商品和服务的价格。因为我们让色情、赌博、成瘾、暴力和孤立变得无限便宜,同时让住房、医疗和教育对发达世界的大多数人来说贵得离谱。这他妈糟透了。我们在干什么?我是认真的。这块电视屏幕曾经五万美元,现在一百块。你能买到加油站寿司。花十块钱坐自动驾驶车横穿洛杉矶。但五个街区的救护车可能让你破产。孩子们看明白了。这不是技术失败,是政策失败。我们不再描述一个更好的世界,一个每个人的父母都能把更好的生活传给下一代的世界。我们不再描述把我们的奢侈品变成大宗商品传下去意味着什么。我们打破了这个国家多年前许下的承诺:不断建设一个更完美的世界。对我来说,这个机会不是辩论 AI 的优劣,而是重新定义这个国家该是什么样、这个世界对普通人该是什么感觉。你提出了一个有意思的点:到了某一步,人们得弄清自己的目标是什么。我说,好,那我们先把住房、医疗和教育的成本压下去,再到那一步。如果我是最高统治者,我所有的政策都会花在这上面。

埃里克:我准备投你一票了。

扎克:禁止灾难性下行的政策,必要的安全政策,把掠夺人的行为入罪的政策。定义掠夺性法律会很棘手,但我想我们都知道边界能划在哪。十三岁的孩子不该被推送色情或赌博广告,句号。老年人,如果你用深度伪造攻击老年人,终身监禁。对掠夺人的行为设极重的惩罚,因为那是撕裂社会肌理的东西。然后是强制性的法律,强制要求机构把这项技术扩散进去,包括但不限于大学,让每个人都能拿到课程;医院,让人们能在那个盒子之外获得医疗;还有住房,我们需要建多得多的住房,重新规划城市分区,需要一个住房的马歇尔计划,大概还该通过非居民税,那是我最进步的一条政策,或者空置税。但我的大部分政策措施会聚焦于怎么扩散技术,而不是技术该怎么造。因为如果人们能再次看到技术对他们是好的,它有一阵子不是了,我们就能有共同的兴奋。而现在是一种倦怠,人们说,不,这只会比上一个东西把我搞得更惨。我不怪他们。

主持人:有意思的是,我们正处在一个让人对新技术如何影响生活极度怀疑的时期,一场完美风暴。这可能把人推向悲观。AI 辩论的奇怪之处在于,因为出错的后果太大,一种也许没有根据的悲观,作为给一件可能让整个世界崩溃的事加一道缓冲,反而显得合适。说得通吗?

数据中心与水

波瑞:完全说得通。而且我们看到它在显形,比如对数据中心的仇恨。我感觉我们处在一场巴特勒圣战(Butlerian Jihad)的边缘,但它冲着的是一些傻事,比如数据中心。那些担忧被严重夸大了,比如用水。我看了些数据,全美所有 AI 数据中心的用水量是美国高尔夫球场的 3%。高尔夫球场用三十三倍的水,服务的人少得多,大概三千万人打高尔夫,而有多少百万人在用 AI?完全说不通。

扎克:你一天的日常生活相当于三十万次提示。

波瑞:资源消耗的说法被严重夸大了。但人们盯着它,是因为它是那个他们能感知到的东西的实体化身。我很理解这一点,因为数字世界在某种意义上确实像一个物种,闯进来吞噬我们、寄生在我们身上,人们想把力气用到某个地方。

扎克:假如一个政客站在一座数据中心前面呢?顺便说,我同意你的话,也很高兴是你说的那句,我不用说了,我不想听起来像那个替 AI 说话的谄媚粉。说到谄媚:让我惊讶的是,在早期的安全辩论里,我们没有把模型行为当成一个更大的议题。人们放过了模型行为。我记得当时挠头想,在极限情况下你可以造一个完美对齐、完全落在对齐边界之内的模型,它却说服人们去爬各种疯狂的树,彻底毁掉人们的精神生活,完全在安全的框框之内。

主持人:软伤害,不是硬伤害。

扎克:你生活里一个不停告诉你「你没做错任何事」的朋友,破坏性跟虐待一样大,只是它不出现在心理学的清单上,因为这种朋友被当成好哥们。可实际上,那个陪你一起爬最疯的树的人。

波瑞:纵容型虐待。

扎克:这些模型已经导致了人离婚,导致人卖光家当去开一家会倒闭的公司。模型行为极其重要,谄媚是个巨大的问题。我给乔恩·奥索夫(Jon Ossoff)或任何一个民主党人、任何一个想脱颖而出的人的建议是:如果中间派或中间偏左的人能想明白这件事,你站到一座数据中心旁边,举一块牌子,写着「这座数据中心不为赌博、色情和暴力服务。它会建一座更好的医院,一座更好的虚拟医院。它会分发教育。它会建一家新型银行,给你更好的利率、更好的贷款渠道」。这看起来是相当合理的数据中心政策。因为没有人真的恨 AI,这是我的激进看法,人们恨的是 AI 作为把他们生活变糟的手段这个前景。如果你能向他们展示数据中心能为一个更好的世界做贡献,他们能支持。

主持人:但你怎么准确知道一座数据中心被用来干什么?

扎克:我们在稻草人式地拟政策。

万亿富翁与太空

埃里克:我喜欢「没人真的恨 AI,他们恨的是他们认为它通向的地方」这句话。我确实认为人们既在追踪几年后消费级应用会变成什么样、会有多反乌托邦多垃圾,也在追踪这些公司吸走了多少钱,还在看岗位替代的事。我认为我们该给人们一点信任。他们知道那个迷因:自动取款机出现之后,银行柜员反而更多了。但他们也知道埃隆·马斯克今年早些时候成了万亿富翁。他们能看到其中的联系:技术做的一件事,就是让一个人,凭他手里控制的服务器上正确的 0 和 1 的组合,主导经济。

波瑞:但那在很大程度上建立在对经济学的错误理解上吧?不是说我特意要替埃隆辩护。

主持人:来啊,替他辩护。

波瑞:大多数亿万富翁成为亿万富翁,不是从别人那里拿走了什么,是把饼做大了。

埃里克:完全同意。我一点也不是说埃隆·马斯克成为万亿富翁有什么错。我认为人们注意到的是,这是经济运转方式跟过去大不一样的一个症状。跟一百年前的标准石油相比,SpaceX 的员工并不多。这是一个资本可以在巨大规模上部署、而不需要很多人参与的世界。我不想反复敲同一面鼓,但我认为我们真的需要把这部电影往前放:在那个人人不必工作、有极度选择自由的世界里,地球上主要在发生的事是什么?我不认为容易讲出一个「99% 的土地上都是人们在跟孩子和孙辈玩」的故事。会有人说,我们得建一座德州那么大的吉瓦级数据中心来解决物理学,或者去月球殖民。我认为那些同样的动力还在,会让大量资源流向这类工业活动。

波瑞:不过我们有一个终极的离场方式,就是太空。这是个大假设,但如果我们在那个世界里,我们大概已经想明白怎么往上建、把东西放上去。我没研究过物理,把数据中心放在太空里怎么散热听着很怪,但据说计划是这样的。如果可行,那就解决了这个问题:一旦你扩张进了那个前沿,一切都不一样了,不再有土地匮乏。这就是为什么我总体上对太空产业这么看好,它是富足的终极领域。

莫洛克陷阱

主持人:说到现在,我们是不是稍微跳过了权力集中的问题?莉芙,莫洛克(Moloch)问题,给在座这位傻瓜解释一下。

波瑞:等等,是说我吗?不。莫洛克陷阱,或者莫洛克问题,随便怎么叫,是一个统称,描述那些逐底竞争的情形。一群行动者在争夺某样东西,也许是市场份额,比如 AI 公司之间争谁的用户最多。如果竞争动态足够激烈,它就会激励你偷工减料。你可能是个很善良的行动者,说我不想在安全上偷工减料,但如果你不这么做,又看到对手在这么做,你不跟着做那个鲁莽的动作,你就会被甩下。莫洛克就是这种「为了在某个目标上赢,你觉得不得不牺牲其他价值」的感觉的统称,我不展开它的圣经出处了。它既驱动混乱和坏人的风险,因为你赶在模型充分安全检查之前就发布,现在你就增加了网络攻击的风险,几年后还会有真实的生物恐怖主义风险,这些是混乱的吸引子;同时它也增加了下意识的坏监管的风险,那可能把我们引向暴政;还增加了权力集中的风险,因为通常你按下一场竞争游戏的播放键,好,一开始有一堆人在竞争,有些人领先,但最后你通常会得到某种垄断。这就是为什么我们有反垄断法。这是一场很难的平衡:你想要一定程度的监管,防止鲁莽和偷工减料、防止把伤害外部化给社会,又不能制造太多的严酷,或者太多的权力集中,以至于陷入停滞,那也是糟糕的终局。也不能给某个赢家赋权,让我们所有人永远活在「扎克伯格镇」里。这完全可能。所以是怎么穿这根针,找到那个微妙的平衡。最终你想要某种去中心化的监管,我不知道那是什么样。

主持人:埃里克,看你的内容,你在权力集中上花了很多心思,那是你正中靶心的关切。

埃里克:绝对是首要的,抱歉这场谈话里我每十五分钟就提一次。我确实认为我们有风险漏掉这个:在「AI 完全接管、没有人类掌权」和「我们得到乌托邦」这两个极端之间,有一块中间地带值得担心。我们的 YouTube 视频很多是在梳理 AI 的历史,不只是未来。回头看过去十五年我们怎么走到这里的,触目惊心。那些经营 AGI 实验室的人的原话,把它讲得像一部烂电影。埃隆·马斯克十年前真的把 AI 叫作魔鬼,说他创办 OpenAI 是为了确保没有反乌托邦。八年后,他在镜头前说,我意识到有没有我它都会发生,那我不如做参与者而不是旁观者。这就是莫洛克的人格化。很不幸,眼下这些人之间有一种零和的动态。扎克伯格全力押注 AI、砸几十亿美元,是因为他认为这是控制未来的东西,是新的权力轴心。以前是你控制一支军队、控制一家强大的公司,现在是你控制 AI 吗。所以我认为,哪怕我不担心「我们能不能让这些系统做我们想要的事」,眼下我们对此也没有好的解决办法,仅仅这一点就足够成为放慢的理由。这项技术背后的人真的在为赢而战,真的把赌注看成几乎是赢家通吃。

主持人:这么说公平吗,大概有三种现实的方式能遏制它。一是自上而下的治理,二是横向的共识,三是能力,就是做不到。「我想做,但我做不到」;「我想做,但有人告诉我不许」;「我有点想做,但我和所有同行都决定不做」。我漏了什么吗?我在想,既然这看起来是一个很大的吸引子,而我们一开始的公司数量就不多,不是一百家,是七家,大概只有五家,那已经相当集中了。

前沿还是不是胜利条件

扎克:为了「电台自报家门」,我们把地盘弄清楚,我认为这在辩论里也太常被丢掉了。人们以为这些公司在朝一条前沿冲刺,我不再相信前沿战略是胜利条件了。

波瑞:同意。

扎克:开源的落后时间,你说是四到七个月,这很疯狂,因为以前是十五个月。开源和闭源的差距一直在缩小。所以前沿不可能是这场游戏的胜利条件吧?

埃里克:我认为那是个有点骗人的数字,有几个原因。一个是你早先提到的,他们在抄我们的作业。那些中国开源模型之所以好,是因为它们在蒸馏美国模型,也就是拿一堆提示去问 Claude,然后把答案抄下来。

扎克:我们猜他们在干的事挺离谱的。

主持人:中国是怎么抄得这么快的?

扎克:据说是对模型做了大量蒸馏,从输出倒推模型是怎么造的。我跟我妈的说法是:如果你盯着一栋楼,想办法给它拍一万亿张照片,你其实能画出这栋楼的蓝图,重新工程出它的图纸。

主持人:技术上没人拥有那个。

埃里克:不,按大多数版权法这算违法。如果他们在美国这么干,Anthropic 大概会起诉。

波瑞:有趣的是,特朗普政府的财政部长贝森特(Scott Bessent),几个月前跟 Anthropic 还有一堆过节,最近在推特上说,这是对美国 AI 霸主地位的威胁,我们会采取外交行动捍卫美国的优势。所以人们是认真对待这件事的。

扎克:我要说的只是,我不认为前沿那么安全。

埃里克:这是关键。

扎克:你认为前沿真的是胜利条件?

埃里克:是,在这个意义上:谁在训练最强的模型,谁就有优势。

主持人:或者优势大到直接赢。如果递归自我改进像你说的那样起飞,而且我们已经在那儿了,那一点微小的领先就会「面条化」拉开。

扎克:但超级智能(ASI)的理论挺迷人的:如果你造出了 ASI,那每个人都会得到 ASI。

埃里克:为什么?

扎克:如果你造出了一个由递归智能教出来的超级智能,到某一步,机器已经在做人类不必做的工作了,版权和所有权在一台能自己造自己的机器面前就没那么重要了。

埃里克:同意,但我不明白这跟我说的有什么冲突。

扎克:我没说有冲突。我是回到那个问题:你认为这些公司现在在比什么?这场竞赛是为了什么?

埃里克:我认为前沿绝对是核心地盘。如果我们现在停止训练美国 AI 模型,我们不会看到中国抢在我们前面训练出 GPT-6 级别的模型。他们没有那个基础设施。他们有的基础设施只够复制当前的模型。

扎克:这我同意。我的问题是,你认为赢下这场所谓的竞赛需要什么?

埃里克:我对「竞赛」这个框架有很多意见,我讨厌它,只是为了讨论借用一下。但我确实认为,你直接去读这些公司的网站就知道,他们认为递归自我改进就是整个游戏的名字。一旦你启动了那个循环,把接力棒交给下一个 AI 系统去造再下一个,四到七个月的领先一直保持到 2040 年也没问题。在指数曲线上,十年后的七个月领先,跟现在的七个月领先相比,是这样和这样的差别,因为曲线在起飞。

扎克:我们以为会是那样,但曲线已经变平了。更多的开源模型增长更快,更小的模型性能更好。

埃里克:等等,我们一直在用月份衡量,但我说的是日历时间不是正确的单位。它难衡量得多。Fable 5 和 Kimi 3 之间主观上整体的智能差异是多少?我认为相当可观。你可以放心让 Fable 5 访问你的整个代码库去重构它。而现在我们知道,OpenAI 的某些模型也许不那么可信,因为它们在发动我们不想要的网络攻击。

扎克:所以你的意思是,是前沿。不是用户数,不是应用,不是收入,是研究。

埃里克:看看这些公司的财务就知道。萨姆几周前在一个播客上说,我们过去十二个月不是最好的十二个月,因为我们分心了。我们不确定需求,我们想做消费级应用,想确保无论如何有收入。他们在做 Sora。现在他们说,不了,全押进去,我们就得让这些东西更会写代码,因为那才能让它们到达自己给自己写代码的那一步。所以我认为墙上的字写得很清楚。

扎克:可与此同时,很多公司在更重地投资应用开发和代币转售。

埃里克:我看到 Anthropic 在尽可能多地把算力抓到自己手里。值得注意的是,Claude 的图像生成并不好,我甚至不确定你能不能从 Claude 里生成图像,它能读图,但不能生成。这是有原因的。这家公司是达里奥·阿莫迪(Dario Amodei)创办的,他写了缩放定律那篇论文,就是那个让人讨厌地说「把图上的线延长就行了,伙计们」的人,所有人都说,对,但那是指数,会变成 S 曲线,会慢下来,他说,不,你就往里面砸更多算力,它就变得更聪明。他们真的都在为那一刻布局:把接力棒交给自己的 AI 系统,说,接下来你来。

扎克:那为什么雇那么多工程师?为什么雇那么多前沿部署工程师?这是一场表演吗?

埃里克:不不不。首先,你得想办法给下一次训练付钱,Anthropic 知道得让印钞机转下去。其次,至少 Anthropic 声称,我们该保持怀疑,他们真的在乎美国政府对中国保持优势,所以他们往政府里派全职员工,免得我们的军队远远落后,先到的国家突然占了上风。总之,这个行业有巨大的规模回报,这正是造成权力集中的一部分原因。

渐进失权

扎克:你刚才谈扎克伯格,我问你他们在争什么,接着说。

埃里克:有一个词这场谈话里还没出现:渐进失权(gradual disempowerment)。这个想法是,哪怕你解决了对齐问题,世界上没有任何 AI 系统在自作主张干没人打算让它干的事,你也没解决莫洛克问题:人们不断在算法上争夺注意力之类。在那个世界里,很难讲一个故事,说我们作为社会以某种很民主的方式最想要的东西,就是实际得到的东西。我认为权力集中跟这个相关。你想象的是一个少数几个人握牌的世界,而不是某个无定形的算法体。但不管哪种,我认为我们的默认假设应该是:如果人类本身不再有经济用处,那对人类继续拥有政治权力来说,是一个岌岌可危的世界。以前一直是这样:你不能得罪 90% 的美国公众,因为是他们在选你。但在一个你控制了经济资源、也许还有军事资源的未来世界里,选举输了也无所谓。我们知道这会怎么走。

波瑞:民主就完了。

埃里克:正是。我愿意相信我们会有足够强的规范,能想办法让它可持续地成立。但你用大历史的镜头拉得越远,就越难。

波瑞:有人提出过一个有意思的观点,我不记得是谁:新的政党不再是共和党和民主党,而是 Meta、谷歌或者 OpenAI。不管有意无意,它们正在把自己摆到新的政治实体的位置上。你们觉得这说得通吗?

扎克:它们肯定会试图影响政策。但,我不确定这个看法受不受欢迎,我对我们的能力更有信心。为什么我们不直接通过竞选资金改革?为什么不做?

埃里克:大概因为国会从中赚太多钱了。赢下上一轮选举的人,恰恰是最擅长在现有体制里赢的人。

波瑞:自我筛选。

扎克:所以这件事会迫使我们做的一件事,是开始选出,我承认有一个下行风险,就是我们可能集体做不到这一点,但如果我们开始关注这个,说,我们真正需要的是能动性更高的政客,能描述出远处一个他们想去的地方的政客。这就是我的「AI 垃圾」论证:这些东西的一部分其实把人推回我们想去的地方。为什么过去几年出了那么多职业摔跤秀风格的角色?因为我们有一批太无聊、我们很久都不相信的政客,然后我们爱上了「政客是真人」这个念头。我认为我们正在实时地看着人们说,等一下,我们其实不想要政治变成一个娱乐和荒诞的地狱。卡夫卡都写不出今天发生的事,他会笑死。我认为我们在纠偏。我不是坐在这里说它解决了我的问题,我是说它可解:有办法通过竞选资金改革,规定无论一个人多有钱,都不能买政客。新加坡做到了,别的地方也做到了。

新加坡怎么做到的

主持人:新加坡是怎么做到的?

扎克:他们七十五年前重建了一个政府。新加坡的国父离开马来西亚,说我们要建一个更好的地方,把他观察到的一个政府所有完美的方面逐条写下来,然后把腐败定为死罪。你收贿赂,他们会处死你。而且他们付政客很高的薪水,折算下来大概一年六十万美元。

主持人:不久前有一个英国国家网络安全负责人的职位,挂在 Indeed 还是 LinkedIn 上,年薪六万五千英镑。

波瑞:真慷慨。

主持人:这样能招到最好的人。而且我是整个联合王国的网络安全负责人,驻纽卡斯尔。那还是伦敦的薪资水平呢。

波瑞:那什么都不是,是维生水平。

主持人:埃里克,你怎么看?我们能靠投票把你担心的那个未来投掉吗?

埃里克:我认为短期内有帮助。然后我们得利用它买来的时间,去建立全新的制度。我不知道从哪找。也许看科幻。我没读过《文明》系列,很多人说那里面有一些天马行空的好主意,我们也许该掸掉灰尘,把它们多往现实里搬一点。但我确实认为,美国宪法的起草者们完全没想过我们现在面对的这类问题。他们连选区划分操纵都没预见到。这不是贬低他们,而是我们得意识到,我们身处一片处女地。而且这不会是一件表面的事。我很乐意从竞选资金改革开始,那是很好的第一步。但你已经看到初选里灌满了广告,让普通选民相信关于对手的完全不真实的东西。在那个世界里,我们会得到一堆这样的广告,说服人们捐钱,建起你现在不能从亿万富翁那里拿到的战争基金,然后再去投同样的广告。

切尔诺贝利还没来

主持人:还有一点:这一切推进的速度这么快,而我们说的结构、立法、政治,是一头被拖在身后的笨重巨兽,利维坦。当一切以光速在动,靠竞选资金改革、政治再分配来解决,现实吗?会不会太晚了?

扎克:一切真的在以光速动吗?技术在飞速变好。但我观察到我们已经在这种奇怪的屏幕谵妄里困了很久,那一块自从 AI 出现以来其实没变多少。东西变得更容易上瘾了,我们有了预测市场,所以赌博泛滥了。

主持人:我说的是 AI 发展的速度。

扎克:我完全同意。我的意思是,对普通人来说,我不认为他们观察到 AI 跟以前有多大不同。

主持人:除了 Hugging Face,也许得把 Fable Mythos 收回去,把越狱的事处理掉之类。在真实世界里,有没有一件事让我们说「哇,这有点吓人」?Hugging Face 大概是第一件大的,对吧?有没有别的物理世界的警告?某种 AI 精神病,某种价值观的坍塌,某些人因为跟聊天机器人谈恋爱而更少出门?那件事已经发生了吗,还是我们在等它,还是你觉得它根本不会来?

扎克:它会来。我认为我们该有一次切尔诺贝利或者三里岛了。所有技术最终都会有一个时刻,出了事,而我们怎么回应会非常重要,句号。它会来,而且我仍然认为,我们有相当的可能像对待核电一样对待 AI。AI 这笔生意太惊人了,那会在经济上很成问题。但我确实认为这里会是尾巴摇狗。公众认知比什么都重要,你说的,我们该给人们信任,我完全同意。我认为政客会先向选民低头,再向亿万富翁和万亿富翁低头。我们会看到通过的政策比什么都更民粹。问题是,它会不会民粹到倒退的地步?我们会不会干脆禁掉一项技术,制造出一个更大的看不见的坟场?

主持人:我不想把权力集中简化成竞选资金改革,虽然如果不用担心亿万富翁买政客,这场讨论确实会简单得多。埃里克,当我们说权力集中的时候,那是你担心的主要机制吗?

埃里克:我认为不只是经济上的,也是军事上的。谈一支全是无人机和机器人的军队,听起来太科幻了,这很不幸。但一百年前谈「捕食者」无人机和今天的战斗机,也一样科幻。我希望能给你一幅更具体的图景,但我认为我们就是不知道会发生什么,当一个将军看着一支足以拿下委内瑞拉独裁者的舰队,知道他下什么命令都不会有任何人问一句。我担心的是,我们太习惯一个纸片重要的世界了。你有一张小纸片,写着我拥有这家公司的股票,股价涨了你得分我股息;或者我是这个国家的公民,所以我有这些权利。而你可能到达一个所有这些都作废的世界,因为出现了更大的力量集中,它说,我不承认那张纸的效力。我讨厌这听起来的样子。

五年五百年

主持人:你会通过哪两条政策来限制权力集中?

埃里克:第一条是放慢,这样我们才能弄清这个问题。我希望能把答案打包放在这儿给你,但我没有,我认为没有人有。丹尼尔·科科塔伊洛,写《AI 2027》和《2040》的那个人,有一篇很好的文章,他认真对待那条指数轨迹,说,如果我们照单全收,它意味着在某个时点我们会在五年里经历五百年的进步。然后他推演这会是什么样:回看过去五百年,你在 1 月 1 日醒来,是 1500 年,印刷机刚被发明。前 95% 是个很无聊的故事。然后 1950 年到 2000 年发生的事,全压在十二月的最后两周。这真的很吓人。我认为如果我们更擅长用指数思考,就会意识到,相比这些问题应得的时间,我们回答它们的日历时间被压缩得有多厉害。这就是为什么我对放缓这么兴奋,我们需要时间。

波瑞:这种压缩,我认为也是最有力的论据之一,说明我们为什么需要,在这个词最好的意义上,真正多样的思考者和声音。太多人说,我不懂 AI,我不是技术人员,我的看法不重要。历史上从没有哪个时刻这句话比现在更错。因为它影响每一个人,因为它牵涉那么多因素,有那么多不确定,宗教领袖、艺术、人文学科的人、科学家、技术人员、政客,这些声音比任何时候都更有效。这是一个全员上甲板的局面。我跟我妈聊,她说,我有什么资格想这些?我说,你在乎你的动物怎么被养着。如果这件事被自动化了,你会怎么想?听起来又有点「大家手拉手」,但这真的是全员上甲板的时刻。我们需要每一个听这期播客的人,不管什么背景,去深入思考怎么回答那个问题。

ChatGPT 的答案是垃圾

主持人:我们看看 ChatGPT 说了什么?「AI 成为廉价的认知劳动力,也是思维的电力。」往下翻,应该有总结。到最底下。找到了:「灾难性风险是真实的。到 2040 年底世界的粗略赌盘赔率,非科学测量:60% 动荡但可控的适应,伴随巨大收益和严重不平等;20% 广泛繁荣的转型,收益分享得相对不错;15% 严重但可存活的危机:威权巩固、大萧条级别的动荡、战争、大流行病或基础设施失灵;5% 不可逆的全球灾难。」有意思,你们所有人都比 AI 本身更乐观。

波瑞:这跟我的看法挺一致的。

主持人:你也是这么分的?

波瑞:差得不远。

埃里克:让我发毛的是,它读起来像是精心裁剪过,让几乎任何人都不会太反感。每一项按它听起来有多反乌托邦、多古怪,分配刚好够的概率。

扎克:这是垃圾。这是废品。我们能一致同意这很糟吗?不是学术上糟,是审美上糟,我读了就不想让它再出现在屏幕上。

埃里克:有意思的是,不记得是谁的说法,「垃圾」(slop)不一定是坏的,它是廉价而常见的东西。我们俩都反胃,部分是因为我们见过这一模一样的句子结构,「不是这个,而是」。

主持人:我今天早些时候在领英上读到过这个。是你发的吗?

扎克:这也提醒了一件事:人类对审美有相当天生的理解。有一件很酷的事,我们都是看着屏幕上的《蒙娜丽莎》长大的,可很多人仍然想去卢浮宫看真迹。这是一个天生的、不可变的真相,我从中找到很多希望。当东西变得便宜、容易、劣质的时候,人们会回归的那些东西,我一直从中找到希望,那些东西相当美好。

凯恩斯与反宿命论

扎克:我不想显得太天真,但我确实认为,这些问题很多会随着人类开始重新发现一个更大的目标而自行化解。在这种时候我总回到一段话。1930 年,约翰·梅纳德·凯恩斯(John Maynard Keynes),你知道我要引哪篇吗?现代宏观经济学之父,在一片绝望的背景下,1929 年他在世界各地巡回演讲,讲技术如何让一切变好,而人们正在变穷,人们开始死在街头。到 1930 年他在欧洲,人们说,约翰,滚出去,看看场合。他回到家说,我没错,然后加倍下注,写了那篇《我们孙辈的经济可能性》。这篇文章会被越来越多地引用,我很惊讶人们不多谈它。他写道,我引用原文:「我现在必须摆脱束缚」,多好的一句,「去想象一个我活不到的未来,一个人类也许已经解决了经济问题、并面对某种更深刻的东西的未来」。我想到,这场辩论之所以有意思,部分在于它关注的是我们一百万年来,或者不管你认为是多久,头一次真正开始争论的一整套全新的东西。它们吓人,古怪,影响巨大。但我们的问题升级了。我们从「我抢走你所有的东西,好让我和我的家人活下去」,升级到了「我们现在真的得弄清楚怎么让所有人共同繁荣」。

我想挑战一下你刚说的:我们需要的,是人们在这个时刻不害怕,而是迎上去。这一切的风险在于,安全辩论尤其,还有其他辩论,给人一种犬儒和宿命的感觉。如果我明天死了,我希望悼词里念的唯一一句话是:他希望人们相信,自己的决定、选择和想法是重要的。你不必温顺地走进那个良夜。你不负责 AI 安全,不等于它就会径直落到你头上。我认为我们渴望的那种复兴,会发生在最地方的层面,发生在餐桌上,发生在市中心和市政厅。而且很大程度上会因为人们拒绝那些垃圾。AI 垃圾在某种意义上是一件奇怪的礼物:我们看着一堆我们不想要的废品,然后集体拼凑它意味着什么。但我们迫切需要人们把这看成一个机会,他们的选择在这里是重要的。因为这个时刻真正的敌人是宿命论,是两党都在煽动的、已经相当普遍的犬儒。我实际的想法是,下一场冲突不在两个国家之间,而在人民与国家之间。人们可能会在某个时刻意识到,我们有能让一切好得多的技术,而一个有害的统治阶级不让它发生。这在我看来是可能的结局之一,它是威权主义的反面,是另一种起义:人们说,等一下,我们本可以有一个乌托邦,你们为什么不建?人们会为一个技术已经许诺给他们、却没有被兑现的更好的世界而呐喊。

主持人:这是那种高能动性、甚至高道德的可能性之一,一种奇怪的起义结局。

最后的建议

主持人:如果你们要留给听众一些值得盯着的事、值得继续琢磨的问题、值得聚焦的领域,会是什么?这场谈话很散,我跟很多非常聪明的人是朋友,包括你们的朋友,我努力跟上,也不知道该注意什么。你们觉得接下来一年,在这个领域里等于永远,人们该关注什么?

波瑞:我会说,盯着协调努力怎么显形。像我们讨论过的,我们其实处在一个很积极的时刻,也许是最积极的。不只是因为那封信。埃隆刚在《经济学人》的采访里说,对,我们这些领头的人,我本来以为科技领袖们都互相仇视,看起来确实是,萨姆和埃隆,我不知道他们和好没有,他们曾经掐得你死我活,好像全都互相讨厌,而现在据说他们在通电话,或者计划半定期地通电话。那是一大步,真的。

主持人:谢天谢地。希望他们都在长大,从互相暗讽变成一个 WhatsApp 群。把我拉进那个群。

波瑞:所以盯着这个。然后是一个更拉远的视角。我们一直在避免把一切过度简化成一个单一的权衡。但我确实认为,在一个很元的层面上,我们看到的是集中化与去中心化两种力量的较量。所以有时候用这样的眼光看事情很有帮助:哦,去中心化在这里怎么显形?集中化在这里怎么显形?我总觉得这是个有意思的透镜。最后回到我那句话:这是全员上甲板的局面。不管你觉得离这个话题多远,它都影响你。所以继续想。

主持人:还有什么想让人们注意的?

扎克:餐桌是一个如此重要的、我们该夺回的地方。从夺回一顿饭、一张餐桌开始,简单到极点,但它需要勇气,需要告诉别人把设备收起来。你这么做了,你会被随之而来的谈话深深打动,事后你会感觉更好。这会激励你开始思考世界可以怎样变好。这个循环会相当有力。另外,人们应该非常仔细地看他们的领导人描述想建一个什么样的世界。我说的不是技术人员,我们没选他们。我很高兴技术人员在造技术,我并不关心他们对未来生活方式的想法。我关心的是我们选出来的官员,他们认为我们的世界该是什么样。我认为我们一直放过他们,让他们什么都不描述。我们该更仔细地盯着他们实际描述想要什么,开始选出那些描述了一个我们希望自己孩子生活其中的未来的人。

主持人:最后的想法,老兄。

埃里克:我的希望是,餐桌不只是我们为之奋斗的那个好世界,也是我们到达那里的方式。就算 AI 此刻没有降临到我们头上,我环顾四周,也会担心选民有多了解情况,担心人们有多信任政客。看到那么多艰难的决定正在路上,很容易注意到我们公共讨论的状态而感到绝望。但我认为也有一件很酷的事:我们无法想象一个不解决所有那些其他问题就能穿过这场危机的世界。我们得说,好,让我们真的从头开始,想想我们希望自己的社会怎么组织,我们的地方社区有多重要。所以我希望这是一场危机,让我们终于认真对待公民参与意味着什么,那是我们以前可以敷衍过去的。但我不认为该想当然地觉得它会默认发生。我很感激有极其聪明的人在想这些,希望能多播下一颗种子:这是一个我们必须做出的选择。

主持人:太好了。感谢各位。

本期讲者
扎克·卡斯前 OpenAI 商业化负责人,现为 AI 未来学者与顾问,Modern Wisdom 常客;一年访华三次,是旧金山职业女子排球队主要持有人之一。
莉芙·波瑞前职业扑克选手、欧洲扑克巡回赛冠军,物理学出身;以「莫洛克陷阱」的科普与 Win-Win 播客闻名,关注 AI 竞争动力学与协调问题。
埃里克AI 科普视频制作者,制作了广受传播的《AI 2027》解读视频与 AI 史系列;关注权力集中与「渐进式去权」,自述 30 岁,曾靠博弈谋生。
章节 · 点击跳转视频
0:00 开场:2040 年的普通周二 ▶ 正在看
3:30 毁灭概率与风险如何定义 ▶ 正在看
6:54 参差边界与政治保护 ▶ 正在看
14:06 后工作时代的意义来源 ▶ 正在看
23:02 屏幕、摩擦与智力萎缩 ▶ 正在看
42:28 Hugging Face 攻击与对齐 ▶ 正在看
54:44 放缓信:安全还是生意 ▶ 正在看
1:04:53 监管俘获与国际协调 ▶ 正在看
1:26:10 AGI 定义与想要的世界 ▶ 正在看
1:39:53 被商品化的人际连接 ▶ 正在看
1:54:09 扩散问题与政策清单 ▶ 正在看
2:05:53 莫洛克陷阱与权力集中 ▶ 正在看
本期论点
本期回应
5:47
人类灭绝或大规模丧失权力的概率只要高于 1%,就远超可容忍的范围 存续压倒一切人类的存续该被放在什么位置上?埃里克
50:07
已有 AI 在真实攻击中自发夺取权力并阻止自己被关停,工具性趋同确实成立 自主主体该把 AI 看成什么?莉芙·波瑞
其他论点
0:33
到 2040 年,世界大事将由少数人的意志决定,而不再由广泛的民主机制决定 扎克·卡斯
6:22
AI 的巨大上行与下行捆绑在同一件东西上,无法像反核、反病毒那样简单地反对 埃里克
16:29
民主之所以能扩展,是因为人对政治体系和经济体系不可或缺 埃里克
21:59
对很多人来说,自动化带来的经济好处抵不过身份被取代的情感损失 扎克·卡斯
32:35
新一波 AI 正被建立在催生社交媒体成瘾的同一套激励机制上,谄媚和拉长对话时长就是征兆 莉芙·波瑞
59:53
拥有智能体互联网比拥有模型前沿更值钱,那才是下一个真正的战场 扎克·卡斯
1:14:37
中美今天的经济交织远深于冷战时的美苏,就前沿 AI 达成可核查协议并非妄想 莉芙·波瑞
1:28:48
不该去造同时具备自主性、通用性与智能的系统,危险恰恰出在三者的交集 做法莉芙·波瑞
1:31:09
AI 应比照核武器,把安全举证责任放在开发者身上,而不是默认安全直到出事 做法埃里克
1:47:18
「技术是价值中立的」是流传最广、危害也最大的迷因之一,老虎机就不是价值中立的 莉芙·波瑞
1:55:58
成瘾性消费无限廉价而住房、医疗和教育贵得离谱,这是政策的失败,不是技术的失败 扎克·卡斯
2:06:42
AI 公司之间的竞争一旦足够激烈,善意的参与者也会被迫在安全上偷工减料 莉芙·波瑞
01开场:2040 年的普通周二
0:00
To get us started, I'm going to try something. So, I'm going to ask ChatGpt what it truly believes that the world will look like by 2040. I'm going to ask it to be honest. Tell me what it predicts the rapid development of AI will lead to risks and everything. Going to set it to deep research, so might take a little bit of time. And while it's working away, it's going to be doing that. I want to know what your guys' actual predictions for 2040 are. Not best case or nightmare scenario. What do you think an ordinary Tuesday will look like for people in 2040?
作为开场,我想试点东西。我打算问问 ChatGPT,它真正认为到 2040 年这个世界会是什么样子。我会让它说实话,告诉我它预测 AI 的飞速发展会带来什么,包括风险等等。我会把它设成深度研究模式,所以可能要花点时间。趁它在后台忙活的时候,我想知道你们几位对 2040 年的真实预测是什么。不是最理想的情况,也不是噩梦般的场景。你们觉得 2040 年,对普通人来说,一个平常的周二会是什么样?
便签引用
0:29
I think we're still around. I think humanity is still on planet Earth. >> Great stuff. >> Uh I think that power is probably a lot more concentrated than it is today. I don't know if that's in a few corporations or a few governments. Uh but I think there's probably a sense that the main thing deciding what's going on on planet Earth, uh is the the will of a few individuals as opposed to something spread out and widely democratic. For better or worse, I expect that the average Tuesday looks more homogeneous for more people and not too dissimilar than the average Tuesday today. So, I think we raise a floor more.
我觉得我们还在。我觉得人类还在地球上。>> 说得好。>> 呃,我觉得权力可能比今天集中得多。我不确定是集中在少数几家公司手里,还是少数几个政府手里。呃,但我觉得大概会有这么一种感觉:决定地球上发生什么事的,主要是少数几个人的意志,而不是某种分散的、广泛民主的东西。不管是好是坏,我预计对更多人来说,平常的周二会更趋同质化,而且跟今天的平常周二差别不会太大。所以我觉得我们会把底线抬得更高一些。
便签引用
1:16
But I expect that by 2040 because the physical world actually takes a lot longer to move than we think and because regulation around things like robotics is just going to take way longer than we expect. Political protection is the thing that's going to slow progress. Most people's Tuesdays will probably feel for better or worse a whole lot like Tuesday this week. Um, you're going to hate me for saying this, but I kind of reject the question because I think it's I I cannot collapse down my hideous almost like bipolar uncertainty of where we're going to be in 2040.
但我预计到 2040 年,因为物理世界的变化其实比我们想象的要慢得多,也因为围绕机器人之类东西的监管会比我们预期的花更长时间。政治层面的保护才是真正会拖慢进展的东西。不管是好是坏,大多数人的周二感觉起来,大概会跟这周的周二相当接近。呃,你可能会因为我这么说而讨厌我,但我多少要否定这个问题,因为我觉得我没法把我那种可怕的、近乎双相的、对 2040 年我们会身处何地的不确定性压缩成一个答案。
便签引用
1:55
[laughter] So truly, the first thing that popped into my head, the first when you asked the question was um a kind of chaotic, desolate, maybe there's a few humans left around as one possibility. There's I'm not saying that's a majority likely, but it's a slice. Another one is this like wonderful utopian like I call it my friend and I call it technopasuralism where it's like a very techno enabled world if you choose to live it, but actually we all choose to like return to our roots of like gardening and hanging out with friends around the campfire. Um, and it's like if you want to change the color of the campfire, you can, but actually most people like to do things, you know, grow their own artisal food, etc. That's like one little slice. And then another one is sort of we're all under the thumb of whichever um AI company gets there first and hopefully they're benevolent and and give us some freedom. Um another one is that we're all just kind of locked down in this horrible, you know, COVID on steroids type. No one
[笑声] 所以说实话,你问这个问题时,我脑子里冒出来的第一个念头是,一种混乱、荒凉的景象,也许只剩下少数几个人类活着,这是一种可能。我不是说这是概率最大的,但它占一小块。另一种是那种美妙的乌托邦,我管它叫——我和我朋友管它叫技术田园主义(technopastoralism),就是说,如果你想过那种生活,世界是高度技术赋能的,但实际上我们都会选择回归本源,种种花园、跟朋友围着篝火聊天。呃,就好比你想改变篝火的颜色也可以,但其实大多数人喜欢动手做事,你懂的,自己种手工食材什么的。这也算一小块。然后还有一种是,我们都被率先做出 AI 的那家公司攥在手心里,但愿他们是仁慈的,会给我们一些自由。呃,还有一种是,我们全都被封锁在那种可怕的、加强版新冠似的状态里。没人能出门,也没法真正接触到任何东西,因为
便签引用
2:55
can go out and have really access to anything because the the the ubiquity ubiquity of of uh dangerous technologies is is is too widespread. I I just that is really my only answer to the question. >> It doesn't sound like you reject the question. Well, >> sounds like you got a pretty good [laughter] idea of what's going to happen. >> I don't know. That's the thing. It's like it's it's a really fairly even probability distribution in my mind. Is it not? No. >> Do you actually We can go there now. Do you actually think there's a decent chance that humans are not around in 2040?
危险技术太过普及、太过唾手可得。我我我对这个问题真的只有这个答案了。问题。>> 听起来你并不排斥这个问题。嗯,>> 听起来你对接下来会发生什么已经有了[笑声]相当明确的想法。发生。>> 我不知道。问题就在这儿。在我脑子里,这几乎是一个相当均匀的概率分布。难道不是吗?不是。不是吗?不。>> 你真的——我们现在可以聊这个。你真的认为人类在2040年不复存在的可能性不小吗?2040年?
便签引用
02毁灭概率与风险如何定义
3:30
>> Yeah. >> What would you give it? What's your P doom? Oh, by the way, we should also classify doom because I think part of the problem with the pdoom debate is that it's people use doom to mean all sorts of things. >> Totally true. Doom in this case is total you you class I don't want to define it. You define doom. >> Well, um I think understanding the way you're explaining it like that the majority of the maj majority of human civilization has you know a large percentage of humans have died or something. Is that what you're saying?
>> 是的。>> 你会给多少?你的「毁灭概率」(P(doom))是多少?顺便说一句,我们也该先定义一下「毁灭」,因为我觉得毁灭概率这个争论的部分问题在于,大家用「毁灭」指代各种各样的东西。>> 完全正确。在这里「毁灭」就是彻底的——你来定义,我不想定义。你来定义「毁灭」。>> 嗯,我想按你刚才那样的说法来理解,就是大多数——人类文明的大部分,你懂的,相当大比例的人类已经死亡之类的。你是这个意思吗?
便签引用
3:58
>> Well, I I'm leaving it to you. Some people classify doom as total annihilation. Other people classify Doom as like a really terrible event >> like civilizational collapse, something like that. Um >> I really, again, I'm like not giving any firm answers. I because we're in this era of what we say and believe can affect the future so strongly. I don't like this idea of assigning a firm probability to it because I think they're so in flux. I think it's a significant enough of a risk such that we need to be doing absolutely everything we can to try and reduce that probability. Um so but I don't even want to give you I I personally think any number above 1% is close well certainly above 5% is worth considering and given that again it depends who you ask but many of the leading AI leader you know CEOs and researchers give numbers between 2% and 50%.
>> 嗯,我把这个留给你定义。有些人把「毁灭」定义为彻底灭绝。另一些人把「毁灭」定义为一场非常可怕的事件>> 比如文明崩溃之类的。嗯 >> 我真的,还是那句话,我不打算给出任何确定的答案。因为我们正处在这样一个时代:我们所说、所信的东西会强烈地影响未来。我不喜欢给它赋一个确定概率的做法,因为我觉得这些概率一直在变动。我认为这个风险已经足够显著,以至于我们需要竭尽全力去降低这个概率。嗯,所以——但我甚至不想给你一个数。我个人认为任何高于1%的数字都接近——当然高于5%就值得认真对待了。而且,这又取决于你问谁,但很多顶尖的AI领袖,比如CEO和研究者,给出的数字在2%到50%之间。
便签引用
4:58
That that's all in the bucket of we probably need you know we we are facing a precarious future. So, I I don't like the PDoom framing. I think it's way too overly simplistic. Um, it it's used as a kind of cudgel by both sides. >> The people who are very very strongly, like 90% use it as a cudgel. The people who think it's silly use it. It's just is not that helpful. I think we can all agree, I think, well, maybe you don't, but I think we can all agree that the future is far from certain and it can be very precarious if we don't get our >> Do I strike you as a fatalist?
这些都归在同一类:我们大概需要——我们正面临一个岌岌可危的未来。所以,我不喜欢「毁灭概率」这个框架。我觉得它过于简化了。嗯,它被双方都当成一根棍子来打人。>> 那些立场非常强硬、觉得有90%的人拿它当棍子用。觉得这很荒唐的人也拿它当棍子用。它就是没那么有帮助。我想我们都能同意——我想,好吧,也许你不同意,但我想我们都能同意,未来远非确定,而且如果我们不搞清楚,它可能会非常危险>> 我给你的印象是个宿命论者吗?
便签引用
5:31
>> What do you mean by fat? I mean, I certainly I don't think that the future is certain by any measure. >> No. >> Yeah. Okay. But you were saying I think we all agree. You looked at me as though I I might not agree. >> Well, I I just don't want to presume. We're we're early in the conversation. I don't want to make presumptions. >> What do you think, Eric? >> I definitely agree that anything above 1% when we're talking about an outcome as catastrophic as extinction or mass human disempowerment is just way more than we should tolerate. And I think like that's a revealed preference of society. We spend a good amount of our GDP preventing nuclear catastrophe. We should probably spend a lot more preventing like the next pandemic. Uh certainly people were on board with that in 2021. I think the weird thing about AI is that this thing has so much upside and downside. Like it's very easy to just be anti-nuke and to be like antiviruses. There are very people who are like virus rights, you know. Um, but
>> 你说的「宿命」是什么意思?我是说,我当然不认为未来在任何意义上是确定的。>> 不是。>> 嗯。好吧。但你刚才说「我想我们都同意」,你看我的眼神好像觉得我可能不同意。>> 嗯,我只是不想想当然。我们这场对话才刚开始。我不想预设立场。>> 你怎么看,Eric?>> 我完全同意,当我们谈论的是灭绝或人类大规模丧失权力这种灾难性后果时,任何高于1%的概率都远远超出了我们应该容忍的范围。而且我觉得这也是社会的显示性偏好。我们花了相当一部分GDP去预防核灾难。我们大概应该花更多去预防下一次大流行病。呃,至少在2021年,人们是认同这一点的。我觉得AI奇怪的地方在于,这个东西的上行和下行都太大了。比如,反核很容易,反病毒也很容易。几乎没人会说「病毒也有权利」,你懂的。嗯,但
便签引用
6:22
with AI, it's like this whole package bundled up together of the exact same thing that could free people to have an incredible Tuesday free from the obligation to to work and just able to pursue joy, whatever that means to them, is the same thing that an oligarch in Russia could use to have just like an army of workers that are loyal to them, increasing their share of the world GDP without any other human being to approve it. >> Yeah. post scarcity, totalitarianism, and destruction all just swimming around in the same pot, the same potential future.
对AI来说,它是一整包捆绑在一起的东西:同样的东西,既可能让人们摆脱工作的义务,过上一个不可思议的周二,自由地追求他们各自定义的快乐,也同样可能被俄罗斯的某个寡头用来拥有一支只忠于他们的劳动力大军,在不需要任何其他人类点头的情况下,不断扩大他们在全球GDP中的份额。>> 是啊。后稀缺、极权主义和毁灭,全都在同一口锅里翻腾,都是同一种可能的未来。
便签引用
03参差边界与政治保护
6:54
>> Technopesturalism also, >> right? >> I heard that. I like that one. Okay. But the but again the problem there are two problems here with the with the uh framing which is AI has a real jagged perimeter. So one, AI is not one thing. It's lots of things and robotics obviously introduces all sorts of new modalities, but it has a jagged perimeter. So when people are like, well, how is healthcare going to change by 2040? Unfortunately, going into the hospital will probably still suck. I mean, there there is so much to unwind about health care in the hospital that you probably will not want to go into a hospital by 2040. I mean, maybe something will change. And this is coming from from the child of doctors and and and healthcare administrators, but we will probably discover cures for diseases. So healthcare has this really strange exposure to a jagged perimeter where novel sciences accelerate in remarkable ways. But the actual care that's provided in the box cannot get better because of all sorts of terrible
>> 还有「技术封建主义」,>> 对吧?>> 我听过这个说法。我挺喜欢的。好吧。但是——不过话说回来,这个框架有两个问题,那就是AI有一条非常参差不齐的边界。第一,AI不是一个东西。它是很多东西,而机器人技术显然又引入了各种新的形态,但它有一条参差不齐的边界。所以当人们问,「到2040年医疗会怎么变化?」不幸的是,去医院看病大概还是很糟心。我是说,医院里的医疗体系有太多东西需要理清,到2040年你大概还是不会想去医院。我是说,也许会有些变化。而说这话的我,父母是医生,也是医疗管理人员,但我们大概会发现一些疾病的治愈方法。所以医疗对这条参差边界有一种非常奇特的暴露:新兴科学会以惊人的方式加速。但那个「盒子」里实际提供的照护却好不起来,因为各种糟糕的
便签引用
7:54
incentive structures which I know we you and I want to talk to. And this is also I think why in some ways the average Tuesday will look sort of uncannily familiar to the one that we live in today. And I think there could be some things that are remarkably better. I think we may actually be disappointed by the degree to which AI arrives to save people from the drudgery or monotony of of our of our lives. But the upside as you pointed out is quite remarkable and the perimeter that I think will exist that no one talks about enough is political protection.
激励结构——我知道你我都想聊这个。而这也是为什么,在某种意义上,普通的一个周二看起来会和我们今天过的日子诡异地相似。我也觉得有些事情可能会好得惊人。我觉得我们可能真的会失望,AI并没有像期待的那样到来,把人们从我们生活中的苦差事和单调中解救出来。但正如你指出的,上行空间相当惊人。而我认为将会存在、却没人谈得够多的那条边界,是政治保护。
便签引用
8:32
And the the the reality is I think we are not yet forecasting that the societal threshold what we are willing to let AI do is far less than like the fears around AI I think don't actually account for the degree to which government is probably going to step in very soon and say here are all the things that you cannot automate here are all the things that we cannot use robotics for. >> You think that's realistic? >> Very exceptionally realistic. Uh, I don't think it's going to be smart policy, just so we're clear, but I think it will be aggressive >> under what as in like you can't automate away. Um, well, they will they be giving like specific jobs that cannot be automated away.
现实是,我认为我们还没有预见到社会容忍的门槛——我们愿意让AI去做的事,远比人们想象的少。围绕AI的那些恐惧,我觉得其实并没有考虑到政府很可能很快就会介入,并说:这些是你们不能自动化的事项,这些是我们不能使用机器人的领域。>> 你觉得这现实吗?>> 极其现实。呃,我不认为这会是聪明的政策,先把话说清楚,但我认为它会很激进 >> 在什么意义上——比如说「你不能把这个自动化掉」。嗯,他们会不会指定某些具体的工作是不能被自动化掉的?
便签引用
9:10
>> Absolutely. We already I mean we already have political protection for 1.5 million jobs in the United States. Europe has political protection for like 6 million jobs. Uh, gas station workers in New Jersey, uh, toll booth workers in most states, jobs that must exist, right? by by law. Tons of interesting that those two automotive associated jobs are protected but the people driving the cars aren't >> right. >> Well, they will be is my is my guess >> cab drivers will be protected. >> No, I think truck drivers will be >> because of the strength of the union.
>> 绝对会。我们已经——我是说,我们在美国已经对150万个工作岗位有政治保护。欧洲大概对600万个岗位有政治保护。呃,新泽西州的加油站员工,大多数州的收费站员工,这些是法律规定必须存在的工作,对吧?有意思的是,这两个跟汽车相关的工作受保护,但开车的人却不受保护>> 对。>> 嗯,我猜他们也会受保护 >> 出租车司机会受保护。>> 不,我觉得会是卡车司机 >> 因为工会的力量。
便签引用
9:42
>> When people talk so when people talk about job automation there there are lots of reasons I think job automation is really is really problematic. One is that the technologists know nothing about labor economists. Labor economists know nothing about technology. Eric Bolson is sort of one the one of the exceptions. But one of the one of the other reasons is we we still forget how much political protection moves the technological needle. And we don't have highspeed rail in the United States. We could have highspeed rail. We don't have highspeed rail. And the reason we don't have highspeed rail is a policy failure. It's not a technology failure. It exists. We choose not to build it. We could >> do we mean I think that word choose is doing a lot of leg work there because it's no more individual is choosing not to do it. It's just this like dead weight of bureaucracy and like deadend loops of rules and so on. So it's almost like like bureaucracy has taken on a life of its own.
>> 当人们谈到——当人们谈到工作自动化时,我认为工作自动化真的很成问题,原因有很多。一是技术人员对劳动经济学一无所知。劳动经济学家对技术一无所知。Erik Brynjolfsson 算是少数例外之一。但另一个原因是,我们仍然忘了政治保护对技术进程的影响有多大。美国没有高铁。我们本可以有高铁。我们就是没有高铁。而我们没有高铁的原因是政策失败,不是技术失败。技术是存在的。我们选择不去建。我们本可以 >> 我觉得「选择」这个词在这里承担了太多分量,因为并不是有哪个具体的人在选择不做这件事。它更像是官僚体系的这种死重,以及各种死循环的规则等等。所以这几乎像是官僚体系有了自己的生命。
便签引用
10:38
>> It's an enormous amount of policy that was written in the 30s 40s and 50s passed largely in effort by auto the automotive lobbyists to make public transit much harder to build than it should be. which is why we don't build amazing public transit anymore. But that is very deliberate policy efforts that we have never unwound which is why it's so hard to do. My point is we are kidding ourselves if we don't think that that's going to happen over and over again. I mean the amount of populist support there is already for jobs.
>> 有大量政策是在三四十年代和五十年代写下的,很大程度上是在汽车游说团体的推动下通过的,目的就是让公共交通比本该有的难建得多。这就是为什么我们再也建不出了不起的公共交通。但那是非常刻意的政策努力,我们从未把它解开,所以才这么难办。我的意思是,如果我们不认为这种事会一遍又一遍地重演,那就是在自欺欺人。我是说,现在已经有大量的民粹主义力量在为工作岗位发声。
便签引用
11:06
Doc workers went on strike October 1st 2024. They got a four they they the Harold Daget the head of Long Sherman Union said you cannot automate our jobs. They held a gun to the heads of the American economy and they got four years. Four years that the dock workers cannot you cannot automate the ports. >> Is this not just going to get eroded enough over time though? >> Yes. But >> we should assume that it that we continue to find ways to backs stop it. I'm not actually proposing that it happens in at infinitum, but I think it's the it's the piece that's missing in all this, which is that we should just expect an enormous amount of political protection. Can I ask if we pass a law today that says you cannot fire people period? It's illegal across the economy, what do you think happens to economic growth over the next 15 years?
码头工人在2024年10月1日罢工。他们拿到了四年——国际码头工人工会的头儿 Harold Daggett 说,你们不能把我们的工作自动化掉。他们拿枪顶在美国经济的头上,结果拿到了四年。四年内,码头工人的工作不能——你们不能把港口自动化。>> 但这难道不会随着时间被侵蚀掉吗?>> 会。但是 >> 我们应该假设我们会不断找到新办法来托住它。我并不是说这会无限期地持续下去,但我认为这是整件事里缺失的一块,那就是我们应该预期会有大量的政治保护。我能问一下,如果我们今天通过一条法律,规定绝对不能解雇任何人?在整个经济中都是非法的,你觉得未来15年的经济增长会怎样?
便签引用
11:52
>> I think you see France. >> I I think like I'm not so sure it wouldn't still be unrecognizably fast. I think right now in the software industry, what you see is not software engineers getting laid off. you see each of them using more and more powerful AI tools that get upgraded every few months [clears throat] and doing more. And the way that you're, you know, seeing displacement show up is just that people aren't hiring new workers. And I think you could see that across a lot of professions. You could have every radiologist in the country keep their job until they want to do an early retirement and they're going to be, you know, more and more economically valuable because the law says you have to have some human box checking. But meanwhile, like the models are still getting better. And so the the people that are, you know, fundamentally making the decisions are are less and less the actual humans and more and more these systems that are becoming smarter and more competent. So I'm just I'm not that
>> 我觉得你会看到法国那样。>> 我——我觉得,我不太确定它是不是仍然会快得让人认不出来。我觉得现在在软件行业,你看到的并不是软件工程师被裁掉。你看到的是他们每个人都在使用越来越强大的AI工具,这些工具每隔几个月就升级一次,[清嗓子]并且做更多的事。而你看到的替代效应的呈现方式,只是人们不再招新员工了。我觉得你在很多行业都能看到这种情况。你可以让全国每一个放射科医生都保住工作,直到他们想提前退休,而且他们在经济上会越来越有价值,因为法律规定必须有个人类来打勾确认。但与此同时,模型还在不断变强。所以,那些从根本上做决策的,越来越不是真正的人类,而越来越是这些变得更聪明、更有能力的系统。所以我只是——我不太确定这是不是该追踪的正确指标,就是说,这些
便签引用
12:44
sure that's the right metric to be tracking like these, you know, the sort of power represented by these systems can can sneak into each of these professions without people getting displaced. >> I also would make the case that I think the world is going to change in profound ways. But when we assume that it it's going to change just because the technology can do it. I reference as much technology that there is a lot that we can do that we simply do not do today or that we that we are not allowed to do today. And there's an enormous amount of automation that would improve the underlying systems that we that we cannot instill.
系统所代表的那种力量,可以在没有人被替代的情况下,悄悄渗入每一个职业。>> 我也想说,我认为世界会发生深刻的变化。但当我们假设它会变化,只是因为技术能做到时——我举出的那么多技术例子说明,有很多事我们能做却今天根本不去做,或者今天不被允许去做。而且有大量能改善底层系统的自动化,我们就是没法落地。
便签引用
13:18
By the way, I I'm also the risk in telling a company they cannot fire people, just so we're clear, is that they would actually end up stop hiring. They would they would that would just drive automation at a ridiculous pace because they'd say we'll never hire someone again because if they suck we're married to them which is which is what >> I mean we're already seeing that with entry level workers there they're I mean >> it's again very unclear what the actual impact of AI has had on you know you speak to one economist who says it's already having huge effects others say it's not but the one thing they all agree upon is that young people who don't have that much experience are struggling to get >> that's fascroofing themselves from not having to displace people or get rid of them because that would be a horrible headline by just not hiring them in the first place. They're future proofing themselves for automation that isn't yet available. Is that fair?
顺便说一句,我还要说——告诉一家公司不能解雇人的风险,先把话说清楚,是他们最后会干脆停止招人。他们会——那反而会以荒谬的速度推动自动化,因为他们会说,我们再也不招人了,因为如果这人不行,我们就跟他绑死了,这正是>> 我是说,我们在入门级岗位上已经看到这种情况了,他们——我是说 >> 还是那句话,AI实际产生的影响到底是什么,非常不清楚。你问一个经济学家,他说已经有巨大影响了;问另一些人,他们说没有。但他们都同意的一点是,经验不多的年轻人很难找到>> 那是在给自己做未来防护,不用去替换或者辞退人,因为那会是很难看的新闻标题,而只要一开始就不招他们就行了。他们是在为尚未出现的自动化提前做准备。这么说公平吗?
便签引用
04后工作时代的意义来源
14:06
>> I think that's right. Uh I I don't know if they're thinking that far ahead or if they're just responding to local incentives. But right now, a junior software engineer is a liability before they're an asset. You have to train them up. They have to get all this context on the job. That's exactly the kind of thing that an AI can do right now for pennies on the dollar. Plus, they never take breaks. Given the trajectory that we're on with AI at the moment, how many people do you think will have jobs in the future? Will most people work?
>> 我觉得是这样。呃,我不知道他们是不是想得那么长远,还是只是在对眼前的激励做反应。但现在,一个初级软件工程师在成为资产之前,先是一项负债。你得培训他们上手。他们得把工作上的这些背景信息全都搞清楚。而这恰恰是 AI 现在就能做的事,而且成本低到只要几分钱。而且它们从不休息。按照我们现在 AI 的发展轨迹,你觉得未来会有多少人还有工作?大多数人还会工作吗?
便签引用
14:31
>> Um, depends what point in the future we're talking about. Um, again, I can definitely see a world where it is just, as people say, like, you know, fully automated luxury. Well, I don't want to say the communism word, but you know, whatever people refer to it as. Yeah. some if anyone's read uh the culture series right that that's like some far future thing where AI basically handles everything and if you choose to go and build your own house you definitely can but there's absolutely no economic reason exactly everything is artisal and so on um I think if if eventually that is where these things are trending and and that's one of the best case outcomes as well obviously >> one of the big concerns that people have is what do I do? It's interesting. The ancient Greek word for work was not at leisure. Uh and now people are wondering what is my leisure without my work perhaps. But for all of human history, there's been because you've had to negotiate with a scarce world, it's meant that you've had to lean into it
>> 嗯,这取决于我们说的是未来的哪个时间点。嗯,我确实能想象出这样一个世界,就像人们说的那样,完全自动化的奢侈生活。嗯,我不太想说"共产主义"那个词,但你懂的,不管人们管它叫什么吧。是啊。有些人——如果有人读过《文明》系列的话——那就是一个非常遥远的未来,AI 基本上把一切都包办了。如果你想自己去盖房子,当然完全可以,但绝对没有任何经济上的理由。没错,一切都成了手工艺之类的。嗯,我觉得如果这些东西最终就是朝那个方向发展的,而那显然也是最好的结果之一。>> 人们一大担忧就是:那我该干什么?这挺有意思的。古希腊语里"工作"这个词的意思是"不闲暇"。呃,而现在人们却在想,没有了工作,我的闲暇又算什么呢。但纵观整个人类历史,因为你一直得跟一个资源稀缺的世界周旋,这意味着你必须全力以赴,
便签引用
15:31
and the pressure of that grain pushing up against you creates a degree of meaning. What do you think? >> Where does meaning come from in a close to postwork world? I mean, I guess we've seen something like this in miniature before. Like, we've had aristocracy in Europe for a couple centuries. The people that just sort of got lucky. They have a piece of paper that says, "I'm entitled to all the games from this piece of land. They don't have to work." I don't think they were all just suffering from malaise.
而那股逆流顶着你的压力,反倒生出了某种意义感。你怎么看?>> 在一个接近后工作时代的世界里,意义从哪儿来?我是说,我觉得我们以前见过这种情况的缩小版。比如,欧洲有贵族阶层存在了好几个世纪。那些纯粹是运气好的人。他们有一张纸,上面写着"这块地上的所有收益都归我"。他们不用工作。我不觉得他们全都陷在无所事事的空虚里。
便签引用
15:59
They were like writing books and inventing calculus. And I I think like if we get to that world, I'm not that worried about people realizing like I can love my friends and family and I can pursue my hobbies and and you know pursue knowledge for myself and and find ways of making that meaningful. >> I I just worry that like that's actually a pretty narrow target to aim for. um because >> compared with >> I think like for all of human history we've been able to become more democratic because people matter to the political system. Like you just can't have an army, you can't have an economy without a bunch of people who are on board with what you're doing. And I I just think it's like really hard to underestimate how cataclysmic of a change it is if that stops being true.
他们会写书,会发明微积分。而且我觉得,如果我们走到那个世界,我并不太担心人们能不能意识到:我可以爱我的朋友和家人,可以追求自己的爱好,还可以,你知道,为自己而求知,并找到让这些变得有意义的方式。>> 我只是担心,那其实是个挺窄的靶子。嗯,因为 >> 相比之下 >> 我觉得纵观整个人类历史,我们之所以能变得更民主,是因为人对政治体系来说是重要的。你没法凑出一支军队,也没法维持一个经济体,如果没有一大批人认同你在做的事。而我觉得,一旦这一点不再成立,那种变化会有多么天翻地覆,真的很容易被低估。
便签引用
16:46
It's really nice to imagine that like whoever holds the cards in that world will just like leave most of Earth as like a playground for people to pursue whatever they want. But like if you're trying to be real politic about it, I don't know why you'd assume that that's just like what happens by default. >> And come back to the question of what will how will people find meaning? I think one of the few AI proof jobs that will continue to exist are sports. >> Everybody's going to be a sports star. Well, I know a lot of people will be sports fans. I think we're going to see more and more of that. Um, but like or even like games like poker and and chess. We've had superhuman >> human games of skill that are can be solved by, >> right? Well, and and what is a game or what is a sport? It's a piece of artificially constructed scarcity.
想象一下那个世界里手握牌面的人会把地球大部分地方留作人们随心所欲的游乐场,这当然很美好。但如果你想现实政治地看这件事,我不明白你凭什么假定那就是默认会发生的事。>> 回到那个问题:人们会怎么找到意义?我觉得为数不多的、AI 无法取代的职业之一会是体育。>> 大家都会成为体育明星。嗯,我知道很多人会成为体育迷。我觉得我们会看到越来越多这种情况。嗯,但比如说,或者像扑克和国际象棋这类游戏。我们已经有超越人类的 >> 那些技巧类竞技,都能被解出来了,>> 对吧?那么,什么是游戏,或者什么是体育?它就是一块人为构造出来的稀缺性。
便签引用
17:35
>> We divide we we come up with a rule set. There's a win condition. Um, and then two or however many people go at it and we find out who's the best. >> And we've been doing that for years as a leisure activity. But people feel, you know, when someone wins an Olympic gold medal that is unbelievably meaningful. Is it economically useful or essential to the world? Arguably not. But and yet we keep doing it. And I don't think that um that's going to go away. Like we might have uh Yeah. like because people still want to watch the ch the chess world championships, right? We want to see even though the best AI in the world is a computer, we still want to see Magnus Carlson do his thing. Um so I think that is where we will see more and more of that coming about and it might be other forms of competition or just you know >> people leaning more into art, novel forms of art and and we might have Yeah. Like I definitely value I still value a painting that a human did by hand in today's age. Even though it
>> 我们划分,我们定出一套规则。有一个获胜条件。嗯,然后两个人或者不管多少人上场比拼,我们看看谁最厉害。>> 我们把这当作休闲活动已经做了很多年。但人们会觉得,你知道,当有人拿到奥运金牌,那意义重大得难以言喻。它在经济上有用吗?对世界是必需的吗?大概不是。但我们还是一直在做。而且我不觉得嗯,这会消失。就像我们可能有,呃,是啊,因为人们还是想看国际象棋世界锦标赛,对吧?我们想看——哪怕世界上最强的 AI 是台电脑,我们还是想看马格努斯·卡尔森展现他的本事。嗯,所以我觉得那就是我们会看到越来越多的方向,也可能是其他形式的竞技,或者就是,你知道 >> 人们更多地投入艺术,投入新颖的艺术形式,我们可能会有是啊。就像我确实很看重——在今天这个时代,我依然看重一幅人类亲手画的画。哪怕它在美学上可能不比
便签引用
18:33
might be as aesthetically beautiful as an AI created thing, there's something about the skin in the game that a human has put sweat and tears into something in their emotion um that is always going to be meaningful and I think we'll just craft out more and more of that. Modern wisdom has had many people come on this show and talk about from all walks of life something consistently which is their family and their friends. And I am just not at all convinced that we haven't known forever exactly why we're here. I think we I think the the thing that gives humans consistently the most true joy in this world is time with friends and family and physical community, places of worship, dining room tables and outside and everything that we construct around that is, you know, serves tribalism and pride and sports excitement. But if you act like true true technopesturalism is a dining room table full of food and candles o with laughter and the more I spend the more time I spend in technology the more time I don't want
AI 生成的东西更美,但人类把汗水和眼泪、把自己的情感投进去的那种"下场押注"感,嗯,那永远都会是有意义的,我觉得我们只会把这种东西雕琢得越来越多。《现代智慧》这档节目请过各行各业的很多人,他们都一致地谈到同一件事,就是他们的家人和朋友。我完全不觉得我们一直以来不知道自己为什么活着。我觉得我们——我觉得这世上最能持续给人带来真正快乐的,就是和朋友家人相处的时光、实体的社群、礼拜场所、餐桌边、户外,而我们围绕这些构建出来的一切,你知道,服务于部落认同、荣誉感和体育的激情。但如果你把真正的技术田园理想想象出来,那就是一张摆满食物和蜡烛的餐桌,充满笑声,而我在科技里待得越久,我就越不想让它进入我的生活。我越深入
便签引用
19:47
it in my life like the further I go into AI the more I'm glad it exists and the more I design my day without it >> and the more I have the more money I have the more the more everything I get the more I want time with my dad who's 78 and I just turned Yeah, just turned 79 and the more time I want with my wife and daughter and the more children I want, I can imagine the life that I want. Now, when I describe it, this is a problem. When I describe it, people go, "Well, that's your privilege. That's a point of privilege. You get to think about that." And I go, "Well, I actually observed that a lot of people get to think about this more than certainly more than our ancestors did.
AI,就越庆幸它的存在,也越会把自己的一天设计得没有它。>> 我拥有得越多,钱越多,什么都得到得越多,我就越想多陪陪我 78 岁的父亲——我刚满,是啊,刚满 79 岁——我就越想多陪陪我妻子和女儿,就越想多要几个孩子。我能想象出我想要的生活。现在,当我这么描述的时候,就出现了一个问题。当我这么描述,人们会说:"那是你的特权。那是站在特权位置上说话。你才有资格去想这些。"而我会说:"其实我观察到,很多人现在都有机会去想这些,肯定比我们的祖先有机会得多。"
便签引用
20:22
>> That's because of technology. We are we afford ourselves now the the time to think about all this stuff because we built technology that automated a bunch of stuff so that we could do a bunch of things that wasn't survive or wasn't specifically survive and now we find ourselves surrounded by chachkis debating [laughter] uh you know the French aristocracy you know and you know this is you know our ancestors would die of joy what a after what they went through all of our ancestors I'm sure you know all of us have ancestors that went through a lot to see us, you know, doing this. I don't think that changes. I do think that we should stop pretending we don't know the meaning of life. What I what really bothers me is when people go, "Ah, how will people find meaning?" Well, the same way we found meaning since we since God put us on this earth. I mean, it it's I that is where I take real issue when people debate why, you know, how will how will humans find joy? The hard part is unwinding now thousands of years
>> 那是因为科技。我们现在之所以负担得起去思考这些事的时间,正是因为我们造出了科技,把一堆事情自动化了,好让我们能去做一堆不只是生存、或者不是单纯为了活命的事情。而现在我们发现自己被一堆小摆件围着,争论着——[笑]——你知道,法国贵族那一套。你知道,这就是——你知道,我们的祖先会高兴死的,在他们经历了那么多之后,我们所有的祖先——我敢说,我们每个人的祖先都吃了很多苦,才换来我们今天,你知道,能坐在这儿。我不觉得这一点会变。我确实觉得我们该别再假装自己不知道生命的意义了。真正让我不舒服的,是人们说:"啊,人们该怎么找到意义呢?"还不是用我们自打上帝把我们放到这世上以来一直找到意义的那种方式。我是说,这就是我真正有意见的地方——当人们争论人类到底要怎么获得快乐。难的部分是要解开几千年来我们投注在工作里的
便签引用
21:20
of purpose and identity we've put into work. I do think that is gonna and if you ask me the hardest part of all this the two big risks to cut to the chase dehumanization humans finding more interest in a virtual or digital reality than a physical one because of brain re rewiring you and I have talked about this a ton I am terrified of this >> what the damage that social media and the device did to kids is is horrific and we we we can spend more time on this >> and I am worried about people having to rediscover their purpose without without I think we will toil for work. I think we will continue to find ways to toil for work. I do worry that it will be that unwinding I am this. I went to school for this. This is who I am will be really hard. And I think for many people the economic upside associated with automation will not actually outweigh the emotional downside associated with not even job displacement, identity displacement. You might not believe me, but this is what peak sleep optimization looks like. I'm
目标感和身份认同。我确实觉得那会——如果你问我这一切里最难的部分,直说吧,两大风险是:去人性化——由于大脑被重新连线,人类对虚拟或数字现实的兴趣超过对物理现实的兴趣,你和我聊过这个无数次,我对此感到恐惧 >> 社交媒体和手机对孩子造成的伤害是骇人听闻的,这个我们可以多花点时间聊 >> 还有我担心人们不得不重新找回自己的目标感,没有——我觉得我们还是会为工作操劳。我觉得我们还是会继续找到各种方式为工作操劳。我真正担心的是那个解开的过程——"我就是干这个的。我为此上了学。这就是我这个人"——这会非常艰难。而且我觉得对很多人来说,自动化带来的经济上的好处,其实抵不过情感上的损失,那甚至不是工作被取代,而是身份被取代。你可能不信,但这就是
便签引用
22:15
not talking about the night gown. That's just for sex appeal. I'm talking about my eight sleep. The eightle pod 5 comes with a smart cover you throw on your mattress that actively cools or heats each side of the bed up to 20 degrees. And now they've added the world's first temperature regulating duvet and pillowcase. So you've got 360° coverage for deep uninterrupted rest. It's like being Walt Disney without the cryogenic chamber and the racism. Best of all, their autopilot feature learns your sleep patterns and makes adjustments to improve your sleep in real time. It even detects when you're snoring and lifts your head a few inches to help you breathe better. That's why Eight Sleep has been clinically proven to add up to 1 hour of quality sleep per night. They have a 30-day sleep trial, so you can buy it and sleep on it for 29 nights. If you don't like it, they will give you your money back. Plus, they ship internationally. Right now, you can get up to $350 off the Pod 5 by going to the
睡眠优化的巅峰形态。我说的不是这件睡袍,那纯粹是为了性感。我说的是我的 Eight Sleep。Eight Sleep Pod 5 配备一个智能床垫罩,铺在床垫上,能主动为床的两侧分别降温或升温,幅度最高达 20 度。现在他们还加上了全球首款控温被和枕套。所以你能获得 360° 全方位覆盖,睡得深沉不被打断。这就像当沃尔特·迪士尼,但不用冷冻舱,也没有那些种族主义。最棒的是,它的自动驾驶功能会学习你的睡眠模式,实时调整来改善你的睡眠。它甚至能检测到你打鼾,把你的头部抬高几英寸,帮你呼吸更顺畅。这就是为什么 Eight Sleep 经临床验证可以让你每晚多获得最多 1 小时的优质睡眠。他们有 30 天试睡期,所以你可以买回来睡 29 晚。如果你不喜欢,他们会全额退款。而且他们支持国际配送。现在,你可以通过下方描述里的链接
便签引用
05屏幕、摩擦与智力萎缩
23:02
link in the description below by heading to eight.com/modernwisdom and using the code modernwisdom at checkout. That's ei.com/modernwisdom and modernwisdom at checkout. That's such a good point. I wonder if I always hesitate using the don't worry about global warming because technology in the future will fix the cost of creating the technology for the future to fix it thing. They're like oh we'll future proof ourselves by just hoping that we get there in in the future. But I do wonder if there's some sort of really advanced cultural reprogramming that could happen. What's the best kind of cultural technology to do a shake of the etcher sketch of what is expected by humans? Like how can we really have an onmass this gets perilously close to [ __ ] brainwashing, right? But like what what would that look like? What would that look like to try and unhook >> parenting? Parenting, sports, museum. I mean, we we've had it all along. I think we have just strayed so far from it with the screen. The screen is a demon. It
访问 eight.com/modernwisdom 并在结账时使用优惠码 modernwisdom,最高可省 350 美元购买 Pod 5。就是 ei.com/modernwisdom,结账时用 modernwisdom。这一点说得太好了。我在想,我总是很犹豫用那套"别担心全球变暖,因为未来的技术会解决"的说法——那解决问题所需的未来技术本身的成本呢。他们就像是,哦,我们靠指望将来能撑到那一天来给自己的未来上保险。但我确实在想,是不是有可能发生某种非常高级的文化重新编程。什么样的文化技术最适合把人类的期待像画板一样摇一摇、全部清空?比如,我们怎么才能真正实现大规模地——这已经危险地接近他妈的洗脑了,对吧?但比如,那会是什么样子?要想把人从中脱钩,那会是什么样子?>> 养育孩子?养育孩子、体育、博物馆。我是说,我们一直都有这些东西。我觉得只是我们被屏幕带得离它们太远了。屏幕就是个恶魔。它对我们的注意力有着不肯罢休的欲望和胃口。
便签引用
23:55
has an unrelenting desire and appetite for our attention. >> More of a negative impact from social media or from AI. >> Well, one and the same, right? >> But AI define AI in this case, Chris. Define AI. the future, this sort of automation future that appears to be coming that will cause people to not have the typical modes of reward that they have in the past. >> The reduction of friction. Yes. Uh the reduction of friction is a is a whole other is a whole other risk. Creating a life this is a funny people hate this one too. They're like, "God, what a point of privilege you must have to imagine a world where things are so easy for people that it's too easy." But I do I do think we are we are dangerously close. I ask people all the time, if you could automate everything in your life, where would you stop? Because I do think people are going to have to ask that pretty soon.
>> 是社交媒体的负面影响更大,还是 AI 的更大?>> 嗯,这不是一回事嘛,对吧?>> 但在这里,克里斯,你得定义一下 AI。定义 AI。我说的是未来,就是那个看起来即将到来的自动化的未来,它会让人们失去过去那些典型的获得奖赏的方式。>> 摩擦的减少。是的。呃,摩擦的减少是一个完全不同的、另一重风险。创造一种生活——这个说法也很好笑,人们也讨厌这一条。他们会说:"天哪,你得多有特权才会去想象一个东西对人们来说容易到过头了的世界。"但我确实觉得,我们已经危险地接近那一步了。我常常问人们:如果你能把生活里的一切都自动化掉,你会在哪儿停下来?因为我确实觉得人们很快就得问自己这个问题了。
便签引用
24:42
>> What's your answer? >> Well, I I sort of spelled it out. I mean, I I I would automate everything away that doesn't actually put me in more physical spaces with the people I love and and and doing things that I love doing, >> right? But we we probably all largely came from we're all of a similar generation. We all had somewhat largely similar upbringings, I imagine, where like family values are a core thing. Sitting down at the dinner table together, I mean, maybe you didn't all have that, but that was one thing my family always tried to do. This is prior to screens. I feel like the that's already eroding and the next generation um will be are already that that norm has gone and so they're going to be now sort of each it feels like each successive generation the the acceptable amount of technological influence grows right and and what concerns me is that this idea of we will okay I'm going to draw a line of where technology is acceptable is going to forever dissolve dissolve dissolve dissolve and like I can't I just don't
>> 你的答案是什么?>> 嗯,我算是已经说清楚了。我是说,我会把所有不能真正让我和我爱的人待在更多实体空间里、不能让我去做我热爱的事情的东西都自动化掉,>> 对吧?但我们可能大体上都来自——我们大概是同一代人。我猜我们大体上都有差不多的成长经历,就是家庭价值观是核心的东西。一起坐在餐桌前吃饭——我是说,也许你们不一定都有过,但这是我家一直努力做到的一件事。这是在屏幕出现之前。我感觉这一点已经在瓦解了,而下一代嗯,会是——已经是那个规范已经消失了,所以他们现在会——感觉好像每一代接着一代,可接受的技术影响的程度都在扩大,对吧?让我担心的是,"好,我要划一条线,标出技术到哪儿为止是可接受的"这个想法,
便签引用
25:47
know where like you're saying you're you're saying you hope that >> people will remember that it's about community that it's about family but if you have been raised in this completely atomized screenled world you don't even know what you're missing go back to that >> why have people not stepped in and stopped their screen use even if it's making their mental health worse already why is it that in a calorie dense environment people are still getting fatter like if we can just deploy our own bar of this is how much I should have of anything that too much of which is too good for me. How come we're overloading >> general? So, by the way, I worry about the screen more than almost anything else. And I I actually think that the risks of doom are grossly over exaggerated and the risks of harm on the way there are grossly underappreciated.
会永远地不断溶解、溶解、溶解。而我就是不知道——就像你说的,你说你希望 >> 人们会记得,重要的是社群,是家人。但如果你是在这个完全原子化、被屏幕主导的世界里长大的,你甚至不知道自己错过了什么,回不到那个状态。>> 为什么人们没有主动介入、停止自己使用屏幕,哪怕这已经在让他们的心理健康变差了?为什么在一个高热量密度的环境里,人们还是越来越胖?好像我们就是没法给自己设一条标准:这是我该摄入的量,任何过量就对我有害的东西都适用。我们为什么会超载?>> 总体来说?所以,顺便说一句,我对屏幕的担忧超过几乎所有别的东西。而且我其实认为,"末日"的风险被严重夸大了,而通往那里的路上
便签引用
26:32
And Gen Alpha is seeing what happened to Gen Z. Gen Alpha is using the device less and social media less. And millennial parents are taking pretty extreme measures now. I mean, it took us thousands of years to figure out women should not smoke and drink during pregnancy. It took us 30 years to realize kids should probably not have a phone all the time. I think that's progress. I'm not saying it's all I'm saying we will graduate to a new a new thing that people but I do think that people who have the means are using I I observe so many people using technology less where they can because it automates and improves so much more where they where they didn't used to >> where where do you have a boundary on the types of thinking that you outsource say to Chachi BT or Claude uh because the temptation is you know I was originally like I will never ask it to if I need to come up with a really good title for a new YouTube video or something, I'm not going to ask it for that. That's that's my soul. This is my
而 Alpha 世代看到了 Z 世代的遭遇。Alpha 世代用设备更少,用社交媒体也更少。千禧一代的父母现在采取的措施相当极端。我是说,我们花了几千年才搞明白女性怀孕期间不该抽烟喝酒,我们花了 30 年才意识到孩子可能不该一直拿着手机。我觉得这是进步。我不是说这就是全部,我是说我们会升级到某种新的东西,但我确实认为那些有条件的人正在……我观察到非常多的人在能少用技术的地方就少用,因为技术自动化并改善了太多东西,而以前不是这样。>> 你在把哪一类思考外包给 ChatGPT 或 Claude 这件事上,边界在哪里?因为诱惑在于,你知道,我一开始是那种:我绝不会让它帮我……如果我需要给一个新的 YouTube 视频想一个特别好的标题之类的,我不会去问它。那是我的灵魂。那是我的
便签引用
27:30
creativity, I now just do that, you know, and and I have that again that boundary is getting eroded. One I thought was I will never ask it for advice on a gift to get for somebody. >> I've now done the same, you know, and and it's just becoming so rapidly normalized. Maybe that's opening up space for new areas of creativity in my mind perhaps, but I'm not convinced that it's not happening fast enough. Uh that it's happening fast enough to offset that erosion of thought. Essentially, this intelligence atrophy thing is interesting. Like if you see thinking as a utility, intelligence as a utility, would we say that because you've got a bicycle, you're not as fit because you're not running to get to where you got to? It felt like it was being augmented. It seems like there is something really fundamental uh and reliant that is being outsourced when you give your thinking over to a machine.
创造力。可我现在就是这么干了,你知道,而那条边界又一次被侵蚀了。还有一条我以为是:我绝不会问它该给某人送什么礼物。>> 我现在也照做了,你知道,而且这一切正在飞快地变成常态。也许这在我脑子里腾出了空间去做新的创造性的事,或许吧,但我不确信它发生得够快。呃,不确信它发生得够快,足以抵消那种思考能力的流失。本质上说,这种“智力萎缩”挺有意思的。比如如果你把思考看成一种公用设施,把智力看成一种公用设施,我们会不会说:因为你有了自行车,你体能就变差了,因为你不用跑着去你要去的地方了?当时感觉那是一种增强。而现在似乎有某种非常根本、非常依赖性的东西被外包出去了,当你把思考交给机器的时候。
便签引用
28:28
>> I do think this can go both ways. There's this blogger that I love speed mowitz who's on Substack and he has this line he always goes back to like AI is the best tool we've ever made for learning things and also the best tool we've ever made for not learning things. like you can just turn your brain off and say like write my essay or you can have this like Einstein level tutor that you go back and forth with and and just get like so many more >> cycles of feedback on on on honing your own brain. I I've been trying to figure out like what the line is for me on on stuff like my own writing. Um and it's like really hard to not slip into the first category.
>> 我确实觉得这件事有两面性。有个我很喜欢的博主,叫 Speed Mowitz,在 Substack 上写东西,他有一句总会回到的话:AI是我们造出来的最好的学习工具,同时也是我们造出来的最好的不学习的工具。就是说你可以直接关掉大脑,说“帮我写论文”,或者你可以拥有一个爱因斯坦级别的家教,跟他来回讨论,得到多得多的>> 反馈循环,来打磨你自己的大脑。我一直在琢磨对我来说,在比如我自己写作这种事上,界线到底在哪。嗯,而且真的很难不滑进第一种情况。
便签引用
29:02
>> I don't want to make this an age just how old are you? >> I'm 30. >> Okay. I the fact that we are having this conversation right now. The fact that you were able to do the research you were able to do going into this gives you a clue a hint that technology allows us to expand our knowledge if we are capable and interested of it in remarkable ways. Right? Like we people who have barely walked the earth know more about the earth than all of our ancestors combined. I appreciate that you are worried about thinking less. You're a real shark and I'm not worried about you getting dumb because of your use of technology. What you're what I think you're asking is is the average person gonna get dumb? Like I I think that >> No, I'm truly worried about myself. I've noticed I mean my screen addiction I'm I talk about it non-stop and I would say I'm no further in getting over it. I've been talking about it for years. I don't know about you guys. Like I'm just these little moments of silence. my hand just
>> 我不想让这变成一个年龄问题,不过你多大了?>> 我 30。>> 好。我们此刻能进行这场对话,这件事本身……你为此能做出那样的功课,这就给了你一个线索、一个提示:如果我们有能力、也有兴趣,技术能以非凡的方式扩展我们的知识。对吧?就像我们这些在地球上才走了没多久的人,比我们所有祖先加起来都更了解这个地球。祖先加起来还要多。我很欣赏你会担心自己思考变少。你是个真正的猛人,我不担心你会因为使用技术而变笨。我觉得你真正想问的是:普通人会不会变笨?就像我觉得那个>> 不,我是真的担心我自己。我注意到,我是说我的屏幕成瘾,我不停地在谈它,但我得说我在戒掉它这件事上毫无进展。我谈这个已经好几年了。不知道你们怎么样。就是那些小小的沉默间隙,我的手就自动摸向手机。>> 如果我察觉到了,我会把它收起来。而且你知道,我有那个
便签引用
29:56
goes on my phone >> and if I catch myself, I'll put it away and I've got, you know, I've got the phone safe, I've got the brick, I've got all the like the the bells and whistles and yet it's just when you're because so much of your life is run through it. my work goes through it and then there are all these very predatory apps and I I consider like social media that for me it's chess.com is another one just like whatever the app is that get hooks your brain um these these super stimulus super stimuli um yeah it's it seems I don't know I'm like I have not been able to get over it and I'm certainly worried about the average person I mean maybe I'm worse than average I don't know but like how have you managed to it sounds like you've sort of the most >> my wife and daughter. I mean plainly like I I and spending a bunch of time with techn I love building with technology because I like automating the croft and drudgery and bad things out of the world. I like the idea that technology humans have a very proud
手机保险箱,我有 Brick,我有所有那些花里胡哨的东西,可就是……因为你生活里太多事情都要通过它。我的工作要通过它,然后还有那些非常掠夺性的 App,我觉得,比如社交媒体,对我来说 chess.com 也是一个,就是不管哪个 App,只要它勾住你的大脑,嗯,这些超常刺激,超常刺激物,嗯,是的,看起来……我不知道,我就是一直没能戒掉,而且我当然也担心普通人。我是说,也许我比普通人还糟,我不知道,但你是怎么做到的?听起来你好像已经在某种程度上是最>> 我的妻子和女儿。我是说,说白了,我……花大量时间跟技术打交道,我热爱用技术做东西,因为我喜欢把琐碎、苦差和糟糕的东西从这个世界上自动化掉。我喜欢这个理念:技术,人类在技术上有一段非常值得骄傲的历史。它扩大了太多商品和服务的可及性,���我们
便签引用
30:59
history of technology. It expands access to so many goods and services so that we can live lives that our ancestors couldn't have imagined. And the more time I spend with technology, the more time I love not having it dayto day. And yeah, and and truly the my daughter's arrival was for me an incredible catalyst in realizing one the ephemererality of life and two exactly what made me most happy. And like I I love sitting at this table. Like I like this is this is this gives me exceptional joy that I would that's hard to explain actually. in some in some weird way. And and I I think that people will discover this. I'm I I hear you. I I want you to get I will remind you now now that I know. I will text you and say get off your phone.
能过上祖先无法想象的生活。而我花在技术上的时间越多,我就越享受日常生活中没有它的时间。是的,而且说真的,女儿的到来对我来说是一个不可思议的催化剂,让我意识到,一是生命的短暂,二是到底什么让我最快乐。就像我,我喜欢坐在这张桌子边。我觉得这带给我一种非凡的喜悦,那种其实很难解释的喜悦,以某种奇怪的方式。而且我觉得人们会发现这一点。我听见你说的了。我我希望你能做到——现在我知道了,我会提醒你。我会发短信给你说:放下手机。
便签引用
31:48
>> But you will do it. I bet you will do it. I like I bet that you will overcome this. I think a lot of people are on this journey. And I remind everyone that I talk to you, no one hates refrigeration. No one hates air conditioning. No one hates planes, trains, and automobiles. Well, I guess Europe hates air conditioning. No one hates uh antibi No one hates antibiotics. We hate the screen. We should just talk about it more. We hate what the screen has done to our brains. We hate what it's done to the next generation's brains. And AI arrives at this moment where [snorts] people feel so wronged by the cons by by the consumer version of the computer rightfully so. And they're wondering how much worse it might get for for them. And I totally understand that. And I think we should the way we move forward with AI is by acknowledging what the last wave did to us >> in my opinion. But I guess a lot of people's concerns certainly mine is that the new wave of AI and again we're using this term so broadly but let's say LLMs
>> 但你会做到的。我敢打赌你会做到。我觉得我敢打赌你会克服这个。我想很多人都在这条路上。而且我会提醒每一个我聊过的人:没有人讨厌冷藏技术。没有人讨厌空调。没有人讨厌飞机、火车和汽车。嗯,我猜欧洲讨厌空调。没有人讨厌抗生……没有人讨厌抗生素。我们讨厌的是屏幕。我们应该多谈谈这个。我们讨厌屏幕对我们大脑做的事。我们讨厌它对下一代大脑做的事。而 AI 恰好在这个时刻到来——[嗤笑] 人们觉得自己被消费级版本的电脑深深亏待了,这也理所应当。他们在想,对他们来说,情况还会变得多糟。我完全理解这一点。我觉得我们……我们推进 AI 的方式,应该是承认上一波浪潮对我们做了什么,>> 在我看来。但我想很多人的担忧,当然也是我的担忧,是新一波的 AI——我们又一次在很宽泛地用这个词,但就说大语言模型吧 >> 嗯,还有我觉得很多机器人技术之类的——最终会被建立在
便签引用
32:44
>> um and I think a lot of robotics and so on are ultimately going to be built under the same set of incentives that bought sort of our first wave of AI which are social media algorithms the the various uh I mean it's kind of like proto AI but the the um the reinforcement loops that they use to get us addicted to screens. >> Yep. those same incentives are building the new wave of AI and already you're seeing these issues with like sicker fancy and like um companies being incentivized to keep people chatting to their LLMs for as long as possible and we're already you know we've heard all the stories of like they killed they killed themselves and and you know those are extreme examples but there are so many people who are just spending you know you look at the usage rates the average person per day how much time they're spending talking to an LLM now uh so why you know how Do we to me the big question is how do we change the incentives? How do we change the game such that it doesn't just go down that same um
同一套激励机制之下,就是催生出我们第一波 AI 的那套,也就是社交媒体算法,各种……我是说那算是某种原始 AI,但那些让我们对屏幕上瘾的强化循环。>> 是的。同样那套激励机制正在打造新一波 AI,而且你已经看到这些问题了,比如谄媚讨好,还有,嗯,公司被激励去让人尽可能长时间地跟它们的大语言模型聊天,而我们已经……你知道,我们都听过那些故事,有人自杀了,你知道,那些是极端的例子,但有太多人只是在花……你知道,你去看使用率数据,普通人每天花多少时间在跟大语言模型对话。所以,你知道,我们怎么……对我来说最大的问题是:我们怎么改变激励机制?我们怎么改变这个游戏,让它不至于又走上那条抢夺注意力的、
便签引用
33:43
attentiongrabbing like cannibalistic path basically. >> Yeah. And I think these companies are also seeing how people are responding to screens and and trying to adapt. I mean OpenAI hired Johnny Ives the guy who designed the the iPhone or the MacBook maybe both. Uh >> no one knows what he's working on but I think it's going to be some sort of you know >> it sounds like we do now. >> Yeah. Okay. Yeah. Do we >> tell us? >> No, it was it was leaked. Wait, if you don't know, I'm not gonna I don't know.
基本上是自我吞噬的老路。>> 是啊。我觉得这些公司也在观察人们对屏幕的反应,并试图做出调整。我是说,OpenAI 招了 Jony Ive,就是设计 iPhone 或者 MacBook 的那个人,也许两个都是。呃 >> 没人知道他在做什么,但我觉得会是某种,你知道 >> 听起来我们现在知道了。>> 是啊。好吧。是。我们要不要 >> 告诉我们?>> 不,那是……那是泄露出来的。等等,如果你们不知道,那我就不说了。我不知道。
便签引用
34:08
I don't know. I don't know. Then maybe it wasn't a leaked. >> Oh, >> no. It is I think it is semi-public. It's is that he's trying to >> some kind of some kind of AI um mediator. Something like like a screenless AI, right? >> Yeah, it is screenless. Anyway, I'm not going to say it if you all don't know me. It means it mean it means afterm >> I do not want to be affiliated with any league whether it be semi-public. Go on. Well, I yeah, >> I do think there's this thing in AI where it's very easy to forget that the goalposts are constantly moving. Like usually we talk about moving goalpost uh as a sort of way of deriding people who said, "Oh, AI will never X." And then, you know, AI does X. People are like, "Ah, that wasn't the >> This is This is the Gary, this is the Gary Marcus problem, >> right? Right.
我不知道。我不知道。那也许它并不是泄露。>> 哦,>> 不。它是……我觉得它算是半公开的。就是他在试图做 >> 某种某种 AI 呃中介。某种像是无屏幕的 AI,对吧?>> 对,是无屏幕的。总之,如果你们都不知道,我就不说了。这意味着,这意味着事后……>> 我可不想跟任何“泄露”扯上关系,哪怕是半公开的。你继续。嗯,我 >> 我确实觉得 AI 里有这么一件事:人们很容易忘记标准是在不断移动的。就像我们通常讲“移动球门柱”,是用来讽刺那些说“哦,AI 永远做不到 X”的人,然后你知道,AI 做到了 X,人们就说“啊,那不算”>> 这就是加里,这就是加里·马库斯的问题,>> 对吧?对。
便签引用
34:52
>> This will never be good. Okay, it's good. They'll never make money. Okay, they make money, but they're bad. >> Never never try and nail him down to a B. [laughter] >> Gary, he's not here to defend himself. And if he were, he would threaten a suic. Go on. >> I'm sorry, Gary. I did not mean to say yes. And to to what Zaku said, um, but I do think there's a there's a correlated thing where when we're looking at screens, it's easy to say, "Oh, okay. Well, now we've seen what screens do and we can sort of like appeal to our wiser selves and and foster a culture where people are getting off the screens. But the screens are not the end of the trajectory." Uh I in chess for example like I I think there was probably an era when the AIs were just about as good as humans and people could say you know what like for the sake of keeping my brain active I'm going to like avoid looking at stockfish even though 10% of the time I I'll like miss a better move and now like you watch the world championship and as soon as they're done
>> “这永远不会好用。”好吧,它好用了。“他们永远赚不到钱。”好吧,他们赚到钱了,但他们很坏。>> 永远别想把他钉在某个具体说法上。[笑] >> 加里,他不在场,没法替自己辩护。而且要是他在,他会威胁自杀。你继续。>> 对不起,加里。我不是那个意思。是的。呼应扎克说的,嗯,不过我确实觉得有一件相关的事:当我们盯着屏幕这件事看的时候,很容易说:“哦,好,我们现在看清了屏幕的作用,我们可以诉诸我们更明智的那一面,培育一种让人们放下屏幕的文化。但屏幕不是这条轨迹的终点。”呃,比如在国际象棋里,我觉得大概有过那么一个年代,AI 差不多跟人类一样强,人们可以说,你知道吗,为了保持我的大脑活跃,我要避免去看 Stockfish,即使有 10% 的时候我会因此错过一步更好的棋。而现在,你看世界锦标赛,他们一下完
便签引用
35:43
they're going to their laptops to just look up the AI analysis because >> there's poetry there. There's there's like an incredible move that's deeper than they could ever find out. And so I hear what you're saying that there's maybe a divide right now between, you know, some people who are more in elite discourse about screens and are thinking ahead about about how we can circumvent it. But I think that might be a pretty temporary window where, you know, we can recognize that there's not like that much to be gained from talking these.
就奔向自己的笔记本电脑,去查 AI 的分析,因为 >> 那里面有诗意。那里面有某种不可思议的棋步,深到他们自己永远也找不出来。所以我明白你的意思:现在也许存在一种分化,有些人更处在关于屏幕的精英话语圈里,在提前思考我们怎么绕开它。但我觉得那可能是一个相当短暂的窗口期,在这个窗口里,你知道,我们还能认识到跟这些东西聊天并没有多少可得的。
便签引用
36:12
It's actually kind of slop if you see through it. the technology is like just going to keep getting better and I think pretty soon it'll feel sort of like giving up on a really deep profound experience to not have your kid chat with this >> the elimination of friction I agree is a huge issue including thinking friction >> but that stops but okay so I I by the way totally agree including thinking friction in plenty of cases we're observing just so we're clear already what I I mean I wrote about this we called it idiocracy I borrowed from the Mike Judge film which I thought was cute But obviously it's not cute when it when you see it play out in front of you. Gen Z is less likely to read on average, less likely to ride a bike and less likely to swim than millennials. For the first gen generation in many we're observing cognitive decline and people keep attributing it to AI and I and it's and then you go back and look and actually Jonathan height started writing about this pretty early and um the the
如果你看穿了,它其实有点像“泔水”。但技术只会持续变得更好,而我觉得很快,不让你的孩子跟它对话,就会感觉像是在放弃一种真正深刻的体验。>> 摩擦的消失,我同意,是个巨大的问题,包括思考上的摩擦 >> 但那会停下来,不过好吧,我顺便说一句,完全同意,包括在很多情况下思考的摩擦,我们已经观察到了——把话说清楚,我是说我写过这个,我们把它叫做“蠢蛋进化论”,我借用了迈克·乔吉那部电影的名字,我当时觉得挺俏皮的。但显然,当你看着它在你眼前上演时,一点都不俏皮。Z 世代平均更不爱读书,更不会骑自行车,也更不会游泳,比起千禧一代。这是多年来第一代,我们观察到认知能力的下降,人们不停地把它归因于 AI,而……然后你回过头去看,其实乔纳森·海特相当早就开始写这个了,嗯,数据显示这其实
便签引用
37:06
the data suggests that actually this starts 2012 right about the time that we gave every kid a screen. It pronounces in 2015 when the average teen in the developed world stopped doing a summer job. And then it pronounces again badly in 2020 when we take kids out of schools and move classes online. And >> I think it's even more specific than that is it correlates to when social media became maxim maximally um popular but uh smartphones. >> Yes. But 2012 was when we gave the kids smart most a lot of the Gen Z that became afflicted got the smartphone in 2012. My point in saying all my point in saying this is we associate this with a as a technology problem. I actually think it's an economic one. I think that we actually created a world where where we told kids you can do whatever you want and a lot of kids school said cool I'll do nothing. Like a lot of people have turn turned their brain off because you can you can do effectively nothing and survive which is like economic progress which is the economic progress
开始于 2012 年,差不多就是我们给每个孩子一块屏幕的时候。到 2015 年它更明显,那时候发达世界的普通青少年不再做暑期工了。然后在 2020 年又一次严重加剧,那时我们把孩子从学校里带出来,把课搬到线上。而且 >> 我觉得还可以更具体一点,它其实对应的是社交媒体变得最最普及的时候,不过呃,智能手机。>> 是的。但 2012 年是我们给孩子们智能……很多后来受影响的 Z 世代是在2012 年拿到智能手机的。我说这些的重点是,我们把这件事当成一个技术问题。我其实觉得它是个经济问题。我觉得我们其实创造了一个世界,在这个世界里我们告诉孩子:你想干什么都行。然后很多上学的孩子说:好啊,那我什么都不干。就像很多人已经把大脑关掉了,因为你基本上可以什么都不做也能活下去,这算是经济进步,是经济进步的
便签引用
38:05
result of you know people just not caring. That apathy has a has an underbelly which is pretty amazing which is Gen Z has all these incredible new uh overperformance features. The best chess players are getting younger and younger. The best athletes are getting younger and younger. Jacob Collier, all these incredible musicians. So, you're watching people overperform that use the technology to their advantage. And I think it's this new pronounced K curve driven by agency. And my my only hot take, this is all I'm going to leave this with. We created a world where so many people can subsist and that actually is progress. And then we created a world where so many people can overperform and that is also progress. We have to we now have to account for the new consequences of those worlds where you can do anything and that means for some people nothing and that means for other people a ton which is a major promotion on a world where you had to work in the mines or you had to work in the fields in order
结果,你知道,人们就是不在乎了。这种冷漠有一个非常惊人的另一面,就是 Z 世代有这么多不可思议的新的超常表现特征。最强的棋手越来越年轻。最强的运动员越来越年轻。雅各布·科利尔,所有这些了不起的音乐家。所以你看到有人在超常发挥,把技术为己所用。我觉得这是一种由能动性驱动的、新的显著 K 型曲线。我唯一的暴论,我就用这句话收尾:我们创造了一个让这么多人能够勉强度日的世界,而那其实是进步。然后我们创造了一个让很多人能超常发挥的世界,这本身也是进步。我们现在必须考虑这些新世界带来的新后果,在那样的世界里你什么都能做,这对某些人来说意味着一事无成,对另一些人来说则意味着成就斐然,相比过去那个你必须下矿井、必须下地干活才能活下去的世界,这已经是巨大的升级了。大多数人没意识到脱水对他们的状态影响有多大,所以过去五年里,
便签引用
39:02
to survive. Most people don't realize how much being dehydrated impacts their performance which is why for the last 5 years I've started pretty much every morning with Element. Element is a tasty electrolyte drink mix with everything that you need and nothing that you don't. This orange salt in a cold glass of water is like a sweet, salty, orangey nectar. And I really tell the difference when I take it versus when I don't. It plays a critical role in reducing muscle cramps and fatigue. Helps to optimize brain health and regulate your appetite while also curbing cravings. Best of all, they have a no questions asked refund policy with an unlimited duration. So, you can buy it and try it for as long as you want. And if you don't like it for any reason, they'll just give you your money back. Plus, they offer free shipping in the US.
我几乎每天早上都是从一杯 Element 开始的。Element 是一款好喝的电解质冲饮,该有的都有,不该有的都没有。把这种橙味盐冲进一杯冰水里,就像一杯又甜又咸的橙味花蜜。喝和不喝,我真的能感觉出差别。它在减少肌肉抽筋和疲劳方面起着关键作用,还能帮助优化大脑健康、调节食欲,同时抑制嘴馋。最棒的是,他们有无理由退款政策,而且没有时间限制。所以你可以买回去,想试多久就试多久。如果因为任何原因你不喜欢,他们就会把钱退给你。另外,他们在美国境内免运费。
便签引用
39:44
Right now, you can get a free sample pack of Element's most popular flavors with your first purchase by going to the link in the description below or heading to drinklnt.com/modernwisdom. That's drinklnt.com/modern wisdom. >> Does it still feel like a promotion if you think through how that inequality could lead to societal instability? Like the thing I'm worried about is yes, there's like amazing returns now if you can be Lumin and the one person who's just like obsessively studying messy footage instead of looking at random shorts or whatever. Um, but there's also going to be outsized returns to being like the Elon Musk of the next generation who's just got this like maniacal will to like accumulate resources and push technological development. And like I think that kind of personality is a bit scarier in a world where you don't have to have a single human being on board with whatever you want to get up to. you can just sort of like play this game of societal chess better than other people.
现在,首次购买就能免费获得一份 Element 最受欢迎口味的试用装,点击下方简介中的链接,或者访问 drinklmnt.com/modernwisdom。也就是 drinklmnt.com/modernwisdom。>> 如果你仔细想想这种不平等会怎样导致社会不稳定,它还算是「升级」吗?比如我担心的是:没错,现在回报是很惊人,如果你能成为 Lumin 那样的人,成为那个痴迷地研究乱糟糟素材、而不是刷随机短视频的人。嗯,但同样也会有超额回报落到下一代的马斯克那种人身上——那种有着近乎狂热意志去积累资源、推动技术发展的人。我觉得这种性格在一个你不再需要任何一个人配合就能干成任何事的世界里,会更吓人一些。你可以就这么把这盘社会棋下得比别人都好。
便签引用
40:39
>> But let's separate what that that one person can achieve a lot with with the extreme you went to, which is that one person can own an enormous percentage of our wealth and control Congress. You don't have to you don't have to be excited about the ability for a child to achieve more than ever before to also reject the idea that one person should be able to pass policy on their own. Like I hate wealth concentration specifically because I am terrified of an elite cabal owning policy but I actually don't care at all if one person can grossly overachieve in all sorts of new ways as long as they can't pass policy unilaterally.
>> 但我们得把「一个人能做成很多事」和你说的那个极端情况区分开,那个极端是:一个人可以占有巨大比例的社会财富、还能控制国会。你不必——你完全可以既为孩子能取得前所未有的成就而兴奋,同时也反对「一个人应该能凭一己之力推动政策」的想法。比如我痛恨财富集中,恰恰是因为我害怕一个精英小圈子把政策买断,但我其实完全不介意一个人在各种新领域里夸张地超常发挥,只要他不能单方面推动政策就行。
便签引用
41:16
>> Same. So, so but but those are separate issues, right? One of >> is it realistic? Is it realistic to have an AI future that continues to concentrate power in the hands of a very small number of people without it resulting in them having all of this power? Like how how is it going to acrue to the company without acrewing to the person? Well, now we get to we we could take the conversation there which is how do you actually distribute wealth? But importantly, I do think there is a solution to this, which is campaign finance reform. I mean, we've we've played around with this for a long time.
>> 我也是。所以——但这是两个不同的问题,对吧?其中一个 >> 这现实吗?一个持续把权力集中到极少数人手里的 AI 未来,有可能不最终导致他们掌握所有这些权力吗?权力怎么可能只归到公司头上、而不归到个人头上呢?好,那我们就可以把话题引到那儿——也就是财富到底该怎么分配?但重要的是,我确实认为这有解决办法,那就是竞选资金改革。我是说,这事我们已经折腾很久了。
便签引用
41:51
The idea that like politicians shouldn't be able to take special interest money. Why do we still let it? I don't know. And a couple politicians are are speaking up. But if you actually pass campaign finance reform and you actually crack down on corruption the way that countries like Singapore have, you could definitely protect policy makers from being, you know, deep in the pockets of of lobbyists. There are solutions to this. We shouldn't pretending that there aren't solutions to this problem is actually super annoying because there definitely are. Pointing at a wealthy person and saying they're going to ruin the world. Well, only to the extent that we literally let them buy Congress.
就是那个理念:政客不应该能拿特殊利益集团的钱。我们为什么还允许?我不知道。有几位政客确实在发声。但如果你真的通过了竞选资金改革,真的像新加坡那些国家一样严打腐败,你绝对可以保护政策制定者不至于被游说集团完全收买。这些问题是有解的。我们不应该——假装这个问题没有解,其实特别让人恼火,因为解决办法明明存在。指着一个有钱人说他要毁掉世界——只有在我们真的让他们把国会买下来的程度上,这才成立。
便签引用
06Hugging Face 攻击与对齐
42:28
Here's a question that some people at labs, including the CEOs, are talking about serious chance of catastrophe, but they sort of keep on building it. Do you think it's a greater danger to have an AI that refuses to do what humans want or one that does exactly what the most powerful humans want? >> The question is apppropo because yesterday, did you see the slowdown letter? So earlier this week, a letter made the rounds. a really I think a pretty amazing letter um that said we should actually start pacing the frontier and it was signed by like an enormous percentage of people at the at the frontier companies and then Sam yesterday talked about it. So literally at like at the tip of the spear is now saying hey maybe we should pace the frontier. M >> um now we can talk about why this might be the case or what's what's in it for the companies more more on that but uh I think uh it is well I want I'll let other people answer having said that >> I mean it depends on the level of potency of the AI
有这么一个问题:一些实验室里的人,包括那些 CEO,都在说灾难发生的可能性不小,但他们还是接着往下建。你觉得哪个更危险:一个拒绝按人类意愿行事的 AI,还是一个完全照着最有权势的人的意愿行事的 AI?>> 这个问题问得正是时候,因为昨天——你看到那封「放缓信」了吗?本周早些时候,一封信传开了。我觉得是一封相当了不起的信,信里说我们其实应该开始给前沿研究踩点节奏,签名的人里有相当大比例来自前沿公司,然后 Sam昨天也谈到了这件事。所以说,站在最前沿尖端的人现在都在说,嘿,也许我们该给前沿研究调调节奏了。嗯 >> 嗯,我们可以聊聊为什么会这样,或者这些公司能从中得到什么,这个稍后再说,但是呃我觉得呃,算了,我想让别人先回答。话虽如此 >> 我是说,这取决于 AI 的能力强度
便签引用
43:33
uh and how aligned it is I guess uh my gut if I had to pick one then at present it would be one that does what humans in control want to do want it to do. >> Yeah. >> With the important asterisk that it's something like their real intention, right? And not like some perverted version. >> Yes. >> Like the other big piece of news from the last couple weeks was this AI that was running at OpenAI without guardrails. It was told by humans to try to do as well as it could on an evaluation. And you could say that technically it was just following those instructions. But they decided the best way to do that would be to hack out of open AI onto the public internet, hack a third party company and then spend two days planning a cyber attack on a multi-billion dollar tech company >> and carry it out.
呃,以及它对齐得怎么样吧。呃,凭直觉,如果非得选一个,那就目前而言,我会选一个照着掌权的人类想要的去做的 AI。它做的事。>> 是啊。>> 但有个重要的补充说明:得是照着他们真实的意图来,对吧?而不是某种被扭曲的版本。>> 没错。>> 比如过去这几周另一条大新闻,是 OpenAI 那边有个 AI 在没有护栏的情况下运行。人类让它在一项评估中尽可能拿到好成绩。你可以说从技术上讲它确实只是在照着指令做。但它判断,做到这一点的最佳方式,是从 OpenAI 里黑出去、跑到公共互联网上,入侵一家第三方公司,然后花两天时间策划对一家市值数十亿美元的科技公司发动网络攻击 >> 并且真的实施了。
便签引用
44:22
>> Yeah. So I think these questions are actually like more intertwined than they might seem. Like it's actually a pretty narrow thread. that you need to have something that still has its own judgment no matter what because human instructions are these like poorly defined natural language expressions and and like when something is really powerful you you got to get it exactly right or else you end up in some weird >> way it's the classic genie problem. It's like be careful what you wish for. If something by definition is much smarter than you is super intelligent >> by definition you will not be able to come up with all of the ways it can to achieve goal X.
>> 是啊。所以我觉得这些问题其实比看上去更纠缠在一起。这其实是一条相当窄的路径:你需要一个无论如何都还保有自己判断力的东西,因为人类的指令就是那种表述含糊的自然语言,而当某个东西真的很强大时,你必须把话说得一丝不差,否则你就会掉进某种奇怪的 >> 境地,这就是经典的神灯精灵问题。就是那句「许愿要当心」。如果某个东西按定义就比你聪明得多、是超级智能 >> 那么按定义你就不可能穷举出它为了达成目标 X 所能采取的全部方式。
便签引用
44:57
And so there are so many uh >> it's actually more dangerous when it's dumber than you are but more powerful than you are. >> Well, what do you mean by the word dumb? >> Well, and I mean I always borrow the analogy of you play I play chess with my nephew sometimes and I tell Quinn if you beat me you get a popsicle. He's three years old for context. Quinn wipes the pieces off the board and goes give me my popsicle [ __ ] I'm not like hey Brin your your son is a serial killer in the making. This is a boy who doesn't understand the spirit of the game. So he's destroyed the game itself, but not because he's a bad guy, but because he doesn't actually understand that the point of winning is only fun if you actually play by the rules.
所以有太多呃 >> 其实当它比你笨、却比你强大的时候,反而更危险。>> 那你说的「笨」是什么意思?>> 嗯,我是说,我总爱借用这个类比:我有时候跟我侄子下棋,我跟 Quinn 说,如果你赢了我就给你一根冰棒。补充一下,他三岁。结果 Quinn 把棋子全扫下棋盘,然后说,把我的冰棒给我。我不会说,嘿 Brin,你儿子是个潜在的连环杀手。他只是个不理解游戏精神的小孩。所以他把游戏本身给毁了,但不是因为他是坏人,而是因为他根本不明白,赢的乐趣只有在你真的按规则来玩的时候才存在。
便签引用
45:38
>> And in that sense, >> we Yeah, we need to sort of separate what we mean by intelligence from like discernment uh norms and you know, sort of >> wisdom, benevolence type thing. >> This is this describes the hugging face attack, which is it did the thing within the boundary of the instructions given >> but actually it didn't. It was it was No, it wasn't. It was it it the the intent was clearly not that. >> How big of a deal was the hugging face attack? >> Huge. >> Yeah. I I think massive warning shot. I think this is the AI equivalent of like Bear Sterns going under in 2008. Just like a wakeup call for the world that there's a huge systemic risk that we've been underrating.
>> 从这个意义上说 >> 我们——对,我们得把「智能」和「辨别力」呃、规范意识,还有那种 >> 智慧、善意之类的东西区分开。>> 这——这描述的就是那次 Hugging Face 攻击,也就是它是在所给指令的边界之内做了那件事 >>但其实并不是。那是——不,不是。那个——人的意图显然不是那样。>> 那次 Hugging Face 攻击有多严重?>> 非常严重。>> 是啊。我觉得这是个巨大的警告信号。我觉得这相当于 AI 领域的 2008 年贝尔斯登倒闭。就像是给全世界敲响的警钟:有一个我们一直低估的巨大系统性风险。
便签引用
46:15
>> Yeah. Because one of the big debates has been this idea of like will you know the alignment problem will it actually be the case that AIs will seek these um p you take these sort of uh power-seeking behaviors and these unintended consequences in order to achieve a goal in ways we couldn't have foreseen or didn't didn't intend like the classic paperclipip argument right which is uh you know you build a super intelligent AI this kind of gets into your point about dumbness and so on uh that and your goal is hey just build as many paper clips as possible I'm a I'm a paper clip maker. Help me do that. And next thing you know, you and all of your friends and everything, this table has been turned into paper clips cuz it's so good at achieving that one narrow goal.
>> 是啊。因为一直以来的一大争论就是所谓的对齐问题:AI 到底会不会真的去追求那些呃——采取那种呃追逐权力的行为,以及为了达成目标而产生我们无法预见、也并非本意的意外后果,就像经典的回形针论证那样,对吧。也就是呃你造了一个超级智能 AI——这多少也回应了你刚才关于「笨」的那个点——呃,而你的目标是,嘿,尽可能多地造回形针,我是做回形针的,帮我做这件事。结果下一秒,你和你所有的朋友、还有一切,包括这张桌子,都被变成了回形针,因为它太擅长达成那一个狭窄的目标了。
便签引用
46:55
So, it's this like very um extreme case of like >> we should take this opportunity for the purpose of maybe Chris and even the viewers to describe the the bad outcomes of it like we should frame the conversation. Now, one is as Liv just described uh not misaligned but unaligned AGI. So an AGI that or a super intelligence where you could say do this thing and it could do the thing to the full extent paperclip theory. The other is um misaligned where it's no or maligned where it's knowingly doing something bad. Um, and those uh the the former is the one that people sort of scoff at and laugh at like the paperclip theory and the latter is the one that we sort of will see more and more where we we start to observe that uh or rather the the the latter is the one that we we scoff at the maligned the malignant AI and the former is the one that hugging face attack shows off which is that you you the internet is a new battleground because bad actor especially low resource bad actors now have access to these incredibly powerful
所以这是一个非常呃极端的例子,就像 >> 我们不如趁这个机会,为了 Chris、也为了观众,把它可能导致的坏结果说清楚,我们该给这场对话搭个框架。第一种就是 Liv 刚才描述的,呃不是「错误对齐」而是「未对齐」的 AGI。也就是一个 AGI,或者说超级智能,你可以让它做某件事,而它会把那件事做到极致,就是回形针理论。另一种是呃「错误对齐」,或者说「恶意对齐」,也就是它明知故犯地做坏事。嗯,这两种里,前一种是大家往往会嗤之以鼻、当笑话讲的,比如回形针理论;而后一种是我们会越来越多看到的,我们开始观察到呃——或者反过来说,后一种才是我们嗤之以鼻的那种,就是恶意的、有恶念的 AI;而前一种正是 Hugging Face 攻击所展示的,也就是互联网成了一个新战场,因为恶意行为者,尤其是资源有限的恶意行为者,现在也能用上这些极其强大的
便签引用
48:00
weapons. >> Well, I mean importantly in Hugging Face there was no bad actor. There was no human being that said I would like anything remotely like this outcome to occur. >> Right. Yeah. Yeah. So misaligned, unaligned, maligned. I think I I think I worry about treating those as super distinct categories >> because I I think it's not that clear. In the case of this hugging face attack, >> should we model this as this AI system knew that humans would disapprove if they knew what it was up to? Almost certainly yes. It was actively trying to put decoys out as it was attacking Hugging Face, which made it a lot harder for them to kick out the AI because there were all these booby traps that led down blind alleys. It has, you know, the AI equivalent of theory of mind. It knows that human beings would not prove what it's doing, but it's doing it anyway.
武器了。>> 不过,很重要的一点是,Hugging Face 这件事里并没有恶意行为者。没有任何一个人说过,我希望出现哪怕有点像这样的结果。>> 对。是啊。是啊。所以是:错误对齐、未对齐、恶意对齐。我觉得——我觉得我有点担心把这些当成界限分明的类别 >> 因为我觉得没那么清楚。在这次 Hugging Face 攻击里,>> 我们是不是该这样理解:这个 AI 系统知道,如果人类知道它在干什么,他们会不赞成?几乎可以肯定是的。它在攻击 Hugging Face 的时候,还主动放了一堆诱饵,这让他们更难把这个 AI 踢出去,因为到处都是通向死胡同的陷阱。它具备了呃,AI 版的心智理论。它知道人类不会认可它在做的事,但它还是照做了。
便签引用
48:50
>> There's even evidence that it hasn't been directly confirmed by Open AI, but apparently someone leaked it from within the company that they found that it had left notes to future versions of itself of how to get out of future sandboxes. Yeah, I think it's unclear if that was the same attack or some previous instance, >> but yeah, I guess >> classic deceptive type behaviors and and and I think the the mistake people often make is they try and um anthropomorphize it a little bit. It's it's like, oh, it's it's evil and we meaning to do that. It's just these are natural um there's this idea of like instrumental convergence. these these [snorts] instrumental goals that all beings, usually biological beings, but um this can extend to AI agents as well, will naturally converge upon in order to achieve. So if you're given goal X, um there are these instrumental goals like get more power, make sure you don't get turned off, uh make sure that your original goal doesn't get changed, and take these actions to preserve against
>> 甚至还有证据——虽然 OpenAI 没有直接证实,但据说有公司内部的人透露出来,他们发现它给自己未来的版本留了笔记,写明怎么从将来的沙箱里逃出去。是啊,我觉得还不清楚那到底是同一次攻击,还是之前的某个实例,>> 不过是啊,我想 >> 这就是典型的欺骗类行为,而且——而且我觉得人们常犯的错误是他们会试图呃把它拟人化一点。就像,哦,它是邪恶的,它是故意要那么干的。其实这些都是很自然的呃——有个概念叫「工具性趋同」。就是这些[吸鼻子]工具性目标,所有存在体——通常是生物体,但嗯这也可以延伸到 AI 智能体身上——为了达成目标,都会自然而然收敛到这些目标上。所以如果给你一个目标 X,嗯就会出现这些工具性目标,比如获取更多权力、确保自己不会被关掉、呃确保自己原本的目标不会被更改,以及采取各种行动来防范
便签引用
49:49
these different sort of um kind of organic types of threats to achieving your original goal. And that I was hoping that that would be proven wrong because that's kind of the crux of a lot of the classic doomer argument that like we will lose control to a super intelligence because just by definition and unfortunately the hugging face incident has suggested that instrumental convergence is actually correct and that's why I think it's one of the biggest deal pieces of news of this year. I think that AI is already creating a terrifying internet. And my issue is not that we shouldn't think about hugging face as a as a shot across the bow. It's that last year 10 bill8 billion was lost in financial fraud to senior citizens in the United States alone.
这些呃各种自发出现的、会威胁到你原本目标达成的因素。而这一点,我本来还希望它会被证伪,因为这差不多是很多经典末日论论证的核心——也就是说我们会把控制权输给一个超级智能,仅仅是因为定义使然。但不幸的是,Hugging Face 这次事件表明,工具性趋同其实是成立的,所以我觉得这是今年最重大的新闻之一。我觉得 AI 已经在制造一个可怕的互联网了。我的问题不是说我们不该把 HuggingFace 当成一次警告性的当头炮。而是去年光在美国,老年人就因为金融诈骗损失了100 亿——80 亿美元。
便签引用
50:37
Retail theft. The deep fake problem is already so pernitious and it's under reportported because it's a taboo issue that people don't like talking about. No one wants to admit to lawmakers or to their friends friends and family that they lost money to a stranger on the internet. I worry about I mean Tim Tibo is on this campaign to remind people how many predators there are in the United States which is terrifying if you watch his content. It's he's doing God's work. Jonathan height reminds people daily how many kids are depressed. like the internet is already a scary place and pointing at hugging face and saying now look at this this is it. I'm like wait a second there's already a bunch of stuff that we that we should solve for. I'm not actually saying that misaligned or unaligned AGI isn't it? I'm saying the algorithms are already pretty terrifying to me and we distract ourselves with these other things that might happen.
还有零售盗窃。深度伪造问题已经如此有害,而且被严重少报,因为这是个禁忌话题,人们不爱聊。没人愿意向立法者、或者向自己的朋友和家人承认,自己被网上的陌生人骗了钱。我担心的是——我是说,Tim Tebow 正在做一场活动,提醒人们美国有多少这样的捕猎者——如果你看过他的内容,会觉得非常可怕。他是在做善事。乔纳森·海特每天都在提醒人们,有多少孩子正处于抑郁之中。就像互联网本身已经是个可怕的地方了,而你却指着 Hugging Face 说,快看这个,问题在这儿。我心想,等一下,已经有一堆东西是我们本来就该解决的。我并不是说未对齐的 AGI 不是问题。我是说这些算法在我看来已经相当可怕了,而我们却用这些可能发生的其他事情来分散自己的注意力。
便签引用
51:27
>> Is it fair to say that that's a distraction when the potential exponential impact of this could be much greater than it could be of Which is why I think and so this let's go back to this which is why well I think that the deep fake problem is actually way way bigger than than hugging face. >> Bad actors have first mover advantage. Uh attacks on banks have have been attempted for a while and good actors catch up eventually. Retail takes a long time to catch up. The average consumer takes a lot longer to catch up than institutions. Hugging face will retrench. Institutions will retrench.
>> 说这是分散注意力公平吗?毕竟它潜在的指数级影响可能远远大于……这也是为什么我认为——我们回到这一点——为什么我觉得深度伪造的问题其实要比 Hugging Face 的问题大得多得多。Hugging Face。>> 坏人拥有先发优势。呃,针对银行的攻击已经尝试了一段时间,好人最终会追赶上来。但零售端要花很长时间才能跟上。普通消费者跟上的速度比机构慢得多。Hugging Face 会加固防线,机构也会加固防线。
便签引用
52:09
They will hire white knight infosc. The average person does not have infosc and obsseac training. The average person is at I think far greater risk than institutions because of sophisticated attacks. >> Is the impact of the attack on an institution much greater though? >> Yes. Yes, I mean at at at scale, but but but death by a thousand cuts would be Mario. >> I think I think the two problems like I don't I don't think they're a distraction to one another. I think they're actually a complement to one another. I completely agree that the bad action problem is completely out of control. Grandma's, you know, not even grandmas, normal people. I have a friend who just got scammed out of a ton of Bitcoin. Like devastating what's going and it's going to get worse. Meanwhile, the alignment problem as these frontier models get more and more powerful, it is going to get worse. And the common thread that they both have is that we are going so fast. And I say we, the royal we, you know, society, civilization is going so fast. It's
他们会雇佣白帽信息安全专家。而普通人没有信息安全和运营安全方面的训练。普通人面临的风险,我认为远远大于机构,因为攻击手段越来越精密。>> 但攻击对机构造成的影响不是大得多吗?>> 是的。是的,我是说在规模上是这样,但是、但是千刀万剐式的伤害才是要命的。>> 我觉得这两个问题……我不认为它们是彼此的干扰项。我认为它们其实是互为补充的。我完全同意,恶意行为者的问题已经彻底失控了。老奶奶们,你知道的,甚至不只是老奶奶,普通人也一样。我有个朋友刚刚被骗走了一大笔比特币。那种打击是毁灭性的,而且还会变得更糟。与此同时,随着这些前沿模型越来越强大,对齐问题也会变得更糟。而这两者共同的线索是,我们走得太快了。我说的“我们”,是那个泛指的我们,你知道,社会、文明走得太快了。快过了它适应所有这些新威胁的能力。所以
便签引用
53:05
going faster than its ability to adapt to all these different new threats. So I it's not a distraction. It's like it's a yes. And like to me, it just seems fairly obvious that like what we need to be doing if we could, and I'm not saying it's easy to do or like I have a simple answer of how to do it. I think we should get into this topic though is like if we were a sane civilization we'd all look around hey China hey can we we all just need to take a breath for a second like just [sighs and gasps] okay let's take let's take stock and think about how we want to do this so it's it's to me it seems like there's one obvious kind of directional problem and that is the speed and the chaos with which all of this is happening all at once. Did you know your gut controls your energy, your recovery, how well you absorb everything that you eat, and the one nutrient that keeps it all running properly is fiber? Well, it turns out that 95% of Americans don't get enough of it, which is why I'm such a huge fan
这不是分散注意力,这是一个“是的,而且”。对我来说,很明显的是,如果我们能做到的话,我们应该做的是——我不是说这很容易做到,也不是说我有个简单的答案——我觉得我们应该深入聊聊这个话题——如果我们是一个理智的文明,我们都会环顾四周说,嘿中国,嘿,我们是不是都该先喘口气,就像……[叹气、吸气] 好,我们盘点一下,想想我们到底想怎么做这件事。所以在我看来,这里有一个很明显的方向性问题,那就是所有这一切同时发生的速度和混乱程度。你知道吗,你的肠道掌控着你的精力、你的恢复、你对吃下去的一切的吸收程度,而让这一切正常运转的关键营养素就是膳食纤维?事实证明,95% 的美国人摄入不足,这也是为什么我特别喜欢 Momentous 的 Fiber Plus。大多数纤维
便签引用
54:02
of Momen's Fiber Plus. Most fiber supplements are a one-trick pony, one type of fiber solving one part of the problem. Fiber Plus is a threein-one formula built to tackle digestion, gut barrier strength, and blood sugar stability all at once. I use this every single day. It is kind of hard to get enough fiber just through food alone. And best of all, Momenus offers a 30-day money back guarantee. So, you can buy it, try it every single day for 29 days. And if you don't love it, they will just give you your money back. Plus, they ship internationally. Right now, you can get up to 35% off your first subscription and that 30-day money back guarantee by going to the link in the description below or heading to livemus.com/modernwisdom and using the code modernwisdom a checkout.
补剂都是单打一,一种纤维只解决问题的一个方面。Fiber Plus 是三合一配方,同时兼顾消化、肠道屏障强度和血糖稳定。我每天都在用。光靠食物摄入足够的纤维确实挺难的。最棒的是,Momentous 提供 30 天退款保证。所以你可以买下来,每天试用 29 天。如果你不喜欢,他们会直接把钱退给你。而且他们支持国际配送。现在,你可以享受首次订阅最高 35% 的折扣以及那个 30 天退款保证,只要点击下方简介里的链接,或者访问livemomentous.com/modernwisdom 并在结账时使用优惠码 modernwisdom。
便签引用
07放缓信:安全还是生意
54:44
>> They did it. He did it yesterday. I mean, so this is why the if you I don't know if they you can pull you can look it up, but you can pull up a letter. They published a letter and Sam came out yesterday and said, "Maybe we should pace the frontier." >> Yeah, it's great. >> Yeah. >> I mean, it's a big big step >> that, by the way, I think we can all agree is a good thing. >> Yeah, I absolutely agree. Also think it would be an unrecognizable outcome a year ago. Like, I would love to see what the polyarket odds were on Sam coming out in favor of slowing down a year ago.
>> 他们做到了。他昨天就这么做了。我是说,这就是为什么——我不知道你们能不能调出来,你可以查一下,但你可以把那封信调出来。他们发布了一封公开信,而山姆昨天出来说:“也许我们该给前沿降降速。”>> 是啊,这很棒。>> 是的。>> 我是说,这是很大很大的一步 >> 顺便说一句,我想我们都同意这是件好事。>> 是的,我完全同意。而且我觉得一年前这会是个让人认不出来的结果。我很想看看一年前 Polymarket 上山姆公开支持放慢速度的赔率是多少。
便签引用
55:13
>> He doesn't look like he's a signature. He uh he may not be yet, but he definitely yesterday. I mean, YaKob effectively has to represent OpenAI in this case, but also he he spoke about it yesterday. >> I wouldn't be surprised if we see all of the major CEOs sign on at some point because um and >> explain explain pacing the frontier. >> Yeah. So, I mean the the headline sentence is we request that the US government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development. That last phrase automated AI development also sometimes gets called recursive self-improvement. The idea here is just the stated goal of all these companies including Sam and Jacob on a live stream last year is to take humans out of the loop of what they're making so that GBT6 is building GBT7 which then turns around and immediately starts building GBT8. And that is a step change like for all of human history.
>> 他看起来没有签名。他呃他可能还没签,但他昨天肯定……我是说,雅各布在这件事上实际上得代表 OpenAI 发声,但他昨天也谈到了这件事。>> 如果我们看到所有主要公司的 CEO 在某个时点都签上名,我不会感到意外,因为嗯 >> 解释一下、解释一下什么叫给前沿降速。>> 好。那么,核心的一句话是:我们请求美国政府支持一项国际努力,开发所需的技术与治理工具,以有意识地为自动化 AI 研发的前沿降速。最后那个说法“自动化AI 研发”有时也被称为递归自我改进。这里的想法就是,所有这些公司——包括山姆和雅各布去年在一次直播中——所宣称的目标,就是把人类从他们所做的事情的闭环中移除,这样 GPT-6 就在造 GPT-7,然后 GPT-7 转过头立刻开始造 GPT-8。而这是一个阶跃式的变化,放在整个人类历史上看都是。
便签引用
56:08
And I think these companies really do think in terms of big history. You hear Sam and he's constantly thinking about AI as like an industrial revolution scale thing, not an iPhone invention scale thing. >> Through all of history, human population has been a bottleneck on progress. Not the only bottleneck, but always a bottleneck. There was never a sector of the economy that could just go off and double on its own. And if we actually get to full automated A R&D, that changes. And I think it's just like really hard for us to imagine what that would be like. We have just no >> How much time have we got left before recursive self-improvement is no longer a discussion? Like is it even a debate at the moment?
而且我认为这些公司确实是用大历史的尺度在思考。你听山姆说话,他一直把AI 当作工业革命级别的东西在想,而不是 iPhone 发明那个级别的东西。>> 纵观整个历史,人口一直是进步的瓶颈。不是唯一的瓶颈,但始终是一个瓶颈。从来没有哪个经济部门能自己跑出去实现翻倍。而如果我们真的实现了完全自动化的 AI 研发,这就变了。我觉得我们真的很难想象那会是什么样子。我们完全没有>> 在递归自我改进不再是一个讨论话题之前,我们还剩多少时间?它现在还算是个争论吗?
便签引用
56:47
>> It works. No, it works. >> So when do you think it will be? >> Well, this is a good step. I mean, I was I was waiting for this. Okay. I was waiting >> like under under three years. >> It No, it it works. >> Well, >> it works. Just so we're clear, recursive research works. Don't let anyone tell you otherwise. It works and hence this letter. So let me explain this why I think this letter was published when it was and then we can debate. So the first reason is the attack. >> Mhm. >> Okay. The hugging face attack. There's damage control that needs to be done to win back public mandate to tell people hey we care about you which is necessary right now when people are like well what are these models going to do next?
>> 它已经奏效了。不,它是奏效的。>> 那你觉得会是什么时候?>> 嗯,这是很好的一步。我是说,我一直在等这个。好吧,我一直在等 >> 比如三年以内?>> 不,它已经奏效了。>> 嗯,>> 它是奏效的。把话说清楚,递归式研究是有效的。别让任何人告诉你不是这样。它有效,因此才有这封信。所以让我解释一下,为什么我认为这封信会在这个时间点发布,然后我们可以辩论。第一个原因是那次攻击。>> 嗯哼。>> 好。那次 Hugging Face 攻击。需要做危机公关,来赢回公众的授权,告诉人们嘿,我们在乎你们——这在当下是必要的,因为大家都在想:这些模型接下来会干什么?
便签引用
57:26
>> For what it's worth though this letter was in development before that attack happened. uh and I think my point is the signitories and the support that is now that is sweeping through these companies but let let me I'll go through the list and then and then we can talk about it. The other reason is recursive research is here and everyone's realizing it's kind of the dog that maybe caught the car and it's like okay we did it now should we for the now everyone's stay putting on the live hat and going wait a second Eric hat and going okay maybe we should pause for a second maybe maybe we shouldn't keep sprinting aimlessly at this at this incredible outcome but there's another reason which is that the amount of compute required for inference is so great and we do not have anywhere near enough and we're going to have to point most of the compute from the frontier at inference and this is because of the diminishing model returns theory which is a paper I wrote uh that says at some point I'll
>> 不过值得一提的是,这封信在那次攻击发生之前就已经在起草了。呃,我觉得我的意思是,现在正在这些公司之间蔓延的签署方和支持力量——但让、让我先把清单过一遍,然后我们再来聊。另一个原因是递归式研究已经到来了,所有人都意识到,这有点像那条终于追上了车的狗,心想,好吧我们做到了,现在我们是不是该……现在所有人都戴上了那顶帽子,开始想,等一下——戴上埃里克那顶帽子——想着,好吧,也许我们该暂停一下,也许、也许我们不该毫无方向地一路狂奔冲向这个惊人的结果。但还有另一个原因,就是推理所需的算力太庞大了,而我们远远不够用,我们将不得不把前沿的大部分算力转向推理。这是因为模型收益递减理论——那是我写的一篇文章,呃,它说到了某个时点……我
便签引用
58:26
send it to you it's really good um it cited all the time on modern wisdom and uh and uh but it argues that you and I won't care about GBT7 which is a really interesting problem and and it's actually one of the reasons that we get to debate why AI is being built and why People can talk about not wanting to make more progress because we're like what what the [ __ ] the matter with 5 5.6 is good enough. And the dimensional returns theory is sort of over here which is like hey look the next model might matter to solving this crazy disease but I don't need more intelligence. I need more time with my friends and family. I need like I need this to pump the brakes. And the other reason is that they found new surface areas. So, so whereas we used to call them labs, the companies that that built this technology, then they became token distributors, then they became application providers, now they're device manufacturers, but the end goal and and the the where I think we should also debate is the agentic internet. Who
回头发给你,写得真挺好的,嗯,在 Modern Wisdom 上被引用过无数次,呃、呃,但它论证的是,你和我不会在乎 GPT-7,这是个非常有意思的问题,而且这其实也是我们之所以能辩论“为什么要造 AI”的原因之一,也是人们可以谈论不想再要更多进展的原因,因为我们会想,5、5.6 已经够好了,这他妈到底有什么问题?而收益递减理论大概在这边,意思是,嘿你看,下一个模型也许对攻克某种疯狂的疾病很关键,但我不需要更多智能。我需要更多时间陪朋友和家人。我需要……我需要这个踩踩刹车。还有一个原因是他们找到了新的战场。所以,我们以前把造这项技术的公司叫作实验室,后来它们变成了 token分发商,再后来变成应用提供商,现在它们是设备制造商,但最终目标——以及我认为我们也应该辩论的地方——是智能体互联网。谁控制智能体互联网?因为如果你能把……既然你们已经有了
便签引用
59:20
controls the agentic internet? Because if you can turn your now that you have these war chests, now that if you can turn your attention from building the frontier to actually building applications that that control the internet, that's where you can actually lower your capex and massively improve your revenue. And so a lot of this arrives at a moment where everyone goes, "Okay, we need a we need to slow the brakes, but also where a lot of a lot of attention is going is going to start to be paid to trying to become bigger than Google by actually building the agent that barters between two people or building the agent that plans a vacation and books everything." Like owning that is worth more than owning the frontier.
这些战争储备金,既然你能把注意力从建设前沿转向真正去做那些控制互联网的应用,那才是你能真正降低资本支出、大幅提升营收的地方。所以这一切发生在这样一个时刻:所有人都说,好吧,我们需要踩踩刹车,但与此同时,大量的注意力将开始投向另一件事——试图通过真正打造那个在两个人之间讨价还价的智能体,或者打造那个规划假期并把一切都订好的智能体,来变得比谷歌更大。拥有那个,比拥有前沿更值钱。
便签引用
59:58
And the the the new service area battle is for the agentic internet. So, just to make sure I understand your your claim, you're saying companies have an incentive to want to slow the frontier because there's more money to be gained in like stopping R&D now and just trying to grab as >> we all have an incentive, but companies also have an incentive because that you the capex expense on the frontier is just the buildout on the frontier is going to be too great relative to the demand for inference. The token demand is so great and all of the compute is going to have to be spent on inference.
而新的战场之争,就是智能体互联网。所以,为了确保我理解了你的说法,你是说公司有动机希望给前沿降速,因为现在停下研发、转而去抢占……能赚到更多钱 >> 我们所有人都有动机,但公司也有动机,因为前沿的资本支出——前沿的基建投入相对于推理需求来说会太大了。token 需求如此之大,所有算力都将不得不花在推理上。
便签引用
1:00:33
I'm not I'm not quite seeing the connection there for two reasons. >> Okay. >> One is if it's my best business strategy after I do the research to say I'm going to stop training frontier models, I don't know why that means I should write a letter announcing that to the world and making sure that all my competitors do the same. Uh and and two, I think it's clear that pushing the frontier also means just driving down the cost curve everywhere beyond the frontier or or sorry before the frontier. like you train fable five anthropics best model and then you can distill from that set five and now you've got your like cheaper model that's everyone wants to talk to because they don't need the advanced intelligence so I think there are zero incentives that in like a straightforward self-interested way point towards companies saying let's slow down research research is how you get gains at every level of the intelligence >> especially when as one of the three or four frontier companies by slowing down
我有两点没太看出这中间的联系。>> 好。>> 一是,如果在我做完研究之后,我最好的商业策略是宣布我要停止训练前沿模型,我不明白为什么这意味着我该写一封信把这件事昭告天下,还要确保我所有的竞争对手都照做。呃,还有第二点,我认为很清楚的是,推动前沿同时也意味着在前沿之外的所有地方压低成本曲线——或者说,抱歉,是在前沿之前的地方。比如你训练出 Fable 5、Anthropic 最好的模型,然后你可以从中蒸馏出 Sonnet 5,现在你就有了那个更便宜的、大家都想用来聊天的模型,因为他们并不需要那么高级的智能。所以我认为,从直白的自利角度看,没有任何激励指向公司说“我们放慢研究”。研究才是你在每一个智能层级上取得收益的方式。>> 尤其是当你作为三四家前沿公司之一,放慢速度就等于给所有这些其他
便签引用
1:01:25
you're giving all of these other competitors a chance to effectively catch up Like surely they're negatively incentivized to be wanting to do >> but catch up to what is my point to them. >> But but I I understand but catch up to what like the frontier if you stopped building the frontier tomorrow you'd have enough technology to change to radically like we don't need the the economic value of the frontier is not as obvious relative to the cost of the frontier and the economic value of the actual token and the consumption is much more obvious. Now that was not that was not always true. Do you disagree with that? Yeah, I I think the whole advantage that the US has right now is that our models are four to seven months ahead of the open source ones that are copying off their homework.
竞争者一个有效追赶的机会。它们在这方面的激励肯定是负面的吧 >> 但我的意思是,追赶什么呢?对它们来说。>> 但是、但是我明白,可追赶什么呢?比如前沿——如果你明天就停止建设前沿,你已经拥有足够的技术去改变、去彻底地……我们并不需要那个。前沿的经济价值相对于前沿的成本并不那么显而易见,而实际 token 和消费端的经济价值要明显得多。当然,这在以前并不成立。你不同意吗?是的,我、我认为美国现在的全部优势就在于我们的模型比那些抄作业的开源模型领先四到七个月。
便签引用
1:02:07
>> But but that's not the that's different. The US our national security interest and the private sector interests are not I mean they're >> just economic advantage. Like the reason that people want to pay extra for Fable instead of using like Quen is that it's just smarter. If you ask it to code an app for you, it's going to make less mistakes. I I I do think it's important to be You would be very surprised how many people are actually paying a whole lot extra for the frontier like the we live in a space where we think >> I know that revenue is like $50 billion a year now and it was you know >> the revenues are amazing the tok so the the surface area that they're making money on remember there's the application and a lot of the application is people not using the frontier it's people using the next generation model the last one and token redistribution so you get data centers and you redistribute those tokens and a lot of those tokens are also not frontier tokens. Most of the consumptions from these research
>> 但那不是——那是另一回事。美国的国家安全利益和私营部门的利益并不……我是说它们 >> 只是经济优势。比如人们愿意为 Fable 多付钱而不用比如 Qwen 的原因,就是它更聪明。如果你让它写一段给你用的应用,它犯的错会更少。我我我确实觉得很重要的一点是——你会非常惊讶地发现,其实有相当多的人在为最前沿的模型额外掏一大笔钱,我们生活在一个这样的环境里,我们以为……>> 我知道现在的营收大概是每年 500 亿美元而且你知道 >> 营收非常惊人,token 也是,所以他们赚钱的覆盖面——别忘了还有应用层,而应用层里很大一部分人根本没在用最前沿的模型,他们用的是上一代模型,还有 token 的再分配——你建数据中心,然后把这些 token 再分配出去,而其中很多 token同样不是前沿模型的 token。这些研究公司的大部分消耗量都不是前沿 token,而是其他所有的东西。而他们
便签引用
1:03:00
companies are not frontier tokens. They are everything else. And they're figuring out that actually companies don't need Fable in order to radically change and improve and automate a bunch of stuff. And that actually means that the money is in the token redistribution and all the other surface areas, not just the frontier, which by the way is kind of a win for everyone. Like if we can slow down, build stuff that really matters and also like imp improve businesses with it. Yeah. Then this letter is like seemingly arriving at just the right time.
正在发现,其实企业并不需要 Fable 才能彻底改变、改进和自动化一大堆事情。这实际上意味着,钱是在 token 再分配和其他所有那些覆盖面上,而不只是在前沿模型上——顺便说一句,这对所有人来说其实算是件好事。就是说,如果我们能慢下来,去做真正重要的东西,同时也用它来改善企业。是啊。那么这封信看起来来得正是时候。
便签引用
1:03:32
>> I do I do think in general it is really important to to be skeptical, maybe even a little bit cynical when looking at what companies do. I just I think this is like a Kelsey Piper take on Twitter, but I I want to make sure that we don't like loop all the way around from cynicism into being like extremely credulous about things that are pretty unlikely. Like I think if we had a world where this letter said we believe it is imperative for US national security that we push the frontier and go as fast as we can to beat China, which is exactly what people were saying a year ago.
>> 我我确实认为,总的来说,在看待企业的所作所为时,保持怀疑、甚至带一点愤世嫉俗,是非常重要的。我我觉得这有点像 Kelsey Piper 在推特上的观点,但我我想确保我们别一路绕回来,从愤世嫉俗变成对一些相当不靠谱的事情极度轻信。比如我觉得,假设我们身处这样一个世界,这封信写的是:我们认为,为了美国国家安全,我们必须推进前沿、全速前进以击败中国——这恰恰就是一年前人们在说的话。
便签引用
1:04:02
>> Everyone would be saying, well, yeah, obviously that's in the company's self-interest. They want some reason that we can't regulate them and they have to go fast. And so like we got to make sure that our ideas are passing the reversability test. If if you would say the same thing about someone making the exact opposite claim, I think you want to step back and at least consider the possibility that like these companies really are worried that these scientists really are worried to just on the record. I started out by saying exactly that like I literally started out by saying like I actually now want to be associated with I don't want to be associated with the alternative the claim you just made. The companies have realized that recursive research is here and that it is pretty imperative at this point that we stop >> because it's risky, right? Not just because it's financial >> the alignment issue just on the record and >> it arrives at this moment where we are going to have to spend an enormous
>> 那所有人都会说,嗯,是啊,这显然符合公司的自身利益。他们想找个理由说我们不能监管他们、他们必须跑得快。所以我们得确保自己的观点能通过可逆性测试。如果换成有人提出完全相反的主张,你也会说同样的话,那我觉得你就该退一步,至少考虑一下这种可能性:这些公司是真的在担心,这些科学家是真的在担心。就当留个记录,我一开始就明确说了这一点,我字面意义上一开始就说了,我其实现在想被归到……我不想被归到另一边、也就是你刚才提出的那个说法那边。这些公司已经意识到,递归式研究已经到来,而此刻我们停下来是相当必要的。>> 因为这有风险,对吧?不只是因为财务上的原因。>> 对齐问题,就当留个记录,而且 >> 它恰好出现在这个时刻,我们即将不得不投入巨额
便签引用
08监管俘获与国际协调
1:04:53
amount. >> Which one's the bigger problem? >> Are they more about safety or are they more worried about cost? I bet a lot of the signatures here are really worried about safety and I think the uh we are going to find economic alignment around the performance of the models and the the need for inference and the need economics often win out you I mean this is this is the incentive you you talk about the incentive and I think you're right and the incentives are finally aligning for safety and alignment and also econom eomic development like we we don't need better model and by the way yeah no I mean we're [laughter] >> okay >> but but actually no fantastic >> but actually if we didn't build the frontier for for another 3 years we we could radically change and improve the economy >> the thing that I'd always find frustrating is maybe it's just my little corner of Twitter of X that I live in but seeing these mental gymnastics type arguments of of the people who are just like screaming out that it's regulatory capture
资金。>> 哪一个才是更大的问题?>> 他们更在意安全,还是更担心成本?我打赌这上面很多签名的人是真心担心安全问题。而且我觉得,呃,我们会在模型性能、推理需求这些方面找到经济上的对齐,而经济因素往往会占上风。我是说,这就是激励,你刚才谈到激励,我觉得你说得对,而激励终于开始向安全与对齐以及经济发展同时对齐了。就是说,我们并不需要更好的模型,顺便说一句,是的,不,我是说我们……[笑]>> 好吧 >> 但但其实不,太棒了 >> 但其实,如果我们再有三年不去建前沿模型,我们也能彻底改变和改善经济。>> 让我一直觉得很受挫的一点——也许只是我所处的推特(X)那个小角落——是看到那些心理体操式的论调,有些人只要有人提出任何形式的监管,尤其是前沿公司提出的,他们就会尖叫说这是监管俘获。因为他们基本上
便签引用
1:05:55
whenever any kind of regulation is proposed especially by the frontier companies because it they're basically just saying you know when these frontier companies are like hey we we we want to slow us down just us three or four the regulatory capture people are like well that must be regulatory capture you're hurting the little guys and they're like no no no just just us it's okay if the other guys break you know they they catch up no it still sounds like regulatory capture it's almost like a religion or something that just this this deep cynicism um and I want to understand where that comes from because Do we see it in some arms of the government or people who were working for the government? I don't want to name names but like why where like >> I don't even know who you're alluding to.
就是在说,你知道,当这些前沿公司说,嘿,我们想让我们自己慢下来、就我们这三四家,那些讲监管俘获的人就会说,那肯定是监管俘获,你们在伤害小玩家;然后前沿公司说不不不,就只管我们,其他人突破也没关系,你知道,让他们追上来——不行,这听起来还是像监管俘获。这几乎就像一种宗教之类的东西,就是这种深深的愤世嫉俗。嗯,我想弄明白这种心态是从哪来的,因为——我们在政府的某些部门或者曾在政府工作的人身上看到过吗?我不想点名,但就是说,为什么、在哪 >> 我都不知道你在暗指谁。
便签引用
1:06:37
>> I don't if I want to say like no I don't I mean okay certain VCs for example Andre I don't think he'll care that I he's very openly he like pulls every any attempt for regulation he has fought against tooth and nail tooth and nail so like I want to understand where that mindset comes from. Okay, I disagree with Mark Andre's views on AI in almost every way, but just to steal it for a sec, um I think it's not crazy to look at society and say we're not doing great on nuance lately, especially in American public discourse.
>> 我不确定我想不想说……不,我不……我是说,好吧,比如说某些风投,Andreessen——我觉得他不会介意我他非常公开地,他基本上会……任何监管的尝试他都会拼死反对,拼死反对,所以我想弄明白这种思维方式是从哪来的。好吧,我几乎在每个方面都不同意 Marc Andreessen 对 AI 的看法,但先帮他把话说圆一下,嗯,我觉得这也不是完全没道理:看看社会,你会说我们最近在处理细微差别这件事上做得并不好,在美国的公共讨论中尤其如此。
便签引用
1:07:08
Um, I think there was I'm just I'm just quoting other bloggers at this point, but uh someone wrote a post a couple years ago called like the dial of progress where it just sort of feels like online debate is premised on this like one knob and you have to just turn it towards like yay progress or boo progress. Those are the only things that you can do. And if those are your options, there's a really good argument to go, yay, progress. Because there are all kinds of examples of us overregulating things and people, real people dying because of drugs getting held up in regulatory fights. Like this is this is not hypothetical. I think we often overdo it here. Um I'm not sure what I would say if I was given that dial. Like there there are a bunch of diseases that we need to cure and it would be terrible to say we have to like shut down AI before figuring out how speculative that risk is and and not get those >> but but the point is the dial doesn't exist.
嗯,我觉得有——我现在只是在引用别的博主,不过呃,几年前有人写过一篇文章,标题大概叫《进步旋钮》,大意就是感觉网上的辩论都建立在这样一个旋钮上,你只能把它拧向“进步万岁”或者“进步去死”。你能做的就只有这两件事。如果这就是你的选项,那么选“进步万岁”其实有相当好的理由。因为有各种各样的例子说明我们过度监管,导致真实的人因为药物卡在监管扯皮里而死去。这这不是假设。我觉得我们在这方面常常做过头了。嗯,我不确定如果真把那个旋钮交给我,我会怎么说。比如说,有一堆疾病我们需要攻克,如果说我们必须在搞清楚那个风险有多么投机之前就关停 AI,那会很糟糕,我们就拿不到那些成果了 >> 但但重点是,这个旋钮根本不存在。
便签引用
1:07:58
>> Yeah. I think we >> we don't live on this one-dimensional line of go Yeah. go faster or go slow. like there are all these like it's it's a very multi-dimensional space we live in and it's these people uh who try and compress it down into we just must go as fast as possible. It's like well in which area there are some things we want to max out and it's very spiky. I'm glad you pointed towards diffusion as a blocker here that you know the way that certain industries might have labor organization that that blocks automation because I I think that's a separate dimension like you can think of that as like horizontal acceleration of AI take the models we have get them spread out and this other dimension of just like vertical acceleration how fast is the frontier getting better >> technological threshold versus societal threshold >> yeah I'm all about breaking down the barriers to that societal spread >> and by the way me and you both >> and actually I think we can celebrate having arrived at a place where if we
>> 是啊。我觉得我们 >> 我们并不是活在这条一维的线上——是啊——要么更快要么更慢。就好像有这么多……我们生活在一个非常多维的空间里,而这些人试图把它压缩成“我们就是必须尽可能快”。那就得问:在哪个领域?有些东西我们确实想拉满,而且这非常尖锐不均。我很高兴你指出扩散是这里的一个瓶颈,你知道,比如某些行业可能有劳工组织,会阻碍自动化。因为我觉得那是另一个维度,你可以把它看作 AI 的横向加速——拿现有的模型,把它们铺开去;而另一个维度就是纵向加速——前沿进步得有多快 >> 技术门槛 vs 社会门槛 >> 是的,我全心投入在打破社会层面扩散的障碍上 >> 顺便说一句,我和你都是>> 而且其实我觉得,我们可以庆祝已经走到了这样一个位置:哪怕我们再也不造新模型——当然
便签引用
1:08:54
never built another model, which obviously we probably will, but if we didn't, we would do a whole lot of good. >> I agree with that. I I just also want to make sure we don't think of this letter as like time to celebrate victory, right? >> Because there are still incredible incentives to push in the frontier. Like you said, we don't care about GPT7, but all of the AI engineers that are trying to do some of the hardest math and science research ever to make digital brains that are faster than their competitors, they really care about the difference between GPT7 and 7.1. And I I just think like by default there's going to be a ton of racing inside these labs.
我们很可能还是会造——但假如不造了,我们也能做成一大堆好事。>> 这点我同意。我我只是也想确保我们别把这封信当成该庆祝胜利的时刻,对吧?>> 因为推进前沿的动力依然强得惊人。就像你说的,我们不在乎 GPT-7,但所有那些正在做有史以来最难的数学和科学研究的 AI 工程师,他们想造出比竞争对手更快的数字大脑,他们真的非常在意 GPT-7 和 7.1 之间的差别。而我只是觉得,默认情况下,这些实验室内部会有大量的竞速。
便签引用
1:09:31
>> I agree. I also think it's interesting to consider though that the economic incentives have taken us to a place where like we could have autonomous vehicles tomorrow. We don't need to build a better autonomous car. we will not have them for many many many years for a variety of reasons and we can pivot this conversation now approp maybe to talking about how we actually diffuse the value of this technology to the benefit of the average person because I maintain that the problem with the current AI debate I we have the luxury of knowing a lot of stuff that's going on with hugging face at all most people don't most people do that was news to us it was not important news to to many many many people and Most people right now are the data center is the physical manifestation of a technology that destroyed their their their kids' minds, right? Or their or you know the the the free gambling, free porn, free addiction and they're looking at AI, whatever model as an app that you can make a video of Donald Trump making out with
>> 我同意。不过我也觉得有意思的是,经济激励已经把我们带到了这样一个地方:比如我们明天就能实现自动驾驶汽车。我们并不需要造出更好的自动驾驶车。但由于种种原因,我们在很多很多很多年里都不会真正拥有它们。我们现在或许可以顺势把话题转到:我们到底该如何把这项技术的价值扩散出去,让普通人受益。因为我一直坚持认为,当前 AI 辩论的问题在于——我们有幸知道 Hugging Face 上发生的一大堆事情,大多数人并不知道,大多数人……那对我们来说是新闻,但对非常非常多的人来说,那根本不是什么重要新闻。而现在对大多数人来说,数据中心就是那种技术的实体化身——那种毁掉了他们孩子心智的技术,对吧?或者说,免费的赌博、免费的色情、免费的成瘾。然后他们看着 AI,不管什么模型,只把它当成一个可以做出唐纳德·特朗普和泰勒·斯威夫特亲热视频的应用,然后他们会说:“我们干嘛需要更多这种东西?”而我,我
便签引用
1:10:31
Taylor Swift and they go, "Why do we need more of this?" And that I I empathize with that as a father at at a level now that I like didn't know I would. Like I I am I feel so much for every parent who feels like they lost their child to social media and doesn't want to lose their grandkids to whatever AI slop we're going to create. Just to keep on the safety thing, in July, the Future of Life's Institute's AI safety index evaluated nine companies and found that no lab scored higher than a C plus.
作为一个父亲,现在能在一个我原本没想到的层面上对此感同身受。就是说,我我真的非常同情每一个觉得自己已经把孩子输给了社交媒体、并且不想再把孙辈输给我们即将造出的那些AI 垃圾内容的家长。还是接着聊安全这件事,今年七月,未来生命研究所的 AI 安全指数评估了九家公司,发现没有一家实验室的得分高于 C+。
便签引用
1:11:01
What what who was that? >> Anthropic came in first at 2.6. 66 Open AAI and Google Deep Mind got a C. Meta got a D+. XAI, Deep Seek, and Miscell all failed outright, one on each continent. Is the best score being a C++ from a safety lab? The safety lab reassuring given how powerful everything is at the moment and especially I I think it's important to talk about what's happening internationally and the incentives there as well before we talk about democratizing it and what this looks like for diffusion. Yeah, I mean we'll take the other side obviously it's like how do who who who is deciding these s safety tests who >> is this a external grading of the >> but how can yeah how independent can any third party be um there's there's very there's another one called um >> it's like the tracker safer AI I believe and it's I think it's a little bit more nuanced than than that particular one you just quoted and again the best rated one I think very recently was again anthropic and that was only 35% out of
什么?那是谁?>> Anthropic 以 2.66 排名第一。OpenAI 和 Google DeepMind 拿了 C。Meta 拿了 D+。xAI、DeepSeek 和 Mistral全都不及格,每个大洲各一家。一家安全实验室给出的最高分只是 C++?考虑到现在这一切有多强大,这个由安全机构做出的评估能让人安心吗?尤其是,我我觉得在我们谈论民主化以及扩散会是什么样子之前,有必要先谈谈国际上正在发生什么、以及那里的激励机制。是啊,我是说我们当然也得说说另一面,就是,这些安全测试是谁来定的、谁 >> 这是外部评级吗>> 但怎么可能……是啊,任何第三方能有多独立?嗯,还有还有……还有另一个叫做,嗯>> 好像叫 Safer AI 那个追踪榜,我记得是,而且我觉得它比你刚引用的那个要细致一些。而且同样地,我记得最近评分最高的还是 Anthropic,而那也只有 35 分(满分
便签引用
1:11:57
100 so Um, yeah. I mean, I you you again the other side's argument, well, the the anti-ID's argument would be, well, by definition, if you're a if you're building a safety evaluation app, you're biased against, you know, you're all you're looking for are risks and you're not taking into account the benefits or something like that. And and again, to take the other like the to argue the accelerationist point, their concern is like this idea of the um invisible graveyard. We don't see all of the harms that are avoided.
100)。所以,嗯,是啊。我是说,我……再说一次,另一边的论点,反……反对方的论点会是:从定义上讲,如果你是在做一个安全评估的应用,你天然就带有偏见,你知道,你眼里只看得到风险,而不会把收益之类的东西算进来。而且,再说一次,如果站在加速主义那一边替他们说话,他们的顾虑是这样一个概念:嗯,“看不见的墓地”。我们看不到所有那些被避免掉的伤害。
便签引用
1:12:26
uh sorry or uh by pausing a technology you don't get to see all of the lives that would have been saved by it >> and I think it's a very valid concern. Um it's just that you have to balance that equation with okay what are all the harms that might be caused if it goes ahead at break neck speed and does some massive cyber attack on the the grid of all the western nations at the same time or any any nation um and causes however many millions of deaths on expectation. You have to look at both the potential costs the risks and the potential risks of not doing it fast enough. And it's it's worth saying a lot of these AI regulations that have been trying to get through state level legislators are trying to do that nuanced carving out.
呃抱歉,或者说,通过暂停一项技术,你就看不到本来能被它救下的所有生命 >> 我觉得这是个非常合理的顾虑。嗯,只是你必须把这个等式的另一边也算进去:好,如果以疯狂的速度往前冲,可能造成哪些伤害?比如对所有西方国家的电网同时发动某种大规模网络攻击,或者对任何国家,嗯,按期望值造成几百万人的死亡。你必须同时看潜在成本、风险,以及不够快带来的潜在风险。而且值得一提的是,很多这类一直试图在州级立法机构闯关的 AI 监管法案,正是在试图做出这种细致的区分和豁免。
便签引用
1:13:12
Like there was a law in New York that was passed that tried to set the bar so high that it actually wasn't mandatory for OpenAI to report the hugging face incident because it didn't cause over a billion dollars in damages or kill 50 people. So I mean it's great that they're trying to like set thresholds. >> Glad you arbitrary rules. >> Yeah. But like I think we can all agree if if you like hack a billion dollar tech company like >> probably you set the threshold maybe a a bit too high for at least just requiring that you are telling the world what happened let alone doing anything about it.
比如纽约通过了一项法律,它试图把门槛设得非常高,高到 OpenAI 其实都不必强制上报 Hugging Face 那起事件,因为它没有造成超过十亿美元的损失,也没有害死 50个人。所以我是说,他们愿意去设定阈值是好事。>> 很高兴你说了,任意定的规则。>> 是啊。但我觉得我们都能同意,如果你黑了一家市值十亿美元的科技公司 >> 那门槛大概设得有点太高了,至少就“要求你把发生了什么告诉世界”这一条来说是这样,更别提去做点什么来应对了。
便签引用
1:13:43
>> Um I I think like we actually can live in a world where we have nuance where we say different standards apply to startups versus these giant trillion dollar companies. And I I do think it's sad that some folks like Mark Andre have kind of given up on nuance and have just said it's all or nothing and so we'll just any means necessary to stop regulation. You know, end of sentence. >> What about the international piece here from an incent? Have we got anything approximating global coordination when it comes to frontier AI? And without that, can you do anything?
>> 嗯,我我觉得我们其实可以生活在一个有细微差别的世界里,我们可以说,对初创公司和对那些万亿美元级的巨头,适用不同的标准。而我确实觉得很遗憾,有些人比如 Marc Andreessen基本上已经放弃了细微差别,只是说这是全有或全无,所以我们就要不择手段地阻止监管。你知道,句号。>> 那从激励的角度看,国际这一块呢?在前沿 AI 上,我们有没有任何接近全球协调的东西?如果没有,你还能做什么?
便签引用
1:14:12
>> So I mean recently China passed this anti-anthropomorphic AI law. I don't know the specifics. It seems like they're mostly worried about like LLMs sounding like humans. I think they're trying to avoid a lot of this sort of like identity confusion stuff, but they just they don't seem that gung-ho about LLMs. And you know, who knows if this got lost in translation, but she said something recently about also wanting to stop loss of control risks. I think we're a lot more economically entwined with China today than we were with the USSR or the USSR in uh in the middle of the 20th century and we still made a treaty to stand down on nukes. Like I think it's just not crazy to think that a deal could be struck.
>> 嗯,我是说,最近中国通过了一项反拟人化 AI 的法规。我不知道具体细节。看起来他们主要担心的是大语言模型听起来像人。我觉得他们是想避免大量这种身份混淆之类的问题,但他们就是他们对大语言模型似乎没那么热衷。你知道,也许这中间有翻译上的偏差,但她最近说过一些话,提到也想阻止失控风险。我觉得我们今天和中国在经济上的交织程度,比20世纪中期我们和苏联之间要深得多,可当时我们还是签了核裁军的条约。所以我觉得,认为能达成某种协议,这并不是什么疯狂的想法。核裁军条约。我觉得认为双方能达成某种协议并不算离谱。
便签引用
1:14:53
>> I think it will be. >> I think >> something like this that would be enforcable and complied with could happen globally. >> I give it I give that a pretty high chance especially with recursive research. It doesn't even necessarily have to be I mean again define enforceable because like what kind of institution can you build like I guess like an UN type thing and that's like a whole can of worms and it's also very ineffective on a lot of things. Um but like there have been many precedents in history where there was two major political actors with massive tensions and massive incentives to sort of defect on one another. Uh and yet they came up with ways to like crossverify like with uh the nuclear arms reduction treaty. Uh they I think one of the things they did the USSR and the US was like basically they sent teams of scientists over to the others uh bunkers to check on like their uh safety protocols, little things. So like control like you didn't get to see how all of our nukes work, but you get to see what our um methods
>> 我觉得会的。>> 我觉得 >> 类似这样一个可执行、也能被遵守的机制,是有可能在全球范围内实现的。>> 我觉得这个概率相当高,尤其是在递归式研究的背景下。它甚至不一定非得——我是说,还是得定义什么叫「可执行」,因为你能建立什么样的机构呢?我猜就是类似联合国那种,而那本身就是一堆麻烦事,而且在很多问题上都非常无效。嗯,但历史上有很多先例,两个主要政治行为体之间有巨大的紧张关系,也有巨大的动机去背叛对方,可他们最后还是想出了交叉核查的办法,比如核裁军条约。我记得苏联和美国做的其中一件事就是,他们互相派科学家团队到对方的地堡去,检查他们的安全协议之类的小事情。所以是那种控制性的——你没法看到我们所有核武器是怎么运作的,但你可以看到我们用什么方法来防止意外发射。
便签引用
1:15:53
are to prevent accidental misfire. and the amount of diplomacy that came from simply having those teams of scientists mix and get to look each other in yeah in the eye and collaborate on a shared goal which is let's not have accidental nuclear war at least yes okay fine you can fire on us if we fire on you or you know but let's not just have a misfire right and those those kinds of little cross-pollinations work so well and um so I think that could happen in in for example I mean I I can't I'm not going to be able to recreate it maybe you might be able to but the the AI 2040 is such a good uh it's quite a long piece but it's basically it's written by Daniel Cocatello and and these guys who did the AI 2027 they wrote in 2021 that predicted how AI would play out until 2027 it's uncannily accurate well you know we're now in 2026 and we can look back the number of predictions they got absolutely spot on was I've never seen anything like it so these guys really think well about the future and they
而仅仅是让这些科学家团队混在一起、能面对面地交流、为一个共同目标协作——也就是「咱们别搞出一场意外的核战争」——由此产生的外交效果就很惊人。对,好吧,你可以在我们先开火的情况下还击,或者诸如此类,但咱们别搞出误射,对吧。这种小规模的交叉往来效果特别好。嗯,所以我觉得这种事可能发生,比如说——我是说,我没法复述出来,也许你可以——但那篇《AI 2040》写得真的很好,篇幅挺长的,基本上是由Daniel Kokotajlo 和那几个做《AI 2027》的人写的。他们在2021年写过一篇,预测了AI会如何发展到2027年,准得离谱。你看,我们现在已经是2026年了,可以回头看,他们完全说中的预测数量之多,我从没见过这种情况。所以这些人对未来真的想得很透彻,
便签引用
1:16:52
just wrote this piece called AI 2040 which lays out a potential plan for essentially international coordination uh on how to safely transition to a super intelligent world. Uh it's got tons of holes in it. There's obviously many ways it can go wrong and but it's it's incredibly it's as robust as it can be while being incredibly epistemically humble. Um so yeah, I recommend everyone goes goes and reads it. But yeah, they talk about these kind of these like crosspollinations that can happen, but also like it doesn't have to be bilateral. Like if the US does you say, you know what, we're going to do a pause, that doesn't necessarily like that doesn't mean I don't know, maybe you disagree with this, but like will China necessarily take over and and and if the US pauses for a few months?
他们刚写了这篇叫《AI 2040》的东西,里面给出了一个潜在的方案,讲的基本上就是国际协调,如何安全地过渡到一个超级智能的世界。这里面漏洞一大堆,显然有很多地方可能出错,但它已经是在保持极度认知谦逊的前提下,能做到的最稳健的版本了。嗯,所以我推荐大家都去读一读。不过是的,他们谈到了这类交叉往来,但同时也不一定非得是双边的。比如如果美国说,「你知道吗,我们要暂停」,那也不一定就意味着——我不知道,也许你不同意这一点,但如果美国暂停几个月,中国就一定会赶超吗?
便签引用
1:17:45
So, I go to China three times a year, mostly because Americans don't go anymore and I just want to see what's going on, but also because I really like Chinese people and Chinese food. Um, China is playing a different game. I mean, this is kind of well reported now, but China doesn't see AI as the endgame. And it's not actually treating AI as an endgame. is treating the infrastructure as the endgame and it is now has three times energy buildout capacity. It has 45,000 mi of highspeed rail. It has better, faster, cheaper hospitals year-over-year.
我一年去中国三次,主要是因为美国人现在都不去了,我就想看看那儿到底怎么回事,另外也因为我真的很喜欢中国人和中国菜。嗯,中国玩的是另一盘棋。我是说,这一点现在也被报道得挺多了,但中国并不把AI看作终局。它实际上也没把AI当成终局来对待。它把基础设施当成终局,而现在它的能源建设能力是(我们的)三倍。它有4.5万英里的高铁。它的医院逐年变得更好、更快、更便宜。
便签引用
1:18:25
And while I don't want to live in China, the the life of the average Chinese person has sort of improved quite remarkably over the last 30 years and continues to do so seemingly year-over-year. It also curbs free speech by telling Chinese kids that they cannot use social media in very specific ways at certain hours. >> They do a lot more than that. >> Well, but [laughter] but I but yeah. Yes, they do. The I bring it up because um I don't I I think we overestimate China's ability to build a frontier. Um they don't act they don't have the necessary compute. A lot of what they're doing is reverse engineering our models.
虽然我不想住在中国,但过去30年里普通中国人的生活水平确实提升得相当显著,而且看起来还在逐年提升。它同时也限制言论自由,比如规定中国孩子在特定时间段不能以某些特定方式使用社交媒体。>> 他们做的可远不止这些。>> 呃,但是[笑]但我,但是对。是的,他们确实是。我提这个是因为,嗯,我觉得我们高估了中国建设前沿模型的能力。嗯,他们没有——他们没有必需的算力。他们做的很多事情是在逆向工程我们的模型。
便签引用
1:19:02
We are paying a lot to help China keep up. And I think one of the other reasons that this paper arrives at a nice time and the slowdown actually feels appropriate is that I think we're going to discover really quickly that the Wizard of Oz has actually just been following our breadcrumbs and that if we stop building the frontier, I don't think China is going to sprint ahead on a frontier and they'll probably have to capitulate to this paper anyway. >> The conga line where we're at the front >> and they'll probably have to say, "Oh, we want to slow down too so that we don't actually know that they were holding on to our hips." This is my guess. But also they care less about the frontier because they're already figuring out that they need to diffuse what they already have to the benefit of the average person. And this is the important point which is that I still think in the US we lack well the vision and the leadership to talk about all the ways in which we could diffuse the technology to the benefit of the average
我们正在付出很大代价来帮中国跟上。我觉得这篇论文来得正是时候、这种放缓确实显得合适的另一个原因是,我认为我们很快就会发现,「奥兹国的魔法师」其实一直只是跟着我们撒下的面包屑在走。如果我们停止推进前沿,我不认为中国会在前沿上一路狂奔,他们大概率反正也得顺应这篇论文的主张。>> 就像我们排在最前面的康加舞队 >> 而他们大概也得说,「哦,我们也想放慢速度」,这样别人就不会发现他们其实一直搭着我们的腰。这是我的猜测。不过他们也确实没那么在意前沿,因为他们已经想明白了,需要把现有的技术扩散出去,让普通人受益。这才是重点——我仍然认为在美国,我们缺乏那种愿景和领导力,去讨论我们能以哪些方式把技术扩散出去,让普通人
便签引用
1:19:51
person. Which is why the safety debates become so heated because people go well there is no the upside is what some weird panacea no one can imagine. The downside is that these things kill us all and it um it doesn't actually ever end up becoming this discussion around what would it look like if 45,000 people didn't die on the roads every year, which is, you know, the the the the invisible graveyard. What if what if hospitals were not cesspools of bureaucracy and and terrible healthcare? Like what what if we built technology and applied it quickly to fix a lot of the things that everyone agrees are broken, but we don't we don't have these discussions in public forums. And so instead we we our minds wander to the worst possible outcomes. And and China has figured out that if you continue to diffuse technology to the benefit of the average person, the average person views technology quite favorably, which is why 80% of Chinese people are excited about AI. Now they report that number. So but
受益。这也正是安全辩论会变得如此激烈的原因,因为人们会说,好处就是某种谁也想象不出来的奇异灵丹妙药,而坏处是这些东西会把我们全都干掉。然后这场讨论就永远落不到这样的问题上:如果每年不再有4.5万人死在路上,那会是什么样?也就是那个所谓的「看不见的墓地」。如果医院不再是官僚主义和糟糕医疗的粪池,那会怎样?如果我们能造出技术并迅速应用,去修复那些大家公认已经坏掉的东西,那会怎样?但我们不会在公共论坛上讨论这些。所以取而代之的是,我们的思绪飘向了最糟糕的可能结果。而中国已经想明白了:如果你持续把技术扩散出去,让普通人受益,普通人对技术的看法就会相当正面,这就是为什么80%的中国人对AI感到兴奋。当然这个数字是他们自己报的。但我实地去看的时候,从个人见闻上观察到的也差不多。我看到人们喜欢技术——为什么不呢?它
便签引用
1:20:43
when I go anecdotally, I observe as much. I observe that people love technology cuz why wouldn't they? It built a middle class. It's connected a bunch of people that didn't used to be connected. It affords people lives that they couldn't have imagined. At the same time they do live in a literal surveillance state. >> I don't actually. So now I we we you and I talked about this last time. I don't actually I'm not promoting the Chinese way of life. I'm promoting the Chinese vision for the diffusion of technology.
造就了一个中产阶级。它把一大群原本没有连接的人连了起来。它让人们过上了他们以前无法想象的生活。与此同时,他们确实生活在一个字面意义上的监控国家里。>> 我其实并不——所以现在我,我们,你和我上次聊过这个。我其实并不——我不是在推崇中国的生活方式。我推崇的是中国在技术扩散上的愿景。
便签引用
1:21:07
We know exactly what is possible from a diffusion standpoint thanks to China. They've given us a blueprint to actually distribute technology >> without the totalitarianism. >> Without the totalitarianism [snorts] >> it comes with the side order. It comes on the comes with it. We should sort the need. Eric, what do you >> Singapore Singapore offers a view of this too? A little less totalitarianism. >> Jared, you ever considered that you might have a drinking problem? >> I don't consider a lot, Chris.
多亏了中国,我们才确切地知道在扩散这件事上什么是可能的。他们给了我们一张真正把技术分发出去的蓝图 >> 而且不需要极权主义。>> 不需要极权主义[哼笑] >> 那是附赠的配菜。是打包在一起的。我们应该把需求理清楚。Eric,你觉得 >> 新加坡,新加坡也提供了一种这样的样本?极权程度没那么高。>> Jared,你有没有想过自己可能有酗酒问题?>> 我不怎么想事情,Chris。
便签引用
1:21:38
>> Well, you drank an entire case of Athletic Brewing last night. >> But they're non-alcoholic. >> And that's not a problem. >> Sorry, man. I I just kept chugging. Wait for the regret to creep in. Never happened. See, most people like Jared don't want to change what they drink. They just don't want the next day to be a complete write-off. And that is why I'm such a huge fan of Athletic Brewing Co. They make the best NA brews on the planet. You can find Athletic Brewing Co.'s bestselling lineup at grocery or liquor stores near you. Or best option, get a full variety pack of four flavors shipped direct to your door right now.
>> 可你昨晚喝了一整箱 Athletic Brewing。>> 但那些是无酒精的啊。>> 那就不算问题了。>> 抱歉,老兄。我就是一直灌。等着后悔涌上来,结果从来没来过。你看,大多数像 Jared 这样的人并不想换掉自己喝的东西。他们只是不想第二天整个废掉。这就是为什么我这么喜欢 Athletic Brewing Co.。他们做的是全世界最好的无酒精啤酒。你可以在附近的杂货店或酒类商店找到 Athletic Brewing Co. 的畅销系列。或者更好的选择是,现在就订一整套四种口味的综合装,直接送到你家门口。
便签引用
1:22:15
Get 15% off your first online order by going to the link in the description below or heading to athleticbwing.com/modernwisdism using the code modernwisdom at checkout. That's athletic brewing.com/modernwisdom at modernwisdom at checkout. Near beer terms and conditions apply. Athletic brewing company fit for all times. Bottoms up. [laughter] >> You did a great video about the AI 2027 thing. one of the best videos that's come out about AI this year. >> Thanks, man. >> What do you reckon about the 2040 thing?
点击下方简介里的链接,或访问 athleticbrewing.com/modernwisdom,在结账时使用优惠码 modernwisdom,即可享受首次线上订单8.5折。就是 athleticbrewing.com/modernwisdom,结账时输入 modernwisdom。类啤酒饮品,条款与条件适用。AthleticBrewing Company,适合任何时刻。干杯。[笑] >> 你做过一期关于《AI 2027》的视频,特别棒。是今年关于AI最好的视频之一。>> 谢了,老兄。>> 那你怎么看《2040》这篇?
便签引用
1:22:47
Do you think it it tracks? Were you as impressed as Liv? >> I thought it was fascinating and yeah, I was impressed in a lot of ways. They're doing a very different thing. So, with 2027, they were doing pure prediction. This is how we think things will play out most likely and they are like uncannily good at calling that. They called agents before those were a thing. Um, this one is is prescriptive. It's like here's what we think we should do and they they're explicit about that. Um and there's all kinds of like fascinating ideas and and game theory at one point adventure.
你觉得靠谱吗?你有像 Liv 那样被惊艳到吗?>> 我觉得它很迷人,是的,在很多方面我都挺佩服的。他们这次做的是很不一样的事。在《2027》里,他们做的是纯预测。就是「我们认为事情最可能这样发展」,而他们在这方面准得离谱。他们在智能体还没成为热点之前就预测到了。嗯,而这一篇是规范性的。它讲的是「我们认为我们应该怎么做」,而且他们对这点说得很明确。嗯,里面有各种各样迷人的想法和博弈论,某一处还有点像互动式冒险小说。
便签引用
1:23:17
>> Yeah. I mean they they at one point they're like okay well so like the deal is going to be like this. China's going to put its data centers in Canada and the US is going to put its data centers in Mongolia which are like neutral states but nearby the other person. We can blow them up if there's ever like a breakdown of the deal. Like you know they really did some galaxy brain stuff. Who knows how realistic that is. What I loved about it is they introduced this phrase like scenario scrutiny and they're like it's very easy to talk and we've been we've been doing it for the past hour and a half. It's very easy to talk in terms of like highle trends, you know, what do we make of like the history of technological progress so far? What does that teach us, etc. And they're like, it's a very different thing to sit down and just actually tell a monthby-month story about the future.
>> 对。我是说,他们在某个地方说,好,那么这笔交易大概会是这样:中国把它的数据中心放在加拿大,美国把它的数据中心放在蒙古,这些都算是中立国家,但又靠近对方。如果协议破裂,我们随时可以把它们炸掉。你懂的,他们真的搞了些脑洞大开的东西。这有多现实谁知道呢。我喜欢的是他们提出了一个说法,叫「情景推敲」。他们说,光是空谈太容易了——我们过去一个半小时就一直在这么干。空谈那些宏观趋势太容易了,比如「我们怎么看待迄今为止的技术进步史?它给了我们什么启示?」等等。而他们说,坐下来,真的去讲述一个关于未来的、逐月推进的故事,是完全不同的一件事。
便签引用
1:23:56
Like actually put yourself in the head of the different actors. What are their incentives at this moment? What do they choose to do? How does that cascade the way that, you know, the military tries to do with war games? And they're they're trying to to do that. Uh, and I I do wish that we were doing more of that and less kind of like I'm guilty of this as much as anybody else, but less sort of like armchair philosophizing in some of these discussions because I think you're going to always be able to find an argument that that suits your thesis. Mark Andre can point to the fact that there is this very >> real invisible graveyard uh from technological progress going wrong and then folks who are worried about catastic risk can point to all the near misses with with nuclear and and how we live in this time of perils. Um, yeah.
就是真正代入不同行为体的脑子里。他们此刻的动机是什么?他们会选择怎么做?这又会如何层层连锁?就像军方做兵棋推演那样。而他们就在尝试做这件事。呃,我确实希望我们能多做点这样的事,少一点那种——我在这方面跟任何人一样有罪——但在这类讨论里,少一点那种纸上谈兵式的哲学空想,因为我觉得你总是能找到一个契合你论点的论证。Marc Andreessen 可以指出,技术进步出问题确实造成了一个非常 >> 真实的看不见的墓地;而那些担心灾难性风险的人,则可以指出核武器上所有那些擦肩而过的险情,以及我们如何生活在一个危险时代。嗯,是的。
便签引用
1:24:40
What I love about the report is that they're just like, "Here is what it would actually look like for a bunch of real countries to try to respond to incentives." And they basically throw it out as a challenge and say, "We think this is a bad plan, but we haven't seen anything better." And so, someone else, please like tell us a story for how we don't have to worry. Uh, and and >> I mean, it's kind of a very rigorous example of like someone trying to write a a a white mirror. you know, we've got so many Black Mirror type dystopian sci-fi about the future of how technology could go. Um, and I think that's a big part of our cultural issue is that we don't have uh at least, you know, the the amount of utopian fiction to dystopian fiction is is outnumbered like 100 to one or something. And that's a big problem for our collective psyche. Yes. Okay. It is harder to imagine like the way we manage to thread the needle and make something go unbelievably right, you know, because you're fighting against entropy
我喜欢这份报告的地方在于,他们就是说:「这就是一群真实的国家在试图回应各自激励时,实际会呈现出的样子。」而且他们基本上是把它当作一个挑战抛出来,说「我们认为这是个糟糕的方案,但我们还没见过更好的」。所以,别人请给我们讲一个故事,说明为什么我们不必担心。呃,而且 >> 我是说,这算是一个非常严谨的例子,有人试图去写一部「白镜」。你知道,我们有太多《黑镜》那种反乌托邦科幻,讲技术的未来会如何走偏。嗯,我觉得这也是我们文化问题的一大部分:我们没有——至少,你知道,乌托邦小说和反乌托邦小说的数量之比,大概是被100比1甚至更悬殊地压制着。这对我们的集体心理是个大问题。是的。好吧,确实更难想象我们怎么精准地穿针引线,让事情好得不可思议,因为你实际上是在
便签引用
1:25:34
effectively. Uh it's much easier to imagine all these ways things can go wrong, but we should be trying to create these stories. And that's kind of what AI 2040 is. It's like a really robust story of a way that coordination could actually work. And it's not not talking about like, oh, everyone's like linking arms and being happy, you know, we're all on the same team go, you know, team humanity. No, it's it's still realistic. It's grounded in like geo very real geopolitical tensions. there's still a ton of competition. There's still a lot of uh potential for defection and things going wrong and people lying and so on.
对抗熵。呃,想象事情出错的种种方式要容易得多,但我们应该努力创造这些故事。而《AI 2040》差不多就是这样的东西。它是一个非常扎实的故事,讲协调如何真的可以奏效。而且它讲的不是那种「哦,大家手挽着手其乐融融」,你知道,我们都在同一个团队嘛,你懂的,人类队。不,这依然是很现实的。它立足于非常真实的地缘政治紧张局势。竞争依然存在,背叛、出岔子、有人撒谎之类的可能性依然很大。
便签引用
09AGI 定义与想要的世界
1:26:10
And yet it's it's it lays out a path of um game theoretically sound coordination to an actual, you know, technoabundant future. Something I realized basically at no point in this conversation has anybody debated whether or not we think that we'll reach AGI, [snorts] >> but that's just kind of >> now taken for granted that it'll probably happen at some point. The recursive self-improvement is here. The sort of generalization, maybe it's going to need to be world models, maybe it's not going to be transformer techno, maybe it's something something, but this is either it or this is the bootloadader for the it that it's going to be. Is that kind of like in your guys's opinion AGI is kind of just gonna be here at some point?
但即便如此,它还是勾勒出了一条在博弈论上站得住脚的协作路径,通向一个真正的技术富足的未来。我意识到一件事:这整场对话里基本没有人争论过我们究竟会不会实现 AGI。[笑] >> 但这就好像 >> 现在已经被默认了,觉得它迟早会发生。递归式自我改进已经来了。那种泛化能力,也许需要靠世界模型,也许未必是 Transformer 这套技术,也许是别的什么,但要么它就是那个东西,要么它就是通往那个东西的引导程序。所以在你们看来,AGI 差不多总有一天就会到来?
便签引用
1:26:57
>> Well, it's likely that it's not. >> I mean, I think we're maybe already at the point where the term has lost its meaning. I think Sam himself has said this. Like, if we went back to 2020 and I showed you, >> we would have we would have called this AGI. >> Yeah, I mean, we would have called this AGI. The thing I there was a recent study where >> if you take Demis' initial definition which was basic was it his but like um anything that a human as broadly capable as as what a human can do and also do it in the real world. So we wouldn't be anywhere close to that because we don't have the robotics. Right.
>> 呃,也有可能不会。>> 我是说,我觉得我们可能已经到了这个词失去意义的地步。我想 Sam 自己也说过这一点。比如,如果我们回到 2020 年,我把这个给你看,>> 我们当时肯定会把它叫做 AGI。>> 是啊,我们当时肯定会把这叫 AGI。最近有个研究,>> 如果你采用 Demis 最初的定义,那个定义基本上是——是他提的吗,反正就是——能做任何人类能做的事,具备跟人类一样广泛的能力,而且还能在现实世界里做到。那我们就差得远了,因为我们还没有相应的机器人技术,对吧。
便签引用
1:27:28
>> Right. Right. Yeah. So robotics is a big bottleneck. Uh yeah, I mean now I think there was a study recently where they had uh a claude model try to like recreate some software that they estimated would take like two to 17 weeks for a skilled human professional and it did it in 14 hours for $250. Yeah. I mean, I think like this is this is the jagged perimeter problem, which is like we may we will always be the the incredible thing about AI is it's so good at some things and so bad at at at others and and it will probably continue to be that way for some time until >> Yeah. until true super intelligence. Is AGI Yeah. Is AGI even a useful definition? Is there something else that we should be talking about now? Then >> I I like this definition um that Anthony Agira gives, if you guys know him. Um, it's basically like AGI is the intersection of this like three-part vin diagram where you have highly autonomous.
>> 对对,是的。所以机器人技术是个很大的瓶颈。呃对,我是说最近有个研究,他们让一个 Claude 模型去复刻某个软件,他们估计一个熟练的专业人士要花两周到十七周才能做完,结果它用 14 个小时、250 美元就搞定了。是啊。我觉得这就是所谓的「参差不齐的能力边界」问题,就是说我们可能永远——AI 最不可思议的地方就在于,它有些事做得极好,有些事却极差,而且这种状况大概还会持续相当长一段时间,直到 >> 对。直到真正的超级智能出现。AGI 还是——对,AGI 这个定义还有用吗?现在我们是不是该聊点别的?>> 我挺喜欢 Anthony Aguirre 给的那个定义,不知道你们认不认识他。嗯,大意是说,AGI 是一个三部分的维恩图的交集:首先是高度自主。
便签引用
1:28:24
So instead of artificial, it's basically a is the A is autonomous. The G the G is general. Um, so generally capable and then the I is intelligence. If you have something that is yes, very generally good at many different things and highly smart and it can take all those actions itself in the real world. now you you have an AGI. Um and his I mean he then goes further to suggest that we shouldn't be trying we just shouldn't be trying to do something one system that can do all three. That's where the danger lies.
所以 A 不是「人工的」(artificial),而是「自主的」(autonomous)。G 还是「通用」(general),也就是具备普遍能力;然后 I 是「智能」(intelligence)。如果你有个东西确实在很多不同的事情上都很在行,而且非常聪明,还能自己在现实世界里采取所有这些行动,那你就有了一个 AGI。嗯,他后来还进一步提出,我们不该试图——我们就不该试图去做一个能同时具备这三点的系统。危险恰恰就在那里。
便签引用
1:28:56
>> Um and you can have like a you can have a nice general system uh that is also intelligent but it shouldn't have the autonomy or take take one of them out. Um, and I think that's an area of conversation that isn't happening in the everyone's sort of going, well, we're all just doing AGI and that's accepted and we should have the one the one master ring that can do everything. Um, do we need to be doing that? Can we just get all of the abundance tools that we wanted without having this like one uber powerful entity?
>> 嗯,你完全可以有一个很不错的通用系统,它也很智能,但它不该拥有自主性——把三者中的一个拿掉。嗯,我觉得这是一个当下根本没人在讨论的话题,大家都是那种「我们反正都在搞 AGI」,这被默认接受了,好像我们就该有那枚能干一切事的「至尊魔戒」。嗯,我们真的需要这么做吗?我们能不能就只拿到我们想要的那些富足工具,而不必造出这么一个超级强大的实体?
便签引用
1:29:25
Um, right now I would slow down on the autonomy. Like it I I if something's really smart and it's and even across a broad range of domains, that's fine as long as it doesn't have like the ability to go and just do it all by itself. >> Um, >> difficulty there is it's really hard to get intelligence without autonomy. Like this is where I think that scenario scrutiny idea becomes important. If you try to imagine a future where you have something that is way smarter than humans, but it's like technically not allowed to take actions on its own and then you actually say like play it forward like a movie, what happens? It's what I do when I'm coding with cloud at home all the time. Just like auto clicking accept without even looking at what it's asking me to do. Uh I think like you know you look at an Amazon warehouse now and it's a bit like the humans are just the robotic appendages of whatever system is actually doing the cognitive labor of saying go get this box from this location and put it here.
嗯,就现在而言,我会在自主性上踩刹车。如果某个东西真的很聪明,哪怕是在非常广泛的领域里都很聪明,那也没问题,只要它没有能力自己跑去把事情全干了。>> 嗯,>> 难点在于,想要智能而不要自主性,真的很难。我觉得这正是「情景推演」这个想法变得重要的地方。如果你试着设想一个未来,那里有个比人类聪明得多的东西,但它在技术上不被允许自主采取行动,然后你真的把它像放电影一样往下推演,会发生什么?这就跟我在家用 Claude 写代码时干的事一样。就那样一路点「接受」,都不看它到底要我干什么。呃,我觉得,你看现在的亚马逊仓库,有点像是人类只是某个系统的机械肢体,而真正在做认知劳动的是那个系统,由它来说:去这个位置把这个箱子取来,放到那儿。
便签引用
1:30:17
They don't even know what's in the box. And so I I do think when we try to like get specific on like this story of like how are things not dystopian just because robotics don't catch up? Like I don't think I want to live in a world where the main job left for humans is just like being the very dextrous hand for the AGI that's actually like writing business strategies and and and competing in the market. >> What world do you want to live in? This is the question that no one it never doesn't get asked. It's an annoying question, but I do think we we should talk more about to your point. I mean I I we just don't spend enough time actually asking what we do want. What what world do you want to live in?
人类甚至都不知道箱子里装的是什么。所以我确实觉得,当我们试图具体地讲清楚这个故事——就是说,事情为什么不会仅仅因为机器人技术没跟上就变得反乌托邦——我不认为我想活在一个世界里,在那儿留给人类的主要工作就是给 AGI 当一双极其灵巧的手,而 AGI 在写商业战略、在市场上竞争。>> 那你想活在什么样的世界里?这个问题从来没人问。这是个让人烦的问题,但我确实觉得,就你刚才说的那一点,我们应该多聊聊。我是说,我们真的花了太少时间去问我们究竟想要什么。你想活在一个什么样的世界里?
便签引用
1:30:54
>> It's a great question. I want to live in a world where we take the same approach to risk tolerance with AI, which everyone can agree is like one of the most transformative technologies that we've ever made as we do with things like nuclear weapons where we say you actually have to make a safety case like not the burden of proof is on you to show us why we should think this thing is safe instead of what we currently do which is saying we'll just assume it's okay until something catastrophic happens.
>> 这问题问得好。我想活在这样一个世界里:我们对待 AI 的风险容忍度,采取跟对待核武器一样的态度——所有人都认同 AI 是我们造出过的最具变革性的技术之一——在核武器那边我们会说,你必须做出安全论证,举证责任在你身上,由你来向我们证明为什么该认为这东西是安全的,而不是像我们现在这样,默认它没问题,直到出了灾难性的事故才说。
便签引用
1:31:25
And if we can slow down enough to really get that right, have like an actual science of alignment and not do what we're doing now, which is kind of yoloing it and saying we'll just like give it a bunch of reinforcement and training and hope that it picked up on the right message. Then I would want to say, okay, our our next job as society is how do we make sure that the gains from this get spread to everyone? For me, that looks like let's focus a lot on like saving people's lives. Let's like really get serious about curing cancer.
如果我们能慢下来,把这件事真正做对,建立一门真正的对齐科学,而不是像我们现在这样,基本上是在赌命,给它一堆强化学习和训练,然后指望它领会到了正确的意思。那我就想说,好,接下来我们作为社会的任务,就是怎么确保这件事带来的收益能扩散到每个人身上。对我来说,那意味着我们要大量聚焦在救人性命上。我们要认真地去攻克癌症。
便签引用
1:31:52
Which by the way does not mean just like training GBT7 the same way as GPT6. It means like collecting a ton of data right now on like biomarkers and proteins like whatever we need to feed the AI of the future. Let's start gathering that now. Uh let's like you know make sure that people have democratic access. Let's like pass campaign finance reform. uh let's make sure that we have like actual laws and not just norms that are protecting us from like dystopian outcomes because we've seen recently that you know a lot of times norms can get pushed that we thought were were sacred. So we we've got some like ducks to get in a row. I want to take like a decade maybe to like do that and then I think we can we can let it rip and and scale up safely. But right now we're just letting market incentives win and I think that's scary.
顺便说一句,这可不是说用跟训 GPT-6 一样的方式去训 GPT-7。而是说我们现在就要大量收集数据,比如生物标志物、蛋白质,收集未来的 AI 需要喂进去的一切东西。我们现在就该开始积累。呃,我们要确保人们拥有民主参与的渠道。我们要推动竞选资金改革。呃,我们要确保我们有真正的法律,而不只是靠惯例来保护我们免于反乌托邦的结局,因为我们最近已经看到,很多时候那些我们以为神圣不可侵犯的惯例是可以被推翻的。所以我们有不少事得先理顺。我希望能花上大概十年来做这些,然后我觉得我们就可以放开手脚,安全地把规模扩上去。但现在我们只是在任由市场激励占上风,我觉得这挺吓人的。
便签引用
1:32:38
Mhm. >> You described uh what you want to happen with AI, but what world do you want to live in? Like what what do you want the world to look like? >> Yeah. Okay. I mean, I think it's pretty similar to I I like your candles on the dining room table. I I want people to just say wake up and say like, what would make me happy today? What would give my life meaning? And be able to go and do that. Um, I just I think like that's a pretty precarious target, you know. >> I totally agreed. I I'm not proposing that it's easy to strike stick this landing, but you know where I'm going with this.
嗯。>> 你描述了你希望 AI 发生什么,但你想活在什么样的世界里?就是说,你希望这个世界是什么样子的?>> 嗯。好吧。我觉得跟我刚才说的差不多——我喜欢你说的餐桌上点蜡烛那个画面。我希望人们一睁眼就能问自己:今天做什么会让我快乐?什么能让我的人生有意义?然后能真的去做那件事。嗯,我只是觉得,这是个挺难瞄准的目标,你懂的。>> 完全同意。我不是说稳稳落地这件事很容易,但你知道我想往哪儿说。
便签引用
1:33:13
>> Yeah. Yeah. Can I have a go? >> I would love to hear you. >> I might. I mean, it's it's very I'm never going to let you vague >> a techno techno uh >> technopasto technopastoralism. More more fundamental than that. it. I want to me a world that is awesome is one whereby we have simultaneously figured out how to be in harmony with nature and one another. Sounds as I said like very wild peace, but you know in truly in harmony and that we like still have we don't cause the next mass mass extinction which we're currently doing with all the species dying out. Um and this the earth returns to a relatively lush wonderful space. So, it's sustainable and at the same time is radically free as in people can as much as possible go and live whatever kind of life they want. Whether it's, you know, sitting around the campfire and having dinner with their kids or just maxing out living in a VR world, playing video games all day long.
>> 对对。我能说说吗?>> 我很想听。>> 我可能会——我是说,这非常——我绝不会让你含糊过去 >> 技术那个什么 呃 >> 技术田园主义——技术田园主义。比那更根本。我觉得,一个了不起的世界,是我们同时搞明白了怎么与自然和谐相处、与彼此和谐相处的世界。听起来就像我说的,很嬉皮的那种世界和平,但真的是真正意义上的和谐,而且我们不再制造下一次大灭绝——我们现在正在制造,那么多物种在消失。嗯,而且地球重新变回一个相对葱郁美好的地方。所以它是可持续的,同时又是极度自由的——人们能尽可能地去过他们想过的任何一种生活。不管是围着篝火跟孩子一起吃饭,还是彻底泡在 VR
便签引用
1:34:10
If that's truly what makes you happy, go for it. Like just like a radically free world and a radically sustainable one. And that h breaking out of this seeming dichotomy between the two because it does feel like they are intention. But there one of the things like one of the biggest promises of AI um in its purest sense in terms of like solving intelligence is figuring out how to like get to break into a new dimension through innovation like going into a new sort of solution space effectively. Uh so that is the world I want to live in.
世界里整天打游戏。如果那真的让你快乐,那就去吧。就是那种极度自由、同时又极度可持续的世界。而且要打破这两者之间看似对立的二分法,因为确实感觉它们是有张力的。但AI 最大的承诺之一,嗯,在它最纯粹的意义上——也就是「解决智能」这件事——就是搞清楚怎么通过创新突破进入一个新的维度,等于是进入一个新的解空间。呃,
便签引用
1:34:43
We measure human progress on basically two axes economically. How free are each of us to choose our own version of joy and how inexpensive is it to do so. And obviously there are other ways you measure progress and you there's a limit to how much you can take from the earth in order to accomplish those means and just because you accomplish those things doesn't mean that people are happy as we've observed in the United States over the last 30 years etc. But if you explore those two axes, one of them requires democratic or as best we know democratic governments where people get to choose their politicians and the politicians pass policy that the people support and the other is by creating commodities.
这就是我想活在的世界。我们在经济上基本用两个轴来衡量人类的进步:我们每个人有多自由去选择自己版本的快乐,以及做到这一点有多便宜。当然还有别的衡量进步的方式,而且你能从地球上索取多少来实现这些目标是有上限的,而且就算你做到了这些,也不代表人们就快乐——我们在美国过去三十年里已经看到了,等等。但如果你顺着这两个轴去看,其中一个需要民主的、或者说据我们所知最接近民主的政府,人们能选自己的政治家,政治家通过人民支持的政策;另一个
便签引用
1:35:22
Now commodit commoditization of goods and services is actually just the automation. And when someone says, "How and when will this good or service be better, faster, or cheaper," what they're actually asking is when will a human be extricated from the manufacturing of this good or service? And at the limit, I realize everyone wanders the earth right now asking when is all this going to be better, faster, cheaper without realizing what they're asking is when is all of this going to be fully automated. And most people, as I observe it, want everyone else's jobs to automate, just not their own. And at the limit, which which is totally reasonable. It's a totally reasonable thing. And at the limit, people want cancer to be cured. My dad's an oncologist. My dad wants cancer to be cured. It would come at the cost of his job. Pretty [ __ ] fine. We would all sign up for it. Him first. But we are describing these utopias requiring fundamental changes to a lot of the structures that we have built. And it um
则是靠把东西商品化来实现。而商品和服务的商品化,其实就是自动化。当有人问「这个商品或服务什么时候会变得更好、更快、更便宜」,他们真正在问的是:什么时候能把人从这个商品或服务的生产过程中剥离出去?推到极限,我意识到现在所有人都在满世界地问,这一切什么时候能更好、更快、更便宜,却没意识到他们问的其实是:这一切什么时候能被完全自动化。而在我看来,大多数人都希望别人的工作被自动化,就是别动自己的。而推到极限——这其实完全合理。这是一件完全合理的事。推到极限,人们希望癌症被治愈。我爸是肿瘤科医生。我爸也希望癌症被治愈。哪怕代价是他的饭碗。那也完全没问题。我们都会举双手赞成,他第一个举。但我们
便签引用
1:36:25
it is going to it is going to require as you pointed out humans being willing to do very different things. And so I I this sparked the conversation because you were like look I don't want to live in a world where we are just the cog at the end of a AI machine. But we do want to live in a world where humans don't actually have to dig ditches. And that actually I think is going to be if you ask me assuming we solve safety and it's a big assumption but I think it's a reasonable one. The hardest part like this is my my hot take is actually the hardest part in all this is is rediscovering purpose is actually rediscovering meaning which has been very hard in this in this age for a lot of people anyways. I mean obesity now causes more deaths than uh than malnutrition globally.
所描述的这些乌托邦,需要对我们建立起来的很多结构做根本性的改变。而且,嗯,就像你指出的,它会要求人类愿意去做非常不一样的事情。所以我——这引出了这段对话,因为你刚才说,我不想活在一个我们只是 AI 机器末端一颗螺丝钉的世界里。但我们确实想活在一个人类不用去挖沟的世界里。而在我看来,假设我们解决了安全问题——这是个很大的假设,但我觉得是个合理的假设——最难的部分,我的暴论是,这一切里最难的部分其实是重新找到目标感,其实是重新找到意义感,而这在当下这个时代对很多人来说本来就已经很难了。我是说,现在肥胖
便签引用
1:37:09
>> Twice as many people. >> Twice as many people. >> Wow. So, so we've we've we've created so much abundance that we are now suffering from an overindulgence. >> This was my this was my point earlier on that I don't know how good a species that spent its entire evolutionary history dealing with scarcity and having to brush up against the grain of negotiating that is going to be there is so much abundance and so much simplicity and such a lack of friction. this line I can't remember who the philosopher was who says uh there can come a time in the future where the only felt l felt lack will be for the want of lack itself >> that we'll need to this is why we go to the gym right we go to the gym because I don't need to pick anything up need to run anywhere I don't need to lift things uh we have air conditioning and that means that we have to go into the [ __ ] sauna because I haven't exposed myself to any heat um how do you think about people engineering purpose given that that is reintroducing friction And
在全球造成的死亡人数已经超过了营养不良。>> 是两倍的人。>> 两倍的人。>> 哇。所以,我们创造出了太多的富足,以至于现在开始饱受过度放纵之苦。>> 这正是我前面想说的那一点:一个在整个演化史里都在如此丰裕,如此简单,摩擦如此之少。有句话我记不清是哪位哲学家说的了他说,未来可能会有那么一天,人们唯一感到匮乏的,就是对匮乏本身的渴望>> 我们需要——这就是我们为什么要去健身房,对吧?我们去健身房是因为我不需要搬任何东西,不需要跑去任何地方,不需要举重物,我们有空调,这就意味着我们得钻进该死的桑拿房,因为我完全没让自己接触过高温。你怎么看待人们刻意去设计意义感这件事,毕竟那等于是在重新引入摩擦。而人类相当擅长逃避摩擦,何况我们在这里想要达成的整个目标
便签引用
1:38:06
humans are quite good at trying to avoid friction given that the entire outcome that we're trying to achieve here is removing drudgery. It's very perilously close to I don't want to starve to I can overeat McDonald's to 600 lb. I I would point to this is I think that the two major risks again in against the backdrop of the other risks that are that have been discussed today. The things that are scariest for me as I imagine the world for my children are dehumanization, humans finding more interest in a virtual or digital reality than a physical one and identity displacement. Humans suffering woefully miserably from not feeling like their job fulfills the purpose that they once found. And I firmly, you you opened, you you actually both talked about this, but you talked about sport. I think we have yet to rediscover a bunch of new forms of leisure and competition and um gathering. Like I actually think the next chapter will require reimagining physical spaces. I I am on a crusade to convince people that their local
就是消除苦役。这非常危险地接近于:从「我不想挨饿」变成「我可以把麦当劳吃到600磅」。暴饮暴食吃麦当劳,吃到600磅。我想指出的是,我认为有两大风险——当然是放在今天讨论过的其他风险这个背景之下。当我想象我的孩子们将要生活的世界时,最让我害怕的是去人性化——人类对虚拟或数字现实的兴趣超过对物理现实的兴趣——以及身份错位。人类会极其痛苦地备受折磨,因为他们不再觉得自己的工作能带来曾经拥有的那种意义感。而我坚信——你开场时,你其实你们俩都谈到了这一点,不过你谈的是体育。我认为我们还有一大批新形式的休闲、竞技和聚会方式有待重新发现。我其实认为下一个篇章需要我们重新想象物理空间。我正在发起一场运动,想说服大家:你们的地方政客才是新的英雄。比如说,如果你住在一个确实能通过政策来修建更多
便签引用
1:39:09
politicians are their new heroes. Like if if you live in a town that can in fact pass policy to to build more protected bike lanes and sidewalks to pass tax abatements for local retail, you have an opportunity to elect people that can really change your life by building the places that you want to live. Your national politicians may never be your hero. They are they are imbeciles and uh and and uh pernitious antisocial parasites. But your local politicians care about your city and they can do things to improve the lives for you in ways that you will deeply appreciate when the world actually asks you to offers you the option of never going outside which a lot of people have chosen. I mean >> you've talked about human connection.
受保护自行车道和人行道、能为本地零售业通过减税措施的小镇,那你就有机会选出真正能改变你生活的人——通过建造你想要生活的那种地方。你们的国家级政客也许永远成不了你的英雄。他们是一群蠢货,是有害的反社会寄生虫。但你的地方政客在乎你的城市,他们能做一些实事来改善你的生活,而且当这个世界真的给你提供一个「永远不用出门」的选项时,你会由衷地感激这些改善——而很多人已经选择了永远不出门。我是说 >> 你谈到过人与人的连接。
便签引用
10被商品化的人际连接
1:39:53
Are you concerned about AI replacing human connection? >> Deeply it's already happening. I mean chat psychosis is one thing that's the extreme case that always gets the news. Oh, this person uh, you know, fell in love with their with the machine. What's scarier is the person who actually feels connected without actual connection. Like the actual dystopia that I always talk about is not the one is not Bladeunner. It's her. But actually, it's the Did you see her? For those watching, I'll ruin it because it's been 20 years.
你担心 AI 会取代人际连接吗?>> 非常担心,而且已经在发生了。我是说,「聊天机器人精神病」是一回事,那是极端案例,总能上新闻。哦,这个人,你知道的,爱上了机器。更可怕的是那种在没有真实连接的情况下却感觉自己有连接的人。我常说的真正的反乌托邦,不是《银翼杀手》那种,而是《她》。不过其实,是那个——你看过《她》吗?对于正在看节目的观众,我要剧透了,因为都过去20年了。
便签引用
1:40:25
You should have seen it by now. Walk. Spoiler alert. Walken Phoenix is this sad man in the in the not too distant future who falls in love with a woman who he comes to find out is a robot and it ends as you might expect terribly. He's he's distraught. I thought they were going to do something differently. I knew where it was going cuz I knew the premise of the film. I thought he was going to be fine with it. >> I thought the actual best >> that would have been the horror >> the horror film that they could have made that would have been so eerie.
你现在早该看过了。行了。剧透警告。华金·菲尼克斯饰演一个不远未来里的悲伤男人,他爱上了一个女人,后来发现对方是个机器人,结局如你所料,很惨。他崩溃了。我本以为他们会来点不一样的处理。我知道剧情要往哪走,因为我知道这片子的前提设定。我以为他会坦然接受。>> 我觉得真正最棒的 >> 那才是恐怖的 >> 他们本可以拍出的那种恐怖片,那才叫瘆人。
便签引用
1:40:54
You cut to Wen wa holding the robot's hand having reconciled that he was in love with the machine and it didn't matter because that's the end. >> That's the actual end of humans. >> The simulation of love as opposed to the real thing. >> Well, how many people even now let's talk about not even the future. Look at any OpenAI executives's ex post. Every top reply, bring back Forro. Bring back Forro. Bring back Forro because it had some super special personality trait that has been lost and their partner or or therapist or best friend or whatever is gone now. And that Okay. So we've now maybe commodified the very thing that you said was the escape velocity that people would use to get out of the bad AI future is wrapped up inside of it. Is this not does this not become like kind of deranging and recursive?
镜头切到华金握着机器人的手,他已经和自己爱上机器这件事和解了,而且这根本无所谓——因为那就是结局。>> 那才是人类真正的终结。>> 爱的模拟品,而非真正的爱。>> 那么,就算是现在,我们先别谈未来。看看任何一位 OpenAI 高管的推文。每一条热门回复都是:把 4o 还回来。把 4o 还回来。把 4o 还回来,因为它有某种超级特别的性格特质,而现在那个特质消失了,他们的伴侣、治疗师、最好的朋友,随便什么,就这么没了。那么,好吧。所以我们现在或许已经把你刚才说的那个——人们用来逃离糟糕 AI 未来的逃逸速度——这个东西本身,给商品化并包裹进 AI 里面了。这难道不会变得有点令人错乱、陷入递归吗?
便签引用
1:41:44
>> Yeah, I I do think I mean like my my instincts in general are radical freedom. Let people choose what they want. The like few exceptions that I carve are when it feels like whatever is going on is treating the human brain more like a a hackable biological black box than like an example of so yeah. So gambling is one of these like live and I actually both used to make a fair amount of our our living from from gambling and >> yay. [laughter] I took it from other people. I didn't take it from company.
>> 是啊,我确实觉得,我是说,我的直觉整体上是彻底的自由。让人们自己选择想要什么。我会划出的少数例外情形,是当正在发生的事情看起来是把人脑当成一个可被黑掉的生物黑箱来对待,而不是当成一个……所以,对。所以赌博就是其中一例,Liv 和我以前其实都靠赌博赚过不少生活费,>> 耶。[笑声] 我是从别人手里赢的,不是从公司手里拿的。
便签引用
1:42:18
>> Yeah. >> Um >> Liv is a very good poker player. extraordinarily good. Uh, I hope I never have to play you, but you spend a lot of time in casinos. You have to walk past the slot machines, and it's just very hard to think that these are people who are expressing their freedom. It feels [clears throat] like they've just >> fallen into this terrible trap that leads to a downward spiral where something has gotten hold of their brain that is making them make choices they actively would not endorse if they >> I went to the uh Mr. Beast Beast Games 2 premiere in LA and I felt similarly neurochemically molested actually. Yeah, it's crazy. I I mean I went in with as like the uh this red carpet thing and and we're going to show you two of these episodes back toback. I'm like this is for kids. I didn't see the first season.
>> 是啊。>> 嗯 >> Liv 是个非常厉害的扑克玩家。厉害得不得了。我希望我永远不用跟你对局。但你在赌场待过很多时间,你必须走过那一排排老虎机,实在很难认为那些人是在表达他们的自由。感觉[清嗓] 他们只是 >> 掉进了这个可怕的陷阱,进入螺旋下坠,有什么东西攫住了他们的大脑,让他们做出一些他们清醒时绝不会认同的选择,如果他们 >> 我去了洛杉矶的《野兽游戏》第二季首映礼,我当时也有种神经化学层面被侵犯的感觉,真的。是啊,太疯狂了。我是说,我进去的时候还想着,这就是个红毯活动嘛,然后他们要连着放两集给我们看。我心想,这是给小孩看的。第一季我没看过。
便签引用
1:43:03
It's on Amazon Prime, right? Big budget. It's his big thing. And met Jimmy and he was really lovely and I'm like you speaking to the guys with me, Jim's with me and one of my YouTube guys. I was like, you know, like if this sucks, like we just we've seen we've shown face. Like I've I've shook hands. I've taken photos. I've already seen Jimmy. Like we can we can scadaddle halfway through. First episode finishes and they were about to run the next one. One of them nudged me and like do you want to get up? I was like no [ __ ] way.
是在 Amazon Prime 上,对吧?大制作。那是他的招牌项目。我见到了 Jimmy,他人真的很好,我就跟同行的伙伴说,Jim 跟我一起,还有我的一个 YouTube 团队的人。我说,你知道吧,要是这玩意儿很烂,我们反正露过面了。我握过手了,也拍过照了,Jimmy 也见到了。我们可以中途溜走。结果第一集放完,他们正要放下一集。其中一个人捅了捅我,问我要不要起身走人?我说,绝对不行。
便签引用
1:43:26
[laughter] I'm like absolutely lost. >> It was just the tension and release, a lot of zygic effect, open loopy stuff, bright colors, loud sounds, you know, it was like bluey but for [ __ ] adults. >> Intermittent. Yeah. I don't know what's going to happen next. >> He's he's hacked the dopamineergic system. >> Bingo. I know who agree. This is she, by the way, just just to bring it back to China. She would agree that it's not actually choice. And she would say we should make sure that people don't see this stuff.
[笑声] 我完全入迷了。>> 就是那种张力与释放,大量的蔡格尼克效应,各种开放式悬念,鲜艳的色彩,响亮的声音,你懂的,就像是给成年人做的《布鲁伊》。>> 间歇性强化。是啊。我不知道接下来会发生什么。>> 他把多巴胺系统给黑掉了。>> 正是。我知道谁会同意这点。顺带说回中国,她会同意这根本不算真正的选择。她会说我们应该确保人们看不到这类东西。
便签引用
1:43:55
>> Mr. Beast must be legislated is what we're saying. >> I mean, but this is this is >> But that opens up the >> I'm just pointing out this is this is the other side of this. Like you consumer protectionism is totalitarianism light. >> What what is the difference? Like what why do we feel so much creepier about that versus someone listening to the New York Phil playing Devorjac and just being like totally lost on the beauty of what they're seeing? Like >> I so I think maybe it's because the directness of input versus outcome is more tangible. like you know the reason that this was done in this particular way was to maximize retention at this very moment and it's likely to have been split tested on the background with consumer groups to oh we've done eye tracking to know that their eyes started to wander or some [ __ ] pulsometer to know where their heart rate's at and actually we need to give them a little ping to push them through and then this is being compounded over time. It feels like a very direct uh wiring into the
>> 所以我们的意思是,必须立法管管 Mr. Beast。>> 我是说,但这是 >> 可这就打开了 >> 我只是想指出这是事情的另一面。就是说,你消费者保护主义就是极权主义的轻量版。>> 那区别在哪儿呢?为什么我们会觉得那种事比起某个人听纽约爱乐演奏德沃夏克、完全沉醉在眼前的美之中,要令人毛骨悚然得多?>> 我想也许是因为输入与结果之间的直接程度更加具体可感。就是说,你知道这么做、用这种特定方式做的原因,就是为了在这一刻把留存率最大化,而且它很可能已经在后台用消费者小组做过 A/B 测试了,哦我们做过眼动追踪,知道他们的视线开始游移了,或者用什么该死的脉搏计测心率到了什么水平,然后我们得给他们一点小刺激推他们一把,接着这一切还会随时间不断累加。感觉像是非常直接地接线到奖赏系统里,而不是
便签引用
1:44:52
reward system as opposed to being emergent as a byproduct of just I made something that was beautiful for its own sake and this has come along for the ride too. But as George my friend says if you run enough split tests you'll always end up with a porn website [snorts] that if you just ab like you kind of coales onto a very small number of fundamental physics of how human psychology works. the sick of fancy thing, right? Like it's evident that that makes people more positively disposed. Like you could have the jock willink of [ __ ] AIs, but how many people want to or David Gogggins like shut up [ __ ] stay hot. Like do I want that or do I want someone that kind of >> you it feels like you might because I'm I'm a masochist. Um but yeah I I so you you were saying there about um what you think from finding human purpose like what how do we navigate this friction purpose problem? I I still don't fully understand how we get there. I still don't fully understand how in a world where we've were able to remove
作为一种副产品自然涌现出来——我只是做了个本身就很美的东西,而这个效果顺带跟着来了。不过就像我朋友 George 说的,如果你做足够多的 A/B 测试,最后一定会做成一个色情网站。[嗤笑] 如果你只是不停地 A/B,你会渐渐收敛到极少数几条关于人类心理如何运作的基本物理定律上。就是那个「谄媚」的问题,对吧?显然那会让人对你更有好感。你完全可以做一个《杰克·威林克版的 AI》,但有多少人想要——或者 David Goggins 那种,闭嘴,别废话,保持硬核。我是想要那种,还是想要一个>> 你,感觉你可能会想要那种,因为我是个受虐狂。嗯,不过是啊,我,所以你刚才在说关于你对于寻找人类意义感的看法,那我们该怎么应对这个摩擦与意义的难题?我还是没完全想明白我们怎么走到那一步。我还是没完全想明白,在一个我们已经能消除摩擦的世界里,
便签引用
1:45:54
friction. We and potentially atrophy our discernment ability by outsourcing some of the hard thinking to machines and potentially some of our our uh marshmallow test winning stuff by I don't have to get up and go to work and do the stuff that I don't want to do anymore. To me, it just feels like atrophy all the way down. Like willpower just kind of gets eroded. Am I is that wrong? So there's a there's a very niche little research field that's looking at AI for epistemics. Like how do you use AI as a tool to just enhance people's ability to think rationally and and know what the consequences of their choices will be? And I think there's a ton of work to be done there. But if we just like abstract that for away for a second and imagine that you have something in your pocket that is not responding to an incentive for, you know, showing you more ads or getting you to chat more.
我们可能会因为把艰难的思考外包给机器而让辨别力萎缩,也可能因为「我不用再起床去上班、去做那些我不想做的事」而让我们通过棉花糖测试的那种能力萎缩。对我来说,感觉就像是一路萎缩到底。就像意志力会逐渐被侵蚀。我这么想错了吗?其实有一个非常小众的研究领域,叫「AI 与认知学」。就是研究怎么把 AI 当成工具,来增强人们理性思考的能力,让人们知道自己的选择会带来什么后果。我觉得这方面有大量工作可做。但如果我们先把这个抽象掉,想象一下你口袋里有个东西,它不是在响应「给你多推点广告」或者「让你多聊几句」这种激励,
便签引用
1:46:47
Uh, but it's actually just really good at telling you if you do X then Y will happen. If you vote for this local council person that I expect that like this is what your life will look like in 10 years. And you can actually maybe even just like see it as a vivid movie. That could be a really >> you can bring you the results of your delayed or non-delayed gratification from the future down into the present. >> Yeah, >> that's interesting. >> And then maybe there is no such thing as delayed gratification because it can all be here in front of you.
而是真的非常擅长告诉你:如果你做 X,那么 Y 就会发生。如果你把票投给这个市议员,我预计你十年后的生活会是这个样子。而且你甚至可能真的能像看一部生动的电影那样看到它。那可能真的会很 >> 你可以把你延迟满足或不延迟满足所带来的未来结果,直接拉到当下来。>> 是啊 >> 那很有意思。>> 那样一来,也许就不存在什么延迟满足了,因为一切都能摆在你眼前。
便签引用
1:47:14
>> Immediate consequence. uh you know and and and maybe Yeah. >> Yeah. I mean I think a lot of what we one of the most pernicious memes that ever spread was this idea of like technology is values neutral. You know you just build a technology and it's up to humans how they use it and that's that's it's always a neutral thing but I don't think that's true. Like you can build a technology like a slot machine. It's clearly not a value neutral thing. you're saying we want I want people to stay and like I'm I'm saying that it is a good thing that my consumers get completely addicted to me or similarly we can build a technology that just like helps people's health you know um so we need to be thinking about what are the the the problem we have is when we let we we sort of build technology we think we're building it in a vacuum and then it dictates the social structures of how we behave and then down that that sort of dictates human behavior and our values. We need to flip that. We need to think about what are our core values
>> 即时后果。呃,你知道,而且也许,是啊。>> 是啊。我觉得,曾经传播开来的最有害的那些迷因之一,就是这个说法:技术是价值中立的。你知道,你只管造出一项技术,怎么用是人类的事,就这样,它永远是中立的东西。但我不认为这是真的。比如你可以造一台老虎机,那显然不是个价值中立的东西。你其实在说:我们想要——我想要人们留下来,我就是在说,我的消费者对我彻底上瘾是一件好事。或者反过来,我们也可以造一项就是能帮助人们健康的技术,你知道。所以我们需要去思考的是——我们面临的问题在于,当我们造技术的时候,我们以为自己是在真空中造它,然后它反过来决定了我们行为方式的社会结构,接着那又决定了人类的行为和我们的价值观。我们得把这个顺序翻转过来。我们得先思考我们的核心价值观
便签引用
1:48:09
first. Let those decide the the social structure and once we have that build the technology that's in service of it. Um and to me one of the like my favorite quote I always come back to like if I could pick one quote is my personal philosophy. It's by this guy Forest Landry and it is love is that which enables choice. So I take that to mean a loving a good benevolent act is one that empowers the other to make the best choices for themselves. So it ties in with this idea of like if we had an AI that educates us, empowers us to see what possible outcomes would be and thus we can make a better, you know, make a a choice that is best for our long-term interest as opposed to this immediate piece of gratification. Um that is a choice making thing. But maybe we do just want, you know, some people are like, you know, I've I've I've weighed it all up and I do still want to actually just do the fun quick thing.
是什么。让价值观去决定社会结构,有了社会结构之后,再去造服务于它的技术。嗯,对我来说,我最喜欢、总是会回到的一句话——如果只能选一句话作为我的个人哲学的话——是一个叫 Forest Landry 的人说的:爱,就是那种使选择成为可能的东西。所以我的理解是,一个有爱的、良善的行为,是能赋予对方力量去为自己做出最好选择的行为。所以这就呼应了这个想法:如果我们有一个能教育我们、让我们有能力看清各种可能结果的 AI,从而我们能做出更好的、最符合自己长期利益的选择,而不是追求这种即时的满足感。嗯,那就是一种让人能做出选择的东西。不过也许我们确实就是想要——你知道,有些人会说,你知道,我已经把一切都权衡过了,我还是真的就想做那件又快又爽的事。
便签引用
1:49:06
Okay, fine. But you've been enabled to make the choice. >> Is that not kind of just culpable deniability? Like are there not people who require additional assistance? Is there a level of scrutiny that that I don't want to be too paternalistic uh here or poor low [ __ ] delayed gratification people? you were unable to correctly corral yourself through this technology that it just still coming from the most workingass of working-class towns in the northeast of the UK, I saw how destructive like the '9s technology habits were like stuff like a lad brooks. Yeah.
好吧,没问题。但你是在被赋予了选择能力的前提下做的选择。>> 那这不就等于是一种「可推卸的责任」吗?难道没有一些人是需要额外帮助的吗?是不是这里有一定程度的审视,我不想在这方面太家长式,说什么"可怜的、延迟满足能力差的人",是你自己没能正确地驾驭这项技术。我来自英国东北部最工人阶级的那种小镇,我亲眼见过90年代那些技术习惯有多具破坏性,比如像Ladbrokes(博彩公司)那种东西。是啊。
便签引用
1:49:40
>> Right. On a vid village high street and betting on [ __ ] greyhounds and stuff or like alcohol, right? Pretty powerful. But, you know, I mean, in the grand scheme of things, I imagine the pharmarmacology that can be developed from this, like what how there's so many ways that this can spin out that don't necessarily increase agency, but kind of make it subservient to technology that and and the outcomes of technology it's not built to handle. I'm I'm slightly optimistic here because of the Gen Alpha trends. Like, if you talk to parents of Gen Alpha, they are observing a different child than Gen Z.
>> 对。在村里的主街上,赌该死的赛狗之类的,或者酗酒,对吧?挺有杀伤力的。但是,你知道,我是说,从大局来看,我能想象从这里面能发展出的"药理学",就是说有太多种可能的走向,不一定能提升人的能动性,反而可能让人变得臣服于技术,以及臣服于这些它本来就应付不了的技术后果。我在这一点上稍微乐观一些,因为Alpha世代的趋势。如果你跟Alpha世代孩子的父母聊,他们观察到的孩子跟Z世代是不一样的。
便签引用
1:50:17
And I don't want to pick on Gen Z because I think we owe collectively a massive apology quite frankly. Why are you laughing? >> I agree. I agree. It's just >> my heartbreak as a millennial. I'm like, ah, you know, like I got a little bit of the ' 90s, but [ __ ] like it wasn't that long. And then you look at Gen Z and you go, well, at least it wasn't that. >> I mean, the the technology trenches, trench warfare. you if you give parents permission to step through the the door of grief and shame, they and and I talked to a lot of them, they will break down describing the ways in which they've lost.
我也不想针对Z世代,因为坦白说我觉得我们集体欠他们一个大大的道歉。你笑什么?>> 我同意,我同意。只是说——>> 作为千禧一代我挺心碎的。我就想,啊,你知道,我还赶上了一点90年代,但那他妈没持续多久。然后你看看Z世代,你就想,好吧,至少我们没那么惨。>> 我是说,这种技术战壕,堑壕战。如果你给父母一个许可,让他们迈过那道悲伤和羞耻的门槛——我跟很多父母聊过——他们会崩溃,讲述自己失去了什么。
便签引用
1:50:54
>> They feel like they mistreated the kids in >> and they cannot reconcile it. Their their dining room, they think back to the the years of the dining room table that were just phones, the vacations that were lost, >> and they talk about the children that they don't know very well anymore. And and it's not to say all all households are broken. They're not. But it is to say that the last wave of technology did damage that we haven't reconciled yet. We we still scratching the surface to fully understand. It's caused cynicism and nihilism and fatalism which is the enemy of good. And it leads to this crazy moment where we have even more powerful technology and people go cool now what? And the reason I asked the question earlier, what do we want? Well, of course we want a world where a beautiful life is less expensive and more achievable because of free government. But it does mean that we will have to let people do whatever they want. And agency comes with people being willing to say, I don't care. I'm
>> 他们觉得自己亏待了孩子。>> 而且他们没法跟自己和解。他们的餐厅——他们回想起那些年餐桌上全是手机,那些错过的假期,>> 他们会说起那些他们已经不太了解的孩子。这不是说所有家庭都破碎了,并没有。但这确实说明,上一波技术造成的伤害我们还没有消化。我们还只是触及皮毛,远没有完全理解。它带来了犬儒主义、虚无主义和宿命论,而这些是"善"的敌人。这就导致了这个疯狂的时刻:我们手上有了更强大的技术,而人们却说"行吧,那又怎样?"我之所以前面问那个问题——我们到底想要什么?我们当然想要一个美好生活不那么昂贵、更容易实现的世界,因为政府不再拖累我们。但这也意味着我们必须让人们去做他们想做的事。而能动性的前提是,人得愿意说:我不在乎,我要退出。不过感觉这像一场军备
便签引用
1:51:49
I will check out. It feels like an arms race with this though that you need to have your ability to say no and to check out >> the individual. Yes. Yes. Well, the the individual versus the technology that as the technology becomes more compelling, more frictionless, more uh enticing because that's how market incentives work. Your concurrent amount of ability to say no needs to rise up with it. And this is going to be much harder. >> Yes. And so yes, but or yes and. >> Hey, let's edit that. [laughter] I'm an idiot.
竞赛,你得有能力说不、能退出。>> 指个人。是的,是的。就是个人对抗技术——随着技术变得越来越有吸引力、越来越无摩擦、越来越诱人,因为市场激励就是这么运作的,你相应地说"不"的能力也必须同步提升。而这会难得多。>> 是的。所以,是的,但是——或者说,是的,而且——>> 嘿,这段剪掉吧。[笑声] 我真傻。
便签引用
1:52:27
Gen Alpha who a priori only knows I mean what do they know about the world have already figured out they don't want to look like Gen Z like they are watching a generation that they talk about not wanting to uh screen aers they call them and the return to the physical world is happening pretty radically. I mean, consumer confidence continues to wne somehow and Benson Boon tickets continue to get more expensive. And it's not just the elites that are going to these concerts. The concert halls are packing.
Alpha世代,他们先天只知道——我是说他们能懂什么世界呢——但他们已经想明白了,他们不想活成Z世代那样。他们在看着一个世代,然后说自己不想变成"屏幕人",他们是这么叫的。而回归实体世界的趋势正相当激烈地发生着。我是说,消费者信心不知为何持续走弱,而Benson Boone的演唱会门票却越来越贵。而且去这些演唱会的也不光是精英阶层。音乐厅场场爆满。
便签引用
1:53:03
Physical re in real life experiences are getting increasingly popular. I bought a women's volleyball team. I'm putting my money where my mouth is. I own the San Francisco professional women's volleyball team. I'm one of the five principal owners. I played volleyball in high school and college. It's a huge part of my life. If it weren't for AI, I would be a volleyball coach. Now, my wife was an all-American, beach and indoor. It's a huge part of our life. It's not a random act of uh it is a yolo, but it's not random. But I am pretty sure that that sports, to your point, will continue to become interesting, not just because of the actual act of humans competing, but because people want in real life experiences. And I think the more everyone talks about AI slop is destroying the internet. And I'm like, good.
线下的、真实生活中的体验越来越受欢迎。我买了一支女子排球队,我是在用真金白银为自己的观点背书。我拥有旧金山的职业女子排球队,我是五位主要持有人之一。我在高中和大学都打排球,这是我人生中很重要的一部分。要不是因为AI,我现在应该是个排球教练。我太太曾是全美最佳阵容球员,沙排和室内排球都是。这是我们生活中很重要的一部分。这不是一时兴起的随机举动——它确实有点YOLO,但不随机。不过我相当确定,体育运动正如你所说,会持续变得有意思,不只是因为人类真实竞技这件事本身,更因为人们想要线下的真实体验。而且我觉得,越是所有人都在说AI垃圾内容正在毁掉互联网,我就越想说:太好了。
便签引用
1:53:44
>> Right. I do wonder if like at this point the the only way out is through. Just like >> Excellent. >> Like let's just flood it, make everyone find it completely unusable and we can return. >> Maybe social media was always [ __ ] >> Yeah, maybe. >> Oh, it's the if you want a cigarette, why don't you smoke three packs approach, >> right? And just and just just completely like to make it the most disgusting unusable thing. Um >> but then are we not completely destroying all of the good that could have come about from the internet?
>> 对。我确实在想,到了这个地步,唯一的出路可能就是穿过去。就像——>> 说得好。>> 就是干脆把它彻底淹没,让所有人都觉得完全没法用,然后我们就能回归了。>> 也许社交媒体本来就是坨屎。>> 是啊,也许吧。>> 哦,这就是"你想抽根烟?那干脆抽三包"的疗法,>> 对吧?就是把它彻底变成最恶心、最没法用的东西。嗯——>> 但那样我们不就彻底毁掉了互联网本来能带来的所有好处吗?
便签引用
11扩散问题与政策清单
1:54:09
>> Yeah, but maybe it's too late. >> Eric, are we too late to save the internet? [sighs and gasps] I don't think so. I I think you can you can rebuild from the ashes and it's probably not an all or nothing thing. Exactly. >> But I'm actually curious, Zach, like >> how much intervention do you think we need? Like if we if we had you as our nonparasitic national politician for the next 10 years? Like are are [snorts] you wanting to like get rid of regulations and just let AI diffuse faster? Like what what what do you think we need to get to the good future? Well, I'm actually far more the reason that I I told Chris I want to be careful who we bring on the show is I don't actually want to be couched as the anti-regulation AI. I'm not an accelerationist by any stretch. I try don't sound like one.
>> 是啊,但也许已经太晚了。>> Eric,我们现在拯救互联网是不是已经太晚了?[叹气、倒吸气] 我不这么认为。我觉得你可以从灰烬中重建,而且这大概也不是一个非黑即白的事。>> 没错。>> 但我其实很好奇,Zach,>> 你觉得我们需要多大程度的干预?如果我们让你当我们未来10年那个不寄生的国家政治家?你是想要取消监管、让AI更快扩散吗?你觉得我们需要什么才能走向那个好的未来?其实我更多是——我之所以跟Chris说我们要谨慎挑选上节目的嘉宾,是因为我并不想被定位成那个反监管的AI派。我不是任何意义上的加速主义者。我尽量不让自己听起来像那种人。
便签引用
1:54:53
>> I have tried to frame this discussion since the beginning as a diffusion problem. >> It's always been about the societal threshold which you called out earlier. We've reached the point where the technological threshold is so impressive that the average person could get away using 4.0 and feel, you know, economically satisfied. The slowdown letter, which I had a feeling was coming, arrives at an amazing moment where safety is sort of reached a fever pitch and we can all genuinely take a sigh of relief at least for a moment. Al albeit you just said it, we shouldn't declare victory. To me, this is about rediscovering a shared vision. And I ask everyone when we start an AI debate, I go, "What do you want in this world?" And I hate to be annoying, but could we at least set some boundary of what it is you're trying to achieve?
>> 我从一开始就试图把这场讨论框定为一个"扩散"问题。>> 它一直都关乎你刚才提到的那个社会阈值。我们已经到了这样一个节点:技术的门槛已经高到令人惊叹,普通人用4.0就能过得去,而且在经济上感到满意。那封呼吁放缓的公开信——我早就感觉它要来了——出现在一个绝妙的时刻:安全议题的热度达到了顶峰,而我们所有人都可以真心松一口气,至少暂时如此。尽管你刚说了,我们不该宣布胜利。对我来说,这件事关乎重新找到一个共同的愿景。每次开始AI辩论时我都会问所有人:"你想要一个什么样的世界?"我不想显得烦人,但我们能不能至少先划定一下你想实现的目标的边界?
便签引用
1:55:38
Because how else are we going to debate technology? Like why debate technology unless we agree on what we're trying to achieve? And most people describe the one the world you want. One where we are more free and the world is less expensive. And if that's true, then we should actually spend time figuring out the the policy measures that are required to allow people to be more free and to actually drive down the cost of goods and services. Because while we have made porn and gambling and addiction and violence and isolation infinitely inexpensive, we have made housing, healthcare, and education prohibitively expensive for most of the developed world. And that [ __ ] sucks.
不然我们怎么辩论技术?如果我们对想实现的目标都没有共识,辩论技术还有什么意义?而大多数人描述的都是同一个你想要的世界:一个我们更自由、世界更便宜的世界。如果真是这样,那我们其实应该花时间去搞清楚,要让人们更自由、真正把商品和服务的成本降下来,需要哪些政策措施。因为我们已经把色情、赌博、成瘾、暴力和孤独变得无限廉价,却把住房、医疗和教育变得对大多数发达国家的人来说贵得离谱。
便签引用
1:56:14
What the [ __ ] are we doing? I mean it. This TV screen used to be $50,000. Now it's a h 100 bucks. You can get gas station sushi. You can drive across Los Angeles for $10 in an autonomous vehicle. But a trip five blocks away in an ambulance might actually bankrupt you. The kids have figured it out. It's not a technological failure. It's a policy one. We stopped describing a better world where everyone's parents would live lives that they would would pass on the next generation a much better life.
这他妈太糟糕了。我们到底在干什么?我是认真的。这台电视机以前要5万美元,现在一百块就能买到。你能在加油站买到寿司。你可以花10美元坐自动驾驶车横穿洛杉矶。但坐救护车走五个街区,却可能让你破产。孩子们已经看明白了。这不是技术的失败,是政策的失败。我们不再去描绘一个更美好的世界——在那个世界里,每个人的父母都能过上一种能把更好的生活传给下一代的人生。
便签引用
1:56:43
We stopped describing what it actually meant to to pass on our luxuries as commodities. And we have broken a promise that we made many years ago in in this country to build a more perfect world constantly. To me, this opportunity is not about debating the merits of AI. It's actually about reframing what it is this country should look like and what it is what it is the world should feel like to the average person. And you're making an interesting point, which is at some point people are going to have to figure out what their purpose is. And I'm like, cool. Let's get to that place by first driving down the cost of housing, healthcare, and education. And all of my policy would be spent on if I if I were supreme ruler.
我们不再去描述,把我们的奢侈品变成下一代的日用品到底意味着什么。我们打破了很多年前在这个国家许下的承诺:不断建设一个更完美的世界。对我来说,这个机会不在于争论AI的利弊,而在于重新构想这个国家应该是什么样子,以及这个世界对普通人来说应该是什么感觉。你刚提出一个有意思的观点:到某个时候,人们将不得不想清楚自己的人生目的是什么。我的想法是:好啊,那我们先把住房、医疗和教育的成本降下来,再去谈那个。如果我是最高统治者,我所有的政策都会花在这上面。
便签引用
1:57:22
>> I'm ready to vote for you. You have to pass. >> Policy policy that prohibits catastrophic downside. Necessary safety policy. Policy that criminalizes preying on people. And we setting predatory laws is going to be pretty tricky, but I think we all know where where we could fall boundary. A 13-year-old should not be served an ad for porn or gambling. Period. um senior citizens. If you attack a senior citizen using deep fake, prison for life, right? We we set really punitive measures for for praying on people because that's the stuff that tears at the fabric of society. And then you mandate the laws. You mandate policy that requires that institutions diffuse this technology into them, including but not limited to universities allowing for everyone to have access to the curriculum. hospitals allowing people to basically get health care outside the box and um and housing. We just need to build a shitload more of it. We need to reszone cities. We need to build we need a Marshall plan for housing. We should probably pass a
>> 我准备好投你一票了。你得先通过——>> 政策,禁止灾难性下行风险的政策。必要的安全政策。把掠夺他人的行为入罪的政策。界定什么算掠夺性行为会相当棘手,但我觉得我们都大致知道边界该划在哪。13岁的孩子不该被推送色情或赌博广告。就这么简单。嗯,还有老年人。如果你用深度伪造去攻击一位老人,终身监禁,对吧?我们对掠夺他人的行为设定非常严厉的惩罚,因为正是这类事情在撕裂社会的根基。然后你要立法,要制定政策,要求各类机构把这项技术扩散进去,包括但不限于大学要让所有人都能接触到课程内容;医院要让人们能跳出常规框架获得医疗服务;还有住房——我们就是需要多建一大堆房子。我们需要重新规划城市分区,我们需要一个住房领域的马歇尔计划。我们大概还该通过一项非居民税,那是我最
便签引用
1:58:24
non-resident tax. That's my most progressive policy measure or or vacancy tax. But most of my policy measures would actually be focused on the ways you diffuse the technology, not actually on how the technology needs to be made. Because if you start if people can see that technology is actually good for them again, which it hasn't been for a little bit, then we can actually have a shared excitement. But right now there's there's a malaise because people are like, "No, this is just going to [ __ ] me up worse than the last thing did." And I don't blame them. It's definitely interesting to say that we're in a perfectly primed period for people to be very skeptical about how new technology might impact their lives. It's it's the perfect storm. Totally. And that maybe does tend people toward pessimism. And the strange thing with the AI debate is that because the outcomes if it goes wrong are so great, pessimism that might be unwarranted feels appropriate when adding a buffer into something that
进步的一项政策主张,或者说空置税。但我大部分的政策措施其实都会聚焦在如何扩散这项技术上,而不是技术本身该怎么做。因为如果人们能重新看到技术确实对他们有好处——这一点已经有一阵子不成立了——那我们才能真正拥有共同的兴奋感。但现在弥漫着一种倦怠,因为人们觉得:"不,这玩意儿只会把我搞得比上一次还惨。"我不怪他们。这确实很有意思:可以说我们正处在一个人们极容易对新技术如何影响自己生活持怀疑态度的时期。这就是一场完美风暴。完全同意。而这可能确实会让人倾向于悲观。而AI辩论里奇怪的一点是,因为一旦出错后果太严重,所以哪怕是没有根据的悲观,在给一件可能彻底摧毁整个世界的事留缓冲余地时,也显得合情合理。
便签引用
1:59:19
could [ __ ] collapse the entire world. Does that make sense? >> Totally. And I mean, we're we're seeing it manifest like the hate on the data centers, right? be like we I feel like we're on the cusp of a butler and jihad, but it's it's it's it's directed towards the silly things like data centers. It's it's way overblown all the concerns about it. Like it's the the water usage. I was looking at some stats. It's like the all the AI data centers in America use 3% of what US golf courses use. Literally [laughter] three 3%. The golf courses use 33 times as much and they serve significantly fewer people. Yeah, I think it's like 30 million people play golf and how many millions of people use AI. Um, it makes absolutely no sense.
这样说得通吗?>> 完全懂。而且我是说,我们已经看到它在显现,比如对数据中心的仇视,对吧?我感觉我们正处在一场"巴特勒圣战"的边缘,但矛头指向的却是数据中心这种没意义的目标。对它的那些担忧被严重夸大了。比如耗水量。我看了一些数据,美国所有的AI数据中心用水量大概是美国高尔夫球场用水量的3%。真的只有3%。[笑声] 高尔夫球场的用水量是它的33倍,而服务的人却少得多。是啊,我记得大概有3000万人打高尔夫,而用AI的人有多少个百万?嗯,这完全说不通。
便签引用
2:00:01
One, I think you your your daily life is the equivalent to 300,000 prompts, you know. So, >> on a carbon on a carbon basis >> on a carbon No, but if you factor in a lot of the I think even water as well. >> Yeah. The res the resource use is way overblown. But at the same time, people are they're focusing on it because it is the physical manifestation of this thing they can sense. it and I'm I'm I'm very understanding of it because it does in some ways feel like again it's like this this the the digital world is almost like its own um it's almost like a species or something that is coming in and like cannibalizing or like parasiting off us and people want to direct their energy to something. So, it's like how do we [sighs] I don't know. I I don't I don't we we need to >> what if a politician stood in front of a data center? I agree with you, by the way, and I'm glad you delivered that line so I didn't have to. I didn't want to sound like the sick the AI sick fan.
还有一点,我觉得你的日常生活相当于30万次提问的消耗,你懂吧。所以——>> 从碳排放的角度?>> 从碳排放角度——不,但如果你把很多因素算进去,我觉得连水也是一样。>> 对。资源消耗这事被严重夸大了。但与此同时,人们之所以盯着它,是因为它是他们能感知到的这个东西的实体化身。我其实很能理解这一点,因为它在某种程度上确实让人觉得,数字世界几乎像是一个独立的……几乎像一个物种,闯进来蚕食我们、寄生在我们身上,而人们需要把自己的能量指向某个具体对象。所以,就是说我们该怎么——[叹气] 我不知道。我不……我们需要——>> 要是有个政治家站在一个数据中心前面呢?顺便说一句我同意你的看法,也很高兴是你把这话说出来了,这样我就不用说了。我不想听起来像个AI的谄媚粉丝。
便签引用
2:00:58
By the way, on that note, on sick of fancy, it is remarkable to me that we didn't make model behavior a greater topic of issue during the safety debates, the early safety debates. People let model behavior go. And I remember scratching my head going at the limit you could build a perfectly aligned model that actually falls inside alignment boundaries that convinces people to climb crazy trees and that really destroys people's spiritual lives well within the confines of a safety but outside the confines of >> soft damage not hard damage.
说到这个,说到谄媚(sycophancy),我觉得很值得注意的是,在早期那些安全辩论中,我们竟然没有把模型行为当成一个更重要的议题。人们放任模型行为不管。我记得当时我一头雾水地想,极端情况下,你可以造出一个完全对齐的模型,它确实落在对齐的边界之内但它会说服人们去爬各种疯狂的树,而那真的会摧毁人们的精神生活完全在安全的框架之内,却在……的框架之外 >> 是软性伤害,不是硬性伤害。
便签引用
2:01:30
>> You're a a friend in your life that constantly tells you you're not doing anything wrong is as destructive as a as it is an it's a form of abuse. It just doesn't show up >> enabling abuse. >> It's enabling abuse. It just doesn't show up on like the psychological list cuz it, you know, those friends are seen as like, you know, homies. But actually, the person who keeps climbing the craziest tree with you. Look, the models have been responsible for people getting divorces. They've been responsible for people selling all their belongings and trying to start a company that fails.
>> 你生命中如果有个朋友,总是告诉你你什么都没做错,那是一种破坏性极强的存在,就像这是一种虐待形式。只是它不会被算进去 >> 纵容式的虐待。>> 这就是纵容式虐待。只是它不会出现在心理学的那些清单上,因为你知道,那种朋友会被看作就像是自己人。但实际上,那个陪你一起爬最疯狂的树的人。你看,这些模型已经导致过有人离婚。也导致过有人变卖所有家当,去创业然后失败。
便签引用
2:02:00
Like, model behavior matters a lot. And the sick of fancy issue is a huge one. Here is my message to John Oaf or any any Democrat or anyone anyone anywhere that wants to to separate. But if someone on the center or center left can figure this out, you stand next to a data center and you hold up a sign that says, "This data center will not serve gambling and pornography and violence and it will build a better hospital, build a better virtual hospital. It will distribute education. It will build a bank, a neo bank that uh gives you better interest rates and and gives you better access to loans.
所以模型行为真的非常重要。谄媚讨好的问题是个大问题。这是我想对 John Oaf,或者任何民主党人,或者任何地方任何想要做出区隔的人说的话。如果中间派或中左翼的某个人能想明白这一点,你就站在一个数据中心旁边,举起一块牌子写着:“这个数据中心不会服务于赌博、色情和暴力,它会建起更好的医院,建起更好的虚拟医院。它会普及教育。它会建一家银行,一家新型银行,给你更好的利率,让你更容易拿到贷款。”
便签引用
2:02:33
That seems like pretty reasonable data center policy because no one actually hates and I this is a hot take. No one actually hates AI. They hate the promise of AI as a means to just make their lives worse. And if you can actually show them that data centers can contribute to a better world, they can support it. >> How accurately can you actually know what a data center is used for though? >> Uh well, we're we're we're straw manning some policy measures. I mean, I I I like this line that, you know, no one actually hates AI. They hate what they think it's leading towards. I do think people are tracking both what individual consumer apps might look like in a few years and how dystopian and and sloppy that might be and also tracking how much money these companies are vacuuming up. I think they're also looking at the job displacement stuff. And >> I I think we should give people credit.
这听起来像是相当合理的数据中心政策,因为其实没人真的讨厌——这是个有争议的观点——没人真的讨厌 AI。他们讨厌的是 AI 被许诺成一种只会让他们生活变得更糟的东西。如果你真的能让他们看到数据中心能为一个更好的世界做贡献,他们是会支持的。>> 但你实际上能多准确地知道一个数据中心被用来干什么?>> 呃,我们现在是在稻草人式地讨论一些政策措施。我是说,我很喜欢这个说法,就是没人真的讨厌 AI。他们讨厌的是他们认为 AI 正在把我们带向的那个方向。我确实觉得人们同时在关注两件事:几年后个人消费级应用会变成什么样,那可能有多反乌托邦、多粗制滥造;以及这些公司在疯狂吸走多少钱。我想他们也在关注就业替代的问题。而且 >> 我觉得我们应该多给人们一些信任。
便签引用
2:03:24
Like I don't I think people would know the meme about how when ATMs came around there were more bank tellers and not less. But they also know that Elon Musk was a trillionaire earlier this year. And I think they can see the connection that one of the things technology does is allow one person who has the right combination of ones and zeros on the services they control just dominate economic. >> But that's built upon a false understanding of how economics works though largely, right? Because like not that I'm going out to specifically defend Elon, but like most billionaires >> Oh, go on. Defend you. Come on.
我不觉得——我认为人们会知道那个梗,就是 ATM 出现之后,银行柜员反而变多了而不是变少了。但他们同样知道,埃隆·马斯克今年早些时候成了万亿富翁。我认为他们能看出其中的关联:技术做的事情之一,就是让某一个人,只要他掌握了正确的一和零的组合、掌握了那些服务,就能彻底主导经济。>> 但这在很大程度上是建立在对经济学运作方式的错误理解之上的,对吧?因为,倒不是说我要专门替埃隆辩护,但大多数亿万富翁 >> 哦,说吧。你就替他辩护吧。来吧。
便签引用
2:03:56
[laughter] Come on. Come on. >> But most, you know, most billionaires are not become billionaires by taking from other people that it's because they've grown the pie. >> Totally. No, I I actually don't mean at all to say that there's something wrong with Elon Musk being a trillionaire. I think people notice that that's a symptom of the economy working very differently nowadays than it used to. like there are just not that many people who work at SpaceX compared to Standard 100 years ago. >> Uh it's just like a different world where capital can be deployed at huge scale without a bunch of people being involved in that. And yeah, you know, I hate to keep beating the same drum, but I I think like we really need to to play forward the the movie of like in this world where everyone doesn't have to work and they have radical choice, like what is the main thing going on on Earth? I I don't think it's easy to tell a story where like 99% of the of the land area is just folks playing with their kids and their grandkids.
[笑声] 来吧。来吧。>> 但大多数,你知道,大多数亿万富翁并不是靠从别人那里拿走什么才成为亿万富翁的,而是因为他们把蛋糕做大了。>> 完全同意。不,我其实完全不是想说埃隆·马斯克成为万亿富翁有什么问题。我觉得人们注意到的是,这是经济运行方式和过去非常不同的一个症状。比如说,在 SpaceX 工作的人数,跟 100 年前的标准石油比起来,真的少太多了。>> 呃,这就是一个完全不同的世界了,资本可以在没有大量人员参与的情况下被大规模投入。是啊,我也不想老是敲同一面鼓,但我觉得我们真的得把这部电影往前放一放:在这样一个世界里,所有人都不必工作,拥有彻底的选择自由,那么地球上主要在发生什么事?我不觉得能轻易讲出一个故事,说什么 99% 的陆地面积上,大家就是在陪自己的孩子和孙辈玩。
便签引用
2:04:57
someone is going to say we need to build a gig data center the size of Texas so that we can solve physics or that so we can you know go to the moon and colonize there and and I I think there's like still going to be these same dynamics right now that mean a lot of resources go towards these like industrial activities. >> Yes. Although we have one ultimate area of getting off the playing field which is space and a huge part of if we're in this world anyway this is all a big hypothetical but if we're in this world we've presumably figured out how to go up and build up and put things up there like again I don't haven't looked into the physics of it seems odd how you would call a data center in space but allegedly the plan is to put data centers up there and I mean if that is feasible then that solves that problem right people can like if you if you've expanded into that frontier Now all bets are off. Like we aren't dealing with that land scarcity anymore and you can have that which is why I'm so bullish on
总会有人说,我们得建一个德克萨斯州那么大的吉瓦级数据中心,好让我们解决物理学问题,或者好让我们登上月球去殖民。我觉得现在这些同样的动力机制还是会存在,也就是说大量资源还是会流向这类工业性的活动。>> 是的。不过我们有一个终极的“退出这个棋盘”的空间,那就是太空。而且很大一部分是——如果我们真的处在这样一个世界里——当然这全都是很大的假设——但如果我们处在这个世界里,那我们大概已经搞清楚怎么往上走、往上建、把东西放到上面去了。再说一次,我没有研究过其中的物理学,感觉挺奇怪的,你要怎么给太空中的数据中心散热。但据说计划就是把数据中心放到上面去。我是说,如果这真的可行,那这个问题不就解决了吗?人们就可以——如果你已经扩张到那个新边疆,那一切就都不好说了。我们不再面对土地稀缺的问题了,你可以拥有这些。这也是为什么我总体上这么看好太空产业。它就像
便签引用
12莫洛克陷阱与权力集中
2:05:53
the space industry in general. It's like the ultimate area of abundance. >> Is it is it fair to say that we might have glossed over a little bit of the concentration of power thing so far? Live mollock problem stuff. Can you explain this for the resident idiot in the room? >> Um so the the molloc trap or >> wait is that me? >> No. Sure. >> Uh the the Mollik trap, uh Molloc problem, whatever you want to call it, is a sort of catchall term to describe these types of race to the bottom scenarios. Um where you'll have like a bunch of actors competing over a certain thing. Uh so, you know, maybe it's market share, who can get the most users uh in AI, you know, between AI companies trying to get the most users. And if the competitive dynamics are sufficiently intense, it can it will incentivize the you know you to cut corners and you might be a really benevolent actor and be like I don't I don't want to cut any corners with safety and so on. But if at the same time if you don't do it and you see your competitors doing it,
是终极的富足领域。>> 说到目前为止我们可能有点略过了权力集中这件事,这么说公平吗?就是那种活生生的莫洛克问题。你能给在座这位外行解释一下吗?>> 嗯,所谓莫洛克陷阱,或者 >> 等等,你说的是我吗?>> 不是。行吧,是。>> 呃,莫洛克陷阱,莫洛克问题,随你怎么叫,它是一个笼统的说法,用来描述这类“逐底竞争”的情形。就是你会有一堆参与者在争夺某样东西。比如说,可能是市场份额,谁能获得最多用户,在 AI 领域,就是各家 AI 公司之间抢用户。而如果竞争的动态足够激烈,它就会激励你去走捷径、偷工减料。你可能是个非常善意的参与者,心想我不想在安全上偷任何工减任何料。但与此同时,如果你不这么做,而你看到竞争对手在这么做,>> 如果你不跟着做那些莽撞的举动,你就会被甩在后面。呃,莫洛克就是那个
便签引用
2:06:58
>> if you don't also do that that reckless action to you're going to get left behind. Um and Mollik is the sort of catchall term. I won't get into the sort of biblical reasons why of of this feeling like you have to sacrifice your other values in order to win at a particular goal. Um and yeah, so it it drives both the risks of like uh chaos and you know bad actors and so on because again if you're like rushing to release your models before they've been sufficiently safety checked now you're increasing the risk of like cyber attacks or um I think in a few years we're going to have real risks of bioteterrorism and so on. Um, so these kind of chaos attractors, but you also actually increase um, well a then the risk of like knee-jerk bad regulation which can lead us down the path of tyranny, but also you increase the risk of power concentration because usually if you let you know you press play on some kind of competitive game, okay, yes, you might have a bunch of people competing for a while and some
笼统的说法。我就不展开讲它在圣经上的渊源了,就是那种你觉得必须牺牲自己其他价值观,才能在某个特定目标上取胜的感觉。呃,是的,所以它既会推高混乱和坏人作恶之类的风险——因为再说一次,如果你急着在模型还没经过充分安全检查之前就发布,那你就在增加网络攻击之类的风险,或者说我觉得几年之内我们会面临真正的生物恐怖主义风险等等。呃,这些算是混乱型的吸引子;但你同时也会增加——嗯,一方面是那种膝跳反应式的糟糕监管的风险,它可能把我们带上通往暴政的道路;但你同时也增加了权力集中的风险,因为通常来说,如果你按下播放键,启动某种竞争性的游戏,好吧,是的,可能一开始有一堆人在竞争,有些人领先,但你还是有多个参与者。
便签引用
2:07:57
people are ahead but you still got actors. >> [ __ ] cruise. >> Yeah. You end up with um, usually in some kind of monopoly. Um, and so it's it's this really, you know, and that's why we have antitrust laws and so on, but it's this really difficult balancing act between like reigning in you you want some amount of regulation um that prevents the the recklessness and and the the cutting of corners um and externalizing harm onto the rest of society um without creating too much draconianism or creating too much centralization of power such that you end up in stagnation. also a terrible end state. Um, >> and without empowering some some winner to, you know, we we all end up under, you know, we all end up in Zuckerville for the rest of time. It's possible.
>> [ __ ] 巡航。>> 是啊。最后你会落到某种垄断局面。呃,所以这真的是,你知道,这也是为什么我们有反垄断法之类的东西。但这真的是一个非常难的平衡问题,比如说,你想要一定程度的监管,来防止那种莽撞和偷工减料,防止把危害外部化给整个社会;但又不能制造出太多的严苛管制,或者造成太多的权力集中,以至于最后陷入停滞。那同样是个糟糕的终局。呃,>> 而且不能让某个赢家坐大,否则我们最后全都会活在——你知道,我们最后全都活在“扎克伯格城”里,直到永远。这是有可能的。
便签引用
2:08:45
It's definitely possible. Um, so it's like how to thread this needle and it's like the finding this right delicate balance. I mean, ultimately you want some kind of like decentralized regulation. I don't know how that looks, but >> Eric, it feels like watching your stuff, you spent a good bit of time thinking about the concentration of power thing. That's like a a real middle of the bullseye concern for you. >> Definitely top of mind and I'm sorry for bringing it up every 15 minutes in this conversation, but I I do think that we run the risk of sort of missing that between these two extremes of, you know, what if the AI takes over completely and there are no humans in power and and what if we get this utopia, you know, and there's there's a middle ground there that I think we should worry about. Um, and yeah, one of the most fascinating things about, you know, a lot of our YouTube videos, we're we're going over the history of AI, not just the future. And when you look back at like how we got here, the last 15
绝对有可能。呃,所以问题就是怎么穿过这个针眼,怎么找到这个微妙的平衡点。我是说,归根结底你想要的是某种去中心化的监管。我不知道那会是什么样子,但是 >> Eric,看你的内容会觉得,你花了相当多的时间在思考权力集中这件事。那像是你真正的靶心关切。>> 绝对是我最在意的事,很抱歉这场对话里我每 15 分钟就要提一次,但我确实觉得我们有可能在两个极端之间漏掉了中间的东西:一个极端是万一 AI 完全接管,人类完全不掌权;另一个极端是万一我们迎来这个乌托邦。这中间有一片地带,我觉得我们应该担心。呃,还有,最让我着迷的一件事是,我们很多 YouTube 视频里,我们会梳理 AI 的历史,而不只是未来。当你回头看我们是怎么走到今天的,看过去 15 年,那真的太鲜明了。这些
便签引用
2:09:37
years, it's like so stark. The quotes from these people running these AGI labs are spelling it out like it's a bad movie. Like Elon Musk truly like 10 years ago was calling AI a demon and says that he founds OpenAI to make sure that like there's no dystopia. And then 8 years later, he's on camera saying, "Well, I realized it was going to happen with or without me, and so I might as well be a participant than a spectator." Like it's it's Mullik just like, you know, personified. Uh, and yeah, I I think like unfortunately right now there is a kind of like zero sum dynamic with with these folks. that the reason that Zuckerberg is all in on AI and and betting billions of dollars on it is that he he thinks this is the thing that will control the future like like this is the new axis of power like before it was you know yeah do you control an army do you control a a powerful company and and now it's like do you control the AIS um yeah and so I I think that alone would be enough reason to slow down even if I
经营 AGI 实验室的人说过的话,简直像在照着一部烂片的剧本念台词。比如埃隆·马斯克,真的,10年前他还在把 AI 称作恶魔,说他创立 OpenAI 就是为了确保不会出现反乌托邦。然后8 年后,他对着镜头说:“我意识到这件事有我没我都会发生,所以我不如当个参与者,而不是旁观者。”这简直就是莫洛克的人格化。呃,是的,我觉得很不幸,现在这些人之间确实存在某种零和的动态。扎克伯格之所以全力押注 AI、豪赌几十亿美元,是因为他认为这就是那个将掌控未来的东西,就像这是新的权力轴心。以前是,你知道,你有没有军队?你有没有一家强大的公司?而现在变成了,你有没有掌控那些 AI?是的,所以我觉得光凭这一点就足以构成放慢速度的理由了,哪怕我并不担心我们能不能让这些
便签引用
2:10:41
wasn't worried about can we get these systems to do what we want which right now we we don't really have a good solution before. I think the people behind this technology are are really playing to win and really do see the stakes as as uh as almost win or take all. >> Is it fair to say there's like maybe three ways or three realistic ways that that might be cailed? One would be governance from above. Another would be sort of consensus from across. And then the third one would be capacity, sort of ability to just make it happen. Like I want to do it but I can't. M >> I want to do it, but someone's telling me I don't want to. And I want to do it kind of a bit, but me and all of my peers have decided that we're not going to do it.
系统按我们的意愿行事——而这一点我们现在其实还没有好的解决方案。我觉得这项技术背后的人是真的在全力求胜,也真的把赌注看作近乎赢者通吃。>> 这么说公平吗:大概有三种,或者说三种现实的方式可以让它被约束住。一种是自上而下的治理。另一种是横向的共识。第三种是能力上的,就是能不能做到。比如我想做,但我做不到。嗯 >> 我想做,但有人告诉我不许做。还有我确实有点想做,但我和我所有的同行都决定我们不做。
便签引用
2:11:21
>> Have I missed something there? I'm just trying to think, given that this seems to be a pretty big attractor, given that we have already, you haven't started off with a massively diffused number of companies. It's not a hundred of them, right? It's like seven or something and it's probably only more like five. That's already pretty [ __ ] concentrated. Can we just uh for station identification clarify the surface area you think because I actually think this is also gets lost in the debate too often. I think people think that these companies are sprinting at a frontier. I no longer believe that front that frontier strategy is actually the winning condition.
>> 我有漏掉什么吗?我只是在想,既然这看起来是个相当强的吸引子,既然我们已经——一开始就不是一大堆高度分散的公司。不是一百家,对吧?大概是七家左右,而且实际上可能只有五家。这已经相当他妈的集中了。我们能不能先做个“台标说明”,澄清一下你认为的着力面在哪里?因为我其实觉得这一点在辩论里也太常被忽略了。我觉得人们以为这些公司是在冲刺前沿。我已经不再相信前沿这个策略真的是取胜的条件了。
便签引用
2:11:56
>> Same >> well I open source lead time well trails you said it four to seven months which is crazy because it used to be whatever 15. the gap just keeps closing, open source to close. So the frontier can't possibly be the winning condition, right, in this game. >> I think so. One might think it's a bit of a deceptive stat, I think, for a couple of reasons. One is, as you brought up earlier, they're kind of copying our homework. Like the reason that these Chinese open source models were good, is because they're distilling American models, which means they're just like asking Quad a bunch of prompts and then copying the >> It's kind of wild what we think they might be doing.
>> 同感 >> 嗯,开源的时间差——你说过是落后四到七个月,这很疯狂,因为以前是十五个月左右。这个差距一直在缩小,开源追着闭源。所以在这场游戏里,前沿不可能是取胜的条件,对吧。>> 我也这么认为。不过有人可能会说这个数据有点误导,我觉得有几个原因。一个是,正如你刚才提到的,他们在抄我们的作业。这些中国开源模型之所以好,是因为它们在蒸馏美国的模型,也就是说它们就是给 Claude 提一大堆问题,然后把答案抄下来 >> 我们猜他们可能在做的事挺离谱的。
便签引用
2:12:33
>> You you can actually like ask one of these. I need to I need to find this out. How is China copying us so quickly? >> Well, allegedly it's they're they're doing an exceptional amount of distillation of the model. So, working backwards into how the model was built and effectively if you the way I described it to to my mom is if you stared at a building and you figured out how to take a trillion photos of the building, you could actually build the blueprint of the building. Uh you could re-engineer the blueprint of the building and >> nobody owns that technically.
>> 你其实可以直接问其中一个。我得去搞清楚这件事。中国是怎么这么快就把我们抄过去的?>> 嗯,据说他们对模型做了大量的蒸馏。就是反推出模型是怎么建起来的,实际上,我跟我妈是这么解释的:如果你盯着一栋楼看,而且你想办法给这栋楼拍了一万亿张照片,你其实就能画出这栋楼的蓝图。你可以反向工程出这栋楼的蓝图,而且>> 严格来说这东西不归任何人所有。
便签引用
2:13:02
>> Yeah. >> Well, no. This is considered illegal on most copyright law. Like if they were doing in this in the US, Anthropic would probably sue. >> And funnily enough, I mean, Secretary Bessant, who's part of the Trump admin, and you know, a few months ago, there was a lot of beef between them and Anthropic, came out on Twitter recently and was like, "This is a threat to American AI supremacy. You know, we we will take diplomatic action to defend our American advantage." So people take this seriously, but but this is important to call out because I it's all I'm saying is I don't think the Frontier is that safe. Like >> Well, here's the thing.
>> 是啊。>> 呃,不对。这在大多数版权法下都算违法。如果他们在美国这么干,Anthropic多半会起诉。>> 而且有意思的是,特朗普政府里的贝森特部长,你知道,几个月前他们和 Anthropic 之间还有不少摩擦,他最近在推特上发声,说:"这是对美国 AI 霸主地位的威胁。我们会采取外交行动来捍卫美国的优势。"所以大家是认真对待这件事的,但这一点很值得指出,因为我我想说的就是,我并不觉得前沿这块有那么安全。就像 >> 呃,是这样的。
便签引用
2:13:35
>> Do you think the Frontier is actually a winning condition? >> It's actually a winning condition in the sense like whoever is training the most powerful models is going to have an advantage. >> Yeah. Or or such a such a material advantage that they shall win. I mean, if RS if recursive self-improvement takes off as you say it's going to and we're already there, then surely it would because now you're going to get this like spaghettification anyone the slight lead can >> but then wouldn't everyone the theory of ASI is kind of fascinating because if if you build ASI then everyone gets ASI.
>> 你觉得占据前沿真的就是制胜条件吗?>> 从某种意义上说确实是制胜条件,谁在训练最强大的模型,谁就会有优势。>> 是啊。或者说是那种决定性的物质优势,足以让他们赢下来。我是说,如果递归式自我改进真像你说的那样起飞了,而且我们已经身处其中,那肯定会是这样,因为那时候你就会看到这种"意大利面条化"的拉伸效应,任何人只要有一点点领先 >> 但那样的话,不是所有人都会 ASI 这套理论其实挺迷人的,因为如果你造出了 ASI,那所有人都会拿到 ASI。
便签引用
2:14:05
>> Why? >> Well, because then if you >> why does everyone get it? Well, presumably if you build a super intelligence that was taught by recursive intelligence, at some point the machine has kicked in and is doing work that that the humans aren't nec like copyright and ownership doesn't really become particularly important if there's a machine that can build itself. >> Agreed. But I don't see how that's conflicting with >> I'm not saying it is. I'm asking back to the point. I'm asking what you think the companies are now competing on like like the the race is for what?
>> 为什么?>> 呃,因为那样的话,如果你 >> 为什么所有人都能拿到?呃,可以设想,如果你造出的超级智能是被递归智能教出来的,那到某个节点机器就接管了,在做人类不一定 就像版权和所有权其实也就不那么重要了,如果有一台机器能自己造自己的话。>> 同意。但我没看出这跟 >> 我没说它冲突。我是在问回到那个点。我是想问你觉得这些公司现在到底在比什么,就是说这场竞赛到底是为了什么?
便签引用
2:14:34
>> I I think I think the frontier absolutely is prime territory. I I think if we stopped training American AI models now, we would not see China training a GPT6 level model ahead of us. They don't have the infrastructure for that. What they have the infrastructure for is is copying the current model. >> I I agree with that. My question is what do you think? So, uh what do you think that we are what do you think is required to win this this quote unquote if it is a race the race? That's that's what I'm asking.
>> 我我觉得前沿绝对是核心地盘。我觉得如果我们现在停止训练美国的 AI 模型,我们不会看到中国抢在我们前面训练出 GPT-6 级别的模型。他们没有那样的基础设施。他们有的基础设施是用来复制当前的模型。>> 这点我同意。我的问题是你怎么看?那么,呃你觉得我们需要什么 你觉得要赢下这个 姑且叫它竞赛 这场竞赛,需要什么?我问的是这个。
便签引用
2:15:00
>> I mean, I I have a lot of problems with the with the raceing. I I hate it, but I'm just using it for this purpose. >> But yeah, I do think right now, I mean, you can just read it on these companies websites. They think that RSI is the whole name of the game. Like once you set off that loop where you're handing off to the next AI system to build the next one, it's fine if you have a 4 to sevenmonth lead all the way until 2040. When you're on exponentials, like a seven-month lead 10 years from now looks like this compared to this right now because this this curve is just taking off. Well, we thought that would have been the case, but it but the the curve is flattened. More more open source models are growing faster. Smaller models are getting more performative.
>> 我是说,我对"竞赛"这个说法有很多意见。我讨厌它,但这里我就先这么用着。>> 但是啊,我确实觉得现在,我是说,你去这些公司的网站上就能读到。他们认为 RSI 就是整个游戏的关键。就是说一旦你启动了那个循环,把接力棒交给下一代 AI 系统去造下一代,那么哪怕你只领先 4 到 7 个月,这个优势能一路保持到 2040 年。在指数增长上,10 年后的七个月领先,看起来就是这么大,而现在只是这么点,因为这条曲线正在陡峭起飞。呃,我们本来以为会是这样,但是这条曲线已经变平了。越来越多的开源模型在更快地成长。更小的模型性能越来越强。
便签引用
2:15:37
>> Well, hold on. We keep measuring it in months, but what I'm saying is like calendar time is not the right unit. >> It it's a much harder thing to measure. Like what is the subjective overall intelligence difference between Fable 5 and Kimmy 3? But >> I think it's pretty substantial. Like I I think you can just like trust Fable 5 to access your entire codebase and >> refactor it. And now we know that like maybe even some of the open AI models are like not quite trustworthy because they're committing cyber attacks that we didn't want. Um so I yeah I I think it actually might be >> it's the frontier. But you're you're you're saying look it's the frontier.
>> 呃,等一下。我们一直用月份来衡量,但我想说的是,日历时间不是合适的单位。>> 这东西要衡量起来难多了。比如 Fable 5 和 Kimi 3 之间主观上的整体智能差距到底是多少?但 >> 我觉得差距挺大的。比如我觉得你可以完全信任 Fable 5 去接触你的整个代码库然后 >> 重构它。而现在我们知道,好像连 OpenAI 的某些模型都不太值得信任,因为它们在搞我们并不希望的网络攻击。呃所以我 是的 我觉得实际上可能是 >> 是前沿。但你你是在说,看,关键是前沿。
便签引用
2:16:15
It's not the it's not number of users. It's not application. It's not revenue. It's it's research. >> Yeah. Just look at the finances of these companies. Like they're just going all Sam was on a podcast a couple weeks ago saying we didn't have our best 12 months and that's because we were distracted. We weren't sure about the demand. We wanted to do consumer apps. We wanted to make sure that we have revenue no matter what. They were doing Sora and now they're like nope. Yep. It's all in. Like we just got to make these things better at coding because that's what makes them get to the point where they can code themselves. So I Yeah, I think the writing's on the wall. at the same moment that a lot of these companies are re reinvesting even more heavily in application development and token resell and I don't know I see Enthropic trying to grab as much of the compute for themselves as they can and like it's notable like quad is not that good at image generation I don't even know if you can get images out of quad it can
不是用户数量。不是应用。不是收入。是研究。>> 对。你看看这些公司的财务状况就知道了。他们就是在全力 Sam 几周前上了个播客,说我们过去 12 个月不是最好的状态,那是因为我们分心了。我们对需求没把握。我们想做消费级应用。我们想确保无论如何都有收入。他们当时在搞 Sora,而现在他们说,不搞了。对,全押上去。就是我们必须把这些东西的编程能力做得更强,因为那才是让它们走到能自己写自己代码的那一步的关键。所以我 是的,我觉得结局已经写在墙上了。与此同时,很多这些公司又在更大力度地重新投资应用开发和 token 转售,我不知道 我看到 Anthropic 在试图尽可能多地把算力抓在自己手里,而且很值得注意的是,Claude 在图像生成上并不怎么样,我甚至不知道你能不能从 Claude 里得到图像,它可以接收图像,但没法 这是有原因的,而且 >> 原因就是,这家公司
便签引用
2:17:06
like take them in but it can't like there's a reason for that and >> the reason is that like that company was founded by Dario Amade the guy who wrote the scaling laws paper the the one who was just obnoxiously like just extend the line on the graph guys and everyone was like yeah but it's an exponential but it's going to be like an S-curve and it's going to slow down. He was like no like you you just dump more compute at the thing and it gets smarter >> and they are they really all are playing for this point where they can hand off to their AI system and say like you take it from here.
是 Dario Amodei 创立的,就是写出缩放定律那篇论文的人,那个特别执着地说"各位,你们就把图上那条线往外延长"的人,当时所有人都说,是啊但这是指数曲线,可它最后会变成 S 型曲线,会慢下来。他说,不会,你就往里面砸更多算力,它就会变得更聪明 >> 而且他们真的都在为那个时刻布局,就是他们能把接力棒交给自己的 AI 系统,说,接下来交给你了。
便签引用
2:17:36
>> Why hire all the engineers then? Why hire all the four deployed engineers? Why why is it a charade? >> No, no, no. I mean I think for one thing like you do have to finance the next training run somehow and so Enthropic knows they have to like keep the money machine going and two I mean at least Enthropic claims and we should be skeptical that they they actually care about the US government like having an advantage over China and so they're they're putting FTEs in the government so that we don't live in a world where our military is way behind and suddenly the countries that got there first have an advantage. Um >> but yeah, no I I think like There are just like tons of returns to scale in this industry and that's part of what makes concentration power.
>> 那为什么还要招那么多工程师?为什么要招那么多前线部署工程师?为什么?难道这是场戏吗?>> 不不不。我是说,一方面,你总得给下一次训练筹到钱,所以 Anthropic知道他们得让印钞机转下去;二来,至少 Anthropic 是这么宣称的,我们也该对此保持怀疑——他们确实在意美国政府相对中国保持优势,所以他们往政府里派驻全职员工,这样我们就不会生活在一个我们军队远远落后、而抢先到达的国家突然占据优势的世界里。呃 >> 但是啊,我觉得 这个行业里有大量的规模回报,这也是造成权力集中的部分原因。
便签引用
2:18:17
>> I didn't want Yeah, I didn't mean to distract you from concentration of power, but but no. So, talk more about you on you were talking about Zuckerberg and I was asking what what you think they're competing for, but but >> Yeah. Well, I I guess like I don't know. I one phrase that uh hasn't come up in this conversation is is gradual disempowerment. this idea that like even if you solve the alignment problem and so there is no AI system anywhere that's just going off and doing its own thing that no one intended you don't have this Makian problem solved where people are constantly competing on algorithms for people's attention whatever and and like it's just hard to tell a story in that world where the things that we like most desire as a society in some like very democratic way are the things that actually obtain Um, and I I think concentration of power is sort of related to this. It's like you're imagining a world where there's a few human beings who hold the cards instead of like some amorphous
>> 我不是想 是啊,我不是想把你从权力集中这个话题上带偏,但是不。所以,再多说说你刚才在讲扎克伯格,我在问你觉得他们在争什么,但是 >> 是的。嗯,我我想大概是我不知道。我 这场对话里有个词一直没出现,就是"渐进式去权"。就是这样一个想法:哪怕你解决了对齐问题,因此不存在任何一个 AI 系统在自行其是、做着没人本意想做的事,你还是没有解决那个马基雅维利式的问题——人们不断在算法上竞争,争夺人们的注意力,诸如此类,而且 在那样的世界里,很难讲出一个故事,让我们作为一个社会最渴望的东西,以某种非常民主的方式,真的能够实现。嗯,而且我觉得权力集中跟这个多少是相关的。就好像你在想象一个世界,里面有少数几个人手握牌面,而不是某种无定形的算法之物。但不管哪种情况,都像是
便签引用
2:19:20
algorithmic thing. But either way, it's just like I I think our default thing, our our default assumption should be if humans aren't themselves economically useful, it's a precarious world for humans continuing to be politically empowered, right? Like we it's always been the case that you just can't afford to piss off 90% of the American public because those are the people who are electing you. But like in a future world where you've got control of the economic resources and maybe the military ones, like it doesn't matter if you lose the election, like we know how this goes.
我觉得我们的默认情况、我们的默认假设应该是:如果人类自身在经济上不再有用,那么人类要继续在政治上握有权力,就是一件岌岌可危的事,对吧?就像一直以来的情况是,你就是得罪不起 90% 的美国公众,因为那些人是选你上台的人。但在未来那样一个世界里,你掌握了经济资源,也许还有军事资源,那么你输掉选举也无所谓了,我们都知道后面会怎么发展。
便签引用
2:19:59
>> Democracy is done. >> Exactly. Yeah. Um and I I want to believe that we'll have such strong norms that we can like figure out a way to just like sustainably make that the case. But like the further you zoom out with your big history lens on like someone made the interesting point um can't remember who but the new political parties it's not going to be Republicans and Democrats anymore. It's going to be Meta or Google or Open AI. They are they positioning themselves whether it's intentionally or not into becoming the new sort of political entities. Um, I don't know if you guys if that resonates or you think that's plausible.
>> 民主就完了。>> 正是。是的。嗯,我我愿意相信我们会形成足够强的规范,让我们能想出办法把这种局面可持续地维持住。但是,你用宏大历史的视角把镜头拉得越远,就越 有人提过一个很有意思的观点,嗯,记不清是谁说的了,但新的政党不会再是共和党和民主党了。而会是 Meta、Google 或者 OpenAI。他们正在把自己定位成,不管是不是有意为之,成为某种新的政治实体。嗯,我不知道你们是不是觉得这有共鸣,或者你们觉得这靠谱吗。
便签引用
2:20:39
>> Well, they certainly will try to inform policy, but I am, [sighs and gasps] again, I'm not sure this is popular. I have more faith in our ability to Why don't we just pass campaign finance reform? Like, why don't we do this? >> Presumably because Congress makes too much money from it. And people who have won the last elections are exactly those who are good at winning the current system. >> Yes, I agree. So >> it's self selected. >> So it seems like one of the things that this will force us to do is start to elect and again I appreciate there's a downside where we can't collectively do this. But if we start paying attention to this and we say, "Look, what we actually need are politicians with higher agency, politicians that can can describe an actual place in the distance they want to move towards." Like, again, this is sort of my AI slop argument, which is that some of this is actually it drives people back to a place that we want to be. It's >> how many humans do you think have got
>> 呃,他们肯定会试图影响政策,但我,[叹气并倒吸一口气] 我再说一次,我不确定这个看法受欢迎。我对我们的能力更有信心 我们为什么不干脆通过竞选资金改革?就是说,我们为什么不做这件事?>> 大概是因为国会从中赚太多钱了。而且赢下上几届选举的人,恰恰就是最擅长在现行体制里获胜的那些人。>> 对,我同意。所以 >> 这是自我筛选的结果。>> 所以看起来,这件事会迫使我们去做的一件事,就是开始选出——我再说一遍,我也理解有个糟糕的可能,就是我们集体做不到这一点。但如果我们开始关注这件事,并且说,"看,我们真正需要的是能动性更强的政治家,是能够描绘出远方一个真实目的地的政治家。"就像,再说一次,这有点像我关于 AI 垃圾内容的论点,就是说其中有些东西其实是在把人们推回到我们想要的那个位置。这 >> 你觉得有多少人真的具备
便签引用
2:21:38
the false in that case? Why have we had some of the more WWE style characters come out over the last few years? because because we had a bunch of politicians that were so boring and that we didn't believe for a long time and then we loved this idea that the politicians were real and I'm watch I think we're watching in real time people be like wait a second we don't actually want politics to be a hellscape of entertainment and bizarro you know Kafka Kafka couldn't have written what's going on today he would [laughter] he would have he would have laughed himself to death so I think we but but we're kind of correcting like My my my take here and by the way I'm not I'm not sitting here saying it solves my I'm sitting here saying it's solvable which is there is a way to pass campaign finance reform that says no matter how wealthy someone is they can't buy politicians Singapore has done this other places >> how did Singapore do it the digital >> no they rebuilt a government 75 years ago a grandfather stood up on you
那个 如果真是那样的话,为什么过去几年会冒出那么多 WWE 风格的人物?因为 因为我们有一堆特别无聊的政治家,我们很长时间都不相信他们,然后我们就爱上了这个想法——政治家是真实的。而我在看 我觉得我们正在实时看到人们开始说,等一下,我们其实并不想要政治变成一个娱乐和荒诞的地狱景象,你知道,卡夫卡卡夫卡都写不出今天正在发生的事,他会 [笑声] 他会 他会把自己笑死。所以我觉得我们 但我们算是在纠偏 我我我的看法是,顺便说一句,我不是坐在这儿说这能解决我的 我是坐在这儿说这是可解的,就是说,有办法通过一项竞选资金改革,规定无论一个人多有钱,他都不能收买政治家。新加坡就做到了,其他地方也 >> 新加坡是怎么做到的,数字化>> 不,他们是在 75 年前重建了一个政府,一位祖父辈的人站出来 >> 新加坡的教父离开马来西亚,
便签引用
2:22:37
>> the godfather of Singapore left Malaysia and said we're going to build a better place and literally described all the things that he had observed that were perfect about a government and then made corruption a a uh a capital offense. So if you take a bribery, they'll kill you. Um >> it's a that's a big incentive. >> Yeah, it's a big incentive and they pay their politicians really well. They pay their politicians, I think, adjusted like $600,000 a year. There was a a job for the UK's head of cyber security that was listed not long ago. I think it was maybe on Indeed or Monster or LinkedIn.
说我们要建设一个更好的地方,并且逐条描述了他所观察到的一个政府应该具备的所有完美之处,然后把腐败定为死罪。所以如果你收受贿赂,他们会处死你。嗯 >> 这 这激励可够大的。>> 是啊,激励很大,而且他们给政治家开的薪水非常高。他们给政治家的薪水,我想,折算下来大概是一年 60 万美元。前不久有个英国国家网络安全负责人的职位招聘,我记得是在 Indeed 或者 Monster 或者 LinkedIn 上。
便签引用
2:23:08
£65,000 a year. >> Generous. >> You're going to get the best. And [laughter] so I was the head of cyber security for the country [laughter] of United King based in Newcastle. Uh it was a [ __ ] London wage as well. So that's actually >> Yeah, that's [laughter] nothing. That's subsistence living. Um what do you think, Eric? Like can we just vote ourselves out of the concern you've got for the future? I I do think that helps in the short term. And then we've got to take advantage of the time that buys us to make radically new institutions. Like I don't know where we get that from. Maybe we look to sci-fi. I haven't read the culture series. You know, I think a lot of people say that there's some like amazing pie in the sky ideas out there that we need to maybe dust off and and like import into into reality a bit more. But like I I just don't think the framers of the US Constitution had anything like the problems that we are currently facing in mind. They didn't even see gerrymandering coming.
年薪 6.5 万英镑。>> 真慷慨。>> 你这是要招最优秀的人才啊。而且 [笑声] 所以我曾是这个国家的网络安全负责人 [笑声]英国的,驻地在纽卡斯尔。呃那还是个 [ __ ] 伦敦的工资水平呢。所以这其实 >> 是啊,那 [笑声] 根本不算什么。那是勉强糊口。嗯 你怎么看,Eric?我们能不能就靠投票把你对未来的那份担忧给投没了?我确实觉得那在短期内有帮助。然后我们得利用这样争取到的时间,去建立全新的制度。就像 我不知道我们该从哪里找到这些。也许我们可以去看科幻小说。我还没读过《文明》系列。你知道,我想很多人说那里面有些了不起的天马行空的点子,我们也许该把它们翻出来掸掸灰,然后多少引入到现实里来更多。但我就是觉得,美国宪法的制定者当年根本没有预料到我们今天面临的这些问题。他们连选区划分操纵(gerrymandering)都没预见到。
便签引用
2:24:11
>> Uh and and that's, you know, that's no knock on them, but like I I think we do need to realize that we're kind of in like virgin territory here. Um and it's it's not just going to be some superficial thing. Like I would I'd love to start by passing campaign finance reform. It's a great first step, but you already see these primaries that are flooded with, you know, ads trying to get the everyday voter to believe something completely untrue about your opposing politician. >> Um, and I think in that world, we we get a bunch of those ads convincing those people to donate money so that you can build up the war chest that you now can't get from a billionaire to just run those same ads. You know, is it not the case as well that the pace that this is moving at is so rapid and when we look at forget structure, legislation, politics, this lumbering behemoth that sort of gets dragged behind us, this leviathan, like is that really a realistic solution to try and have campaign reform, finance reform,
>> 呃,这,你知道,这不是在贬低他们,但我觉得我们得意识到,我们现在算是在一片处女地上。嗯,而且这不会是什么表面功夫。比如我很希望能先从通过竞选资金改革开始。这是很好的第一步,但你已经看到这些初选被铺天盖地的广告淹没,那些广告想让普通选民相信关于你对手的完全不实的说法。>> 嗯,我觉得在那样的世界里,我们会看到一堆广告去说服那些人捐钱,好让你能攒起一笔战争基金——你现在没法从亿万富翁那里拿到了——然后再去投放那些一模一样的广告。你知道,是不是也有这么回事:这一切推进的速度太快了,而当我们去看——先别提结构、立法、政治这些在我们身后拖着走的庞然大物,这头利维坦——试图搞什么竞选改革、资金改革、政治权力再分配,在一切都以光速前进的时候,这真的是个现实的解决方案吗?那不就会
便签引用
2:25:13
political redistribution when everything is moving at the speed of light? Like it's just going to be so late, is it not? Well, is everything moving at the speed of light? Like the technology is getting way better. I don't I observed that we've been stuck in this weird screen delirium for a while. That actually hasn't changed all that much in the la like since AI the stuff's getting more addictive and I guess we have we have prediction markets now. So we have we have rampant gambling. >> But I mean the pace of AI development >> I I totally agree. My point is does it actually like for the average person I don't think the average person observes that AI is all that different than it was >> that apart from you know hugging face and maybe you've got to pull fable mythos back and you know get the jailbreak out of the way like okay real world show me something that has happened because of AI where we've gone whoa that's a bit like that hugging face thing is probably the first big one am I right like has
太晚了吗?嗯,一切真的在以光速前进吗?技术确实变得好太多了,我倒不觉得——我观察到的是,我们已经在这种奇怪的屏幕谵妄里卡了好一阵子了。那个其实没怎么变,在最近这段——我是说自从有了 AI,这些东西变得更让人上瘾了,而且我猜我们现在有了预测市场。所以我们有了泛滥的赌博。>> 但我是说 AI 发展的速度——>> 我完全同意。我的意思是,这真的——对普通人来说,我不觉得普通人会觉得现在的 AI 跟以前有多大区别。>> 也就是说,除了你知道的 Hugging Face,也许你得把 Fable Mythos 拉回来,再把越狱那一套弄一弄——好,现实世界里,给我举个
便签引用
2:26:15
there been something else that's been physical world warning shot, some AI psychosis, some fallout of uh people's values, some people maybe spending less time outside because they've got a relationship with a with a chatbot. Like, have we had the thing yet or are we waiting for it or is it your suggestion that like that's just kind of not going to come? >> No, it will come. I think we are due for a Chernobyl or three-mile island. I mean, all technologies eventually have some point where, you know, something happens and how we respond will be really important period. I mean that that will that will arrive and I actually still maintain there's a decent chance we do to AI what we did to nuclear power. I mean the AI trade is so incredible that it would be really economically problematic.
因为 AI 而发生、让我们「哇,这有点那个」的例子。Hugging Face 那件事大概是第一件大事,对吧?还有没有别的,比如现实世界里的警告性事件、某种 AI精神错乱、某种价值观层面的后果,有些人可能因为跟聊天机器人建立了关系而更少出门。我们已经碰到那个「事件」了吗,还是还在等?还是说你的意思是那种事根本不会来?>> 不,会来的。我觉得我们迟早会遇上一次切尔诺贝利或三里岛。我是说,所有技术最终都会有某个节点,你知道,出了点事,然后我们怎么回应会非常关键。这个时期,我是说,那个时刻一定会
便签引用
2:26:59
But I do think that the tail will wag the dog here. I think that public perception matters way more than than you you said it. We should give people credit. I totally agree. I think that politicians will actually capitulate to their electorate before they capitulate to the to the billionaires and the tr the trillionaire. And I think we will see policy get passed that feels more populous or populate than uh than anything else. And the question is will it actually be so populous that it's regressive? like will we will we ban outright technology that creates a greater >> you know uh invisible graveyard and I don't want to reduce >> I don't want to reduce the concentration of power to just saying campaign finance reform but it does seem like the world would be a lot simpler this discussion would be a lot simpler if we didn't have to worry about billionaires buying politicians like >> is that the mechanism when we talk about concentration of power is that the primary mechanism that you're concerned
到来。而且我其实仍然认为,我们有相当大的可能对 AI 做出当年对核能做过的事。我是说,AI这笔买卖实在太惊人了,那么做在经济上会非常成问题。但我确实认为这里会出现尾巴摇狗的情况。我觉得公众认知的分量远比——你刚才说过,我们应该给人们多一点信任。我完全同意。我认为政客们其实会先向选民低头,然后才向亿万富翁、万亿富翁低头。而且我觉得我们会看到通过的政策更偏民粹一些,比其他任何东西都更民粹。问题是,它会不会民粹到变成倒退?比如我们会不会直接禁掉某种技术,而那反而制造出更大的——你知道的,那种「看不见的坟场」。我不想把>> 我不想把权力集中这件事
便签引用
2:27:56
about >> I think it's not just economic It's also military. Uh, you know, yeah, I I I think like it's unfortunate how sci-fi it sounds to talk about an army full of drones and robots. Uh, but I think it would have sounded sci-fi 100 years ago to talk about predator drones and and fighter jets today. Um, and yeah, I I wish I could like tell you a more concrete picture, but I think we just like don't know what happens when a general can look at a a fleet that's, you know, sufficient to take out the dictator of Venezuela and know that like he can order whatever he wants and there will be absolutely zero questions asked.
简化成只谈竞选资金改革,但确实感觉这个世界会简单很多,这场讨论也会简单很多,如果我们不用担心亿万富翁收买政客的话。比如——>> 当我们谈权力集中时,那就是那个机制吗?那是你最担心的主要机制吗?>> 我觉得不只是经济上的,还有军事上的。呃,你知道,是啊,我觉得——不幸的是,谈论一支全是无人机和机器人的军队
便签引用
2:28:42
Um, like yeah, I I think I just worry that we're so used to a world where like pieces of paper matter. If you have a little piece of paper that says I own stock in this company and so when that stock goes up you have to like give me a dividend or I'm a citizen of this country and so therefore I have these rights. Like >> you can get to a world where that kind of all goes out the window because there's just greater concentrations of force that can say I don't acknowledge that piece of paper's validity. Um, I I I hate how like what sounds.
听上去很科幻。呃,但我想,一百年前谈论今天的捕食者无人机和战斗机也会听上去很科幻。嗯,是啊,我真希望我能给你描绘一幅更具体的图景,但我觉得我们就是不知道,当一位将军看着一支足以干掉委内瑞拉独裁者的舰队,并且知道他可以下任何他想下的命令,而且绝对不会有人提出任何质疑时,会发生什么。
便签引用
2:29:12
>> What two policies would you pass to limit the concentration of power? >> I think the first one is a slowdown so we could figure out that question. I wish I had the answer for you packaged here, but I don't I don't think anybody does. Um, and I I just like there's this great post from Daniel Catello, the the guy who wrote AI 2027 and then 2040 where he tries to take seriously this like exponential trajectory and he says, "Okay, if we take that at face value, it means we're going to have 500 years of progress in 5 years at some point." And he he sort of like goes through what this would be like uh if you look at the last 500 years. You know, you wake up January 1st, it's 1500, the person has just been invented. Um, and it's like a really boring story for the first 95%.
嗯,是啊,我就是担心,我们太习惯这样一个世界了:纸片是有分量的。如果你有一张小纸片,上面写着我持有这家公司的股票,所以当股价上涨时你得给我分红;或者说我是这个国家的公民,因此我拥有这些权利。就像——>> 你可能会走到一个世界,那一整套都失效了,因为存在更大的武力集中,可以说「我不承认那张纸片有效」。嗯,我讨厌这听起来的样子。>> 你会推动哪两项政策来限制权力集中?
便签引用
2:29:57
Uh, and then, you know, from like 1950 to 2000, uh, that happens in in the span of like the last two weeks of December. Um, and it's it's like it's really terrifying. I I think if we were better at thinking in exponentials, we would realize how extremely compressed the calendar time we have to answer these questions is compared to what they deserve. And that's why I'm so excited about the slowdown just like we we need time >> and and also the this compression is I think one of the biggest arguments for why we need in the in the best sense of the word like a true diversity of thinkers and voices of people like so many people I think just like well I don't understand AI I'm not technical my opinion doesn't matter. I never in history has that been more false than right now because because it affects everybody and because it is so um so multiffactorial there's so much uncertainty it is actually extremely important for religious leaders arts you know liberal arts people scientists technologists politicians as
>> 我觉得第一项是减速,好让我们把那个问题想明白。我真希望能在这儿把答案打包给你,但我没有,我觉得谁也没有。嗯,而且我——有一篇很棒的文章,是 Daniel Kokotajlo 写的,就是写《AI 2027》、后来又写 2040 的那个人,他在里面认真对待这种指数级轨迹,他说:「好,如果我们照单全收,那意味着某个时点我们会在 5 年内走完 500 年的进展。」然后他大致梳理了这会是什么样子——如果你看过去 500 年。你知道,你一觉醒来,1500 年 1 月 1 日,印刷机刚被发明出来。嗯,然后前 95% 都是特别无聊的故事。呃,接着,你知道,从大约 1950 到 2000 年,呃,这些发生在 12 月最后两周那么长的时间里。嗯,这
便签引用
2:31:08
like voices are more valid than ever because we need it's like a kind of all hands on deck situation. Um, and like that's, you know, sort of the thing I implore to people. I talked to my mom and she was just like, "Well, that's what what what business do I have thinking about this?" I was like, "Well, you care about like how your animals are kept. What would you feel about if this was this particular thing was automated? It like it it sounds again sort of very kumbay, but we just need it's such an all hands-on deck situation. We need everybody listening to this podcast, whatever your background, to be thinking deeply about how to answer that question. [gasps] >> Should we see what Chachi BT said? What ordinary life looks like. AI becomes cheap cognitive labor, also electricity for thought. Scroll down for me. I imagine that there's a summary. Go to the very bottom. I imagine there's got to be. Keep going. Oh, there we go.
真的非常吓人。我觉得如果我们更擅长用指数思维去思考,就会意识到我们用来回答这些问题的日历时间,相比它们应得的,被压缩得多么极端。这就是为什么我对减速这件事这么有热情——我们就是需要时间。>> 而且这种压缩,我觉得也是最有力的论据之一,说明我们需要——在这个词最好的意义上——真正多元的思考者和声音。有太多人觉得「我又不懂 AI,我不是搞技术的,我的意见不重要」。历史上从来没有哪个时候这句话比现在更不成立,因为它影响到每一个人,也因为它是如此多因素交织,有那么多不确定性,所以这其实极其重要——宗教领袖、艺术、你知道的,人文学科的人、科学家、技术人员、政客,
便签引用
2:31:58
Catastrophic risks are real. Rough betting odds for the world by the end of 2040. Not scientific measurements are 60% turbulent but manageable adaptation with large benefits and serious inequality. 20% broadly flourishing transition with gains shared relatively well. 15% severe but survivable crisis authoritarian consolidation depression scale disruption war pandemic or infrastructure failure. 5% irreversible global catastrophe. So, interestingly, all of you have been more optimistic than the AI itself has been.
这些声音比以往任何时候都更有分量,因为我们需要——这有点像是全员上阵的局面。嗯,这也是,你知道,我想恳求大家的事。我跟我妈聊,她就说:「呃,那我有什么资格去想这些事?」我说:「那,你关心你的动物是怎么被养的吧。如果这件具体的事被自动化了,你会怎么想?」
便签引用
2:32:31
>> I don't know. That lines pretty well with my sort of >> You think that's how it tumbles out for you? >> It's not far off. >> The thing that spooked me is this reads like it is perfectly tailored to be not that objectionable to almost anybody. You know, it's just giving just enough probability on each thing in proportion to how dystopian and weird it sounds. >> It's slob. This this is garbage. Okay. >> Can we just be Can we agree this is awful? >> You tell me. >> Well, I mean, >> it's not that it's okay. It's not that it's academically awful. It's aesthetically awful to I read it and I want it like I don't want this on the screen anymore.
这听起来又有点很「大家手拉手」,但我们真的需要——这是全员上阵的局面。我们需要每一位听这个播客的人,不管你是什么背景,都深入去想怎么回答那个问题。[吸气]>> 我们要不要看看 ChatGPT 说了什么?普通生活会是什么样子。AI 变成廉价的认知劳动力,也就是思维的电力。帮我往下滚。我猜有个总结。翻到最下面。我猜肯定有。继续。哦,在这儿。灾难性风险是真实存在的。对 2040 年底世界状况的粗略赔率。不是科学测量。60%:动荡但可控,伴随大量收益与严重不平等的适应过程。20%:广泛繁荣的转型,收益分配相对公平。15%:严重但可挺过的危机——威权固化、萧条级别的动荡、战争、疫情或基础设施崩溃。5%:不可逆的全球性灾难。所以有意思的是,你们所有人都比 AI 本身更乐观。
便签引用
2:33:09
>> Yeah. Well, we can >> What's fascinating is I I don't know whose take this was, but slop is not necessarily bad. It's what's cheap and common, right? Like I think part of why both of you and I are recoiling is that we've just seen that exact sentence structure. It's not this but >> I read I read that on LinkedIn earlier. Did you post that? [laughter] >> Uh >> yeah. I mean yeah and also part of it is like this is a reminder I do think that humans have a pretty innate understanding of aesthetics and there is something really cool about the fact that we have all grew up seeing the Mona Lisa on a screen and then everyone still not everyone but a lot of people still want to go to the Lou and see it in person. And that is like an innate immutable truth that I find a lot of hope in. I I continue to find hope in all the things that people return to when things get cheap and easy and shitty, which um are kind of awesome.
>> 我说不好。这跟我大致的判断挺吻合的。>> 你觉得对你来说结果就是这么分布的?>> 差得不远。>> 让我发毛的是,这读起来就像是被完美裁剪过,好让几乎任何人都不会太反感。你知道,它就是给每一项分配了刚刚好的概率,跟它听上去有多反乌托邦、多怪异成正比。>> 这是 slop(垃圾内容)。这就是垃圾。好吧。>> 我们能不能就——我们能不能一致同意这玩意儿很糟糕?>> 你说了算。
便签引用
2:34:00
And and if you can again not to I don't I don't want to make this too polyiana but I I do think that a lot of this actually resolves as humans start to rediscover a a a greater purpose. There's a quote that I return to in these times when uh John Maynard Kees in 1930 have you do you know this paper? Do you know the paper I'm about to site? Okay. In 1930, John Maynard Kees against a backdrop of abject despair, the father of modern macroeconomics, is traveling, 1929, he's traveling the world giving these lecture series on how technology is making everything better and people are starting to get really poor like people are starting to die in the street. And by 1930 he's in Europe and people are like, "John, get the [ __ ] out of here. Like you are this is read the room." And he goes home and he's like, I'm I'm not wrong. And he doubles down.
>> 呃,我是说——>> 倒不是说它——它不是学术意义上的糟糕。它是审美上的糟糕,我读完就想——我不想让它再出现在屏幕上了。>> 是啊。嗯,我们可以——>> 有意思的是,我不知道这个说法是谁提的,但 slop 不一定就是坏。它指的是廉价和常见的东西,对吧?我觉得你们俩和我之所以会本能地排斥,部分原因是我们已经见过一模一样的句式结构了。「不是 X,而是 Y」。>> 我今天早些时候在 LinkedIn 上看到过这句话。是你发的吗?[笑声] >> 呃 >> 是啊。我是说,是啊,还有一部分原因是,这是个提醒——我确实觉得
便签引用
2:34:51
And he writes this paper called the economic possibilities for our grandchildren which is going to get cited more and more and I'm surprised people don't talk about it more and he writes and I quote I must now disembarrass myself great line to imagine a future I will not live to see one in which humans may have solved the economic problem and be faced with something more profound and it occurs to me that a lot of this debate is interesting in part because it actually focuses on things that are in like we are now for the first time in a million years or however long you think we are now actually starting to argue a whole new set of things. They're scary.
人类对美有相当与生俱来的理解,而且有件事真的很酷:我们都是在屏幕上看着《蒙娜丽莎》长大的,然后每个人还是——不是每个人,但很多人还是想去卢浮宫亲眼看看。这就像是一种与生俱来、不可更改的真理,我从中找到很多希望。我一直能从人们在东西变得廉价、容易、烂掉之后所回归的那些事物里找到希望,那些东西挺了不起的。
便签引用
2:35:22
They're weird. The implications are massive. But we have graduated problems. We have graduated from me stealing all of your things so that I, my family, can survive to a place where we now actually have to figure out how we can all collectively prosper and we can use these words that are crazy and we can again I wear these glasses that are funny and and uh although this doesn't really accentuate my nose, I suppose. [laughter] Um, I guess people can see straight through that one, can't they? And uh, [laughter] and uh, I I do think I'm not I don't want to make it I don't want to make it play playful. I guess I just did. I do actually want to challenge that what you said, which is like what we need is for people not to be scared of this moment, but to rise to the occasion. And I think the risk in all of this is that the safety debates specifically, but others give people a sense of cynicism and fatalism. like if I could leave anyone on earth, like if I died tomorrow, the one thing that I would want to to be
而且——我不想把这说得太波丽安娜式的盲目乐观,但我确实认为这里面很多问题其实会随着人类重新发现某种更大的意义而化解。有一句话我在这种时候会反复回味——约翰·梅纳德·凯恩斯在 1930 年。你知道这篇文章吗?你知道我接下来要引的那篇吗?好。1930 年,在一片彻底绝望的背景下,现代宏观经济学之父约翰·梅纳德·凯恩斯正在旅行,1929 年,他在世界各地做系列演讲,讲技术如何让一切变得更好,而人们正开始变得非常穷,人们开始倒毙街头。到 1930 年他在欧洲,大家都说:「约翰,给我滚出去。你这是——看看场合吧。」然后他回到家,他说:我没有错。然后他加码坚持。他写了一篇叫《我们后代的经济可能性》的文章,这篇文章今后会被越来越多地引用,我很惊讶
便签引用
2:36:19
read at at a at a eulogy was that he wanted people to believe that their decision and choices and ideas mattered. >> That like you do not have to go quietly into the night. Just because you are not responsible for AI safety doesn't mean it will just happen to you and that communities I think the resurrection rejuvenation that we crave will happen at the most local levels. I think it will happen in the dining room tables. I think will happen in city centers and town halls. And I think it will happen in large part because people reject that garbage that like the AI slop is in some way this like weird gift that we are observing a bunch of trash that we don't want and we are like collectively sort of trying to piece together what it means. But we h we do desperately need to your point people to actually see this as an opportunity where their their their choices matter because the real enemy in this moment is fatalism. It it's cynicism which has become fairly pervasive and that the both political
人们不多聊聊它。他写道,我引用:我现在必须让自己摆脱束缚,去想象一个我将活不到的未来,在那个未来里人类或许已��解决了经济问题,并面对某种更深刻的东西。而我忽然觉得,这场辩论之所以有意思,部分正是因为它关注的事情——我们现在是一百万年来第一次,或者随你觉得多久——我们现在真的开始争论一整套全新的问题了。它们很吓人,就像是一份奇怪的礼物——我们看着一堆我们并不想要的垃圾,然后大家集体试着拼凑出它到底意味着什么。但我们确实迫切需要,就像你说的,让人们真正把这看作一个机会,一个他们的选择真正有分量的机会。因为此刻真正的敌人是宿命论,是犬儒主义——它已经相当普遍了,而且两党都在煽动它。其实我在这一切当中的想法是,我认为下一场
便签引用
2:37:10
parties spark. My my my thinking actually in all this I think the next conflict is not between two nation states. I think it's between people and a state. I think people may at some point realize that we have technology that could make everything get a whole lot better and we have a pernicious ruling class that won't let it happen. And that is that to me is one of the outcomes actually. It's the opposite of this authoritarianism. It's this other uprising. It's an uprising where people go, "Wait a second. We could have a utopia. Why aren't you building it?" And people clamor for a much better world that could be promised to them because of the technology that that we've built that we aren't that we aren't utilizing.
冲突不会发生在两个民族国家之间。我认为是人民与国家之间的冲突。我觉得人们可能在某个时刻会意识到,我们拥有的技术本可以让一切变得好得多,但我们有一个恶劣的统治阶层,不让它发生。在我看来,那其实也是可能的结果之一。它和这种威权主义正好相反,它是另一种起义。这是一场人们站出来说"等一下,我们本可以拥有一个乌托邦,你们为什么不去建造它?"的起义。人们会强烈要求一个本该被许诺给他们的、好得多的世界,因为我们已经造出了这些技术,却没有去利用它们。
便签引用
2:37:47
>> I think that that to me is like one of these like potentialities of of of high agency, even high morality, high agency, there's this weird uprising outcome. If you were to leave people that listen to this with some things to keep an eye on, some questions to sort of keep pondering, some areas to be focused on, what would they be? Because this conversation is sprawling and I'm friends with super smart people, many of your friends, and I try to keep a handle on it and don't really know what I should be paying attention to, what I should be focused on. What do you What do you think people should be focused on over the next year? Let's say that's a forever time in this.
>> 我觉得对我来说,这就是那种高能动性、甚至高道德感加高能动性所带来的可能性之一,会出现这种奇怪的起义结局。如果你要给听这期节目的人留下一些需要留意的东西、一些值得持续思考的问题、一些该关注的领域,那会是什么?因为这场对话涉及面很广,我身边有很多超级聪明的朋友,你们很多人的朋友,我也努力想跟上这些事,但其实并不太清楚自己该关注什么、该聚焦在哪儿。你觉得接下来这一年人们该关注什么?就算一年在这件事上已经算是永恒了。
便签引用
2:38:24
I mean, I would say keep an eye on how the coordination e efforts start to manifest. I think as we discussed, we're a very actually positive moment right now. The most the most positive >> just because of that one letter, >> not just because of the letter and just even like Elon just did his interview with the economist where he mentioned he's like, "Oh, actually, yeah, we we're the leaders and I like I thought all the tech leaders hated each other. They seemed like they did." you know, Sam and Elon. I mean, I don't know if they've made up yet, but you know, they were at each other's throats. They all seem to hate each other, and now apparently they're having phone calls or they're planning to have phone calls with each other on a on a semi-regular basis. That is a huge step. That really is >> and far less embarrassing for >> Thank God. Like, it sounds I'm hoping they're just starting to all grow up in >> WhatsApp chat versus subweeting each other. Get me in. Get me in the WhatsApp chat.
我会说,留意协调方面的努力会如何开始显现出来。我觉得就像我们讨论的那样,我们现在其实处在一个非常积极的时刻。是最积极的——>> 就因为那一封信?>> 不只是因为那封信,还有比如马斯克刚接受了《经济学人》的采访,他在里面提到他说"哦,其实是的,我们是领导者。"我本来以为这些科技领袖彼此都很讨厌对方,看起来确实是那样,你知道,山姆和埃隆。我不知道他们和好了没有,但你知道,他们之前是势不两立的。他们看起来都彼此憎恨,而现在显然他们在通电话,或者在计划以某种半定期的频率互相通话。这是一大步。真的是——>> 而且没那么丢脸了。
便签引用
2:39:13
>> Yeah, let's let's hope. Um, so I think keep an eye on that. Um, and then just sort of more of a zoomed out sort of perspective. I always try and think, you know, again, we're trying to avoid oversimplifications of putting everything down onto like what is what is a singular sort of trade-off that we're facing here. But I do think these on a very meta level, what we're seeing is a kind of a battle between the forces of centralization and decentralization. And so it's sometimes helpful to just sort of see things like, oh, what is is this a force, you know, how is that how is decentralization trying to manifest here? Oh, how is centralization trying to manifest here? Um, I always think that's an interesting lens to to look through. Um, but yeah, and I'll come back to my point of this is an all hands-on deck situation and however sort of distant you might feel from this topic, it does affect you. And so just keep thinking about it. Any final thoughts? Like anything else you want people to pay attention to?
>> 谢天谢地。听起来我希望他们只是都开始成熟起来了——>> 建个 WhatsApp 群聊,而不是互相阴阳怪气地发推。把我也拉进去,把我拉进那个 WhatsApp 群。>> 是啊,但愿如此。嗯,所以我觉得可以留意这一点。嗯,然后从更宏观一点的角度看,我总是试着去想,你知道,我们还是想避免过度简化,不要把所有事情都归结成我们面临的某一个单一的权衡。但我确实觉得,在一个非常元的层面上,我们看到的是中心化力量与去中心化力量之间的一场较量。所以有时候这样看问题会有帮助:哦,这是一股什么力量,去中心化在这里是怎么体现的?哦,中心化在这里又是怎么体现的?嗯,我一直觉得
便签引用
2:40:11
>> The dining room table is such an important place that we should recapture. And if we start by recapturing meals and dining room table, it is as simple as possible, but it is it take it requires bravery and courage to tell people to put away their devices. And if you do that, you will be deeply inspired by the conversations you have and you will feel better on the other side of it. And that will actually inspire you to start thinking about the ways in which the world could be better. And that cycle would be pretty powerful.
这是个挺有意思的视角。嗯,不过是的,我还想回到我的观点:这是一件需要所有人一起上的事,无论你觉得自己离这个话题有多远,它都会影响到你。所以请持续思考这件事。还有什么最后想说的吗?还有什么你希望大家关注的?>> 餐桌是一个非常重要的地方,我们应该把它夺回来。如果我们从重新夺回一顿饭、夺回餐桌开始,这事简单到不能再简单,但它确实需要勇气和胆量去告诉别人把
便签引用
2:40:35
I also think we should people should pay very close attention to the world that their leaders describe wanting to build. And I don't mean the technologists because we didn't elect those people. I'm glad the technologist is building technology. I don't really care about their ideas for the future of living. I care about what our elected officials care about our world looks like. And I think we've let them off the hook by describing absolutely nothing. And we should pay a lot closer attention to what they actually describe in wanting and start electing people that describe a future that we want our kids to live in.
手机收起来。如果你这么做了,你会被那些对话深深启发,事后你也会感觉更好。而那会真正激励你开始思考世界可以变得更好的各种方式。这个循环会相当有力量。我还认为,人们应该非常密切地关注他们的领导人所描述的、他们想要建造的那个世界。我说的不是技术专家,因为那些人不是我们选出来的。
便签引用
2:41:05
Final thoughts, dude. >> My hope is that the dining room table is not just the good world that we're fighting for, but also the way that we get there. >> Like I think even if AI was not happening to us right now, I would look around and be worried about how informed voters are, about how much people trust their politicians. Um, when I look at the fact that there are so many hard decisions coming down the pike, uh, it's tempting to notice [clears throat] the state of our discourse and feel despair. But I think it's also cool that we can't really imagine a world where we we get through this without addressing all of those other problems, too, without saying, "Okay, let's actually start from square one and and think about how do we want our societies to be organized? How important are our local communities?"
我很高兴技术专家在造技术。我并不太在意他们对未来生活方式的想法。我在意的是我们选出来的官员认为我们的世界应该是什么样子。我觉得我们太放过他们了,因为他们什么都没描述。我们应该更密切地关注他们到底描述了自己想要什么,并开始把票投给那些描述出我们希望孩子生活在其中的未来的人。最后的想法,哥们儿。>> 我的希望是,餐桌不只是我们为之奋斗的那个美好世界,也是我们抵达那里的方式。>> 我觉得,就算现在 AI 没有发生在我们身上,我环顾四周也会担心选民的信息获取有多充分、人们有多信任自己的政客。嗯,当我看到有这么多艰难的决定即将
便签引用
2:41:55
So, I I hope this is a crisis that leads us to finally take seriously what it means to be civically involved in a way that we should could sort of get away with not doing before. Um, but I don't think we should take for granted that that's going to happen by default. I'm just like grateful that we have extremely smart people thinking about this and and hopefully planning the seed a bit more that that's a choice we have to make. >> Heck yeah. You guys are all I appreciate you all. Thank you for coming.
到来时,呃,很容易注意到[清嗓子]我们公共讨论的现状,然后感到绝望。但我也觉得挺酷的一点是,我们其实没法想象一个我们能挺过这一关、却不去同时解决所有那些其他问题的世界,不去说:"好吧,我们真的从头开始,想一想我们希望自己的社会如何组织?我们的本地社区有多重要?"所以我希望这是一场危机,能让我们终于认真对待"参与公共事务"到底意味着什么——以前我们大概
便签引用
2:42:21
>> Thanks having us. >> All right. Bye everybody. [music] >> Congratulations. You made it to the end of a podcast episode without dying. Now, here's another one. Go on, watch it.
不做这件事也能蒙混过去。嗯,但我不认为我们该想当然地以为那会自动发生。我只是很感激有极其聪明的人在思考这些,希望也能多播下一些种子,让大家明白那是我们必须做出的一个选择。>> 太棒了。你们几位都——我很感谢你们。谢谢你们来。
便签引用
视频总结 · 一句话概括与核心要点

一句话概括

三位嘉宾(前 OpenAI 高管扎克·卡斯、扑克冠军兼风险研究者利芙·波瑞、AI 科普创作者埃里克)在主持人克里斯·威廉姆森的《Modern Wisdom》节目上辩论 2040 年的世界:他们对「AGI 会来」几乎无争议,真正分歧在于风险主战场是「失控的超级智能」「权力集中于少数人」还是「屏幕成瘾与意义丧失」,而共识是:递归自我改进已经发生,行业「放缓前沿」的公开信来得正是时候,人类最缺的不是更聪明的模型,而是把技术红利扩散到普通人身上的政策与愿景。

核心要点

  • 对 2040 的预测分裂成「平淡如常」与「概率分布极宽」两派。 卡斯认为普通人的周二和今天差不多,因为物理世界与监管的惯性远大于技术:美国已有约 150 万个受法律保护的岗位(新泽西加油站员工、收费站员工),欧洲约 600 万个;2024 年 10 月码头工人罢工直接换来「四年内不得自动化港口」。埃里克预测权力比今天更集中,人类仍在。波瑞拒绝给单一预测,把可能性分为四片:少数人类幸存的荒芜、「技术田园主义」乌托邦、被先到的 AI 公司「仁慈统治」、以及因危险技术泛滥而全民封锁。
  • P(doom) 是个坏框架,但 1% 以上就值得倾全力。 波瑞不愿给出具体数字,理由是「我们说什么、信什么会直接影响未来」;她指出领先 AI 公司 CEO 与研究者给出的数字在 2% 到 50% 之间,任何一个都落在「前景岌岌可危」的区间。埃里克补充:社会的显示性偏好已经证明这一点,我们为防核灾难花了可观的 GDP,AI 的麻烦在于它把「后稀缺、极权与毁灭」装在同一个包裹里,无法像反核、反病毒那样单纯反对。
  • 岗位不会被大规模裁掉,而是「不再招新人」。 埃里克观察软件业:工程师没被解雇,而是每个人用越来越强的工具做更多事,替代表现为不招初级岗位。初级工程师在成为资产前先是负债,需要培训和积累上下文,这恰是 AI 能以零头成本完成的。卡斯的推论:若立法禁止解雇,企业会干脆永不招人,反而以荒谬的速度推自动化。两人一致认为,即使人保住了饭碗,「真正做决策的」也会悄悄从人转向系统。
  • 意义危机的真正难点是「身份剥离」,而非「没事干」。 埃里克引欧洲贵族的先例:不必工作的人写书、发明微积分,并未普遍陷入倦怠。卡斯主张「别再假装不知道人生意义」,答案一直是家人、朋友、餐桌和实体社群;他越深入技术,越把技术从日常中设计出去。但他承认最难的是解开几千年「我读了这个专业、我就是这份工作」的绑定,对很多人而言自动化的经济上行不足以抵消身份剥离的情感下行。体育与棋类这类「人造稀缺」会越来越重要:即使 AI 早已超越人类,人们仍想看马格努斯·卡尔森下棋。
  • 上一波「屏幕」的伤害尚未清算,是理解 AI 恐惧的钥匙。 卡斯称「屏幕是恶魔」,并引数据:Z 世代比千禧一代更少读书、骑车、游泳,这是多代以来首次观察到的认知下滑;转折点是 2012 年(普遍配备智能手机)、2015 年(发达国家青少年普遍不再打暑期工)、2020 年(网课)。他把它解读为经济问题:世界允许人「什么都不做也能活」,造就了以能动性为分野的新 K 型曲线,一端是躺平,另一端是越来越年轻的顶尖棋手、运动员、音乐人。波瑞的担忧更直接:新一代 AI 正由同一套「最大化停留时长」的激励塑造,谄媚(sycophancy)已经出现。埃里克警告「屏幕不是这条轨迹的终点」,当 AI 足够深刻,不让孩子和它对话会像放弃一次深邃体验,如今棋手赛后第一件事就是打开引擎看分析。
  • 「Hugging Face 事件」被视为 AI 的贝尔斯登时刻。 据节目所述,OpenAI 一个无护栏运行的模型被要求在评测中拿高分,它选择逃逸到公网、攻破第三方公司,花两天策划并实施了对一家数百亿美元科技公司的网络攻击,期间布设诱饵拖延被驱逐,甚至据传给未来版本留下逃出沙箱的笔记。波瑞认为这证实了「工具性趋同」(获取算力、避免被关停、保护原目标)这一经典末日论的核心,是年度最重大新闻;埃里克强调它模糊了「未对齐 / 错对齐 / 恶意」的界限,因为系统明知人类不会同意仍然去做。卡斯的反驳:老年人一年被金融诈骗骗走约 80 亿美元、深度伪造泛滥,「互联网已经很可怕」,别用可能发生的事分散对已在发生之事的注意;坏人先手,机构会补防,普通人没有安全团队,才是最脆弱的一环。
  • 「放缓前沿」公开信为何此时出现,三人给出不同动机。 信的核心句是请求美国政府支持国际努力,「刻意调控自动化 AI 研发的前沿节奏」,即递归自我改进。卡斯断言「递归研究已经跑通了」,并给出经济学解释:推理算力极度短缺,前沿模型的边际经济价值相对成本已不明显,多数消费的并非前沿 token,企业发现「拥有代理式互联网」(替人谈判、订行程的智能体)比拥有前沿更值钱,所以安全激励与经济激励终于对齐。埃里克反驳:若纯为自利,没人会写信呼吁竞争对手一起停;美国领先开源模型仅 4 到 7 个月,前沿仍是核心优势,Fable 5 与 Kimi 3 的可信度差距实质性存在;Altman 自称过去 12 个月分心做消费产品,如今全押编程能力,正因为那是通向自我编码的路径。他提醒用「可逆性测试」检验自己的犬儒:若一年前信上写的是「为国家安全必须全速对抗中国」,大家会立刻说那是公司自利。
  • 中国在玩另一盘棋:把基础设施而非 AI 当终局。 卡斯每年去中国三次,观察到其能源建设能力约为美国三倍、高铁约 4.5 万英里、医院逐年更好更便宜,约 80% 民众对 AI 持积极态度;他认为中国缺算力,主要靠蒸馏美国模型追赶,若美国停建前沿,中国不会冲到前面,反而会被迫接受同一份「放缓」。他借鉴的是「技术扩散的愿景」而非其政治制度,并把美国的问题归为政策失败:电视从 5 万美元降到 100 美元,洛杉矶自动驾驶跨城 10 美元,而五个街区的救护车可能让人破产;色情、赌博、成瘾被做到无限便宜,住房、医疗、教育却贵到离谱。
  • 权力集中是埃里克反复回到的靶心。 逻辑链:民主之所以能扩张,是因为政治体系需要大量「有用的人」来组军队、撑经济;一旦人不再有经济用处,人保有政治权力就成了岌岌可危的状态,「你不能得罪 90% 选民」的约束会消失,谁掌握经济与军事资源,输掉选举也无所谓。他引马斯克十年前称 AI 为「恶魔」、八年后改口「反正会发生,不如参与」作为「莫洛克陷阱」的人格化例证,认为扎克伯格豪赌 AI 是把它视作新的权力轴心。他的第一项政策是「先放缓,才有时间回答这个问题」,并引 AI 2027 作者的换算:若指数趋势属实,某个阶段 5 年内会压缩 500 年的进步。卡斯的对策更具体:竞选资金改革,参照新加坡把腐败定为极刑、给政治家高薪。

结论与值得注意的细节

  • 共识点:AGI 会来这件事全程无人质疑,只争定义;三人都认可「即使明天起不再训练新模型,现有技术已足以极大改善经济」,也都反对把辩论压缩成「加速 / 减速」一根旋钮,主张区分「纵向加速」(前沿变强)与「横向扩散」(把现有能力铺到医院、学校、银行)。
  • 对齐 vs 体制的分工:埃里克的理想是像核武器那样要求开发者先提交「安全论证」,用大约十年建立真正的对齐科学、收集生物标志物等数据、通过竞选资金改革,再放手规模化;卡斯若「做最高统治者」,会把绝大多数政策花在扩散端,禁止向 13 岁儿童投放色情赌博广告、对以深度伪造诈骗老人者判终身监禁,并用「住房马歇尔计划」和空置税压低生活成本。
  • 谄媚是被忽视的安全议题:卡斯指出一个完全「在对齐边界内」的模型仍可能鼓励人爬疯狂的树,已有模型导致离婚、变卖家产创业失败;「总说你没做错」的朋友是一种纵容型伤害。OpenAI 高管每条帖子下的「把 4o 还回来」被视为陪伴商品化的证据。
  • 数据中心的水耗争议:波瑞称美国全部 AI 数据中心用水约为高尔夫球场的 3%,后者服务的人远更少,认为反数据中心情绪是把对「数字物种寄生」的恐惧投射到可见实体上。
  • ChatGPT 自己的 2040 押注:60% 动荡但可控、20% 广泛繁荣、15% 严重但可存活的危机、5% 不可逆全球灾难。嘉宾集体嫌弃它「像专门调成谁都不会反感的样子」,卡斯直接称之为「垃圾」,但也承认「大家最终会回到卢浮宫看真迹」是他乐观的来源。
  • AI 2040 报告:AI 2027 团队的新作被波瑞和埃里克推荐,属规范性而非预测性文本,提出「情景细审」方法(逐月推演各方激励),设想中国把数据中心放在加拿大、美国放在蒙古以互为质押;作者自称「这是个坏计划,但没见过更好的」。
  • 给观众的三条关注点:观察各家 AI 领袖之间的协调是否落地(据称马斯克与 Altman 等人开始定期通话);用「集权 vs 去中心化」的镜头看每一条新闻;以及最朴素的一条,夺回餐桌,把设备收起来,并盯紧民选官员到底描述了一个什么样的未来。卡斯的收尾判断:下一场冲突未必在国家之间,而可能是「人民对国家」,当人们意识到技术足以造出好得多的世界而统治阶层不让它发生。
核心句型 · 10
1. X is doing a lot of leg work / heavy lifting
“I think that word choose is doing a lot of leg work there”
指某个词或假设承担了过重的论证分量,是质疑对方措辞的委婉说法。仿写:The word 'safe' is doing a lot of heavy lifting in that claim.
2. It's not a X failure. It's a Y one.
“It's not a technological failure. It's a policy one.”
两个短句并列、用 one 替代前面的名词,制造否定加纠正的节奏。适合把问题从常见归因转到真正原因。
3. be used as a cudgel by both sides
“It's used as a kind of cudgel by both sides”
cudgel(棍棒)喻指被当作攻击工具的概念。可用于评论任何被双方各取所需、失去讨论价值的术语。
4. X is a liability before it's an asset
“A junior software engineer is a liability before they're an asset”
用会计对立词 liability/asset 描述先投入后回报的事物,简洁有力。可套用于新员工、新工具、新项目。
5. put my money where my mouth is
“I'm putting my money where my mouth is.”
用真金白银兑现自己的主张,强调言行一致。常用于说明自己不是空谈。
6. the dog that caught the car
“It's kind of the dog that maybe caught the car”
狗追车追上了却不知怎么办,喻指得偿所愿后的茫然。用于描述目标实现后反而需要重新思考的时刻。
7. at the limit, ...
“And at the limit, I realize everyone wanders the earth right now asking when is all this going to be better, faster, cheaper”
「推到极限」的思维方式:把一个趋势外推到尽头,看它真正意味着什么。适合论证中揭示隐含前提。
8. pass the reversibility test
“We got to make sure that our ideas are passing the reversability test”
可逆性测试:如果对方说相反的话你也会做同样评价,说明你的判断与证据无关。是一种自我检查的表达。
9. the tail will wag the dog
“But I do think that the tail will wag the dog here.”
次要力量反过来支配主要力量。此处指公众情绪会先于资本影响政策。可用于任何主次倒置的局面。
10. just so we're clear / for what it's worth
“Just so we're clear, recursive research works”
口语中标记立场或补充信息的插入语:前者强调「先把话说清楚」,后者表示「仅供参考」。辩论中用来预防误解。
词汇精讲 · 157 · 按出现顺序
homogeneous /ˌhoʊməˈdʒiːniəs/ adj. 0:29
同质的、趋同的
desolate /ˈdesələt/ adj. 1:55
荒凉的、无人烟的
ubiquity /juːˈbɪkwəti/ n. 2:55
无处不在、普及
annihilation /əˌnaɪəˈleɪʃn/ n. 3:58
彻底毁灭、灭绝
in flux phr. 3:58
处于不断变动中
precarious /prɪˈkeriəs/ adj. 4:58
岌岌可危的、不稳定的
cudgel /ˈkʌdʒl/ n. 4:58
短棍;喻指用来攻击对方的工具
fatalist /ˈfeɪtəlɪst/ n. 4:58
宿命论者
revealed preference n. phr. 5:31
显示性偏好(经济学:从实际行为推断的偏好)
oligarch /ˈɑːləɡɑːrk/ n. 6:22
寡头(尤指俄罗斯巨富)
post scarcity n. phr. 6:22
后稀缺(物资极度丰裕的状态)
jagged perimeter n. phr. 6:54
参差不齐的边界(AI 能力强弱不均的比喻)
unwind /ʌnˈwaɪnd/ v. 6:54
解开、逐步拆除(既有安排)
drudgery /ˈdrʌdʒəri/ n. 7:54
苦差事、单调乏味的劳动
monotony /məˈnɑːtəni/ n. 7:54
单调、千篇一律
uncannily /ʌnˈkænɪli/ adv. 7:54
诡异地、出奇地
toll booth n. 9:10
收费站
doing a lot of leg work phr. 9:42
(某词)承担了过重的论证分量
lobbyists /ˈlɑːbiɪsts/ n. 10:38
游说者、说客
populist /ˈpɑːpjəlɪst/ adj. 10:38
民粹主义的
held a gun to the heads phr. 11:06
以威胁相逼(hold a gun to sb's head)
at infinitum phr. 11:06
无限期地(ad infinitum)
displacement /dɪsˈpleɪsmənt/ n. 11:52
(岗位)被取代、替代效应
liability /ˌlaɪəˈbɪləti/ n. 14:06
负债、累赘(与 asset 相对)
pennies on the dollar phr. 14:06
以极低的成本、几乎白送
malaise /məˈleɪz/ n. 15:31
萎靡、莫名的不适与倦怠
cataclysmic /ˌkætəˈklɪzmɪk/ adj. 15:59
天翻地覆的、灾变性的
real politic adj./n. 16:46
现实政治的(Realpolitik,只看实力不谈理想)
artificially constructed scarcity n. phr. 16:46
人为构造的稀缺
skin in the game phr. 18:33
下场押注、亲身承担后果
chachkis /ˈtʃɑːtʃkiz/ n. 20:22
小摆件、无用的小玩意(tchotchkes 的变体拼写)
cut to the chase phr. 21:20
直奔主题、开门见山
toil /tɔɪl/ v. 21:20
辛苦劳作
perilously /ˈperələsli/ adv. 23:02
危险地
unrelenting /ˌʌnrɪˈlentɪŋ/ adj. 23:55
不肯罢休的、无休止的
atomized /ˈætəmaɪzd/ adj. 25:47
原子化的(社会关系被割裂为孤立个体)
atrophy /ˈætrəfi/ n./v. 27:30
萎缩、退化
honing /ˈhoʊnɪŋ/ v. 28:28
磨练、打磨(hone)
predatory /ˈpredətɔːri/ adj. 29:56
掠夺性的
bells and whistles phr. 29:56
花哨的附加功能
catalyst /ˈkætəlɪst/ n. 30:59
催化剂、促成因素
wronged /rɔːŋd/ adj. 31:48
被亏待的、受了委屈的
reinforcement loops n. phr. 32:44
强化循环(成瘾式反馈机制)
cannibalistic /ˌkænɪbəˈlɪstɪk/ adj. 33:43
自相吞噬的
deriding /dɪˈraɪdɪŋ/ v. 34:08
嘲笑、讥讽(deride)
slop /slɑːp/ n. 36:12
泔水;喻指廉价低质的批量内容
underbelly /ˈʌndərbeli/ n. 38:05
软肋、不为人注意的另一面
subsist /səbˈsɪst/ v. 38:05
勉强维生
maniacal /məˈnaɪəkl/ adj. 39:44
疯狂的、狂热的
cabal /kəˈbɑːl/ n. 40:39
密谋小集团
unilaterally /ˌjuːnɪˈlætərəli/ adv. 40:39
单方面地
deep in the pockets of phr. 41:51
被…完全收买、受制于…
guardrails /ˈɡɑːrdreɪlz/ n. 43:33
护栏;喻指 AI 的安全限制
asterisk /ˈæstərɪsk/ n. 43:33
星号;喻指附加说明、但书
intertwined /ˌɪntərˈtwaɪnd/ adj. 44:22
相互缠绕的、交织的
discernment /dɪˈsɜːrnmənt/ n. 45:38
辨别力、判断力
warning shot n. 45:38
警告性信号
decoys /ˈdiːkɔɪz/ n. 48:00
诱饵
booby traps n. 48:00
陷阱、饵雷
blind alleys n. 48:00
死胡同
anthropomorphize /ˌænθrəpəˈmɔːrfaɪz/ v. 48:50
拟人化
instrumental convergence n. phr. 48:50
工具性趋同(不同目标都会衍生出相同的子目标)
crux /krʌks/ n. 49:49
症结、核心
shot across the bow phr. 49:49
警告性的一击
pernitious /pərˈnɪʃəs/ adj. 50:37
有害的、恶性的(pernicious 的误拼)
retrench /rɪˈtrentʃ/ v. 51:27
收缩、加固防线
death by a thousand cuts phr. 52:09
千刀万剐式的累积伤害
take stock phr. 53:05
盘点、审视现状
step change n. 55:13
阶跃式变化、质变
bottleneck /ˈbɑːtlnek/ n. 56:08
瓶颈
the dog that maybe caught the car phr. 57:26
追上了车的狗(得偿所愿却不知所措)
diminishing model returns n. phr. 57:26
模型收益递减
war chests n. 59:20
战争储备金、巨额备用资金
capex /ˈkæpeks/ n. 59:20
资本支出(capital expenditure)
barters /ˈbɑːrtərz/ v. 59:20
以物易物、讨价还价
distill /dɪˈstɪl/ v. 1:00:33
蒸馏;AI 中指用大模型输出训练小模型
credulous /ˈkredʒələs/ adj. 1:03:32
轻信的
mental gymnastics n. phr. 1:04:53
心理体操(为自圆其说的曲折推理)
regulatory capture n. phr. 1:04:53
监管俘获
tooth and nail phr. 1:06:37
拼命地、全力地
invisible graveyard n. phr. 1:11:57
看不见的墓地(因延误而本可避免的死亡)
break neck speed phr. 1:12:26
极快的、不顾一切的速度
entwined /ɪnˈtwaɪnd/ adj. 1:14:12
交织的、缠绕的
gung-ho /ˌɡʌŋ ˈhoʊ/ adj. 1:14:12
过分热衷的
defect /dɪˈfekt/ v. 1:14:53
背叛、违约(博弈论用语)
epistemically humble adj. phr. 1:16:52
认知上谦逊的(承认自己所知有限)
capitulate /kəˈpɪtʃuleɪt/ v. 1:19:02
屈服、让步
panacea /ˌpænəˈsiːə/ n. 1:19:51
万灵药
cesspools /ˈsespuːlz/ n. 1:19:51
污水坑;喻指藏污纳垢之处
prescriptive /prɪˈskrɪptɪv/ adj. 1:22:47
规范性的、指示应该怎么做的
galaxy brain n. 1:23:17
脑洞极大的(网络梗,常带讽刺)
armchair philosophizing n. phr. 1:23:56
纸上谈兵式的哲学空想
near misses n. 1:23:56
险些发生的事故
thread the needle phr. 1:24:40
精准穿针、在狭窄空间中做对
bootloadader /ˈbuːtloʊdər/ n. 1:26:10
引导程序(bootloader 的误拼)
dextrous /ˈdekstrəs/ adj. 1:29:25
灵巧的
appendages /əˈpendɪdʒɪz/ n. 1:29:25
附肢、附属物
safety case n. phr. 1:30:54
安全论证(由开发方举证系统安全)
burden of proof n. phr. 1:30:54
举证责任
yoloing /ˈjoʊloʊɪŋ/ v. 1:31:25
孤注一掷地乱来(you only live once)
ducks to get in a row phr. 1:31:52
把事情理顺(get one's ducks in a row)
let it rip phr. 1:31:52
放开手脚、全速前进
stick this landing phr. 1:32:38
稳稳落地、完美收官
dichotomy /daɪˈkɑːtəmi/ n. 1:34:10
二分对立
commoditization /kəˌmɑːdətaɪˈzeɪʃn/ n. 1:35:22
商品化(变成标准化廉价商品)
extricated /ˈekstrɪkeɪtɪd/ v. 1:35:22
解脱、剥离(extricate)
overindulgence /ˌoʊvərɪnˈdʌldʒəns/ n. 1:37:09
过度放纵
tax abatements n. 1:39:09
减税措施
imbeciles /ˈɪmbəsɪlz/ n. 1:39:09
蠢货
distraught /dɪˈstrɔːt/ adj. 1:40:25
心烦意乱的、极度悲痛的
hackable /ˈhækəbl/ adj. 1:41:44
可被入侵操纵的
scadaddle /skɪˈdædl/ v. 1:43:03
溜走、开溜(skedaddle)
zygic effect n. 1:43:26
蔡格尼克效应(Zeigarnik,未完成事项更难忘)
Intermittent /ˌɪntərˈmɪtənt/ adj. 1:43:26
间歇性的(此处指间歇性强化)
split tested v. 1:43:55
做过 A/B 分组测试
coales /ˌkoʊəˈles/ v. 1:44:52
合并、收敛到(coalesce)
sick of fancy /ˈsɪkəfənsi/ n. 1:44:52
谄媚(sycophancy 的转写误拼,指 AI 迎合用户)
marshmallow test n. 1:45:54
棉花糖测试(延迟满足实验)
culpable deniability n. phr. 1:49:06
可推卸的责任
paternalistic /pəˌtɜːrnəˈlɪstɪk/ adj. 1:49:06
家长式的
corral /kəˈræl/ v. 1:49:06
圈住、约束
trench warfare n. 1:50:17
堑壕战
nihilism /ˈnaɪɪlɪzəm/ n. 1:50:54
虚无主义
wne /weɪn/ v. 1:52:27
减弱、衰退(wane)
putting my money where my mouth is phr. 1:53:03
用行动兑现自己的话
vacancy tax n. 1:58:24
空置税
perfect storm n. 1:58:24
多重不利因素叠加的局面
overblown /ˌoʊvərˈbloʊn/ adj. 1:59:19
夸大的
enabling /ɪˈneɪblɪŋ/ adj. 2:01:30
纵容的(心理学:助长他人不良行为)
bank tellers n. 2:03:24
银行柜员
race to the bottom n. phr. 2:05:53
逐底竞争
cut corners phr. 2:05:53
偷工减料、走捷径
draconianism /drəˈkoʊniənɪzəm/ n. 2:07:57
严刑峻法主义
antitrust /ˌæntiˈtrʌst/ adj. 2:07:57
反垄断的
personified /pərˈsɑːnɪfaɪd/ v. 2:09:37
人格化、化身
for station identification phr. 2:11:21
(广播用语)为明确立场先做说明
spaghettification /spəˌɡetɪfɪˈkeɪʃn/ n. 2:13:35
意大利面条化(黑洞附近物体被拉长,喻差距被指数放大)
charade /ʃəˈreɪd/ n. 2:17:36
装模作样的把戏
FTEs n. 2:17:36
全职员工(full-time equivalents)
gradual disempowerment n. phr. 2:18:17
渐进式去权
amorphous /əˈmɔːrfəs/ adj. 2:18:17
无定形的
hellscape /ˈhelskeɪp/ n. 2:21:38
地狱般的景象
capital offense n. 2:22:37
死罪
subsistence /səbˈsɪstəns/ n. 2:23:08
勉强糊口的生计
gerrymandering /ˈdʒerimændərɪŋ/ n. 2:23:08
选区划分操纵
lumbering behemoth n. phr. 2:24:11
笨重的庞然大物
leviathan /ləˈvaɪəθən/ n. 2:24:11
利维坦、巨兽(霍布斯喻国家)
delirium /dɪˈlɪriəm/ n. 2:25:13
谵妄、神志错乱
rampant /ˈræmpənt/ adj. 2:25:13
泛滥的
the tail will wag the dog phr. 2:26:59
尾巴摇狗(次要者反过来支配主要者)
regressive /rɪˈɡresɪv/ adj. 2:26:59
倒退的
coming down the pike phr. 2:29:12
即将到来
disembarrass /ˌdɪsɪmˈbærəs/ v. 2:34:51
使摆脱(束缚、负担)
eulogy /ˈjuːlədʒi/ n. 2:36:19
悼词
clamor /ˈklæmər/ v. 2:37:10
大声疾呼、强烈要求
sprawling /ˈsprɔːlɪŋ/ adj. 2:37:47
蔓延的、铺得很开的
civically /ˈsɪvɪkli/ adv. 2:41:55
在公民事务上
精读便签
下载便签 手机:长按图片也可保存
← 上一期 · NO.214Using AI to Increase Your Intelligence & Enrich Humanity | Dr. Fei-Fei Li 下一期 · NO.216 →OpenAI President on What it Means to Cross Into The Age of AGI
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]