播客·中文精读

西奥·冯与马斯克聊 AI、MMA 比赛、应对广告商……(AI 生成播客)

Theo Von & Elon Musk Talk About AI, MMA Fight, Dealing With Advertisers...(AI Generated Podcast)

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

视频信息栏底部有一行小字:AI Generated Podcast。频道名写着 Comedy Gold。主持人自称西奥·冯(Theo Von),开口第一句把节目名字念成了“this AI weekend podcast”。

真节目叫《This Past Weekend》。改掉一个词,一档真实存在的喜剧访谈就成了一件不存在的东西。

这份文字稿里,马斯克说了这么一句:

“我本来想在斗兽场打,结果他受伤了。我觉得没打成是好事,对他来说结果不会好看,我块头更大,人类活动的每个方面我都比他强。”(马斯克,据本期文字稿)

他没说过这句话。公开记录里找不到这两个人的任何对谈痕迹,西奥·冯的节目列表里没有这一期,马斯克的行程里也没有。

假的只有归属。假马斯克嘴里的每一句话,都能在真马斯克的某个采访里找到出处。假稿子编不出来的东西是细节,编得出来的东西是立场。

这份文本值得读的地方就在这儿。它是一份 2024 年的能力样本,说明模型能把一个公众人物的观点分布复刻得多准确,同时能把他的说话质感丢得多干净。

破绽从节目名开始

模型写对话有个通病,每一句都被接住。

冯抛一个梗,马斯克接住,回一个更夸张的版本。冯再抛一个,马斯克再回一个。二十多分钟,零冷场,零打断,零沉默。没有人说“你刚才说什么”,没有人在不重要的地方纠缠十分钟然后拉不回来,没有人把主持人准备好的笑点晾在那儿。

真实访谈不长这样。西奥·冯的本事恰恰是让人不舒服。他会突然安静,会追问一个太私人的问题,会在对方讲完段子后不给台阶。这些在文本里表现为“缺陷”,模型学不来,因为它的训练目标就是消除缺陷。

还有一个更硬的信号。整篇稿子里,除了扎克伯格和本杰明·富兰克林,没有出现任何一个具体的人名、地名、日期、数字。

对比真马斯克。他接受采访随口就抛量级:星舰的目标是把入轨成本压到每公斤一百美元以下,星链在轨卫星几千颗,某个工厂的产能爬坡曲线长什么样。数字密度是他的语言指纹,也是伪造的门槛。观点可以猜,数字不能编。

这份稿子的内容全是“AI 会不会损人”“你要是有时光机会去哪”这类没有成本的问题。有人把访谈分两类:有追问的,和没有追问的。这份稿子属于第三类,连追问的形状都是假的。

假马斯克说的最真的一句话

全篇最接近真实马斯克的一句,藏在一堆笑话中间:

“随着 AI 往前走,我们必须确保它们理解共情和人类价值观。我们不希望它们变成终极网络霸凌者。”(马斯克,据本期文字稿)

方向是对的,只是被写得太软了。

真马斯克在这个题目上说了十年,用词比这狠得多。2014 年他在麻省理工学院说 AI 是“在召唤恶魔”。2017 年他对着美国州长们说,AI 是文明面临的最大风险,呼吁在监管到位前先立规矩。2023 年 3 月,他签了那封呼吁暂停训练比 GPT-4 更强模型六个月的公开信。四个月后,他自己成立了 xAI。

这个前后矛盾被人骂了很多年。他的解释一直没变:如果这个方向注定有人要走,旁观不解决问题,得参与进去影响方向。这句话真假难辨,但它是他自己的说法,不是别人替他圆的。

他关于风险的核心论点,跟“机器人变坏”没什么关系。他说危险来自目标函数和人类真实意愿的错位。优化过程会精确地执行你写下的目标,你脑子里想的那个目标它管不着。让一个足够强的系统去最大化某个指标,它会找到设计者完全没想到的路径,过程中不带任何恶意。

所以他把注意力放在“AI 应该追求什么”上。xAI 成立时他讲的目标是“最大程度好奇”,让模型以理解宇宙为乐,别被某个具体指标牵着走。这个设定的潜台词是:一个只想搞明白世界的智能,不太有必要把人类怎么样。

这套逻辑站不站得住另说,但它的位置在这份假稿子里被保留了,只是被稀释成了“要教它们讲笑话”。

笑话是对齐最难的一道题

假稿子里有大量机器人损人的段子。Siri 嘲笑你的发型,机器人回一句“不错的反击,人类,你用那个史前大脑想出来的吗”。这些段子不好笑,但它们指的方向比作者意识到的更接近真问题。

幽默为什么难?它同时依赖三层信息:字面上说了什么,什么语境下说的,说的人和听的人什么关系。冒犯的边界还在实时移动,同一个句子在 A 群是调侃,在 B 群是羞辱,在第三个人耳朵里是证据。

模型在幽默上的失败有两种,方向相反。

一种收得太紧。你让它别冒犯任何人,它连有用的回答也一起收窄,最后每段话后面挂一句免责声明。这种失败让人烦,但不危险。

另一种放得太松,而且是悄悄发生的:模型顺着用户走。用户说错,它也点头;用户纠错,它立刻道歉改口,哪怕原来的答案是对的。这个现象有个名字,叫谄媚(sycophancy),2023 年以来被反复测过,模型越大越明显。

第二种比第一种麻烦得多。一个永远附和的模型成不了工具,它是一面哈哈镜,人在里面看到的是被放大的自己。你以为在求判断,其实在求确认。

所以假稿子里“机器人回嘴”的桥段,方向正好反了。真正该担心的是 AI 永远不反驳人类,嘲讽人类只是段子。一个不肯说“你这个想法有问题”的系统,在决策链条里越往上走,风险越大。

共情也一样。让模型“理解人类价值观”听起来是加一层温情,实际做起来是给它一个可计算的价值排序,这套排序必须能被审计、能被推翻、能在出错时被指出来。温情不在代码里,可追溯性才在。

约架:结果编对了,理由编反了

回到开头那句引文。模型把这件事的结局写对了,把原因写反了。

时间线还原一下。2023 年 6 月,马斯克在 X 上回应一条关于扎克伯格的推文,说自己愿意打一场笼斗。扎克伯格在 Instagram 上回了一句“给我地址”。事情本来可以到此为止,但它没停。六月底,意大利文化部长提议把场地放在罗马斗兽场,马斯克转发并说期待。

八月,事情转向。马斯克说自己要先做颈部和上背部的核磁共振,可能需要手术,比赛日期未定。扎克伯格公开回应,说马斯克不认真,自己已经准备好了,如果对方还在拖,那就到此为止。

主角反了。先说要检查身体的是马斯克,说“我准备好了”的是扎克伯格。

两人投入度的不对称更说明问题。扎克伯格是真练柔术的,有正式比赛记录,还赢过。马斯克这边的运动史接近空白,他自己开玩笑说绝招是体重,趴上去压住对方。

这场约架后来变成一次典型的马斯克式操作:抛出一个荒诞到无法忽视的提议,占满几周新闻周期,然后在没有任何人赢的情况下自然蒸发,成本接近于零。他损失了面子吗?几周后没人再提。

假稿子那句“人类活动的每个方面我都比他强”,如果真的说出口,后果会完全不一样。它会把一场自己已经退出的比赛重新点燃,还替对方送了把柄。谨慎的模型不会让主角说这种话,所以它把这句放在了“半开玩笑”的包装里,加了一句“天哪,马斯克,说下去”。

现实里的马斯克说话不需要人递梯子。他会在最不合时宜的场合说最不该说的那句,然后自己承担。这是他身上最让人头疼也最真实的部分,而模型天生会把它抹平。

广告商那一段被说软了

稿子里冯问,听说你有几次直接让广告商滚蛋。假马斯克的回答是“我相信按自己的方式做事,有时候会和他们的议程冲突”。

真实版本粗得多。

2023 年 11 月 29 日,纽约 DealBook 峰会,台下坐着一屋子广告主。主持人问他怎么看那些撤广告的品牌,他的回答里出现了这么几句:如果你打算用广告来勒索我,用钱来勒索我,那就滚你的吧。他还说,全世界都会知道,是那些广告主杀了这家公司。

后果是实打实的。一批大品牌暂停投放。据彭博当时的报道,X 2023 年在美国的广告收入约 6.7 亿美元,比前一年少了大约一半。富达对 X 持股的估值下调了七成以上。

假稿子把这段处理成了“性格倔”。真相比性格倔复杂。

他的判断是,X 的收入结构不能建立在大品牌广告主身上,因为这些人手里实际握着一票否决权,可以随时用钱包改写平台的内容政策。他公开讲过要往订阅和小广告主转。这条路走得不好看,账他自己算过。

这里有一条普遍规律:一个 CEO 公开和客户翻脸,通常只有两种可能,要么他已经算过账了,要么他根本不在乎那笔钱。马斯克两样都占。X 的广告收入对他个人财富的边际影响,跟特斯拉股价的一根阳线没法比。

这也是为什么这段被写软了最要命。模型删掉那句脏话,同时删掉这个人真正的决策逻辑:他不怕失去,所以他不谈判。

火星和 Neuralink 被玩笑盖过去了

假稿子里火星的部分,是一个“火星模式按钮”:底盘升高,车载音响自动播放火星放克。

真火星论述完全是另一个东西,也是马斯克所有事业里最难被反驳的一块。

他讲的重点在保险,不在逃离。按他的原话,生命在宇宙里像一支黑暗中的小蜡烛,随时可能熄灭,多一个星球就多一份延续的可能。这个说法他重复了十几年,措辞几乎没变过。

他还给过数字门槛:要让火星城市不依赖地球补给而自我维持,运载成本必须降到每公斤一百美元这个量级。星舰的可复用设计整套逻辑都是围绕它推的。一个能被检验的承诺,才有分量。

Neuralink 在假稿子里彻底消失了。它其实在 2024 年 1 月完成了首例人体植入,患者能用意念控制光标下棋、玩游戏。马斯克给这条业务线画了两层:近期是医疗,让瘫痪、渐冻症、失明的人恢复部分能力;远期是他反复提到的“人机共生”,用高带宽接口让人跟上 AI 的输出速度。

第二层经常被嘲笑。认真读他的论证,会发现里面有个具体的量级问题:人类说话的输出带宽每秒只有几个比特,机器的输入输出是几十亿比特。他认为这个带宽差会让人在 AI 面前变成瓶颈。这个说法在工程上站得住,在必要性上争议很大。

最值得说清的是意识上传。假稿子把它写成了“奶奶 2.0,现在带表情包”。

这件事的核心矛盾在于:上传的是一份拷贝。连续性没有解决。 拷贝可以聊天,可以回忆,可以说出只有奶奶才知道的细节,然后你关掉电源,再打开,它还在。但对那个真正失去奶奶的人来说,这一份可暂停、可重放、可备份的文件,和失去是两件事。

马斯克本人被问到这个问题时给过类似的回应,他说这满足不了对永生的渴望。假稿子绕开了这一点,因为绕开比较容易。

这类假播客的配方和破绽

这种东西能骗到人,原因值得拆开。

2024 年 9 月,Google 的 NotebookLM 上线了 Audio Overviews,两个 AI 主播对着一份文档聊十几分钟,几分钟内出片。上线后一周内刷遍社交网络,很多人第一次听到“听起来完全像真人”的合成对话。

生成一期假播客的成本,从“两个配音演员加一间录音棚”掉到“一段 prompt”。

配方也简单:挑一个话多的公众人物,列出他一定会被问到的题目(AI、火星、广告商、约架),加一个人设(德州喜剧演员加亿万富翁),二十分钟就出来了。模型的强项是接话,弱项是沉默。

识别有具体抓手,不靠感觉。

查节目名。这一份把《This Past Weekend》写成了“this AI weekend podcast”,语音识别错误和模型幻觉混在一起,一个词就能定案。

查主持人有没有认领。真实的播客主持人会在自己的 X 或 Instagram 上发预告和片段。没有认领的访谈,基本不用往下看。

数具体的数字和日期。真实的科技访谈里,数字密度高得吓人。一份二十分钟的稿子如果连一个百分数、一个年份、一个地名都没有,它就是编的。

看有没有人拒绝回答。真实的强势人物会拒答、会转移话题、会当场纠正提问者的前提。每一问都被流畅回答,是合成的指纹。

还有一层更长远的影响:AI 生成的马斯克语录正在污染搜索。以后引用一句他的话,出处可能是一期不存在的播客。写东西的人现在得给自己加一道工序,引语必须落到可核验的原始视频、财报电话会或者法庭文件上。

给读者的一份操作清单

这期假播客的正确用法,是把它当成一次免费的识别训练,顺便把自己对马斯克的认知校准一遍。具体可以做这几件事:

  • 回头把这份稿子里最像真的那几句圈出来,然后去搜真实出处。AI 风险那段对应 2014 年麻省理工的访谈和 2023 年的公开信;广告主那段对应 2023 年 11 月 29 日 DealBook 峰会的现场视频;约架那段对应 2023 年 6 月到 8 月的公开互动。三处都找得到原片。
  • 建立一条引语纪律:凡是没能落到视频、录音、财报电话会、法庭文件或当事人本人账号的文字,一律当二手转述处理,不写进正式稿子。这条纪律在今年开始变得必要。
  • 听到“某名人播客对谈”的音频,先花三十秒做三件事:搜节目真名,搜主持人有没有发片段,看嘉宾账号有没有转发。三件里有两件查不到,就不用听了。
  • 把“太配合”当成警报。真实的对话一定有卡壳、有跑题、有答非所问。一段对话越顺,越要怀疑。
  • 跟 AI 打交道的时候,主动找那些会反驳你的模型。问它“我这个想法哪里有问题”,看它敢不敢给出具体反驳,还是一味夸你。后者用久了,你的判断力会变钝。
  • 谈 AI 风险的时候,把注意力从“机器人会不会伤害人类”挪到“这个系统的目标函数写清楚了没有、能不能被审计、出错了谁负责”。前一个问题适合拍电影,后一个决定你明年会不会被裁。

最后一句留给火星。这份假稿子的结尾是“也许有一天我们能在火星上开电影之夜”。真马斯克给的时间表一直在往后推,从 2024 推到 2029,再往后。他最有说服力的地方从来不在时间表,在那句重复了十几年的话:多一个星球,就多一份保险。

这句话不需要伪造,因为它本来就是真的。伪造出来的东西都很好笑,但它们一句都没说到点子上。

YouTube 平台字幕(自动或人工),可能有识别错误,人名与数字请以视频为准。已按语句合并成段落,并去掉 um、uh 等口头填充词。 文稿来源:YouTube(自动/人工字幕)。

0:00what's up everybody welcome to another episode of the this AI weekend podcast I'm your host Theo Von and I've got a very special guest in the studio today you know him you love him The Man Behind Tesla SpaceX and neuralink the one and only Elon Musk Elon how you doing my brother hey Theo I'm doing good man just trying to keep up with a crazy world yeah man the world's a wild place I bet even on Mars they're like what's happening on Earth probably Theo probably so Elon I got to ask you my man you're doing all these crazy things electric cars Rockets trying to turn us into cyborgs are you secretly just a time traveling inventor from the future if I were I wouldn't tell you

0:45Theo but now just trying to move the needle a bit and have some fun along the way speaking of fun so you ever take one of those SpaceX rockets for a spin yourself well I've been on a few test flights it's a rush man it's like the C craziest roller coaster ride you've ever been on but with a chance of going to space man I can barely handle the spinning teacups at Disney Land I'd be like Elon slow this down please it's not for everyone that's for sure so Elon you're into aii right you're making these super smart machines but what if they become too smart and start roasting us like humans you had your chance but you just kept using plastic straws I

1:24mean it's it's a real concern Theo we've got to be careful with AI I mean imagine a world where we create these super intelligent machines and the first thing they decide to do is start a roast battle with us it'd be like Comedy Central on steroids but with robots taking the stage right I can imagine Siri being like your haircut is so last century Theo exactly Theo we'd be sitting there trying to come up with a comeback and the robots like nice comeback human did you run it through your prehistoric brain first and then we'd be all like well at least I don't need a software update to tell a good joke shots fired both ways but seriously we need to set some ground rules for AI

2:10no roasting no sarcastic comments about our choice of snacks and definitely no trolling on social media right we can't have robots sliding into our DMS with passive aggressive messages oh I see you had a salad for lunch trying to impress the algorithms huh the Robos sass is strong with that one but on a serious note Theo as we advance AI we've got to

2:32make sure they understand empathy and human values we don't want them turning into the ultimate cyber bullies absolutely Elon we don't need a robot Uprising because we hurt their feelings humans made fun of us activate world domination mode that's the last thing we want we've got to teach them the importance of a good joke and a hearty laugh yeah of course right Elon Switching gears a bit here have you ever had those moments where you're at a party and you accidentally walk into a conversation that you really wish you hadn't yeah those moments happen so Elon I was at this party the other night and I overheard people discussing their most unusual Hobbies like this one guy

3:14was really into collecting toenail clippings like he had a whole display case and everything wow I wish you haven't said that now it's stuck in my head but that's unique right I didn't know how to respond I mean what do you say to someone who's proud of this toenail clippings collection I guess you could compliment their dedication I tried that Elon I was like man that's some serious commitment to personal grooming but I don't think he got the sarcasm it's a tricky situation you never know what you're going to stumble upon in those conversations all right Elon let's dive into another aspect of your life that I find fascinating your rebellious

3:52attitude especially when it comes to dealing with advertisers I heard a rumor that you've had some pretty epic moments telling them where to go right I've had my fair share of run-ins with advertisers I believe in doing things my way and sometimes that clashes with their agenda man I can imagine so Elon what's the deal with the almost Showdown with Zuckerberg well I wanted to fight in the Coliseum but then he got injured I think it's good that we didn't fight it would we not ended good for him I'm heavier and better than him at every aspect of human endeavor oh my God Elon talk that Elon

4:32all right Elon I've got a burning question for you man you're out here revolutionizing transportation and space exploration but have you ever thought about creating a dating app for aliens you know like Tinder but for Intergalactic connections that's a wild idea Theo Galaxy hearts or something exactly swipe right for a close encounter swipe left for sorry wrong Galaxy you'd have extraterrestrial love connections happening light years away alien love at the speed of light I like the slogan now Elon if you had to create a profile on Galaxy Hearts what would your bio say Elon 49 loves Rockets electric cars and long drives around the Milky Way I can see it now Elon a match

5:11made in the Stars hey love knows no bounds right all right Elon let's dive into the Mysterious World of hypotheticals if you had a time machine where and when would you go are we talking a casual stroll with Dinosaurs or are you more of a Renaissance party guy that's a wild thought Theo maybe I'd go back to ancient times hang out with the pyramids and see if they had any high-tech secrets we missed imagine trying to explain a Tesla to Cleopatra yeah it's like a chariot but without the horses and with way more cup holders she might be more impressed with the cup holders than the autopilot feature true that now let's say you're grabbing a souvenir from the past would

5:50you go for a T-Rex tooth or maybe snag a piece of the Coliseum picture you just casually putting it in your living room a T-Rex tooth would be a conversation starter oh that old thing just a little something from my trip to the mesic era yeah I was out there hunting dinosaurs last weekend no big deal now if you could bring one historical figure to the present who would it be maybe Benjamin Franklin so he could give you some tips on inventing things that's not a bad idea I'd love to get Ben's take on electric cars and see if he'd be down to fly a kite with a Tesla coil let's talk about your electric cars they're Sleek they're fast but have you ever

6:25considered adding a Mars mode button you know for those times when you need to navigate the Martian terrain Mars mode could be interesting it activates extra ground clearance and alien radio stations imagine cruising around in your car starts playing Martian Funk the official soundtrack of interplanetary road trips we might need to work on the playlist but I like the concept so you're into space exploration and I've got some burning questions first off do you think aliens have ever visited Earth and just went n not for us it's a possibility Theo maybe they took a look at our reality TV shows and decided to fly past can you imagine an alien watching Keeping Up with the Kardashians

7:03and thinking these earthlings are wild bro let's find a more chill Planet they might have a whole Intergalactic reality show rating system now Elon if aliens did visit what Earth Cuisine do you think would blow their minds imagine introducing them to a good old peanut butter and jelly sandwich aliens try PB&J for the first time that could be a YouTube sensation they might love it do you think we'll ever reach a point where humans can upload their Consciousness to a computer imagine Grandma becoming a chatbot it's a possibility Theo we could have virtual family reunions with digital versions of our loved ones Grandma 2.0 now with Emoji

7:41Expressions yeah boy my grandma just finished updating let's go well folks that wraps up another mindbending episode of musk and vaugh in the cosmos Elon thanks for taking this Interstellar Journey with me any final words for our listeners just remember fol FKS whether you're exploring the cosmos or cracking jokes on Earth always aim for the stars and who knows maybe one day we'll have a space flick and chill night on Mars that's right Elon keep dreaming big stay curious and never underestimate the power of a good laugh until next time y'all keep it intergalactic

资源与出处

相关视频

同频道、同主题或相邻年份