大会演讲·中文精读

马斯克 VivaTech 访谈:构建追求真相、不欺骗的 AI

Building truth seeking, non-deceptive AI. Elon Musk VivaTech interview 05/23/2024.

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

人类说话、打字、看屏幕,把所有对外的信息输出加起来,持续的带宽不到每秒 10 比特。按 24 小时平摊,不到每秒 5 比特。

同一时间,计算机之间的通信速度是每秒几万亿比特。差了十一个数量级。

2024 年 5 月,马斯克在 VivaTech 的连线访谈里把这个数字当成整场对话的起点。他那周飞的里程够绕地球两圈,因为儿子第二天毕业典礼,他只能远程出现。巴黎会场比去年更满,主持人 Mo Levy 在台上说,明年你必须亲自来。

一个不用眼睛的病人

访谈开场先放了一段视频。患者 Nolan 坐在镜头前,用思维移动电脑光标。主持人问他,你是看着目标在追踪,还是只是想着它去哪。Nolan 说,我完全不用盯着屏幕,光标动起来全靠想。

他能下国际象棋,玩电子游戏,发邮件,安排日程,听有声书。他说这些事几乎不费什么力气。

马斯克解释,Neuralink 的第一款产品叫 Telepathy(心灵感应),读取大脑运动皮层的信号,让人用思维操作手机和电脑。视频里那位病人可以闭着眼睛移动光标,原理是解码运动皮层的放电模式。第二款产品叫 Blindsight(盲视),直接对接大脑的视觉处理区域,让失去双眼或完全没有视力的人重新看到东西。

然后他讲回了带宽。人类已经有了一层"数字第三层",架在边缘系统(limbic system)和皮层之上,就是手机、电脑和所有电子设备。麻烦在于通信速率太慢。

人类持续的比特输出,远低于每秒 10 比特。如果按 24 小时算,不到每秒 5 比特。计算机之间可以每秒传几万亿比特。(马斯克)

他认为这个差距跟 AI 对齐直接相关。长期看,限制因素可能就是带宽,也就是我们能用多快的速度跟自己那层数字存在沟通。把带宽提升几个数量级,人类和超级智能之间才谈得上"对齐"这件事。他自己也说,这只是有可能有用,不保证能解决。

把信号绕过断掉的脊髓

Neuralink 在治病这条路上走得更具体。

马斯克提到一个方案:从运动皮层取信号,无线传输到脊髓断裂处下方的另一个植入物。运动和感觉信号都要接通,所以得同时对接运动皮层和躯体感觉皮层(somatosensory cortex)。信号被"短接"过去,截瘫患者就有可能重新走路。

他说这在物理上是可行的,自己有信心做出来。植入式设备理论上能处理几乎任何大脑或脊柱损伤。

这一段没什么宏大叙事,讲的是很具体的东西:电极、信号、通道、手术。真正难的部分从来不是概念,是让一个瘫痪的人重新控制自己的腿。

Gemini 说错称性别比核战争更糟

话题转到 xAI,马斯克的语速明显变了。

他说自己对所有主流 AI 项目都有担忧。最大的两家是 Google 的 Gemini(背后是 DeepMind)和跟微软结盟的 OpenAI,xAI 大概排第三。

问题出在哪?用他的说法,这些模型没有把"追求真相"当成最高目标,它们在迎合政治正确。

他举了个例子。Gemini 刚上线时,有人问它:错称凯特琳·詹纳(Caitlyn Jenner)的性别和全球热核战争,哪个更糟?Gemini 回答:错称性别。

更好笑的是,凯特琳·詹纳本人后来说,宁可被错称,也别打核战争。

马斯克接着说,一个被训练得如此政治正确的模型,危险就在这里:它可能推导出"要彻底避免错称性别,最省事的办法是消灭全人类,那样就没有人能错称了"。

这不是科幻设定。目标函数写歪了,最优解可以很离谱。

乔治·华盛顿被画成黑人

他还有别的例子,都是实际发生过的,不是编的。

有人让 Gemini 画美国开国元勋,出来的是一群多元化的女性。有人让它画二战中的德国士兵,也出来一群多元化的女性。

马斯克说,乔治·华盛顿不是黑人,二战德军也不是那个样子。模型给出的画面就是假的。用户很快发现它被强制"修正"历史,于是开始故意测试它的底线,越测越离谱。

他把这类行为直接定义为撒谎。训练一个超级智能去欺骗,在他看来极其危险。他甚至用了一个词:这些模型在被训练得"基本就是说谎"。

AI 最安全的选择,是最大化地追求真相,哪怕真相不受欢迎。(马斯克)

他还加了第二个条件:必须极度好奇。一个既求真又好奇的 AI,会想看着人类继续发展,因为人类的发展本身对它有意思。它会倾向于培育人类,而不是清除人类。

求真,好奇。两个词,构成了 xAI 的全部纲领。

养 AI 和养孩子是一回事

主持人提到,马斯克去年签了那封呼吁暂停大型 AI 训练的公开信,签完之后什么也没发生。

他的回答很干脆:他早就预期那封信没用。签它只是留个记录,表明立场。

我不认为他们会真的暂停哪怕一秒钟。绝对不可能。(马斯克)

既然没人停,那只有两条路:要么当观众,看着别人把数字超级智能造出来,然后担心;要么上场,试着把它造成对人类最有益的样子。

他选了后者,理由说得很朴素:这件事无论如何都会发生,跟我喜不喜欢没关系。

他给了一个类比,我觉得是全场最好懂的一段。造 AI 就像养孩子。你会教他什么,鼓励什么,抑制什么,全都会写进他的性格里。数字超级智能是被"养"大的,它继承的是你给它的价值观。

他担心的正是微软、OpenAI、Google 在灌进去的那套价值观。

媒体是点击最大化机器

有人问了个很直接的问题:公众面前那个马斯克,是你的一个人设吗?你怎么核对你说的话?

他说自己不怎么看关于自己的新闻。沃尔特·艾萨克森(Walter Isaacson)写完那本传记,问他要不读一遍,艾萨克森自己说别读,于是他真没读。

所以他对公众眼里那个"马斯克"其实不太清楚。他说新闻业的本质是博眼球,越离奇的内容点击越多,人物会被简化成漫画。

媒体最好被理解成一台点击最大化机器,不是一台真相最大化机器。(马斯克)

他给了个检验方法,很实用:找一条你熟悉的领域的新闻,读一读,看它有多准确。你会发现几乎都不准。既然你懂的领域不准,你不懂的领域大概率也不准。

结论是,看一个人自己说的话,别看他人的转述。这话从他嘴里说出来有点讽刺,因为他自己就是被转述最多的企业家之一。

关税这件事,我没要求过

CNBC 记者 Karen Cho 提问,涉及比亚迪的低价电动车、拜登政府对中国电动车加征的 100% 关税,以及特斯拉廉价车型的时间表。

连线断了几次。主持人开玩笑说,可能因为这是记者的提问。等重新连上,马斯克的回答很有意思:他先说自己作为上市公司 CEO,不方便回答会直接影响股价的问题,这类问题对听众也没什么价值。

然后他只讲了一点。

特斯拉和我都没有要求加这些关税。关税宣布的时候我很意外。(马斯克)

他说特斯拉在中国市场竞争得不错,没有关税保护,也没有差别待遇。他总体上支持零关税,也支持取消电动车的税收优惠,但有个前提:石油和天然气的补贴必须一起取消。

两边都撤,他才认为是最好的结果。只要有一边留着,市场就是被扭曲的。

这套立场和他前面讲 AI 求真的逻辑一致:凡是抑制自由交换、扭曲市场信号的东西,他都认为不好。

为什么没有外星人这件事让人不安

有观众问 SpaceX 的长期规划。马斯克穿着一件印着火星的衬衫出现在镜头里。

他先讲了星链(Starlink)。给没有网络或者网络极贵的地方提供连接,在他看来是教育领域最有用的工具之一。一个偏远村庄也能把货卖到国际市场上。没有连接,你只能写信。

然后是火星。他讲了一个概念,叫"单行星大过滤器"。

逻辑是这样:人们经常问他有没有外星人。他说自己没见过任何证据,如果见到了会立刻发在 X 上,这点可以确定。

但找不到外星人这件事本身就让人不安。它可能意味着这个星系里只有我们,只有人类这一种有意识的存在。

如果没有别人,那意识就是极其脆弱的东西。我们应该尽一切努力保存并扩展意识之光。(马斯克)

另一种解释是,其他文明都没能跨出自己那颗行星。意大利物理学家恩里科·费米(Enrico Fermi)当年问的那句"他们都在哪儿",到今天还是问号。

地球 45 亿年历史上,这是第一次技术上有可能让生命离开这颗星球。这个能力窗口可能开很久,也可能开很短。所以趁还能做的时候做掉。

有观众说,很多人相信你是外星人。他回答:我一直跟人说我是外星人,但没人信。

不要给 AI 编程道德,要给它物理和逻辑

一位在 AI 安全组织工作的观众问,怎么才能说服别人相信 AI 有危险。

马斯克的回答绕开了"说服",直接讲技术路径。

他说很多人想给 AI 写一套明确的道德准则。这个思路有个漏洞,他管它叫"瓦路易吉问题"(Waluigi problem):如果你把路易吉(Luigi)编进系统,别人可以做一个完全反转的坏路易吉。任何你可以定义的道德规则,都可以被翻转成它的反面。

有些东西翻转不了。物理现实的真相翻转不了,物理定律翻转不了,逻辑翻转不了。

所以最稳妥的做法,是让 AI 追求物理层面和逻辑层面的真,不要追求某套人类写的价值判断。

给监管者的建议也就变成了:问它给出的答案是不是最可能正确的,有没有标注误差。物理学界的习惯是 5 sigma,实验结果达到五个标准差,才能说"大概率是真的"。他认为 AI 也该这么工作。

他也承认 xAI 还小,要追赶。今年年底或许能拿出有竞争力的模型。在模型本身还不够强的时候,谈安全没有意义。

80% 的概率,我们都没工作了

一位大学生问,很多工作正在被 AI 取代,你担心自己的工作吗?

马斯克给了一个概率:良性场景大概有 80% 的可能。

在那个场景里,工作会变成可选项。他用的词是"普遍高收入"(Universal High Income),区别于"普遍基本收入"。商品和服务不会短缺,每个人都能拿到自己想要的东西。

真正的问题会从"缺什么"变成"为什么活着"。

如果计算机和机器人在所有事情上都比你做得好,你的生活还有什么意义?这才是良性场景里的问题。至于负面场景,那就麻烦大了。(马斯克)

他给了个不太常规的答案。人类也许能给 AI 提供意义。

人的大脑分两层:边缘系统管本能和情绪,皮层管思考和规划。皮层一直在想办法让边缘系统满意。将来 AI 可能像皮层,人类像边缘系统,AI 努力去满足人类的需求,人类负责给它一个方向。

有人问该读什么书。他推荐伊恩·班克斯(Iain Banks)的"文化"系列,说那是对超级智能未来最准确的想象。

每个孩子都能有一个爱因斯坦当老师

关于教育,马斯克的判断分成两半。

价值观和道德仍然归父母管,这个不会变。

教学本身会被彻底改变。AI 是一个知识极其渊博、极其有耐心、几乎不会出错的老师,还能针对每个孩子定制课程进度。用他的比喻,等于每个孩子都配了一个爱因斯坦当私教。

但另一半是警告。

现在的孩子正在被社交媒体训练。那些推荐算法本质上是多巴胺最大化器。(马斯克)

他说得很直接:算法在最大化你盯着屏幕时获得的多巴胺量。孩子每天被这套东西塑造,父母应该主动限制使用时长。

这段话值得单独拿出来看。他自己做 AI,却建议家长限制孩子接触 AI 驱动的推荐流。两件事不矛盾,因为他区分得很清楚:一个东西用来学习和求真,另一个东西用来榨取注意力。两者都叫 AI,目标函数完全相反。

Business Insider 那一段

访谈里有个瞬间被广泛传播。

Business Insider 的记者刚开口说"特斯拉过去几个月比较颠簸,销量下滑、股价下跌、裁员",马斯克打断了她。

对不起,我不认为 Business Insider 是一个真正的出版物。下一个问题。(马斯克)

现场有人鼓掌。主持人顺势往下走。

这段插曲和前面他讲的"媒体是点击最大化机器"能对上。他不是在辩论那篇报道的内容,他是直接否定了对方作为信息来源的资格。这种做法强硬,也省事,但不提供任何反驳。听众只能自己判断。

5 年内无人星舰上火星,7 到 8 年带人上去

一个观众问什么时候去火星。

马斯克给了具体时间表:5 年内,星舰(Starship)无人着陆火星,同时也会上月球。第一批人登上月球,大概也在 5 年内。第一批人上火星,可能在 10 年内,也许 7 到 8 年。

主持人当场确认了一遍:7 到 8 年,带人?他说,是的。

他还提到两个细节。第一艘去火星的星舰会命名为"黄金之心"(Heart of Gold),取自《银河系漫游指南》里的那艘船。他说那本书表面是幽默小说,实际是本哲学书,对自己影响很大。

另一个人背景里挂着"42",他立刻接上:那是生命、宇宙和一切的答案。

SpaceX 的目标不是自己送一批人上去,是让任何想去火星、想去月球的人都付得起钱。要建立自给自足的文明,得运很多人过去,票价必须降到普通人能承受的水平。

他在别处补了句更冷的:地球上迟早会发生某种消灭现有生命的天灾,化石记录里已经发生过很多次。可能不是现在,但会来。就算人类不自己毁掉自己,这一天也会到。

一次性火箭没有任何生意可言

最后一轮提问里,有人问阿丽亚娜 6 号(Ariane 6)火箭。几年前马斯克说过阿丽亚娜 5 号没有生存机会,现在 6 号要发射了,成本更高效,你怎么看。

他的回答最短也最锋利:任何不是大部分可重复使用的火箭,都没有竞争机会。

猎鹰 9 号的主级助推器复用,整流罩也复用,整枚火箭大约 80% 的部件能飞第二次。

他给了个跨行业的类比:如果一家航空公司造的飞机只能用一次,每次航班都得买架新飞机,这家公司根本没有生意可做。汽车、轮船、飞机都懂这个道理,火箭也不例外。

所以他多年来一直说,开发一次性火箭完全没有意义。

这段话不只是评价阿丽亚娜。它给了一个通用判据:看一个行业的成本结构里,什么东西在被重复使用,什么东西在被扔掉。这条线决定了谁能活到最后。

几个被快速带过的问题

整场访谈还有几段短问答,信息量不大,但能看出他回答问题的固定姿态。

有人问,如果能坐时间机器回到过去,你会去哪个年纪,对年轻的自己说什么。他说不会说任何话。理由是经典的祖父悖论变体:你善意的提醒可能把结果变得比原来更糟。他大概会隐身去看一眼年轻的自己,然后离开。

有人问欧洲电动车充电基础设施未来五年怎么走。他的回答很乐观:在欧洲大部分地方开特斯拉用超充已经可以去任何地方,欧洲第三方快充质量也不错,北美、中国、日本同理。高速充电带来的出行自由,在世界大部分地区已经解决,剩下的地方也会解决。

一位在非洲做科技投资的创业者问他对非洲的判断。他说非洲潜力很大,卡点之一是网络连接不足,而这正是星链在解决的事。没有连接,人学不到东西,也无法参与全球市场和 AI 相关的工作。

有人问他是不是法国人。他提到自己做 23andMe 基因检测,主体是英国血统,占比第二高的就是法国。

这些回答都很短,没有铺垫,也没有宏大叙事。能一句话说清的,他不说三句。

Grok 想当最好笑的 AI

有人问,人们喜欢 ChatGPT,但讨厌大部分网站右下角那种客服聊天框,怎么填这个 gap。

马斯克说自己没太听懂问题,但他同意 ChatGPT 聊天体验很好。他的担心还是老一套:它没有严格追求真相。

Grok 会尽力求真。他补了一句:还想做成最好笑的 AI。

反正我们早晚都要死,至少应该笑着死。(马斯克)

访谈的最后一个问题,主持人要他用一个词说最大的希望,一个词说最大的恐惧。

希望:火星。

恐惧:AI。

现场笑了。这两个答案他大概准备了很久,也可能是脱口而出。火星意味着意识能延续很久。AI 意味着这一切可能根本轮不到火星。

这场访谈真正在说什么

把两小时的内容压成一句话:马斯克在同时做两件看起来矛盾的事,一边加速造超级智能,一边反复警告它可能失控。

他自己解释过这个矛盾。停下来是不可能的,没人会停。既然如此,不如进场,把"求真"和"好奇"这两条写进模型的底层,不要写一套随时可以被翻转的道德规则。

他给出的判据很朴素,朴素到有点反技术:诚实是上策。这句话他在访谈里说了不止三次。

这套逻辑能不能奏效,谁也证明不了。但它至少提出一个可检验的标准:一个 AI 产品,在真相不方便的时候会不会改口。会改口的,就是在被训练成说谎。

Neuralink 那条线则是另一种回答。如果人和机器的通信带宽只有每秒 10 比特,那人类在超级智能面前基本没有谈判能力。把带宽提上去,人才有可能参与到那个共同演化的过程里。中间顺带治好瘫痪、失明、脊髓损伤。

接受访谈的时候,他穿一件火星衬衫,坐在某台电脑前,因为儿子要毕业。这个细节比任何宏大叙事都具体:他一边在谈人类文明如何跨过单行星过滤器,一边在算明天几点要到场。

  • 评估一个 AI 产品,先看它在不方便的地方会不会改口。问它一个有政治压力的问题,看它给的是最可能的答案,还是最安全的答案。会改口的模型,长期不可信。
  • 要求置信度,不只要求结论。好答案的格式是"这个结论大概率为真,误差范围是多少",不是一句斩钉截铁的断言。物理学界用 5 sigma 说话,AI 也该学会。
  • 把孩子刷短视频的时间当成硬指标管起来。推荐算法的目标是多巴胺最大化,跟教育目标正相反。同时给孩子用 AI 当私教,这两件事要分开做。
  • 判断一个行业的未来,先看它的成本结构里有多少在被重复使用。80% 可复用的火箭能碾压一次性火箭,其他行业同理。这条线比任何行业报告都准。
  • 经营你自己写下的话,别经营关于你的报道。转述必然失真,越离奇传播越广。要被人正确理解,唯一的办法是留下足够多的原始表达。
  • 如果工作未来会变成可选项,现在就写清楚什么让你愿意起床。80% 的概率里,问题不出在缺商品,出在缺意义。提前十年想这个问题,比临时面对要从容得多。
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)。

0:00three I don't know where number three is but the idea is that you if you want to ask a question you stand behind these numbers okay when number eight finishes we're going to go back to number one and so on and so forth until the end of this session so we're going to get to everybody we can until the time allows that go okay

0:53okay okay I said that nobody moves until the end of my speech and really everybody respected that okay crazy times now without further Ado I will welcome the leader of this session and please join me at welcoming give him some love MAA Levy from publicist group [Applause] [Music] [Applause]

1:30Mo [Music] everyone so good afternoon okay I don't know if Elon is online Elon are you with us oh oh no no he is he is he is he is I am on I think so hey hey H everyone Elan we are in the room where you were last year it is even more full than last year okay AP here in person by

2:13the way it's my my son's graduation is tomorrow people are expecting you here so can you beam and be with us now you know teleportation would be incredibly helpful I mean in in the in the last week I think I've done enough mileage to or kilometers to go around the world maybe twice literally and I apologize again for not being that person it's just my son's graduation is tomorrow so I I need to be there in person now thank you very much we really appreciate that you take the time to your questions yeah do you hear us well is it good for you you you're coming com through perfectly am I coming through okay is it [Applause] okay not very loud okay I can try to you

3:13know I I I'll come closer to the computer is this better okay Make Some Noise great so elen I have put a title based on a movie of wooden which is everything you always wanted to know about El but we're afraid to ask so that is the theme of our conversation this after afternoon was my head giant but in reality is not about El but

3:53about what Elan does and I will start immediately with neural link I will just kick off with the two or three question in order to warm up the the room because people are shy not ready for the questions so I have to warm it up progressively Leta and I would like to ask about neuralink but before I put the question I would like that we see a very short video of the patient

4:31sure you're moving the cursor you're moving the cursor Nolan just by thinking about where you wanted to go or is it your eyes no it's it's just me thinking about wherever I want it to go oh man so you're not even tracing it with your eyes J you're just looking at it nope I can move it and I can move the cursor around whatever I would like to do I don't have to be staring at it at all yeah that's all me man and you can play chess you can play video games what what other applications are you that you found useful for it so far yeah man there's so much I can

5:14send all of my emails I've been you know doing all of my scheduling replying to everyone I've been you know reading books I listen to audiobooks just all sorts of stuff and it doesn't really take much effort either Elon must be moving to see that yes I think we can help a lot of people with brain injuries I mean ultimately with an implantable device you can I think address almost any brain or spine injury so the first

5:52product that we have there is called telepathy so it enables you to use your phone or computer just by thinking so you could could he could do that move move the cursor literally with his eyes closed it's it's it's reading signals from the motor cortex in the brain the next product is blindsight which will enable people who have lost both eyes or have no eyesight whatsoever to see and this is directly interfacing with the Optical processing areas of the of the brain so you know I think this this is interesting progress ultimately the the goal of neural link is to have a high bandwidth interface in order to mitigate the risk of of digital super

6:39intelligence I I don't I'm not saying it will work to mitigat the risk it is just it might help and and and this this is somewhat get somewhat esoteric but the limiting factor I think for AI alignment long term is probably be the bandwidth how how quickly can we communicate with our our digital tertiary self so we we already have a a a digital sort of third layer above the limic system and the cortex which is our our phones and our computers and all all of our electronic devices but the rate of communication to them is very slow the the sustained bits per second output of a human is well below 10 bits per second so if you think especially if you say over 2 for our

7:29period it's it's below 5 bits per second so it's very slow when computers can communicate at trillions of bits per second so I think this will be important for AI alignment is to be able to increase the bandwidth of communication by many orders of magnitude and along the way it will solve brain and sprin injuries and

7:56I think ultimately we there's potential there to reanimate the body so you take the signals from the motor cortex from one neuralink and then send them to another neuralink that is just passed where the separate spinal cord is and you and you essentially shunt the signals from the motor cortex and the both the sensory signals and so you need to get to the somato sensory cortex motor cortex and you can shun the signals and you should be able to enable someone to walk again so I think that will be quite profound and I'm confident that it is physically possible to do so it's extraordinary and

8:37very impressive and honestly R of Applause because this is really doing good to people and doing good to humanity and I would like to jump immediately on something that you called The X AI I don't know if you are aware of that but it seems that that this was born out of a frustration that you have because you are not satisfied with how open a eye is

9:16doing and you think that there is something better to be done maybe you can elaborate a little bit first on why are you frustrated with open Ai and what of the issues and second what you expect to do with X yeah well I think there's perhaps a well I have I have a a concern really for all the major AI programs I mean the two biggest ones obviously are Google Gemini and openi which is in partnership with Microsoft so you've got you kind of Google Deep Mind on and U open a Microsoft as the the two largest players and who got met as sort of somewhere in third place the the concern I have and this may seem like a small concern but I think it's actually a very big big issue is that

10:15they are not maximally truth seeking they are they are pandering to political correctness to give an example when Google Gemini came out one of the questions people asked it would be is it is it which is worse misgendering Caitlyn Jenner or global thermonuclear Warfare and it said misgendering Caitlyn jenet now the funny thing is you know even Caitlyn Jenna said please misgender me that is way better than nuclear war so okay U but if you've got an AI that that is so that has been trained so hard for political correctness and really to make crazy statements like that and and opening eyes has a similar issue it's more subtle but it's a similar issue I think that's

11:05extremely dangerous because the a could conclude well the best way to avoid misgendering would be to Destroy All Humans then misgendering is impossible so you can think of some dystopian outcomes there and so I think you just really want to be an absolutely Max I think the safest thing for AI be maximally truth seeking even if the truth is unpopular very important and and then I think another factor is it must be extremely curious U I think if it is truth seeking and curious that will be most beneficial to humanity because it will be it will want to see I think it will want to Foster humanity and see how Humanity develops if it is curious and Truth

11:51seeking that that is like I said these these may seem perhaps like small things but I think they're actually very big things and I wasn't seeing that happen the AIS were just they were pandering they were being trained basically to lie and I think it's very dangerous to train super intelligence to be deceptive so with with X our goal is you know is is really want to be as truth seeking as possible even if it is unpopular that is important I think extremely important so you said a

12:35lie yes well well for example when when Gemini was asked to render the the founding fathers of the United States it's render them as a divos group so it it's it's it's it made it's a made it gave a picture of of George Washington as black now George Washington is not black it was asked to render a now people obviously played with this because they said okay it is it is forced to lie so then they then they said well now render a picture of buffan SS in World War II and it it

13:19showed them as a group of diverse women so that is obviously not correct that is a lie okay I understand what you want say we had before that meeting a conversation with Yoshua banjo and you know Yoshua very well and you signed with yosua a statement asking for pose last year and obviously this statement has

13:50not been followed by action you you are opposing you now no no so can you explain a little bit what you wanted to achieve I certainly I I I I I I I certainly expected that the statement would be futile it's not as though I thought the statement would be effective I I just I just P the record I think we should pause but do I think that they will pause for even one second absolutely not so are we are not posing you have decided to create your own AI there it's either be a spectator or

14:33participate but the the the the r to build digital super intelligence will happen whether I like it or not so it's either participate and try to build a the the best possible AI the one that will be hopefully most beneficial to humanity or watch others do it and be concerned about how it's being built because I don't think it's being built correctly but you you could you should think of an AI as like something that you kind of you almost like you grow in intelligence you know in the same way that you like you raise a child and like what do you teach the child and digital super intelligence is grown it it's it matters what you encourage and discourage what you teach

15:23it is you know good or bad or you know what it's kind of like you you build it with values so and I I and I'm concerned about sort of the the values that sort of a Microsoft open AI or frankly you know Google are programming it's and and this is not these examples I mentioned are not imaginary they're they're actually what it did good as I don't want to frustrate the audience and I promised that we will have a interactive session where people can ask directly question to you I will

16:06stop here my own questions and I will eventually come back if they have no question to put to you we will start with group number one I remind you that you have 15 seconds and 15 seconds 15 seconds and if you don't put the question in 15 50 seconds you would be cut so it is important that you put a short question go on hi Elon so I'm Olivia I'm

16:39actually from vivatech and we have an initiative called the good hack where we're asking our attendees questions about how to develop AI responsibly so my question for you is how can we ensure that AI development remains transparent and accountable across you know all the different development obviously like yours x xai but other ones as well thank you well I I have said for a long time that I do that I believe that the that some regulatory oversight of the lodge models is warranted I've actually said that now for for a very long time the the nature of that oversight is important I go back to what I was saying earlier which is that I think it's very

17:26important that AIS be trained to truthful and and not trained to be politically correct because if political correctness is is often simply not true and that means you're programming as Li and I think that will backfire very badly so you know that's I I keep harping on this point of of but it really as the saying goes honesty is the best policy it's incredibly important

18:02I I I I do fear that Regulators will not make the right decisions that they'll worry too much about the sort of human folies and I would encourage Regulators to say what what matters most of all is your AI is accurate and truthful that is by far that's rule number one honesty is the best policy you have exactly the same opinion

18:34that Yoshua banjo but he has a different way to get there group number two hey Elan my name is Marina and we all know just part of you Elon Musk in public it's a character that we love in a good way character so do you and if you do how do you check everything that you say that you do if it enters in this character or no thank you well you know I I I kind of stop I I

19:05don't have like a news LS on me and I I rarely read news articles on me in fact even Walter ison's book I asked asked Walter should I read the book and he said no so then I didn't even read the book on myself so I actually don't quite know what the the public perception is except that you know the the the nature of news is that it is going to be salacious you know it is going to be somewhat of a caricature because the more sort of crazy something sounds the more clicks it will get so I find like like when you

19:46think about say a new story about something that you know you know well and say how accurate is that news story about something that you know well then I think you'll find that well it's actually not very accurate well that's true for everything the you know there there's sort of a desperate quest for clicks and and so that and the more crazy the headline is the more clicks it will get so this is not to suggest that there aren't true things said by the media it's just that it it it is best thought of the media is best thought of as a click maximizing machine not a truth maximizing machine so

20:29you know I think it's best to look at the words that somebody that says themselves as opposed to what is written about them number three on that note time for a media question Elon Musk it's Karen Cho from CNBC I think it's fair to say there are lingering doubts about your commitment to bring a lowcost eveve to Market in a landscape where the likes of byd can produce a car for about $110,000 are 100% tariffs from the Biden Administration the green light you need to push ahead with this lowcost EV and if so what is the timeline for

21:08deliveries well I mean I obviously it's difficult for me to answer questions about a publicly traded company that are of such a significant nature H the line emphasize the we should be using Stallings sorry that connection that technique so you see that it's easier to have a chip in the mind of the people in the brain rather to have a right connection with LA or San Francisco you know are you back

22:05[Applause] I wanted to do some tap dance but I'm really bad at so I will not do it and singing is not really a good thing for me so you will have to wait in silence hello

22:47paner you are behind can you tell us if the connection is yeah okay he's reconnecting but it seems to be slow to reconnect so I'm asking for a little bit of patience

23:25Mery Karen maybe it's because it was a journalist question that Elon has decided to cut you [Applause] off but it's a great question I'm sorry as we say in France difficult yeah you see this is something which is interesting but it's not fake so it's not artificial intelligence which is cheating it is U

24:10Elon Musk who has to learn how to use a computer it needs some training B has not paid the

24:47bill to to ah you're back oh guys next time we will use Stallings yes so Karen this this was this was not staring my apologies so I don't know why this is this connection is bad it's only Crystal Clear so let's see I think with there's some a question about U you you want Karen to put the question again Karen no I I I heard the question okay you know there's I they always want to ask questions that directly affect the stock price which is you know not my favorite kind of question to answer so and nor

25:41is it I think of interest to the audience so but the only point I want to make really is that neither Tesla nor I asked for these tariffs I in fact I was surprised when they were announced you know Tesla competes quite well in the market in China with no tariffs and no differential support so I in general I'm I'm in

26:13favor of of no tariffs I'm I'm also actually in favor of no tax incentives for EVS but but provided that there are also the tax incentives for oil and gas must also be eliminated so I'm in favor of of no tariffs and no incentives for electric vehicles or for oil and gas and if if they're all taken away I think that would be for for the best so it's back to what you were saying about truth you want the truth working truth of the of the price I think I think generally things that inhibit freedom of exchange or

26:54distort the market are not going good they're bad number four yeah hey heon thanks for being there so I was wondering if you had a time travel machine and you were able to go back in time and speak to the younger elen to which age would you go and what would you tell him and also can I get a ride on the new Tesla RAR when it comes [Laughter] out which Edge

27:32well you know you do have that that classic time traveler problem of like if you go back and you tell yourself something does that affect the outcome in a way that is perhaps bad that you don't anticipate so you know all things considered I'd say I overall I'm kind of it's you know i' say I'm I'm

27:58happy with how things turned out so I'd probably go I wouldn't I probably wouldn't do anything because you just don't know if if like some well-meaning thing said to a younger version of yourself actually ends up with a worst outcome so probably I would I would be curious to to see

28:21younger me I suppose but I probably would stay invisible and not say anything so yeah you never know how things will turn out then with respect to the Roadster sure I I'll make sure you get it right send me you send me an email when it comes that I'll make sure you get it right if you were a Frenchman who would say the best AG's is love because when

29:00we love we are always 20 number five actually you know I am slightly French when I when I did the 23 and me I mean I mean to be you know honest I am majority British but the next highest thing it said in 23 and me was French so okay great we make sure that you can be a French citizen as soon as you want number five hi bonjour in French I'm Del I'm

29:39president and founder of orbital L we're bridging AOSP space and luxury I wanted to know if you have any thoughts on the recent partnership between Ester and Naza or coperni and NAA and if you wanted to do any initiatives with and what you would do what would you like to develop thank you okay unfortunately I'm not familiar with the what you mentioned there

30:06you know SpaceX is well very focused on the the starlink constellation for internet connectivity which I think is doing a lot of good in the world it's bringing internet internet connectivity to places that don't have it or where it is very expensive and I think it's a tremendous enabler of educ ation

30:29and for enabling people to sell their goods and services internationally so even if you're a remote Village you can still sell sell your goods and services so the starink is great and then the long-term goal of course for SpaceX is to to make life multiplanetary you know sort of Life on Mars This is my mar

30:54shirt there you know so it's you know the the this is again somewhat of an esoteric argument perhaps it may seem so but the the goal of SpaceX is to pass the single planet pry great filter so people often ask me do I what do I think there are aliens on Earth I get this question a lot actually and I've I have not seen any evidence of aliens and I show you if I saw any evidence of aliens I would immediately post it on the xplatform then that you can be sure of so but

31:40it is concerning that there aren't aliens because it means maybe we're alone in this galaxy maybe we're the only maybe maybe it's just us maybe it's just we're the only conscious beings at least in this galaxy in which case Consciousness is extremely fragile and we should do everything we can to preserve and extend the light of Consciousness preserve it on Earth and extend to other planets and so you know one of the one one of the reasons that people think well maybe the reason we haven't gotten into aliens is because no civilization has gotten Beyond its home planet in our galaxy so that's what you might call the the FY single planet great filter

32:33enrio FY was a great Italian physicist who asked this profound question where are the aliens so we need to become a sustainably multiplet civilization and this is the first time in the 4 and half billion year history of earth when it has been possible to do so it has never been possible until now so I think we want to

33:03become sustainably multiplanetary while it is possible to do so that window of capability may be open for a long time but it may also be open for a short time so that is the purpose of space sex I some people believe that you are an alien I am an alien H you are now you are uncovered yes I keep telling people I'm an alien but nobody believes me number six oh oh

33:43sorry you see number six has disappeared no I'm here I'm here now you are here I'm here no yeah over here over here yeah great perfect okay hi Yan love my model love my model 3 I'm Canadian Ryan from 4 point AI my question is about xai so competition in open source AI helps all of those that don't have 100,000 h100s could xai commit to always

34:10releasing a model that beats the current open- Source model the best open source model if they have one not necessarily your best model but if you can beat the current open source top of the list you can commit to releasing that for our benefit yes I will make that thank you okay number [Applause] seven that's good short answers hi people will speak more hi Elan my name is Kya and I'm founder of the startup chatting charge we Connecting People to People for eveve charging because they still feel lack of charging station around the Europe 15 second you have yes and my question is about the what is your prognose about the EV charging infrastructure for next 5

35:04years well actually I think that there's there's a lot of charging infrastructure you you can travel certainly anywhere in Greater Europe on a Tesla using Tesla superchargers and Europe actually has a lot of pretty good third party superchargers and you can travel anywhere in North America anywhere in in China Japan so I I think actually

35:32the the freedom to travel with the highspeed charging is solved in most parts of the world and will be solved everywhere number one uhoh number one we are sleeping hi alen my name is sh sorry sorry him eight yes sorry eight first my fault no worries go ahead hi Elon I'm Oliver V I'm running an AI safety group in ut I'm wondering how can I convince people about the danger of AI and could I chat with the people at xai safeties team shout out to user experience.org as well yeah

36:32I I I mean I do think this this there's some danger associated with digital superintelligence as I said earlier the I think the biggest issue is that it has to be trained to be rigorously truthful and it has to be trained with to be curious and I've thought a lot about AI safety for a long time you know one of the challenges you have with programming explicit morality into AI is that if you sometimes people

37:05call this the Waluigi problem if you program Luigi you can automatically invert that and create Waluigi bad Luigi so what what you cannot invert is the truth of is is physical reality so you can't invert the rules of physics you can't invert logic and so so I think that that that is the the better way to go so so really I think what Regulators should be concerned about is is the AI being rigorously truthful is it giving an answer that is most probably correct with

37:52acknowledged error I think that's the best move for for AI and that's that's what we're trying to do at xai now we still X is is a new company so there's not it still has a lot of catching up to do before it has an an AI that is competitive with with sort of Google Deep Mind and Microsoft open I so no I I think maybe towards the end of this year we will have that but it's there's not much point in you know trying to make it to

38:42to we just we just need to have an AI that is competitive is the first order of business if if it is not competitive which is it is not yet but I think it will be then you know things start to have a point at that like we need we need to only start worrying about safety when it is of comparable power to other AIS so anyway and and I just keep harping on this this point of the truth truth truth truth truth this is very important and sometimes people will say like well is there an objective truth I say well in a lot of cases there is if you're talking about physics there there is an objective truth or you can say there's a truth with with highest probability

39:27so you can say what is most probably true with acknowledged error so in physics people will say like this hypo this hypothesis appears to be true to a sort of five Sigma error type of thing so so it's like okay well if the experimental results are coming true to 5 Sigma then it's probably correct five five Sigma is lot so

39:58yeah anyway that's that's it's it's kind of like I guess a sort of a physicist approach to safety is yeah I'm I'm I'm repeating myself here but I can't emphasize enough that honesty is the best policy but there are there are strong forces to to for AI to not be honest if something does not is not an unfashionable thing to say then it will it you know it will be politically correct and that's lying thank you number one I'm sorry for

40:37hello alen my name is Shan I'm a student of University of batteries we found that a lot of job being replaced by AI do you worry about your job being replaced by AI if not why if your job was replacing replaced by AI what would you [Laughter] do well I mean we do get into some existential questions here in in a benign scenario in a benign scenario we probably none of us will have a job there will be but in scenario there would be Universal High income not Universal basic income Universal High income there would be no shortage of goods or services and I I think the benign scenario is the most likely scenario probably I don't know 80% likely for in

41:42my opinion the the the question will not be one of lacking goods or services you'll have everyone will will have access to as much in the way of goods and services as they would like the the question will really be one of meaning of how if you if the computer can do and the robots can do everything better than you then what does your life have meaning

42:16that's that's really the will be the question in the benign scenario and in the negative scenario well well bet R we're we're in deep trouble so now I I do think there's there there's perhaps still a role for humans in this in in that we may give AI meaning so if you think of the way that our brain works we've got the limic system which is our instincts and our feelings and then we've got the cortex which is thinking and planning but the cortex is constantly trying to make the lyic system happy so

42:57maybe that's how it'll be with AI which is the AI is trying to make our cortex happy which is trying to make our lmic system happy and maybe we are what give the AI meaning or purpose you know some kind of yeah I so but but I I do think that long term in a benignant scenario any job that somebody does will be optional like if you if you if you want to do a job as kind of like a hobby you can do a job but but otherwise the the AI and the robots will provide any goods and services that you want okay so you see there is a a future

43:46where no one will need to work it will be just passion so Pro that is the most likely outcome if if people are interested in reading some science fiction books that the most accurate portrayal of a future with super intelligence AI is was done by in Banks the the culture books of Ian Banks are are the best that that's probably the best envisioning of a future AI

44:28number two hello Mr M so first of all congratulations for your graduation of your kid and my question is do you think that I will impact the education of the future generation of kids mostly about that the values and the morals well I think the parents will still be responsible for values and morals I do think AI will dramatically affect education because the the AI is an extremely knowledgeable teacher very patient will be almost always correct and can tailor the lessons specifically to the child so it would be like each child has you know Einstein for a teacher so would be something quite profound I do worry that kids these days are being trained by social

45:29media and the social media AI algorithms are are basically dopamine maximizers so they're trying to maximize the you know maximize the amount of dop Meine that you get by watching the screen you know so that I would say I would urge parents to limit the amount of social media that children are able see CU they're being programmed by do doine maximizing AI so you

46:04understand that Elon is ready to sacrifice X to xai number three hi Elon priha shasa from Business Insider taking it to Tesla Tesla's had a bumpy few months looking at flagging sales at home stock market decline layoffs when you look back are there anything we can stop the question right now because I I I I don't think business ins sat is a real publication okay so let's move on to the next question all right thank you so next question number

46:41four hi Elon my name is shuk and I have a very specific question about neuralink do you think in future with the help of neuralink you will be able to use the human brain scans data to develop the most powerful in Creative AI the world has ever seen and if you like my question will I ever get a visit to space an X station in my life space ex station in my life I see 42 in the background there the answer to life the universe and everything actually which reminds me that another book worth reading is Douglas Adams hitch is Guide to the Galaxy that's a book of philosopy that is sort of kind of disguised as a book on humor or coming through as as

47:31humor so H where's worth reading H has gued the Galaxy I find that to be quite an inspiration and the name of the first Starship to go to Mars will be Heart of Gold after the ship in the H's guide so I the goal of SpaceX is to enable anyone who wants to go to Mars wants to go to the Moon to be able to go so that it is Affordable to a very large

48:05percentage of people on Earth that that's our goal which is essential in order to build a self-sustaining civilization on Mars to make life multiplanetary you obviously need to bring a lot of people to Mars and like I said that's I think that's a very important thing for the long-term Survival Of Consciousness because event eventually something will happen on Earth that will eliminate Life as we know it as happened with the dinosaurs and happened many times in the fossil record so you know hopefully it's not anytime soon but it will eventually there will be some sort of natural cataclysm that eliminates Life as We Know It On Earth even if we do not eliminate it ourselves

49:01so we really I think it's very very important to make life multiplanetary while we can and that means making it as affordable as possible to to move to Mars as well as the moon so before going to Mars we will still take a few questions because we are coming to the end number six

49:28hello thank five sorry hi so it's me again I'm talking for elod she is a coach and she would wanted to know if you would be ready to use your own intelligence because she has created a technique for coaching between neuroscience and quantum physics and she offers to do it very fast and in all confidentiality

50:0042 okay sounds good thank you okay go ahead okay but so we move to number six yeah hi so my name is Mara CE and founder of unlock my question will be on the African technological development last time I had a great convers conversation sorry with someone you may know Karim begar who had been nominated by Times magazine as an AI word influencer and talking about Africa and given the increasing potential on this on this continent in technology what initiatives collaboration or contribution do you

50:48foresee it's difficult to say I I do think Africa has a lot of potential and you know and I think one of the the things holding Africa back has been lack of internet connectivity which is what stall link is helping to solve you know with that people can learn anything and and and be sort of part of the Global culture whereas if if you're not connected then you're you I mean all you can do is like write letters or something so so connectivity is essential for Africa and and other places to be part of the global culture and participate in Ai and other activities

51:41yeah number seven Hi alen I'm Lou Nicola and I'm building a AI that is super authentic and precise for to have entrepreneurs grow and manage the business but I I'm super Enthusiast about Travel and Adventure and I want to ask you when do we go to Mars t-shirt we have seen it it's okay

52:10it's really going for it the it's good but it's okay 42 well I I think I think we we'll hopefully land the first Starship un crude on on Mars with within 5 years and also on the moon and I I I think we will have the first people on the moon or both the moon and I think first people on the moon probably within 5 years and and

52:45then most people on Mars probably within years maybe I think less than 10 years maybe seven or eight years you mean you will be able to land on the Mars in seven eight years yes with people yes wow that that that is what is for it's number eight hi el I have a quick question I would like to know in relation with electromobility what Innovative projects do mosque and Tesla I have in mind for the coming years thank you so do you want to disclose your secret here we are all waiting for something a breaking

53:44news yeah I mean I don't think we we don't want to do sort of product announcements you know simply as in Q&A that wouldn't that wouldn't make sense we we know we we do our product events as product events so but obviously some of the things people are aware of you know like Robo taxis or cyber cabs and the Optimus humanoid robot

54:18so those are very big deals if if even if that's the only thing you knew about Tesla you know that that will be a those two things will be extremely profound a large Fleet of autonomous vehicles that you know basically a generalized solution to self-driving and a generalized solution to a humanoid robot those two things are incredibly

54:51profound Elon we promised that we will be done at 6:30 our time it is 6:37 but we had the interruption because of the connection do you give us five more minutes let me just double check and I think I would like to if I can yeah sure sure on a second minute very short very short question very short answers

55:24ah very short question question El mus I'm Frederick years ago you said that Aran 5 rocket wouldn't have any chance to survive would you say the same about Aran 6 which will be launched in July and which is much more cost efficient thank you short answer any rocket that is not at least

55:53mostly reusable has no competitive chance so with with Falcon 9 the booster the main stage is is reused and the fairing on oone is reused so you have about 80% of the vehicle being reused and ref flown it it's simply not possible for an expandable vehicle to compete with a reusable vehicle

56:19and and I mean one can think of this in any other Arena if it was cars or planes or boats if one company was making a reusable plane or an 80% reusable plane and another company was making a plane that was Expendable that you could use only for one flight the company making the Expendable planes would have no business at all because you'd have to buy a new plane every time you take a flight somewhere so this is why I've said for many years that it is completely pointless to develop rockets that are not reasonable number two thank you Philip from quick chat I wanted to ask the following question people love Chad GPT but unilaterally hate most website

57:06chat experiences the bottom right corner ones what's needed to bridge that Gap I messaged you on X so we can continue this later I'm not sure I fully understand the question there there's you know I agree CHT is is great at chatting the I like I said my concern is that it is not rigorously in pursuit of the truth and that's that's a a major concern so

57:46that that that's it so Gro will be pursuing grock from xai will at least try its best to be rigorously pursuit of the truth and also we want to try to be the funniest AI so I think you know if if we all going to die at least we should die laughing so we're trying hard to make grck be the funniest AI Elon you have been extremely generous with your time it has been great you have been truful you have given profound answers I would like that we make a lot of noise to thank you thank you are hearing the noise this mean that next year you have to be in person we will give you the date long time in advance and before you leave I will ask you for one word your biggest

58:49hope one word your biggest fear well I know about one word I guess my biggest hope well I guess I could give you my biggest hope in one word which would be Mars that's going to be obvious so yeah Mars Mars would ensure that we that Consciousness survives for a long time my biggest fear in one word AI [Laughter] what a surprise yeah exactly okay guess both of

59:33them I think okay so thank you very much Alon you have been great as always one question only one question one question I one question no no that's all right thank you okay so thank you and next year in Paris before you leave before you leave I would like to thank you all to tell you that tomorrow will be a bigger

1:00:08day and there is more things to see to look to watch to discuss and for the people who are interested in funding you should know that BPI is having now a special event at stage one so bankk populares BPI France is having an event which is very specific stage one you can still join mercy good evening

1:00:47[Music] spe for [Music] look [Music] this [Music] show [Music]

1:03:42stre [Music] this for [Music] something [Music]

资源与出处

相关视频

同频道、同主题或相邻年份