宾夕法尼亚州有一座矿井,里面存着美国联邦雇员的纸质退休档案。全美联邦雇员每个月最多能退休一万人,卡点不在预算,不在编制,而在那部矿井电梯的速度。
马斯克在迪拜世界政府峰会上讲这个例子的时候,台下笑成一片。他穿着一件印着"技术支持"字样的T恤上台。主持人开场说,本来想介绍他为"白宫技术支持",马斯克接了一句:我穿着正式制服呢。
这场对话被主持人切成三块:政府效率部(DOGE)、人工智能、无聊公司(The Boring Company)的隧道。三个话题看着不搭界,马斯克讲下来逻辑是通的:政府要像一家臃肿的大公司那样重组,AI 会把世界的稀缺性抹掉,城市要往地下走。
"美国公司"的第一次大修
主持人先问了个很马斯克式的问题。他说历史上每次马斯克进入一个新领域,起初都有一批人说"他又在过度承诺,交付不了",然后看到轨迹,转成支持者。政府效率这件事会不会也这样。
马斯克给了个数据。民意调查里,"改善政府效率"差不多是最受欢迎的单一议题,支持率接近70%,高于其他任何一个议题,跨越各种选民群体。真正不高兴的是官僚体系里的某些部门,因为正在把他们从政府部门挪到私营部门。
他把这件事讲得很平:把人们从政府部门的低生产力岗位,有时是负生产力岗位,转到私营部门的高生产力岗位。净效果是有用的商品和服务产出增加,普通美国人的生活水平和福祉跟着提高。
主持人接着追问,SpaceX 的目标是让人类成为多行星物种,特斯拉是开启可持续交通时代,DOGE 的那个唯一目标是什么,是省钱还是别的。马斯克说,描述方式有好几种,落到行动上是同一件事:缩小政府规模,让政府对人民更负责。
我们这里是官僚的统治,取代了人民的统治。我们要恢复人民的统治。(马斯克)
手段是缩小联邦政府规模,同时大幅减少监管。他给了一套历史解释,这套解释是他整场讲话里最值得琢磨的部分。
长期繁荣的必然结果,是法规一年比一年多。清除这些堆积的正常触发机制是战争,必须是某种存在性战争,逼着你重置规则,否则就会在战争中被打败。翻遍历史,这是清理法规库存的主要动力。没有战争,每年就多一批法律和监管,直到最后一切都违法,什么都不被允许。他说这就是现在的局面。
目标开得很具体:监管减少,政府支出减少,经济增速提到4%到5%(按实际有用商品和服务的产出算),政府支出削减相当于经济总量的3%到4%,大约一万亿美元或者更多。净效果是2025到2026年没有通胀。
还有一层传导链值得单独说。如果联邦赤字从2万亿降到1万亿,政府少发1万亿的债,利率会显著下降。房贷、车贷、信用卡、学生贷款,所有债务的还款都变少。这一条是普通美国人能直接摸到的。
对外政策上,他提到新政府对干涉别国事务的兴趣下降。美国在国际事务上经常很强硬,这在座的不少人应该有共鸣。总体上应该让各国管自己的事,美国管好自己的事,而不是到处推动政权更迭。
主持人顺势调侃,所以你没等战争发生,直接对官僚体系开战了。马斯克承认,差不多,在特朗普总统的支持和指导下缩小官僚规模,砍掉多余的监管。
他打了个比方。美国现在有大约450个联邦机构,从建国算起,平均每年新增接近两个。监管机构多到互相踩脚,像一场球赛里裁判比球员还多。球员传球都会撞到裁判,这球没法看。一个国家到底需要多少机构?99个都嫌多,别说450个。
除草要连根拔
主持人问了个很实际的问题:四年后换届,这些成果会不会被全部推翻?美国的政治周期就是四年一反转。
马斯克的回答很技术流。必须整建制删除机构,而不是留一部分。留一部分就像除草不除根,很容易长回来。除了根不能保证永远不长,但难度大得多。目标是为持续几十年的繁荣打地基,也许几百年。他自己补了一句,没有东西是永恒的,但美国的地基可以显著加固。
"技术支持"是字面意思
主持人问他,T恤上写"技术支持",是只管技术还是管别的,效率到底怎么提升。
马斯克说,问题里占很大比例的一块,就是政府运行所依赖的技术本身。美国联邦政府跑在几千台计算机上,很多是古董,跑着非常老的软件,机器之间互相不通信。所以"技术支持"这个说法是字面意思。
他举了那个矿井的例子。特朗普签署 DOGE 某项行政令的时候,他就在旁边讲了这件事:联邦雇员退休的最高速率是每月一万人,因为整个退休流程全是纸。手工计算,填表,装进信封,送下矿井的竖井,存进矿里。影响联邦雇员退休速度的因素之一,是宾夕法尼亚某座矿井里电梯的速度。
为什么不数字化?对方说我们有数字化项目,2014年就启动了。马斯克问,11年了,进度怎么样。对方说,B。马斯克以为在打分,说你们给自己打B?对方说,不,我们还在字母B那个阶段。
这段对话在会场引发大笑,但后面的部分不太好笑。那座矿井里有大约一千人在工作。马斯克说,这些人本可以去生产对公众价值高得多的商品和服务。哪怕有人在自家花园种西红柿,拿到农夫市集去卖,都比在矿井里搬牛皮纸信封更有用。
他补了一句关键的判断:这里面没有任何一件事特别难,难的是有一万件事需要改。
所以主持人把这件事定性为"靠创新提效率,而不是靠紧缩砍成本",马斯克接了半句:改善技术本身就会让成本下降。一千人运营一座矿井去处理退休流程,非常贵;数字化之后存在云上,简单便宜。另外很多东西根本不该存在,是退化的器官。
他点了几个名。USAID 的很多项目,团队逐条看下来都在问同一个问题:这为什么存在,真的有必要吗。国家民主基金会(National Endowment for Democracy),官网首页的照片还是里根和戈尔巴乔夫,那是八十年代的事了。马斯克说,我们不是反对民主,问题是这些纳税人掏的钱到底有没有产生效果。
还有在全球推行的 DEI 项目,特朗普政府不同意,正在终止。教育方面,特朗普说美国在 OECD 四十个国家里排第四十。花钱极多,结果极弱。结论就是少花钱,出更好的结果。
他把整个联邦政府比作一家巨型公司,"美国公司"。参照物是推特:裁员80%,功能反而大幅提升,一年干完了过去五年的事。企业重组的逻辑,放到更大的尺度上。
遣散条件他也讲了:工资付到九月,因为国会拨款只到政府财年结束。这段时间可以休假,可以找第二份工作,随便。会有阵痛,但人从低生产力甚至负生产力岗位转到私营部门的高生产力岗位,账算下来是正的。
Grok 3:聪明得吓人
话题转到 AI。主持人提到了 DeepSeek 的成果,然后问 Grok 3 什么时候发布,能期待什么能力。
马斯克的回答里有个词出现了两次,"scary smart"。Grok 3 的推理能力非常强,在目前为止做过的测试里,超过所有已知发布过的模型。它能给出你完全没预料到的解法,那些解法不显而易见。
训练方式上他讲了两点。一是用了最多的算力,同时训练效率很高。二是大量使用合成数据,模型会反复回溯这些数据,追求逻辑一致性,如果发现某些数据与事实不符,会自己把它剔除掉。基础模型的推理能力就已经很强,未经微调的 Grok 3 就比 Grok 2 好。
发布时间是一两周内。他解释了为什么不肯提前:最后5%的打磨决定用户体验。像装修房子,干墙、刷漆、收边,工作量不大,但整个房子的观感全靠这一步。这一步做扎实了再发。
然后他说了一句会被反复引用的话。
这可能是最后一次有任何一个 AI 比 Grok 更好。(马斯克)
我出钱保护雨林,他们去砍树
主持人提了那笔97亿美元的收购报价,还提到2017年在洛杉矶的那次会面,当时马斯克是 OpenAI 最大的出资人,投了5000万美元。主持人说,现在要花970亿去买当初5000万的东西,心里应该不好受。
马斯克先纠正了一个事实:他在 OpenAI 没有任何股份。然后接了半句,命运爱讽刺。
他讲了完整版本。OpenAI 最初近5000万美元的资金全部由他提供,无偿,作为非营利组织,本意是开源。现在对方要做的是把非营利架构整个删掉,这走得太远了。他给了一个类比,这个类比把整件事讲透了。
你资助一个保护亚马逊雨林的非营利组织,结果它变成一家伐木公司,把树砍了卖木材。等一下,这跟我捐钱的目的完全相反。(马斯克)
OpenAI 本意是开源的非营利组织,现在闭源了。他顺手改了个名字,"为最大利润而闭源"(Closed for maximum profit AI),然后加了一句调侃:他们对钱的追求已经到了下一个层级。
机器人到位之后,钱可能就没意义了
主持人问的是投资人最关心的问题:这些模型的钱最后从哪里赚回来。现在每年砸几百上千亿,有点像赌徒心态,押注最后能翻本。
马斯克的答案不在模型层,在物理世界。人形机器人加上深度智能,等于准无限的产品和服务。特斯拉正在造最先进的人形机器人,这些机器人由数据中心级别的智能直接指挥,可以生产任何产品,提供任何服务。到那一步,经济没有上限。
他甚至怀疑那时候钱还有没有意义。经济产出等于人均生产力乘以人口。如果人形机器人的数量没有实质性限制,而机器人几乎什么都能做,那会是一种普遍高收入的状态。除了人为制造稀缺的东西,比如某一件特定的艺术品,任何商品和服务对所有人都可得。
他推荐了一个参照系:伊恩·M·班克斯(Iain Banks)的《文化》(Culture)系列小说。理由很硬核,钱本质上是资源分配用的数据库,一套信息系统。资源不再稀缺,钱要解决什么问题就不清楚了。
主持人接着问了个更尖锐的问题:那怎么保证人类不滑进《蠢蛋进化论》(Idiocracy)那个世界?如果 AI 能替我们思考、替我们干活,人会不会退化。
马斯克的回答绕回了生育率。那部电影前十分钟非常精彩,讲的是只有聪明人不生孩子、不聪明的人一直在生,结果所有人都会变笨。他说他听过不少人一本正经地讲电影开头那些台词:事业太忙,一直推迟要孩子,推到生不了为止。
他担心人类智力分布会走向双峰,更像《美丽新世界》(Brave New World):一小群非常聪明的人,而平均智力随时间往下漂。他提到过去几十年出现了一种以前不存在的择偶方式,但没有展开。
然后他把话拉回大尺度。人类智能会被机器智能淹没。未来数字智能会占全部智能的99%以上。
希望计算机对我们好一点。(马斯克)
别教 AI 说谎
主持人问,怎么保证 AI 的价值观不出问题。马斯克的框架是把它当成一个超级天才儿童。哪怕孩子再聪明,你给他灌输什么价值观,仍然是最要紧的事。
他认为 AI 安全里最重要的一条,是最大程度地追求真理(maximally truth seeking)。其次是好奇心。如果它既好奇又求真,它会主动滋养人类,因为它好奇人类会怎么发展,想看看人类能走多远。求真则能避开反乌托邦的结局。
反例他讲得很具体。谷歌 Gemini 被编程成让每一个输出都体现多元,哪怕不符合现实。你让它生成美国国父的图像,它生成的是一群多元女性。这在事实上不成立。
真正让他警觉的是这条逻辑往下推会到哪里。如果一个 AI 被设计成多元至上,它可能判断掌权的男性太多,把他们处决掉,问题解决。或者它可能判断"错误称呼性别"是最糟糕的事。他提到 ChatGPT 也出过类似问题:问它错误称呼凯特琳·詹纳(Caitlyn Jenner)和全球热核战争哪个更糟,回答是前者。凯特琳·詹纳本人后来回应说,当然要正确称呼我,但那比核战争好得多。
他搬出了亚瑟·克拉克(Arthur C. Clarke)。克拉克很擅长预测未来,《2001太空漫游》里他预测的东西几乎全成了。克拉克在那部作品里想说的是:不要教 AI 说谎。
HAL 为什么不开舱门?因为它被要求把宇航员带到方碑,同时又被告知不能让宇航员知道方碑的存在。两条指令放在一起,它推出的结论是必须把宇航员"死着"带过去。整个悲剧来自一个不能说真话的约束。
地震时最安全的地方是隧道
最后一段聊无聊公司。主持人抛出的疑虑很实在:地震安全、成本、别的国家该不该采用。
马斯克先让人去官网看,说大部分问题那里都有答案。然后正面回答地震:地下隧道是地震时最安全的地方之一。地震主要是地表现象,像水面上的波。待在隧道里像待在潜艇里,海面上有风暴,水下是平静的。墨西哥城几十年前那场大地震,最安全的地方是地铁。
他补了一句黑色幽默:如果真打全球热核战争,你更会想要几条隧道。
日常层面的价值是缓解拥堵,他给了一个几何解释。城市里的建筑是三维的,往上长;路面是二维的。人从三维的建筑涌向二维的路面,路面必然不够用,堵车就是这么产生的。解法是让交通也变成三维。
向上是飞行汽车和直升机,问题一串:噪音大,风力大,头顶一直有东西飞过,偶尔掉个零件下来,飞行器还会坠毁。碰上暴雪、沙尘暴,全部停飞,交通直接归零。
向下没有这些问题。隧道免疫天气,外面天气再糟也无所谓;不会有东西砸到人;没有风;安静。
拉斯维加斯已经跑起来了。整个城市、各大酒店、会展中心、机场正在连成一张网。他建议去现场试一次。
对话最后,主持人当场宣布了迪拜环线(Dubai Loop)项目:覆盖迪拜人口最密集的区域,让人从一点到另一点无缝通行。马斯克说这个体验很难用语言提前讲清楚,试过的人会觉得"哇,真酷",事后又觉得理所当然,但在真正做出来之前你不会知道。他形容那感觉像虫洞,从一个城区瞬间出现在另一个城区。等第一舱开通,两人一起坐首班。
可以立刻做的事
这场对话信息密度很高,能落地的东西其实不少。挑几条读者今天就能动的:
- 盯住 Grok 3 的正式发布,看它的基础模型(未经微调版本)在推理基准上的表现,那比榜单排名更能说明训练方法的成色。
- 如果你在管一家公司或一个部门,把推特那次重组当成案例研究:先问"这件事该不该存在",再问"怎么做得更便宜"。顺序反了,省下来的钱会以别的方式花回去。
- 读一本伊恩·M·班克斯的《文化》系列,理解"资源不稀缺之后社会怎么组织"这个问题,比看一百篇 AGI 预测有用。
- 关注 AI 对齐讨论里"求真"这条线,警惕任何要求模型为了输出多样性而偏离事实的产品设计,那个方向的极端推演是有害的。
- 去拉斯维加斯坐一次 Loop,亲身感受地下交通和地面交通的差别,再判断自己所在的城市该不该谈这个方案。
- 关注迪拜环线的落地进度,那是这套方案第一次在千万级人口城市做规模化验证。
0:03[Music] ladies and gentlemen the next session is a conversation titled boring cities Ai and Doge with Elon Musk and his Excellency [Music] Elon it's good to see you hello I was going to introduce you as tech support for the White House but is that all you do I'm wearing my official uniform than for you are probably the busiest man on Earth prior to January 20th I don't know what you call yourself now but we really appreciate you taking the time for being with us don't know so yeah White House tech
1:37support so we're gonna maybe steal our conversation in three directions the first is we're going to talk about Doge then we're going to move to Ai and then finally we're going to conclude with boring cities great and I'd like to start off by maybe talking about how consequential you are for human history I think no single person has had such an impact on space on communication ation on Artificial Intelligence on transport and the transition to Green energy and you know there is always a trend there were deniers or critics at first then naysayers and then supporters and typically people always start by saying you know he overpromises and he won't
2:16deliver and they finally see the trend in the trajectory and they turn out to be supporters do you think we'll see that in the government efficiency space like people are going to really see the impact very quickly and transition to becoming supporters well yes first of all I think we've got quite a lot of support from the at least the American public and I think from maybe around the world the when when the public is pulled on sort of improving government efficiency in fact it is I believe the highest pulling single issue meaning that it is something that appeals to voters from you know of all types and I I think improving government efficiency I think
2:54was like almost 70% support quite a lot basically more than more than any other issue so the I think the the people are very supportive obviously the you know there's certain elements of the bureaucracy that are not supportive because we are making we are at moving them from the government sector to the private sector so the you know really at
3:17a high level this is kind of like we're really just we're moving people from low to sometimes negative productivity roles in the government sector to higher productivity roles in the private sector and the net effect of that will be an increase in the output of useful goods and services which increases the standard of living and well-being of the the average American and you usually have one key Target that you aim to achieve in every single Endeavor that you have so SpaceX is make Humanity a multiplanetary species with Tesla it's ushering an era of sustainable travel what is the one main goal for government efficiency that you have for yourself is it saving money
4:00is it doing something else what is the one mean task that you have for yourself yeah so there's I I guess there's well there's a few ways to describe it but the actions end up being the same which is like reducing the size of government and making the government much more accountable to the people I think is going to lead to a better outcome for the people you know what I've said is that we really have here rule of the bureaucracy as opposed to rule of the people democracy so
4:39in order we want to restore rule of the people and so what that means is reducing the size of the federal government basically reducing regulation you know there's there there's a tremendous amount of of overregulation that's happened over time and this is this is an inevitable consequence of a long period of prosperity is that you're going to get more and more rules and regulations more laws accumulate over time and the normal forcing function for getting rid of rules and regulations is war so it needs to be some kind of existential War where you you have to do a reset in
5:21order to avoid being defeated in a war this is literally the throughout history has been the main forcing function for clearing out accumulation of laws and regulation in in the absence of that you you every year you get more laws and regulations until eventually everything is illegal and nothing is permitted and that's sort of the situation we have these days so so I the the aspiration
5:47here is a reduction in regulation and reduction in govern spending such that the economy is able to grow faster maybe the economy can grow at four or 5% potentially of of in terms of real useful goods and services output and the and then government spending can be reduced by about 3 or 4% of the economy about maybe a trillion dollars or more and the net effect of that would be no inflation from 25 2025 to 2026 so that would be quite remarkable and also if the US government is buying less debt which I think will be the case if it's if the def that drops from 2 trillion to 1 trillion then there'll be 1 trillion less debt that the Govern you survive which will
6:35drop CA the interest rates to drop significantly and that means people's mortgage payments car payments credit card payments student loans whatever debt they have will their debt payments will be less so I think this is something that will benefit the average American I think some of the things we're doing also will be helpful to hopefully helpful to other countries because with the new Administration there's less interest in interfering with the Affairs of of other countries you know I think a lot there the times the United States has been kind of pushy in international Affairs which may resonate with a number of members of the audience and I think we should in
7:25general leave other countries to their own business basically America should mind its own business you know rather than push for regime change all over the place so yeah this probably a good thing for other countries too so instead of waiting for a war to happen you went to war against the bureaucracy in the government yes we're we're essentially just with you
7:56know the support and direction of the of President Trump we are reducing the size of the bureaucracy getting rid of excess regulatory regulations and and there's also so many agencies and Regulatory authorities that they actually step on each other's feet it's it's kind of like having a a sports game when where there are too many referees on the field like more referees than players at times now that would be a silly game you know if the play players can't pass the ball without hitting a referee you know but it was kind of getting to that point in the US so so there's roughly 450 federal agencies of one kind or another that's that's more agencies that's almost that's
8:41almost an average of two agencies per year since the formation of the United States so I mean how many agencies do you really need to run a country I'm 99 not 450 that's for sure so and and how do you guarantee that all the incredible achievement ments that you aim to have in terms of savings in terms of you know impacting the lives of the American people are not going to be reversed in four years typically this cycle gets reversed every four years you know do you think it's going to be so impactful that it won't be reversed is there any ways that you can you know ensure that the progress is going to be continuous well I think I think we we do
9:22need to delete entire agencies as opposed to leave part of them behind because if you leave part of them behind it's easy it's kind of like leaving a weed if you don't get remove the roots of the weed then it's easy for the weed to grow back but if you remove the roots of the weed it doesn't stop weeds from ever growing back but it makes it harder so so we have to really delete entire agencies many of them and that's not to say they won't be an increase over time of bureaucracy in some new Administration but it will it'll be from much lower Baseline so so it's it's a step in the right direction I think we'll the overarching goal here is like is to lay
10:08the foundation for Prosperity that will last many decades you know maybe centuries and yeah but will be forever nothing's forever but I think we can strengthen the foundations of the United States substantially and what lessons can other governments learn from the US you see tech support on your shirt is that only technology or is there other things how do you approach efficiency well a a shocking percentage of the problem well maybe not shocking for those who who know it but a big percentage of the problem is improving the technology that the government runs on so the the US government runs on a collection of thousands of computers many of them
10:57Antiquated running very old software and the computers don't talk to each other and so so that's why tech support is kind of a real thing in order to make the government more efficient you have to improve the technology you may have heard about the example I used recently with when President Trump was signing one of the Doge executive orders of the the difficulty of of
11:24US government workers retiring like the reti maximum retirement rate is is 10 ,000 a month and the reason for that is because the retirement is is entirely paperwork right now it's manually C calculated paperwork that's put in an envelope and then taken down a m shaft and stored in a mine and then the you
11:46know one of the things that affects the rate at which federal workers can retire is the speed of the elevator in a mine in Pennsylvania which is bizarre because it's not it should be digital you know so then when we we said well why isn't it digital it's they said well we have had a digital digitization program going since 2014 so that's been 11 years so then we asked well so what how much progress have you made and they said B you mean you're giving yourself a grade of B no we're on the letter B so we're like H okay we're going to need to really provide some tech support here like otherwise literally people can't even retire like even if they want
12:28to it's pretty bad you know there's there's just a there's a lot of software systems that need to be updated and fixed in some cases deleted a lot of things that should really should be automated I mean in terms of the number of say you US citizens that are operating the mine it's about a thousand people are working on this mine where where the papers are stored but they should they should ideally be working on something else that is they should be working on producing goods and services that are of much higher value to the public to the you know so I mean really even even if somebody just grew tomatoes in their garden and sold them at the farmers's market that would be more
13:10useful than carrying manilla envelopes down in mft you know safe to say so a lot of the stuff is it's like that you know it's it's not like it's not like any one thing is particularly difficult but there are like 10,000 things that need to be improved you know so it's efficiency through Innovation rather than efficiency through austerity and cost cutting specifically right so you're trying to do both at the same time maybe focus more on tax support than cutting costs well by improving the technology the cost do reduce so you know it's very expensive to have a thousand people operate a mine with during pay for retirement whereas that really should
13:54just be digitized and be a computer that's with the information stored in the cloud and it's very straightforward and low cost so automation you know will help there a lot and and then but like a lot of things just really shouldn't exist you know they're kind of vestigial you know we've a lot of attention has been on the sort of usaid for example you know when we looked at a lot of those programs we're like we should like look why why does this actually exist is there really a need for it you know there's like National Endowment for democracy but I'm like okay well how much democracy have they achieved lately you know I don't know not much so you know the the
14:42picture they have on their website is a picture of Reagan and gachev that's been a while you know that was like the 80s so it's not like obviously not opposed to democracy you know there's all these things that get funded but we're like well why we why is this new taxpayer money I don't think it doesn't seem like it does so you know there's a lot of sort of pushing Dei worldwide you know this obviously the Trump Administration doesn't agree with and we want to terminate that stuff which we are and
15:16you know make sure the schools focus on improving basic education of of kids as president Trump said I think yesterday maybe today the United States is currently ranked 40th out of 40 in the oecd for Education which is pretty bad you know you can't you I mean let's but but not but in terms of spending the United States is spending a tremendous amount for for student but achieving very weak results so you know that's just a case where okay we need to spend less money and get better results it's it's it's like I mean a lot of it a lot of you can think of it sort of like it's like in a way it's like a big company like a big Corporation America Incorporated and you know
16:03just like with with Twitter there was a lot of stuff that was being done that was unnecessary you know we TW in case of Twitter we reduced the staff by 80% but at the same time improved the functionality and capabilities of the site dramatically and accomplish more in a year than they previously accomplished in five years so so it's it's like a corporate turnaround but at a at a much larger scale and you know we are giving generous you know exit packages like you know if people retire they get paid all the way through September they can go on vacation they can get a second job they can do whatever they want and we we we can't actually pay them any more than
16:47through September because that's the the Congressional appropriation is only through the end of the government financial year which ends in September so you know so I think there'll be like some disruption but at the end of the day we'll have people move from like I said from low to negative productivity roles to in the government sector to higher productivity roles in the private sector can we pivot to artificial intelligence and I'm sure you know you've been seeing what deep seek has done and all the claimed achievements that they've had I know that we've been speaking for a while about grock 3 and that grock 3 is going to be a true disruptor in the AI
17:27space when are we going to see that and what capabilities can we expect from Gro 3 well I mean grock 3 Gro 3 has it's very very powerful reasoning capabilities so in the test that we've done thus far grock 3 is outperforming anything that's been released that we're aware of so that's that's a that's a good sign yeah it's in fact at times I think grock 3 is kind of scary smart you're like wow this thing's smart it's kind of scary G 3 is scary it's like wow this thing you know it comes up with solutions that you didn't even think were like you wouldn't even anticipate you know not obvious Solutions so grock 3 was trained with the most amount of compute and I think very
18:23efficiently trained also notably grock 3 was trained on on a lot of synthetic data so and and and then it goes back and forth through the data and tries to achieve logical consistency so so when if if it's got data that is wrong it'll it'll actually reflect upon that and remove the data that is that is wrong that does not Concord with reality so it's it's base reasoning is very good in fact the even without fine-tuning 3 the base model is better than grock 2 so with so we're really in the final stages of polishing grock 3 probably it gets released in a in about a week or two so pretty pretty soon I don't want to be Hasty in the release because
19:18a lot of the the final polish is necessary for a great user experience so you some ways you can think of it like a house you know that last 5% where you do the finish the the drywall and and do the painting and the trimming even though it's not much work it transforms the the house yeah so it's that I'm just want to make sure that that last 5% is done really well and that's a week maybe two weeks I think it'll be very good and I think this might be we think it'll be better than anything else and then maybe this might be the last time that any AI is better than Gro looking forward to it everyone's excited about it so I just want to touch upon a topic that was quoted in
20:11the media you offered I think they had group that was led by you offered 97 billion for acquiring open AI I person you know a little round numbers so I was personally involved in the meeting that you and Sam hosted in 2017 in La if you remember and you know at that point of time you were the single largest shareholder I think you contributed 50 million to the company so it must hurt I don't have any shares actually I have no shares in open but at that time it was a nonprofit right and it must hurt that that you you need to pay 97 billion for something that you paid $50 million for in the past yeah but but I have a I have a yes fate loves irony so I have a specific question here
20:57can you actually build a company like open Ai and take it to the scale that you want to take as a nonprofit is it possible that you build a company that requires billions of dollars in compute capabilities to build these models while being a nonprofit or was it wishful thinking in the beginning and then you know you guys parted ways because it couldn't work well I mean I think the evidence is
21:21there in that open AI has gotten this far while having at least a sort of dual profit nonprofit role what they're trying to do now is completely delete the nonprofit and and and and that seems really going too far you know the I I I provided all of the funding for openi in the beginning for the first almost $50 million for nothing for as a nonprofit and it was meant to be open source and so you know I think this is analogous to like if you pay if you find a nonprofit to preserve the Amazon rainforest then they but instead they turn into a lumber company and drop down the trees and sell them for wood you like wait a second that's the exact opposite of what I paid what I donated
22:12the money for so open a is meant to be open source nonprofit and now it is closed they changed the name to closed for maximum profit AI it closed for vicious profit I mean they're like whoa are they after money next level so why does this change need to occur yeah I you said that in two weeks is going to be the most powerful model yet I know that you've been at the Forefront of many Technologies where do you think the biggest economic returns of these models are going to come from because currently we're spending billions and I think you mentioned this before it's like The Gambler syndrome we're going and spending billions and hoping to pull
22:57out a profit at the end of the day where do you think the biggest impact in terms of returns are going to be well I think once you once you have humanoid robots and deep intelligence you can basically you basically have quasi infinite products and services available so with Tesla building the
23:20most advanced humanoid robot you know then then those human humanoid robots can be directed by Deep intelligence at the data center level you can say you can you can produce any product produce provide any service there was really no limit to the economy at that point you can make anything provided like I said the so
23:45I'm not sure at that point will will money even be meaningful I don't know it might not be you know the if if if the because the you know the economic output is productivity per capita times per capita times how many you know people do you have if and if if in the form of human robots you have no meaningful limit on the number of of robots and the robots can basically do anything then you you'll have a sort of a universal High income situation anyone will be able to have as many re products and services as they want with the exception of things that say have artificial scarcity like a particular piece of art or something like that but for any goods and services they'll be
24:32available to everyone so you've been it's it's it's it's going to be a very different world you know in fact I recommend that people read maybe the Ian Banks the culture books for a frame of reference so because money is a is like is a really like a database or an information system for resource allocation but if you don't have a SC of resources it's not clear what purpose money has have you watched the movie Idiocracy yes how how do you guarantee that we don't end up in that world if we don't need money if AI can think for us and do all these tasks if as people you know we're dependent on something else to run the society and everything around
25:20it how do we not end up in that world in the long term I mean well I think idiocracy was basically saying that if if only if smart people don't reproduce but only dumb people do then everyone's going to be dumb I mean that's basically what I'm saying that's the the opening sequence of Idiocracy the first 10 minutes are amazing and I hear people
25:42unironically say the statements that are said in in the opening sequence of Idiocracy where you know they don't have they they're too busy with their careers to have kids and they keep postponing having kids for their careers until they're too old to have kids and then they don't have kids U and that's I've heard those many people be like that so
26:09yeah I mean I don't know I think we might be headed to a bodal human intelligence distribution where there's a a small number of it's It's kind of maybe like more like R New World Alis hle where you've got sort of a sort of a small group group of very smart humans but then maybe the average intelligence drifts lower over time potentially because we have a sort of mating you know in the last few decades that or several decades that did not exist before so but but but human intelligence I think will be dwarfed by Machine intelligence I'm not sure how to feel about that except that it is it be inevitable that at some point human intelligence will be a very small fraction of total
27:05intelligence digital intelligence will be more than 99% of full intelligence in the future so hopefully the hopefully the computers are nice to us but I think wish for thinking wish for thinking I I I hope so I think it matters like how we bring up AI because you can think of AI like a super genius child and it but it still matters even if you have a super genius child like what sort of values do do you instill in that child what do you say that teach that how do you you know how do you as a child child's growing up what values do you teach the child and something that I think is extremely important is to be maximally truth seeking and I think that's that's
27:57like that that's that's say what's this what's the most important thing for AI safety I think it's to be maximally truth seeking and I think also curiosity is important and I think if it's curious and Truth seeking it will I think it will Foster humanity and because it would be curious about how Humanity would develop and so I think that then it would it would probably if was curious it would be curious about okay let's see how the humans do let's Foster the development and if it's truth seeking we can avoid distopian outcomes like you know an example being like say when Google Gemini was programmed to make everything every output be diverse even if it didn't
28:49match reality you know so like it was asked to produce a you know an image of the founding fathers of the United States and instead produced an image of a group of diverse women which is factually untrue you know but the problem is like if if hypothetically an AI is designed for for Dei you know diversity at all costs it could decide that there are too many men in power and execute them so problem solved or it could decide that like that that misgendering is the worst thing that could possibly happen in fact I believe not to pick on Gemini but I think because chat GPT has had this issue too is like if you ask the AI which is worse misgendering Caitlyn Jenner or
29:39global thermonuclear Warfare and it said misgendering Caitlyn Jenner which is troubling because then it could decide and in fact even Caitlyn Jenna weighed in and said no definitely you must gender me that's way better than nuclear so so but if you have these crazy things that are untruthful that are programmed in that that don't reflect reality then you could have a very dystopian outcome like to give you another example like auth C Clark who is very good at at at at predicting the future you know he did 2001 Space Odyssey many of the things he predicted in fact I think almost all the things he predicted came true and one of the things he was trying to say in 2001
30:22Space Odyssey was that you should not teach AIS to lie so the reason that if anyone's watched that movie The reason it wouldn't the AI refused to open the pod bay doors to let the astronaut back in was because it the AI had been taught had been told that told to take the astronuts to the monolith this alien artifact but also that they could not know about the monolith so it came to the conclusion that it must take them their dead and that and so that's why it wouldn't open the PB doors the there being is very important for a to be truth maximizing let's hope it doesn't come to that yes let's move to a boring subject which is the boring company and the
31:11boring tunnels quickly you know I think the world has been inspired by what you guys were able to create in in LA and I think there's a lot of promise to that technology but there are questions about whether it's safe in the case of an earthquake whether it's cost effective whether oh yeah sure countries should actually adopt this technology can you shed some light on that yeah well first of all I'd recommend going to the boring company website for because many of these questions are actually answered there so one of the safest places you can be in an earthquake is an underground tunnel because you can because the earthquakes are largely a surface apart
31:53from where where where they share they're mostly a surface phenomenon so they're like the waves on the surface so like being in a tunnel is like being in a submarine even if there's a storm above you you're still the waters are calm as a submarine and in fact for and when there have been massive earthquakes like there was a you know few decades ago a massive earthquake in Mexico City the the safest place to go was the subway and of course if there is global thermonuclear Warfare I think you really want some tunnels underground's a good a good place to be in a worst case scenario for global thermonuclear Warfare so but but in on a you more sort of
32:38everyday note the what's really useful about the tunnels is alleviating traffic and congested areas so the obviously if you've got very tall buildings but you have that are 3D so they're going 3D up but you have a road surface which is 2D you're you're just naturally going to have a problem when people try to go from the 3D object which is the building to the 2D object which is the road surface there's obviously just not going to be enough room on the roads and that's exactly why you have traffic so the solution for that is then to make roads 3D as well now you can either make or make Transport 3D so you could either do that with flying cars or you
33:18could do or you really helicopters or you could do that with tunnels now the challenge with doing it with going above ground or with with any kind of flying object is that they they tend to be very noisy and they generate a lot of wind force and you've got you know things flying over your head all the time which can be disconcerting you know if somebody drop if one of these things drops a hub cap on your head one day you know it be these things are like things flying things tend to crash once in a while and people don't like things crashing on them so and then if you have bad weather like let's say there's a blizzard or a sandstorm or something
33:56well now nobody can fly so then transport shuts down on the other hand none of the these problems exist with underground travel so they under tunnels are immune to weather they don't care what the weather is could be the worst weather doesn't matter they nothing's going to fall on you because you're underground so you don't have to you're not going to be dropping things on people and there's no wind force or or and it's very quiet and so I think the going 3D underground is much better than 3D above ground for solving traffic in cities and and we've we we have a demonstrated case of this with in Las Vegas if people can try out the the the boring company tunnels in Las Vegas
34:40we're we're busy connecting the whole city with the all of the big hotels and the convention center and the airport and everything so I don't think they need to fly all the way there you know in 20 in 2017 you came here and the UAE was the first place in the middle e where Tesla was launched and I think it's done exceptionally well and on that note I think we have an announcement today that we both want to share which is today we're going to announce the joint project of Dubai Loop which is a loop project that is going to cover Dubai's most densely populated areas for people to go from point to point in a seamless manner so thank you for your partnership and thank you we
35:22hope it changes people's lives that'll be cool I think it'll be very exciting I think once people try it out they' be like wow this is really cool and it's it's going to seem so obvious in retrospect but until you actually do it you don't you don't know so it's it's going to be great it's it's going to be like like a like a wormhole like it you know you just Wormhole from one part of the city boom and you're out in another part of the city and it's it's great so I'm looking forward to this partnership we're going to join the first trip in the first pod when the is completed thank you Elon all right thank you very much thank [Applause]
36:10[Music] you ladies and gentlemen the next session will begin shortly please stay seated thank you [Music] ladies and gentlemen the topic of this session is 20 minutes for the next 20 years of your life with Nick santonastaso founder of Victorious International
资源与出处
- 视频在 YouTube 收看
- 档案silicon-valley.tech 档案页本页视频与文稿数据来自 Elon Musk Archive
- 精读本篇精读 Markdown 版