播客·中文精读

马斯克如何解决问题 | Lex Fridman 播客

Elon Musk's approach to problem-solving | Lex Fridman 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

全世界最聪明的一批工程师凑在一起,最容易犯的错是什么?

马斯克的答案不太客气:他们会把一个根本不该存在的东西,优化到极致。

这句话出自 Lex Fridman 播客的一次对谈。马斯克在那里把自己那套「第一性原理算法」(first principles algorithm)从头讲了一遍,从质疑需求一直讲到自动化。他说自己把这五步当咒语念,因为哪怕念了这么多年,栽跟头的次数还是多到他不想回忆。

五步听起来平平无奇:质疑需求、删除、简化、加速、自动化。每一句都像工程师入职手册的第一页。

真正有分量的地方在判断标准。他给的每一条都能立刻拿来检验自己有没有做对,中间没有模糊地带。这套东西后来被用在孟菲斯那座十万张 GPU 的超算集群上,抽象的方法论就此变成一堆正在发热的电缆,和一台正在抖动的电网。

第一步,默认递到手上的需求是蠢的

马斯克的原话是,需求这个东西,无论出自多聪明的人之手,都必然带着一定程度的愚蠢。

你要做的第一件事是减少需求的数量。不管给你需求的人有多聪明,那些需求在某种程度上都是蠢的。(马斯克)

为什么必须从这一步起手?顺序错了,后面全废。

你必须从这一步开始,否则你可能给一个错误的问题找到了完美的答案。(马斯克)

「给错误的问题找到完美的答案」,这句话能描述出很多高效团队的真实处境。执行力无可挑剔,工程能力一流,交付速度快得惊人,最后做出来的东西没人用,或者压根没必要做。

第一步的目标是把问题本身变得尽可能不那么错。急着找答案的人,容易找到一个漂亮的错误答案。

还有个细节容易被滑过去:马斯克说的是「减少需求的数量」,他没有说「把需求想得更清楚」。数量优先于质量。先把一堆可有可无的需求砍掉,剩下的那几条才配进入讨论。

现实里的需求是怎么长出来的?某个部门要一张报表,某个客户抱怨了一句,某个竞品上线了一个功能,某次事故之后加了一道检查。每一条单独看都站得住,叠在一起就成了一头谁也不敢碰的怪物。没有人回头问:这条需求今天还成立吗?提它的人当时基于什么信息?那个信息现在还准不准?

马斯克的做法是默认它们都错,逼着提需求的人重新论证一遍。这个过程让人不舒服,因为大部分人把「被质疑」理解成「被否定」,一位工程师花了两个月做的功能被一句「这个能不能不要」问住,情绪很难不冒出来。可如果不这么干,团队就会在错误的方向上跑得飞快,而且跑得越顺越危险,因为没人有时间停下来看看方向。

第二步的判据:删到有人喊疼

需求理清楚之后,第二步是尝试删除。删掉那个零件,删掉那道流程,删掉那个环节。

这一步听起来极其显然。马斯克说人们经常忘记去做,一个步骤存在得越久,越没人想过它还可以不存在。

他给出的判断标准硬得很:

如果你没有被逼着把删掉的至少 10% 加回来,说明你删得不够。(马斯克)

正常的直觉是:删了一些东西,后来没出问题,说明删对了,任务完成,可以收工。马斯克把这个判断翻了个面。删完一件都没加回,意味着删得太保守,里面还留着一堆不该留的东西。

在这套标准里,「没有被追着加回来」读作一条证据:删得还不够狠。

推论很实在。做删除决策的时候应该主动多删一点,把「至少十分之一会加回来」当成预算内的成本,提前记在账上,别当成事故。这件事得说出来,还得当着大家的面说。

他在团队里就是这么操作的,提前打招呼:

我跟他们提前说清楚:对,我们删掉的某些东西,后面会加回来。有些人会被这个说法吓到。(马斯克)

Lex 在现场见过这种反应,说能看见大家脸上的痛。马斯克回答得干脆:确实,所有人都会痛那么一下。

删东西为什么疼:边缘系统在替你踩刹车

要理解这套方法为什么难落地,得先看人的大脑在删除这件事上怎么运作。

马斯克的解释是,边缘系统(limbic system)会系统性地误导人。

我们会记住那些删掉之后又需要的东西,而且记得特别清楚,带着一种刺痛感。人们会记得三年前忘了装某个东西,结果惹了麻烦。于是他们过度纠正,往里塞太多东西,把事情搞得过于复杂。(马斯克)

这套机制好理解。做对一百次删除,没人会记住。做错一次删除,后果会长期扎在脑子里。人对损失的记忆强度远高于对收益的记忆强度,这是神经系统出厂就带好的设置,改不掉,只能识别出来之后硬扛。

组织的默认状态于是变成加法。每一次事故加一条规则,每一次投诉加一道流程,每一次失败加一层审批。没有人负责减法,因为减错了要担责,减对了没人记得,中间没有任何奖赏。

马斯克把这种应对称为「皮层对边缘系统的压制」(cortical override to limbic instinct)。理性能算出多删是对的,情绪会拽着人不让删,只能靠前额叶硬压下去。

他还有一句补充,把这件事的性质点得很透:

如果你保守到从来不需要把任何东西加回来,那你显然留了一大堆根本不需要的东西。(马斯克)

这句话可以当尺子用。翻一翻手上那个流程,过去三年有没有删掉过任何一个环节?一个都没有的话,多半说明这套流程从来不许被删,跟它完不完美关系不大。

优化、加速、自动化,顺序颠倒就是白干

删完之后才轮到简化和优化。马斯克把这一步排在第三位,没有排在第一。

第三件事才是尝试去优化它、简化它。这些话我说出来的时候听起来都很显然,但我犯这些错误的次数多到我不愿意记。(马斯克)

优化必须排在删除后面,理由是优化一个不该存在的东西纯属浪费,而且比浪费更糟。被优化过的东西看起来更专业、更精致、更不可或缺,于是更难被砍掉。一个能自动跑批的冗余报表,比一张手工填的冗余报表活得久得多。

第四步是加速。马斯克的说法是,任何一件事,无论你觉得能多快,它都能更快。

加速有个严格的前置条件:

你不应该在删除和优化之前去加速。否则你就是在加速一个不该存在的东西,这很荒谬。(马斯克)

第五步才是自动化。

说到这一步,他讲了自己最惨的一类经历,把顺序走反过无数次:

我好多次是倒着走的:我先自动化了某个东西,然后加速它,然后简化它,最后把它删掉了。我干这事干到烦了。所以我才有了这套咒语。(马斯克)

自动化之后再删除,那种感觉相当难受。Lex 接了句「那一定很痛」,马斯克回:

对,感觉就是,哇,我在这儿浪费了好多功夫。(马斯克)

这段自嘲提供的信息量最大。他讲的这套流程来自反复被打脸之后的总结,并非一个从一开始就想明白的理论。他甚至说自己得把它当成咒语来念,因为光靠记会忘,尤其是在赶进度、压力大的时候。

把五步连起来看,它约束的是人的冲动。工程师的本能是优化,管理者的本能是自动化,公司的本能是加速。这三件事都很爽,都能带来看得见的进展,都能写进季度汇报。删除不爽,删除看不见成果,删除还会挨骂。

所以删除被放在第二位,三件最容易让人上瘾的事全部排在后面。这套算法的重心在拦人,拦那些聪明、勤奋、方向跑偏的人。

孟菲斯的十万张卡,卡在电上

播客里 Lex 提到自己刚去过孟菲斯,看了那座超算集群。他说那里有一种极强的驱动力,所有人都在简化流程、理解流程、持续改进、不停迭代。

马斯克的回应很冷静。

它还没跑起来。所以我不想现在就开香槟。(马斯克)

他接着说几小时后还有一个跟孟菲斯团队的电话会,因为「我们遇到了一些功率波动的问题」。

这句话把对话从方法论拽回了工地。一座几周内拔地而起的超级计算集群,卡住它的东西既不是芯片供应,也不是算法,是电。

Lex 问现阶段最难的部分是什么,马斯克只回了几个词:极端功率抖动(power jitter)。

抖动从哪来?马斯克解释得很形象。这类集群做的是同步训练,所有机器在毫秒级别保持同步。

这就像一支交响乐团,而乐团能在亚秒级从最响切换到最安静。然后整个电力系统就疯了。(马斯克)

具体到数字上:

如果电网突然看到每秒好几回、每回 10 到 20 兆瓦的巨大跳变,这完全不是电力系统预期会看到的东西。(马斯克)

一座大型工厂的用电曲线是平的,机器开起来就一直转。一座超算集群在训练时的曲线是尖的,成千上万张 GPU 在同一毫秒开始计算、同一毫秒停下、同一毫秒交换数据,然后进入下一轮。整个集群像一个呼吸急促的巨型生物,每秒好几回把十几兆瓦的负荷甩出去再收回来。

电网的设计假设里没有这种用户。

第一性原理在现实里就是这么运作的。五步算法可以背得滚瓜烂熟,需求层面可以删得干净利落,最后仍然会撞上一个物理约束:电力的瞬时响应速度跟不上算法的同步节奏。这种约束删不掉,只能设计储能、调整训练节奏、改供电方案,一条一条啃。

他还顺口提了一句训练终于跑起来的时间,上周一凌晨,大约 4:20,随后补了句可能是 4:22。Lex 笑说宇宙又开玩笑。这个时刻在美国亚文化里有点特殊含义,两个人心照不宣。

玩笑之外有个细节别放过:他没有说「我们搞定了」,他说的是「它还没跑起来」。一座被媒体反复报道的超级集群,在当事人嘴里仍然停在未完成状态,墙上仍然挂着一个具体的技术问题,几小时后还要开会。

瓶颈是线,而大脑本身也是一捆线

Lex 问到一个更基础的问题:这座集群到底难在哪。马斯克的答案有点出人意料。

大型训练集群的瓶颈往往是布线。(马斯克)

原因在于这套系统的工作方式。要让所有 GPU 协同训练,整个集群需要支持远程直接内存访问(RDMA,Remote Direct Memory Access),一台机器可以直接读写另一台机器的内存,绕过 CPU 和操作系统。

整个系统就像一个巨大的大脑。任意一张 GPU 都要能和十万张里的任意一张通信。这个布线布局疯狂到难以想象。(马斯克)

十万张卡,两两之间都得能高速对话。「连线」这个词已经覆盖不住这个工程量。

Lex 说自己走在那个机房里,感觉像在人类肉眼可见的尺度上看到了一个大脑,那地方整个就是个大脑。

马斯克接的话更有意思:

人脑其实也是这样,很大一部分脑组织就是线缆。灰质是计算,白质是线缆。你大脑里很大比例就是线缆。(马斯克)

这个类比站得住。神经科学里反复提到的数字是,人脑白质占比相当高,髓鞘包裹的轴突负责把信号从一个区域传到另一个区域,铺得满脑子都是。计算单元和连接线路的比例,在一台超级计算机和一颗人头里,受着相似的约束。

Lex 那句「我们走在了一个大脑里面」,其实比听上去更准确。让一个系统变强,一部分功夫花在增加算力上,更大一部分功夫花在让这些算力互相说话上。后者更难,因为它买不来,只能一根一根布出来,接错一根就要花几小时去诊断。

老板亲自去插光纤,为了让「这很蠢」能被说出口

播客里有段对话容易被忽略。Lex 提到,马斯克在孟菲斯做的事,是把每个人正在做的步骤都亲自走一遍。

马斯克解释了自己的动机:

一线的人在做什么,我至少会亲自做几次。接光纤、诊断一个坏掉的连接。(马斯克)

他做这些没有在体验生活。他给出的理由指向一个更根本的问题:如果没有人真正理解整条流程,就没有人有资格判断哪一步是蠢的。

五步算法的第一步要求质疑需求,第二步要求删除步骤,这两件事都建立在一个前提上:有人知道这个需求怎么来的,这个步骤到底在干什么。流程一旦被切成很多块,每块由不同的人负责,每个人只知道自己那一小块,整条链上就没人敢说「这一步可以删」。说了也没人信。

所以他到现场,把所有人正在干的活过一遍。信息在整条链上流通起来,每个环节的人都知道上下游在做什么,「这一步很蠢」这句话才有可能被说出口,也才可能被听见,然后才有可能真的删掉。

这跟他的管理风格是一体的。他反复讲过,最危险的状态是所有人都点头,连一个反对的声音都没有。一座几周内建起来的集群,如果没人敢在老板面前说「这个环节是多余的」,那五步算法就只是播客里的一段漂亮话,落不到任何一根电缆上。

Lex 在那座机房里走了一圈,感受到的「所有人都在简化、都在迭代」,大概就来自这里。流程被摊在所有人面前,谁都能看见全貌,谁都算得出哪一步是白费的。

把五步算法翻译成自己的动作

整场对话里能拿走的东西,是那几条非常具体、可以立刻用来照镜子的规则。「第一性原理」这四个字本身帮不了什么忙,它太抽象,谁都能说。

马斯克自己承认,这五步他念了很多年,还是经常犯错。于是他把它做成咒语,做成清单,做成每次开会都要过一遍的动作。越是聪明的人,越容易缺一道防止自己在熟悉的路上滑过去的护栏。

要把这次对谈变成能执行的动作,大概是这样几条:

  • 拿到任何任务,别急着排期。把需求清单摊开,逐条问「这条是谁提的,他当时基于什么信息」,先砍掉三分之一再动手。砍不动的地方,往往就是没人敢质疑的地方。
  • 把「主动多删」写进流程,当成设计参数。设定一个明确的删除比例,比如这一轮砍掉 20%,并且提前告诉大家其中会有十分之一加回来。加回来记在预算里,别算成失误。
  • 检查手上正在做的优化项目,先问一句「这个东西能不能整个删掉」。把一个不该存在的东西做得更精致,浪费比做错更大,因为它以后更难被砍。
  • 上线前必答两问:有没有在自动化一个还没被简化过的东西?有没有在加速一个还没被质疑过的东西?两个都是肯定的,先停下来重排顺序。
  • 定期回头翻那些已经自动化、已经跑得很顺的系统。它们最容易变成没人敢动的黑箱,也最容易在某天被发现根本不需要。
  • 亲自干几次最底层的活。接一次线、跑一次数据、走一遍客服流程,一小时内你通常能找出三层以上没人敢提的冗余。
  • 算峰值,别只算平均值。孟菲斯的教训很直接,同步带来的瞬时抖动会击穿所有按平均值设计的配套系统。团队、预算、服务器、供应链,都有这个毛病。

最后一条是整场对话里最不显眼、也最容易被跳过的地方。马斯克讲了一整套关于删除和简化的哲学,最后卡住他的,是一秒钟里几次十几兆瓦的功率跳变。方法论能帮你砍掉多余的步骤,砍不掉物理规律。

分清哪些东西可以删、哪些东西必须扛,这大概是这套算法真正的终点。

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)。

0:02It's very difficult. - Can you just speak to what it takes for a great engineering team for you? What I saw in Memphis, the supercomputer cluster is just this intense drive towards simplifying the process, understanding the process, constantly improving it, constantly iterating it.

0:23- Well, (laughs) it's easy to say simplify, and it's very difficult to do it. I have this very basic first principles algorithm that I run kind of as like a mantra, which is to first question the requirements, make the requirements less dumb. The requirement is always dumb to some degree.

0:46So if you wanna start off by reducing the number of requirements, and no matter how smart the person is who gave you those requirements, they're still dumb to some degree. You have to start there because otherwise, you could get the perfect answer to the wrong question. So try to make the question the least wrong possible.

1:04That's what question the requirements means. And then the second thing is try to delete whatever the step is. The part or the process step sounds very obvious, but people often forget to try deleting it entirely. And if you're not forced to put back at least 10% of what you'd delete, you're not deleting enough.

1:31And somewhat illogically, people often, most of the time, feel as though they've succeeded if they've not been forced to put things back in. But actually, they haven't because they've been overly conservative and have left things in there that shouldn't be. And only the third thing is try to optimize it or simplify it.

1:57Again, these all sound I think very obvious when I say them, but the number of times I've made these mistakes is more than I care to remember. That's why I have this mantra. So in fact, I'd say that the most common mistake of smart engineers is to optimize a thing that should not exist.

2:17- Right. So like you say, you run through the algorithm and basically show up to a problem, show up to the supercomputer cluster and see the process and ask, "Can this be deleted?" - Yeah, first try to delete it. Yeah. - Yeah, that's not easy to do. - No, and actually, what generally makes people uneasy

2:41is that you've gotta delete at least some of the things that you'd delete, you will put back in. But going back to sort of where our limbic system can steer us wrong is that we tend to remember, with sometimes a jarring level of pain, where we deleted something that we subsequently needed.

3:02And so people will remember that one time, they forgot to put in this thing three years ago and that caused them trouble. And so they overcorrect, and then they put too much stuff in there and over complicate things. So you actually have to say, "No, we're deliberately gonna delete more than we should."

3:20So we're putting at least 1 in 10 things, we're gonna add back in. - And I've seen you suggest just that, that something should be deleted and you can kind of see the pain. - Oh yeah, absolutely. - Everybody feels a little bit of the pain. - Absolutely, and I tell 'em in advance, like, yeah, some of the things that we delete, we're gonna put back in.

3:42And that people get a little shook by that. But it makes sense because if you're so conservative as to never have to put anything back in, you obviously have a lot of stuff that isn't needed. So you gotta overcorrect. This is, I would say, like a cortical override to Olympic instinct.

4:02- One of many that probably leaves us astray. - Yeah. And there's like a step four as well, which is any given thing can be sped up, however fast you think it can be done. Like whatever the speed is being done, it can be done faster. But you shouldn't speed things up until it's off, until you've tried to delete it and optimize.

4:21Otherwise, you're speeding up something that shouldn't exist is absurd. And then the fifth thing is to automate it. - Damn. - And I've gone backwards so many times where I've automated something, sped it up, simplified it, and then deleted it. And I got tired of doing that. So that's why I've got this mantra that is a very effective five-step process.

4:45It works great. - Well, when you've already automated, deleting must be real painful. - Yeah, that's great. It's like, wow, I really wasted a lot of effort there. - Yeah. - I mean, what you've done with the cluster in Memphis is incredible, just in a handful of weeks. - Yeah, it's not working yet.

5:04So I don't wanna pop the champagne corks. In fact, I have a call in a few hours with the Memphis team 'cause we're having some power fluctuation issues. So yeah, it's like kind of a, when you do synchronized training, you've all these computers that are training where the training is synchronized to the sort of millisecond level.

5:38It's like having an orchestra, and the orchestra can go loud to silent very quickly at subsecond level. And then the electrical system kind of freaks out about that. Like if you suddenly see giant shifts, 10, 20 megawatts several times a second, this is not what electrical systems are expecting to see.

6:01- So that's one of the main things you have to figure out the cooling, the power, and then on the software as you go up the stack on how to do the distributed compute, all of that, all of that. - Today's problem is dealing with extreme power jitter. - Jitter, power jitter. - Yeah.

6:19- That's a nice ring to that. So that's, okay. And you stayed up late into the night as you often do there. - Last week, yeah. - Last week? - Yeah. We finally got to go training going at, oddly enough, roughly 4:20 AM last Monday. - Total coincidence. - Yeah, I mean, maybe it was 422 or something.

6:42- Yeah, yeah, yeah. It's that universe again with the jokes. - Yeah, exactly, just love it. - I mean, I wonder if you could speak to the fact that one of the things that you did when I was there is you went through all the steps of what everybody's doing, Just to get a sense that you yourself understand it and everybody understands it so they can understand when something is dumb or some something is inefficient or that kinda stuff.

7:05- Yeah. - Can you speak to that? - Yeah, so look, I try to do, whatever the people at the front lines are doing, I try to do it at least a few times myself. So connecting fiber optic cables, diagnosing a faulty connection, that tends to be the limiting factor for large training clusters is the cabling.

7:24So many cables, because for a coherent training system where you've got RDMA remote, direct memory access, the whole thing is like one giant brain. So you've got to any connection. So any GPU can talk to any GPU out of 100,000. That is a crazy cable layout. - It looks pretty cool.

7:54- Yeah. - It's like the human brain, but like at a scale that humans can visibly see. It is brain. - Yeah. I mean, the human brain also has, a massive amount of the brain tissue is the cables. - [Lex] Yeah. - So like the gray matter which is the compute, and then the white matter which is cables.

8:15The big percentage of your brain is just cables. - That's what it felt like walking around in the supercomputer center is like, we're walking around inside the brain. We'll one day build a super intelligent, super, super intelligence system.

资源与出处

相关视频

同频道、同主题或相邻年份