Naval : 复杂系统源于简单设计的迭代 || Complex Systems Emerge From Iterations On Simple Designs

本文转自: https://nav.al/iterate
仅做个人收藏,版权归原作者所有

尼维:我们都见过SpaceX火箭的Raptor发动机图片,如果你仔细观察各种版本,它们从易于调整变得难以调整。因为最新版本的发动机部件数量少,你几乎无法对其进行任何改动。

早期版本的发动机有成千上万的部件,你可以随意更改其厚度、宽度、材料等。而当前版本几乎没有多余的部件可供调整。

纳尔:复杂性理论中有一个理论,每当我们在自然界中发现一个复杂系统在运作时,它通常是由一个非常简单的系统经过反复迭代而产生的。

我们最近在人工智能研究中就能看到这一点——人们只是将非常简单的算法与越来越多的数据结合。它们不断变得更聪明。

但反向操作则效果不佳。当你设计一个非常复杂的系统,然后试图用它构建一个功能完整的大型系统时,它往往会崩溃。因为其中的复杂性太多。因此,很多产品设计都是通过不断迭代自己的设计,直到找到一个简单有效的方案。而往往我们会在其中添加一些不必要的部分,之后又必须回过头来从这些杂乱中提取出真正的简洁。

这一点在个人计算领域也能看到,macOS仍然比iOS更难使用。iOS更接近操作系统这一概念的柏拉图式理想。不过,基于大型语言模型(LLM)的操作系统可能更接近——因为它能用自然语言进行交互。

最终,你必须去除一些东西才能实现扩展。Raptor发动机就是这样一个例子。当你弄清楚哪些部分有效后,就会意识到哪些是不必要的,然后去除它们。

这也是马斯克的一个重要原则。他基本上认为:在你优化一个系统之前,这是最后才做的事情。在你开始尝试让某物更高效之前,首先要质疑需求。

你会问:“为什么这个需求存在?”

乔根森新书里提到的埃隆方法之一是,你首先要找到这个需求的来源。不是哪个部门提出的,而是某个具体的人提出的。

是谁说“这就是我想要的”?

你要回去问:“你真的需要这个吗?”

然后你消除这个需求。一旦去除了不必要的需求,你就会拥有更少的需求。现在你有了部件,然后你尝试尽可能多地去除这些部件,以满足绝对必要的需求。

之后,你才开始考虑优化,思考如何最高效地制造这个部件并将其放置在合适的位置。最后,你可能会进入成本效率和规模经济等层面。

将一个优秀产品从零带到一的关键人物,通常是那个能将整个问题装在自己脑海中的单个人——通常是创始人。他们需要能够理解:为什么这个部件在这里?如果部件A被去除了,那么部件B、C、D、E及其需求和考虑又会如何变化?

这就是对整个产品拥有整体视角的体现。

你可以在Raptor发动机的设计中看到这一点。埃隆给出的例子让我印象深刻——他试图让这些玻璃纤维垫片在特斯拉电池上更高效地生产。于是他去了生产线,发现这个过程太慢,就干脆把睡袋铺在地上,留在那里。他们尝试优化用于将玻璃纤维垫片粘贴到电池上的机器人,试图更高效地完成这个任务或加快这条生产线。他们确实做到了一些改进,但速度仍然令人沮丧。

最后,埃隆说:“为什么会有这个需求?为什么我们要在电池上放玻璃纤维垫片?”

电池工程师回答:“实际上是因为降噪,所以你得去和噪音与振动团队谈谈。”

于是埃隆去找了噪音与振动团队。

他问:“为什么我们要放这些垫片?电池的噪音和振动问题是什么?”

他们却说:“不,不——其实没有噪音和振动问题。它们放在这里是因为电池起火时的热量。”

然后他回到电池团队,问:“我们真的需要这个吗?”

他们回答:“不需要。这里没有火灾问题,也没有热保护问题。这是过时的做法。其实是噪音和振动问题。”

他们各自都按照自己被训练的方式做事——按照过去的方法。他们通过测试安全性来验证,用麦克风测试噪音,然后决定不需要这些垫片,于是去除了这个部件。

这在非常复杂的系统和设计中经常发生。

有趣的是,每个人都说自己是“通才”,这其实是他们逃避成为“专家”的一种说法。但真正需要的是“通才”——一种能够掌握各种专业领域、至少达到80%熟练程度的通才,从而做出明智权衡的人。

尼维:我认为人们要获得这种通才能力——能够掌握任何专业领域——的方法是,如果你要学习某样东西,如果你要上学,就去学习那些具有广泛影响的理论。

纳尔:我会进一步简化,只说学习物理学。

一旦你学习了物理学,你就是在学习现实如何运作。如果你有坚实的物理学背景,你就能掌握电气工程、计算机科学、材料科学、统计学和概率学,甚至数学,因为它们都是应用性的。

我几乎在任何领域遇到的最优秀的人,都有物理学背景。如果你没有物理学背景,也别着急。我也有失败的物理学背景。你仍然可以通过其他方式达到,但物理学能训练你与现实互动,而它又是如此严苛,以至于能把你所有不切实际的假想都击碎。

相比之下,如果你在社会科学领域,你可能会有各种荒诞的信念。即使你掌握了一些社会科学中使用的抽象数学,你可能只有10%的真实知识,而90%都是错误的。

物理学的好处在于,你可以学习相当基础的物理学知识。你不需要深入到夸克和量子物理等层面。你只需学习一些基本的物理概念,比如球体沿斜面滚动,这其实是一个很好的入门。

不过我认为任何STEM学科都值得学习。如果你无法选择学习什么,或者已经错过了学习阶段,那就与他人合作。实际上,最好的人并不只是学习物理学。他们是动手实践的人,是建造者,是不断创造东西的人。动手实践的人总是处于知识的前沿,因为他们总是使用最新的工具和部件来打造酷炫的东西。

因此,是那些在无人机成为军事工具之前就制造竞速无人机的人,是那些在机器人成为军事工具之前就制造战斗机器人的人,是那些想要在家中拥有电脑,而不满足于去学校使用电脑的人。这些人真正理解事物,也最快地推动知识进步。
—————

Nivi: We’ve all seen the pictures of the Raptor engine for the SpaceX rockets, and if you look at the various iterations, they go from easy-to-vary to hard-to-vary. Because the most recent version just doesn’t have that many parts that you can fool around with.

The earlier versions have a million different parts where you could change the thickness of it, the width of it, the material, and so on. The current version barely has any parts left for you to do anything with.

Naval: There’s a theory in complexity theory that whenever you find a complex system working in nature, it’s usually the output of a very simple system or thing that was iterated over and over.

We’re seeing this lately in AI research—you’re just taking very simple algorithms and dumping more and more data into them. They keep getting smarter.

What doesn’t work as well is the reverse. When you design a very complex system and then you try to make a functioning large system out of that, it just falls apart. There’s too much complexity in it. So a lot of product design is iterating on your own designs until you find the simple thing that works. And often you’ve added stuff around it that you don’t need, and then you have to go back and extract the simplicity back out of the noise.

You can see this in personal computing where macOS is still quite a bit harder to use than iOS. iOS is closer to the Platonic ideal of an operating system. Although an LLM-based operating system might be even closer—speaking in natural language.

Eventually, you have to remove things to get them to scale, and the Raptor engine is an example of that. As you figure out what works, then you realize what’s unnecessary and you can remove parts.

And this is one of Musk’s great driving principles where he basically says: Before you optimize a system, that’s among the last things that you do. Before you start trying to figure out how to make something more efficient, the first thing you do is you question the requirements.

You’re like, “Why does the requirement even exist?”

One of the Elon methods in Jorgenson’s new book is you first go and you track down the requirement. And not which department came up with the requirement; the requirement has to come from an individual.

Who’s the individual who said, “This is what I want.”

You go back and say, “Do you really need this?”

You eliminate the requirement. And then once you’ve eliminated the requirements that are unnecessary, then you have a smaller number of requirements. Now you have parts, and you try to get rid of as many parts as you can to fulfill the requirements that are absolutely necessary.

And then after that, maybe then you start thinking about optimization, and now you’re trying to figure out how can I manufacture this part and fit it into the right place most efficiently. And then finally, you might get into cost efficiencies and economies of scale and those sorts of things.

The most critical person to take a great product from zero to one is the single person—usually the founder—who can hold the entire problem in their head and make the trade-offs, and understand why each component is where it is.

And they don’t necessarily need to be the person designing each component, or manufacturing or knowing all the ins and outs, but they do need to be able to understand: Why is this piece here? And if Part A gets removed, then what happens to Parts B, C, D, E and their requirements and considerations?

It’s that holistic view of the whole product.

You’ll see this in the Raptor engine design. The example that Elon gives that I thought was a good one—he was trying to get these fiberglass mats on top of the Tesla batteries produced more efficiently.

So he went to the line where it was taking too long, put his sleeping bag down, and just stayed at the line. And they tried to optimize the robot that was gluing the fiberglass mats to the batteries. They were trying to attach them more efficiently or speed up that line. And they did—they managed to improve it a bit, but it was still frustratingly slow.

And finally he said, “Why is this requirement here? Why are we putting fiberglass mats on top of the batteries?”

The battery guy said, “It’s actually because of noise reduction, so you’ve got to go talk to the noise and vibration team.”

So he goes to the noise and vibration team.

He’s like, “Why do we have these mats here? What is the noise and vibration issue?”

And they’re like, “No, no—there’s no noise and vibration issue. They’re there because of heat, if the battery catches fire.”

And then he goes back to the battery team like, “Do we need this?”

And they’re like, “No. There’s not a fire issue here. It’s not a heat protection issue. That’s obsolete. It’s a noise and vibration issue.”

They had each been doing things the way they were trained to do—in the way things had been done. They tested it for safety, and they tested it by putting microphones on there and tracking the noise, and they decided they didn’t need it, and so they eliminated the part.

This happens a lot with very complex systems and complex designs.

It’s funny—everybody says “I’m a generalist,” which is their way of copping out on being a specialist. But really what you want to be is a polymath, which is a generalist who can pick up every specialty, at least to the 80/20 level, so they can make smart trade-offs.

Nivi: The way that I suggest people gain that polymath capability—being a generalist that can pick up any specialty—is if you are going to study something, if you are going to go to school, study the theories that have the most reach.

Naval: I would summarize that further and just say study physics.

Once you study physics, you’re studying how reality works. And if you have a great background in physics, you can pick up electrical engineering. You can pick up computer science. You can pick up material science. You can pick up statistics and probability. You can pick up mathematics because it’s part of it—it’s applied.

The best people that I’ve met in almost any field have a physics background. If you don’t have a physics background, don’t cry. I have a failed physics background. You can still get there the other ways, but physics trains you to interact with reality, and it is so unforgiving that it beats all the nice falsities out of you.

Whereas if you’re somewhere in social science, you can have all kinds of cuckoo beliefs. Even if you pick up some of the abstruse mathematics they use in social sciences, you may have 10% real knowledge, but 90% false knowledge.

The good news about physics is you can learn pretty basic physics. You don’t have to go all the way deep into quarks and quantum physics and so on. You can just go with basic balls rolling down a slope, and it’s actually a good backgrounder.

But I think any of the STEM disciplines are worth studying. Now if you don’t have the choice of what to study and you’re already past that, just team up with people. Actually, the best people don’t necessarily even just study physics. They’re tinkerers, they’re builders, they’re building things. The tinkerers are always at the edge of knowledge because they’re always using the latest tools and the latest parts to build the cool things.

So it’s the guy building the racing drone before drones are a military thing, or the guy building the fighting robots before robots are a military thing, or the person putting together the personal computer because they want the computer in their home and they’re not satisfied going to school and using the computer there. These are the people who understand things the best, and they’re advancing knowledge the fastest.

发表回复