OpenAI研究员Gabriel Petersson：高中辍学自学数学与机器学习，如何自学与求职

比如，想象你在一家书店里，可以把嵌入向量想象成店里的不同书籍。

Like, imagine you're in a bookstore and you can imagine the embeddings being the different books in the store.

然后你可以想象，你知道的，所有这些。

And then you can imagine, you know, all this.

它会将所有与AI相关的内容与现实世界概念联系起来，这让像我这样的人很容易理解。

And it will connect everything that has to do with AI to like real world concepts, which makes it really easy to to reason about for for for someone like like me.

听起来你现在可以学习任何主题，只需要ChatGPT就够了。

So it sounds like any sort of topic you can learn now and all you need is ChatGPT.

你一开始只需要问：'嘿，我需要先了解哪些基础知识？'

And you start with just asking like, hey, what are the preliminary things I need to understand about this?

然后你可能会顺着其中一条线索深入，对吧？

And then you might pull on one of those threads, right?

最初当你研究视频模型时，你会想：'好吧，图像生成模型就像扩散模型，比如稳定扩散。'

First, when you were investigating video models, you're like, Okay, image generation models are like diffusion models, like stable diffusion.

然后你会想：'好吧，扩散模型到底是怎么工作的？'

And you're like, Okay, how the frick does a diffusion model work?

然后你可以让它给你解释，也许还能生成代码示例。

And then you would have it explain it to you, maybe generate code samples.

但每个环节你都会进一步追问。

But every aspect of that, you would then inquire further.

比如，我不理解这部分。

Like, I don't understand this part.

这是什么？

What what is this?

为什么这个被采用到这种模型架构中？

Why is this adopted to this model architecture?

哦，为什么要这样做？

Oh, why is this done this way?

好的。

Okay.

这个数学原理是怎么运作的？

How does that math work?

我是说，我读过你在x.com上非常受欢迎的帖子。

And I I mean, I read like your posts on on on x.com, which are very popular.

看起来你能够以一种持续提问直到完全理解的方式使用AI。

And it it looks like you're just able to use AI in a way where you continuously query until you have full understanding.

当你真正理解后，几乎可以像费曼学习法那样重新解释——最好的学习方式就是讲解事物，但现在你可以用AI来实现这一点。

And then when you do have full understanding, you almost like re explain similar to how Feynman the best way to learn is to to explain things, but now you can do it with AI.

对吧？

Right?

所以当你学习扩散模型时，在深入探讨某个非常技术性的话题后——你可能连梯度是什么都不知道，然后它会向你解释微积分或线性代数，你就能掌握这些知识。

So when you're learning about diffusion models, after going through a deep dive on some very technical topic where you might not even know what gradients are, and then it'll explain you calculus or some linear algebra and you pick that up.

但之后你需要向模型解释这些内容，然后它会澄清或指出你不理解的不同方面，你不断重复这个过程直到完全掌握。

But then you would explain it back to the model and then it will then clarify or like see different aspects that you don't understand and you keep repeating that until you have a very strong grasp.

在我看来这有点像递归填补空白，如果要用一个词来总结的话。

I see it bit as like recursive gap filling, if I would like summarize it in one word.

这里需要的技能是清楚知道自己知识体系中的空白在哪里。

It's like you need like the skill that's required here is knowing what gaps you have in your knowledge.

比如说你有一个AI模型或任何你想学习的东西，但要意识到自己哪部分没真正理解。

Like say you have an AI model or like whatever else you want to learn and understanding when you don't really understand a part.

这实际上相当困难。

It's actually pretty hard to do.

这需要你自己去训练和练习。

Like, it's something you need to train up and practice on yourself.

比如，等一下。

Like, wait a second.

我真的理解这部分了吗？

Do I really understand this part?

所以这是你需要具备的一个能力。

And then so that's one thing that you need.

第二点至少在你开始提问时，你需要对'顿悟时刻'有强烈的感知信号。

The second thing at least when you start asking questions, you need to have a real strong signal for when it clicks.

当你感觉'啊，这下我懂了'的时候。

When you're like, ah, there it clicked.

是的。

Yeah.

我能够从根本上理解这件事为何如此。

I can just understand like fundamentally why this thing is as it is.

其他人该如何学习

How would someone else learn how to

如何借助AI学习？

learn with AI?

这是个非常好的问题。

This is a very good question.

首先，要改变那种认为AI是替你完成工作的误解，转而明白应该利用AI来明确帮助你学习。

I mean, first of all, just change like the misconception of AI being used to do the work for you to instead you know, use the AI to explicitly help you learn.

不要仅仅用它来完成工作，而是要真正从中学习。

Like, you don't you don't just use it to get work done, actually learn from it.

当你转变这种思维方式时——虽然目前看来还不太普遍，但正变得越来越常见——你就已经具备了达成目标的大部分条件，对吧？

I mean, the moment you just switch that mindset, seems still fairly uncommon but it's becoming more and more common all the time, you have most of the things to get there, right?

要想真正精通，首先如我所说，要清楚自己知识上的空白点，理解彻底掌握某个概念时的感受，并且不断想出各种技巧——比如你会注意到JGPT总是以相当标准的方式回应，而你的学习方式可能并不完全符合它的回应模式，因为它要确保每个人都能有良好的体验。

And then to become really good, first of all, like I said, know when you have gaps in your knowledge, understand what it feels like when you fundamentally grasp something, and you know, constantly come up with all these hacks like you you notice the JGPT will respond in a fairly standard way and your way of learning is probably not exactly what it responds like because it wants to, you know, make sure everyone has a good experience.

没错。

Yep.

但你可能希望它以另一种方式回应。

But you probably want it to respond in another way.

我经常告诉你，比如要极其直接具体。

I very often tell you, for example, be extremely direct and concrete.

始终向我展示你生成代码的所有中间状态和形态。

Always show me all the intermediate states and the shapes of the code you produce.

确保我能直观地理解为什么会这样发生。

Make sure to to make sure I have, like, a really intuitive understanding of why it happens.

如果你不确定，一定要向我展示各种选项、别人尝试过的方法、为什么这个有效而另一个不行。

And if you're unsure, make sure you show me options and what others have tried and why this works and why something else didn't work.

这样你就能逐渐擅长提出那些能让你顿悟的问题。

And you start becoming good at asking these questions that give you the moment.

尽可能快地达到那个顿悟时刻。

As fast as possible, you wanna get to the moment.

对。

Yeah.

就像你第一次理解线性代数，或者第一次明白反向传播原理时，那种清晰的'啊哈'瞬间。

Like the first time you understood linear algebra or the first time you understood what back propagation works, you probably had a very clear like, oh, wow.

终于开窍了。

It finally clicked.

而追求这些顿悟时刻，并让它们尽可能频繁地出现，对吧？

And to chase these clicks and to make them appear like as frequent as possible, Right?

这就像是你的效用函数。

That's like kind of your utility function.

太疯狂了。

That's crazy.

在当今时代，要想保持竞争力并成为顶尖表现者。

It's like in modern day, in order to stay competitive and to be like top performing.

当你真诚地观察一个人时，他们能像你一样凭借快速查询信息的能力迅速跻身行业顶尖。

And when you look at someone, like honestly, they can they'll be at the top of the field pretty quickly like how you've done it just by the rate of being able to query for information.

没错。

Yep.

这大概就是当下最重要的技能了，对吧？

And that's probably like the most important skill now, would you Yeah.

而且，你知道，培养这种能力是另一个非常重要的积累过程。

And and and, you know, building up this like this is another very important, like build up.

当你脑海中产生疑问时，务必立即将其输入Chateappiti。

The moment you have a question in your head, make sure to get it into Chateappiti.

这一点非常难做到。

This one is very hard.

我记得这花了我...我那位一起创业的表哥当时就说：'老兄，Chateappiti上线了'。

This took me I remember my my cousin, the same cousin I started a company with, he was like, dude, Chateappiti is out.

那时候还是在...该怎么说来着？前前期阶段？

This is like pre pre like the what it's called?

就像，你回想这本书刚问世时，它就像个游乐场一样。

Like, you think of this book back in the day when it was just like a playground.

那是在GPT-3的超级早期阶段，远在ChatGPT出现之前。

It was like super early in GPT three, like before ChatGPT days.

是啊。

Yeah.

他还说，你为什么现在还没用上这个？

And he was like, why are you not using this yet?

你不是整天都在写代码吗？

Like, you're writing code all the time?

我就说，好吧，我试试看。

Like, yeah, I'm gonna try it out.

而且你知道，他每个月都在催我。

And, you know, he kept on pushing me every month.

我花了差不多一年时间才真正开始建立联系，意识到——哦，我遇到这个问题时可以用它

And it took me like a year until I really started connecting, like, oh, I have this problem.

我得问问Shashi Biti。

I need to ask Shashi Biti.

这太常见了，就像你经常遇到的人那样。

And it's so common, like, you you meet people all the time.

你在讨论中，人们总有各种问题，或者你和别人一起工作，他们也有各种疑问，你懂的。

You're in a discussion, and people have all these questions, or you sit co working with someone, they have all these questions, you know.

然后你就会觉得，每次有任何问题，任何时候需要猜测什么，就直接问Shachibati。

And you're like, you just ask It's like every time you have any kind of question, any time you need to guess about something, just constantly ask Shachibati.

就像，它总是在那里。

Like, it's it's always there.

是啊。

Yeah.

这几乎不费什么力气。

It's very low effort.

当然，你有一个非常简单的方式，可以随时询问Shachibati你好奇的任何事情，然后就能，你知道的，掌握世间所有知识。

Like, sure you have a very simple way to just ask Shachibati about anything you ever wonder, and you'll just, you know, have all knowledge in the world.

是的。

Yeah.

但关键之处在于几乎沉迷于那种能够快速达到领悟或内化某个概念的瞬间。

But the important part is like almost getting hooked on like how fast you can get to that moment of realizing or internalizing something.

以及能够巧妙地引导ChatGPT，不是泛泛而谈，而是让它提供非常具体或不同的类比，或者任何最适合你学习方式的表达形式

And the skill of being able to prop chat g b d, not in a generic way, but in a way where it'll give you very concrete or different analogies or whatever form factor that works best with your learning style

对。

Yeah.

从而让你理解并内化它。

For you to then understand and internalize that.

这确实很难。

Which is really hard.

或者说，我觉得自己挺笨的。

Or I mean, I think I'm pretty dumb.

所以有时候当我向GPT提问，它解释后，我就感觉

So it's like it's like sometimes when I ask the GPT stuff and it explains it and I'm like

我我不明白。

I I don't understand.

我不明白。

I don't understand.

这太难了。

It is just too hard.

然后你试了一次又一次，最后只能说，好吧。

And and you try again and you try again and you're like, okay.

当你真正抓住要点时，你会说，哦，好吧。

And then you really grab you're like, oh, you're like, okay.

是时候提升提问技巧了，对吧？

Prompting skills time, right?

然后你会说，好吧。

And and you're like, okay.

如果世界上这些功能不存在，如果那些从未存在过会怎样？

What if these these features in the world didn't exist and what if that never existed?

他们是否仍会发明这个东西，并用12岁小孩能懂的方式解释给我听？

Would they still have invented this thing and explain it to me like I'm 12?

要知道，生成展示分布情况的图表，这些正是我需要理解的关键信息。

Know, generate graphs showing the distributions that I need to know to really understand this.

要知道，有无数创意方法可以从Chateappeti中提取你所需的信息。

You know, there's so many creative ways you can use to to really extract the information you need from Chateappeti.

我认为我学到的很多东西，特别是通过早期模型如Chateappiti的经验——它一直在不断进步。

And I think a lot of the things I've learned, especially like with previous models, like Chateappiti is becoming so much better all the time.

但大约一年前模型还没这么强大时，有些知识如果我不真正掌握信息提取技巧，可能根本学不会。

But like like a year ago when the models weren't as strong, some of the things I've learned is like, I probably couldn't learn them if I didn't like really know how to extract the information.

比如我可以把问题问上一千遍，让它改写一千遍。

Like I I could ask the question a thousand times and make it rephrase it a thousand times.

我就是无法理解。

I just didn't wouldn't understand.

这就是为什么这个要从小学就开始教Shashi Bati。

This this is why this is teach Shashi Bati in in, like, from elementary school.

这就像，你知道的，一种全新的语言。

This is like, you know, a a new language.

这就像，你知道的，你仍然需要生活中的其他一切，比如创造力、能动性等等，但知识已经进入一个全新的时代。

This is like, you know, you still need all the other things in life, like creativity and agency and like all these other things, but knowledge is a completely new era.

你无法将这与任何其他事物相提并论。

Like you can't compare this with anything else.

一个非常具体的例子，因为人们似乎没有意识到AI将如何急剧地改变世界。

Very concrete example of this, because people don't seem to realize how AI, how abruptly this will change the world.

比如现在，我正在做一份传统上大家都认为需要博士学位的工作。

Like currently, I'm doing a job which traditionally everyone would agree you need like a PhD for.

对吧？

Right?

有一群人没有博士学位也做到了，但如果你五年前告诉别人，说在顶级AI实验室里会雇佣一个没有...你知道的...暂时没有相关资质的人。

There's a bunch of people who have done it without a PhD, but like if you told someone like five years ago like, oh yeah, at one of the top AI labs, someone will be hired who hasn't, you know the Got a thing for a while.

而他唯一拥有的就是...他在其他领域做过很多很酷的事情，但对这个领域一无所知。

And the only thing he had was that he had was like they done like all these very cool things on other areas, but they didn't know anything about this thing.

人们会说，不可能。

People are like, no.

那那那是不可能的。

That that that's not possible.

对吧？

Right?

但现在的情况是，我只需使用Chateapiti就能完成传统上需要多年经验的人才能做的工作。

But we are now in a scenario where I can do the job, traditionally only, you know, downward people have done it for like multiple years, just by using Chateapiti.

这太疯狂了。

That's insane.

世界在Chateapiti的推动下会发展得多快啊。

Like, the amount how fast the world will develop with Chateapiti.

你可以随心所欲地进行任何领域的研究。

Like, you can just do research in anything you want.

如果你想开始做生物研究，或者硬件研究，你完全可以立即着手去做。

If you wanna start doing bio research, you wanna start doing like hardware, you can just go and do things.

这简直不可思议。

It's just incredible.

是啊。

Yeah.

这真的像是世界GDP的两位数增长，仅仅来自大型语言模型。

This is this is really like a double digit increase in world GDP, like just coming from large language models.

而且任何人都能做到。

And anyone can do it.

他们

They

知道如何使用ChatGPT。

know how to use ChatGPT.

没错。

Yeah.

每月只需20美元。

It's $20 per month.

要知道，这些思维模型在编程和理解事物方面表现得非常非常出色。

And you know, the thinking models are like really, really good when it comes to like coding and like understanding things.

你如何利用ChatGPT来学习，同时构建出世界顶级的视频模型之一？

How do you use ChatGPT to learn as you build out like one of the world's best video models?

这非常简单。

It's very simple.

很多人问我这个问题，他们总是很困惑，比如，你到底是怎么做的？

A lot of people ask me this and they're always confused like, okay, what do you actually do?

比如，你会怎么做？

Like, what do you do?

而且，你知道，他们想象——我不知道他们想象什么，但肯定是些非常非常特别的东西。

And, you know, they imagine I don't know what they imagine, but, know, something very, very special.

对吧？

Right?

而且这其实相当简单。

And it's it's fairly simple.

你看视频时会觉得，哦，这部分看起来不太理想。

You know, you look at the video and you're like, oh, this part of the video doesn't look very good.

于是你去稍微调整模型架构，或者修改数据之类的，然后训练模型，查看结果，盯着视频看一会儿，发现这些视频效果更好了，那很棒。

So you go and you change the architecture in the model a bit or you change the data or something and you know, you you train the model, you look at the results, you stare at videos for a while and you're like, oh, these videos were better, that's great.

这部分会合并到主分支。

This this goes into to master.

然后你就这样循环操作。

And then you just do that on a loop.

对吧？

Right?

接着你会想，好，下一个要修复或尝试的是什么？

And you're like, okay, what's the next thing that I wanna fix or the next thing I wanna try?

这正是AI真正擅长的领域。

And that's where like AI is really good.

对吧？

Right?

因为这就像，哦，我遇到了这个具体问题。

Because it's like, oh, I have this specific problem.

你好，AI。

Hello AI.

你看，这是我的全部代码库。

You know, here's my entire code base.

告诉我10个可以改进它的点子。

Tell me 10 ideas of what I can do to improve this.

对吧？

Right?

它会给你一堆建议。

It'll tell you a bunch of ideas.

它会推荐你可以阅读的论文。

It'll refer to papers you can read.

它能完成所有这些非常棒的事情。

It'll do all these, like, really great things.

它会给你一大堆非常适合头脑风暴的点子。

And it will give you, like, a bunch of ideas that it's really good to brainstorm with.

你知道，你可以把这些想法带给你的同事，和他们讨论，他们都是非常优秀的人。

You know, you can bring all these ideas to your colleagues and, talk to them who are just, extremely good.

而且，是的，这相当直接明了。

And, yeah, it's, like, fairly straightforward.

哇。

Wow.

它是如何找到其他研究论文来建议你探索的？

How does it find, like, other research papers to to suggest you to explore?

它只是

It just

知道这些。

knows about that.

它，就是知道。

It just, knows.

是的。

Yeah.

这就像是四点零版本，我觉得即便是更早的模型，你只需打印出链接并点击它们。

It's it's like four point o just pretty I think even earlier models, you just print out the links and press the link.

找一些讨论这个主题的论文。

Find some papers talking about this.

然后，你知道的，显然我不会逐字逐句地读这些论文。

Then, you know, obviously I don't read the papers word for word.

你刚才说的是GBT吗？

You know You said GBT?

我会给出我的指令，比如：'给我列出这篇论文与众不同的地方'。

I have my my instructions like, want to, you know, give me a list of things this paper did differently.

因为很多时候，论文会采用一些我已经熟知的技术，并在此基础上引入新内容。

Because oftentimes, the paper, they take some technique that I already know about and they introduce some new things to it.

我就直接问它：'好的，与其他研究相比，给我一个具体清单，明确指出他们相比之前的研究做了哪些不同的事。'

And I just ask it, okay, compared to the other thing, tell me a list and be extremely concrete of exactly what they did that compares to the previous thing.

这是个非常棒的总结方法。

And that's a really good summarization.

很多时候你会觉得，这篇论文可能通不过。

And often at times, you know, you're like, ah, this paper probably wouldn't make it.

不值得尝试，直接看下一篇就好。

It's not worth trying out and you can just go to the next one.

或者你会觉得，这篇论文真的很棒。

Or you're like, ah, this paper is really good.

我只有在决定要实施时才会深入阅读论文。

Like, I only read the paper in-depth if I actually decide to implement it.

然后可能会在遇到bug时再仔细研读。

And then I probably will will read it when I'm if I have bugs.

我大概就是把所有代码一扔，说'把这个实现到我代码里'。

Like, I'm probably just like throwing all my code and be like, hey, implement this into my code.

对吧？

Right?

我就直接复制粘贴进去。

I just copy paste it in.

哦，哇。

Oh, wow.

你知道，我肯定会仔细阅读代码的。

And you know, I I obviously make sure to like really read through the code.

是的。

Yeah.

我认为绝不能随便扔代码进去，这极其重要。

I think it's extremely important that you can't just throw in code.

你不能凭感觉写代码吗？

You can't just vibe code?

不行。

No.

我我我不是那种凭感觉写代码的人。

I'm I'm I'm not the vibe coder.

我对代码有着非常明确的观点。

I'm I'm very opinionated when it comes to code.

不。

No.

听起来你采取了一种截然不同的方式，比如，如果你真的想构建一些非常具体的东西，就需要理解所有细节。

It sounds like you have, like, the a very different approach where, like, if you actually wanna build, like, you know, really concrete things, like you need to understand everything.

对吧？

Right?

是的。

Yeah.

尤其是当你处于某个领域的前沿时。

Because especially if you're pushing like a forefront of any field.

没错。

Yeah.

我的意思是，我想理解所有的基础知识。

I mean, I want to understand all the foundations.

我觉得人们的第一反应通常是，哦，你只是想走捷径。

I think the the first reaction people have is like, oh, you just wanna take shortcuts.

你并不是真的想理解事物。

You don't really wanna understand things.

你只是觉得自己能靠一堆AI技术蒙混过关。

You just think you can slip out a bunch of AI slope.

对吧？

Right?

而我认为这是

And I think this is

完全不对的。

not correct at all.

就像我

Like I

我想走捷径，那是

I wanna take shortcuts, that's

确实如此。

for sure.

但我想要通过捷径来理解所有基础知识。

But I wanna take shortcuts to understand all the foundations.

是啊。

Yeah.

这件事非常重要。

And that's very important this thing.

看起来人们要么站在一边说'让AI代劳所有工作'，

Like it seems like either you're on the camp like, okay, AI slope, do all the work for me.

我永远都不想工作。

I never wanna work.

要么就认为必须上大学，因为大学垄断了基础知识，必须由教授来传授。

Or you're in the camp, you need to go to college, they have a monopoly on all the foundational knowledge, you need to have this taught by a professor.

而教授其实并不处于中间立场，对吧？

And a professor is not really in between, right?

我是说，你需要所有这些，而AI很棒。

Mean, you need all these things and AI is great.

你应该在所有事情上都使用它。

You should use it for everything.

我们应该使用

We should use

它来理解一切，是的。

it to understand everything Yeah.

训练它就像是人类与AI的共生关系，旨在增强你的大脑和能力。

And train it's like human AI symbiosis in terms of like just enhancing your brain and enhancing your ability.

所以你现在在斯德哥尔摩。

So you're in Stockholm.

你离开了你的第一个创业公司。

You left your first startup.

你是怎么来到旧金山的？

And how did you find your way to San Francisco?

你做了什么？

What did you do?

是的。

Yeah.

我一直知道自己想继续在初创公司工作。

I always knew I wanted to continue to work in startups.

我一直把目光投向旧金山，因为我知道，我认识的所有最优秀的人都搬到了这里。

And I'll always had my sights on San Francisco because, know, all the best people I knew had moved here.

所有那些，你知道的，人们谈论的最好的公司都在这里。

All the, you know, all the best companies people were talking about were were here.

我意识到，或许我应该尽可能高效地快速学习。

And I noticed that like, probably I should just like super optimize for learning as fast as possible.

遗憾的是那时候还没有ChatGPT，不然想想我现在会是什么样子，如果当初学习时就有它的话。

This was sadly pre like just imagine where I would have been now if I had ChatGPT when starting to learn things.

亿万富翁了。

A billionaire.

是啊。

Yeah.

而且那时候，你能做的最好的事就是和最优秀的人共事。

And and like, back then the best thing you could do was to work with the very best people.

所以我就是这么尝试的。

So that's what I tried to do.

那么怎样才能和最优秀的人共事呢？

And so how do you work with the best people?

嗯，你要尽可能多地接触不同的公司。

Well you talk to as many companies as possible.

你要确保——明白吗——面试那个正在面试你的人，对吧？

You make sure you know, interview the person interviewing you, right?

比如：你做过些什么？

Like what have you done?

你们团队做代码审查吗？

Like do you guys do pull requests?

你们会认真审查我的PR吗？这样我才能真正知道自己犯了哪些错误？

Do you make sure to really review my PR so I actually know what mistakes I do?

后来我成功加入了几家拥有非常、非常优秀工程师的公司。

And I managed to join a couple companies with like really, really talented engineers.

我还特别注意保持自由职业者的身份工作。

I also made sure to to be like I generally worked as a contractor.

人们常犯的最大错误就是在职业生涯初期，在同一家公司待得太久。

Like the biggest mistakes people do is that they stay with the same company for way too long, early in their careers.

这绝对是我在人们职业生涯中看到的最大错误。

That's like by far the biggest mistake I see in people's careers.

听起来你高中辍学一年后，在第一家公司之后，就一直在寻找你认为最优秀的团队、工程师或合作伙伴。

So it sounds like a year after dropping out of high school, you know, after your first company, you just kept finding the best teams or the best engineers or the best people that you thought, you know, you could work with.

你与他们共事一段时间，学习经验后，又不断寻找新的机会或更好的团队，然后你...

And you work with them for a bit, you learn what it could, and then you kept finding new opportunities or like better teams and you

是的。

Yeah.

我过去只接受合同工角色，以确保在工作地点选择上能保持高度灵活性。

I used to only take contractor roles to make sure that I that I could be very mobile in the places I work with.

你努力寻找能与最优秀人才共事的最佳工作环境。

You try to find the best places to work with with the best people.

是的。

Yeah.

你尽可能与他们紧密合作。

You try to work as closely with them as possible.

确保对你所从事的工作有主见，这样就不会只做别人不愿做的任务，否则你就学不到东西。

Make sure you're opinionated about what you're working with so you don't only get to do like the tasks no one else wants to do because then you're not learning.

务必真诚感谢那些审阅你代码的人。

Make sure you really show appreciation for the people reviewing your code.

没错。

Yeah.

因为那是最好的资源来源

Because that's the best source of Or

获取反馈。

getting feedback.

总的来说就是获取反馈。

Getting feedback in general.

要知道，主动寻求反馈。

And you know, hunt feedback.

我是说，告诉人们，嘿，我真的很喜欢你的评审。

I mean, tell people, hey, I really like your review.

你能帮我逐条评审所有反馈吗？

Can you just review every single feedback on mine?

人们会感到震惊。

People will like shocked.

就像，哇哦。

Like, oh, wow.

我从没听说过有人喜欢反馈。

I'd never heard someone liking feedback before.

是的没错。

This is yeah.

当然，我

Of course, I

确实如此。

do feel Yeah.

这很罕见，因为人们通常都会回避，因为他们已经完成了学业，现在在工作了。

It's rare because people usually shy away because they already did their schooling and now they're working.

但是没错。

But Yeah.

作为一个年轻人，没有什么真正的荣誉，你就像个无名小卒，没有学位

As a young person with like no really accolades, you're like a middle you're you're nobody with no degree

对。

Yeah.

没错，对你来说学习的方式就是加入最好的团队，然后非常灵活，同时似乎还要不断寻求反馈。

Right, you the way for you to learn is like join the best teams and then be very nimble, but also relentlessly seek feedback it seems.

是的，你知道的，直接打电话给他们。

Yeah, and you know call them up.

打电话给他们，然后说，嘿，那个评审很棒，现在我们一起在电话里过一遍所有评论，对吧？

Call them up and be like, hey, that was a great review, now let's go through all the comments together on a call, right?

你能学到很多，就像追问后续问题一样，比如问题出在哪里，要成为一名真正优秀的工程师是极其困难的。

You learn so much and you just like ask follow-up questions like what's the issue, like there's like Becoming a really good engineer is extremely hard.

这个领域非常广泛。

It's like such a wide area.

有很多基础原理和直觉需要你去理解。

There are like so many like first principles things and intuitions you need to understand.

当你了解它们时它们相当简单，但学习起来可能非常困难。

And they're pretty easy when while you know them, but they can be very hard to learn learn.

如果有人直接告诉你这些，而你又擅长掌握它们，那真是...

And to have someone just straight up tell them to you and you being good at like picking them up is like such a such a In the

就像很久以前，你只能从现有的工程师或老师那里获取这些，但现在你还有AI可以做到这一点。

same way that whereas long before, you know, you only can do that from maybe like an existing engineer or maybe a teacher, but now you also have AI who can do that.

现在的AI

The AI Now

你可以在任何公司随时进行这个操作。

you can do this on demand at any company.

可以直接开始

Can just start in

你自己的公司。

your own company.

可能是凌晨4点，你还在熬夜编程或制作东西。

Could be like 4AM and you've been up coding or making something Yeah.

或者正在写论文或研究什么，你仍然可以向AI寻求反馈。

Or like writing a paper or researching something and you can still ask AI for feedback Yeah.

并解释为什么这个决定更好？

And explain you why was this why is this a better decision?

我经常这样做。

I do this all the time.

我觉得，当你发现生活中某件事特别有效时，就应该最大限度地利用它。

I I think like when you've found something in life that works really well, you should exploit it like to the maximum.

比如每天问ChatGPT一百个问题，对吧？

Like ask ChatGPT a 100 questions per day, right?

是啊。

Yeah.

我总是开着ChatGPT的标签页，写完代码就扔进去问：这样行吗？这样好吗？这样不错吧？

Like I always have tabs open with ChatGPT, I write code, throw it in there and I'm like, is this good, is this good, this is good.

有没有什么漏洞？我还能怎么改进？

Are there any bugs, what can I do better?

你懂的，何乐而不为呢？

You know, why not?

它可能会告诉你：哦，目前看起来没问题。

I mean, it'll probably tell you, oh, it looks fine right now.

但有时候会说：啊对，这里有个bug，或者你可以用这种方法实现，然后推销那个Simta。

But something's like, oh, yeah, there's a bug or oh, yeah, you can do it in this way and sell that Simta.

对吧？

Right?

你就是在持续学习。

You just constantly learn.

是啊。

Yeah.

如果你每天真问上一百次，那就相当于一百个精心构思的问题或后续追问。

If you're doing it like literally a 100 times a day, that's like a 100 well thought out questions or follow-up questions Yeah.

这样你就能超越世界上99.9%的人了。

And you're just able to outpace ninety nine point nine percent of people in the world Yeah.

作为一个高中辍学生。

As a high school dropout.

没错。

Yeah.

还应该补充一点，我认为人类身上依然有太多宝贵建议值得汲取。

And it should be added like, I think there's still so much valuable advice to be be had from humans.

要知道，当涉及到观点之类的事情时，如果你想想模型是如何训练的，它们基于互联网上的所有数据进行训练，而这些数据五花八门。有时候它们可能会产生奇怪的观点，对吧？

There's like still, know, when it when it comes to like opinions and and things like if you imagine about how how how models are trained, they train all data on the Internet and there's a bunch of different And you know, sometimes they might have weird opinions, right?

我是说，与真实且最优秀的人合作仍然具有很大价值。

I mean, there's still a lot of value in working with like the real and best people.

是的。

Yep.

但现在你通过AI能获得其中95%的价值。

But you can get like 95% of that now, which is the AI.

所以要与最优秀的人合作，获取他们的反馈，同时无论走到哪里都要不断向AI提问，从而对你想要解决的任何问题和想要学习的任何概念建立深刻理解。

So work with the best people, get feedback from them, but also constantly query AI wherever you go to build very deep understandings of any problem you wanna solve and any concept you wanna learn.

没错。

Yep.

你当时正在向这些资深工程师学习。

You were learning from these senior engineers.

你在不同的公司做合同工，比如不同的YC公司或各种不同的机会。

You're contracting at different companies, like different YC companies or like different all all these different opportunities.

如果你没上过学，也没有高中文凭，是怎么最终来到美国的呢？

How did you end up coming to America if you didn't go to school, if you didn't have a high school diploma?

是的。

Yeah.

最开始是加入了一家叫DataLand的公司。

It started out with joining a company called DataLand.

我们做的有点像Airtable，但性能更强，更以开发者为中心，嗯。

We're doing kind of like a air table, but way more performant and like developer first and and yeah.

可以说是一个可扩展的Airtable。

A scalable air table, you could say.

那是个非常重要的决定。

And that was a very important decision.

我在那里和一位极其优秀的工程师共事。

I I I was working there with with an engineer who's extremely talented.

他非常喜欢教导别人。

And he just loved to teach people.

哦。

Oh.

而且他喜欢追求完美的代码，这对我来说再好不过了，因为你知道，我写代码时他会给每个PR（代码审查）写上百条评论。

And he'd love having perfect code, which is perfect for me because, you know, I write code and he'd just do like a 100 comments per PR.

哇。

Wow.

有一次我不得不打电话问他，'嘿，你觉得这个怎么样？'

And I had to call him at a time and be like, hey, what do think like this?

他非常擅长从第一性原理出发，解释为什么某段代码要那样写。

And he would be really good at like explaining the first principles reasons for why some color was written in a certain way.

后来有段时间，你知道，我在瑞典远程工作，而他们在纽约，我当时就想，'我真的想去美国'。

And at some point, you know, I was working remotely from Sweden and they were in in New York and I was like, yeah, I really wanna go to The US.

我想这就是我最初想去美国的契机。

And I think this is where I first wanted to go to The US.

最终没能成行是因为公司转型做了其他业务，所以我决定离开。

This ended up not happening because the company pivoted and to something like so something else I decided to leave.

但那算是我的第一次。

But that was like my first.

我在那里启动了一个流程，申请了一种叫J1签证的东西，更像是你可以称之为实习签证的类型。

Like, I started a process there, something called a j one visa, which is more like a you could say like an internship visa.

是的。

Yep.

因为我们当时都挺年轻的。

Because we were all pretty short.

就像，是的，我拿不到O1签证。

Like, yeah, I can't get an o one visa.

根本不可能。

There's like no way.

而且你知道，你不需要赢得诺贝尔奖，但需要满足各种杂七杂八的条件。

And it's, you know, you don't need to win the Nobel Prize, so you need like all these random things.

我当时就说，对啊。

I was like, yeah.

根本不可能。

There's no way.

比如，我怎么可能做到这个？

Like, how how how could I do this?

然后我花了很多时间思考接下来要做什么。

And then I ended up spending a lot of time trying to figure out what I wanted to do next.

这就像是我去旧金山的时候，持ESTA签证待了几个月，到处和人交流，试图搞清楚大多数人在这里做什么。

And this is one of the like, when when I went to San Francisco, and I was here on the ESTA visa for a couple of months just talking to people, trying to figure out, like, okay, what's most people do here?

比如，最酷的公司有哪些？

Like, what's coolest companies?

最后我加入了Midjourney。

And there I ended up joining Midjourney.

加入Midjourney后，感觉就是，对。

After joining Midjourney, was like, yeah.

好的。

Okay.

现在我或许我们可以搞定O1签证。

Now I maybe we can do o one.

结果发现O1签证有非常多创造性的获取途径。

And and turns out that o one visa, there's like so many creative ways you can get an o one visa.

非常多创造性的方式。

Very many creative ways.

比如，我们申请O1签证时就用到了我在Stack Overflow上的帖子。

For example, one thing we used for my o one visa was my Stack Overflow posts.

我记得我表哥说过：你在浪费时间回答一堆Stack Overflow问题。

I remember my cousin telling me, oh, you're wasting your time answering a bunch of Stack Overflow questions.

我当时就想：你不懂。

I was like, you don't know.

说不定哪天就能派上用场呢。

Maybe it's helpful at some point.

结果证明，Stack Overflow的帖子确实可以作为证明材料——你看，这不就用上了，伊万。

And turns out, Stack Overflow posts can be counted as, yeah, here we have it, Ivan.

这是我在Stack Overflow上关于此事的帖子，你可以

Here's my post about it on So you can

利用Stack Overflow的帖子来满足学术出版标准。没错。

use Stack Overflow posts to get the academic publishing criteria Yeah.

为你的O-1签证。

For your o one.

这是完全合理的。

Which is legitimate.

比如，我的帖子有数百万次的浏览量。

Like, I I have, like, millions and millions of of of views.

很多同行会评审你的帖子。

A lot of, peers will review your your posts.

他们非常严格。

They're very strict.

他们会给不实内容点踩并删除。

They will downvote and remove anything that's not true.

如果你获得赞同票，说明你帮助了一群人，这正是标准所在。

And if you get upvotes, you you're helping a bunch of people, which was the criteria.

比如，你帮助过别人吗？

Like, have you helped people?

我觉得像GitHub或Stack Overflow这样的平台，确实是为自己辩护的一种非常有创意的方式。

I think like with GitHub or like Stack Overflow, it's definitely a very creative way to argue for your own one.

是啊。

Yeah.

那么你是怎么...Midjourney是最大也是最好的人工智能图像生成公司。

And so how did you Midjourney is one of the biggest and best is the best AI image generation company.

你是怎么最终在那里工作的？

How did you end up working there?

是啊。

Yeah.

还挺有意思的。

Was kinda interesting.

我是说，要在职业生涯中确定性地达到某个目标极其困难。

I mean, it's extremely hard to deterministically go somewhere in your career.

某种程度上你想做的事，你知道，非常老套，就是大家都在说的那些。

And kinda what you wanna do, it's, know, very cliche, what everyone's saying.

对吧？

Right?

但就像是，你需要在各处都创造许多小机会，明白吗？

But it's like, you wanna have, like, a bunch of small chances everywhere, right?

你就是要广撒网。

You just wanna go wide.

你要展示自己做过的项目，确保有非常出色的演示作品，尽可能多接触人，比如参加活动并当场请人引荐，确保他们就在活动现场完成介绍。

You wanna post things that you've done, you wanna make sure you have really good demos, wanna reach out to as many people as possible, like go to events and like ask people for intros and make sure they do the intro at the event.

最关键的是要立刻行动，比如：'哦你现在就想介绍我？'

Really first, it's like, okay, oh you want to intro me?

那就现在当场介绍吧。

Well yeah, let's do it right now.

一个区别。

A difference.

采取行动以确保其真正发生。

Action so that it actually happens.

是的。

Yeah.

你确保结果，这是一个高度自主性的举动。

You make sure the results It is a very high agency move.

没错。

Yeah.

同时要非常明确你能为他们提供什么价值，并确保他们明白你不是无名小卒。

And also be very clear with how you can give them value and make sure they understand that you're not a nobody.

对吧？

Right?

我曾经是无名小卒，但当你展示任何东西时，比如，哦，我做了这个——例如我做了一个叫Fast Grid的东西。

I was a nobody but the moment you show anything at all like, oh, I made this like something I made for example is this thing called Fast Grid.

这就像是一个性能超强的网页表格，你看，每次当你和那些你认为能帮到你的人交谈时，你得确保向他们展示你的价值所在，对吧？

It's like a really performant web table and you just showed you know, anytime you talk to someone that that you think is relevant that can help you, you know, make sure you you show them that you're relevant, right?

哦，我做了这个超酷的东西，你应该看看。

Oh, I built this really cool thing, you should see it.

他们总是会说，哇，这真的太酷了。

And they're always like, oh wow, this is really cool.

然后突然间，你会发现他们有一堆创业的朋友。

And now suddenly, you know, they have a bunch of friends that start startups for example.

现在他们想把你介绍给这些人，因为他们见识过你的真本事。

And now they want to interview you to them because they have seen that you know things.

就像每个人都想帮你一样。

Like everyone wants to help you.

前提是你要先确保自己能展示出价值，这样他们才能从你这里获得社交资本

If you first can make sure you know that that that you can show them because they can get get a bunch of like you know social value from you know

做个引荐。

Doing an intro.

是的。

Yes.

为了介绍。

For intro.

哦，你雇了我介绍给你的人。

Oh, you hired someone I introduced to you.

对吧？

Right?

而且他们是个好员工。

And they're a good hire.

如果你是个无名小卒，如果你来自某个不知名的地方，比如瑞典的某个偏远小镇。

If you are a nobody, if you are from the middle of nowhere like how, you know, you're from the middle of nowhere in Sweden.

如果你是个无名小卒，你要如何向重要人物展示你的价值？

If you're a nobody, how would you go about showing your value to someone important?

我给大家的首要建议是做一个极其简单的演示。

The number one thing I recommend to people is making a demo that is super super simple.

实际上，由于诸多原因，要制作一个好的演示确实非常困难。

It's actually really hard to make a good demo for a lot of reasons.

大家都觉得难是因为他们认为需要制作一个复杂的演示，而自己又缺乏相关技能。

Everyone thinks it's hard because they need to make a demo that is hard and they don't have the skills.

这种想法大错特错。

This is very not true.

你可以做得非常简单，甚至不需要太多编程知识就能做出非常酷炫的演示。

You can make very simple like, you don't need that much code knowledge to make a really cool cool demo.

制作演示最难的部分是确保人们能在三秒内明白你会编程。

The hard part of making a demo is making sure that people understand why you can code within three seconds.

要知道，每个岗位可能都有上百名申请者。

You know, you have like a 100 like applicants for something.

如果你只提交一个链接，他们点开时你只有一次机会，对吧？

If you apply with one link and they press the link and you know you have one shot, right?

关键是要确保你制作的演示足够惊艳，让人一眼就能看懂内容——这确实很难。

Like making sure you build a really cool demo where people understand what they're looking at which is really hard Yeah.

而且要让人们明白你是个非常优秀的工程师，这确实很难，但这就是关键所在。

And where people understand that you're a really good engineer which is really hard, but then you're there.

我是说，这就是他们想看到的全部。

I mean that's all they wanna see.

公司只想赚钱，你向他们展示如何赚钱、证明你会编程，他们就会雇佣你。

I mean companies just wanna make money, you show them how to make money that you can code and they'll hire you.

然后你可能会说，但他们只招有学历的人。

And then you might say, oh, but they only hire people with degrees.

确实，因为根本没人向他们证明过自己能胜任这份工作。

Well yeah, because literally no one has ever showed them that they can do their work.

他们会说，哦，我有过这些实习经历。面试官就会问，好吧，你在那里做了什么？

They're like, oh, I had this internships And the interviewer would be like, okay, what did you do there?

哦，我优化了流程，将效率提高了30%。

Oh, I streamlined pipelines and made things 30% more efficient.

面试官就会说，好吧，这根本说明不了任何问题。

And the interviewer was like, okay, well, that tells me literally nothing.

好的，还有别的吗？

Okay, what else?

你都做过些什么？

What have you done?

哦，我上了哈佛，成绩优异。

Oh, I went to Harvard, I have the best grades.

但我还是不知道你能不能胜任这份工作，对吧？

Well, I still don't know if you can do the job, right?

哦，但我有很多课外活动，我是辩论冠军。

Oh, but I have all this extracurricular, I was a debate champion.

你开始列举所有这些父母和周围人会告诉你的东西。

You start going on about all these things that the parents will tell you, people around you will tell you.

这些都不重要。

Nothing matters.

唯一重要的原因是没人能证明自己能做事。

The only reason it matters is because no one can show that they can do anything.

于是他们开始听这些类似代理的东西。

So then they start listening to these like proxy things.

一般来说没人在乎。

No one generally cares.

现在确实有些人在乎。

Now there are people who actually cares.

对吧？

Right?

这些人是谁？

Who are these people?

首席运营官永远不会在乎。

The CO will never care.

他们永远不会在乎。

They will never care.

他们只想着赚钱。

They just wanna make money.

对吧？

Right?

这很棒。

Which is great.

你只需要想，嘿，我能赚钱。

You just, hey, I can make money.

哦，太好了。

Oh, great.

这里有个任务。

Here's a task.

明白吗？

Know?

一切都很完美。

Everything's perfect.

当你离CO越远，事情就越困难，因为人们开始失去为公司做最佳选择的动力。

When the further away from the CO you comes, the harder it becomes because people start losing incentives to do the best thing for the company.

那么取而代之会出现什么情况呢？

And instead, what comes up instead?

他们只是不想搞砸。

Well, they don't want to fuck up.

他们只是不想输。

They just don't want to lose.

是啊。

Yeah.

所以他们要怎么招人，如果招错了人就拿不到钱？

So how do they hire someone that if they are a bad hire, they will not get any money?

你会看传统资历，比如他们是否毕业于顶尖学校。

You go through conventional accolades like they went to the top school.

正是如此。

Exactly.

哦，他们毕业于顶尖学校。

Oh, they went to the top school.

我怎么能知道他们不行呢？

How could I know they were bad?

对吧？

Right?

是啊。

So Yeah.

这样招聘者就不会犯错了。

Then the recruiter doesn't make a mistake.

对吧？

Right?

所以这就是，你知道，要避免的情况。

So that's the you know, the thing to to avoid.

比如要避开那些对公司没有利益驱动的人。

Like avoid people who have no incentives in the company.

所以通常要避开公司内部的招聘人员。

So generally avoid recruiters at a company.

是啊。

Yeah.

他们甚至都不懂技术。

Like they're not even technical.

他们根本分不清好坏。

They can't even know if they're good or bad.

他们只会关注那些间接信号。

They will only go on all these like proxy signals.

所以人们总说需要学历，因为这就是他们评判的全部标准。

And that's why people say I need a degree because you know, that's everything they say.

人们不知道其实可以直接和人交流。

People don't know that you can just talk to people.

去参加技术人员的活动就行，我认识的每家初创公司都想雇佣那些主动性强、学习能力强的人。

Just go to an event with tech people and like every single start up I know wants to hire people who have high agency and can learn things.

没错。

Yeah.

说真的，如果你真的很擅长使用ChatGPT，你在某个活动上看到这些人，主动过去跟他们交谈，给他们一些建议，然后说'嘿，我们能不能免费合作一周试试看？'

Literally, if you're really good at using ChatGPT, you saw one of these people at a random event, you went up and talked to them, and you give them some advice, and you're like, yeah, I'd like to can we try working together for a week for free?

这会超级有趣。

This would be super fun.

就像，我刚刚想到的这些随机点子可以和你一起尝试，你这边不需要任何承诺，也不用花时间，就当免费测试一下我是否靠谱。

Like, have these random ideas I just came up with that I can work with with you that are, you know, no commitment from your side, no time from your side, just like get a free data point if I'm good or not.

对。

Yeah.

而且他们100%会答应。

And a 100% of them would say yes.

他们会说'哦，太棒了'。

They're like, oh, great.

我什么都不用做就能看出你水平如何。

I don't need to do anything and I can see if you're good.

没错。

Yeah.

比如说，如果你是一个通常懂很多东西的人，甚至不需要懂太多。

Like, if you're generally a person who knows things and like, not not even knows things.

如果你只是一个会使用ChatGPT的聪明人，你明天就能找到工作。

If you if you're just a smart person who can use Chateappetik, you can get a job tomorrow.

然后人们就会说，但这有风险。

And here's where people are like, but it's a risk.

对吧？

Right?

我上不了大学。

I won't get into college.

我做不了所有这些事。

I won't do all these things.

其实根本没有风险。

There is no risk.

你甚至可以用零风险的方式来做这件事。

And you can even do it the risk free way.

直接申请大学。

Just apply to college.

你可以去上大学。

You can go to college.

对。

Yeah.

申请大学。

Apply to college.

等你上了大学，再申请工作。

And when you're in college, apply to jobs.

对吧？

Right?

而且完全没有风险。

And there's zero risk.

你只需要多花些时间申请其他工作。

You just put some extra times into applying to other jobs.