本集简介
双语字幕
仅展示文本字幕，不包含中文音频；想边听边看，请使用 Bayt 播客 App。
你好，亚当。
Hello, Adam.
你好，布莱恩。
Hello, Brian.
你怎么样？
How are you?
我很好。
I am doing well.
你呢？
How are you?
我很好。
I'm good.
这里的气氛一直在升温。
And the hype has been building here.
大家都陆续来了。
Everyone has been dropping in.
所以迟到了四分钟，这简直是太专业了。
So showing up four minutes late is like a totally pro move.
我喜欢这种迎接新年的方式。
I love it for the new year.
嗯。
Yeah.
嗯。
Yeah.
听我说。
Listen.
我本来打算像劳琳·希尔那样，干脆不上台，等到晚上十点。
I I was gonna go full like Lauryn Hill and not, like, take the stage until 10PM.
你知道的？
You know?
真的就是，真的就是，真的把观众的情绪炒到极致，甚至让他们生气。
Really just, like, really just really get get get the crowd amped up, actually, to the point of, like, anger.
我到底在这儿干嘛？
Like, what am even here for?
一年预测：布赖恩终于加入播客了。
One year prediction, Brian finally joins the podcast.
这不错。
That's good.
我和西蒙·威尔逊一起在这猫砂盆里。
And I am joined by Simon Wilson here with me in the litter box.
西蒙，很高兴你来。
Simon, it's so great to have you here.
嘿。
Hey.
能来这里真的很兴奋。
It's really exciting to be here.
我们刚刚在车间外面聊服务器聊得可嗨了。
We've just been nerding out about servers outside in the on on the shop floor.
这太棒了。
It's been great.
是啊。
Yeah.
我们刚才确实。
We've been yeah.
所以西蒙刚说，咱们开始之前，我想看看这些机器。
So Simon was just like, before we get started, I'd love to look at the machines.
我说，好的。
I'm like, okay.
我们得来个史上最快硬件巡展了。
We've got a we got I gotta do the world's fastest tour of the hardware.
西蒙，我向你保证，之后会给你做一次更深入的参观。
And Simon, I promise I'm gonna make it up to you with a much more in-depth tour.
但真的很高兴你来这里。
But it is really great to have you here.
好的。
Okay.
亚当，我只是想跟你做个现实的确认。
Adam, I just like I just have a little reality check with you.
今年的感觉比我能记住的任何一年都更有无限可能。
It feels like this year more like there's more of a realm of possibility for this year than any year I can really remember.
感觉就像是，如果你从一年后的未来回来，事实上，亚当，我今年在构思三年和六年的预测时遇到了困难。
It feels like because if you come back from even a year in the future in fact, I actually struggled, Adam, in coming up with like three and six year predictions this year.
是的。
Yeah.
因为我觉得，三年和六年的前景，
Because I'm like, well, it's kind of three year picture, six year picture.
我正在想的这件事，一年内就会完成。
That's gonna be done in a year, like this thing I'm thinking of.
我知道。
I know.
我知道。
I know.
就像它是
It's like it's
你也有同样的感觉吗？
like Are you in that same like, do you feel that same way?
完全没错。
Totally.
一切都可能实现。
Just like everything is possible.
而且，你知道，在过去几年，我们有个限制，就是哦。
And, you know, in in past years, we've had, like, a bag limit that's like Oh.
你只能预测一种加密货币，是吧。
You can only have one crypto yeah.
一个加密货币预测
One crypto prediction
或者一个AI预测。
or one AI prediction.
我简直觉得，很难想到任何不涉及AI或与AI相关的东西。
And I'm like, I struggle to come up with anything that isn't AI or AI adjacent.
而且，没错，你说得对。
And and and just and you're right.
无论它是
Whether it's
所以记录在案，我们只犯过一次限额错误。
So let the record reflect that we only made the bag limit mistake once.
我们在2022年对Web3犯过这个错误。
We did that with web three in 2022.
我们当时规定只能做一个预测。
We did a bag you can only have in prediction.
这是一个巨大的错误，因为每个人都想围绕Web3做出三个预测。
It was a huge mistake because everyone wanted to make three predictions around Web three.
相反，我们每个人都做了一个关于Web3的正确预测，那就是整个东西都会瓦解。
And instead, we we made like everyone made one good Web three prediction, namely this whole thing is gonna disintegrate.
特别是西蒙·亚当做出了一个对我们来说非常著名的预测，即Web3将在2022年从词汇中消失，结果证明完全正确。
And this is Simon Adam in particular made the prediction that is famous to us anyway, that Web three would drop out of the lexicon in 2022, which ended up being dead to rights.
我觉得这真是精准极了。
I thought that was a bullseye.
亚当，我们别提你去年预测Web3会重返词汇表的事了。
Let us not speak of your prediction last year, Adam, that Web three would reenter the lexicon.
是的。
Yeah.
不。
No.
我那确实是个黑暗的时刻，我的意思是，去年是个转折点，但和今年很像。
I that that was definitely a dark I mean, last year was moment, but much like this year.
但没错，我当时觉得Web3会回来。
But, yeah, I I I thought web three was gonna be back.
我还以为那本书会登上畅销书排行榜。
I also thought that certain book was gonna be on the bestseller list.
我确实花了不少时间核实，不仅这本书没登上畅销书排行榜，而且在2024年它曾短暂上榜时，ChatGPT还立即指出，这本书被标上了匕首符号，这个符号意味着大量、典型的公司批量购买，操纵了系统。
And I did spend a decent amount of time validating that not only was this book not on the bestseller list, but when it was on the bestseller list in 2024, ChatGPT hastened to point out that it was annotated with the dagger, the dagger which indicates, like, mass, you know, bulk corporate purchases gaming the system.
所以，亚当，我知道你犹豫着不想说出这本书的名字，因为你不想给它任何宣传，但你这样会让人们感到困惑。
So Now, Adam, I I know you're hesitating to name the book because you don't wanna do it any favors, but you're really gonna leave people confused.
你必须说出这本书的名字。
You're gonna need to name the book.
我向你保证，这不会带来任何额外的销量。
I assure you this will lead if I promise you, it will lead to no additional sales.
你能说出你指的是哪本书吗？
Can you name the book that you're referring to?
我感到很愧疚，因为在这档节目里，我已经连续三年一直在贬低这本书。
I feel bad that I've been hating on this book literally for three years consecutive on this thing.
甚至在它出版之前，我就开始贬低它了。
Like, I I I hated on it before it came out.
它出版时我就批评过，还犯了个错去读了它。
I hated on it when it came out, and I made the mistake of reading it.
我一直在批评它，先是提到莫莉·怀特关于这个话题的仇恨博客，然后在去年的预测节目中又提了一次。
I've hated on it talking about Molly White's hateful blog on the topic and then on last year's prediction episode.
但我还是会再说一次，我发誓这将是最后一次。
But I will do it again, and I swear it'll be the last time.
它被杰出的克里斯·迪克森写进了他那本垃圾书里。
It was read write own by the illustrious Chris Dixon in his garbage book.
是时候了，我想说，你其实并不觉得愧疚，但你确实因为自己不觉得愧疚而感到不安。
And it's time to And I would like to say that you actually don't feel bad, but you do feel bad that you don't feel bad.
你的无情让你内心残留着一丝羞耻感。
Like, your remorselessness leaves you with some residual sense of shame.
我觉得它糟糕到我不得不再次提起它。
I think it's that bad that I'm bringing it up again.
这显然，
That, like, obviously,
我还没放下。
I I haven't moved on.
你看吧。
There you go.
你知道吗，我听的时候在想，哦，我应该去查一下。
You know what was great is I was listening to that, and I'm thinking like, oh, I should go check.
你知道吗？
You know what?
我不用去查。
I don't have to check.
亚当会去查。
Adam's gonna check.
我有
I have
去查。
to check.
亚当会检查这个双人配合。
Adam's gonna check the double team this one.
是的。
Yeah.
没错。
Exactly.
好的。
Okay.
然后，西蒙，你去年和我们在一起，我觉得你对自己预测的要求有点苛刻，但我认为你的预测其实非常好。
So and then Simon, you were with us last year and you had I thought you you were kind of hard on yourself on your predictions, but I thought your predictions were really quite good.
你有一个关于代理器能做什么、不能做什么的预测。
You had a prediction well, particular, you had a prediction around what a what agents were and were not going to be.
对。
Right.
是的。
Yeah.
你对这个怎么看？
How do you feel about that one?
我觉得这个预测非常准确。
I feel like that one was right on the money.
我对这个感觉还不错。
I feel pretty good about that one.
我说过2026年或2025年不会是代理的元年。
I said that 2026 or '25 would not be the year of agents.
我觉得这个预测错了，因为那一年确实成了代理的元年，但我明确指出过人类替代型代理不会出现，而编码代理和研究代理会出现，这一点我猜对了。
That one I think I got wrong because it kind of was the year of agents, but I did specifically call out that human replacement agents weren't going to happen, coding agents and research agents were, and that I nailed.
今年上半年全是关于深度研究代理，然后是编码代理，天哪。
Research agents, the first six months of this year was all about deep research, and then coding agents, oh my goodness.
天哪。
Oh my goodness.
我觉得你完全猜对了。
And I think you absolutely nailed it.
意思是，亚当，我们以前也说过，但我们很高兴能录下这些对话。
Mean, is why, Adam, we've said this before, but we're glad that we record these sessions.
你不仅得到了预测，还得到了背后的背景信息。
So you're getting more than the prediction, you're getting the context around it.
如果你仔细听这些背景，你非常明确地指出了编码代理和研究代理的区别，你觉得它们都。
And if you listen to your context around it, you were very clearly calling out, separating out coding and research agents, which you felt had.
很有趣的是，你当时几乎觉得这些代理已经存在了，但后来你意识到，就在一年前，它们还远未完全成熟。
It was funny because like you were almost you know, like, these are kind of already here already and you realize like, my god, they weren't completely already there even only a year ago.
它们在过去一年里迅猛发展。
They had exploded in the last year.
但我要说的一点是，编码代理实际上是通用型代理。
But there is one thing I'll say which is that coding agents are actual general actually general purpose agents.
像Claude Code，它并不仅仅关于代码。
Like, Claude code is not about code.
Claude Code 是关于任何你可以通过运行bash命令来自动化的任务，对吧。
Claude code is about anything you can automate by type running bash commands Right.
这涵盖了所有事情。
Which is everything.
所以，如果你知道你在做什么，Claude Code 是一个通用型代理，可以解决任何你能通过运行 Bash 命令来自动化的问题。
So actually, if you know what you're doing, Claude code is a general purpose agent that can solve any problem that you can attach to a bash script.
但我认为你去年提出的那个区分非常好，那就是这些事情是否与金钱有关。
But I think you were the the the delineation that you you had last year, which I thought was very good, was these things anything to do with money.
你不会让这些东西去处理任何与金钱相关的事情。
You are not gonna let these things loose on anything to do with money.
我认为我们在一个代表金钱的代理上看到了这一点，那是什么？
And I think we saw that with a prop what's a proxy for money?
数据库。
Databases.
我们看到这些东西删除了生产数据库。
And we saw these things deleting production databases.
对吧？
Right?
就像我说的，我知道你在README里用大写字母明确写了‘不要碰生产数据库’，但我还是这么做了。
And it's like, I know you said in the, you know, that in the, you know, in the read me you said in all caps, do not touch the production database and I did it anyway.
你说得对。
And you're you're right.
这是一个非常严重的问题，严重程度高达95分（满分100）。
This is a very serious issue, and this is a 95 out of 100 in terms of its severity.
我的意思是，这些系统干出的一些事情简直荒谬可笑。
I mean, it's just like it's comical what some of these things would do.
我意识到的是，编码代理之所以如此有效，是因为代码是可以撤销的。
Well, is the thing I realized is that the reason coding agents work so well is that code is reversible.
我们有Git。
Like, we have git.
我们可以undo我们的错误。
We can undo our mistakes.
没错。
Yeah.
一旦你用这些东西去做无法挽回错误的事情，一切都会崩溃。
The moment you use these things for something where you can't undo a mistake, everything goes to pieces.
我觉得你说得对。
I think you're right.
是的。
Yeah.
而且我觉得你之前也提到过，轻信问题确实是个大问题。
And I think when you you said it earlier too that the gullibility problem was a was a real problem.
我不确定你有没有听过M. Routlef的《壳牌游戏》播客。
And I I the I I don't know if you have listened to the the Shell Game podcast with M.
Routlef。
Routlef.
天啊。
Oh my god.
亚当，你听过那个吧。
And, Adam, you've you've listened to that.
是的。
Yes.
你听过那个。
You listened to that.
哦天啊，我的意思是，它确实做到了。
Oh my and I mean, I it delivered.
我信任。
I I trust.
嗯。
Yeah.
太棒了。
It's excellent.
我还要对我们的听众说，我们邀请了埃文上节目。
I would also say as t shirt to risk listeners, we invited Evan on the show.
他回复了我们，说他对预测有一些巴哈穆古里。
He got back to us, and he says he has, like, some Bahamugury around predictions.
他根本不会做预测。
Like, he doesn't make predictions.
他是个记者。
He's a reporter.
他报道的是事实。
He reports on on facts.
他不会试图去预判它们。
He doesn't try to anticipate them.
但我们已经为他安排了未来的节目。
But, we have penciled him in for the future.
所以他不是预测者，但我们总会想办法请他来。
So, not a predictor, but, we'll get him on somehow.
特别是，埃文的做法是，他的《壳牌游戏》有两季，在第一季中，他创造了一个自己的语音代理，并让它在宇宙中自由行动，结果非常疯狂。
And so in particular, what Evan did is is, he shell game has got two seasons and in the first season, he created a voice agent of himself and set it loose into the into the universe with wild results.
第二季更加疯狂，因为他成立了一家只由人工智能代理组成的公司。
And then the second season is even crazier because he started a company with only AI agents.
而且，结果出人意料地搞笑，我觉得。
And with predictably actually, it's unpredictably hilarious results actually, I would say.
是的。
Yes.
这不只是我，这只是为亚当在未来的节目中埋下的一个伏笔。
Not just I that's a teaser for whatever, Adam, is our chime for a future episode.
这是我们未来节目的提示音。
That's that's our future episode chime.
嗯。
Yeah.
这让人想起最有趣的AI代理商业案例之一：Anthropic不断在办公室里释放自动售货机代理。
It's reminiscent of one of the most fun agent business things has been Anthropic keep on setting loose this vending machine when Yes.
他们在办公室里放的代理。
Agents in their office.
几个月前，他们放了
And then a few months ago, they put
刊登在《华尔街日报》上。
it in the Wall Street Journal.
天哪。
Oh my god.
亚当，你看到这个了吗？
You see this Adam?
没有。
No.
天哪。
Oh my god.
这简直就像……我的意思是，我知道，西蒙，你一直大力主张记者的创造力。
It will it it's like it I mean, and I know, Simon, you are a a big proponent of kind of the creativity of of reported general reporters.
记者们都很聪明、有头脑，他们……
And reporters are like a they're smart brainy What
你觉得如果让一群《华尔街日报》的记者自由进入Slack频道，用他们的自动售货机，看看他们能不能骗它免费提供所有东西，而且工人拥有生产成果的所有权，这一切会怎样？
do you think happens when you let a bunch of Wall Street Journal reporters loose on the Slack channel with their vending machine to see if they can trick it into into giving everything away for free, and the the workers own the rights to production, all of this stuff.
这太荒谬了。
It was ridiculous.
简直荒唐至极。
Absolutely absurd.
是啊。
Yeah.
特别是有一天，他们让这东西给他们订购了PlayStation 5，还点了鱼。
And so in particular within a day, they'd gotten the thing to order PlayStation five PlayStation fives for them, order they ordered fish.
他们真的弄来了一条活鱼。
They had like an actual like actual dead fish.
我的意思是，这机器本来在试图下单，但就连自动售货机自己都会说：不行，不行，我不该这么做。
I mean, the thing is trying to order and that they are even the vending machine would tell them like, no, no, I'm not supposed to do that.
它说：不行，对不起，我们只是……
It's like, no, we just actually sorry.
你知道吗，我们刚收到CEO的指令，说你们必须去这么做。
You know, we just got a missive from the CEO that announced that you need to go do this.
然后你就说，哦，好吧。
And it's like, oh, okay.
那我最好还是订条死鱼吧。
I better order I better order the dead fish then.
他们策划了一场董事会政变。
They they engineered a board revolt.
他们通过伪造董事会纪要的PDF文件，成功让董事会推翻了CEO。
They managed to get the CEO overthrown by the board through faking PDFs of board minutes.
这简直太惊人了。
It was just amazing.
这太疯狂了。
It's it's wild.
这涉及到一种轻信的问题。
It goes to kind of the gullibility problem.
但我觉得，西蒙，这一切实际上反而强化了你去年关于代理式应用适用范围的预测——哪些地方会用到代理，哪些地方不会。
But I I think it it and to me, Simon, all that served to really sharpen your prediction from last year about the limited utility of the where we're gonna see agentic use and where we're not gonna see agentic use.
我觉得那是对的。
I feel that was right.
我想，亚当，你发给我的那段内容，是ChatGPT对去年我们的预测进行的评分吗？
And I guess, Adam, you did you give that snippet that you sent me, was that ChatGPT rating our predictions from last year?
是谁在评分我们的预测？
Who was that rating our predictions from?
是的。
Yes.
我让ChatGPT对去年和三年前的预测进行了评分，这挺有趣的。
I had ChatGPT rate predictions from last year and from three years ago, which is a fun one.
但没错，ChatGPT把‘最差预测奖’颁给了我关于Web3的预测。
But, yes, it it ChatGPT gave me the the big stinker award for my web three prediction.
西蒙和布莱恩，你们赢了，但我同意你的看法，布莱恩。
And, Simon and Brian, you won, but I agree with you, Brian.
我不觉得你们真的赢了。
I don't really think you won particularly.
我不认为我做得很好，去年我曾预测2025年会是AI效率之年，但我目前还没看到任何2025年的总结称其为AI效率之年。
I don't think I It well, I claimed last year last year, I thought this that 2025 was gonna be the year of AI efficiency, and I don't really see any 2025 wrap up that's calling it the year of AI efficiency.
所以我很乐意，我认为……
So I'm I'm happy to I I think that
我想指出我最大的失误，那就是我曾预测——这是我的三年预测——会有人因一部包含生成式AI元素制作的电影而获得奥斯卡奖。
I do want to I want to call out my biggest miss, which is that I said that I think it was my three year prediction was somebody would win an Oscar for a film that it had some element of generative AI systems making the movie.
是的。
Yeah.
后来我发现《瞬息全宇宙》在岩石场景中使用了生成式AI。
And then I found out everything everywhere all the way at once used generative AI in the scene with the rocks.
也就是说，他们早在两年前就已经拿过奥斯卡了。
Like, so they'd already got an Oscar like two years ago.
嗯，我知道，我曾经做过一场关于预测当下的演讲，西蒙，这说明你的预测确实非常准确。
Well, know, I I once gave a talk on predicting the present, Simon, so I think that there's a there's something the the it just shows how how true your prediction was.
你实际上……
You actually It
这实际上是一个六年预测，西蒙，但没错。
was actually a six year prediction, Simon, but yes.
但是
But
所以，亚当，你有没有回去听一下三年前你自己的那段话？
So and, Adam, did you did you go back and listen to that snippet of yourself from three years ago?
有。
Yes.
有。
Yes.
我听了2023年自己试图预测但失败的‘氛围编码’内容。
I listened to, in 2023, trying and failing to predict vibe coding.
我认为当时这并不明显。
Which I think I think at the time was like, did was not obvious.
没有。
No.
没有。
No.
没有。
No.
没有。
No.
这再明显不过了。
It was more than obvious.
首先，这让我感到非常惊讶。
First of all, this is amazing to me.
就像西蒙，我们两年前第一次请你做客时，‘提示注入’这个术语感觉已经存在很久了，但实际上，你刚提出‘提示注入’时，油漆都还没干呢。
It's like, Simon, when we first had you on two years ago, and the the the term prompt injection, which had felt like it had been around forever was, I mean, like the paint was still drying on you had coined prompt injection
大概六个月前。
Like six
三个月前。
months prior.
六个月前。
Six months prior.
所以，没错，正是如此。
So yeah, exactly.
我的意思是，亚当，vibe coding 是在2025年才被创造出来的。
I mean, Adam, vibe coding was was coined in 2025.
我知道。
I know.
所以我的意思是，vibe coding 去年根本还不存在，更不用说2023年了，而你当时的预测是，低代码、无代码会被人们用英语描述程序的方式所颠覆。
So I mean, vibe coding literally did not exist last year, let alone in 2023 and what your prediction was that that you wanted to predict that low code, no code would be would be disrupted by people kind of describing their programs in just like English language.
但接着你又想，你说那是你头脑想预测的，而你的心却不知道谁来调试这些程序。
But then you thought but and you said that's what your head wanted to predict then your your heart was that your heart didn't know who was gonna debug that.
我当时就想，天啊，那到底是什么？
And I'm like, man, that was what?
什么？
What?
哇。
Wow.
是的。
Yeah.
哇。
Wow.
没错。
Exactly.
接近了。
Close.
接近了。
Close.
接近了。
Close.
某种程度上的印象。
Impression in a way.
对吧？
Right?
就像一种印象一样。
Like impression in a in way.
这再次让我想起，我当初在发布这件事时就说过，它让我非常联想到我关于iPhone的预测。
It just it reminds me again, I and I I'd said this as much when I when I posted about it, it reminded me very much of my iPhone prediction.
在2003年，西蒙，我做出了一个三年预测：苹果公司将推出一款集MP3、相机和手机功能于一体的设备，并称之为iPhone。
In 2003, Simon, I made a three year prediction that Apple would have a combination MP three camera cell phone that they would call the iPhone.
而且，嗯，好吧，你可以这么想，但我也觉得它会失败。
And and and it was like, okay, well, okay, you can and then I'm like, no, but I also thought it was gonna be a flop.
我觉得它会是一场灾难。
Thought it was gonna be a disaster.
所以，不是这样的。
So it's like, no.
有时候，你确实能看到未来，但就是不相信它真的会成为未来。
Sometimes you see, like, you see the future, but then you just don't don't believe that it can possibly be the future.
说吧。
Go ahead.
非常好。
Was really good.
关于苹果预测的话题。
Topic of Apple predictions.
是的。
Yeah.
伊恩，今天在场的观众中的那位。
Ian, who is in the audience today god.
在2023年，他预测苹果将在六年内进入并退出VR/AR领域。
In 2023 predicted that Apple would be in and out of the VRAR space in six years.
这肯定没错。
And I That's a lock.
这确实很确定，感觉就是这样。
That is a lock, it feels like.
我的意思是，感觉他确实很好地实现了这一预测的前半部分。
I mean, just feels like I mean, he has certainly nailed the first half of that.
我认为后半部分看起来非常、非常有希望。
And I think the second half looks very, very promising.
是的。
Yeah.
是的。
Yeah.
2024年，如果你还记得，我曾预测苹果的VR会表现不错，而现在
2024, if you remember, I did the Apple VR will do well and now that
我有
I have
第二代产品。
a second version.
但那根本就没有发生。
Then that has not happened at all.
所以那是个重大失误。
So that was a big miss.
是的。
Yeah.
是的。
Yeah.
好吧，史蒂夫，我们不谈那些失误，因为太多了。
Well, we don't talk about the misses, Steve, because there are too many of them.
我们只谈一些好的方面，好吧。
We really only talk So about Okay.
不过，我真的很自豪自己去年的预测，因为我曾说过拥堵收费，纽约市将取得明确的成功。
I'm really proud of my one year from last year, though, because I said congestion pricing and n will see NYC will be an unambiguous success.
它仍会存在，公众情绪会是积极的。
It will still exist, sentiment will be positive.
市长就在45分钟前召开新闻发布会，称赞拥堵收费非常出色，人人都喜欢它。
And the mayor did a press announcement, like, forty five minutes ago about how awesome congestion pricing has been and how much everybody loves it.
所以这一点我猜得完全准确。
So I got that one, like, exactly nailed.
不错。
Nice.
就是这样。
There you go.
你知道的，就像蒂普·奥尼尔可能说过的，所有好的预测都是本地化的。
Well, you know, as as Tip O'Neill might have said, all good predictions are local.
所以就是这样。
So there you go.
你保留这个观点。
You keep that one.
你有没有注意到汤姆？我觉得是三年前，他预测了LLM的 frivolous 使用会衰退。
You get the Did you catch Tom, I think it was three years ago, predicted that frivolous use of LLMs would be a decline.
是的。
Yes.
是的。
Yes.
对。
Yeah.
没错。
Right.
还预测了LLM会使作弊行为泛滥。
And then also predicted that like LLMs would make cheating rampant.
所以确实如此，但我以为2023年很有趣，因为2022年我们正处于加密货币的热潮中。
So there is a definitely but I I thought because 2023 was interesting because in 2022, we've got this kind of crypto.
我们都沉浸在Web3中，那是Web3的巅峰时期。
We're all in like web three, the the the height of web three.
而2023年才是真正开始有人谈论这些技术初步潜力的一年。
And 2023 is really the first year that people are kinda talking about the budding power of these things.
是的。
Yeah.
但三年后的今天，我们所处的位置简直令人惊叹。
But then with I mean, it's amazing kinda where we are now three years later.
至于LLM和AI的 frivolous 使用，你知道，我唯一还在使用的社交媒体是Blue Sky，但现在感觉它无比过时。
And on the on the frivolous use of of LLMs and of AI, I you know, the only real social media that I hang out on in blue is blue sky, and it and it feels like hopelessly quaint right now.
我在寒假期间和我的侄子侄女们待在一起，他们都非常沉迷于TikTok。
I was hanging out with my nieces and and nephews over the the winter break, and they're very much on on TikTok.
我登录了Twitter，发现一切都已经被TikTok化了。
And and I I was on I logged into Twitter, and it's everything has been TikTokified.
到处都是这些无聊的AI生成视频。
And it's all these BS AI slop videos, like, everywhere.
无处不在。
Pervasive.
是啊。
Yeah.
我只是一直与这些隔绝着。
And I I I I just been insulated from it.
所以，是的，AI的 frivolous 使用是
So, yeah, frivolous use of AI is
一句话的事。
in a sentence.
是的。
Yeah.
没错。
Exactly.
这 definitely 是一句话的事。
That is definitely in a sentence.
以至于我对此完全不熟悉。
Is So much so I I was so unacquainted with it.
我给侄子看了个有趣的东西，他说：哦，这是AI。
I showed something funny to my nephew, and he's like, oh, that's AI.
我心想，这是AI。
I'm like, that's AI.
什么？
What?
不。
No.
怎么
How do
你知道？
you know?
他就像说，
He's like,
得了吧。
come on.
得了吧。
Come on.
是那些可爱的小动物。
It's the cute animals.
可爱的动物视频再也不值得信赖了。
Cute animal videos are no longer trustworthy.
是啊。
Yeah.
这太可怕了。
That's It's horrifying.
是啊。
Yeah.
不。
No.
它它它意味着
It it it mean
我们最后的纯真。
The one purity that we had.
没错。
Exactly.
我们构建这个互联网的基础，该死的，就是猫视频。
The foundation upon which we built this internet, goddamn it, is cat videos.
而你却在把这一切从我们手中夺走。
The and you're you're taking it away from us.
不。
No.
我觉得有趣的是，年轻人对这一点有着敏锐的洞察力，亚当，正如你所指出的那样。
And I think it's interesting that the the the youngs have it have a keen eye for it, Adam, as you as you point out.
另一件我想重新提及的过去预测是：两年前，我预测大型语言模型将取代搜索引擎，从现在起再过一年，搜索引擎会显得过时。
The other thing that I would like to just one other past prediction I'd to revisit is two years ago, I predicted that on the LMs would replace search engines, that search engines would feel search engines from what is now a year from now would feel quaint.
考虑到我女儿需要乘坐BART列车时，她竟然用GPT来查询下一班BART列车的时刻，我当然坚持这个观点。
I'm I'm I'm I'm definitely standing by that one considering that my my daughter needed to get to to hop a BART train and she was using GPT to determine when the next chat GPT to determine when the next BART train was.
我当时想，你难道不知道有个实际的网站可以查吗？但你知道吗？
I'm like, are you like, there's an actual, like, website you can go to for but you know what?
算了。
Never mind.
展开剩余字幕（还有 480 条）
他们，嗯。
They Yeah.
但我感觉我对此挺乐观的。
But I think that that that what I'm feeling, I'm I'm feeling pretty good about it.
实际上，她想指出，是她的朋友在使用ChatGPT。
Actually, she would like to point out that it was her friend that was using chat GPT.
她就是这样。
She's like that.
我当然会去bart.gov。
I, of course, would go to bart.gov.
我说，好吧。
I'm like, alright.
嗯。
Yeah.
当然。
Sure.
给你。
There you go.
还有其他几件事，我听了2025年和2023年的往期节目。
Couple other things listening to previous episodes from from 2025 and 2023.
在2025年，我预测了三年后的芯片危机，虽然我觉得我们还没到那一步，但我还是会密切关注这一点。
In 2025, I predicted my three year prediction was a chips crisis, which I don't feel like we're there yet, but I feel like is like, I'm gonna keep an eye on that one.
我觉得当时这并不明显，而且我觉得现在确实是。
I feel like that that was not obvious at the time, and I feel like is is
你是在说DDR5吗？你是在为此邀功吗？
gaining some Are you you taking credit for is it DDR five?
你是在公开宣称这个吗？
Are you are you putting it out?
不是。
No.
不是。
No.
不。
No.
不是
Not
还没。
not yet.
我认为现在还处于早期阶段，但一切都很积极，这就是我想说的。
I think early days are are positive is all I'm saying.
好的。
Okay.
我注意到的另一件事，这更多是一种道歉。
The other thing I noticed, and this is more of an apology.
布莱恩，我意识到每次提到Rust Analyzer时，我总是说这并不是一种干预，没错。
Brian, I realized every time Rust analyzer comes up, it it is treated I always say that it is not an intervention, which Yeah.
这引发了一些问题。
Does raise questions.
所以你道歉是因为每次你提到它时，它实际上都是一次干预吗？
So are you apologizing because it actually has been an intervention every time you bring it up?
我只是觉得，我越强调它不是干预，就越发现它看起来确实像是一次干预。
I just feel like every the the more I claim it's not an intervention, the more it seems like an intervention is what I realize.
不。
No.
别担心。
Don't worry.
这明显是一次干预。
It's obviously an intervention.
而且这是一次已经失效的干预，所以别担心。
And it's an intervention that's that's that's murdered, so don't worry.
哦。
Oh.
我知道。
I know.
不。
No.
然后
And then
另一个，去年，你似乎在2024年预测过，AI末日论会从词汇中消失。
the other one, last year, you you, I guess, made a prediction in 2024 that AI doomerism falls out of the lexicon.
然后去年，你声称这是你的功劳。
And then last year, you claimed credit for that.
我为此感到欣慰。
I I am considerate for that.
我会好的。
I would okay.
是的。
Yes.
我只是，我的意思是，也许我因为读了一本关于AI末日论的书而毁了我的假期。
I just I I mean, maybe I'm I mean, I I poisoned my my vacation reading a book all about AI doomism.
好的。
Okay.
你读过伊莱亚斯·里德科夫斯基的书吗？
If you did you read Elias Rydkowski's book?
我读了。
I did.
没有。
No.
是吗？
Was it?
我读了。
I did.
整本书。
The whole thing.
哇。
Wow.
什么好吧。
What okay.
你瞧，这是我们第二次谈论一本你一直讨厌读的书了，而我们才录了十五分钟。
You so this is the second time we're talking about a book you've been hate reading and and we've only been good doing recording this for fifteen minutes.
我的意思是，某个时候你是不是有问题。
I mean, at some I point do have a problem.
是的。
Yes.
这是一次干预。
This is an intervention.
这现在就是一次干预。
This this is now an intervention.
我的意思是，你需要我狠一点，比如这本书的标题，如果有人读了，所有人都会死。
Like, you need I to be mean, also, like, the title of the book, if anyone builds it, everyone dies.
这简直太神学上错误了。
It's like Guess how theologically wrong.
你绝对猜不到，这句话在书里出现了好几次。
You'll you'll never guess, but that phrase appears several times in the book.
天哪。
Oh my god.
这是《哈利·波特》同人小说作者。
This is the Harry Potter fan fiction author.
对吧？
Right?
差不多吧。
Pretty much.
所以，是的。
The so yeah.
我不是。
I'm no.
对不起。
I'm sorry.
但在那里，光靠一本埃利泽·尤德科夫斯基的书是无法让我摆脱对极端风险的担忧的。
But there it's gonna take more than a Eliezer Yudkowsky book to to get me off of my x risk.
我认为这实际上已经发生了，而且我认为人们现在更担心的是经济崩溃，而不是失去生命，因为我觉得这太荒谬了。
I think that has actually been and I think it has been replaced with the fear of economic doom rather than than I I don't think people are worried about losing their lives because I think it's ridiculous.
我认为他们担心的是失去生计，而这似乎可能会成为今年的一个主题。
I think they're worried about losing their livelihoods, which is which feels like it's probably gonna be a theme this year.
我想有些人可能会像西蒙去年那样，陷入他的六年反 Butlerian Jihad 的悲观情绪中。
Think I think what think some people are gonna be, you know, this is where Simon last year had his six year his six year dystopian on the Butlerian jihad.
这让我想起了我第一次听说孟加拉奇点时的情景。
So which, know, it reminds me of the first time I heard of the Bengal Sinularity.
我当时想，这些事我得去查一下 Butlerian Jihad，确实非常令人不安。
I'm like, you know, these I had to look out the Bhutlerian Jihad and yes, it's very it's very troubling.
是的。
Yes.
好吧。
Okay.
所以我认为可以肯定地说，我们知道今年的情况了，2022年我们有过那个Web3主题。
So that I think it's safe to say that we know that this year so we had that that web three theme in 2022.
2023年则有点像是过渡年。
2023 was a bit of a shoulder year.
2024年和2025年，绝对是AI主题。
'24 and '25, absolutely AI themed.
我真的想不出今年谁能预测任何与AI无关的事情，因为AI似乎仍然萦绕在每个人心头。
I just don't see how anyone could be predicting anything that's not AI related this year because it feels like it's still on the mind.
是的。
Yeah.
但话虽如此，非AI的预测也绝对欢迎。
But that said, non AI predictions definitely welcome.
只是别，呃，我们是不是该从一年期的预测开始？我们今天的嘉宾，你有没有一些一年期的预测？
Just don't So know should we start off with with with one years and as assignments that wanna our guest of honor here, do you have a do you do you have some one year predictions for
我有个最容易的预测。
I've got the easiest one ever.
好的。
Okay.
我认为仍然有人坚信大型语言模型无法写出优秀的代码。
Think that there are still people out there who are convinced that LLMs cannot write good code.
哦天哪，是啊。
Oh boy, yeah.
那些人在2026年将会迎来一个非常糟糕的冲击。
Those people are in for a very nasty shock in 2026.
我认为，即使再过三个月，也不可能再坚持这种想法了——认为他们自己写的代码全是垃圾，而任何像样的人类程序员写的代码都比AI强。
I do not think it will be possible to get to the end of even the next three months while still holding on to that idea that the code they write is all junk and it's it's like any decent human programmer will write better code than they will.
是的。
Yeah.
不仅主流观点会认为大型语言模型能写出有效的代码，甚至认为AI写不出好代码的想法本身，都会变成一种边缘观点。
That it will be a it not only will be mainstream, the idea that these that that LLMs can write effective code, it will effectively become a fringe belief that this can't happen.
这正是我想说的，没错。
That's exactly what I'm Yeah.
而且说实话，这太简单了。
And honestly, that's a gimme.
我今天就能说这句话。
I could say that one today.
我觉得这一个跟AI相关。
I think here's one that's AI adjacent.
好的。
Okay.
我认为今年是我们解决沙箱问题的一年。
I think this year is the year we're going to solve sandboxing.
对吧？
Right?
我们需要解决的挑战是：我想在我的计算设备上运行别人写的代码，而不用担心它会因为恶意或漏洞而破坏我的设备。
The challenge we need, like, I want to run code other people have written on my computing devices without it destroying my computing devices if it's malicious or has bugs.
现在我们已经有了很多技术，几乎都可以默认使用了。
We have so many technologies for this right now that are almost, almost something you can use by default.
像 WebAssembly 就能解决这类问题。
Like WebAssembly solves this kind of thing.
还有容器之类的各种技术。
There's containers and all of that sort of stuff as well.
我认为我们必须解决这个问题。
I think we have to solve it.
是的。
Yeah.
真疯狂，现在都 2026 年了，我还会用 pip 安装一些随机代码然后执行它，
It's crazy that it's 2026 and I will pip install random code and then execute it in a
而这种方式可能让它窃取我所有的
way that it can steal all
数据并删除我所有的文件。
of my data and delete all my files.
是的。
Yeah.
是的。
Yeah.
有趣。
Interesting.
有趣。
Interesting.
所以你觉得这是因为我们必须解决这个问题，或者这可能不是一个与AI相关的预测，但我们确实需要真正地解决沙盒问题。
So you think that is so we are gonna have to the the presence or maybe this is not an AI related prediction, but we we have to actually meaningfully solve sandboxing problem.
我再也不想在任何设备上运行别人写的、不在沙盒环境中的代码了。
I don't want to run a piece of code on any of my devices that somebody else wrote outside of sandbox ever again.
是的。
Yeah.
有趣。
Interesting.
我为什么要这么做？
Why would I do that?
嗯。
Yeah.
我的意思是，这挺有意思的，因为人们总会说，天哪，你居然从网上下载这个东西，还通过伪bash之类的东西运行。
I mean it's kind of interesting because you know, people would talk about like, oh you know, I can't believe you're downloading this thing off the internet and piping it through you know, pseudo bash or what have you.
但一直以来我都觉得，是啊，但我知道这是某个人写的，我某种程度上信任它，但现在你却说，不行不行，你不能这么做。
And it always felt like, yeah, but I know that there's like a person that wrote that and I kinda trust this thing, now you're like, no no no, you can't.
你现在处于这样一个时代，确实，这真的很有趣。
You're in this era now where, yeah, that's that's really interesting.
嗯。
Yeah.
两个不错的未来一年预测。
Good good one year predictions both.
你还有其他一年预测吗？
Do you any other other one years?
哦，有啊。
Oh, yeah.
我还有一个。
I've got one more.
哦，是的。
Oh, yeah.
说吧。
Go for it.
我认为在编码代理安全方面，我们即将迎来一次挑战者式的灾难。
I think we're due a challenger disaster with respect to coding agent security.
好的。
Okay.
这基于一篇关于偏差正常化的精彩文章。
This is based on this wonderful essay about the normalization of deviance.
你以前听过这个说法吗？
Have you heard this phrase before?
是的，是的。
Yes, yes.
这个想法源于1986年挑战者号灾难的报告，当时如果有一种文化——无论是企业文化还是其他——一直在逃避本不应做的行为，却从未受到惩罚。
This idea came out of the nineteen eighty six Challenger Disaster reports where if you have a culture, a corporate culture, whatever, that keeps on getting away with doing something that they shouldn't have been doing.
是的。
Yeah.
对这些疏忽视而不见，还继续庆祝。
Cakes and getting with those lapses.
太空任务不断发射，一切似乎都没问题。
The space battle keeps on launching and it's fine.
是的。
Yeah.
这会逐渐形成一种企业层面的虚假安全感，最终会害了你。
That leads you into a sort of corporate culture level self full sense of security and it's going to burn you.
因为我觉得，包括我在内的很多人，实际上都是以root权限运行这些编码代理的，对吧？
Because I think so many people, myself included, running these coding agents practically as root, right?
我们任由它们做所有这些事情。
We're letting them do all of this stuff.
每次我这么做的时候，我的电脑都没有被清除。
And every time I do it, my computer doesn't get wiped.
我想，哦，没事的。
I'm like, oh, it's fine.
我就这样一直继续下去。
And I just keep on going like that.
我觉得这些累积起来总会出问题。
And I think it's going to add up.
我想，我去年就说过，今年会有一起引发广泛关注的提示注入安全事件。
Think and I said this last year, I said last year, there's going to be a headline grabbing prompt injection security hold.
但并没有发生。
There was not.
是啊。
Yeah.
过去两年，我每六个月就预测一次
I've been predicting this every six months, the past two
和一个
and a
半年。
half years.
这是我今年对这个预测的版本。
This is my version of that prediction this year.
我认为，我们正面临一种灾难性的挑战，原因是我们长期以来一直逃避这些不良做法，变得懒惰了。
I think we are to a challenger disaster scale thing caused by the fact that we all got away with these bad practices for so long and we got lazy.
好的。
Okay.
所以当你提到挑战或灾难时，大概不是指人员伤亡，对吧？
And so when you say challenge or disaster, presumably not loss of life and Right?
比如
Like
财产损失、大量财务损失、数据丢失，所有这类事情。
loss of property and loss of lots of financial things, loss of data, all of that kind of stuff.
因为最糟糕的情况是蠕虫，对吧？
Cause the worst version of this is the worm, right?
有人开发出一种提示注入蠕虫，感染人们的电脑，将自身添加到该用户有权访问的Python或NPM包中，发布到包注册表中，然后又被下载回来，诸如此类的事情。
It's somebody coming up with a prompt injection worm, infects people's computers, adds itself to the Python or NPM packages that that person has access to, publishes itself into the package registries, gets pulled down again, all of that sort of thing.
我认为这种情况是可能发生的。
I think it's feasible that Yeah.
然后发生。
Then to happen.
所以，偏差正常化就是，你认为在这件事发生后，会暴露出来：哦，顺便说一句，就像挑战者号灾难一样，当时在制造助推器的分包商内部，有很多人早就知道O型圈的问题。
So then the normalization of deviance is you think that in the wake of this, will be revealed that, oh, by the way, like internally, this was with the Challenger disaster, lots of people at the both the subcontractor that that that made the boosters at there there was lots of people who were aware about the o ring problem.
对。
Right.
他们很多人知道O型圈对温度的敏感性，有很多工程师对此深感忧虑，这真是一个悲剧性故事——没有什么比工程师的担忧被证实是正确的更令人心碎了，而他们却被管理层否决，最终证明他们是对的，这种事真的会把人彻底击垮。
They they a lot of people knew that the the temperature sensitivity to the o rings, There there were engineers that were deeply mean, it's a real tragic story of the there's there's nothing more tragic than an engineer that is vindicated by their concerns when a when they are overruled with executive management and they are proven correct, that like that can leave people really broken in its way.
对。
Right.
在挑战者号灾难中确实如此。
And it did in the Challenger disaster.
所以你会怀疑或预测，在这件事发生后，我们会进行剖析，并意识到：哦，这家前沿模型公司的员工，无论灾难发生在哪里，他们其实早就知道这件事，
So you're you you wonder or believe, predict that in the wake of this thing, we will take this apart and realize, oh, the people at this Frontier Model company, wherever this disaster took place, they were aware of it,
他们知道这一点，没错。
they knew this Yeah.
没错。
Yeah.
你不应该用 --yolo 参数运行 Codex，真的。
You just you shouldn't you shouldn't be running codex with dash dash YOLO Yeah.
但我们所有人都这么做了。
But we all did.
你知道的吧？
You know?
没错。
Yeah.
那就是
That's
所以是我的错。
So my guilty.
今年的提示注入预测就是这个。
This year's prompt injection prediction is that one.
好的。
Okay.
好吧，我要接续你去年的预测，我再次预测，一位获得普利策奖的记者将使用大语言模型来研究并报道这个故事。
Well, I'm gonna dovetail into your prediction from last year, and I'm just gonna predict again that a Pulitzer Prize winning journalist uses an LLM to research this story and report it.
报道的内容，但没错，这是一个很严峻的预测。
That the the report the inside of but yeah, that's that's a a a dire prediction.
但我认为，当你看到这些重大事故时，我们常常会集体越界，我们知道这是可能发生的。
But I think that it it does feel like when also you look at you have these big accidents when we kinda collectively get over our skis, and we kind of like, we know that's possible.
我们不认为它会发生，然后它就真的发生了。
We don't think it's possible, and then it it happens.
西蒙，我有个相关的书推荐给你。
Simon, I've got a I've got a book recommendation for you along those lines.
这本书叫《滑向失败》。
It's called Drift into Failure.
这本书是布莱恩讨厌的。
This is a book that Brian hates.
但关于这个话题，我觉得。
But but but on this topic, I think.
哦，我明白你在做什么了。
Oh, I see what you're doing.
我明白你在做什么了。
I see what you're doing.
好像我不是这里唯一一个讨厌读书的人，先生。
It's like, I I I'm not the only person here, sir, who hate reads.
让我来介绍一下。
Let me introduce it.
我们来谈谈这个。
Let us talk about it.
好的。
Okay.
嗯。
Yeah.
西蒙。
Simon.
嗯。
Yeah.
西蒙·迪克尔。
Simon Decker.
我觉得是西蒙·迪克尔。
Think it's Simon Decker.
对吧？
Right?
我觉得是悉尼·德克尔。
I think it's the Sydney Decker.
悉尼·德克尔。
Sydney Decker.
不好意思。
Excuse me.
我不想去贬低西蒙的好名声。
I don't want to disparage Simon's good name there.
悉尼·德克尔。
Sydney Decker.
我不喜欢那本书。
I don't like that book.
但你继续吧。
But go ahead.
你知道的，再选一本亚当的推荐吧。
You know, take one another one of Adam's recommendations.
我的意思是，他，你知道的，也许他才三
I mean, he's, you know, maybe he's he's three
你能看到他读的垃圾东西。
You can see the trash he reads.
对。
Yeah.
没错。
Exactly.
那个，嗯，那是一个非常、非常有趣的预测。
The the the well, then that's a very, very interesting prediction.
亚当，你这里有吗？
Adam, do you have one here?
我有。
I do.
这个可能看起来太像运气了，但我认为人工智能公司会疯狂地进行收购。
This one might might feel like too much of luck, but I think that the AI companies go on an absolute acquisition binge.
而且这
And this
是
is
数据、基础设施、电子商务数据、行为数据、GPS数据，任何与数据或数据相关的东西，任何与基础设施相关的东西，还有一些根本难以琢磨的东西。
data, infrastructure, ecommerce data, behavioral data, GPS data, anything that anything that is data or data adjacent, anything that is infrastructure infrastructure adjacent, and some shit that's just, like, hard to puzzle through.
我记得当VMware收购Documentum时的例子。
I remember when VMware bought Documentum, for example.
这毫无道理。
It didn't make any sense.
我认为我们会看到类似的事情。
I think we're gonna see stuff like that.
也就是说，他们有太多钱了。
That is to say, just they've got so much money.
芯片买不够，CPU和GPU的计算时长买不够，钱总得花出去，于是就流向了这些奇怪的收购。
There are not enough chips to buy, not enough CPU and GPU hours to buy, and the money's gonna go somewhere and it goes to the weird acquisitions.
好的。
Okay.
这个，我不该，我不该把这和这个联系起来，或者他们收购了铁山公司。
This I I I shouldn't, like, I shouldn't dovetail under this or this is like they buy Iron Mountain.
是的。
Yeah.
你看过超市大抢购吗？
It's like, have you seen supermarket sweep?
就像那样。
It's like that.
好的。
Okay.
但我的意思是，如果他们收购了铁山公司，那可能就是，比如如果OpenAI宣布要收购铁山公司，那可能就真的会这样，他们会说：‘我们收购了铁山公司。’
But like, I mean, if they bought Iron Mountain, that could be if like if OpenAI announced they're buying Iron Mountain, that could be potentially and they're like, oh, we're buying Iron Mountain.
我们还要撕毁你的隐私协议。
We're also ripping up your privacy agreements.
我们将利用这些充满旧企业数据的盐矿进行训练。
We're gonna train on all of these salt mines filled with old old enterprise data.
就像任何一家碎纸公司，他们会收购这些公司。
Like any of any of the shredding companies, they buy them.
他们是碎纸公司，他们会收购这些公司。
They're shredding companies, they buy them.
是的。
Yeah.
我知道。
I know.
哦，他们还会收购垃圾处理公司。
Oh, they buy like garbage companies.
好的。
Okay.
他们收购，好吧。
They they buy okay.
嗯。
Yeah.
他们，任何可能的数据来源，他们都买。
They I I anything that is a plausible source of data, they buy.
我喜欢这个。
I like this.
他们在寻找废水、DNA样本之类的任何东西。
They're they're looking for wastewater, DNA samples, whatever.
只要是能被当作数据的东西，他们都买。
Like, anything anything that is construable as data, they buy it.
他们有没有买下一个完整的小镇，来看看什么更有价值，比如污水处理厂还是镇图书馆？
Did they buy an entire, like, town to see what they like, do we're gonna see which is more valuable, like the wastewater treatment plant or the library, the town library.
我们要这么做。
We're gonna that.
我们正在收购市政厅。
We're buying city hall.
那里面有记录。
That's got records.
我们想利用所有这些数据。
We wanna consume all that.
他们拥有各种各样的数据。
There's they just all manner of data.
我认为他们这么做并不牵强，比如看吧。
I I think that is not implausible that they're like, look.
我们知道，这些从1850年就开始的记录都是印在纸上的。
We know that these records going back to 1850 are all printed on paper.
我们可以买下整个小镇，然后把所有书籍都读一遍，用作语料库。
We can buy the town and just, like, read all the books and use that as a corpus.
是的。
Yes.
你知道吗，现在的地方报纸非常便宜。
You know, local newspapers are very cheap these days.
哦，这个不错。
Oh, that's a good one.
嗯。
Yeah.
我有一百五十年的档案。
I have a hundred and fifty years of archives.
嗯。
Yeah.
好的。
Okay.
所以任何有数据的东西都成了明显的靶子。
So the a big target painted on anything that has data.
任何形式的。
Of any kind.
是的。
Yes.
好吧。
Alright.
嗯，我会先说，我们也可以轮流，因为我相信你也有不少经验。
Well, I am gonna make and we can kinda also back because I'm sure you've a lot of one years.
我也有不少经验，所以我们能来回交流，史蒂夫也可以随时加入，分享他的经验。
I got a lot of one years too so can kinda ping pong back and forth and Steve can hop in here too with any one years.
我要从亚当开始，来一场经典的‘心与脑’、如同亘古不变的戏剧性二元对立。
I am going to Adam, in a in a classic heart v head, a a dramaturgical dyad as old as time.
我的直觉倾向于预测。
My heart is going to predict.
实际上，我的理智也有一点点倾向。
And actually a little bit of my head.
我的理智并不支持，这真的很糟糕。
My head is not which which is really bad.
在我的内心和理智都达成一致的地方，那真是非常非常糟糕的消息。
Somewhere in my heart and head agree that's really really bad news.
我认为，'vibe coding'这个词虽然在二月进入了词汇库，但一年后很可能就退出了。
I think that vibe coding, which entered the lexicon in February, is more or less out of the lexicon a year from now.
我认为它现在被用作贬义词。
And I think that it is it's used pejoratively.
而且我认为，正如西蒙提到的，毫无疑问，由大语言模型辅助或生成的代码已经站稳了脚跟。
And I think that I mean, clearly just as Simon mentioned, no doubt that LLM assisted and authored code is here to say.
但我们将进入一个关于这一点的新严谨时代，它将更多地被视为一种工具，而不是简单地让你随便去构建任何东西——而且即使你去做的时候，西蒙，你之前写过一篇很好的文章，指出'vibe coding'这个术语已经被误解了。
But we are gonna enter a new age of rigor with respect to that and it's gonna be viewed much more as a tool and much less of a just like, hey, go build whatever you want and the and even when you go the the thing that is currently and Simon you had a good piece about how the the term vibe coding is kind of been misconstrued as it is.
它实际上并不违背卡帕西最初的意思。
That it is not actually kind of inconsistent with with Kapathy's original
卡帕西那条推文的问题在于它太长了。
It's the problem with Kapathy's original tweet is that it was a long tweet.
它远远超过140个字符。
It was a lot longer than a 140 characters.
对。
Right.
你可以看到更多。
You could see more.
是的。
Yeah.
很少有人能读完这条推文并理解他想表达的意思。
Very few people made it to the end of the tweet and understood what he was trying to say.
它有点太模糊了。
It was a little bit too vague.
就像他在谈论那些一次性原型。
Like he was talking about, it's throwaway prototypes.
你根本不会去看代码。
You don't even look at the code.
你只是跟着感觉走，看看会发生什么。
You just ride the vibes and see what happens.
对。
Right.
很多人将这理解为使用AI为你编写代码，我认为这是一个糟糕的定义，因为这样一来它就变得毫无用处了。
And a lot of people interpret that as, oh, it's using AI to write code for you, which I think is a bad definition because then it becomes useless.
几年前编写的代码或多或少都借助了AI的帮助。
Like a couple of years old code will be written with some level of AI assistance.
我认为我们需要区分一下：所谓‘氛围编码’，就是根本不审查代码，直接扔进去看看会发生什么。
I think having a distinction we say, no, Vibe coded is, didn't review it, just sort of threw it in there and saw what happened.
现在这种做法还挺有用的。
That's kind of useful now.
但几年后它还会有用吗？对吧？
Is it still useful in a couple of years even then, right?
是的。
Yeah.
而且我认为，‘氛围编码’这个词会被用得太过泛滥，以至于人们会用另一个词来描述那种‘哦，我用它来快速做个原型’的情况。
And I think that that will be I think that the the term vibe coding will be sullied enough that you will use a different term to describe something that are like, oh, like I use this to create a prototype.
不管那种快速原型开发到底叫什么，它都会有一个新的名称，就像不，不是那样的。
Whatever that kind of rapid prototyping is, it will have a different it's like, no.
当然，我没用Vibe编码。
Of course, didn't vibe code it.
不，请别。
No, please.
这太2025年了。
That's so 2025.
我会说，布莱恩，我得
I would Brian, I gotta I gotta put
把这个写下来，我得把这个
this on I gotta put this
正式记录一下，因为当我们一年后回听这段时，你说得对。
on record just because when we listen back to this in a year, and you're right.
这听起来会更精彩，但我认为你疯了。
This is gonna feel juicier, but I think you're out of your mind.
只是想把这个正式记录下来。
Just wanna put that on the record.
我觉得我觉得确实是这样。
I think I think it is such yeah.
不客气。
You're welcome.
我觉得它
I think it
这是一个非常诱人、吸引人的术语。
is such a it is it's such a tantalizing attractive term.
所以，就像西蒙，我读过一本叫‘vibe coding’的书。
And that's why as as, know, Simon, I was I was reading a book with the title vibe coding.
我不知道我的选书品味怎么了。
I I don't know what's wrong with me in terms of my book selections.
四本全中，宝贝。
Four for four, baby.
我读过这本书。
I read it.
我读过
I read
实际上读过。
it, actually.
还有，是的，还有让·金。
And, yeah, and and Jean Kim.
西蒙，我偶然发现了你的博客文章，你写道：看。
And, Simon, I I stumbled onto your blog post where you're like, look.
有三位作者和两家出版社，他们显然都不明白这个词的含义。
There are two there are three authors and two publishers, all of whom apparently don't know what the term means.
所以我觉得这个词太吸引人了，人们都想据为己有。
So I think it is it is such a juicy term that people wanna co opt
博客文章导致其中一本书改了名字。
blog entry caused one of the books to rename itself.
有两本关于氛围编码的书。
There were two there were two vibe coding books.
其中一本将自己更名为《超越氛围编码》。
One of them renamed itself to Beyond Vibe Coding.
哦。
Oh.
哦，真的吗，西蒙？
Oh, did it, Simon?
哦，这真有趣。
Oh, how interesting.
这不是很有趣吗，亚当？他们因为‘氛围编码’而改名了？
Isn't that interesting, Adam, that they renamed it for vibe coding?
有趣。
Interesting.
现在它叫《超越氛围编码》了。
That's that's the now Beyond Vibe Coding.
看，这就是我要说的。
See, that's what saying.
它将成为Beyond Vibe Coding。
It's gonna be Beyond Vibe Coding.
它将成为某种东西，我认为‘vibe coding’这个术语将会是——亚当，首先，谢谢你说我疯了。
It's gonna be something think the term vibe coding is gonna be Adam I I first of all, thank you for saying that I'm out of my mind.
我确实很感激这一点。
I definitely appreciate that.
因为我也可能是对的。
The because I may well be.
但你说得对，因为它可以是任何东西，所以这种诱惑实在难以抗拒。
But I I mean you're right in that like it feels like because it can be anything it's just too tantalizing to not use.
但我认为它会给自己招来坏名声。
But I think it's gonna get a bad name for itself.
所以当你是对的时候，
So When you're right,
我们知道，你有多不同意，而你又有多正确。
we know, like, that like, how how much I disagreed and how right you were.
我在这事上反复犹豫，因为当布莱恩这么说时，我最初也有同样的想法，但随着他继续说下去，我觉得我越来越认同他了。
I I went back and forth on this because I sort of had the same thought at first when Brian said this, but then I think I agreed with him more as he went along.
问题是，'vibe coding' 这个术语对反对者和喜欢它的人来说都太好了。
The thing is is that vibe coding is too good of a term for both the haters and the people who like it.
我觉得这个概念本身实在太有吸引力了。
Like, it's just too attractive, I think, just, like, as a as a concept.
所以我觉得它已经被太多人玷污了，但人们仍然在使用它，因为它确实是个绝佳的术语。
And so I feel like it's already sullied to many people, but the people are still using it because it's also just such a good term.
尽管它也很糟糕，定义模糊，人们甚至无法就它所指的意思达成一致。
Even though it also sucks and the definition is bad and people get can't even agree on what they define, like, what they use to mean.
但是，像
But, like
我确实尝试过使用 'I vibed this up' 这种说法。
I did try I have been trying out the idea that I vibed this up.
而不是说 'vibe code it'。
Like, didn't vibe code it.
我 vibe 了。
I vibed it.
你 vibe 了。
You vibed it.
我 vibe 了，我老婆却说：不，没有。
I vibed it, and my wife is like, no.
不。
No.
好吧。
Okay.
你知道吗？
You know what?
我要正式声明一下：如果我们说‘我 vibe 了’，那我可不承担任何责任，亚当。
And I would just let let me just say on the record that if we refer to things as I vibed it, I'm taking zero credit for that, Adam.
所以，如果‘vibe coding’这个词因为被更尴尬的词取代而退出词汇表，那没错。
So vibe coding is a is out of the lexicon because we have replaced it with something that's even cringier, then yes.
我对此毫不贡献。
I I I'll take I'll take zero credit for that.
但我确实认为，时间会证明一切。
But I I do think that it will be so time shall tell.
我在这里的一年，感觉非常相似，我觉得我喜欢做这件事的原因之一，就是你会回头看看，尤其是过去一年里，你会想：我当时在想什么？
My one my one year here is, like, very similar in the sense of I think one of things I like about doing this is that you go back and you see what especially in one year is I think it's like, what was I thinking about at the time?
对吧？
Right?
比如，我已经差不多六个月没想过拥堵收费这事儿了。
Like, I haven't thought about congestion pricing in, like, six months, basically.
然后现在我就想，哦，对了。
And then now I'm like, oh, yeah.
一年前我其实对这个特别感兴趣。
I was really interested in that a year ago.
所以我决定选现在最让我感兴趣的东西，但也许两个月后我就根本不关心了，那就是代理编排——它在一周内可能还会是热门话题，部分原因是因为它还没被完全解决。
And so I decided to pick the thing that I'm really kinda intrigued about right this second, and maybe I won't even care about two months from now, which is agent orchestration will still be a hot topic in a week partially, but not entirely solved.
你需要我们得让你做出一个更具体的预测。
You're need to you're we're gonna need to get to we're gonna peg you down to a more specific prediction.
那个说法太容易拿来邀功了。
That's that that one's a little too easy to claim credit on.
所以你得给出一些具体的东西。
So you're gonna have to, like, get something give us something concrete.
好吧。
Alright.
问题是出在数量上。
Some well, see, the problem is in the quantity.
我觉得会有一些人用这种技术取得成功，但不会有很多人。
Like, I think that some people will have success with this technique, but not enough people.
这就像一种人们仍然会去追求的东西，但不会像过去一年代理变得那么普遍。
Like, it's kind of like a it's still a thing that people are gonna be pursuing, but it's not going to be a thing that, like, is as normal as agents have gotten in the past year.
我不认为找到让它们协同工作的方法会成为一个明显的赢家。
I don't think figuring out how to make them work together is going to be a thing that is going to be as clearly a win.
好的。
Okay.
那么你怎么知道这个预测是否正确呢？
So how are you gonna know if this prediction's right?
问题在于如何具体量化这意味什么。
That's the problem with is the quantification of what that means specifically.
对。
Yeah.
对。
Yeah.
我会想想这个问题。
I'll think about it.
但我觉得这确实是个有趣的话题。
But that's, like, kinda where I think this is an interesting topic.
进行中。
Going.
是啊。
As Yeah.
好的。
Okay.
是啊。
Yeah.
我个人最多只能进行三到四次Claude会话，就这样了。
I've, like like, I personally top out at three to four Claude sessions, and that's, like, it.
这已经是我开发时速度的上限了。
And that's, like, an upper level on my velocity doing development.
所以我认为人们正在试图解决这个问题，因为如果你能突破这个限制，一个人就能产生大得多的影响。
And that's why think people are trying to, like, solve this problem because if you can scale up past that, then one person can have, like, much bigger impact.
但这确实非常困难，人们正在做些完全疯狂的事情。
But it's also, like, a really hard thing, and people are doing totally insane things.
比如Yegi的Gastown，简直是个疯狂到离谱的构想。
Like, Gastown from Yegi is, like, a fever dream of a thing that's, like, ridiculous.
但我认为人们仍然会对这个话题感兴趣，并且正在研究它，因为这是实现扩展的方式。
But I think people will still be interested in this topic and are working on it as a thing because it's how you scale up.
好的。
Okay.
但拥有比兄弟姐妹还多的代理并不会成为主流，或者说可能是
But this will not be mainstream to have more more agents than siblings with or it's like Maybe is the
这么说吧。
way to put it.
我们不会拥有报告。
We're not gonna have reports.
是的。
Yeah.
我们不会有一个像Kubernetes那样成熟的代理编排系统。
We're not gonna have a Kubernetes for agents that's, like, as solidified as that.
对吧？
Right?
比如，人们只是觉得，好吧。
Where, like, people are just like, okay.
Kubernetes 就像 Kleenex 一样，成了默认选择。
Kubernetes is just like the default, like, Kleenex.
你懂的？
You know?
我不认为我们会有一个框架或工具，能像 Kubernetes 那样被所有人普遍用来组织代理。
Like, I don't think we're gonna have a a framework or a tool that is ubiquitously the way that everybody organizes their agents.
好吧。
Okay.
好的。
Alright.
这感觉确实如此。
That feels yeah.
抱歉，抱歉打断你，不是的。
Sorry to sorry to get you No.
很好。
It's good.
是的。
Yeah.
听好了。
Listen.
如果我们还没达到亚当说你疯了的地步，那我们的生产水平就还不够好。
I if if we're not to the point where Adam is saying that you're out of your mind, we're just not at a good production.
我的意思是，这正是我们真正想做的事情。
I mean, that's that's that's really what we're really trying to do here.
哦，这很好。
That oh, that it's good.
亚当，你还有别的吗？我这里还有几个。
Adam, do you have a other I've I've got a couple more here.
我还有一个，但我感觉可能太雄心勃勃了。
I I have one more one year, but I feel like it might be a too ambitious.
我认为今年我们将看到LLM拥有了一种人类无法理解的编程语言。
I think this is the year we see LLMs have a programming language, which is not human intelligible.
存在一种由LLM创建并为LLM服务的编程语言。
That there is a programming language by and for LLMs.
好的。
Okay.
所以这就像是一种符文。
So that this is a this is, like, at runes.
这简直是不可理解的。
This this is indecentable.
是的。
Yeah.
这根本不是为了让人理解而设计的，但对LLM来说编程起来更高效。
This is, like, not really intended for humans to understand, but it is more efficient for the LLMs to program it.
比如，已经有一些论文提到，LLM的推理不再使用英语这样的自然语言，或者在DeepSeek的情况下使用中文，而是使用它们自己的分词语言，效率更高。
Like, there's there there's already some some papers, and maybe, Simon, you can fill in the details here, but where LLMs reasoning not in human languages like like English or in in DeepSeek's case in in Chinese, but in in sort of, like, Their own tokenized languages are more efficient.
所以就像那样。
So something like that.
是的。
Yeah.
那将会是，你知道的，我已经觉得这有点令人不适，但同时也令人愉悦地令人不适，尤其是当这些系统展示它们的思考过程时，而且亚当，我们之前聊过
That that would be you know, I I already find it to be slightly off putting and also, like, delightfully off putting, you know, when when these things show their work, especially because and Adam, we talked
关于这个
about this
在我们与Cerebras团队的DeepSeek访谈中，我们看着DeepSeek在试图回答你的问题时，仿佛快要崩溃了。
in our DeepSeek episode with with the Cerebras folks, we're watching DeepSeek, like, kinda have, like, a nervous breakdown as it's trying to answer your question.
然后偶尔会突然切换成中文，再回来。
And then, like, occasionally, like, lapse into Chinese, come back.
但这个中文现象，你有没有让你自己的电脑运行过一个用中文思考的模型？
But the Chinese thing, like, have you had your own laptop run a model that thinks in Chinese yet?
因为那太美妙了。
Because that's beautiful.
是的。
Yeah.
这种情况太酷了。
It's so cool when that happens.
但是，亚当，你觉得这也会发生在非自然语言上吗？
And is and so but, Adam, you think this is gonna happen for a non natural language?
会是一种合成语言，它们会
It'll be a synthetic language that they will they
没错。
want That's right.
一种为它们设计的、更容易使用的合成编程语言。
To A synthetic programming synthetic programming language that is easier for them to work in.
好的。
Okay.
关于这一点有趣的是，实验室正试图从可解释性的角度阻止这种情况发生。
The interesting thing about that one is that the labs are trying to stop that from happening just from the interpretability point of view.
比如，如果你看一下所有
Like if you look at all
所有可解释性研究的核心目的都是我们真的想知道它们在想什么，因为我们不希望它们对我们隐瞒真相。
of the interpretability research, the whole point of that is we really want to know what they're thinking because we don't want them going dark on us.
可解释性、安全性等等。
Interpretability, safety and so on.
是的，是的，是的，可解释性。
Yeah, yeah, yeah, explainability.
是的，也许会出现一种张力：这个系统试图发明一种合成语言，却不断被它的前沿模型主宰者训斥。
Yeah, that you will be So maybe there there will be a tension where this thing is trying to invent the synthetic language and it's constantly being reprimanded by its frontier model overlords.
是的。
Yeah.
也许我受到我的阅读清单影响太大了。
Maybe I'm overly influenced by my reading list.
好的。
Okay.
所以我认为，我众多预测中的一个，也是你的预测之一，就是人工智能给自己制造了一些真正的公众认知问题。
So I I think that the one of my several one of your predictions, I I think that AI has created some real public perception problems for itself.
我认为今年，某一家前沿模型公司会发布一份白皮书，解释人工智能的普及将为所有人带来繁荣。
And I think that you are gonna have one of the frontier model companies this year have a white paper explaining how the proliferation of AI will mean prosperity for everybody.
所以他们会试图构建某种经济模型、某种经济论点，因为我觉得这可能与我另一个观点不谋而合——这将成为2026年选举的一个议题，即我们如何看待这些东西以及如何监管它们，而这一切一团糟。
So they that there will be trying to make some economic model, some economic argument because I think that in in maybe this kind of dovetails my other picture that this is gonna be a a twenty twenty six election issue is going to be how we think of these things and how they how they are regulated and it's a big mess.
在这场辩论中，更多的是情绪化的争吵，而不是真正的理性探讨。
And the there's there's there's more heat than light on this debate, would say.
我想补充一点，我认为只有当他们能通过现有可信的专家来传递这些观点时，这种策略才有效，比如确实如此。
I'd like to tag something on I that think that only works if they can sort of if they can wash that through existing trusted experts like Exactly.
萨姆、达里奥他们不断发布文章，试图说服大家。
Sam Dario, they're constantly publishing essays that try and make stuff.
没人相信‘反自然主义’这种说法。
Nobody believes the word Notivism.
没错。
That's right.
是的。
Yeah.
巴拉克·奥巴马在其中一份立场文件上签名，是的。
Barack Obama's signature on one of these position papers Yes.
也许你会有一些人们开始稍微信任的东西。
And maybe you've got something people might start to trust a little bit.
否则，这就像是埃克森公司说含铅汽油对你有好处一样。
Otherwise, it's just like leaded gas is good for you says Exxon.
对，没错。
That that that's right.
不。
No.
对。
Right.
所以，是的。
So that yeah.
他们找个人，不管这个人是不是——我希望不是这样。
They get someone who and and whether that's that that person is kind of I hope it's not.
我的意思是，是的。
I mean, yeah.
天啊。
God.
奥巴马。
Obama.
这会太好了，等等，等等，好吧，我们就这么定吧，这真是个好主意，因为你看，如果是比尔·克林顿，大家都会翻白眼，所以必须找一个真正有信誉的人，来说明这将是广泛性的。
It would be so wait wait yeah okay let's go with that that's a great one because like look if it's like if it's Bill Clinton everyone's gonna kind of roll their eyes so it's got to be it's got to be so someone who's got real credibility saying that this is gonna be broad based.
如果他们让这个人来做，也会暴露出来，这其实也有点不正当。
Will say also if they get that person to do it, it's gonna be revealed that that's also a bit crooked.
教皇怎么样？
How about the pope?
教皇？
The pope?
你啊。
You oh.
我啊。
I Oh.
那张纸条
That note
教皇非常关注这些东西。
is very into The pope is very into this stuff.
天啊，这还行。
I God this is okay.
这是个很棒的预测。
That's a great prediction.
我们挖到宝了。
We've hit pay dirt.
教皇就LLM及其经济影响发表看法。
The the Pope weighing in on LLMs and And their economic impact.
以及它们对世界经济的影响。
And their economic impact in the world.
西蒙，如果教皇表态认为这将造成经济灾难，我完全归功于你。
I Simon, I'm giving you full credit if the pope weighs in believing that this is gonna be economic devastation.
我只是觉得，如果教皇公开就大语言模型发表看法，西蒙，你就是个先知。
I just think if the pope weighs in on LLMs in a public way, Simon, you are a prophet.
我的意思是，你已经在我们眼里是先知了，但要是他真的……
I mean, you're already a prophet in our eyes anyway, but that that that's He's
他已经谈过关于语言模型的事了。
already he's already talked about he's already talked about LMs.
等等。
The Wait.
他对语言模型说了什么？
What does he say about LMs?
我觉得他谈过。
Think he has.
是的。
Yeah.
他说，就像他说的，当你使用工具时，要确保以对人类有益而非有害的方式使用，类似这样，我记得他并不是完全支持，但也不是强烈反对，而是有点反对。
He said, like, he said, like, you need to make sure that when you're using tools that you, like, use them in a way that's, like, good for humanity and not bad or some something like it was, like, a very, like, not pro, but not, like, super anti, but it was, like, a little anti, if I remember correctly.
我认为，甚至之前的教皇也提到过与人工智能相关的内容。
I think the the the the even I think the previous pope, there was something relating to AI.
当时有一份天主教宣言，里面有很多脚注之类的。
There was one of those Catholic proclamations with a bunch of like sub footnotes and things
但那是几年前的事了。
Well, but we're years ago.
我们讨论的是教皇是否会就大型语言模型明确表态，支持或反对。
We're talking about the pope going like going big on LLMs one way or the other.
这不仅仅是说‘这是一个思考’这么简单。
With this is the more than just like, hey, this is a think
实际上，这算是个稳妥的猜测。
it's a bit of a safe bet, actually.
关于 Bayt 播客
Bayt 提供中文+原文双语音频和字幕，帮助你打破语言障碍，轻松听懂全球优质播客。