铁路：原生云代理——杰克·库珀

对吧？

Right?

随着时间推移，这些东西必须变得高效得多。

Over time, those things are going have to get a lot and lot more efficient.

自己搭建硬件，能省下不少成本，让这些体验更扎实，这道理你懂的吧？

You can get a lot of almost like back of the napkin balance sheet margin, whatever you want to call it to kind of make those experiences, like, solid by building your own metal, right?

所以，回到之前说的，我们每次都在努力往更深的地方挖，就是为了优化这个体验。

And so, kind of to the earlier point of, like, we've always tried to go a little bit deeper every time to make that experience.

这一切都是为了尽可能让更多人享受到这种独特的体验。

It's all in the service of offering that differentiated experience to as many people as, like, humanly possible,

你知道的。

you know?

是的。

Yeah.

你在新加坡有个数据中心。

You have a data center in Singapore.

是的。

Yeah.

现在每个其他地区我们都有两个，包括新加坡。

So we have two in every other region now, Singapore.

我们将在第三季度新增第二个。

We're adding a second one in Q3.

没错。

So, yep.

那感觉是怎样的呢？

So, like, what's it like?

我是从来没建过数据中心。

I mean, I've never built a data center.

是的。

Yeah.

那我们就得去其中一个了

Then we'll have to, like, It's go to one or

去找Equinix，说：嘿，我要一些资源。

go to, Equinox and say, hey, I want some Yeah.

是的。

Yeah.

没错，就是这样。

So so yeah.

我的意思是，我可以找Equinix帮忙吗？

Mean, I can I can bring Equinix?

Equinix。

Equinix.

是的。

Yeah.

是的。

Yeah.

Equinix。

Equinox.

我的意思是，

I mean,

你也应该为你的身体选择Equinox。

you should also Equinox for your body.

为你的软件选择Equinix。

Equinix for your software.

我的意思是，你完全可以把数据中心放进蒸汽房里，让它热起来，随便怎样都行。

I mean, you can put a you can put a data center in the steam room and get nice and hot or whatever.

不过，是的。

But, yeah.

对。

Yeah.

你 basically 就是过去说：嘿，听着。

You basically just go and you say, hey, listen.

我要电力，还要一个机柜。

I want power, and I want a cage.

他们就说，太好了。

And they're like, great.

给你。

Here.

这就是你要的东西。

This is what it's gonna be.

然后你租下这个机柜一段时间，再往里面放机架和服务器，最后连上网络。

And then you rent the cage for a period of time, and then you have to fill the cage with racks, servers, and then hook up Internet to it.

对吧？

Right?

实际上就这些了。

That's realistically all

接下来你就自己处理其他所有事情。

the And then you handle everything else.

对吧？

Right?

是的

Yeah.

剩下的都由你来处理。

You just handle everything else.

对吧？

Right?

那从成本上来说，和云服务相比呢？

So And, like, what's the math versus, obviously, the clouds?

对

Yeah.

它们

They're

适合你

good for

。

you.

如果我们用云服务，转向金属设备后的投资回报期大约是三个月。

Our our payback period when we go to to Metal, if we rent it in the cloud, our payback period is about three months.

这数字简直爆了。

It goes crazy.

太离谱了。

It's nuts.

没错。

Yeah.

这相当于四年折旧下来的硬件价值。

And that's like four years worth of like depreciated hardware.

对吧？

Right?

我觉得接下来你会看到大量这种计算资源紧张的情况，因为各大云服务商都在大量采购设备。

And so I think it's like, you're gonna see a lot of this almost like compute crunch, so to speak, because a lot the hyperscalers are buying up a lot of stuff.

我们直接跟原始设备制造商、经销商，还有像超微、戴尔这些厂商合作，确保这些设备能顺利运行。

Like, we're working directly with OEMs and, like, resellers and, like, directly with people who are, like, building these machines, like Super Micro, Dell, all of those other things to go in and get these things working.

不过，上游那边还有很多供应链方面的问题。

But, you know, upstream, there's, like, a bunch of supply stuff.

说来有趣，我们上一轮融资时，资金还没完全花在服务器上，甚至到现在，我们融到的钱还比不上银行里的现金加上服务器的估值，因为服务器的价值其实涨了，主要是内存价格普遍上涨。

You know, we it it was funny because when we raised our last round in between basically deploying the capital for the servers and actually, I think even even now, the amount of money that we've raised is less than the amount of money that we have in the bank plus what the value of the servers are because the servers have actually appreciated in value because RAM has gone up in general.

对吧？

Right?

所以硬件这些东西现在值钱得有点离谱。

So it's kind of nuts just in terms of like how valuable hardware and all of this stuff is.

对吧？

Right?

如果你看看那些超大规模云服务商，今年光资本支出就投了八百亿美元，明年只会更多。

And if you look at especially a lot of that, like hyperscalers, like what they deployed, like $80,000,000,000 of, like, capital expenditures, like, this year and and, like, into next, it's gonna be, like, more in in general.

对吧？

Right?

整个基础设施的建设规模都大得惊人。

There's a massive, massive scale, like infrastructure build outs.

你一看就会觉得，这太疯狂了。

And you can look at that by like, wow, that's crazy.

他们花的钱，比曼哈顿计划还多得多。

They're spending, like, way more than the Manhattan Project.

但话说回来，如果每个人都要同时运行几十个、几百个代理，那花更多钱也是应该的。

But, like, again, if you go back to every person is going to run dozens, hundreds, whatever of agents in parallel, You should spend more than that.

你根本

You have no

想象不到要多少算力，才能让这个体验跑起来，哪怕你已经极致高效，哪怕你共享资源，哪怕你每一步都做对了——这还不包括推理部分。

conceptual idea of how much compute is required to go in and make that experience happen, even if you're deeply efficient, even if you're sharing resources, even if you're doing all of these things correctly, and that doesn't even count inference.

你怎么

How do you

规划基础设施的扩张？

plan on the build out?

我的意思是，增长曲线这么陡，你一上线，服务器利用率就能立刻达到100%吗？

Like, I mean, the growth chart is so vertical that, you know, like, are you usually a 100% utilization rate as soon as you're live with these tracks?

或者你们提前规划了多远？

Or like, how far ahead are

你们？

you?

对。

Yeah.

所以我们仍然保留云服务，主要用来应对突发需求。

So, so like, we still maintain like cloud presence for like bursting essentially.

我们跟AWS、GCP，还有其他几家云服务商都有合作，需要的时候随时可以租用。

And so what we can do is, you know, we work with AWS and GCP and a few of those, you know, other clouds, like we can just rent.

一旦我们自己的机房有了空间、电力之类的资源，就可以把那些负载慢慢迁回本地，对吧？

And then the moment we kind of get space or power or whatever, you almost just like compact those off the cloud, right?

因为我们当初是从云上起步的，后来才搭建了系统，方便把服务迁移到自己的硬件上。

Like we, because we started on the clouds and then we built a system to allow us to migrate to our own metal.

所以并没有什么限制说不能反复这么做，而这恰恰就是我们目前的做法。

And so there's nothing that says you can't just continually do that again, which is exactly what we do right now.

对吧？

Right?

所以我们从来不希望陷入计算资源受限的境地。

And so we never want to be in a spot where essentially we are, you know, compute constrained.

对吧？

Right?

年初的时候，我们确实遇到了计算资源紧张的情况，因为当时依赖的那家上游供应商没法按我们需要的速度提供配额。

And at the start of the year, like we actually got to a point where we were compute constrained because the one upstream provider that we were actually working with wasn't able to give us quota at the rate that we needed to.

而且硬件性能也偏慢。

And the hardware was like slower.

对吧？

Right?

所以我们不得不做了不少调整。

And so we had to do a bunch of different stuff.

我花了一个周末，重新搭建了整个网络覆盖系统，以便能同时连接五个不同的云平台。

I spent a weekend rebuilding our entire network overlay essentially so that we could straddle five different clouds.

对。

Right.

没错。

Yeah.

Oracle、AWS、我们自己、GCP，还有一家，对吧？

Oracle, AWS, ourselves, GCP, and like one other one, right?

现在我们能做得更多了，对吧？

And we can do more than that now, right?

但那时候我们真的被迫把实例挤得特别紧，因为根本拿不到足够的计算资源。

But we got into a spot where like, we were just trying to like pack instances tight because we couldn't get the amount of compute that we needed.

对吧？

Right?

这特别糟糕，结果导致我们出现了一些可靠性问题，虽然现在这些问题都过去了，但根源就是资源紧张。我之前发过一条推文，本来是想指出这个问题，结果不小心牵连到了一些无辜的人，惹上了麻烦。

And it was really unfortunate because as a result, like some of we had a few like reliability kind of things, which are now kind of past us, but it was all a result of this kind of like, there's a tweet that I made where like, you know, I've I like gotten in trouble because I was trying to point it out, but I accidentally caught this super base folks in the crossfire.

那条推文是说，要按现在模型的需求来获取计算资源，简直难上加难，而且只会越来越难。

But like the tweet was about there's it's really, really difficult, and it's gonna become more and more difficult to acquire compute at the rate that these models need to acquire compute.

对吧？

Right?

结果我们自己也遭了报应，这其实挺公平的，毕竟我当初那么直白地指出来。

And we got bit by it, which is, you know, fair and reasonable in the karma scheme of me, you know, trying to point it out.

所以，是的。

So, yeah.

考虑到你可能无法随时拥有自己的硬件，你怎么看待定价问题？

How do you think about pricing, knowing that you might not have your own metal available at all time?

你是假设，如果不得不上云，就得给自己留出更高的利润率吗？

Like, are you pricing assuming that you'll need to like pay yourself extra margins if you had to end up going in the cloud?

因为我们建了自己的数据中心，硬件业务的利润率其实挺高的，大概有70%。

Because we've built out our metal data centers, like our margins on metal are like quite high for the like 70%.

所以，如果我们想以合理的速度扩张，完全可以拿硬件业务的利润来补贴云业务。

And so we can actually deeply subsidize the cloud business if we want to scale at a reasonable rate.

因此，从运营角度看，我们其实有好几种灵活的手段来推动增长，这挺有意思的。

And so we have a few different, like, actually very fun from, like, an operations perspective because you have a few different levers on how you can go and scale it.

你得有那些真正带来利润的物理服务器。

Have, like, the metal which actually, like, makes your margins.

还有云资源弹性扩展之类的。

You have the cloud burst, etcetera.

你还可以用贷款来购买服务器。

You have debt you can use to, like, buy servers in general.

所以这其实是个挺有意思的操作问题：我们手头有这么多现金。

So it's a very interesting, like, operational, like, problem to basically say, like, okay, we have this much cash.

当然，你还可以再筹一笔风险投资。

Oh, and then you have obviously venture capital that you can raise on top of it.

对吧？

Right?

所以你现在有这么多现金。

And so you have this much cash.

那我们该筹多少资金？

How much money should we raise?

如果我们能像扩展计算能力一样快速提升收入，而且继续让别人轻松地构建和部署，那我们到底能多快把这套系统部署下去呢？

How, like, how quickly can we go and deploy it, etcetera, if we can scale revenue as basically as quickly as we can scale compute, provided we continue to make it trivially easy for people to go and build and deploy in that.

你把这个循环闭环越快打通，对资本的运营越高效，业务增长就越快——这本质上就是某种线性部署速度的问题。

Like, the faster you can close this loop and the more operationally excellent you are with the capital, like, just the faster your business can It's just a basically like straight linear kind of like deployment rate on some of that stuff,

你懂的。

you know?

我觉得做基础设施的初创公司用债务融资，这个手段大家用得太少，了解得也不够。

I think infra startups raising debt is a tool that people don't utilize enough or know enough about.

天啊。

Oh my god.

你能说说这方面的情况吗？

What can you tell us about that?

意思是

Mean Is

这是用你的GPU做抵押的吗？

this secured against your GPUs or what?

是的

Yeah.

就是用我们的硬件作抵押。

It's just secured against our hardware.

对吧？

Right?

那你们能拿到什么利率？

And so What rates do you get?

贷款方都是谁？

Like, who are the lenders?

我们

We

就按当前的基准利率加一点，而且等债务减少后还能再融资。

just pay, like, prime at whatever it is, plus, like, we can refinance any of the debt as it goes down.

从这个角度看，条件挺不错的。

Like, the terms are pretty good from that perspective.

我觉得问题在于，推特上根本没多少细微差别。

I think, the unfortunate thing is like Twitter has no nuance or whatever.

所以他们就觉得风险债务一无是处之类的。

So they're like venture debt bad or whatever.

其实不是这样的，就像所有事情一样——这根本不是风险债务。

It's like, well, no, like as with all things like- It's not venture debt.

对。

Yeah.

没错。

Or Yeah.

这是数据中心债务，对吧？

It's data center debt, right?

不过我觉得，确实有些工具和领域，你可以非常明确地避免把某一种工具当万能锤，比如把风险投资当成解决一切问题的手段。

But like, yeah, I think there's specific tools and specific areas where you can be very, very deliberate about not just using one specific tool as a hammer, like venture capital as a hammer for everything.

你得主动去了解，弄清楚它到底是怎么运作的，比如说

You just have to kind of, like, go out and explore it and figure out how it kind of, like,

风险投资的融资成本是最高的。

flows for VC is the most expensive financing you can get.

对。

Yeah.

没错。

Yeah.

顺便说一句，我觉得大家对风险投资在融资这件事上的理解完全错了。

I I think, incidentally, I think also people think about VC completely wrong from a raising capital perspective.

说说看，风险投资错在哪。

Like, tell us how this VC

是错的。

is wrong.

对。

Yeah.

我觉得大多数人会这么想：好吧。

Well, I I think most people are like, okay.

那我该怎么在那个时候，尽可能从最合适的人那里融到最多钱呢？

Well, how do I raise as much money as possible from, like, whoever is, like, probably the best that I can get at that point in time?

我觉得这个想法其实差不多是对的。

And I think that's, like, kind of close to right.

但我觉得你真正该做的，或者我们过去一直努力去做的，是想想怎么用这笔股权换到一个几乎算得上是不公平的优势，因为这时候你给出的股权，要么是最便宜的，要么是最贵的——前提是你公司未来会越来越好。

But I think what you should be doing, or at least what we've tried to go in and do is, like, try and figure out what almost unfair advantage you can buy with that equity because it's the cheapest equity or it's the most expensive kind of equity you're gonna give away at that point in time, assuming your company's gonna get better and better and better.

你该怎么利用这一点，去找到一个特别出色、能跟你完美互补的人合作？

And how do you use that to, like, go in and work with somebody who is stellar and who's going to go in and compliment you?

对吧？

Right?

比如说，像Serious这样的投资人，对吧？

Like, you know, yeah, like Serious

真是幸运。

So lucky.

没错。

Yeah.

对吧？

Right?

说实话，真不错。

Like, you know, great.

我从来没创办过公司。

I've never started a company.

Lucky的Race。

Race Race from Lucky.

他提的建议特别好。

He's got good advice.

我随时都能给他发消息。

I can text him all the time.

他回得特别快，等等。

He's really fast, etcetera.

太棒了。

Like, awesome.

对吧？

Right?

然后你就接着往前走，后来跟Unusual的John和Jordan合作了。

Then you kind of like move on and and you kind of like, you know, worked with, you know, John and Jordan at Unusual.

对吧？

Right?

他们说，嗯，你大概知道怎么打造产品。

And they were like, yeah, you roughly know what you're doing in building a product.

我们就主要让你自己发挥，需要建议的时候随时找我们。

Like, we're just going to mostly like leave you alone and be totally available for advice.

太棒了。

Amazing.

非常好。

Awesome.

一路做到A轮。

Get to series A.

做生意简直就是一团乱麻，毕竟我们根本不知道怎么把业务做大。

Business is a total, you know, operational tire fire, right, because we just don't know how to scale a business.

对吧？

Right?

去跟艾丽卡合作吧。

Go and work with Erica.

还有，乔丹在红点资本，这倒是意外之喜。

And, you know, Jordan's over at Red Point, so bonus.

我们可以一直跟他们合作。

Like, you know, we get to work with them continually.

对吧？

Right?

现在我们融了TQ和FPV的钱，开始往企业市场拓展了。

And then now moving into, like, you know, raised from TQ and FPV, like, we're moving into the enterprises now.

对吧？

Right?

然后顺势推进下去。

And, like, feeding into there.

对吧？

Right?

所以每一步，我们都在考虑，这个时候该和谁合作？

So every step of the way, we've kind of moved towards, like, who can we partner at at this specific time?

谁能帮我们打开下一阶段的突破口？

Who's gonna help us unlock that next section of the journey?

因为你知道吗？

Because guess what?

我真的不懂企业销售。

I just I don't know enterprise sales.

我大概能凭直觉判断，作为工程师，我觉得我们需要哪些功能，好在我们内部有很棒的人能帮我们。

I can roughly like eyeball it and be like, yeah, as an engineer, I think these are the kind of features that we're gonna roughly go in and need, and we have some wonderful people who are gonna help us internally.

但真正要和那些人谈的时候，得在董事会层面达成共识，让他们觉得：没错，我们目标一致，这正是我们该做的事。

But you really wanna work with those people at the boardroom dynamic level are going be like, Oh yeah, we're all aligned, that's obviously what we want to go in and do.

我们可以把时间花在思考怎么赢得这场胜利上。

And we can spend our time basically saying, How we win this?

而不是纠结于战略问题。

Versus bickering about strategy.

对吧？

Right?

不，我刚才只是翻出了些资料。

No, I just had to pull out

一些非常漂亮的数据中心图表。

some beautiful data center charts.

我觉得你之前做过其他的。

I feel like you've done others.

但我就是没找到。

I just couldn't find them.

不过这些也不错。

Well, these are good.

意思是，它们看起来都差不多。

Mean, like, they all kinda look the same.

就是一堆服务器放在机架里。

Like, they just they're servers in a rack.

这是我们的机箱。

This is our box.

对。

Yeah.

没错。

Exactly.

这是我们的机箱。

This is our box.

对吧？

Right?

就这个机箱。

This box.

你是想看看更多的机架吗？

It's like, do you wanna see more racks?

哦，对啊。

It's like, oh, yeah.

你懂的吧？

It's like, you know?

那是什么

What's the

杰克·库珀签名版？

Jake Cooper signature edition?

是的。

Yeah.

我们其实内部已经有计划了。

It's we have we actually have we have plans internally.

对，这会很有趣。

Yeah, so it'll be fun.

我们今年会搞一些不同的促销活动和特别企划，挺有意思的。

We've got a few different promos that we're gonna do and like stunts for the year, so those will be fun.

是的。

Yeah.

在结束这部分之前，你发过一条关于太空数据中心的推文。

You had a tweet about data centers in space just before we wrap up this section.

为什么不在太空建数据中心呢？

Why no data centers in Why

你为什么这么讨厌这个想法？

do you hate so much?

好吧。

Okay.

并不是说太空不能建数据中心，实际上我觉得这问题能解决。

So it's not no data centers in space because actually, think, like, my hot take is, like, I think this is solvable.

只是我还没见过谁真正把它做出来罢了。

I've just never seen anybody solve it, right?

因为你得考虑，比如说

Because you need to, like

不，不，不。

No, no, no.

你刚才问，你怎么在真空中散掉那么多热量？

You said you said how are how are you gonna dissipate that much heat in a vacuum?

你这是在提出一个物理上的主张。

You're making a physics claim.

是的。

Yeah.

对。

Yeah.

没错。

Yeah.

因为我就没见谁证明过，怎么在真空中把那么多热量散掉。

Well, because because I haven't seen anybody, like, prove how you're gonna go and dissipate that much heat in a vacuum.

对吧？

Right?

这并不意味着它不可能实现。

Like, it it doesn't mean that it's not possible.

只是说还没人真正把它展示出来。

It just means that, like, nobody's kind of put it up there.

抱歉？

Pardon?

Astrophage。

Astrophage.

火星那个东西。

The Martian thing.

明白了。

Okay.

你运气真好。

You're very lucky.

是的

Yeah.

对

Yeah.

这说得通

That's fair.

不过，确实

But, yeah.

我不确定

I don't know.

我的意思是，从整体上看，这可能是行得通的

I mean, it could work in general.

对吧

Right?

但我认为很多人，顺便说一句，这大概就是你得去做的事情

But I think a lot of people and I I think, incidentally, this is probably what you have to sort of do.

这简直是本末倒置了。

It's like they're putting almost the cart before the horse.

就好像说，哦，我们要把数据中心搬到太空去。

It's like, oh yeah, we're gonna put data centers in space.

然后呢，怎么搞？

It's like, okay, but how?

好像我们还有点时间，慢慢把这事琢磨出来。

It's like, well, we have some period of time to basically figure it out.

对吧？

Right?

就像《火星救援》里那样，他们说，咱们怎么才能跟那个东西对接上呢？

It's like it's like, you know, in The Martian where they're like, oh, how are gonna like intercept with

是的。

the Yeah.

返回地球的宇航员。

Restropatriates.

哦，明白了。

Oh, okay.

对。

Right.

问题是，我们该怎么实现呢？

It's like, how are we gonna do that?

嗯，我们总会找到办法的。

It's like, well, we'll figure it out.

我们还有时间去琢磨清楚，你说是吧？

We have however long to go in and figure that out, you know?

是的。

Yeah.

赌人类的创造力挺奇怪的，因为你只能盲目相信这事最终能解决。

A bet on human invention is weird because you just have to blind trust that it can be solved.

百分百。

A 100%.

对吧？

Right?

我觉得物理学里有些基本原理和边界，可能根本行不通。

I feel like physics and there's some first principles, bounds that you can put on, like, maybe not.

是的，你知道的。

Yeah, know,

对吧？

right?

比如你这要求的是时间旅行，或者违背了某些基本的热力学定律。

Like maybe you're asking to travel time here or like, it breaks some fundamental thermodynamic law.

是的。

Yeah.

而且我并不

And I don't

不知道风投是怎么做的，顺便说一句，他们怎么判断哪些事根本不可能，纯属骗局，而哪些事虽然听着疯狂，但其实有可能呢？

know how VCs do this, like, incidentally too, because it's like, how do you know what's, like, basically not possible and, like, is a grift versus, like, is possible, but, like, sounds completely insane.

对吧？

Right?

然后你就想，哦，厉害啊。

And you're like, oh, cool.

比如说，我们要把数据中心放到太空里。

Like, you know, we're gonna put data centers in space.

嗯，好吧。

It's like, okay.

这事儿到底成不成，纯看运气，谁也说不准。

Coin flip as to whether that's, like, one or the other, you just don't know, I guess.

我想，十年后就能看明白了。

And I guess you'll know in like ten years.

是的。

Yeah.

不错。

Cool.

这真是个有趣的循环。

That's a fun cycle.

好的。

Okay.

明白了。

Okay.

是的。

Yeah.

所以。

So

我们再回来说说智能代理。

Moving back to agents.

对。

Yep.

我觉得你做的那种分支处理、快速启动和协调，正好就是智能代理所需要的前期准备。

I think the branching that you do, the fast spin up and orchestration, it's kind of like the pre work that happened to be exactly what agents want.

代理和人类的需求有什么不同？

What do agents want differently than humans?

代理和人类的需求有什么不同？

What do agents want differently than humans?

我觉得它们希望有能力对内容进行版本管理。

I think they want the ability to version things.

其实也没那么不一样。

So, it's not like actually that different.

只是在实现方式上有些细微的差别，对吧？

Like there's just almost like slight deviations in terms of how it kind of materializes, right?

所以代理需要一种能逐步测试更改的方法，对吧？

So agents want a way to be able to go in and test changes incrementally, right?

像我们工程师会用功能开关，对吧？

Like we have feature flags as like engineers or whatever, right?

那它们为什么不能直接用功能开关呢？

Like, is there any reason why they can't just use feature flags?

对吧？

Right?

我不这么认为。

I don't think so.

我觉得其实完全可以直接这么做。

Like, I think there's ways that you can just go in and do that.

对吧？

Right?

他们需要版本控制。

They want version control.

我们能不能用 Git，或者不用 Git？

Is there ways we can use Git or not Git?

我觉得这一点现在还真的说不准。

I think that one is, like, realistically, completely up in the air.

对吧？

Right?

我觉得，最终会有一种超出 Git 的方案出现，用来长期管理这些内容的版本。

And I do think that's something ultimately outside Git will emerge in terms of how we're gonna go and version a lot of these things over time.

他们需要可观测性。

They need observability.

你得能查清楚某个时间点发生了什么，哪些步骤出错了，还有追踪记录、日志、指标这些信息。

You need to be able to go in and essentially query what happened at what point in time, which steps failed, traces, logs, metrics, all of those other things.

他们还需要网络计算和存储资源。

They need like network compute and storage.

他们得能写文件、保存文件、修改文件，支持文件系统快照，诸如此类的功能。

They need the ability to write files, save files, iterate on files, snapshots file system, all of those other things.

对吧？

Right?

所以我觉得，我们大致需要的这些功能，和智能代理所需要的其实非常相似。

And so I think a lot of the stuff that we roughly needed is like very, very kind of in line with a lot of the stuff that agents also need.

对吧？

Right?

所以分支和分叉这些事，其实没什么不同。

And so like the branching and forking stuff, like, it's not different.

我们只是比以前快了上千倍而已。

Like, we're just moving a thousand times quicker than than we used to.

有些东西看起来好像需要完全颠覆性的改变，但实际上只是需要比现在的好上太多。

And so some of these things, like, look like you really need, like, something massively, massively different, but it's just you need something massively better than what currently existed.

对吧？

Right?

你需要编排能力。

You need orchestration.

你需要一个远超 Kubernetes 的解决方案。

You need something massively better than cube.

对吧？

Right?

你需要网络支持。

You need like networking.

你可能需要一个比Envoy更强大的方案。

You need something probably better than Envoy.

对吧？

Right?

这问题会一直渗透到整个技术栈——如果工作负载本身没太大变化，只是被极度压缩了，因为你得同时处理成千上万的任务，那哪些前提假设得改？

Like, and it just goes all the way down the stack essentially in terms of, well, if the workload profile doesn't change so much as it gets like massively, massively compressed because you need to do thousands of these things, what assumptions change?

对吧？

Right?

etcd肯定扛不住。

Like, etcd is gonna melt.

对吧？

Right?

你得找别的东西来替代它。

Like, you know, you need to replace it with something.

对吧？

Right?

然后你可以一路往下分析整个系统，说到底，

And then I think you can go all the way down the stack and basically say, okay.

这部分得换，那部分也得改，还有这一块也得重新设计。

Well, that part has to change, and that part has to change, and that part has to change.

有趣的是，这种超指数增长的趋势意味着，你的系统必须设计成能随时替换掉这些组件，因为新的瓶颈随时可能出现——比如你把并行代理的效率提升到极致之后。

And the interesting thing about the kind of like super exponential curve is that you have to build your systems in such a way where you can rip out those parts at any point in time because a new bottleneck might emerge because, you know, you start getting really, really good at like parallel agents.

代理。

Agents.

对吧？

Right?

这时候新的瓶颈就出现在这儿了。

And then that's kind of where the new bottleneck is.

对吧？

Right?

然后这就又拖垮了系统里的另一个部分。

And that breaks a different part of your system.

对吧？

Right?

所以我觉得这跟人类一直以来需要的东西其实很像。

So I think it's very much like similar kind of stuff that kind of like humans have needed.

只是你需要的是千倍规模的版本。

You just need at a 1000x scale.

对吧？

Right?

那在智能代理的时代，代码审查该怎么进行呢？

So like, how do you code review in the age of the agents?

对吧？

Right?

我想这更像是一个

I guess it's more of

问题。

a question.

你让更多的代理去处理那些它们不该做的任务。

You tell more agents who don't.

是啊，对吧？

Yeah, right?

但接下来，谁来检查这些漏洞和类似的安全问题呢？

But then like, who reviews things for like CVEs and like all of those other things?

能派上用场的代理。

Can get kids.

更多的代理。

More agents.

好，明白了。

Great, okay.

然后，迟早我们会遇到推理能力的瓶颈。

And then that's how we hit the inference wall at some point.

对吧？

Right?

你可以不断往这个问题上堆代理、代理、再堆代理。

And you can continually throw agents and agents and agents at that problem.

对吧？

Right?

但说实话，我觉得能堆上去的代理数量终究是有限的。

But like, you know, I think there's, I think there's a limit to like the amount of agents you can kind of throw at a problem.

不过你早就开始做了，那时候CLI还没流行呢，我想。

You started, though You already had a CLI before it was cool, I guess.

有

Has

顺便说一句，CLI一直都很酷。

the CLI's always been cool, by the way.

不过，是的。

But, yeah.

你所暴露出来的功能，整体上有没有什么变化？

How has the shape of, like, what you're exposing change, if at all?

是的。

Yeah.

所以我觉得 CLI 会变，因为我们考虑的是，怎么给 Claude、Codex、Chat 或者其他任何模型找个抓手。

So, so I think the CLI changes because the way that we think about this is, like, how do you give Claude or Codex or Chat or, like, whatever, like any of these models almost like a handhold.

对。

Mhmm.

从本质上说，CLI 就是一条命令，对吧？

And, like, a CLI is a single command when you think about it, right?

比如，你就是要部署，或者干点别的什么，对吧？

It's like, okay, well, you're going to do deploy or whatever, right?

你要查日志，或者做别的事情，对吧？

You're going to get logs, you know, whatever, right?

那些对人类来说特别烦琐的事，对智能体来说根本不是事儿。

Like things that were prohibitively annoying to humans are not actually prohibitively annoying to agents.

它们其实特别顺手，对吧？

They're really, really nice, right?

所以，如果我给你一个命令行工具，跟你说：嘿，你猜怎么着？

And so if I wanted to hand you a CLI and I said, hey, guess what?

这个命令行有四十个参数和六百个选项。

The CLI has 40 arguments and 600 flags.

你肯定会说：哇，这也太离谱了吧。

You'd be like, wow, that's crazy.

一般来说，我根本用不到这么多功能。

Like, I'm never gonna use all those things in general.

对吧？

Right?

但你把这个交给一个智能代理，跟它说：嘿，这里有四十个参数和六百个选项。

But you hand it to an agent and you say, hey, that's 40 arguments and 600 flags.

它会说：太好了，这正是我需要的。

You'd be like, oh, yeah, this is excellent.

你知道的。

You know?

我手上有太多入口可以随时进去调整和优化了。

Like, I I have so many handles that I can go in and kind of like work on with this.

对吧？

Right?

所以我觉得，如果你打算通过这种方式向代理开放功能，最好尽可能多地提供入口，让它们能获取信息、查询动态数据，然后尽快完成这个闭环。

And so I think incidentally, if you're going to go in and try and expose things for agents over over that mechanism, you wanna just basically have as many handles as possible where they can get information, query additional dynamic information, and then see how it can close that loop, like, as quickly as possible.

现在大多数问题其实都集中在：怎么才能最快地完成这个闭环？

Most of the kind of like problems right now are actually just how do you close loop as quickly as possible?

代理在哪儿卡住了？你又该怎么去消除这些障碍？

Where does the agent get stuck and how can you go and kind of remove that?

这就是为什么日志监控特别重要——如果你能知道代理在CLI里卡在哪儿，比如发现有12%的用户因为某个问题偏离了正常流程，

That's why incidentally, like telemetry is very, very important because if you can tell where the agent gets stuck from the CLI and you say, hey, listen, like 12% of people are actually getting deviated from the happy path because of this thing.

然后你加了个参数，把这个比例降到了2%，那你实际上就大大提升了大量用户的闭环效率，对吧？

And now I go and add this arg and that drives it down to two percent, you've massively increased the like rate of the loop closing for a lot of people in general, right?

所以我们思考的不只是CLI，而是整个仪表盘里每一个环节都是这样。

So that's kind of the way that we think about not just the CLI, but every point in the dashboard, right?

就像是用户从听说 Railway 开始的整个体验。

Like it is a user journey from, I hear about Railway.

然后你去部署一个项目。

I go and get something deployed.

等到第一次看到绿色构建成功，那种时刻。

I get my first green build, whatever moment.

看到一个端点。

I see an endpoint.

看到一些日志。

I see some logs.

看到各种东西。

I see whatever.

接着你就进去不断调整，对吧？

And then I go in and iterate, right?

然后一遍又一遍地迭代，这个循环会一直持续下去，永无止境。

And then I go in and iterate loop is indefinite and infinite until the end of time.

对吧？

Right?

其实就是用户想部署一个新的应用。

It's basically like user wants to deploy a new thing.

用户想部署一个新的 PostgreSQL 实例。

User wants to deploy a new Postgres instance.

用户想修改自己的代码。

User wants to change their their their code.

用户希望持续不断地迭代。

User wants to iterate all all over time.

对吧？

Right?

所以如果你专注于这些迭代循环，找出阻碍它们快速完成的瓶颈，我们内部经常说的一点是：你绝对不能再让等待计算资源成为常态。

And so if you just focus on a lot of those iteration loops and and figuring out what's blocking that loop from closing as quickly as possible, like one of the things we talk about internally is you never ever, ever want to be waiting on compute anymore.

你该等的，永远是智能的决策。

You always want to be waiting on intelligence.

对吧？

Right?

如果你一直在等计算资源，那说明那里有个瓶颈，必须得彻底解决，因为迟早这个瓶颈会大到让别的流程冒出来，把这一大堆东西都改掉。

And if you're waiting on compute, there's a bottleneck that needs to be destroyed there because at some point that bottleneck will be so, so, so large that some other workflow will kind of emerge to go in and change a lot of that stuff.

顺便说一句，我们其实做了一个特别棒的产品，你只要推送代码，它就会自动构建，还有其他所有相关操作都帮你搞定。

I think incidentally, we've built a really, really awesome product where you can push code and then you build the code and all those other things.

对吧？

Right?

但那种推送-拉取之类的循环，我从根本上觉得，它迟早会消失。

But that push pull whatever kind of like loop, I just fundamentally believe it's gonna go away.

对吧？

Right?

你会到达这么一个阶段：你在生产环境做个小改动，就能瞬间更新整个基础设施的版本。

Like it's you were gonna get to a point where you make a small change in production, that changes version across your entire kind of infrastructure.

你同时在和数据库、整个基础设施的副本一起工作，改完之后一合并，立刻就生效了。

You're working alongside, you know, copy and write versions of your your database, all of your infrastructure, and then you merge it in and instantaneously, it's it's like live.

对吧？

Right?

因为这正是循环系统的终极目标。

Because that's like the holy grail of loops.

对吧？

Right?

但那种推送、拉取、重新构建的流程，正是我们正在彻底消除的摩擦点。

But that like push pull rebuild thing, right, is a point of friction that we are like removing entirely from our loops.

是的。

Yeah.

速度非常快。

It's incredibly fast.

所以如果还没试过，真的推荐你试试。

So if anyone hasn't tried it, like Yes.

没错。

Yeah.

这种快速反馈真的很棒。

That fast feedback is great.

我的看法是，Railway 以前因为它的可视化面板出名，那个面板能展示你的基础设施，但那主要是给人看的。

You know, my hot take is that Railway was kind of famous for its canvas, which sort of visualizes your infrastructure unless you manipulate it visually, but that was for humans.

对。

Yeah.

实际上，现在到了下一阶段的发展，Railway 的命令行工具比那个你出名的可视化面板更重要了。

And actually now, for the next phase in growth, like, Railway, CLI is more important than canvas, which is what you're famous for.

对。

Yeah.

我觉得那个可视化面板挺有意思的，因为它本质上只是用来展示你随着时间的变化情况。

So I think the Canvas is funny because like, it's actually just a mechanism to show you changes over time.

不过你说得对，我们过去确实经常把它当输入工具，但未来它的定位会更偏向于输出。

But I think you're totally right in the sense that like, we have previously used it a lot as an input and its goal moving forward is actually a lot more like an output.

我的意思是，你以前会打开面板，做些改动，然后一堆其他事情也就跟着发生了，对吧？

What I mean by that is you would go to the canvas and you make some changes and all these other things, whatever, right?

你看到这些变化后，你的代理或者基础设施就会随着时间慢慢演变，对吧？

And you see them and, you know, your agents or your infrastructure would evolve over time, right?

现在你只需要一堆代理，它们能直接用 CLI，随时去做出这些改动。

Now you just have a bunch of agents that like, they have access to CLI and they can go in and make those changes in general.

对吧？

Right?

所以，Canvas 其实不再是一个让你去动手操作的输入工具，像以前那样：‘哦，太棒了，我该怎么弄？’

And so the canvas actually, instead of becoming this like input thing where you're like, oh, cool.

我该怎么去实现这个改动？

Like, how do I go in and make this happen?

它现在更像一个输出展示。

It's actually just more of an output thing.

它本质上是在告诉你当前的信息——就是个仪表盘。

It basically says what information- Dashboard.

是的。

Yeah.

这时候，人需要看到哪些信息，才能做出合适的决定，比如是否批准这个请求？

What information does the human need at this point in time to make suitable decisions about about, like, control requests of, do I approve this?

要不要批准？

Do I not approve this?

对吧？

Right?

实际上，到了这一步，画布就只是这么个东西了，对吧？

Like, that's realistically all the Canvas becomes at that point in Right?

而且我觉得这点很重要，很多做画布体验的人反而忽略了这一点。

Also a way and I think this is important, and I think this is lost on a lot of people who are, like, building some of these, like, Canvas experiences.

它必须像一个能锚定上下文的支点。

It has to be almost like an anchor for your context.

必须像风暴中的避风港。

It has to be like a port in the storm.

你得把它想象成一层层的结构，就像文件系统一样，一步步找到下一个目标。

It has to be like you have to think basically about it as, like, layers and, a file system almost to, like, get to the next spot.

对吧？

Right?

所以你的所有基础设施都在这儿，这就是为什么Canvas一开始只是一个项目。

And so you have all your infrastructure and like, this is why the Canvas starts as like, it's just a project.

对吧？

Right?

然后你有个逐层深入的图表，对吧？

And then you have a drill down chart, right?

也就是说，你要把整个系统拆解成各个服务、模块、函数或者代码片段，因为你不仅要自己心里清楚，还要在Canvas里完整呈现出来，让别人也能看懂，跟你的思路同步，才能跟上节奏。

Like, it's like, I'm breaking down into these services or this like section that just is like a function or code or anything else like that, because you want to actually be able to represent the entire thing, not just in your head, but in this canvas so that other people can also get that representation so that they can think on the same wavelength as you so that they can move as quickly.

对吧？

Right?

我觉得很多公司，尤其是规模大了之后，问题就出在这些上下文全靠个人脑子记着。

I think a lot of orgs, especially as they scale, they get in trouble because all that context lives in somebody's head, basically.

然后就有人问：这个微服务到底是怎么运行的？

And then it's like, oh, how does this microservice work?

我完全搞不懂。

It's like, I have no idea.

去问那个人吧。

Go ask this specific person.

对吧？

Right?

现在还出现了很多产品类别，专门解决这类上下文查找的问题。

And then you have entire categories and classes of products that are built around like, how do you do context discovery at all these things?

我觉得，如果你能建立一个非常清晰的层级结构，能无限嵌套服务、代码和各种组件，那才能慢慢搭建起这样的体系，你明白吗？

And I think a lot of that stuff just gets melted in terms of if you can have a really, really solid hierarchy and you can infinitely nest services, infinitely nest code, infinitely nest all these things all the way down, that's what allows you to kind of build these kind of like structures up over time, you know?

我觉得这也是我们未来能构建更庞大系统的关键，我之前写过一些关于这种‘超结构’的东西，就是那种大到难以想象的系统。

And I think it's also what's going to allow us to like build, I've written a bit about this, like these like hyper structures, like things that are way, way bigger.

你看金门大桥，你是不是也会想，这到底是怎么建起来的？

And like, you know, you look at the Golden Gate Bridge and you're like, how, how did we build that?

就像网上那个梗说的，我们到底是怎么造出这些东西的？

Like, you know, we've, there's that whole meme of like, oh, how do we build this?

我们把这项技术给丢了。

Like, we lost the technology.

我们不知道怎么做。

We don't know how.

现在真的不知道该怎么做了。

We don't know how anymore.

对吧？

Right?

说白了，某种程度上确实如此，因为当初建造这些东西所需的那些协调方式，早就变了。

It's like, well, yeah, I mean, to some extent, yes, because a lot of the coordination that built those things like has evolved, right?

而且已经不一样了。

And like has changed.

我们还失去了很多过去那种搭建结构的技巧，现在什么都往Slack里一塞了事。

And there are new things that we've lost almost like some of the art of like building that structure as we've just like jammed everything into Slack.

对吧？

Right?

我们现在就是，所有事情都靠 Slack 来处理，简直——

And we're just like, everything happens through Slack and it's just-

但你们在 Discord 上也有交流，所以——

But you have anything in Discord, so

对。

Yeah.

这其实说的是一回事。

Well, that's the same point.

其实无所谓。

It doesn't really matter.

就是不停地发消息、打断，发消息、打断，发消息、打断。

It's just like message passing and interrupts, message passing and interrupts, message passing and interrupts.

对吧？

Right?

你是说，应该有个更系统、更结构化的方式？

Like- So you're arguing that there should be something better, more structured?

比Slack还好？

Than Slack?

是的。

Yeah.

对。

Yeah.

好吧。

Okay.

当然了。

For sure.

我觉得Slack不行，顺便说一句，我觉得Discord也很差。

I think Slack I think and incidentally, I think Discord is awful too.

这就相当于我妈的测试标准。

This is the equivalent of my mom test.

对吧？

Right?

就是说，你目前用了什么方法来搭建你的解决方案？

Like, what have you done that has your solution

来解决这个问题的？

to this?

所以我们内部自己开发了一款叫Central Station的工具，能把所有用户的相关信息都整合到一起。

So internally, we've built a tool called Central Station that allows us to go in and aggregate all the context from all of our users.

所有的用户反馈、客户支持工单，凡是这类信息，都会被归类整合到我们称为“聚类”的分组里。

So every piece of feedback, every piece of customer support, every single thing like that gets aggregated into what we call like clusters.

如果正在发生突发事件，或者类似的情况，我们现在就能去确认有多少用户受到了影响，以及其他各种相关信息。

If you have an incident brewing or like anything else like that, now we can go and determine how many users are affected, all of those other things, etcetera.

然后我们可以基于这些信息，专门开辟一个讨论区。

And then we can actually break off a discussion based on that.

我觉得这样做其实更有帮助，也更准确。不像以前那样，只有几个长期存在的频道，让你纠结到底该把这条消息发到哪个频道里。

And I think a lot of that is actually a lot more helpful and more correct in terms of instead of, like, having just these, like, long running channels where you're just like, which channel should I put this thing in?

对吧？

Right?