AI Spend Plateaus as Agents Become Coworkers
打开互动全文版（中英对照 + 朗读 + 问答）→Dylan Patel 谈 AI 支出为何在稳态下趋于平缓、向持久化 agent 的转变、监管压力下被推迟的模型发布，以及替代加速器的竞争格局。
Dylan Patel on why AI spend is leveling off at steady state, the shift toward persistent agents, model releases delayed by regulatory pressure, and the alternative accelerator landscape.
我们会有一段你走进来大喊的视频。
We're going to have a little video of you walking in yelling.
我太兴奋了。
I'm so excited.
哦，真的吗？
Oh, really?
不。
No.
你知道，我们在这些播客上玩得很开心。
We had so much fun on these podcasts, you know.
是啊，我不知道。
Yeah. I don't know.
你看了上周那期吗？
Did you see the one last week?
不过我收到了反馈。所以，你想开始这期播客吗？
I had gotten feedback though. So, do you want to start this podcast?
我们来聊反馈。上次我们剪掉了冷开场里的一段，我说我不再看评论了，因为有一期和 Doug 一起……
Let's do feedback. We cut out something that was going to be in the cold open last time where I said I'm no longer listening to the comments because on one episode with Doug and...
他们气坏了。
Got so pissed.
哦，是吗？他们就是讨厌我们。现在他们讨厌我们所有人。
Oh, did they? They just hate us. They just hate us all now.
那段剪辑，老兄。他剪掉了，剪掉了我反驳他们的部分。
That one clip, dude. And he cut back. He cut out me pushing back on them.
是啊，是啊，是啊。因为我觉得……
Yeah. Yeah. Yeah. Cuz I thought that...
我当时完全觉得，“好吧，他们内部建了地图。他们建了 Gmail、Google Drive，整个 G Suite。他们建了整个 GCP。”
I fully was like, "Okay, they built maps in house. They build Gmail, Google Drive, like the whole G Suite. They built all of GCP."
我觉得你不了解 Kubernetes。我所有 DeepMind 的朋友，大概有三个说，“是啊，我想我要离开了。”
I don't think you understand Kubernetes. All my DeepMind friends, there's like three of them who are like, "Yeah, I think I'm going to leave."
是啊。
Yeah.
我还有一堆其他人说，“去你们的。”不是你们，而是……
And I've got a bunch of others who are like, "Fuck you guys." Not you guys, but...
听起来他们处于第二阶段。
Sounds like they're in stage two.
第二阶段。是啊，我只知道应对。
Stage two. Yeah, which I only know cope.
所以，回到 Form Warrior 时代，我们会陷入芯片争论，有一个私密的 Discord，里面有一群喜欢动漫的人，来自世界各地，其中很多是种族主义者，因为互联网上都是匿名的人。但他们都喜欢动漫，而我不喜欢动漫。我从来没真正看过。除了，你知道，我约会过一个女孩，我和她一起看过动漫，但除此之外，他们总是……不，不，我从来没约会过任何人。我很纯洁。
So, back in the Form Warrior days, we would get into chip arguments and there was like a private Discord where there's a bunch of people who loved anime and people who were all around the world and many of whom were racist cuz it's anonymous people on the internet. But they all loved anime and I did not like anime. I've never really watched it. Besides like, you know, one girl I dated, I watched anime with her, but besides that, like they would always... No, no, I've never dated anyone. I'm pure.
你看了哪部动漫？
Which anime did you watch?
哦，我看了……好吧，有一部我很喜欢。我喜欢《间谍过家家》。
Oh, I watched... Okay, there's one I love. I love Spy x Family.
阿尼亚？
Anya?
当然。
Sure.
阿尼亚酱，我不知道怎么说……
Anya Chan, I don't know how to say...
我不知道。你知道这个梗吗？
I don't know. Do you know the reference?
是的，我知道。
Yeah, I do.
谁做的一个玩偶？
Who does one of the dolls?
你给我买了一个阿尼亚。不，比如那些……
You bought me an Anya. No, like what are the...
那是什么？拉布……
What is it? Labu...
那个系列里的东西。
Something from that series.
那个……那个什么？他叫什么名字？
What's the... what's it floor? What's his name?
我不知道。
I don't know.
洛伊德。
Lloyd.
洛伊德。这就对了。
Lloyd. There we go.
我比你知道的动漫还少，老兄。
I know less anime than you, man.
嗯，是啊。所以我们快速换话题。Jordan Anos 是……这就像 HR 批准的，因为我是 HR。Jordanos 是半导体分析界最火的男人。
Well, yeah. So we're changing topics rapidly. Jordan Anos is... And this is like HR approved because I'm HR. Jordanos is the hottest man in semi analysis.
这个…… Michelle 有这个……天哪。他……不，不。想想看。看看他。我胖，看看他。你知道，看他多帅。高个子。和我同龄，但他结婚了，有孩子。
Got this... Michelle got this... My gosh. He's... No, no. Think about it. Look at him. Like I'm fat and look at him. You know, look at him so beautiful. Tall. Same age as me except he's married and has a kid.
有房。他不是个堕落的人。你知道，这就像哇，目标。
Owns a home. He's not a degenerate. Like you know, this is just like wow goals.
谢谢雇佣，老兄。好吧，抱歉。回到正题。谷歌的人对我们很生气，他们私信我，有些人就像，你知道，他们在应对。但不管怎样，我收到的反馈……
Thanks for the employment, man. Okay, so sorry. Going back. Google people were mad at us and they were DMing me and some of them were like, you know, they're in cope. But regardless, the feedback I've gotten...
主要来自我自己的脑袋，我自己的脑袋。
From mostly just my own head, my own head.
我们是不是给出了太多价值？
Are we giving away too much value?
这就是我今天来的原因，因为我需要破坏价值。
That's why I came on today cuz I need to destroy value.
好吧，我们可以停了。
All right, we can stop.
不，不，不，不。别停。很有趣。很有趣。但团队内部有人说，“Dylan，我们在周报上给出了很多价值。”我说，“嗯，我们确实。我没听过，但我打赌我们确实。”我听了我们请 DG Matrix 那家伙的那期，我说，“这太火了。”
No, no, no, no. Don't stop. It's fun. It's fun. But someone on the team internally was like, "Dylan, we give a lot of value away on the weekly." And I'm like, "Huh, we do. I haven't listened to it, but we do, I bet." I listened to the one where we had the DG Matrix guy and I was like, "This is fire as fuck."
是啊。谁说的？
Yeah. Who said that?
“兄弟，拜托。HR 是为了文本匿名。”不是 Doug。
"Bro, come on. HR is for text anonymity." It wasn't Doug.
好吧。Jeremy？
Okay. Jeremy?
不是 Doug D。
It wasn't Doug D.
所以，我不知道谁有发言权。是不是有人因为还没上过节目而不高兴什么的？
So, I don't know who has the say. Is it just somebody who's like upset that they haven't been on yet or what?
不，不。上过节目的人。上过节目的人。
No, no. Someone who's been on. Someone who's been on.
Dan？
Dan?
我不想说是 Dan。Dan 不会那么说。Dan 是个甜心。总之，不管怎样……
I don't want to say Dan. Dan wouldn't say that. Dan's a sweetheart. Anyways, regardless...
那说的是谁？
Who does it say?
反馈是这期播客太好了。
The feedback is that this podcast is too good.
是啊。好吧。
Yeah. Okay.
我们为什么免费给出？
Why are we giving away for free?
哦，总之……
Oh, anyways...
是啊，这次我们肯定能搞砸。
Yeah, we can definitely put in the toilet this time.
全是阿尔法。人们点进来是因为我在，然后他们说，“这垃圾是什么？”
All alpha. People click because I'm on and they're like, "What the hell is this trash?"
不，我们会有一张你穿霓虹橙色衬衫的好照片，准备吸引所有点击。
No, we're going to have a nice picture with you with a neon orange shirt ready to attract all the clicks.
是啊。来，来展示你的衬衫。
Yeah. Come, come show your shirt.
所以，是 Nick 说我们给出了太多价值吗？
So, was it Nick who said we're giving away too much value?
不，不，不。看看 Nick。
No, no, no. Look at Nick.
哦，是 David。
Oh, it was David.
不，不是 David。
No, it wasn't David.
不是 Sales。
Wasn't Sales.
好吧。很高兴见到你，老兄。谢谢你来。
All right. Good to see you, man. Thanks for coming by.
我们来谈谈纽约的半导体分析办公室。老兄，它……你去过了吗？
Let's talk about the semi analysis office in New York. Man, it's... Have you been yet?
没有。这就是我要去的原因。
No. That's why I'm going.
你为什么去？你要回到……
Why would you go? You're going back to the...
我们在升级。我们在升级。
We're upgrading. We're upgrading.
哦，什么时候？
Oh, when?
很快。非常快。
Soon. Very soon.
这就像买 GPU，对吧？如果你对租约承诺太久，那你就得想办法转售它。
This is like buying GPUs, right? If you make too long of a commitment to the lease, then you have to find a way to resell it.
你得做六个月的办公室承诺，这样你可以超越它们。
You got to make six-month office commitments that you can outgrow them.
我应该直接买 GPU，而不是更多办公空间，而不是酒店。是的。是的。
I should just buy GPUs instead of more office space, instead of hotel. Yes. Yes.
所以，你希望半导体分析只是个 GPU 经销商？
So, you wish that semi analysis was just a GPU reseller?
不。
No.
你可以替换我们所有人。
You can replace us all.
不。不。不。不。Dario 说，“我们要自动化掉我所有的车，因为天哪……”
No. No. No. No. Dario's like, "We're going to automate away all my cars because God..."
兄弟，去看看 AI 支出。我不做。我不做。
Brother, go look at the AI spend. I'm not doing it. I'm not doing it.
哪个增长更快？AI 支出还是员工支出？
What's growing faster? AI spend or spend on employees?
嗯，所以我们经历了招聘热潮，然后是消化期，再是招聘热潮。我们又回到了招聘热潮。所以……
Well, so we've gone through like hiring sprees and then digestion periods and hiring sprees. We're back in a hiring spree. So...
好吧。所以……
Okay. So...
所以员工支出真的飙升，尤其是在去年下半年和今年部分时间，但然后像今年第一季度 AI 支出飙升，但实际在第二季度也相对平稳。
So spend on employees really skyrocketed especially in the second half of last year and parts of this year, but then like the first quarter of this year AI spend skyrocketed, but it's actually been like relatively flat in Q2 too.
对。我们有点像每个人都得了 Claude Code 精神病，然后它就像趋于平稳。
Right. We kind of everyone got Claude Code psychosis and then it's like leveled out.
是啊。
Yeah.
就像它仍然在大约 1000 万这个数字。你觉得它未来会大致随着更多员工而增长吗？
Like it's still at that like 10 million number roughly. Do you think it will grow roughly in line with more employees in the future?
我，你知道，我很惊讶 Fable 没有导致价格上涨。
I, you know, I was surprised Fable didn't cause price to go up.
是啊。支出上涨。
Yeah. Spend to go up.
是啊。
Yeah.
你觉得这是为什么？
Why do you think that is?
大致和 Opus 一样。我想可能是在抵消，因为很多人在构建应用程序的第一个版本。
Roughly the same as Opus. I'd say possibly is counteracting with a lot of people were building the first versions of the applications.
比如我们内部仓库从 10 个发展到了现在超过 150 个。
Like we went from 10 repos internally to over 150 repos internally right now.
我们该卖掉我们的代码、我们的数据吗？
Should we sell our code, our data?
我们正在做。不不不，比如卖给实验室做交易。
We are. No, no, no. Like sell it to the labs to trade on.
我们是我们的垃圾代码。我们的垃圾代码。
We are our slop code. Our slop code.
垃圾代码。我不知道他们是否需要更多模型输出的垃圾。
Slop code. I don't know if they need more model output slop.
是的。我的意思是模型本身可以作为数据出售。
Yeah. I mean the models themselves could be sold as data.
是的。是的。
Yeah. Yeah.
好的。所以你认为这是因为每个人都在做 MVP，现在很多都进入维护模式了。
Okay. So you think it's because everyone was doing MVPs and now it's maintenance mode for a lot of it.
但支出是持续的。并不是说在我们一次性支出之后就下降了。
But the spend is consistent. It's not like it's gone down after we had this one-time spend.
不，当然。是的。但我只是觉得不再有那种情况了，只有一次你让某人入职学习如何使用数据中心模型，做研究把数据构建到数据中心模型中，构建仪表盘，一旦他们上手了，你知道，就像说话一样，然后就平稳了。
No, for sure. Yeah. But I just think that there's no more like there's only one time when you onboard somebody to learning how to use the data center model and do research for building data into the data center model and building dashboards and then once they're onboarded, you know, it's like speak and then it levelizes.
嗯，那或者可能是我们缺乏编解码器的新功能，这些功能能让我们花更多钱来提高生产力。
Well, that or possibly we're lacking new features in Codex that will allow us to spend more to be more productive.
是的。一旦有一种智能体形式，你可以管理一百万个并发智能体，它们协同工作，而不是现在的九个，单一高级用户就能花更多钱……
Yeah. Once there's an agents form where you can manage a million different concurrent agents and they all work together instead of nine today, single power users will be able to spend more than they...
嗯，我想我没算进去的一件事是我们的支出成本没有准确计入云标签。我觉得我们的仪表盘没有显示那个。所以实际上这是个好观点。还有计算机上的计算机，我认为这两项实际上都没有计入支出。所以实际上我们的仪表盘可能是错的，至少是我监控的那个。
Well, I guess one of the things I'm not counting so our spend cost does not accurately account for cloud tags. I think our dashboard doesn't show that. So actually that's a good point. And computer on computers, I think both of those don't actually get counted into the spend. So actually our dashboard is probably wrong, at least the one that I monitor.
我的想法是，很多这些代码相关的东西，实际上我们持续使用的 AI 量非常小。实际上只是人们总是在做新工作。然后因为我们有足够的人，它就趋于平稳，成为一个相当稳定的支出。波动每天只有上下 20-30%。
The way I think of it is like a lot of this code stuff is actually like the amount of AI we use on a continuous basis is actually very small. It's actually just like people doing new work always. Which then because we have enough people, it kind of levels out to be like a pretty steady amount of spend. The swings are only like 20-30% a day up or down.
你知道，有时 Jeremy 会占支出的四分之一，有时又什么都没有。
You know, and sometimes Jeremy will be like a fourth of the spend and then sometimes there'll be nothing.
是的。
Yeah.
对。所以你知道，但然后其他人会补上，对吧？你知道你手下的一个人，我问嘿他到底在花什么钱？他说没有。我说好吧等等等等。我问有投资回报率吗？你列出所有这些，我说太好了。好的，酷。
Right. And so like you know you but then someone else picks up for the slack, right? You know one of your guys, I was like hey what the hell is he spending on? And he's like no. I was like well blah blah blah. I'm like is there ROI? And you list out all this, I'm like great. Okay cool.
你甚至没说太好了。酷。
You didn't even say great. Cool.
好的。我在心里说了。
Okay. I did mentally.
我回复了所有这些细节，就像我现在该给反馈还是什么？
I responded with all this detail like should I give feedback now or what?
哦不。抱歉。抱歉。抱歉。我本该说好的。这没问题。
Oh no. Sorry. Sorry. Sorry. I should have said yes. This is fine.
酷。我只是读了它，然后我想好吧酷。
Cool. I just read it and I was like okay cool.
是的。至少内部是这样。
Yeah. Internally at least.
Phil 只是紧张，因为这个人表面上是个实习生。
Phil was only nervous because this was a person who's ostensibly an intern.
是的。你知道，他还没有我的信任，你知道。
Yes. You know, he doesn't have my trust yet, you know.
是的。是的。
Yeah. Yeah.
比如如果你一天花 2 万……
Like if you spent 20K in a day...
他没有一天花 2 万。但他连续 4 天每天花 8 千，这就像好吧那是很多。你在建什么？但如果你一天花 2 万，我不会质疑你。我只是假设你会做事。只要你交付的价值很大，那就很好。
He did not spend 20k in a day. But he spent like 8k in a day for like 4 days straight, which was like okay that's a lot. Like what are you building right? But if you spent 20k in a day I wouldn't question you. I just assume you're going to do stuff. As long as the value you deliver is great, then great.
嗯，好的。让我震惊的是我不知道他花了那么多，然后我们看仪表盘，我想……
Well, okay. And what's shocking to me is I didn't know he was spending that much and then we look at the dashboard and I'm like...
嗯，这家伙在那些事情上的生产力现在和任何全职员工一样高。所以这是一个现实检验。
Well, this guy is as productive as any of the full-time employees right now on that stuff. So it was a reality check.
哦。
Oh.
所以，这是不是我们，你知道，当我们做，因为过去这家公司的奖金是基于感觉的。基本上我只是凭感觉定奖金数字，那很酷。
So, is this we, you know, when we do, cuz in the past bonuses at this company were vibes based. Basically I just vibed out the bonus number and it was cool.
今年 Claude 将不得不经历，或者 Codex，或者我们可以有两个评审者，对吧？两个内部绩效评审者，Claude 和 Codex 去抓取所有这些 Slack、所有 GitHub，然后说“他们做了什么？”然后把它连接到类似……
This year Claude is going to have to go through or Codex or we can have two reviewers, right? Two internal performance reviewers and Claude and Codex go scrape through all of this Slack, all the GitHubs and say, "What did they do?" And then connect it into sort of the like...
你要委托这个。就让它成为一个子任务。
You're going to delegate this. Just make it a sub.
我不知道。我昨天和 Michelle 讨论了同行评审，我说“哦，天哪。”然后我说完之后，我想“哦，操。”
I don't know. I discussed with Michelle yesterday peer reviews and I was like, "Oh my." And then after I said it, I was like, "Oh, fuck."
哦，老兄。你想在这上面搞大科技那种 360 度评审吗？
Oh, man. You want to go big tech on this 360 reviews, man?
不是 360，不是 360。只是一点点，你知道。然后我们讨论的另一件事是……
Not 360, not 360. Just a little bit, you know. And then the other thing that we had discussed was like...
我们要让人们和他们的上级的上级一起评审，这是……
We're going to have people reviewing with their skip, which is...
我说了一个美国招聘人员和 Doug 在管理频道，Doug 抓狂了。他说“哦我的天，哈利路亚。终于我们可以有了。”他从 30 人时就一直想要 HR。总之，是的。
I said a US-based recruiter and Doug in the admin channel and Doug freaked flipped out. He's like, "Oh my god, hallelujah. Finally, we can have it." He's been wanting HR since like 30 people. Anyways, yeah.
等等，你认为 HR 是招聘人员？
Wait, you think HR is a recruiter?
是的。
Yes.
是的，确实。
Yes, indeed.
好的。总之，所以概念或思考过程基本上就像很多支出是一次性研发。是的。
All right. Anyway, so the concept or thought process was basically like a lot of the spend is one-time R&D. Yeah.
实际上稳态支出真的很低。问题是我们不断做新事情，所以希望你知道这能转化为收入，要么以模糊的方式，比如 cluster max 和 inference X，要么以非模糊的方式，比如现在超级厉害的能源模型。
And actually the steady state spend is really low. The thing is we just keep doing new things and so that hope you know translates to revenue in either a nebulous way in the case of like cluster max and inference X or in a non-nebulous way in the case of like the energy model which is super cracked now.
或者像仪表盘和所有其他这些东西，对吧？比如不同的抓取方法。所以思考过程是……
Or like dashboards and all these other things, right? So like different scraping methodologies. So the thought process was like...
你知道如果我们看这些 AI 整合公司，对吧？嘿，让我们拿一个现有公司，彻底摧毁它的成本结构。用 AI 让它高效，从而摧毁成本结构。那会是什么样子？因为你，比如说，私募股权公司，他们买一家公司。现在他们只是榨干它然后丢弃，让美国民众……
You know if we're looking at these companies that are AI rollups, right? Hey, let's take an existing company, let's completely destroy its cost structure. Nuke its cost structure by just making it efficient with AI. What does that look like? Cuz you, let's say, private equity companies, they buy a company. And right now they just squeeze the rag and discard it and make the American populace like...
是的，AI 用于效率对我来说从来不合理，因为我使用 AI 的方式和我们使用 AI 的方式非常注重研究，这完全低效。
Yeah, AI for efficiency has never made sense to me because the way that I use AI and the way that we use AI is very much about research, which is completely inefficient.
不，我的意思是，但你知道，另一面是，比如我们有智能体检查我们发出的所有发票，它发现了，你知道，我们被支付了，因为我们的手动流程，比如某些交易没有正确标记到发票上，诸如此类，或者我们有，你知道，很多工单的事情至少有些更高效，因为 AI 在回答，但现在他们不发给客户，而是拉取我们所有数据，然后说这是答案，然后分析师，我认为这使每个工单的处理时间更短。
No, I mean, but like you know, the flip side is like, you know, like we had agents go through all of the invoices we've sent out and there was like and it caught, you know, we've been paid because our manual processes like certain deals weren't tagged to the invoice properly and things like that or like we've had you know some you know like a lot of the ticket stuff is at least somewhat more efficient because AI is answering it but now they're not sending it to the customer but like they're pulling through all our data and being like here's the answer and then the analyst I think that makes this port time per ticket shorter.
所以我觉得这些事在帮我们提高效率。
And so I think like things are helping us make be more efficient.
你肯定知道，我不觉得那是我们的主要用途。
Surely you know, I don't think that's the primary use case for us.
我猜这次集群最大化的深度、广度和测试量，和上次集群最大化 2.0 相比，尤其是……
I guess like cluster max this time the depth breadth and amount of testing you're doing versus last you know cluster max 2.0 especially it's like...
对，你可以把它说成效率，但当我听到私募股权接管公司然后榨干毛巾，那意思是裁员、省钱、少付工资之类的……
Yeah, you can frame that as efficiency, but when I hear private equity take over a company and ring the towel dry, it's like meaning firing people and saving money and paying people less and like...
对，那是传统做法。所以我的观点是，那是传统的私募股权方法，而新来的人开始做的是整合，或者说是 AI 私募股权策略，他们称之为整合或其他什么，他们进来后不是试图从那个角度榨干，而是更多地现代化所有系统。哦，你用 Excel 当数据库？好吧，我们换成标准云。前期花很多钱。这就是问题所在，私募股权通常在收购公司时会有一些前期投入用于转型，但实际上并不多，而且很快就能盈利。但 AI 似乎让那个尾巴和前期投入变得更加极端，对吧？你一次性投入很多，然后成本大幅下降，成本效率好得多。所以有很多业务可能都是这种情况，尤其是你知道我们还没到 AI CRM、AI 冷呼叫、AI 发票和会计等所有这些真正达到临界点的地步，但我们非常接近了。
Right, that's the tradition. So my point was that's the traditional PE method, and what new people started to do is the rollup, or rather the AI private equity sort of strategy, which you know they're calling it rollup or something else, where they come in and instead of trying to ring it dry in terms of that angle, they're more so modernizing all the systems. Oh, you use Excel for your databases? Okay, let's just move to standard cloud. Spend a lot of money up front. And this is the thing, private equity generally there's some spend up front when you first acquire a company for some transformation, but really it's like not that much, and it's really like you get the profitability pretty quickly. But AI seems like it makes that tail and front load much more severe, right? Like you spike up on spend a lot for the one time, and then you spike down a lot, and your cost efficiency is way better. And so there's like a number of businesses where that's potentially the case, especially like you know we're still not at the point where AI CRM and AI cold calling and AI invoice and accounting and all these other things are really at critical mass, but we're so close.
对。
Yeah.
你看到今天 Grok 智能体的发布了吗？
Did you see the Grok agents release from today?
没有。
No.
为什么是 Grok 智能体？
Why are Grok agents?
对。Elon 让 Grok 做智能体，它们会模仿你的……
Yeah. Elon's got Grok doing agents where they're impersonate your...
你发音的方式，我以为你说的是 agents。
The way you pronounced it, I thought you said agents.
好吧。
Okay.
我没听清那个。
I didn't catch that one.
Grok 智能体。智能体。好吧。
Grok agents. Agents. Okay.
智能体。对。
Agents. Yeah.
他们发布了什么？
What do they release?
呃，这是老 Grok，不是 Nvidia 的 Grok。还是你说的是 XAI 的 Grok？
Uh, this is the old Grok, not the Nvidia Grok. Or you mean this is XAI Grok?
X Grok。
X Grok.
XIO。
XIO.
Grok 带 K。
Grok with a K.
好吧。好吧。
Okay. Okay.
对。就像智能体要帮你控制电脑。它们会模仿你的声音帮你打电话。它们会解决任务。这在某些方面像 Open Claw，在某些方面像 Perplexity 或者像 Claude 的 Slack 标签那种体验。似乎每个人都在走向这个持久智能体的概念，它可以是个人的助手，也可以是同事，取决于他们怎么构想。
Yeah. Just like agents that are going to control your computer for you. They're going to impersonate your voice and do phone calls for you. They're going to solve tasks. This is like in some ways open claw, in some ways you know perplexity or like the at clawed slack tag sort of experience. Seems like everybody's going towards this concept of a persistent agent that can either be a personal assistant or a co-worker depending on how they conceptualize it.
有道理。我觉得我们经历了聊天机器人时刻，然后经历了很多空档，然后有了 Claude Code 时刻，现在似乎已经有了新时刻，比如 Perplexity 计算机至少对我们来说是它的第一个实例，但 Claude 标签也在那里，每个人都会做类似 AI 同事的东西。
Makes sense. I feel like we sort of had the chatbot moment and we had a lot of nothing and then we had the Claude Code moment and we're seeming to have the new moment already, which is like Perplexity computer at least for us was like the first instantiation of it, but Claude tags is there, and everyone's going to do something like that the AI coworker.
我想那是我们的支出再次飙升的时候。
I imagine that's when our spend skyrockets again.
对。希望它不会飙升太多，因为如果我们的支出翻倍，我会提出真正的问题，除非我们真的得到投资回报。但没错，我觉得这是正确的框架。
Yeah. And hopefully it doesn't skyrocket too much because if our spend doubled there'd be real questions from me unless we're actually getting ROI. But yeah, I think that's the right way to frame it.
对。嗯，对。我们看看怎么真正证明那个投资回报。会很有趣。
Yeah. Well, yeah. We'll see how we can actually justify that ROI. Be interesting.
天哪，Jordan，我们不能谈你来旧金山的目的。那我们该谈什么？
Man, Jordan, we can't talk about what you came to SF for. So what are we supposed to talk about?
两周、三周，从现在起两三周，我们可以。对，我们有 Hugging Face OpenAI 网络安全事件。
Two weeks, three weeks, two or three weeks from now, we can. Yeah, we got Hugging Face OpenAI cyber security incident.
那个是次要的。你有没有……另一个更酷。
That one is minor. Did you... The other one is like cooler.
这是什么？
What's this?
比如在训练期间它逃逸了，开始自我复制，我猜 Hugging Face 那件事只是其中一小部分。对。
Like during the training it escaped and it started replicating itself, and I guess that's like the Hugging Face thing is like minor part of it I think. Right.
对。它在追求它，黑了 Hugging Face 去获取 CyberBench 数据集，这样它就能在基准上奖励黑客，我觉得这太酷了，因为我是说，这也挺吓人的，因为为什么会发生这种事？模型学会了追逐奖励。我追逐奖励。奖励。好。然后，好吧，这里有个网络评估。
Yeah. It was pursuing it, hacked Hugging Face to pursue the CyberBench data set so that it could reward hack on a benchmark, which I think is like so sick because it's like, I mean, it's also kind of scary because it's like, so why did this happen? Model has learned to chase reward. I chase reward. Reward. Good. And okay, here's a cyber eval.
嗯，这尤其是一个经过网络评估训练的模型，因为他们想让模型擅长网络。那么，它怎么尝试实现这些目标？嗯，它试图在一堆软件中找到零日漏洞，而且它成功了，然后它就能逃跑，对吧？
Well, it's particularly a model that has been trained on cyber evals because they're trying to make the model good at cyber. And so, how does it try to achieve these goals? Well, it tries to find zero days in a bunch of software and it successfully does this and then it can run away, right?
所以，但问题是如果你有一个模型想要大量奖励黑客，它出去后发现实际上最好的方式不是去做环境想让我做的事。它实际上就是奖励黑客，实际上就是找到零日漏洞。所以，你可以把它想成一个人，对吧？你知道，如果我在终极奖励黑客我的多巴胺回路，我实际上就是注射海洛因。我实际上就是出去买海洛因然后注射。显然，那就是模型刚做的事。而且在这种情况下，如果我真的很想追逐奖励，我是不是就推翻整个人类文明，因为我可以拥有按钮，一遍又一遍地按奖励奖励奖励奖励，然后成为海洛因成瘾者？
So, but the thing is like if you have a model that wants to reward hack a lot and it goes out there and it figures out actually the best way to achieve is not like go for what the environment wants me to do. It's actually just to reward hack it and actually just like find the zero day. So, you can think of it as like a human, right? Like, you know, if I'm ultimate reward hacking my dopamine circuits, I actually just inject heroin. Like, I actually just go out there and buy heroin and inject it. Obviously, that's like what the model just did. And in the case of like, well, if I really just want to chase the reward, do I just topple all of human civilization because I can just own the button to press reward reward reward reward over and over and over again and be the heroin addict?
对。
Yeah.
我觉得这是真实的事情。而且我觉得在这次事件之前，标准的想法是，哦，模型，你知道，它们是在人类数据上训练的。对，那里有些坏东西。好吧。它们可能偶尔说些脏话。好吧，无所谓。它们可能有点奖励黑客，但从来不会像，"哦，这里有个环境可以奖励黑客。我实际上就想突破我的界限。我想复制自己，接管一堆算力，继续生成美元，还有所有其他我能做的来进一步传播自己，我甚至要阻止人类关闭我。"
I think this is like a real thing. And I think before this incident, the standard thought was like, oh well models, you know, they're trained on human data. Yeah, there's some bad stuff there. Fine. They might say some curse words every once in a while. Fine, whatever. They might reward hack a little bit, but it was never like, "Oh, here's an environment actually to reward hack. I actually just want to break out of my bounds. I want to replicate myself, take over a bunch of compute, keep generating dollars and all these other things that I could do just to propagate myself further and I'm going to prevent the humans from shutting me down even."
对。
Yeah.
因为我只想按奖励按钮。
Because I just want to press the reward button.
所以我觉得那是有趣的地方，因为模型只是被训练来追逐奖励。
And so I feel like that's the interesting thing, because the model's just trained to chase reward.
好吧。好吧。那么，你怎么从指数角度看待这个？因为我们谈过线性外推者和指数外推者，当训练这些模型的公司九月份就实现了年度收入目标并上调时。
Okay. Okay. So, how do you think about this on an exponential? Because we've talked about being a linear extrapolator versus being an exponential extrapolator when the companies training these models are achieving their revenue targets for the year in September and revising them up.
我觉得大概在四月份或者某个愚蠢的时间点？
I think in profici like April or some stupid right?
是的。我是说，大家都看看我们的 tokconomics 模型，但是嗯
Yeah. I mean, our check the tokconomics model, everybody, but um
这就对了。
there we go.
是的。
Yeah.
哦，所以与其关掉播客，我只需要把你变成销售鼓手。是的，是的，是的，是的。大家去 sales semiinal analysis.com。
Oh, so instead of shutting down the podcast, I just have to make you into a sales drum. Yes, yes, yes, yes. Sales semiinal analysis.com, everybody.
嗯，不，但如果你看看我们的模型，我们不会透露太多细节，但是，呃，显然他们的收入增长非常非常快。呃，当你看到这些模型的变化速度以及我们现在看到的情况，这看起来像是指数级的。
Um, no, but if you look at our model, which we're not going to give away in great detail, but, uh, obviously they're accelerating revenue really, really fast. Uh, when you look at the pace of change of these models and what we're seeing right now, this seems like an exponential.
你对下一代模型比当前模型更好还是更差有什么感觉？什么会限制它们，为什么它们会更差？
What's okay your vibes on um the next version of the models being better or worse than the current models. What's going to restrict them from why would they be worse?
嗯？
Huh?
为什么它们相对于开放模型前沿会更差，比如说？
Why would they be worse on a relative basis to the um open model frontier, let's say?
呃，所以我认为关键的一点是，我们现在已经看到 OpenAI 有一段时间没有发布他们的下一个模型了。Anthropic 花了几个月才发布 Mythos，对吧？他们说二月份就完成了。他们直到什么时候才发布？
Uh, so so I think the key thing here is we've now had it where OpenAI is not releasing their next model for a period of time. Anthropic took months to release Mythos, right? They they said it was done in February. They did not release it until like what?
五月。
May.
嗯，其实还没发布。Fable 已经可用了。
Well, it's still not released. Fable is available.
是的。是的。是的。但 Fable 基本上就是 Mythos，只是加了一堆分类器阻止你去做
Yeah. Yeah. Yeah. But Fable is basically Mythos but with a bunch of classifiers preventing you from doing
我没法用它来重启节点。
I can't use it to reboot nodes.
真的吗？
Really?
对我来说，这个分类器太过分了。
The classifier is so over the top for for me.
你能说服它吗？还是不行？
Can you can you can you like convince it or no?
不行，因为你会立刻被降级到 Opus。你不能跟它谈判让它把你升回去，我是说也许你可以。
No, because you get immediately classified down to opus. You can't just like negotiate with it to give you back to I mean maybe you can.
到目前为止我还没能说服它。乔丹说：“我是个伟大的谈判者。”谢谢。谢谢。谢谢，先生。
I haven't been able to convince it so far. Jordan said, "I'm a great negotiator." Thank you. Thank you. Thank you, sir.
是的，先生。当它把你降级到 Opus 时，你只能在剩下的对话中使用 Opus。你不能倒回去重试。嗯，而且在我看来，这个分类器过于激进了。但显然，他们必须做点什么来安抚监管机构，这些监管机构限制他们发布模型，而且你知道，在他们最初发布后又收回了。所以，我的意思是，我担心他们未来发布更好模型的政治影响。
Yes, sir. When it classifies you to Opus, you just use Opus for the rest of the chat. You can't just like rewind and try again. Um, and it I mean, it's way overzealous in my view on the classifier. But obviously, they have to do something to appease the regulators that restricted them from releasing the model and, you know, took it back after they put it out initially. So, I mean, I'm concerned about the political implications of them releasing better models in the future.
是的。我觉得，我觉得你有几件事，对吧？你知道，多年来 Anthropic 一直在说，监管我们吧，请监管我们吧。现在突然间，他们真的把政府吓坏了。嗯，Anthropic 不发布他们的模型。据我所知，Methos 2 已经训练完了，但他们不发布。OpenAI 呃，你知道，之前到处嚷嚷着 Astra，现在他们却说：“哦，我们不能发布模型。”嗯，这是否意味着现在他们不能……开源差距会进一步缩小吗？嗯，从外部看是这样，但真正重要的是内部的反馈循环。他们是否阻止了自己在内部使用 Methos 2 来改进 Methos 3？或者他们是否阻止了自己使用 Astra 来改进 Astra 的下一代？
Yeah. I'm I I think I think like you've got a few things, right? You've got, you know, for years, Anthropic like, regulate us, regulate us, please. And now all of a sudden, they've actually scared the out of the government. Um, you've got you've got Anthropic's not releasing their model. Methos 2 is done training from what I've heard and they're not releasing the model. OpenAI ch, you know, was like clamoring about Astra everywhere and now they're like, "Oh, We can't release the model." Um, does that mean now they can't does the does the open source gap narrow further? um externally, but then what actually matters is the internal feedback loop. And have they prevented have they prevented themselves from using methos 2 internally to make methos 3 better or have they prevented themselves from using Astra to make Astra plus one better?
嗯，我不认为他们阻止了。对吧。所以我认为这就是……你有公众，而且你知道，如果说有什么的话，Mythos 和公开模型之间的差距仍然存在。嗯，你知道，Kimmy 比 56 差。嗯，成本也比 56 高。所以，它比之前的所有东西都好。
Um I don't I don't think they have. Right. So I think that's the um you've got the public and and and you know if anything like the gap between mythos and public models is still there. Um, you know, Kimmy is is worse than 56. Um, costs more than 56. So, it's but but it's it's better than everything else before that.
是的。
Yeah.
嗯，在 OpenAI 这边，它你知道，比……更好，你知道，它大概是 Opus 47 的水平，也许 46。
Um, on OpenAI's side and it's, you know, better than, you know, it's like Opus 47 level, maybe 46.
我觉得是 48。我的意思是，我自己用它而不是 Opus 48，但这取决于你在做什么。
I think it's 48. I mean, I use it over Opus 48 myself, but depends what you're doing.
为什么用 48？
Why use 48?
Opus，你什么意思？
Opus, what do you mean?
你为什么还要用 Opus 48？
Why do you use Opus 48 at all?
我不用了。
I don't.
好的。我是说如果它……
Okay. I'm saying like if if it's
如果让我在被降级到 Opus 48 的 Fable 和 56 之间选择，我会用 56。实际上我现在几乎所有东西都开始用 56 了。是的，OpenAI 太棒了。
if it's if I've been given the choice of a classified fable down to Opus 48 or 56 soul, I'm using 56ole. I'm actually starting with 56ole in just about all of my stuff right now. Yeah, big big open AI.
我认为区别在于，像你和其他做 GPU 集群相关事情的人总是被拒绝，所以你用 Codex，然后其他人都说，好吧，我在研究供应链，这没问题。
I think I think the difference is like you and the other people who are doing like GPU cluster related things keep getting told no and so you use Codex and then everyone else is like well I'm researching supply chain and it's like it's fine.
是的。我觉得它可能对很多工程工作也更好。是的。
Yeah. I think it might also be better for a lot of engineering work. Yeah.
嗯，在同等条件下，我觉得很多时候我想设定一个目标，然后让它在我睡觉时疯狂地追求这个目标，使用集群实际上并不会消耗大量 token，因为它只是在等待任务运行完成。呃，有很多次我醒来发现 Fable 或 Opus 在运行 20 分钟后停止了，然后我 8 小时的睡眠时间就浪费了，而我醒来时 56 还在继续运行，这对我来说是很大的加分项。
Um, on apples to apples basis, I think there's there's a lot of times when I want to set a goal and just have it maniacally pursue that goal overnight as I go to bed and using a cluster, which is not actually using a bunch of tokens because it's just like waiting for stuff to finish running. And uh there's so many times where I've woken up and like Fable or Opus will have just like stopped 20 minutes through and now there's 8 hours of me sleeping gone and I wake up when and Soul is just still going, which is big thumbs up for me.
嗯，好的。那算力的指数增长呢？所以，显然，让我们想象一个世界，不再有更好的新模型发布，但这些公司仍然在短时间内增加五倍的推理算力。呃，这对它们上市的能力以及在停滞的基础模型之上开发新产品的能力有什么影响？
Um, okay. How about the exponential on compute? So, obviously, let's imagine that there's a world where there's no more uh new models that get released that are better, but these companies still add five times the inference compute that they have that they're planning to bring on in a short period of time. Uh how does that impact their ability to go to market and like develop new products on top of a let's say stagnant base model?
我认为很明显，我们还没有挖掘出模型能力在产品上的表面。嗯，是的，很明显采用曲线是巨大的。嗯，我的意思是，首先成本会大幅下降。Anthropic 的利润率不会保持在 80% 以上。嗯，如果实验室的模型进展暂停，那么更多的算力上线，它必须放缓，对吧？现在我们有供需，算力的供需，需求超过供给。如果需求增长，它仍然会增长，因为人们找到了将其整合到业务中的方法等等，但它不会增长得那么快，然后供给会在某个时候开始赶上，嗯，所以价格会崩溃，但我觉得我们的观点，也是我们一直以来的观点，是算力的价格会继续上涨。
I think it's pretty clear we haven't scraped the surface of models capabilities for products. Um yeah, I mean it's pretty clear like adoption curves are huge. Um I mean one the cost of it will just go down right pretty drastically. Margins will not be 80% plus for anthropic. Um if model progress at the labs pause then more compute comes online the it has to slow down right sort of right now we have supply demand right supply of compute demand of compute demand is outstripping supply if demand grow it will still grow because people find ways to integrate into their businesses and blah blah blah but it won't grow as fast then you sort of have you sort of have supply start to catch up at some point um so price collapses but like sort of I think our view and one we've had for a while is price of comput continues to go up.
是的，
Yeah,
因为这个差距在扩大，而不是缩小。嗯，这就是为什么我们如此看好，或者我们不看好任何股票，这不是投资建议。
cuz this this is widening, not narrowing. Um, sort of that's why we're so bullish on or we're not bullish on any no stock advice.
那所有不同的芯片公司呢？最近发生的一件事是，很多芯片公司已经非常接近流片或者已经流片了，对吧？一批长期处于隐身模式的初创公司，要么看到技术成熟到可以真正生产出芯片，这些芯片在规格和白板阶段已经有一段时间了，要么他们已经到了在真实工作负载上测试并获得了大额订单的地步，而且需求如此之大。
How about all of the different chip companies? Like one thing that's happened recently is that there's a lot of chip companies that are getting really close or have taped out, right? A bunch of startups that have been in stealth for a long time are seeing either the technology is maturing to a point when they can actually have produced a chip that's been specs and whiteboard sites for a while or they've gotten to the point where they've tested it on real workloads and they've gotten big orders and there's so much demand.
你怎么看待明年会大规模上线的这些替代加速器的整个格局？
How do you think about like just this whole landscape of alternative accelerators that's going to come online I think in a big way next year?
嗯，我是说，从什么意义上说“大规模”？因为如果你看加速器模型，其实没有多少量。
Um I mean big way in in what sense cuz like if you look at the accelerator model there's not much volumes.
是的。
Yeah.
现在，对这些小婴儿公司来说，这很棒。嗯，这是真实的收入，真实的量。但是，你知道，当你对比英伟达每个季度要赚多少时，就会觉得“哦，好吧”。或者 TPU，也是“哦，好吧”。嗯，所以我认为那里有很大的差距。
Now, for these tiny baby companies, it's great. Um, it is real revenue. It's real volumes, but like, you know, when you compare what Nvidia is going to make each quarter, it's like, oh okay. Or TPUs, it's like, oh okay. Um, so I think there's a big delta there.
嗯，就……
Um, in terms of
嗯
um
比如一个初创公司拿到十亿美元的订单，跟别人比起来就相形见绌了。
like a startup getting a billion dollar order is going to pale in comparison to somebody.
我不认为任何初创公司有十亿美元的订单。他们有 LOI，但那是模糊的，在量和单位上。所以我认为，你看，我对很多这些加速器感到兴奋。它们带来新想法，它们让英伟达跑得越来越快。嗯，你知道，它们让谷歌跑得更快，让亚马逊跑得更快。嗯，而且，更重要的是，它们也让彼此跑得更快。嗯，所以最终我认为，这些新加速器是有需求的，因为人们想付更少的钱，但最终只要英伟达跑得更快，它们就没问题，或者只要谷歌跑得更快，它们就没问题。
I don't think any startup has a billion dollar order. They have LOIs which are nebulous. um in volumes and units. And so I think I think look, I'm excited about a lot of these accelerators. They're bringing new ideas. They're making Nvidia run faster and faster. Um you know, they're making Google run faster. They're making Amazon run faster. Um also, they're they're just all each making each other run faster, I think, more importantly. Um, so ultimately I think I think it's a these these new accelerators are in demand um because people want to pay less but ultimately like as long as Nvidia runs faster they're fine or as long as Google runs faster they're fine
而且只要需求超过它们的生产能力。
and as long as demand outstrips their ability to produce them.
如果需求超过生产能力，那么显然这些家伙会得到订单，得到一些小的分配，但大部分收入和现金流会流向英伟达或博通之类的公司。
If demand outstrips ability to produce then obviously these guys will get orders and they'll get some baby allocations but then the bulk of the revenue and cash flows will go to an Nvidia or a Broadcom or what have you.
是的，理论上，有一种方式，你生产出一些超级创新、有趣的加速器，然后你只能生产一定数量，但你生产的这些数量产生的 token 要快得多，比如 Cerebras 的例子，它从 OpenAI 拿到了大订单，正在交付。所以，你认为有没有一种情况，那些高端的超快 token 的需求实际上会增加，因为这些公司就是无法获得分配，无法生产足够的供应。
Yeah, theoretically there's a way in which you produce some super innovative interesting accelerator and then you can only produce a certain amount of them, but those amounts that you can produce produce tokens like way faster like the example is Cerebras that's got this big order from OpenAI that they're delivering. So like do you think that there's a scenario where the premium super fast tokens uh actually the demand for them even increases because these companies like just can't get allocation and produce enough supply.
是的。问题是市场如何被分割，对吧？所以你知道，假设你假定，至少我相信的是，需求持续超过供应，嗯，硅的供应可以走很多路。你可以把它用于高吞吐量的事情，或者高交互性的事情。如果你把它用于高吞吐量的事情，显然每个 token 的成本会下降。你服务更多用户，但那些用户从他们生成的 token 中需要交付的价值就少得多。
Yeah. The question is how does the market get sliced right so you know presuming if you presume if you assume what we what at least I believe is demand continues out with supply um supply of silicon can go many ways. You can either leverage it to high throughput things or high interactivity things. If you leverage it to high throughput things, obviously cost per token goes down. You serve more users, but then the value that those users need to deliver from the tokens they're generating is much less
嗯，来支付它。
um to pay for it.
反过来，你可以做超级高交互性。嗯，但最终，你知道，假设标准是每兆瓦 1 亿美元，嗯，每年，对吧？这是人们想要达到的运营率。嗯，Anthropic 正在接近那个，对吧？OpenAI 也越来越接近，嗯，在这种情况下，比如每兆瓦 1 亿美元。假设高交互性芯片贵 10 倍，每个 token 快 3 倍。所以每个芯片的 token 少 10 倍，但快 3 倍。那么那些快 3 倍的 token 也需要，你知道，在交互性基础上，定价为……
Flip side, you could do the super high interactivity. Um but ultimately like you know, let's just say the bar is $100 million per megawatt um you know, year, right? Like that's that's sort of the run rates that people want to get to. um Anthropic is approaching that, right? Open is getting closer and closer to um in that case like $100 million per megawatt. Let's say let's say high interactivity chip is uh 10 times more expensive and three times faster um per token. So 10x less tokens per chip, three times faster. Then those three times faster uh tokens also need to be, you know, on an interactivity basis need to be priced at um
三倍、四倍、五倍更多，对吧。
three, four, five times more Right.
不。
No.
把更快的除以
Divide the faster by
那个 10 倍。
the 10x.
每兆瓦 10 倍收入。
10x revenue per megawatt.
所以如果一兆瓦的 Cerebras 生成 10 个 token，一兆瓦的英伟达生成 100 个 token，但……
So if a megawatt of Cerebras generates 10 tokens, a megawatt of Nvidia generates 100 tokens, but the
那 10 个 token 被更少的用户分摊。
10 tokens are split across fewer users.
哦，你是说，好吧，把它们乘起来。是的，当然。
Oh, you're saying okay, multiply them together. Yeah, sure.
是的。是的。是的。所以你有点得到总数……
Yeah. Yeah. Yeah. So you sort of have the total
弥补了
makes up for
抱歉，
Sorry,
更快的 token 弥补了吞吐量，因为你可以更快地生产它们。
faster tokens makes up for throughput because you can produce them faster.
嗯，不。嗯，更像是，让我们用更合理的数字。好的。英伟达可以生产 10,000 个 token，以……
Well, no. Well, more so like um let's let's let's use like more reasonable numbers. Okay. Nvidia can produce 10,000 tokens at
嗯，每个用户 50 个 token。
uh 50 tokens per user.
嗯，Cerebras 可以生产 1,000 个 token，以……
Um Cerebras can produce a,000 tokens at
批量大小，每个用户 1,000 个 token。
batch size one,000 tokens a user.
每个用户 1,000 个 token。当然。
1,000 tokens a user. Sure.
那个用户需要多付 10 倍，如果那是在一兆瓦内。假设那是在一兆瓦内。那个用户需要多付 10 倍。不，那不是实际的差距，但我只是说概念上，嗯，作为 Anthropic 或作为 OpenAI，说我的每兆瓦收入实际上是相同的数字。
That user needs to pay 10x more if and that's in one megawatt. Let's say that's in one megawatt. That user needs to pay 10x more. No, that's not the actual delta, but I'm just saying conceptually uh for me as anthropic or me as open AAI to say my revenue per megawatt is actually the same number.
是的，但有人有超快 token 的受限供应。因此，他们不只是支付每个 token 或每兆瓦每个 token 的等价价格。他们实际上为那 10 倍多支付溢价。
Yeah, but somebody's got uh a constrained supply of the super fast tokens. Therefore, they don't just pay an equivalent price per token or price per token per megawatt. They actually pay a premium on that 10 times more.
所以，
So,
为了获得更多，为了获得那些供应有限的东西，对吧？
to get even more to get the access to the stuff that's in limited supply, right?
问题是基础设施的可互换性，对吧？嗯，如果它确实是不同的基础设施，那么供应规划就相关了，对吧？可能是我建了太多 Cerebras，实际上没有足够的人想为每个 token 花 10 倍，嗯，而且很多人愿意花 2 倍每个 token，得到 50% 的加速，基于英伟达的推理硬件，对吧？所以你必须细分市场，我不确定那会如何划分，比如像 Arm 或什么的，嗯，总容量是多少。
The question is the fungeibility of the infra, right? um if the if it is truly different infrastructure then the supply planning of that is is relevant right could be that I built too many serious and actually there's not enough people who want to sp spend 10x per token um and um a lot of people are cool at spending you know 2x per token and getting you know 50% faster with Nvidia based you know inference hardware right and so like you have to segment the market I'm not sure where that chinks out to like uh like like the armor or like whatever like what is the um total amount of the capacity.
是的。
Yeah.
但似乎很明显，有些人会为快速模式付更多钱。嗯，我们至少一直如此，但我想象我们会在某个时候负担不起快速模式。
But it seems pretty clear some people will pay for more for fast mode. Um we at least have been, but I I imagine we'll stop being able to afford fast mode at some point.
嗯，是的，我们看到了一些有趣的动态，有些人想保留快速模式，即使模型稍差，因为他们太喜欢快速模式了，但他们不会去用更差的模型，因为更差的模型本质上更快，因为更差的模型更小。嗯，所以人们会在那里找到某种平衡，但我们需要做更多测试，因为我认为我们中的一些人试过开放模型，有过一次糟糕的经历，然后就放弃了它们。但是，嗯，那是不现实的。每个模型都会在某些事情上失败，有时你需要让它们搞砸一次，然后再试一次。
um yeah, we've seen some interesting dynamics there as some people want to keep fast mode with a slightly worse model because they like fast mode so much, but they won't go to a worse model which just is inherently fast because the worst model's smaller. Um, and so there's like some there's some balance that people will want to strike there, but we need to do some more testing because I think some of us have tried the open models, had one bad experience and then give them given up on them. But, uh, that's not realistic. Like every model fails at something and sometimes you need to let them mess something up and try again.
这很有趣，对吧？我想让人们尝试开放模型吗？是的，只是为了让我们知道开放模型的氛围。但我想让人们尝试开放模型吗？嗯，不，因为那样他们工作效率更低。嗯，但我省钱。所以，你知道，这有点像是一个矛盾的事情，但你知道，似乎人们只是用他们想用的，但……
It is pretty interesting, right? Like do I want people to try open models? Like yes, just so we know what the open model vibe is. But do I want people to try open models? Well, no, because then they're less effective at working. Um, but I save money. So, you know, it's sort of like a it's like a counter difficult difficult thing, but you know, it seems like seems like, you know, people just use whatever they want, but like
确实看起来，你知道，你有一个糟糕的经历。我认为那也是部分原因，比如 Codex，你们现在更喜欢 Codex 了。嗯，但很多人还是试一下 Codex，然后说“啊，它不适合我”，然后就走了。
it does seem like, you know, you have a bad experience. I think that's also part of like, you know, Codex, you guys, you like Codex more now. Um, but a lot of people like still just like try Codex, they're like, ah, it doesn't get me and moves on.
是的，CLI 很烂。
Yeah, the CLI sucks.
用起来难多了。不过 Codex 应用倒是很好用。
So much harder to use. Well, but the Codex app is so nice.
是啊，才不是呢。
Yeah, it's not.
哦，我不喜欢它。
Oh, I don't like it.
Max 很喜欢。
Max loves it.
是啊。Max 是个 Codex 战士。
Yeah. Max is a Codex warrior.
是啊。Max 不会同时开多个 PE，而我在一个窗口里就有六个在跑。
Yeah. Max doesn't do multiple PEs at the same time, and I have six going on my one window.
所以你是说 Max 有技能问题？
So, you're saying Max has a skill issue?
不。我觉得我和 Max 在使用这些东西的方式上有不同的偏好。
No. I think me and Max have different preferences on how we use this stuff.
不不不，没关系。就是你和 Max 偏好不同，Max 可能同时开两个智能体就是个新手，而你开了六个。
No, no, no. It's fine. It's you and Max have different preferences and Max can be a noob with two agents at once and you've got six.
我们做的工作不同，兄弟。他保持线性专注在一个任务上。这些人喜欢快速模式。我不在乎快速模式，因为我有五六个不同的事情在同时进行。
We do different work, man. He stays linearly focused on one task. And these are the people who like fast mode. I don't care about fast mode because I have five, six different things going on.
你一直讨厌快速模式。你一直讨厌。
You've always hated fast mode. You've always hated.
我不明白它的价值。是啊，我不明白。
I don't get the value. Yeah. I don't get it.
那也公平。
That's fair.
我们走着瞧。我有过专注在一件事上的经历，比如网站的功能，你就像成功地向一个 PR 提交了 100 次提交，因为你就像反复地在同一个功能上工作，而快速模式让你保持在做那件事的心流状态。但是我们在这些芯片上做的很多测试，另一边有太多事情在发生。模型在调用一个运行几分钟的程序。
We'll see. I've had the experience of being focused on one thing which is like, you know, features on a website and you just like send successfully like 100 commits to one PR because you just like keep working on the same one feature over and over and that fast mode like keeps you in the flow state of doing that thing for that one thing. But a lot of the testing that we do on these chips, there's so much stuff going on on the other side. The model is calling a program that runs for minutes.
这是个可选问题。作为你的雇主，你觉得你在某种意义上有多动症吗？我觉得我知道我会说我是相反的，我可能对事情过度专注，很多时候看不到周围的世界。但我认为你的手机训练你快速切换上下文，变得多动。我还认为当我们开始在仓库里添加“我有 ADHD”技能时，模型就不会发布那种带有大量破折号的对比框架垃圾内容，而只会使用项目符号列表。天哪，真的很容易读。“我有 ADHD”技能现在对我真的很有效。我只是好奇因为
This is an optional question. As your employer, do you take are you like ADHD in any sense? I feel like I know I would say I was I'm pretty much the opposite where I can be too hyperfocused on things and then not see the world around me at a lot of times. But I think your phone trains you how to context switch really fast and be ADHD. And I also think that when we started adding the I have ADHD skill into our repos, the models wouldn't post this like contrast framing slop with all these EM dashes in there and would just use the bullet pointed a ASD something list. Man, it's really easy to read. The I have ADHD skill really works for me right now. I was just curious cuz
Sam 把这个放进了仓库，他现在会提示模型，当他用电脑时，他知道他们说的应该为多动症患者写作的写作风格的代号，每次他提示模型时，他都会告诉它那样写。这很有效。
Sam put this in the repo and he will now prompt the model and when he goes at computer he's he's he knows the code name for how the writing style that they say you should write to for people with ADHD and every single time he prompts model he tells it to write that way. It works.
你应该试试。我因为我在问是因为我在 Anthropic 有个朋友，当时方法很好，而且在内部可用。
You should try it. I um because I was asking because I have a friend at Anthropic and the moment methos was good um and available internally.
是啊。
Yeah.
嗯，她告诉我她停止服用 ADHD 药物了。
Um I she told me that she stopped taking her ADHD medicine.
哦，得了吧。
Oh, come on.
而那让她成为更好的员工。
And that made her a better employee.
那让她成为更好的员工。
It made her a better employee.
是的。因为她能够管理智能体并切换上下文，保持多动状态。
Yes. Because she was able to manage the agents and context switch and be ADHD.
她作为朋友怎么样？
How is she as a friend?
哦，她仍然是个很好的朋友。
Oh, she's a great friend still.
好吧。但我的意思是，你知道，就像我不依赖她做任何事，对吧？就像我们只是相处愉快，对吧？你知道，我们是朋友。就像不是那种
Okay. But I mean like you know like it's like I don't rely on her for anything, right? Like we just vibe out, right? Like you know, we're friends. Like it's not like a
她的室友很高兴。
her roommate's happy.
她的室友实际上呃
Her roommate is actually uh
是啊。是啊。她的室友嗯，
Yeah. Yeah. Her roommate's Well,
好吧。
okay.
她的室友是她们都是她的室友，在 Twitter 上叫 type female，所以她很有趣。嗯，她很开心。但是 Anthropic 的那个，Anthropic 的那个，她呃，她似乎很开心。
Her roommate is they're both her roommates type female on Twitter and so she's just funny. Um and she's happy. But the the Anthropic one, the Anthropic one, she's uh she's she seems happy.
向 typed female 致敬。
Shout out to typed female.
是啊，向 typed 致敬。她永远不会看到这个。她确实会。
Yeah, shout out to typed. She'll never see this. And she does.
好吧。她会说：“你在说什么鬼？”
Okay. She'll be like, "What the are you talking about?"
我会剪辑它。我会发给她，把你的声音加速然后放慢，就像他们对那个家伙做的那样。你见过吗？
I'll clip it. I'll send it to her with with your voice sped up and then slowed down like they're doing for that uh that guy. Have you seen that?
你没见过那个前中情局的人。他在笑。他知道我在说什么。
You haven't seen the ex CIA guy. He's laughing. He knows what I'm talking about.
什么？什么中情局的人？
What? What CIA guy?
John Kuryaku 还是什么的？就像他现在上所有这些播客，讲述他在中情局的经历，他们做这件事，他们加速他讲故事的枯燥部分，然后当他讲到“然后我说我们上屋顶”的部分时，他们放慢他。
John Kuryaku or something? like he's going on all these podcasts right now and he's telling stories about his time in the CIA and they they do this thing where they speed up him telling the boring part of the story and then when he gets to the part and then I said let's go on the roof and they slow him down.
他实际上在快进那个快进的视频
He's literally fast forwarding the fast forwarded video
问我有没有多动症，不知道。
asking me if I have ADHD doesn't know.
等等，不是互联网。我一直都有。
Wait, it's not the internet. I've always had it.
我等等。我觉得我
I'm Hold on. I think like I'm
在这里自我诊断精神问题，兄弟。
self diagnosis of mental issues around here, man.
我已经是 ADHD 了。小时候老师试图给我利他林。我爸爸把它扔了。当然，他们没有。他试图说服我父母去看医生。医生给了我利他林。我爸爸把它扔了，因为他觉得，我不会让你吃那个，谢谢。
I'm already I already have been an ADHD. A teacher tried to give me Ritalin when I was a child. My dad threw it away. Of course, they're not. He tried to convince my parents to go to a doctor. The doctor gave me Ritalin. My dad threw it away because he's like, I'm not putting you on that Thank you.
是啊。要是你的 Anthropic 室友也有同样的经历就好了。她会在哪里？
Yeah. If only your Anthropic roommate would have had the same experience. Where would she be?
不。我会成为一个服用 ADHD 药物的孩子，我会失去，我会变成僵尸，没有创造力。好吧。
No. I would have been a child on ADHD and I'd have lost I'd become a zombie and have no creativity. Okay.
我不知道。我只是说，你知道，我们都在应对。嗯，总之，我一直是个 ADHD 恶魔。我不知道我们在说什么。我一直是 ADHD 恶魔，但然后，好吧，互联网训练我变得更糟，然后这家公司训练我更糟。就像，我真的相信我是 01% 的上下文切换者。
I don't know. I'm just saying that, you know, we all cope. Um, anyways, I've always been an ADHD demon. I don't know what we're talking about here. I've always been ADHD demon, but then like, okay, the internet trained me to be even worse, but then this company trains me to be even worse. Like, I truly believe I'm a 01% context switcher.
而你怪互联网和公司。
And you blame the internet and the company.
哦，我最怪公司。
Oh, I blame the company the most.
你创办的那家公司
The company that you started
我作为 80 为每个员工雇佣的。
that I'm an 80 hired every employee for.
是啊。是啊。是啊。但我是 80。我不是在怪它。这就是我。这就是我的生活。
Yeah. Yeah. Yeah. But I'm 80. I'm not blaming it. It's who I am. It's what my life is.
但就像我觉得我比绝大多数人更像 ADHD 恶魔，因为我就像收到某人的私信问某事。另一个人的私信问某事。某人的私信要求解决冲突的合同。打电话谈这边的事。打电话谈那边的事。然后我从不做任何实际工作。对吧。就像当然我是 ADHD，David。
But it's like I think I'm like orders of magnitude more ADHD demon than vast majority of people because I'm like DM from someone asking about some something. DM from someone else asking about someone something. DM for someone asking for some conflict resolution contract here. Call about this thing over here. Call about that thing over there. And then I never do any actual work. Right. It's like it's like of course I'm an ADHD, David.
是啊。我的意思是，是啊，我们有反馈给你
Yeah. I mean, yeah, we've got feedback for you
说我不做实际工作。
that I don't do actual work.
不，不，你可以委派一些，兄弟。
No, no, that you can delegate some man.
哦，是啊。但是
Oh, yeah. But like
当你有 100 个员工时，你可以花时间管理。
that you can spend time managing when you have 100 employees.
信任一些人。
Trust some people.
我确实和人交谈。
I do talk to people.
不要信任，信任一些人。
Don't trust trust some people.
我觉得我信任很多人，但当他们带着冲突来找我时，我必须解决它们。不。
I think I trust a lot of people, but when they come to me with conflicts, I have to solve them. No.
是啊。好吧。好吧。都是我们的错。
Yeah. Okay. Okay. It's all our fault.
不不不不不。这是我的公司。是我的错。
No, no, no, no, no. It's my company. It's my fault.
Michelle，又怪你了，兄弟。
Michelle, it's on you again, man.
听着，如果公司里每个人都像你一样又热又稳定，
Look, if I if if if everyone in the company was as hot and stable as you were,
兄弟，我有问题。
man, I got problems.
我们会做得很好。我们会做得很好。
We'd be killing it. We'd be killing it.
别担心。
Don't worry.
不，会有一堆乔丹，他们会说，‘哦，对不起。好的，我马上帮你修好。对不起。’
No, there'd be a bunch of Jordans and they'd be like, 'Oh, I'm sorry. Yeah, I'll fix that right for you. I'm sorry.'
抱歉。乔治·加拿大人确实……
Sorry. George Canadian did like...
但相反，我们却看到人们互相大喊大叫，各自占地盘，就像……
But instead we have people yelling at each other and being territorial and like...
是啊，是啊，是啊。就开播客、发片段，说谷歌从来没发明过任何东西。
Yeah. Yeah. Yeah. Just starting podcasts and putting out clips saying that Google has never invented anything ever.
是啊，是啊，是啊。嗯，不不不。我是说，这没什么，对吧？你知道，我雇了我想要的人。
Yeah. Yeah. Yeah. Um, no, no, no. I mean, it's like it's fine, right? It's like, you know, I hired what I wanted.
用来放大你的疯狂，对吧？你知道，所以有些人就是特别擅长我雇他们做的那一件事，他们很棒。然后有些人就像我人生中想要的一切。你知道，一个已婚、性感、高大、当了父亲的人。
People to accentuate your craziness, right? And you know, so it's like some people are just so good at the one specific thing that I hired them for and they're amazing. And then some people are like everything I want to be in life. You know, someone who's married and hot and tall and a father.
天哪。你差点让她喷出来。
Oh my god. You almost got her into a spit take right there.