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。
某种程度上，OpenAI 在消费端太成功了，反而成了他们的干扰。
To some extent, may have been a distraction for OpenAI that they were so successful in consumer.
如果这算是一种仁慈的了结，那苹果可能还帮了他们一把。
If this was a mercy killing, Apple may have done them a favor.
要是推特有什么动静，哈利，你肯定第一时间冲上去。
If there was action on threads, j Harry, you'd be there.
要是火星上有动静，哈利，你肯定也在那儿。
Action on Mars, Harry, you'd be there.
每个有点脑子的首席信息官，都会搞个便宜的开源模型来压一压这股疯劲。
Every company with a CIO who's half awake is gonna have a cheap token model to hand to stop this madness.
真糟心。
Shit.
我得给埃文找份工作。
I'm putting Evan out a job.
我最少得让他有口饭吃。
The least I can do is keep him alive.
叛国之举从来难成，可这到底是为什么？
Treason does not succeed, but what's the reason?
要是真成功了，谁还会管它叫叛国？
If it does succeed, no one calls it treason.
刚进董事会的时候，你不能同时投资两家竞争对手。
In the early stage when you're on the board, you can't invest in two competitors.
到了后期，从结构上讲，你非得这么做不可。
In the late stage, structurally, you have to.
要是你不想只值一个X，那就赶紧行动吧，别等来不及了。
If you don't wanna be worth one x, like, do something before it's too late, man.
这里是与我一起的20家风投，我是哈里·斯蒂宾斯。
This is 20 VC with me, Harry Stebbings.
这是你每周必听的节目。
It is your weekly listen.
本周科技圈发生的所有大事，这里一网打尽。
Everything you need to know that's gone down in tech this week.
今天议程的主讲人是罗里·奥德里斯科尔、杰森·兰普金，还有我。
Rory O'Driscoll, Jason Lampkin, and me are pining on the agenda for today.
第一，苹果起诉OpenAI窃取商业机密。
So number one, Apple sues OpenAI for trade secret theft.
第二，Meta在编码模型领域反击，扎克伯格也重新回到了X平台。
Number two, Meta fights back on the coding model front, and Zuckerberg is back on X.
第三，SK海力士以265亿美元的估值在纳斯达克上市，把计算和基础设施层带到了公众视野。
And then number three, SK Hynix prices a $26,500,000,000 NASDAQ listing, bringing the compute and infrastructure layer to the public domain.
但在进入今天的内容之前，你想过没有，有什么东西在商业领域的扩散速度跟AI一样快？
But before we dive into the show today, what's one thing in business that's spreading as fast as AI?
AI风险。
AI risk.
你们团队注册的每一个新工具，每个供应商开启的AI功能，每一次新集成，说实话，都可能埋下隐患。大多数安全体系根本跟不上AI这种爆发式的发展节奏。
Every new tool your team signs up for, every vendor that turns on AI features, every new integration, each one, I'm sorry to say, is an opportunity for something to go wrong, and most security programs weren't built for AI's pace of growth.
这时候，Vanta就能派上用场了。
Well, that's where Vanta comes in.
Vanta 是目前使用最广泛的智能信任平台，已有超过一万六千家快速发展的公司，比如 Ramp、Cursor、Harvey 等，都在用它确保随时都能通过审计。
Vanta is the number one agentic trust platform used by over 16,000 fast moving companies like Ramp, Cursor, Harvey, and more to ensure they're always audit ready.
现在，Vanta 也开始帮助像你这样的公司。
And now Vanta's helping companies like yours.
留意那些在你供应商、AI 工具和整个系统之间出现的风险。
Watch for the risks that show up between orders across your vendors, your AI tools, and your whole environment.
怎么做呢？
How?
Vanta 的智能代理就像一个 24 小时在线的 GRC 工程师，默默后台工作，发现隐患、起草解决方案，还能把供应商审核时间缩短多达一半。
Well, the Vanta agent works like a twenty four seven GRC engineer in the background, finding issues, drafting fixes, and cutting vendor agreements time by up to 50%.
无论你是快速增长的初创公司，还是全球性大企业，Vanta 都能帮你自动化安全与合规流程，建立并证明你的可信度。
Whether you're a fast growing startup or a global enterprise, Vanta's here to help you automate your security and your compliance and earn and prove trust.
我的听众可以享受专属优惠。
My listeners get a special offer.
当然了。
Oh, yes.
特别优惠。
A special offer.
在 vanta.com/20vc 可享受 1000 美元折扣。
$1,000 off Vanta at vanta.com/20vc.
访问 vanta.com/20vc，立减 1000 美元。
That's vanta.com/20vc for $1,000 off.
Vanta 负责合规，Deal 则帮你全球招聘。
While Vanta keeps compliance covered, Deal helps you hire globally.
创始人用 Deal 加速创业扩张。
Founders scale startups faster on Deal.
突破地域限制，持续成长。
Grow without borders.
Deal 处理全球招聘的复杂环节，让你专心发展业务。
Deal handles the hard parts of global hiring so you can stay focused on growth.
几分钟内即可为任何国家设置薪资发放。
Set up payroll for any country in minutes.
随时随地招聘，签证办理快速搞定。
Hire anyone, anywhere, and get visas handled fast.
Deal 一站式帮你搞定入职、人力、IT、雇主代管、福利和合规，所有初创公司快速扩张所需的服务，全都在一个平台快速完成。
Deal takes care of onboarding, HR, IT, EOR, benefits, and compliance, everything your startup needs to scale quickly, all done fast in one place.
正因如此，包括 Airwallex、Eleven Labs 和 Intercom 在内的四万多家快速增长的企业都信赖 Deal，让他们能加速推进，专心搞产品。
It's why more than 40,000 fast growing companies like Airwallix, Eleven Labs, and Intercom trust Deal to move fast and get back to building.
访问 deal.com/20vc。
Visit deal.com/20vc.
就是 deal.com/20vc。
That's deal,deel,.com/20vc.
Deal 负责组建团队，Framer 负责搭建网站。
While Deal builds the team, Framer builds the site.
如果你的团队想要一个看起来精致考究、但又能快速上线的网站，Framer 就是为你量身打造的。
If your team wants a website that looks and feels handcrafted but is still fast to ship, Framer is built for exactly that.
我特别喜欢这一点。
Here's what I love.
Framer 是专为创作者、团队和注重细节的企业打造的智能网站构建工具。
Framer is the pro AI website builder for creators, teams, and businesses that care enough to get every detail right.
AI助手能弥合创意构想与实际可用网站之间的差距，因为所有工作都在网站实际运行的环境中完成。
The agents close the gap between AI generated ideas and production ready website work because it all happens where the site actually lives.
内容直接落在画布上，随时可编辑，等团队准备好了就能一键发布。
It lands on the canvas, stays editable, and can be published when the team is ready.
你可以在这里自定义代码组件、创建和管理CMS内容、优化SEO设置，所有操作一站式搞定。
So you can build custom code components, create and manage CMS content, optimize SEO settings, and ship everything all in one place.
AI助手带来的是速度和规模。
The agents bring speed and scale.
而你负责品味、判断和掌控全局。
You bring taste, judgment, and control.
它同样适用于企业级需求。
It's an enterprise level solution too.
提供高级托管、企业级安全防护和99.99%的可用性保障，这就是为什么Perplexity、Miro等全球顶尖品牌都选择在Framer上构建网站。
Premium hosting, enterprise grade security, 99.99% uptime SLAs, which is why the world's leading brands like Perplexity and Miro build in Framer.
想了解更多如何让您的网站发挥更大价值？欢迎联系Framer专家，现在就免费开始创建，访问 framer.com/20vc，享Framer Pro年费方案30%折扣。
Learn how you can get more out of your site from a Framer specialist get started by building for free today at framer.com/20vc for 30% off a Framer Pro annual plan.
访问 framer.com/20vc，立享30%优惠。
That's framer.com/20vc for 30% off.
framer.com/20vc
Framer.com20vc.
具体规则和限制可能适用。
Rules and restrictions may apply.
您已到达目的地。
You have now arrived at your destination.
各位，能回来真是太好了。
Guys, it is so good to be back.
和往常一样，我们今天要聊的内容很多。
As always, there is a lot for us to discuss.
不如我们先聊聊苹果这家公司，它市值已突破万亿美元，而这次它又把另一家万亿美元公司——OpenAI——告上了法庭，指控对方窃取商业机密。
Why don't we start with the trillion dollar company that is Apple, as it brings in another trillion dollar company in the form of OpenAI, where Apple sues OpenAI for trade secret theft.
简单说，苹果一位干了二十四年的老员工，把机密信息带给了OpenAI，泄露了本不该外泄的内容。
Essentially, one of Apple's long term twenty four year vets was taking secrets to OpenAI and sharing information that he shouldn't have been.
这事儿非常严重，我们马上就能跟听众分享一些关键结论。
This is a really serious thing, and there's actually some conclusions that we can share with listeners in a second.
但就事实而言，这位二十四年的元老是唐唐，他在诉讼中被点名了；而实际被指控窃取机密的，是那位在苹果干了六年的员工程璐。
But on the facts, the twenty four year veteran is Tang Tang, and he's named in the suit, but the person who is alleged to have taken the secrets was the six year Apple employee, Cheng Lu.
唐唐在离开苹果后，成了OpenAI硬件部门的负责人，指控称他怂恿员工把苹果的机密带过去。
Tang Tang is in charge of the hardware division at OpenAI after having been a super senior executive at Apple, and the allegation for him is that he encouraged hires to bring this stuff over.
稍微有点不一样。
Just slightly different.
没错。
Absolutely.
他还鼓励员工把实际的硬件部件带过去，当面展示。
And he encouraged them to bring actual parts for show and tell.
正是如此。
Exactly.
说真的，我觉得关键在于，听好了，首先想给所有听众一个建议。
You know, I think the big picture here is in particular I mean, listen, let's start with the advice to anyone listening.
绝对别干这种事。
Don't ever do this.
那个直接把东西带过去的人，我今天早上看了诉状，41页，没全看完，但读了相当一大部分。
The person on point, the individual who brought that stuff across, I read the lawsuit this morning, 41 pages, didn't read it all, but got through a good slug of it.
里面有很多实实在在的证据表明，这个人确实干了这事，他完蛋了。
There's a lot of fact based evidence that the person who took this stuff did it, and he's screwed.
他最先会发现的是，打官司根本没人讲情义。
And the first thing he's going to discover is there's no gratitude in litigation.
你以为自己是在帮新公司招人，或者想讨好新东家，结果他们转头就会把你彻底出卖。
You think you're caught helping your new employee or trying to get a job, they're going to burn you so fast your head's going to turn.
他们还会装出一副震惊的样子，就像《卡萨布兰卡》里克劳德·雷恩斯演的那样：‘天啊，没想到这儿居然有人赌博！’——而你，会被彻底抛弃。
But they're going to say, like Claude Rains in Casablanca, I'm shocked and appalled to find that there's gambling going on here, and he's gonna be left high and dry.
找工作面试时，千万别干这种事，否则你肯定吃不了兜着走。
Don't ever, as someone interviewing for a job, ever do this because you're gonna be screwed.
这人完蛋了。
That guy's toast.
对。
Yeah.
我还真没想过这一点。
That's that I hadn't thought about that.
你以为自己是在帮公司。
You think you're helping your company.
他已经被炒了。
He's already lost his job.
别管他能拿多少补偿了。
Forget what his comp exit package is.
他早就走人了，对吧？
He's already gone, right?
你根本不是在帮公司。
You're not helping your company.
再往上就是你提到的那位先生，谭棠，他也有不少问题。
And then the next level up is the gentleman you mentioned, Tan Tang, who is he also has some issues.
情况没那么明确，因为没有确凿的证据，但也不是空穴来风，好几个人都说过，是他鼓励我这么做的。
It's not as clear because there's no fact based evidence, but it's not just hearsay because a bunch of folks have said he did it, he encouraged me to do this.
就连和张流的邮件往来里，也多次提到‘我们是被要求这么做的’。
And even in the email traffic with Chang Liu, there was a lot of, hey, we were asked to do this.
他现在面临的风险是，接下来会有大量的取证和调查。
So the risk he's up against is that in, there's now going to be a ton of deposition, a ton of discovery.
如果真有邮件证据显示他发过信息，或者怂恿别人在来OpenAI面试时带点前公司的资料，那他也完蛋了。
And if there's any smoking gun that has him emailing or encouraging people to say, hey, bring in information when you're coming in for your next interview here at OpenAI, then he's screwed too.
同样的事情也会发生在他的身上。
And the same thing will happen.
等你说到最高层，你看苹果公司，从某种角度说，他们其实挺高兴的，毕竟有四百个苹果员工跳槽去了OpenAI。
Now when you get to the highest level, look, Apple in one sense is quote unquote happy, because they're really pissed that 400 people from Apple have gone to move to OpenAI.
过去这么多年，苹果一向特别反感员工跳槽去竞争对手那里，我觉得他们甚至有点过分，总是施压其他公司别挖人。
And traditionally, the years, has always hated, in my view, almost to the point of not being appropriate, people leaving their place and going to someone else and really leaned on other companies to not hire.
对。
Right.
我觉得，就像我说的，这种类似十五年前史蒂夫·乔布斯那种事，确实有点过头了。
And I think, as I said, at some point that gets inappropriate to see this whole Steve Jobs kind of issue fifteen years ago.
但关键是，他们并不是在想：‘你想要点什么，对方犯了个错，这下我们抓到把柄了’这种逻辑。
But the bottom line is they're not looking at this and going, you know, way when you want to get something for something, and then the other guy makes a foot fault and there now we have them.
他们知道自己手里有那个基层员工，很可能也掌握了工程副总裁的把柄，接下来就会步步紧逼，把所有相关人等都叫来录口供，看看这事到底牵扯多深。
So they know they have the junior guy, they probably have the VP of engineering and they're going to twist the knife, depo everyone and see how high up it goes.
他们很生气，而且手里有筹码。
They're pissed and they've got leverage.
所以现在处境很被动，因为说到底，他们真正在意的不只是那些机密——你知道的，他们关心这些技术，但更核心的是，OpenAI一直在说要搞硬件设备。
So it's a tough place to be because really what they want isn't, I mean, you know, they care about these secrets, but what they're really dealing with is, you know, the big picture comment is OpenAI has been talking about building a hardware device.
这正是他们从苹果挖走四百人的原因，苹果因此暴怒，现在有了主动权，一定要把事情查个水落石出。
That's obviously why they're hiring 400 people from Apple And Apple's pissed, and now they've got some leverage and they're going to find out what's going on here.
是啊，正如罗里说的，这事确实挺棘手的。
Yeah, you know, it's a tough one to Rory's point.
我们每天都在招聘有行业经验的人，不管是做运营还是做投资，这事儿再正常不过了，对吧？
We all have history in operating or investing experience where you do hire folks for their domain expertise every day, right?
这其实就是职业的流动性。
That's employment portability.
至于商业秘密的界限到底在哪，其实挺微妙的，法律上是这样，可能Rory会跟我争辩，但现实中更复杂，因为不是每个人都会被起诉。
And where the line is for trade secrets, it's actually somewhat nuanced, not only in the law, maybe Rory would challenge me, but in practice because not everyone gets sued.
实际上，偷商业秘密的事儿天天都在发生。
People are stealing trade secrets all the time.
但这么做确实不对，对吧？
They shouldn't be, right?
销售员带着客户通讯录跳槽的时候，得格外小心。
The salesperson's got to be careful when they take the Rolodex.
偶尔还真有人因此栽了跟头。
Once in a while, they get burned for that.
有时候，销售代表跳槽到你投资组合里的竞争对手，公司就会立刻发火。
Once in a while, a sales rep actually they lash out when one goes to a competitor in your portfolio.
他们被起诉是因为带走了真实的通讯录，而不是脑子里记的那些联系人。
They get sued for taking their actual Rolodex, not their conceptual one.
离职时下载客户管理系统，甚至黑进旧文件夹，或者干点别的什么把苹果公司的文件偷走——这种事真有人干，还被逮住了。
They get sued for downloading the CRM when they leave and Jesus Christ hacking into the old I folder or whatever the hell he did to take documents from Apple after you leave.
这种事情确实会发生。
It happens.
但这种行为也太过分了。
But that's too much.
对吧？
Right?
不过也不是每个人都会被起诉，这事儿确实挺难界定的。
But still, don't always get sued and it's tough.
但如果我们连行业专家都请不到，总不能指望人人都自己发明大语言模型吧。
But if we can't hire domain experts, sometimes we can't all invent LLMs ourselves.
有时候，我们确实需要一点外部的帮助。
Sometimes we need a little help.
杰森，你对这个挺兴奋的。
Jason, you're excited about it.
我觉得这恰恰说明你根本没必要这么做，退一步说，加州在雇佣关系上可是全美最开明的地区之一，对吧？
I actually think it actually makes the point that you didn't need to do this because stepping back, California is one of the most liberal places in the country in terms of employment, right?
那里根本不存在竞业禁止协议。
Which is, you know, there's no non competes.
就算有那种条款，法院通常也不认的。
There are kinds of non competes are typically unenforceable.
还有一种叫‘不可避免披露’的法律原则，你知道吧？
And there's this doctrine called inevitable disclosure, right?
意思是你在一家公司学到的知识，跳槽到另一家公司后用了，加州法院会说，这根本不可避免，因为这属于你专业技能的一部分。
Which is if you acquire knowledge in one company and you go to another company and you use that knowledge, the California courts are pretty friendly and they'll say it's inevitable because it's part of your trade.
所以他们甚至会说，我们没法阻止你把在A公司学到的本事带到B公司去。
So they'll even say, we can't stop you using the tricks you learned at company A at company B.
我给你举个很实际的例子。
And let me give you a very practical example of that.
Anthropic 是最受益于无法执行竞业禁止协议和必然披露原则的公司。
Anthropic is the biggest beneficiary of the inability to enforce non competes and the doctrine of inevitable disclosure.
因为那七名员工离开时，完全是凭自己的本事走的，他们比这 guy 聪明多了。
Because those seven employees who left, they basically walk out of them because they're way smarter than this guy.
他们可能什么都没带，连一张纸都没拿。
They probably didn't bring anything, No pieces of paper.
更别说偷代码了。
No stolen code.
但第二天一早，他们就开始说：我们上周刚想到这个点子，现在还能继续想，脑子里的东西都是我们自己的。
But next morning, they start up and they say, we just were thinking about this last week, and we can keep thinking about this, and everything in our brain is owned by us.
全美还有很多州是不允许这么做的。
And there's lots of states in the union where you can't do that.
我好奇 Anthropic 能不能在马萨诸塞州成立。
I wonder if Anthropic could have been started in Massachusetts.
大概不行。
Probably not.
对吧？
Right?
不对。
No.
你说得完全对。
You're exactly right.
虽然加州作为雇主挺糟糕的，但我认为加州的做法是对的。
Which and I think California's even though it sucks as the employer, I think California's been right.
奇怪的是，这竟然是加利福尼亚州少有的政策胜利，对湾区的创新帮助极大。
Oddly enough, the rare policy triumph for the great state of California, It's been really good for innovation in the Bay Area.
对吧？
Right?
你本来就有一个对员工非常友好的环境。
So you already have that very friendly employee thing.
而且正如杰森所说，既然这种人才流动是不可避免的，那何必冒险去偷那些根本不需要的东西呢？
And given to Jason's point, given you have that given you have that inevitable, why put yourself at risk by stealing some shit that you just don't need?
嗯，我原本的问题是没问题的。
Well, my question was okay.
直白点说，这些人都被坑了。
So these individuals are screwed, to be blunt.
这对OpenAI的时间表意味着什么？
What does it mean in terms of timeline for OpenAI?
会不会造成实质性的延误？
Does it cause meaningful delays?
会不会影响他们按原计划推出硬件产品？
Does it impact their ability to ship their hardware product when they wanted to?
要是让我猜的话，硬件项目恐怕悬了。
Listen, if I had to breathe the tea leaves, hardware is probably on the bubble.
这是萨姆当初推动的一个计划，当时一切顺利，OpenAI在竞争对手面前有绝对优势，对吧？
This was one of Sam's initiatives that made sense when everything was working and OpenAI had an unassailable lead over its competitors, right?
现在呢，你知道的，我们砍掉了Sora。
Now everything's, you know, we chopped Sora.
我们当初干嘛要花大价钱收购TBPN？明明六个月前看起来那么棒。
Why the hell did we buy TBPN as great as it seemed six months ago, right?
当时六十几亿美元买硬件，还能请到传奇的乔尼·艾维，这买卖听起来挺值，结果团队把所有东西都给拿走了。
Hardware seemed like a great deal for six billion dollars when I could get the legendary Johnny Ive, but team stole everything.
这纯粹就是个干扰项，只会不断烧钱，根本没意义。
What's the point of like this is just an effing distraction for something that's going to ham it's going to hemorrhage cash.
我全力支持。
I'm all in.
如果回溯六十周前，我确实坚定看好无处不在的AI。
Well, if we go back sixty weeks ago, I'm all in on persistent AI around us.
扎克伯格搞的新眼镜之类的东西，但这个项目恐怕也快得砍了。
So Zuckerberg, like with his new eyewear and all that, but this has got to be almost on the cutting board.
就跟苹果的汽车一样。
It's like the Apple car.
差不多该砍掉了。
It's got to be close to being cut.
我觉得这场诉讼发生的原因，可能是想给它最后一击，逼着公司砍掉它。
I think the reason this lawsuit may have happened is to push it over the line of getting cut.
直接砍了吧，各位。
Just kill it, guys.
虽然不是直接的因果关系，但真有可能成为压垮它的最后一根稻草——咱们不如先搁置一年再说。
Like it's not directly cause and effect, but it really could push it over the line to just let's let's put a pause on this for a year, boys.
咱们先搁置一年吧。
Let's put a pause on it for a year.
我一开始没跟风投入，不是因为我不认同这个观点，而是我完全同意。
I didn't pile in on that initially because I feel I pile in on this point, but I totally agree.
我记得他们签那个个人合约的时候，我就觉得这事儿太蠢了。
I remember when they did the individual deal, I was like, this is stupid.
有时候吧，老话讲得没错：抓住重点，别跑偏。
Sometimes, you know, the, the, the cliche keep the main thing, the main thing.
这里的重点是，大语言模型在编程方面实在太厉害了，由此产生的经济价值，远超其他所有东西。
The main thing here is that LLMs are really amazing at code and the amount of economic value that can be created from that dwarfs everything else.
我看过专栏作家本·汤普森的观点，他说某种程度上，OpenAI 在消费端太成功，反而成了干扰。
I've seen some of the columnist Ben Thompson even articulate the view that to some extent it may have been a distraction for OpenAI that they were so successful in consumer.
因为真正创造价值的地方，其实并不只是在企业市场，而是在企业市场的编程领域。
Because whether the value is all clearly been created is not just within enterprise, but within enterprise for coding.
我同意你的看法，杰森。
And I'm with you, Jason.
如果这真是一次仁慈的终结，苹果可能反而帮了他们一把。
If this was a mercy killing, Apple may have done them a favor.
别纠缠了，放慢脚步，别再干这种事了。
You know, move on, slow it down, stop doing this shit.
想想一年前，当 OpenAI 还是个消费级公司的时候，我知道你觉得这说不通，但退一步看。
If you think about a year ago when OpenAI was a consumer company, I know you thought it didn't make sense, but step back.
如果你是个消费级公司，这一切其实都合情合理。
It actually all makes sense if you're a consumer company.
我们该不该进军硬件领域？
Should we go into hardware?
总归有一天，你会开始认真思考这个问题。
At some point you do, you think about it.
我们是不是该把媒体也当作一家消费公司来运作？
Should we go into media as a consumer company?
对。
Yeah.
比如说，我们花一亿多美元买下一个顶级播客，也不是不可能。
I mean, we could buy a top podcast for 100 and something million.
最坏的情况，也就是这笔投资没成功。
Like, worst case, it doesn't work.
如果你纯粹是面向消费者的公司，这正是你持续扩张的方式。
Like, if you're a pure consumer play, this is how you keep building.
你一层一层地往里加，像同心圆那样不断扩展，对吧？
You add to this, you add to the layers of the concentric circle, right?
但如果你是个企业级的AI模型，这么做就完全说不通了。
As an enterprise death march LLM, it makes no sense.
这真让人想不通。
It's baffling.
对吧？
Right?
让别人去开发那种别在衣服上的小徽章吧。
Let someone else build the pin the pin that clips to your shirt.
你提到扎克伯格了。
You mentioned Zuck.
扎克伯格重回X，还为Mews Spark 1.1推出新政策——这是Meta第一次向开发者收取使用自家模型的费用。
Zuck returns to x, and he returns to x for mews spark 1.1, and the first time ever that Meta start charging developers to use their own models.
有些人觉得，这次发布意义重大。
It was a very meaningful release, some people think.
亚历山大·王特别指出，这个定价相比OpenAI和Anthropic太过激进了。
Alexander Wang called the pricing of it, in particular, very aggressive versus OpenAI and Anthropic.
正如我所说，扎克伯格打破了三年的沉默，在X上发布了这个消息。
As I said, Zuckerberg broke a three year silence on X to launch it.
我们当初是怎么看待Spark发布的？
How did we think about the release of Spark?
这对Meta在对抗OpenAI和Anthropic的斗争中，算不算一个重要的进展？
And was it a meaningful progression for Meta in the fight against OpenAI and Anthropic?
是的。
Yes.
我想先岔开一下，聊聊实际是怎么操作的。我很好奇，扎克伯格内部的判断标准是什么——到底得有多好，他才会骂骂咧咧地把东西发到竞争对手的社交平台上？
And I'd like to come back to the how it's actually let's do on a tangent first, I wonder what the internal Zuckerberg algorithm was, which is how good does something have to be before I spitting and cursing, put something on the competitor social media product?
我的意思是，也许公关的人跑来跟他说，新出的Glass太棒了。
I mean, you know, maybe the comms person goes in and says, I think the new Glass release is awesome.
我们要不要在X上发一下？
Should we use X?
他直接说：别搞了。
He's like, fuck, no.
但最后，他还是说了：所有开发者都在X上。
And then finally, he's like, all the developers are on X.
如果我们现在推出Spark，那简直就是在自取其辱。
If we're going to launch Spa, we just look like idiots.
然后他就说，好吧，去那个搞笑公司发点东西吧。
And then he's like, okay, type something on the clown car company.
不过算了，也无所谓。
Anyway, that's fine.
但我之前根本没想过。
But I don't Before think
你先别急着跳过，亚历山大·王在X上发了27条，但在Threads上只发了3条，结果网上很多人都气炸了。
you move off that, Alexander Wang has done 27 posts on X to three posts on threads, and it actually got people quite pissed off online.
我觉得内部可能也...
And I think internally
他可能根本忘了还有Threads这回事，我也是直到前几天才想起来。
He probably forgot they had threads as I as did I for until the other day.
我吗？
Did I?
是吗？
Did It's
据说已经有四亿用户了。
got 400,000,000 users apparently.
差不多吧。
Sort of.
对。
Yeah.
因为它是和 Instagram 整合的，你想躲都躲不开。
Because it's integrated into Instagram, you can't help but see it.
但这不代表真的有互动。
It doesn't mean there's any real engagement.
关键是活跃度。
Activity is the word.
怎么才算互动呢。
How engagement.
咱们说实话吧。
Let's be honest.
咱们三个都是老粉了，真正在玩的可不是Threads。
As three inveterate inveterate Twitter junkies here, guys, we know where the action is, and it ain't on threads.
明白吗？
Okay?
说实话，要是Threads真有动静，哈利，你早就冲过去了。
Let's be honest, If there was action on threads, Harry, you'd be there.
那当然。
I would.
就算火星上有动静，哈利你也会冲过去。
If there was action on Mars, Harry, you'd be there.
咱们还是说重点吧。
So let's focus on the real point.
毕竟，股价涨得不错，因为这儿确实发生了不少事。
I mean, and the stock jumped nicely because a ton of things happened here.
他们推出了 Spark 1.1，然后就从这里撤了，免费提供。
They shipped Spark 1.1, they move away from here, it's free.
我老提醒自己，忘了的时候就想想，当初那会儿既不炫酷，也不是开源，只是开放权重，但他们后来转头开始收 API 费了。
As I always remind myself, when I forget, it wasn't glam, it wasn't open source, it was open weight, but they moved away from that to charging on an API bit.
换句话说，他们跟其他前沿实验室走的是一条路。
In other they've embraced the same business model as the other frontier labs.
这产品还不错。
It's a decent product.
我作为普通用户用过。
I've used it as an end consumer.
其实我拿它跟 Claude 做过对比，看了一些回答。
Actually, you know, I compared and contrasted to Claude and some answers.
表现挺不错的。
It was pretty good.
但说实话，真正的考验根本不是普通用户用起来怎么样。
But of course, the real test isn't already the end user.
真正的考验是写代码。
The real test is coding.
他们公布的基准测试成绩不错，但过去他们的成绩没这么好。
And, you know, they reported good benchmarks, but in the past, their benchmarks haven't been as good.
待会儿我们再聊基准测试的事。
And we'll talk in a second about benchmarks.
但从大局来看，关键在于他们采用了和别人一样的商业模式，用低价产品打入市场，给竞争对手造成了巨大冲击。
But fundamentally, the big picture point is in the game, using the same business model as everyone else, entering the marketplace with the low price product, big hit to competition.
我的意思是，这相当于在和对方的资产负债表硬碰硬。
I mean, it's, you know, going toe to toe with the balance sheet to afford it.
五年或十年后，作为市场第四名，以这个价格销售，能实现正回报吗？
Will it be ROI positive over five or ten years to be the fourth player in the market selling at that price?
我不确定。
I don't know.
但如果你是OpenAI或者Anthropic，肯定不希望这事发生。
But, you know, if you're OpenAI and Anthropic, you'd have preferred this not to happen.
我觉得这条推文挺有意思的。
It seems to me the tweet's fun.
很明显，他们想在评估和推文评估上赢一把。
Clearly, they wanna win the eval game and the eval tweet game that's out there.
这确实能吸引眼球。
It gets a tent like it clearly gets attention.
你在评估榜单上的位置，本身就容易引来关注。
Where you are on the evals gets attention.
这价值很高。
It's very valuable.
为了这个，扎克伯格也值得上推特走一趟。
It's worth going on Twitter for Zuck.
在我看来，只要还没这么做，大家迟早都会去堆令牌数。
My Captain Obvious view is everyone's going to be token maxing if they aren't already.
每个人都会撞上上限的。
Everyone's going to hit the limits.
昨天我第一次在最高档套餐下用完了Claude的配额，得想办法解决了。
I finally hit my Claude limits yesterday for the first time on my max plan, and I had to decide what to do.
每个组织都会从‘随便用、别省着’转向预算管控，不管你是自动限制、手动控制，还是选个下拉菜单，最终都得用个更便宜的模型。
And as every organization goes from burn all the tokens you want, baby, to budgets, however you do it, whether it's automated, whether it's manual, whether it's picking a dropdown, everyone's going to have to have a cheaper model.
也许你的产品里需要一个更经济的流程版本。
It may be in your product for cheaper workflows.
也许是为了写代码用。
It may be for coding.
所以现在，这俨然是一场争夺低价市场的较量。
So at the moment, this seems like a battle for the cheap seats.
而且这场低价市场的竞争还特别激烈。
And it's a big battle for the cheap seats.
咱们走着瞧吧，看最后会怎样。
And we'll see what happens.
我们平时不提，但我觉得Anthropic的Haiku真是个相当不错的工具。
We don't talk about it, but I think haiku from Anthropic is a pretty damn good product.
他们只要调一下 Haiku 中 Opus 的使用比例，一天之内就能把它做得更好。
And in one day they could make it better just by by adjusting the slider of how much opus you get in haiku.
我实际用的 Haiku 量特别大，用来处理便宜又简单的任务，效果非常好，成本还不到一分钱的十分之一。
I literally use a massive amount of haiku and for like cheap, simple stuff, it's really good and it's a tenth of a cent.
所以现在这场竞争聚焦在了这个第二梯队上。
And so there's a battle for this second bucket.
这里的利润空间有多大？
How much of the margin is there?
收入又能有多少？
How much of the revenue?
谁想在这块儿赢下来？
Who wants to win in it?
我不清楚。
I don't know.
所以我觉得这挺有意思的。
So I think it's interesting.
表面上看，这场竞争发生在前沿领域，但我认为真正较量的是B级梯队——那里仍有大量使用量，但利润空间到底有多大，现在还不确定。
It appears to be a battle in the frontier bucket, but I think it's a battle in the B tier bucket where there's still a lot of volume, but how much margin of revenue is TBD?
我觉得你说得对。
I think that's correct.
我同意你的看法。
I agree with you.
我觉得Databricks那篇论文，还有我们待会儿要聊的几篇其他论文，都提到了模型分层和分梯队这个概念。
And I think the Databricks paper and a couple of other papers that we'll talk about have talked about this concept of buckets and tiers of models.
你说得真好，杰森，完全说中了。
I like what you're saying, Jason, you're exactly right.
当成本压力摆在面前时，你那句话特别打动我：每家公司都会建立内部的分级体系。
When the pressure on spend is out there, the sentence you said that really resonated with me is every company is going to have an internal tiering.
他们可能都从OpenAI或Anthropic采购，但只要CIO稍微有点脑子，就一定会配一个廉价的模型来控制这种无节制的消耗。
They might all buy it from OpenAI or Anthropic, but every company with a CIO who's half awake is going to have a cheap token model to hand to stop this madness.
某种程度上，这就为这类模型留出了位置。
And to some extent, means there's a slot for that.
有意思的一点是，大家都知道Meta签了那个合同。
A fun interesting one would be, guys know that obviously Meta did that contract.
现在他们也开始像SpaceX那样，短期对外提供算力了。
They're starting to offer their capacity now just like SpaceX on a short term basis.
挺值得琢磨的是，他们到底会定什么价格。
It'll be fun to figure out at what price.
我不确定他们按现在的token价格卖，真能赚更多钱。
It's not clear to me that they would make more money selling at these token prices.
说不定把算力租出去反而赚得更多。
They might actually make more money renting their capacity out.
这事儿挺有意思，得想想算力到底该给谁最划算。
So it'll be interesting to think about compute to the highest user.
等我们聊到Grok和SpaceX的时候，再回头讨论这个。
You know, we'll actually have that discussion again when we come to Grok slash SpaceX.
这产品定价挺激进的，说真的，他们干得不错。
But this is an aggressively priced product and, you know, good on them.
你说得对，杰森。
You're right, Jason.
价格已经压到最低了。
It's pushing at the low end.
是的。
Yeah.
他们可能会见机行事。
They'll probably play it by ear.
如果这个项目更成功，他们就能获得更多算力。
I mean, if this is more successful, then they get more capacity.
如果效果差一些，就把资源让给竞争对手。
If it's less, then they give it to their competitors.
他们大概还是会见机行事。
I mean, they'll they'll they'll probably play it by ear.
我们之前提到过Databricks的那篇论文。
We mentioned the Databricks paper.
这篇论文引起了很大关注。
It got a lot of attention.
我们能继续聊这个吗？
Can we can we actually stay on that?
你能总结一下发生了什么，以及为什么重要吗？
Can you summarize what happened and why it's important?
有些内容其实很明显。
Some of the stuff is obvious.
我通读了一遍，一开始想抓住重点，其实就是说，成本按每个词算根本没什么用。
I read it all and trying to the big picture at the start was the basic point is, you know, you know, cost per token is really not a useful metric.
真正该看的是完成每个任务花了多少钱。
It's kind of cost per completed task was the first thing.
办成一件事到底要花多少成本？
How much does it cost to get things done?
对吧？
Right?
而且有些模型虽然单个token成本低，但推理用的token反而更贵，你还没法预估他们到底会用多少。
And because one of the is some of these things that have low cost per tokens, their reasoning tokens are more expensive than their basic tokens, and then you don't know how many they're going to use.
这里面涉及的因素太多了。
There's a whole bunch of things.
所以第一个关键点是，真正该衡量的应该是完成每个任务的成本。
So the the the first picture point was cost per completed task is really the thing you have to assess.
他们提出的第二个观点很有意思：不同任务会形成一条帕累托曲线，也就是说，每种任务都有最适合它的模型。
Then the second point they made, which is interesting, is you end up with a Pareto curve for different kinds of tasks, which is the best model for this guy.
这正是杰森想表达的。
And this is Jason's point.
会有一批复杂度很高的任务，必须用高性能模型来处理。
There will be a big brainiac frontier set of tasks for which you use them.
一旦你把任务成本量化了，就会发现不同模型在不同任务上的成本效益差异很大。
Once you quantify the cost per task, then the second thing you're going to have is different models are more cost effective at different tasks.
有时候，一个便宜但需要无限推理的模型，反而不如直接花钱用更贵的模型来得划算。
And a cheap model that just requires infinite reasoning mightn't be as good as paying up for the more expensive model.
所以你可以想象出这条价值的帕累托曲线。
So you can imagine this Pareto curve of value.
第三点， broadly speaking，是你所使用的工具——也就是背后的基础设施——会极大影响整个系统的效率。
And then the third point, broadly speaking, was the harness you use, in other words, your infrastructure around that, massively impacts kind of how efficient that all happens.
所以这中间有很多奇怪的地方，你看到那座桥了吗？
So it's quite a lot of just weird So saw the bridge?
Databricks，是的，他们确实这么做了。
Databricks, yeah, they did.
没错。
Exactly.
说实话，我真没想到会这样。
Yeah, mean, I'm shocked exactly.
我简直震惊了，居然有家公司明明在帮你管理模型，却说管理模型很重要。
I'm shocked to discover that company that effectively helps you manage models says that managing models is important.
但公平地说，Jason，他们确实拿出了扎实的证据来支持这个观点。
But to be fair, they produce the goods, Jason, in terms of facts to back it up.
这很合理。
And it makes sense.
我觉得不该只看每次发布新模型时他们给出的每百万输入输出令牌的成本，
I think moving away from because every time someone announces the new model, they they kinda give you the cost per input, million input and output tokens.
我觉得更应该关注每个任务完成的成本，作为CIO，我完全能理解这种思路。
I think the emphasis on cost per completed task, it's just I can totally see that making sense if I'm the CIO.
我今天早上看了ClickHouse的Aaron Katz发的一篇帖子，说他们自二月以来的AI支出增长了60倍。
I I like the post this morning from Aaron Katz at Clickhouse who said that their AI spend is up 60 x since February.
哦，真的吗。
Oh, yeah.
你看到这个了吗？
Did you see this?
我看到了。
I did.
我看了那篇帖子。
I saw the post.
展开剩余字幕（还有 480 条）
是的
Yeah.
他还说，人工智能正在从根本上改变人们对数据平台的期望，如果我们不能深入理解它，就无法为AI打造最好的数据平台。
And then he said AI is fundamentally changing what is expected from a data platform, and we can't build the best data platform for AI if we don't deeply understand it.
我们正在投入资源组建一支在AI理解和创新方面真正领先的团队。
We're investing in building a team that's truly leading in understanding and innovating with AI.
六十倍
60 x.
我的意思是，一旦你开始做点复杂的事情，支出涨六十倍其实挺正常的。
Well, I mean, I mean, once you do anything complicated, it's easy to go up 60 x.
对吧？
Right?
别再只设计一个页面了，干脆让Claude在后台帮你彻底重做整个网站。
Instead of doing one design for one page, your website, have Claude design redesign your entire website in the background.
这样五分钟就把你整个月的token配额耗光了。
That'll blow your token budget in five minutes for the whole month.
很快就耗完了。
It'll be gone.
这些复杂的流程会消耗大量令牌。
These really complex workflows consume a lot of tokens.
即使令牌价格下降，流程和输出却变得越来越丰富、越来越复杂。
Even as they get cheaper, the workflows, the outputs are getting richer and richer and more and more complicated.
所以，令牌耗尽这个问题，早就从少数人的困扰变成了普遍现象，对吧？
So not only is token maxing, you know, gone from niche to interesting, right?
每个开发者都能消耗更多令牌。
Every developer can consume more tokens.
普通开发者会这么做，只是为了装样子，假装自己在干活，尽管他们的代码提交全都被拒了。
The mediocre developers can do it to be performative and pretend they're working even though none of their PRs are accepted.
而顶尖的开发者，真的能同时运行十到二十个代理，7x24小时不停。
And the great ones literally can run 10 or 20 agents 20 fourseven.
你去问问那些真正沉迷AI的顶尖开发者吧——虽然我不喜欢这个说法——他们现在几乎每天24小时都在写代码。
I mean, you talk to the best developers now that are AI pilled and I hate that term, okay, they are coding twenty four hours a day now.
没错，而且这上瘾。
Okay, they are because and it's addictive.
就跟打游戏一样。
It's like video games.
现在我能一次性开发出十个功能，一百个功能。
Okay, now because I can build 10 features, 100 features.
我能做以前根本做不到的事，而且快得惊人。
Now I can do things I couldn't I can do so quickly.
现在都晚上十点了。
So I'm it's 10:00 at night.
我不再刷短视频了，改成了疯狂写代码。
I'm checking instead of instead of doom scrolling, I'm I'm doom coding.
对吧。
Right.
我通宵干，早上醒来第一件事就是去看看那个流程跑得怎么样。
And I'm going all night and I wake up in the morning and I want to check-in on that workflow.
而且不只是检查一个功能，要是它能把我所有功能都给建好了呢？
And instead of checking in on one feature, what if what if it could have built all my features?
对啊。
Right.
再说一遍，我不喜欢这个词。
And so and again, I hate the term.
我认识的开发者，没有谁会嫌前沿级别的令牌不够用。
I don't know any developer that couldn't consume even more frontier level tokens.
我们只会想要更多。
We would just want more.
对啊。
Right.
你别的什么事都别干了。
You can do nothing else.
你可以把正在做的东西，做成更高级的版本。
You can do more prime versions of whatever you're working on.
你可以直接多做几个版本。
You can just do more versions.
这样能省时间，效果也更好。
It saves you time and it's much better.
我干嘛非得在A、B、C之间选一个呢？
Why do I have to choose A, B or C?
干脆全做出来，再分别优化成A加、B加和C加。
Just build them all and then build A prime, B prime and C prime.
对。
Right.
咱们再往前想一步。
And then let's actually go further.
把它们全部署到生产环境、预发布或开发服务器上。
Let's put them all up on production, on staging or the dev server.
看看它们各自表现怎么样。
Let's see how they all work.
你知道吗
You know what?
咱们再进一步
Let's go further.
咱们用一百万条测试用例来跑一遍
Let's run a million test cases against it.
对吧
Right?
哦，这得花两万美元的调用费用
Oh, well, will cost $20,000 in tokens.
好吧，其实不用
Okay, well, won't.
如果可以的话，我会先缩小范围
I'll narrow down the what if I could?
这样覆盖更全面，对吧
It would be better coverage, right?
你可以调用任何顶级开发者，消耗的令牌量是他们现在的好几倍。
You can consume any top tier developer and can consume an order of magnitude more tokens than they are now.
创意是无限的。
The ideas are limitless.
我同意。
Agreed.
我觉得我想再深入聊聊这个点。
And I think I want to stay on this one for a while.
我觉得有意思的是，杰森，以前从来没有任何产品能做到这一点，这也是为什么这成了一个管理难题。
I think what's interesting about that, Jason, there's never been a and this is why this is a management challenge.
以前从没出现过这样的产品——能按需提供服务，让个体员工不花一分钱，就能替他们完成工作，还让他们表现得特别出色。
There's never been this product before, which on a totally on demand basis can allow for an individual worker with no cost to them can basically do their work for them and make them look amazing.
是的，公司花一百美元，你想想看，如果你是个开发者，你完全可以自己花两周时间干这事，但你只要按一下神奇按钮，就会有大笔开销算到公司头上，却不会算在你个人头上，还能让你轻松搞定任务，那我当然选后者。
Yeah, the company is earning $100 If you think about it, if you're a developer and you're sitting there going, I could crank this thing for the next two weeks myself, or I can press the magic button, and somewhere out there a whole bunch of expense will accrue to my company, but it's not been taxed on me, and it can get me a long way down, I'm going to do that.
如果我是实习生，比如今天早上我得整理一堆参考资料，我肯定直接用Magic Claude帮我总结了，对吧？
If I'm an intern, I mean, you know, summarizing references in our shop this morning, I'm going to get the Magic Claude summary, right?
那为什么不呢？
And why not?
说得明白点，有时候这能为公司带来巨大价值——你省下了一位员工一小时的时间，他每小时工资500块，而你只花了20块。
And to be clear, sometimes that's going to be wildly value accretive to the company because you saved an hour of someone's time that you're paying, you know, $500, then you paid $20.
你肯定会为他这么做感到高兴。
You're ecstatic they did that.
但正如杰森指出的，如果对此没有限制，迟早你会遇到花600块去省500块的情况。
But as Jason points out, if there's no governor on that, eventually, you're gonna hit the moment where you spend $600 to save $500.
现在的问题是，人们该怎么掌握好这个度？
What's going on right now is how do people get to grips with that?
我简单说一下，虽然可能听起来很基础，我知道有些观众听了会笑我，但真的很重要。
Let me just I I sort of hit, but let me just give you a simple even if it's very basic, I know some folks will make fun of me that watch this, right?
或者听众也会这么想。
Or listen to it.
但我也是花了一段时间才想明白的。
But it took me a while.
我现在完全用上Claude的设计了。
Now I'm all in on Claude design.
它变得特别好用了。
It got really good.
好吧。
Okay.
刚上线的时候，我花了挺长时间才适应。
It took me for a while when it launched.
我一开始还以为它只是个玩票的功能。
I I wasn't sure if it was a play feature.
我现在真的很喜欢它。
I love it now.
我觉得它比以前更厉害了。
I think it's gotten better.
但就像别人说的，我花了好久才明白，它消耗的token比其他任何工具都多。
But as folks said, and it took me a while to get it, it consumes more tokens than anything else I do.
这消耗量太大了。
It is massive.
昨天我发第一条消息时，才过了五分钟，就有人提醒我说：‘杰森，你得等到下午三点才能用额度了。’
And so yesterday for my first message, I'm like five minutes in, it's like you got to wait till 3PM Jason.
得等到下午三点才能用你的积分。
You got to wait till 3PM to use your credits.
其实就改一个页面而已，却用掉了这么多积分。
And that's to do one basically redesign one page in essence because it uses so many.
我更希望的是这样：想象一下。
What I would love to do instead, imagine this.
这是我的整个网站，不管是什么产品。
Here's my entire website for any product.
一共一百页，不管多少页吧。
Okay, it's 100 pages, whatever it is.
每天晚上，自动生成更优的版本。
Every night, come up with better versions.
每天晚上。
Every night.
所以我早上醒来，就会逐个浏览你昨晚给我网站提的那些改动建议，挑出一百个里最好的两三个，而不是浪费资源。
So I'll wake up in the morning and I'll and I'll doom click through your ideas from my website and I'll pick the two or three of the 100 that are better instead of burning.
你想想看。
Like, think about it.
我花二十分钟就耗光了所有积分，就为了做一个页面。
I burned through my tokens in twenty minutes building one page.
如果每晚都要重做一百个页面，那得消耗多少积分？
How many tokens would it create for every night to redo 100?
对吧？
Right?
Rory，你帮我算算这笔账。
Help me with the math, Rory.
这得多出好几个数量级的积分消耗，对吧？
It's a lot of it's orders of magnitude more tokens, right?
我觉得，每个产品人员、每个设计师都能这么做，轻松消耗比现在多一百倍的令牌，连眼睛都不用眨一下。
I think literally every product person, every designer could do that and consume 100x tokens they do now easily, without even blinking.
没错。
Agreed.
这会影响你用 Figma 的方式吗？
Does it change your Figma usage?
我觉得不会，不过别深究这个了——我觉得 Claude 的设计能力如果继续这样投入下去，我们就能明白，虽然令牌消耗巨大，但你可以不用设计师，或者只用半个设计师、外包设计师，也能推进很多，因为你完全可以搭建出整套设计框架，还能生成人像。
I don't think so, but not to go down to rat hole, I do think Claude designed, if they keep investing in it like they have and we could figure out the massive token consumption is that you can get further without a designer or you can get further with half a designer or an outsourced designer or something because you can build your whole design schema and build people.
你肯定会笑话它。
You're going to make fun of it.
你还是能一眼看出这是 Claude 做的设计。
You can still smell a design a Claude design app.
你能看到那些元素的痕迹，但确实能走得很远。
You can see the elements, but you can get pretty damn far.
在产品层面，它确实挺聪明的。
It's pretty smart on the product side of how to do it.
所以这并不意味着不需要设计师，但可能意味着我们不再需要那种差劲的设计师来做出V1版本了，对吧？
And so it doesn't mean you don't need a designer, but it may mean that we don't need that crappy designer anymore to get us to a V1, right?
根本没必要再请这种人了。
There's just there's just no need to have it.
那这对Figma有影响吗？
So does that harm Figma?
没有，但我觉得在AI时代，作为CEO，一旦发现你的低端客户开始流失，就得警惕了。
No, but I think in the age of AI, every time you start losing the bottom of your market as a CEO, be nervous.
这种变化不会体现在当季的财务数据里，因为那是低端市场。
It won't show up in that quarter's numbers because it's the bottom of the market.
那些都是小客户。
It's the ankle biters.
你早就往高端市场走了。
Like you've already gone up market.
Figma现在盯的都是百万美元级别的大单，对吧？
Figma is tracking million dollar deals, right?
谁会在乎那点小单子被Claude Design抢走了呢？
Who the hell cares to the low end one seat deal it lost to Claude Design?
但随着时间推移，这些小客户会慢慢往上走。
But over time, those ankle biters move up the leg.
他们就像食人鱼一样。
You know, they're the piranhas.
先吃掉膝盖，再啃到大腿。
They take over the knees and the thighs.
很快你就只剩下大客户了，但你的客户漏斗已经缺了一块。
And pretty soon you retain the big customers, but you've lost a portion of your funnel, right?
如果你丢了底层客户，就等于丢了一部分客户漏斗。
If you lose the bottom, you lost a portion of your funnel.
他们不会等着说，等我长大了再用Figma。
They don't all wait and say, you know what, I'll graduate to Figma when I'm bigger.
这就是Salesforce现在面临的危险。
And this is where Salesforce is at risk.
像我这样喜欢Salesforce、用无头模式的人，但下一代人可能根本不会转向Salesforce，因为他们压根就没从那儿开始。
People like I love Salesforce who run it headless, but the next generation of folks may not graduate to Salesforce because they never started there.
他们一开始就是用智能代理工具。
They started agentically.
他们可能永远都不会转到Salesforce上来。
They may never graduate there.
如果我是CEO，我就担心再也没人会升级到我们的平台了。
If I'm a CEO, I'm at risk that folks won't graduate into anymore.
工作流工具的风险在于，哪怕只有10%到20%的人不再升级，也会引发增长缓慢的恶性循环。
Workflow tools are at risk that even if just 10%, 20% less folks graduate into, it leads to a death spiral of slow growth.
所以，对上市公司来说，我最关心的只有一件事：他们每年新增客户数量有没有增长15%以上？
That's why, you know, for public companies, the only thing that really matters to me is are they growing net new logos 15% or more a year?
如果有，那他们迟早能找到出路。
If they are, they'll figure it out, right?
如果你每年新增客户能增长15%，甚至20%，那就没问题。
If you're growing your net new customers 15, even 20% a year.
但一旦下滑，说实话，你每年提价也撑不了多久。
But as soon as it falls, man, you could I mean, you can raise price every year, but it only takes you so far.
所以拉姆利的答案是，我确实担心，但这对Figma这种企业级业务根本构不成威胁。
So Ramly answer is it does worry me, but it's not a threat to this sort of enterprise Figma play at all.
但如果我现在要创业，那确实是个不错的起点。
But if I are building a startup today, that's a right start.
罗伊，你觉得我漏掉了哪个更有趣的点？
Roy, what point did you think was more interesting that I didn't mention or glossed over?
你刚才说，那根本不是他们想表达的意思。
You said that actually wasn't the point they made.
只是当你思考市场时，有些事情会让你忍不住琢磨。
It just things you find yourself puzzling on as you think about the market.
我现在就是看这些公司的市场反响，简直太惊人了，对吧？
I'm just looking at the level of traction on these companies right now, which has just been astonishing, right?
我的意思是，Anthropic今年年初还值90亿美元，现在可能已经涨到500亿美元了。
I mean, the Anthropic explosion to what now might be 50,000,000,000 from 9,000,000,000 at the start of the year.
我看了半分析的那份拆解，确实其中有两三百亿美元是Claude的代码部分，但说实话，我觉得API部分里也有很大一块是代码。
You know, I saw the semi analysis breakdown and yeah, two or 3,000,000,000 of that is Claude code, but the real truth is I think a huge portion of the API part is code as well.
我一直都在想，当某样东西增长得这么快，势头这么猛，我就会问，到底什么能改变这种趋势？
And I, you know, I've just been wondering, you know, when something's growing that quickly, momentum, I think I'll say, what could cause it to change?
我觉得真正可能影响它的只有两个因素，一个是大规模竞争，这个我们已经聊过了。
And I think that, you know, there's really only two things is that one is mass competition, which we've been talking about.
如果所有人都疯狂降价，价格直接砍半甚至降到三分之一，那确实会看到你提到的这种低端token定价，大概能降到70%左右。
And those, if everybody dive bombed the price and it went down by a factor of two or three, yeah, it's 70%, the kind of the low end token pricing you're seeing here.
这种情况会真的发生吗？
Is that likely?
我不太确定。
And I I don't know.
我觉得这事不一定能成。
I don't know if that's gonna happen.
那如果势头继续下去，接下来的瓶颈会是什么？
So then, you know, if the momentum continues, what's the constraint?
在所有这些后期增长型投资中，你最终会发现，总可服务市场才是关键指标。
And one of the things you figure out in all these very late stage growth investing is in the end, the TAM is the criteria.
当你看这里花的钱，跟整个软件工程领域的规模相比，数额真的非常大。
What you start to look at the dollars being spent here relative to the size of software engineering, and it's pretty huge.
还记得我们之前讨论过吗？它会不会占到软件工程的20%？
Remember, talked about, is it going to be 20% of software engineering?
这跟杰森说的其实是一回事，对吧？
And kind of to Jason's thing, right?
大家都说，有三千万开发者呢，完全没问题。
Everyone goes, oh, 30,000,000 developers, it's all fine.
但如果你去查美国劳工统计局的数据——我上周末查了——美国实际只有180万开发者，其中只有20万在软件公司工作，另外60万在科技公司，比如惠普、IBM，还有整整一百万在像摩根大通、美国银行这样的金融机构工作。
If you go to the BLS, which I did last weekend, and say there were only 1,800,000 developers in The United States, of which only 200,000 of them work in software companies, another 600,000 work in tech companies, you know, HP, IBM, and a million of them work in, literally, JPMorgan, BofA, right?
这才是真正的总支出规模。
That's a total spend.
这个岗位的中位数年薪大约是14万美元。
Median wages are about 140 ks for that role.
总支出大约是两千五百亿美元，如果OpenAI和Anthropic这些企业的业务中大部分是软件相关，那它们已经接近占到百分之二十了。
It's a total spend of around $250,000,000,000 And if most of this OpenAI and Anthropic enterprise web, if a large percentage of that is software, they really are close to 20% already.
它们早就达到了。
They already hit it.
我有时候真想问，这到底发生了什么？
Part of me says, oh my god, what's happening?
因为这些公司并没有裁员百分之二十啊。
Because there hasn't been 20% of layoffs of these companies.
我的意思是，这个数字相对于美国整个软件开发的支出来说，实在太庞大了。
It's a huge my point is this: it's a huge number relative to the total software coding spend in The US.
我坦率地说，我对这个现象还没有得出结论。
And I freely admit, I don't have a conclusion from that.
我在想，这会不会是美国以外的收入？
I'm looking at it going, is it ex US revenue?
我看了Semianalysis的分析，觉得应该不是。
I read the semi analysis, I don't think it is.
我觉得大部分钱都花在美国本土。
Think a lot of it is in The US.
虽然美国也有采购用于海外的项目，但这个数字相对于你所能想到的最大数字——比如美国在软件工程上的工资支出——依然高得惊人。
There's some purchasing in The US for use in overseas, but it's an astonishingly high number relative even to the biggest number you can think of, which is the wages being spent on software engineering.
接下来十二个月，这个趋势会怎么发展，还挺值得观察的。
It'll be interesting to see how that shapes out in the next twelve months.
这个分析很有价值，因为它点出了一个悲观但现实的情况：不管这些工具多厉害，公司能在这上面花的钱终究是有上限的。
It is a great analysis because it's sort of, you know, it is definitely a bear case that no matter how great these tools are, we may soon hit the ceiling in terms of how much companies just on an absolute sense are willing to spend here.
你不可能花掉比收入还多的钱。
There is a you can't spend more revenue than you take in.
就算你的毛利率是100%，情况也一样。
And that's if you have 100% gross margins, right?
这背后有实实在在的物理限制。
There are just physical limits.
就算你把预算重新分配到IT部门，也还是有极限的。
Even if you reallocate budget to IT, there are just limits.
你说得对，杰森。
You're right, Jason.
而且说实话，我都不太愿意这么讲，因为我自己有个一贯的原则。
And you know, I almost don't like to posit it because one of my rules of thumb is this.
我常开玩笑说，简单来讲，牛顿运动定律是成立的。
I always joke, simplistically put, Newton's laws of motion are true.
运动中的物体会保持运动状态。
Things in motion stay in motion.
年增长十倍的东西，可能会降到八倍、六倍、四倍，但不会突然戛然而止。
Things growing at 10x year on year might decline to 8x or 6x or 4x, but they don't come to a grinding stop.
最惊人的是，当你以十倍的速度增长时，哪怕明年降到六倍、后年降到四倍，数字依然非常庞大。
And the amazing thing is when you're going at 10x, even 6x next year and 4x the year after that is a huge number.
我的意思是，今年你营收已经做到五六十亿了，明年再翻六倍，那就是三千亿。
I mean, you're at fifty, sixty build this year in GAAP revenue and you 6x next year, that's 300,000,000 billion.
这比全球软件开发的总工资还要高，对吧？
That's more than the total wage spent on software development, right?
所以你一看就会觉得，明年可能会发生两种情况之一。
So you look at it and go, one of two things is going to happen in the next year.
要么这件事的放缓速度，比你见过的任何事情都要快。
Either this thing decelerates faster than anything you've ever seen decelerate.
别以为这种情况会是第一种可能。
And don't think that's option A.
就像杰森说的，很多CFO现在都在想：我们以前花四百万在科技上，全是人力成本；现在花八百万，只有一半是工资。
Option B, to Jason's point, a whole bunch of CFOs are literally going, we used to spend 4,000,000 on tech, all of it in salaries, and now we spend 8,000,000 on tech, 4,000,000 on salaries.
天啊，我们得做点什么了。
WTF, we got to do something.
所以，正如我所说，这还没定论。
So it's just, as I say, no conclusion.
我只是在7月15号这天，脑子里冒出来的这些想法。
I'm just things I'm thinking about in July 15.
挺值得看看接下来会怎样。
It'll be interesting to see.
你就是杰森。
And you're Jason.
这些都没问题。
There are there are fine.
我的意思是，成为人类历史上增长最快公司的代价，可能是你比任何其他公司都更快地触及总可寻市场，这确实值得观察。
I mean, it may well be that the prize for becoming the fastest growing company in human history is you may hit TAM faster than any other company in human recorded history, and it'll be interesting to see.
你可能已经触及了手头资金的上限。
You may have hit the limits of how much money there is in the till.
还有另一个角度，我随便提一下，大家别笑话我。
You know, there's just one other vector, just to add it up for fun, and again, don't mock me if folks want to.
我们能花在工程上的钱到底有多少？
There's how much can we spend on engineering?
还有一套算法可以算一算。
And then there's another math you can do.
好的。
Okay.
每一段软件都在迅速变得具有自主性。
Every piece of software is becoming agentic very, very rapidly.
明白了。
Okay.
那软件支出是多少呢？
So what's software spend?
一万五千亿美元？
A trillion and a half?
你手里的数据更准确。
You have better than numbers.
一万亿美元左右。
A trillion plus or minus.
对。
Yeah.
一万亿美元。
A trillion.
所以我们大概可以得出一个结论，人们或许能接受将10%的毛利花在AI上。
So think we're coming to the rough conclusion that folks may tolerate, say, a 10% gross margin spend on AI.
咱们先假设这一点，好吗？
Let's just assume that for a minute, right?
那这就意味着又有1400亿美元可以流向Anthropic和其他公司，对吧？
That's another $140,000,000,000 that can go to Anthropic and Friends, right?
这比你刚才算的数字要多得多，但这个数字本身也有上限。
So that's significantly more than the math you just did, but it itself has a ceiling.
对吧？
Right?
不可能把全部1.4万亿美元都花在这上面。
It can't be all of the 1,400,000,000,000.
以现在的状况来看，10%这个比例还挺合理的。
10% feels about right today.
140%？那不可能。
There's 140%.
所有软件支出都会把营收的10%用在令牌上。
All the software spend is going to put 10% of their top line into tokens.
首先我完全同意，因为之前你问过这个问题，我当时答错了，脑子短路了，后来才想明白——我那个销售和市场助手到底花多少钱运行？
First of all, I totally agree because I mean, you asked that question a while back and I got the answer wrong because I was an idiot and then I processed it right, which is how much does it cost to run my little sales and marketing agent?
结果比你想象的低得多。
And it was way less than you thought.
这跟我们现在讨论的直接相关，原因就在这儿。
And the reason this is relevant here is this.
事实证明，除了编程之外，大多数代理型软件的令牌成本，其实像你说的那样，完全可控。
It turns out that outside of coding, most of the token costs for agentic software outside of coding are manageable to your point.
它们占10%，不是50%。
They're 10%, they're not 50.
所以你说得对，杰森，如果软件行业在非编程领域花上十亿、甚至万亿，那又是一千亿的开销。
So you're right, Jason, is that if the software industry is doing a billion in other kind of non coding stuff, a trillion, that's another 100,000,000,000 right there.
你说得没错。
You're right.
像Salesforce这样的公司，完全可以拿出营收的10%来购买token。
Like Salesforce could comfortably pay 10% of its revenue for tokens from someone.
他们不会付40%，因为他们的运营利润率只有22%。
They're not going to pay 40 because they only have 22% operating profit.
这相当于现在就有一千亿美元的收入触手可及。
That's a $100,000,000,000 of revenue that's almost accessible now.
接下来，这一切都会转向基因领域，对吧？
And then it's, I mean, it's all going to genetic, right?
我需要一个比我更聪明的人来判断，但这个数字在Anthropic的燃料消耗里占多大比例，我也不清楚，但肯定不小，一百亿美元呢。
So that I need someone smarter than me, but that could be a huge, I don't know what percent of Anthropic Fuel is, but it's a lot that 100,000,000,000.
没错，你说得对。
No, you're right.
其实就是那三大类。
It's it's with those really three buckets.
第一类是编程相关的。
There's the coding bucket.
然后是代理型软件的百分之十抽成，就像十年前亚马逊对软件收税一样，所有人都得用亚马逊，给我七个百分点。
You're Then there's the agentic 10% software tax, just like Amazon took a tax on software ten years ago, everyone's going to run on Amazon, give me 7%.
现在你说对了，所有人都会用大语言模型，那就得给我百分之十。
Now you're right, everyone's going to run using LLMs, give me 10%.
这又是一千亿美元的收入。
That's 100,000,000,000 more.
接下来你开始讨论，像 CoWork 这样的工具，会不会像当年的 Microsoft Office 一样，取代月薪两万美元的知识工作者？
And then the last thing you start talking is, does co work replace the knowledge worker at $20 like Microsoft Office?
但你说得没错，从宏观角度看，我还在琢磨这些数字，得再仔细算一算，回头给你反馈，不过你现在讨论的确实已经是天文数字了，远超其他人估算的规模。
But you're right, mean, the big picture, and I'm still wrestling through and trying to look at those numbers and I'll report back, but you're definitely at the stage where you're talking big ass numbers well to everyone else's numbers.
咱们走着瞧吧。
And we'll see.
这种向法律和生物科学领域的拓展，是不是因为他们已经意识到，在当前市场里，潜在的总市场规模快到顶了？
Is that movement into legal biosciences not just an appreciation of the fact that they are potentially hitting TAM ceiling limits in where they are already?
目前没有任何证据表明他们遇到了天花板，虽然从逻辑上讲，我确实希望有人别总说‘他们早就触顶了’。
There's no evidence that they are, even though logically, again, I want because I don't want someone to say, Hey, we always said they're hitting TAM.
所有现有的趋势都表明，他们并没有。
All the traction momentum seems to say they're not.
我看了看实际的市场，觉得在医疗领域，说来有点奇怪，我觉得Anthropic的创始人，就像DeepMind的创始人一样，真正驱动他们的其实是想治好癌症这种愿望——每次聊到他们的大模型，大家总拿这个当例子。
I just look at the actual market and I go, I personally think that on the medical side, I think oddly enough, I don't know if that's as much as I think the founders of Anthropic, just like the founder of DeepMind is very much motivated on the medical side by you know, the desire to everyone always references their world where cancer is cured when they're talking about their LLM.
所以我觉得，人们投身生命科学，三分之一是冲着市场扩张，另外三分之二其实是：我让这么多人失业了，至少得想办法让他们活下来。
So I think actually people's interest in doing life science is one third TAM expansion and two thirds, shit, I'm putting everyone out of job, the least I can do is keep them alive.
我想做点有意义的事，因为我想找到价值。
I want to do good because I want to have meaning.
这本质上是一种对意义的追寻，从DeepMind开始，一路延续到现在。
It's a search for meaning, you know, starting with DeepMind and going from there.
如果你觉得其他创始人不羡慕DeepMind、不羡慕德米斯那种拿诺贝尔奖的成就，那你就太天真了。
I mean, if you if you don't think the other founders wish they had a Nobel Prize like like DeepMind, like Demise, you're crazy.
那一定很棒。
That must be good.
那感觉一定很好。
That must feel good.
等等。
Wait.
等等。
Wait.
等等。
Wait.
杰森可能会因为……获得诺贝尔和平奖。
Jason's gonna get a Nobel Peace Prize for
不是每个人都能拿诺贝尔奖。
Everyone's not gonna get a Nobel Prize.
非常好。
Very good.
他和自己的公司一起努力过。
He's worked with his companies.
明白吗？
Okay?
但杰森确实该得。
But Jason should.
对。
Yeah.
没错。
Exactly.
影响投资组合。
Impact the portfolio.
我只等着拿到一份像样的offer，来自Anthropic的任何一份offer都行。
I'm waiting just to get an offer, a decent offer from any offer from Anthropic.
到现在连一个都没收到。
I haven't gotten a single one.
没人愿意请我去担任任何首席职位。
No one's offered to bring me in as a chief anything.
这简直是个笑话。
It's a joke.
杰森，你想要多少报酬？
What would your price be, Jason?
二十万。
200.
哇哦。
Wow.
税后要是低于这个数，就不值了。
After tax is not worth it otherwise.
而且我也不想给老板打工。
Plus, don't really wanna work for the man.
所以，如果我真的要接，最低也得是二十万，不然就算了。
So that'd be the I I if I really wanna do it, 200 would be the minimum I'd still
杰森，我得提醒你一下，根据我对Anthropic公司结构的了解，你其实不是给老板干活，是给女老板干活。
strongly the need to say to you, Jason, that based on my understanding of the Anthropic org structure, you in fact don't work for the man, you work for the woman.
对吧？
Right?
黛尼拉掌管一切。
Daniela runs everything.
这
It's
我觉得这是个中性词。
it's a gender neutral term, I I would say.
说到你的
To your
观点，对吧？
point, right?
接下这份工作，然后
Take the job and
200。
200.
这就对了。
There we go.
我们有个报价了。
We have a price.
而且还有风险。
Plus there's risk.
我可能五年都做不下去。
There's I might not work out for five years there.
也就是说，我有可能根本不会接手。
Like, there's a small chance I don't take.
杰克，这话不说出来会伤人，但我得说实话。
Jake, there's no other way of saying this without being hurtful.
我觉得你过不了性格测试。
I don't think you'll pass the personality test.
我的意思是，我知道我自己肯定过不了，但他们那边挺注重关怀和氛围的。
I mean, I know I wouldn't, but they're doing good, caring, all that stuff.
我觉得你直接就挂了。
I I think you just flunk out.
是的。
Yeah.
但有些人觉得我挺拼的，罗里，即使到了我这个职业生涯阶段。
But they some people think I'm a little driven, Rory, even at this point in my career.
对。
Yeah.
你确实很拼，但他们筛选的是那种愤世嫉俗的人。
You're driven, but they they they filter out for cynicism.
你根本过不了这一关。
You just won't make it true.
你不够理想主义。
You're just not idealistic enough.
我看看。
I look.
我还没拿到offer。
I haven't gotten the offer.
所以你是说，没错。
So you yeah.
从实际来看，你说得对。
Empirically, you're you're correct.
对吧？
Right?
我已经做了63期播客了。
I've done 63 of these podcasts.
我写过一万篇博客文章。
I've written 10,000 blog posts.
但Anthropic从来没给过我任何offer。
I haven't gotten one offer from Anthropic.
这事儿根本不会发生，是吧？
It's it's not coming, is it?
这是年龄歧视。
It's age discrimination.
你说得对。
You're right.
支撑基础模型的基础设施，比如SK海力士，还有它的纳斯达克上市。
The infrastructure that powers foundation models, SK Hynix, NASDAQ listing.
你想聊这个吗，Rory？
You wanna go there, Rory?
很好。
Great.
我的意思是，确实如此？
I mean Yeah?
不是。
No.
我只是说，很好。
I'm just saying great.
我有点紧张，要
I'm nervous to
现在提个话题，杰森，他就皱眉头，我心想：天哪。
suggest a topic these days, Jason, because he just scowls, and I'm like, oh, fuck.
不行。
No.
不行。
No.
不行。
No.
我以为是
I thought that
他是个赢家。
He's a winner.
不行。
No.
很明显，大家都知道，内存行业就三家大公司。
Obviously, the stuff everyone knows, there are three memory companies.
过去一年，他们从人工智能资本支出热潮中获益巨大。
They have been a huge beneficiary of the AI CapEx boom in the last year.
SK海力士的股价已经涨了六倍。
Those stock SK Hynix is up six x.
三家里面有两家在韩国，分别是三星和SK海力士。
Two of the three are based in Korea, Samsung and SK Hynix.
美光公司总部在爱达荷州，美国制造。
Micron is based in Idaho, Go America.
说到底，这属于寡头垄断，他们赚得盆满钵满。而且说实话，美国人想买韩国股票特别麻烦，所以即使现在SK海力士的ADR在美国上市交易，股价还是比在韩国本地高20%，就是因为美国人开不了韩国的证券账户，买不了本地股票。
And the truth is it's an oligopoly, they've made out like bandits, and it's actually very hard to buy stocks in Korea, which is why even today, interestingly enough, after the ADR is traded in the US stock exchange, it's still trading at a 20% premium to the same stock in Korea because it's just hard for an American to open a brokerage account in Korea, buy the Korean thing.
所以我觉得他们能进入资本市场，这事儿挺好的。
So I I think it's just good that these guys are accessing the capital markets.
所以对所有人来说
And so for everyone to
大家都知道，SK海力士在纳斯达克上市，作为外国公司大规模融资，市场反应热烈，收盘时股价飙升了13%。
know SK Hynix, Nasdaq listing, large seller by a foreign company, had a great reaction, pops 13% at close of day.
然后又回落了。
And then back down.
说实话，我最近几天过得特别糟，昨天和前天都挺难熬的，对吧？
And then I mean, I've had a very difficult day, I think yesterday, couple of days, right?
是的，确实容易让人觉得，天啊，AI资本开支的泡沫要破了。
Yeah, it'd be tempting to go to some kind of, my god, the AI CapEx bubble is busting.
迟早会的。
And at some point, it will.
但我认为这里面有很多技术性因素在起作用，毕竟，两家内存公司主导了韩国股市，这并不意外。
But I actually think a lot of that there's a lot of weird technical stuff going on there because no surprise, the two memory companies dominate the Korean stock market.
三星和海力士这两家公司，市值加起来占了整个韩国股市的60%，差不多是韩国GDP的120%。
They are the two companies, Samsung and Hynix, are 60% of the stock market and roughly 120% of the GDP in terms of market cap.
它们的体量实在太大了。
They're just huge.
股价波动也特别剧烈。
They're highly volatile.
韩国股市里散户的交易特别活跃。
There's lots of retail action in Korea.
所以韩国股市的日内波动非常大。
So there's just a lot of intraday volatility in the Korean stock exchange.
那边的财富波动简直像赌场一样疯狂。
It's called kind of crazy casino level wealth there.
对吧？
Right?
尽管最近几天股价下跌了，但很难说清楚这到底是DRAM的正常回调，还是纯粹在玩一些奇怪的交易把戏？
So even though the stocks moved down in the last couple days, it's not as clear as it fundamentally, oh, DRAM's correction, or is it just weird kind of trading shit?
至于DRAM这块，说实话我没啥明确看法，但那三家公司的市盈率都在五到八倍之间。
But I think on the DRAM thing, I mean, the fun bet there is I don't have an opinion on it, but those three companies are all trading at, you know, five to eight PEs.
也就是说，股价便宜得离谱，任何一点利好都能让股价涨起来，这就是看涨的理由——这种行情能持续下去。
In other words, dirt cheap and 1¢ in any kind of screen, and that's the bull case, this is going to last.
而看跌的理由是，这类企业历来属于资本周期性行业，一旦新增产能上线，价格就会崩。去年价格涨了六倍，才让利润率从2023年的负数飙升到70%，成了全球最赚钱的公司。
And the bear case is these have traditionally been capital cyclical businesses, and the minute more capacity comes online, prices have gone up six x in the last year, which is why operating margins have gone from negative in '23 to like 70%, the most profitable companies on the planet.
三星现在是全球最赚钱的公司，但迟早会回落。
Samsung is the most profitable company on the planet right now, and they're going to go back down.
这就是其中的利与弊。
So that's the kind of pro and con.
我觉得这次周期可能会比往常长一点，但最终还是会回调的。
I think it'll last a little longer than the typical cycle, but in the end, it'll I mean, at some point, it does correct.
内存厂商不可能永远保持70%的净利率——不是毛利，是净利率。
You're not going to have the memory makers earning 70% net margins, not gross margins, net margins forever.
也许罗里，你可以质疑一下这个观点。
Maybe Rory, you'll challenge the connection.
我觉得挺有意思的是，IBM 这周表现很差，这可是一家老牌科技公司。
I thought it was interesting that IBM this week, pretty old tech company, right?
我记得好像是19世纪就成立了。
I think founded in the 1800s or something like that.
结果业绩大幅不及预期，股价直接跌了20%。
Mean, they had a huge miss, stock crashed 20%.
其中一部分，说实话，可能只是借口。
And some of it, listen, is probably an excuse.
当我们业绩大幅下滑时，作为CEO就得找些理由来解释。
We have to make up excuses as CEOs when we have a huge miss.
但他们确实在整合内存资源。
But they blend memory.
他们说，内存占用了CEO预算的很大一部分。
They said memory is taking so much of the CEO's budget.
大家都在抢购内存，生怕价格再涨上去。
There is such a scramble to buy memory before it goes over up.
他们根本没从我们这儿买主服务器或者服务器设备。
They weren't buying our mainframes or our servers or anything from us.
CEO们慌了，赶在内存价格进一步上涨前疯狂采购。
CEOs went into a panic to buy memory before it would get even more expensive.
我肯定这是个借口，但里面也确实有点道理——这笔钱总得从别处省出来吧？
I'm sure it's an excuse, but I'm sure there's a germ of truth that that money's got to come from somewhere, right?
IBM说这笔钱是从我们这里挤出来的。
And IBM said it came out of us.
我觉得你说得对，杰森。
I think you're right, Jason.
其实我不这么认为。
I actually don't.
我觉得不止是德国式思维，钱总得有个出处。
I think there's more than a German Look, money has to come from somewhere.
如果你在大笔投入大语言模型，同时又因为这些模型推高了内存价格，导致你不得不花更多钱买内存，那你肯定不想在季度最后一天被排在采购清单的末尾——我敢打赌，CFO正盯着账本问：咱们能砍掉哪一笔？
And if you're spending a lot on LMs and then if you're spending and you're having to spend even more on your memory needs because the LLMs themselves are bidding up the price of memory, You don't want to be the last item on the purchase order list on the last day of the quarter, because I'm sure there's a CFO sitting there going, what can we cut?
因为这个季度我们的资本支出总额是有限的。
Because we only have a finite CapEx budget this quarter.
你刚才说的关于SaaS模块供应商的事，我也同意。
And it's what you said also about what's going to happen to the module SaaS vendor.
他们不是要完全砍掉预算，而是会说：各位，我们今年的技术预算是一千万美元，去年在AI上只花了100美元，今年涨到了三百万，所以我们得省出两百九十万出来。
It's not that they're going to vibe coat it away, it's as much as, you know, guys, we have a tech budget of $10,000,000 and last year we spent $100 in AI, this year we're spending $3,000,000 so we need $2,900,000 in savings.
结果发现，我们其实不需要五个生产力应用。
And it turns out we don't need five productivity apps.
等你真正用上 Teams，你会喜欢上的。
You'll love Teams when you get to know it.
也许我们也会继续用现有的，而不是从你们这儿买那个模块。
And maybe we'll stick with what we have too rather than buy that module from you.
没错。
So, yeah.
而且今天 IBM 的情况确实大幅下滑了。
Mean and IBM did come up pretty drastically today.
Roy，我想岔开话题说点别的，因为你刚才说，除了 AI 还有别的世界。
I'm gonna go off on a bit of a tangent, Roy, because you said, oh, there's a world outside of this like AI.
这周早些时候，我和 Jason 在聊天，聊到 Jason Calacanis 的投资动向，他说了这么一件事。
Me and Jason were messaging at the earlier this week about Jason Calacanis' syndicate or investing activity where he basically said, hey.
因为大家需求强烈，我们决定把投资重点从早期项目转向一些大公司的后期成长机会。
Due to popular demand, I'm moving where we'll be investing from very early to later stage growth opportunities in some of the big names.
这是否反映了当前风投市场的现状，也说明了时代的变化？
Is this symbolic of where the venture market is at today and a sign of the times?
杰森，既然我们刚才在聊这个，不如你先说说，让罗里也
Jason, given it was us talking about it, why don't we start with you so Rory can
能有时间
have some
我觉得他说这话挺有意思的，因为你知道，杰森确实有不少不错的投资。
I just thought it was interesting for him to say that because, actually, you know, Jason's got some pretty good investments.
他做了很多投资，对吧？
He does a lot of investments, right?
他手里的项目都不错。
So he's got some pretty good ones.
我还以为他的意思是，我们要新设一个附属基金呢。
I thought the message would be we're adding an annex fund, right?
或者我们要再加点投入。
Or we're going to do a little bit more here.
但说他几十年来投入了如此多精力，去打造一条替代路径，大量开展联合投资，最终全部转向成长型项目，这意义非凡。
But to say that he was put so much energy over decades, right, into creating alternative path to do massive number of syndicated investments to apparently tilting all into growth.
在一个细分领域里，这确实是时代变迁的强烈信号。
In a niche, it's a huge sign of the times.
这比慢慢调整更说明问题，是彻底的转向。
It's a huge sign of times rather than easing into it.
当你手握好几个十亿美元级别的成功项目，还拥有实质性股权时，却选择完全转型，这在我看来非同小可。
Completely switching when you've got multiple billion dollar winners that you have material ownership in is, to me, it's a big deal.
对吧？
Right?
我们已经知道，这种机会就像坐等送上门的傻瓜，持续了六十多周了。
I mean, seats for suckers, as we've known for sixty something weeks.
真正能说明问题的，是YC放弃早期投资，而不是别的什么。
This is the this is the other than YC dumping early stage, that would be a real sign.
对吧？
Right?
YC发个帖子说只做成长期融资，这真能掀翻整个行业吗？
YC issuing a going on threads and telling everybody we're only doing growth rounds that that would that would rock the world?
但我觉得这至少是时代变化的信号。
But I thought it was at least a sign of the times.
现在所有东西都在涨。
Everything's up.
当所有项目退出估值都能超过两百亿时，我甚至有个投资，投了一亿美元，花了好多年才做到现在这个地步。
And when everything can exit north of 20,000,000,000, I even me like I have an investment with $100,000,000 position that took me years and years and years to get to right now.
结果一看，有人进来直接投一亿美元，我就想：这合理吗？
And then just watching a growth round where someone comes in and invests 100,000,000 like, what?
我何必经历那么多波折、熬那么多年、在董事会里孤军奋战，别人一纸支票就甩一亿美元？
Why did I bother all the drama, all the years, all the being the only guy at the board meeting when someone just drops $100,000,000 check?
如果这种项目能再翻十倍，那我早早就冲进去，到底图个啥？
If that if that can grow an order of magnitude, what's what's the point of being the guy any earlier than there?
对吧？
Right?
不，没什么存在主义的意义。
No no existential point.
因为他热爱这份事业。
Because he loves the craft.
我不爱这行。
I don't love the craft.
谁会真的爱这行？
Who the hell loves the craft?
大卫·弗兰克尔。
David Frankel.
我昨天刚采访过他。
I interviewed him yesterday.
他确实爱，这人真的热爱这行。
He does love that man loves the craft.
更多
More
他可太爱这行了，特别享受。
fun, He like loves the craft so much.
是的。
Yeah.
对。
Yeah.
让我换他来上这个播客吧。
Replace me with him on this pod.
那家伙就算住进养老院，照样能谈成大单。
That man that man will be doing deals from the old folks home.
对吧？
Right?
而且还能拿到好项目。
And he'll be getting good ones.
而且还能拿到好项目。
And he'll be getting good ones.
我会把钱交给他。
I would give that man my money.
我真的很喜欢大卫。
I love David.
我也一样。
So do I.
刚才我是在为他热爱这份事业辩护，我其实想说，
Please that was me defending his love of the I like, I
我觉得他是真心热爱这门手艺。
feel like he loves the craft.
不过要说得更清楚一点，我明白你的意思。
But to give it good framing, I hear you.
但这个说法——早入行还是晚入行，
But this whole idea of you know, venture early versus venture late.
对，这确实是个关键问题。
Yeah, well, that's really the question.
你怎么看这个问题？
How do you think about that?
对吧？
Right?
我能给个老生常谈的答案，你也知道，我一向喜欢留个带点讽刺的套话——投资里，当一件事看着太容易的时候，恰恰就是你该避开的时候。
And, you know, the trite answer I could give, and you know, I always like to leave the snarky trite answer is generally when something looks easy in investing is precisely the time you shouldn't do it.
现在晚期投资看起来特别简单。
And late stage looks very easy right now.
这大概就是我骨子里的逆向思维吧。
So that's kind of the natural contrarian in me.
但我觉得，哈利，真正关键的是你得考虑三件事，对吧？
But I think really what's going on, Harry, is there's three different things you have to think about, right?
我在这儿列出来。
And I put them out here.
第一，私人公司融资有个大的长期趋势，就像我们说的，过去二十年，早期项目的规模其实没怎么变大。
One is, there's this big secular trend in private company financings, which is, as we said, the early stage business hasn't gotten that much bigger in the last two decades.
但更重要的是，由于公司上市变少了，像OpenAI这样的公司又迅速壮大，导致晚期投资的规模扩大了五倍。
But on top of that, a 5x bigger late stage business has emerged as companies, a) haven't gone public as much, and b) in the case of something like OpenAI Tropics just became bigger quicker.
因此，晚期投资领域又多出了一块全新的业务。
So there's been this new business for late stage on top.
回头看看十五年前，如果你要选一个领域：一个是已经存在三十年、有两百个强劲对手的成熟行业，另一个是当时还不存在、竞争者寥寥的新领域——在其他条件相同的情况下，从那时起，晚期投资显然看起来是最容易入手的选择。
And there's no doubt, looking back fifteen years, if you said you can enter the one business that's been around for thirty years and has 200 very good competitors, or you can enter this new business which didn't exist before and has relatively few competitors, all other things being equal, that late stage opportunity in 2005 from then on looked, in retrospect, looked like the easiest place to play.
所以，这确实是一个长期的趋势。
So there's definitely this kind of secular trend.
你还没说到第二点或第三点呢，可你已经摇头了，Harry。
You look like not even getting on to two or three, but you're shaking your head already, Harry.
不是的。
No.
不是的。
No.
我完全明白你的意思。
I I I totally get it.
我接触过不少持怀疑态度的年长者，他们对这事说得挺尖锐，但我也不是在对你这么说。
I I I've had a lot of cynical, kind of respectfully older people talk to me about this, and I'm not suggesting that to you, by the way.
你本身就很 cynical，Harry，不过这没关系。
I'm cynical a lot of that in you, Harry, so you are, but it's okay.
明白。
Yeah.
我每次都会看到这种现象，真的。
And and I was like, oh, I'm seeing this every cycle.
每一轮我都看得清清楚楚。
I see this every cycle.
别说了。
Stop.
我得直接回你一句，因为我刚才说的本来就是长期趋势。
I'm going to push right back in your face, because I didn't even I actually made a secular comment.
我马上要说的才是周期性的问题。
I'm going to make a cyclical comment in a second.
对吧？
Right?
我刚才说的是一个非周期性的观点，这一点非常明确。
I made a secular comment, which is really precise here.
其实我是在同意你的看法，哈里。
What I I'm actually agreeing with you, Harry.
我想说的是，这个类别在以前根本不存在。
I'm saying the secular trend is this category didn't exist.
但现在，它会在每个周期中都出现。
And now it will exist across cycles.
我刚入行的时候，风投圈里根本没有‘无一人参与’这种说法，别说1994年了，就连2004年，也没人会写一亿美元的后期投资，可现在这已经成了常态。
This new business called no one in venture when I started, now forget '94, even in 2004 was writing $100,000,000 late stage checks, and now it's a thing.
这是一种全新的金融产品——给估值已经上十亿美元的公司投一亿美元的私人资金。
It is a new class of financial product, private $100,000,000 checks in companies already worth a billion dollars.
这是新出现的现象。
It's a new thing.
所以我不是在当老古董，这纯粹是周期性的。
So I'm not doing the old guy, it's just cyclical.
我待会儿就来谈这个。
I'm actually going to do that in a second.
但现在我想说的是，回头一看，这确实成了风投领域一个全新的类别。
But right now, I'm saying, in retrospect, it actually is a whole new category of venture.
它并不是在取代早期风投。
It's not replacing baby venture.
而是在原有基础上新增了一个完整的类别。
It's kind of adding a whole new category on top.
这个世界需要这个类别，因为这些公司都没有上市。
And the world needs that category because these companies aren't public.
所以像Altimeter、Thrive，还有Andreessen那些做后期投资的，其实都替代了本该上市的公司，对吧？
So the altimeters, the thrice, the Andreessen late stage part of their business, all those guys have replaced what would have been public companies, right?
所以这并不算愤世嫉俗。
So that's kind of not cynical.
现在，我做个 cynical 的评论：在长期趋势之外，确实还叠加了一个周期性因素。
Now, the cynical comment I'll make is on top of a secular change, you do also have a cyclical thing.
毫无疑问，离公开市场越近，这个周期性波动就越明显——有时候后期投资看起来特别棒，有时候又简直一塌糊涂。
And there's no doubt that the closer you are to the public markets, the more the cyclical thing happens, which is there's times when the business late stage business looks amazing, and there's times when it looks crap.
这主要取决于市场环境的变化，尤其是公开市场的表现。
And it's really very much dictated to by change plus the public markets.
但我个人觉得，长期趋势比周期性波动重要得多。
So there is I actually think the secular part is more important than the cyclical part.
你只需要对周期性的波动保持清醒认识就行。
You just got to be aware of the cyclical stuff.
2022年，其实是个特别适合投钱的年份。
In 2022, turns out was a great time to write checks.
2020年呢，就没那么理想。
2020, not so much.
那些在2022年抄底投了低价轮次、然后迅速加码的人，现在觉得自己聪明透了。
Mean, the people who did that kind of cheap round and ramp in 2022, they feel pretty damn smart.
所以，不，我的意思是，虽然高峰时期确实很难熬。
So no, I mean, it's just but there's no doubt at the height, it gets tough.
我注意到的一点是，我们以前从未见过二级市场像今天这样流动性和成熟度这么高，能这么轻松地退出像现在这些知名高增长公司。
The thing that I I do see is just I don't think we've ever seen the secondary market be as liquid and mature as it is today in terms of ability to get out of great named, high growth companies like we can today.
就拿我手里的所有头部项目来说，现在随时都能退出，每天都有买家上门。
Like, in all of my top names, I can get out of them today with ease, and every single day I have buyers for them.
我从没见过这么活跃的二级市场。
I've never seen such a liquid secondary market.
所以，为什么是现在，这一点特别关键。
So the why now is always really important.
退出回报的规模比以往任何时候都大。
Outcome scenarios are bigger than ever.
在很多情况下，这些退出机会发生的速度也比以往快得多。
The speed of those scenarios happening faster than ever in in a lot of cases.
你的难题是，四年里有六百亿次意外。
Your curse is 60,000,000,000 accidents in four years.
而且在这个过程中，你有机会参与比以往任何时候都更活跃的二级市场。
And then along the way, you have the chance to have secondaries far more liquid than ever.
整个生态系统的重大变化，让如今的情况远胜过以往的周期。
There are meaningful changes to the ecosystem, which make now today significantly better than prior cycles.
我同意。
I agree.
需要澄清一下，这并不是说公司上市更晚了，因为实际上这些公司成立的时间并不长。
It's worth pointing out just to be precise that that's not a stay private for longer comment, which sometimes people say, because in fact, these companies haven't been around a long time.
你说得对，真正不同的是，现在有些公司五年内就能从零做到上百亿美元的价值。
You're right, that what is and it's a different phenomenon is that you're now having companies go 0 to billions of dollars in value creation in five years.
所以并不是它们没时间上市。
So it's not a state they don't have time.
有人会说，SpaceX 私营了很长时间。
I mean, people say, look, SpaceX stayed private a long time.
但 OpenAI 还没来得及把系统搭建好，对吧？
OpenAI hasn't had time to get its systems together yet, right?
它们不是上市，而是因为没必要急着上市。
They're not public because they're staying private for longer.
它们上市，是因为五年前根本没什么动静，对吧？
They're public because, Jesus, five years ago they were doing nothing, right?
我同意你的看法。
So I agree with you.
现在出现了一种新型企业，赢家很少，四千家初创公司里只有四十家真正重要；如果你希望成为其中之一，又错过了早期机会，那在后期投一笔钱，确实是参与这场游戏的一种方式。
There's this new class of businesses, there's a small number of winners, and if there's 4,000 early stage companies and only 40 of them matter, and you want to matter, and you miss them at the early stage, there's no doubt that sticking some money in the later stage is one way to play the game.
是的。
Yeah.
但你也可以投入更多资金。
But you can also move more money.
所以，我觉得杰森的做法完全合理。
So, yes, I thought it was entirely logical from Jason.
说实话，我觉得他这么做再合适不过了。
I think it's exactly what he should be doing, to be honest.
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。