Grok 4惊艳亮相，苦涩一课，埃隆的第三方势力，AI浏览器之争，最高法院支持总统裁员决定

我有一个非常好笑的故事要告诉你，Jason。

I have a very funny story to tell you, Jason.

你去哪了？我一直试着发短信给你。你一直不在线。发生什么事了？你去哪了？

Where have you been? I've been trying to text you. You've been offline. What's going on? Where have you been?

我一直在拼命工作，但昨天我不得不去准备一些周日的会议，这些会议我不能告诉你。不过我和Nat去了帕萨拉克（Pasa Laqua），它位于科莫湖（Lake Como），这地方太棒了。那里的环境非常漂亮，酒店也很惊艳。如果你有机会去科莫湖的话，无论如何都推荐你们去帕萨拉克（Pasalacua）。

I've been working feverishly, but yesterday, I had to go to prepare for some meetings that I have on Sunday, which I can't tell you about. But Nat and I went to Pasa Laqua, which is in Lake Como, which is an I mean, it's stunning. The grounds are stunning. The hotel is stunning. If you have a chance to go to Lake Como anyways, this is us at Pasalacua.

那边那个漂亮的女士是谁？是这家店的老板还是什么人？那是

Who's the beautiful woman there? Is that the woman who owns it or something? Is that

女王吗？不是。最精彩的部分是我们玩得特别开心。你知道他们有一个留言簿可以写留言吗？对吧。

the queen? That's not. But the best part is we had such a good time. You know how they have, like, a registry book to leave a message? Sure.

所以我留了一条信息。

So I I left a message.

来了，这里真是一个超越所有期待的绝妙之地。

Here we go. A truly magnificent place above and beyond any expectation we have.

往下看，帮我看看下面的内容。

Go below go below that stuff for me.

谢谢你们，我们把一切都安排妥当了。一切都很顺利，Berg。

Thanks for We took everything to free. Took everything to free, Berg.

太好了，真棒。Jason，那些衣架。好，行李袋。

Great. Awesome. Jason, the hangers. Okay. The bags.

洗衣袋。

The laundry bags.

你有没有拿到

Did you get to

长袍，

base the robes,

拖鞋绳？所有东西。太棒了。好了，听我说。

the slipper ropes? Everything. Fantastic. Alright. Listen.

你要

You're gonna

发送

have to send

账单给自由鸟们。这简直令人难以置信。听好了。我们本周有一个非常精彩的小组讨论。现在正值夏季。

a bill to the free birds. It's absolutely amazing. Listen. We've got a great panel this week. It's the summer.

事情进展缓慢。有些人很忙。我想我们的恐慌症王子、亲爱的科学苏丹他正在响铃处待命。萨克斯很忙，这周无法参加。代替他的是另一位出色的贝宝校友，而且我得说一句，还是共和党支持者的海斯·鲁鲍格。

Things are slow. Some people are busy. I think our prince of panic attacks, our dear sultan of science is he's at the beep. Saxe is busy. Couldn't make it this week In his place, another brilliant PayPal alumni and, dare I say, GOP supporter, Heath Roubaughey.

先生，你好吗？

How are you, sir?

再次与您相聚感到非常荣幸。

Pleasure to be with you again.

很高兴见到你。我猜你现在应该是在美丽的佛罗里达，或者意大利某个地方吧？对吧？

Nice to see you. And I'm assuming you're in gorgeous Florida or somewhere in Italy. Yeah?

其实我现在在纽约。

I'm actually in New York.

哦，我的家乡。那里安全吗？情况还好吗？我妈没追着你在街上跑吧？没有

Oh, my hometown. Is it safe? Is it okay? Mom don me chasing you down the street? Not

暂时没有，不过挺安全的。

yet, but it's safe.

你有没有没收你的

Did you seize your

很安全。你的资产呢？

It's safe. Your assets?

没有？安全。是的。现在很安全。我们11月4号见。

No? Safe. Yeah. It's safe right now. We'll see you on November 4.

你知道，正如你可能听说的，7月4日是有记录以来第一次纽约当天没有发生枪击或谋杀事件。所以目前来说情况相当不错，但我们也许会很快离开纽约。

You know, as you probably heard, on July 4 was the first time in recorded history that there were no shootings or no murders in New York on that day. So right now, things are in pretty good shape, but we maybe we maybe leaving New York quickly.

是的。如果你在那里有房产的话，最好考虑卖掉，因为曼达米会把它没收，然后给你改成药店。没错，到时候就是曼达米药店了。特拉维斯·卡兰尼克又回到我们身边了。

Yeah. You're gonna probably wanna sell that place if you got one there because Mamdami is gonna seize it and turn it into a drugstore for you. Yes. It's gonna be the Mamdami drugstore. Travis Kalanick is back with us.

最近怎么样，亲爱的？

How are doing, bestie?

挺好的，挺好的。

Pretty good. Pretty good.

是的。这是你第二次参加这个圆桌讨论。是的，也是第三次上这个节目。当然了，你还在峰会上发过言。

Yeah. Second appearance here on the roundtable. Yeah. And third time on the show. Of course, you spoke at the Summit.

你一直在忙Cloud Kitchens的事情吧？是的吧？有很多令人兴奋的事情在进行中。

You've been busy with Cloud Kitchens. Yeah? Lots of exciting things going on.

哦，有很多事情。很多很多事情。机器人正在接管一切。我们正在推出机器人，正在部署机器人。

Oh, lots of stuff. Lots of stuff. The robots the robots are taking over. We're we're rolling out we're rolling out robots.

是的。TK，你能告诉我们你现在是否和 Pony AI 有合作吗？或者这只是猜测？

Yeah. TK, can you tell us what you're doing with this pony AI or not? That's speculation.

看，你知道的，很明显，正如我们在美国所了解的那样，当然是这样。等等，当然要等等。

Look. You know, obviously, it's autonomy as we you know, in The US, we have, of course Wait.

有没有人可以给那些可能不太了解情况的人介绍一下已经宣布的内容，或者至少说明一下为什么

Do wanna just frame for people that don't that may not be up to speed what was announced or at least Why

你自己来说说吧？

don't you frame it?

你为什么不

Why don't you

Pony AI 是一家自动驾驶公司，专注于无人驾驶技术。它是为数不多的几家真正将汽车开上路的公司之一。他们总部位于中国，在中东地区也有很多业务。此外，他们还与一家名为Uber的配送公司达成了合作，这家公司你可能听说过。

frame AI is an autonomous company doing self driving. It's one of the few players that actually have cars on the road. They're based in China. They've got a lot of operations in The Middle East. They've got a deal with a delivery company called Uber, which you might be familiar with.

好的，那么

Okay. So

看，这个交易基本上是你要和Uber合作，授权Pony的技术，并且实质上创办一家竞争对手，我想，针对Waymo和特斯拉？

look. Well, deal was basically that you would partner with Uber, license in the Pony technology and essentially start a competitor, I guess, to Waymo and Tesla?

让我来谈一下这个问题。好的，那么在美国，我们有Waymo。我们在旧金山、洛杉矶、奥斯汀看到Waymo，很快会进入迈阿密、亚特兰大和华盛顿特区，他们甚至在谈论纽约。

Let me work on this one. Okay. So so in The US, we have Waymo. We see the Waymos in San Francisco, Los Angeles, Austin, coming soon to Miami, coming soon to Atlanta, coming soon to DC. They're They're even talking about New York.

特斯拉有点像你知道的那样，他们用一种艰难的方式来做这件事，典型的埃隆风格。比如，让我们以一种根本性的方式去做，彻底一点。但目前还不清楚什么时候能越过临界点。当然，他最近在奥斯汀启动了一个半试验性质的项目，但除此之外没有其他替代方案。因此，一些希望确保存在替代选择的人就主动联系了我们。

Tesla's sort of like the, you know, they're doing it the hard way, you know, classic Elon style. Like, let's let's do this sort of in a fundamental holy shit, let's go all the way kinda approach. And it's unclear when it gets over the line. Of course, he launched sort of a a semi semi pilot of sorts in Austin recently, but there's no other alternatives. So what happens is is some of the folks who are interested in making sure there are alternatives have reached out.

他们联系了我，并开始了一些不同的讨论，因为他们觉得，Travis，你以前很早就涉足自动驾驶领域，在2014年初的时候推动了Uber的自动驾驶项目。也许这里可以做些事情来创造一些选择权。比如，我对食品方面当然非常感兴趣。我经常谈到自动化卷饼是个大事，因为如果你能实现厨房的自动化，食物生产的自动化，然后还能实现食品物流的自动化，就能大幅降低食品成本，这当然是我非常关心的事情。当然，也有些人希望看到移动领域的自动驾驶。

They they've reached out to me and there are different discussions they get going because they're like, Travis, you did autonomy way back in the day, got the Uber autonomous stuff going in 02/2014. Maybe there's something to do here to create optionality. Maybe, like, I'm, of course, very interested on the food side. I talk about autonomous burritos being a big deal because if you can automate the kitchen, the production of food, and then you can automate the sort of, logistics around food, you take huge amount of costs out of the food, out of what's going on in food, and that's of course near and dear to my heart. There's folks, of course, that want to see autonomy and mobility.

这是个现实的问题。也许或者可以说，如果你解决了自动驾驶问题，你可以将其应用到这两个领域。所以有很多人对运输东西感兴趣，运输食品，运输人员。如果有一些自动驾驶技术可能我会参与其中，它可能适用于许多不同的领域。所以我收到了一些主动的提议。

That's a real thing. It may be that or I would say if you get the autonomy problem right, you can use it to apply to both problems. So there's a lot of folks interested in moving things, moving food, moving people. And if there is some kind of autonomous technology that maybe I get involved in, it might apply to a bunch of different things. And so I've got some inbound.

就这样说吧，目前还没有真正的交易达成，但确实有一些主动的提议。我认为有关这些提议的一些消息可能会或不会发生。这样说可能是最合适的。刚才我说得比较啰嗦，下次我会尽量简洁一些。

Let's just put it that way. There's no there's no real deal right now, but there is definitely some inbound. And I think there is some news about some of that inbound that may or may not be occurring. That's probably the best way to put it. As long winded, I'll try to tighten that up next time.

不，不，我觉得首先在这里给大家一个整体的介绍非常好。感谢你与我们分享。大家都知道你一直在做一个碗构建器（bowl builder）。

No. No. I think it's great to get the overview here first on all in. Thank you for sharing it with us. And everybody knows you have been doing a bowl builder.

是的，Lab 37，我想是这个名字。你可以把它显示在屏幕上。不确定它的现状如何，之后我会让你继续提问，Chamath。但我认为这里有一个相当有趣的概念，就是把碗建造器放进一辆自动驾驶汽车里。

Yeah. Lab 37, I think it's called. You can throw it up on the screen. Not sure what the status of it is, and then I'll let you go, Chamath, with your follow-up question. But I I think there's a pretty interesting concept here of the bull getting built and then put into a self driving car.

那台机器看起来很大，但实际上只有60平方英尺。那张照片让它看起来像个怪物。它其实是一台60平方英尺的机器。想象一下经营一个类似Sweetgreen品牌或Chipotle品牌的餐厅，让它为那些看到后会问‘这是什么东西？’的人们带来生机。

Now that machine looks huge, but it's actually 60 square feet. That picture makes it look monstrous. It's a 60 square foot machine. Like, imagine running, like, a Sweetgreen like brand or a Chipotle like brand and just making it so it comes to life for people who who, you know, are like, hey. What is this thing?

想象一下，你只需在线下单，就能准确地得到你想要的碗装食物。实际上，这台机器可以同时运行多个品牌，并且已经在这样做了。你可以定制自己想要的碗，选择任何配料。如果你看底部，看到那些小的白色砖块吗？那是用来在分配器下方运送碗的装置。它会填满配料，然后给碗加上酱料，再盖上盖子。

Imagine you just order online exactly the kind of bowl you want, and actually, this machine could run like many brands at the same time and and does. You build the bowl you want, whatever ingredients, it sort of if you look at that bottom, you see those little white bricks at the bottom? That's what carries the bowl underneath dispensers. It fills up. The machine puts it sauces the the bowl, then it puts a lid on it.

接着，机器将碗放入一个袋子中，在袋子里放上餐具，密封袋子。袋子通过传送带传送到另一台机器，我们称之为AGV（自动导引车），它将碗送至前厅。碗被放入储物柜中，DoorDash或Uber Eats的快递员会在摄像头前晃动他们的应用程序，对应的储物柜就会打开，里面就是他们应该取走的食物。这样一来，就省去了很多我们所谓的组装成本。

It takes the bowl, puts it in a bag, puts utensils in the bag, seals the bag, and the bag goes down a conveyor belt where then another machine, what we would call an AGV, takes the bowl to the front of house, the bowl gets put into a locker, the courier via DoorDash, Uber Eats courier, will wave their app in front of a camera and it will open up the locker that has the food that they're supposed to pick up. So it just it takes out a lot of what we would call the the cost of assembly, which is

而且减少了错误。对吧？

more And reduces mistakes. Right?

很难

It's hard for

我们会犯错。是的。

us to make a mistake. Yeah.

我们知道每种配料的确切克数吗？你能得到的就是你本应得到的分量。因此，你能获得更高品质的产品，同时大大降低了成本。你可以想象，最终也会有相应的快递员参与其中，比如说实现自主配送的卷饼。

Do we know exactly how many grams of every ingredient are put in? That's exactly what you're supposed to get. And so you get a higher quality product. It takes a lot of the cost out. You imagine ultimately that's going to be there are gonna be couriers with that as well that, you know, like to say autonomous burritos.

比如，Waymo会不会运送一个卷饼？特斯拉会不会有一台运送食物的机器？或者是否还有其他公司最终从事这种物品的自动化配送？重点在于，我们现在所处的位置是有客户的。这些客户将在本季度开始部署这套系统，这非常有趣。在我们的配送厨房里，人工成本约占收入的30%。对于成功的餐厅来说，可能占到30%、35%。而在传统实体餐厅中，这个比例甚至更高。

Like, is a Waymo gonna carry a burrito or is Tesla gonna have a a machine that carries food or, you know, is there another another company that ends up doing, you know, sort of the the the things, the the the the autonomous delivery of things. And the point is is, well, where we are right now is we've got customers. And so those customers are starting to deploy this quarter, and it's pretty interesting. I mean, you can in our delivery kitchens, the cost of labor is about 30% of revenue. That's what the successful guy let's say 30%, 35% of revenue in a in a brick and mortar in a brick and mortar restaurants, it's even higher.

当他们使用我们的机器时，人工成本只占收入的7%到10%之间。

K? When they're running our machine, it's between 710% of revenue.

太棒了。当你再去掉配送的成本，现在每个人都能拥有自己的私人厨师了，这也是你最初为Uber设想的愿景。没错。人们可能不知道Uber最初的标语，但那正是你的想法：每个人都有一个私人司机。

Amazing. Then you take out the cost of the delivery, you know, and now it's becoming everybody can have a private chef, which was your original vision for Uber. It was Yeah. People don't know the original tagline, but it was your your everybody has a private driver.

“每个人都有私人司机”是Uber最初的口号。基本上，基础设施已经存在了。我在你们最近的一次活动中提到过，我想是在All In峰会上，Jason。就像在汽车出行领域，道路早已存在，车辆也已经制造出来，而人们每天98%的时间并没有使用他们的汽车。

Everyone's private driver was the original for Uber. Basically, the infrastructure was already there, and I said this on, you know, one of your recent I I think it was at the All In Summit, Jason. But, like, in the mobility cars, you know, transport space, the roads were already there. The cars were already built. People weren't using their cars 98% of the day.

因此，现有的基础设施已经可以让人便捷高效地提供这种服务。但食物方面却没有这样的基础设施。比如，是的，餐馆有闲置产能，Uber Eats 就是这样利用的。但如果要说让一个城市中所有餐食的30%都由某种服务来准备和配送的话。

So the infrastructure is already there to get people around to do this as a service and do it very efficiently and conveniently. With food, the infrastructure is not there. Like, yes, restaurants have excess capacity. That's what Uber Eats utilizes. But to go and say, like, let's make 30% of all meals in a in a city sort of prepared and delivered by a service.

目前并没有这样的基础设施，所以你必须去建设它。我们公司的使命就是打造更好的食品基础设施。这包括房地产、软件以及机器人技术，以一种超级高效的方式进行食品生产和配送。

The infrastructure is not there, so you have to build it. So our company the the mission is infrastructure for better food. So that's real estate, that's software, and robotics for the production and delivery of food in a super efficient way.

好的，基思，你怎么看？有什么问题要问吗

Alright. Keith, what are your thoughts? Any questions for

嗯，他现在不在这里，但这难道不是大卫·弗里德伯格几年前尝试做过的事情吗？

Well, he's not here, but isn't this what David Friedberg tried to do a few years ago?

是的，上一次全员会议的时候提到了这个。是的，我参加的上次会议上就讨论过。没错。

Yeah. This came up on the last all in. Yeah. The last one I was at. Yeah.

是啊，伊扎。伊扎。当时的问题是我告诉弗里德伯格，人们不想吃藜麦，你得加点牛排进去，也许再来一块三文鱼。

Yeah. Itza. Itza. The problem was I told Friedberg, people don't wanna eat quinoa. You gotta put a a little steak in there, maybe a piece of salmon.

但他一开始坚持，最后才勉强同意让人们加一点蛋白质。不过，是的，这是一个非常棒的愿景。

But he was kinda think eventually he relented and let people have a little bit of protein. But, yeah, so it's such a great vision.

等等，他是作为素食殉道者去世的吗？

Wait. He he died as a vegan martyr?

我想

I think

这家企业是以殉道的方式倒闭的。

the business died as a martyr.

嗯，呵呵

Well, he he

被带去死。有种说法是

was led to die. There's a

有很多人死在了那座山上，但归根结底，如果你要进入自动化领域，你必须实现端到端的自动化。我的意思是，比如有一些披萨公司曾经出现过又倒闭了，这些自动化的披萨公司会说我们有一台披萨机器，大家都会觉得哇，这太棒了。你花一百万美元买了一台披萨机，左边有个人在给机器添加原料。

lot of people that have died on that hill, but the bottom line is if you're gonna get into automation, you have to it has to be end to end automation. What I mean by that is, like, there are pizza there are pizza companies that have come and gone, automated pizza companies, where it's like, we have a pizza machine, and everybody's like, yeah. This is amazing. And you have a guy, you have a million dollar pizza machine. And then on the left, you have a guy feeding ingredients into the pizza machine.

右边还有一个人把做好的披萨拿出来装盒。所以与其让一个人来做披萨，你现在换成了一百万美元的机器加上两个人来制作披萨。当你看到这些食品生产的机器人或食品组装设备时，你需要从整体来看，并问自己：它是否适合餐厅现有的生态系统？它能否实现从头到尾的全链条自动化？比如，我们有个东西就像之前看到的那个机器，员工先准备好食材放进机器里，然后他们

And on the right, you have a guy taking the pizza out and then putting it in a box and doing all this. So instead of one guy making pizzas, I have a million dollar machine and two guys making pizza. And so when you look at these, like, robotic food production machines or food assembly machines, you have to look at the full stack and say, does it work with the ecosystem that exists in a restaurant? And does it go full stack from, you know, like like, we have this thing where that machine we saw earlier. The staff preps the food, they put the food in the machine, and then they

离开了，对吧。

leave. Right.

他们就走了。这家餐厅可以在无人看管的情况下运行好几个小时。

They're gone. This restaurant runs itself for many hours without anybody there.

但这可能是麦当劳、汉堡王或者塔可钟，没人能分辨出来。

But this could be McDonald's, Burger King, and Taco Bell. Nobody would know.

就在那里，那台机器是一台装配机，对吧？食物是由人工准备的，然后由这台机器进行组装。对于Chipotle（连锁餐饮品牌）或Sweetgreen（沙拉连锁店）来说，这就是他们的主要劳动力成本，对吧？

That right there, that machine is a it's an assembly machine. Right? The food is prepped by humans and then assembled by this machine. For a Chipotle or a Sweetgreen, this is like a a majority of their labor. Right?

你去Chipotle吃午饭的时候，那里可能有10个员工，而你还是要排队。而这台机器每小时可以完成300碗餐食。所以你可以想想，这就是所谓的装配线。

You go up to a Chipotle, there's like 10 guys at lunch and you're still in line. That machine right there does 300 bowls an hour. Right? And so you go, okay. That's the this is what's called, like, the assembly line.

其实就是前面那条生产线，你基本上是在上面组装各种东西。有时候他们可能会称之为制作线。随着时间推移，你会发现会有垂直的生产线向它供应食物。对吧？你会看到一条生产或制作线连接到这条装配线上，然后你会感叹，哇哦。

It's just that front line where you basically assemble things. I I think sometimes they'll call it the make line. What will happen over time is you'll have perpendicular lines going into it where you're producing food. Right? So you'll have a production or make line going into an assembly line here and then you go, oh, wow.

所以你有一个能分发汉堡夹面包的装置。这就是分配器，也就是组装部分。

So you have it something that dispenses burgers on buns. That's the dispenser. That's the assembly.

对。但这基本上就像是强化版的《异星工厂》（Factorio）。

Right. But It's like factorio on steroids, basically.

是的。然后问题是，你怎么烹饪这个汉堡？我称之为状态变化。所谓状态变化就是食物的烹饪过程。而组装则是如何把它组合起来并装盘。

Yeah. And then it's like, how do you cook that burger? That's what I call that's what we call state change. So state change is the is the cooking of the food. Assembly is the, like, how do I put it together and plate it.

这不会崩溃吗？比如，如果你每小时产量是300个，你说过是的，来自那一台机器。

Doesn't this collapse? Like, for example, if you have a yield of 300 per hour, you said Yes. Out of that one machine

是的。

Yes.

很快你就能计算出拥有一家占地面积更小、仓库里有五个这种设备的价值。是的，通过无人机配送或汽车配送，你不需要实体基础设施。那么你不就制造了一片房地产的荒地吗？你如何重新利用所有这些房地产？

Very quickly, you can impute the value of having a smaller footprint store with five of these things in a faceless warehouse Yeah. With drone delivery or cars. You don't need the physical infrastructure. So then don't you create a wasteland of real estate, or how do you repurpose all the real estate?

思考这个问题的方式是，目前在美国，大概90%的餐食是在家里吃的。现在可能略低于这个数字，我们假设是85%。

Well, the way to think about it is like 90%, well, it's probably a little lower than that right now. Let's say 85 of all meals in The US are at home. They just are.

嗯。

Mhmm.

其中绝大多数餐食都是在家做的。像Uber Eats和DoorDash这样的平台，目前只占所有餐食的1.8%或2%，比例非常小。因此，你正在使用房地产和基础设施来准备并配送餐食到人们家中。

And a vast majority of those meals are cooked at home. So, you know, like Uber Eats and DoorDash, they represent, like, 1.8% or 2% of all meals right now. It's very tiny. Right? So what you're doing is you're using real estate to and infrastructure to prepare and deliver meals to people at their homes.

因此，并不是说餐馆都不存在了。我们仍然想去餐馆吃饭。我们仍然想外出活动。在新冠疫情期间我们意识到了这一点，之前我们也知道这一点。

And so it's not restaurants still exist. We're still gonna wanna go to restaurants. We're still gonna wanna go outside. We learned that during COVID. We knew it before.

我们肯定是在之后才知道的。所以我不认为这真的像房地产那样具有破坏性。这是将一件我们过去自己做的事情变成一个能更高质量完成的服务。我喜欢这么说，你不必有钱才能健康，只需要基础设施来降低成本。因此，你正在以服务的形式做一件我们以前自己做的事情。从长远来看，你会想，杂货店的故事在哪里？

We definitely know it after. And so I don't it's not really like a decimating real estate situation. It's taking a thing we used to do for ourselves and creating a service that does it higher quality. You know, sort of I like to say you don't have to be wealthy to be healthy and just infrastructure to get that cost down. And so you're doing something as a service that we I used to do at think in the super long run, you're like, what where's the story on grocery stores?

如果你设想一下，二十年后，我想大家都会同意，将有机器为每个人制作非常高质量、高度个性化的餐食。

If you go to like, in in twenty years, I think everybody agrees you you will have machines making very high quality, very personalized meals for everybody.

这对基思来说会很好，因为他根据Instagram上的信息，连摄入热量都要精确到五卡路里。

This would be good for Keith because he measures stuff down to, like, five calories based on his Instagram.

基思，你的体脂率是多少？7%、8%？

Keith, what's your what's your what's your body fat? Like, 7%, 8%?

大概是10%。

It's, like, 10.

直接打开他的Instagram就能看到，他发过四次。哦天哪。

Just open his Instagram. Can see he posted four. Oh my god.

我今天已经提了好几次了

There's my times today about

他的体脂率。

his body fat.

他对自己的10%感到非常恶心。

So disgusted with himself at 10%.

是按满分10分算的。但我实际上认为这个愿景自然延伸下去，也许最理想的情况是，每个家庭都拥有一个私人厨师，一个家里的机器人，专门负责这种个性化服务。因为人们确实想在家做饭，但就是没有时间。

It's, like, out of 10. But I actually think the vision of this actually, the natural implication and maybe the home run version of this is everybody has a private chef in their house. A robot in their house that actually does this personalized because people do want to cook at home, but they don't have the time.

是的。当然，空间和基础设施方面，但是伙计，这些配送服务一直在向有钱人收费，让他们这么做。对吧？他们提供这些疯狂的送餐服务，每天要花200美元，而这只是将这种服务抽象化并普及到每个人而已。而且，伙计，正如你提到的Shammaf，当有这种空白空间时，人们会变得很有创意，这些空间到底会发生什么。

Yeah. Of course, space and infrastructure, but, man, these delivery services are charging rich people do this all the time. Right? They do these crazy meal delivery services for $200 a day, and this is just gonna abstract it down to everybody. And, man, people get creative when there's that empty space to your point, Shammaf, about what happens to all this space.

我在八九十年代住在纽约的时候，在我住的翠贝卡和西切尔西地区，很常见的是租个店面，前面放一个小建筑师事务所，后面住人。很多人当时都在灵活利用房地产资源。我们国家仍然需要五百万到一千万套住房，而商场已经被这样改造了。我一直看到商场被改造成大学和创意空间。波士顿有一个商场，他们把第二层和第三层改成了艺术家的工作室公寓。

When I lived in New York in the eighties and nineties, it was common to in Tribeca, in West Chelsea where I lived, to take storefronts, put your little architect's office in the front, and live in the back. And many people were hacking real estate. We still need five, ten million homes in this country, and they're already doing this with malls. I I I keep seeing malls being turned into colleges and creative spaces. One of them in Boston, they turned, like, the Second And Third Floor into studio apartments for artists.

所以你知道，有志者事竟成。我们可以利用这些空间，Simon。

So, you know, where there's a will, there's way. We could use the space, Simon.

是的。Chamath说的这个趋势，以及房地产的发展方向，我们称之为互联网美食广场，比如说你在亚马逊上，它就是一家无所不包的商店。现在想象一下食物领域的亚马逊。

Yeah. Where this goes with Chamath saying and where the real estate goes is we call it the Internet food court where, you know, you're on Amazon. Right? It's the everything store. Now imagine that for food.

然后想象一下你有一个8000平方英尺的场地，基本上任何食物都可以在那里制作出来。

And then imagine you have an 8,000 square foot facility where basically anything can be made.

任何食物都能做出来。

Anything can be made.

因为如果你看到的那个机器有18种食材分配器、10种不同的酱料，你可以理解。那如果变成50个或100个食材分配器呢？如果你有多个配备100个食材分配器的机器呢？

Because if you have that machine you saw has 18 sort of dispensers for food, 10 different sauces, you get the idea. Now what what about when it's 50 or a 100 dispensers for food? What if you have multiple machines with a 100 dispensers for food?

那太疯狂了。

That's crazy.

从组合数学的角度来看，可能实现的食物种类数量呈指数级增长。因此，互联网美食广场就是我们对未来发展的愿景。

You can the combinatorial math in terms of what's possible, what could be made sort of, you know, goes exponential. And so the Internet food court is sort of the vision for where this all goes.

这是‘苦涩教训’的另一个例子。

Another example of the bitter lesson.

很苦，对吧。嗯，我们今天有很多事情要处理。在开始之前，先简单交代一下：今年九月于洛杉矶，再次举办All In峰会，allin.com上有更多详细信息。这次嘉宾阵容非常强大，我们即将开始公布演讲者名单。

The bitter yeah. Well, when we get to that, I guess, today in a very full docket. Before we get to that, just a little bit of housekeeping here. September in Los Angeles, the All In Summit again, allin.com slash yada yada yada. Lineup is stacked, and we're gonna start announcing the speakers.

人们一直恳求我们公布演讲者名单。我不清楚，或许

People have been begging us to announce the speakers. I don't know. Maybe

你得保留一些悬念。小心点，谨慎点。

you You gotta hold some back. Careful, careful.

已经确认了几位嘉宾，但我们确实邀请到了非常棒的演讲者。这次活动将会非同凡响。

Called a couple back, but we got some really nice speakers lined up. It is gonna be extraordinary.

这是目前为止最好的一次。我的意思是，做得不错。这一年我们就该这么做。嗯

Is the best one yet. I mean Well done. Year. We would do this. Well

干得漂亮。没错，没错。每年我们都会有一点小恐慌，就像你知道的那样，我们会请到很棒的演讲嘉宾。

done. Yeah. Yeah. Every year, we have this little bit of panic. Like, you know, we're gonna get great speakers.

而且，老天，这周嘉宾们陆续确认了。这次真的会非常精彩。几乎和我身后这瓶美味的龙舌兰酒一样精彩。来我们的网站tequila.allin.com下单吧，夏末开始配送。

And, man, they started flowing in this week. It's gonna be extraordinary. Almost as extraordinary as this delicious tequila behind my head here. Get the all in tequila tequila.allin.com. Deliveries begin late summer.

接下来

Moving

移到一边。你甚至都看不出那是龙舌兰酒。没错。是的。

to the side. You can't even tell it's tequila. That's right. Yeah.

好啦，听我说。

Alright. Listen.

哦，哇。

Oh, wow.

本周有很多话题要讨论。显然，人工智能继续成为我们行业的重大新闻，而且有充分的理由。我们的老朋友埃隆在周三晚上发布了Grok四。有两个版本，基础模型和重型模型，基础版每月30美元。重型模型每月300美元，它有一个非常独特的功能。

Lots to discuss this week. Obviously, AI is continuing to be the big story in our industry and for good reason. Our bestie, Elon, released Grok four Wednesday night. Two versions, base model and a heavy model, $30 a month for the base. $300 a month for this heavy model, which has a very unique feature.

你可以使用多代理功能。我实际上在几周前参观XAI时看到了这一点，多个代理共同处理同一个问题，并且它们可以同时进行，然后互相比较彼此的工作成果。这有点像学习小组，通过共识得出最佳答案。非常有趣。根据人工智能分析基准，Nick，你可以把这个调出来看一下。

You can have a multi agent feature. I got to see this actually when I visited XAI a couple of weeks ago, where multiple agents work on the same problem, then they and they do that simultaneously, obviously, and then compare each other's work. And it gives you kinda like a study group, the best answer by consensus. Really interesting. According to artificial analysis benchmarks, you can pull that up, Nick.

Grok四的基础模型已经超越了OpenAI的o三Pro、谷歌Gemini的2.5 Pro，成为最智能的模型。这包括大约七种不同的行业标准评估测试。你可以去查一下，包括推理、数学、编码等。但请注意，这是书本上的聪明，不一定是实际应用中的智慧，所以并不意味着这些系统真的能够推理。当然，在X（以前称为Twitter）上也出现了一些争议，当时XAI说了一些各种疯狂的话，可能需要更果断地进行红队测试。

Rock's four base model has surpassed OpenAI's o three Pro, Google Gemini's 2.5 Pro as the most intelligent model. This includes, like, seven different industry standard evaluation tests. You can look it up, but reasoning, math, coding, all that kind of stuff. This is, you know, book smarts, not necessarily street smarts, so it doesn't mean that these things can reason. And, obviously, there was a little there was a little kerfuffle on X, formerly known as Twitter, where XAI got a little frisky and was saying all kinds of crazy stuff and needed to maybe be red teamed a little bit more decisively.

你们很多人知道Grok四是在Colossus上训练的。那是埃隆一直在建设的那个巨型数据中心，我们在这里展示了图表，Chamath。你在群聊中给我们发了一个Rich Sutton于2019年写的博客文章《苦涩的教训》的链接。我们会在这里把它调出来供大家查看，并将其放入节目备注中。

Many of you know Grok four was trained on Colossus. That's that giant data center that Elon's been building, and we showed the chart here, Chamath. You sent us a link to the bitter lesson by Rich Sutton in the group chat. That's the 2019 blog post. We'll pull it up here for people to take a look at and put it in the show notes.

也许总的来说

Maybe just generally

是的。你的

Yeah. Your

对埃隆进展如此之快的反应，以及图表显示了他追赶的速度。我认为人们并没有预料到他会领先，但我们现在确实到了这个地步。

reaction to both how quickly Elon has and that chart showed you, how quickly Elon has caught up. And I don't think people expected him to take the lead, but here we are.

在我们开始之前，Nick，你能展示一下埃隆关于他们在AGI基准测试中表现如何的推文吗？那真是令人难以置信。两点。第一点是从2023年三月开始，不到两年半的时间，这个团队所取得的成就以及他们相对于其他人的领先地位，正如这张图所展示的那样。

Before we start, Nick, can you please show Elon's tweet about how they did on the AGI benchmark? It's absolutely incredible. Two things. One is how quickly starting in March of twenty twenty three. So we're talking about less than two and a half years, what this team has accomplished and how far ahead they are of everybody else as demonstrated by this.

第二点是埃隆做出的一个根本性的架构决策，我想我们现在才完全意识到这一点，而这一决策也映射到他在特斯拉所做的一个架构决策。很可能我们以后会发现他在SpaceX也做了类似的决定。这篇由Rich Sutton撰写的名为《苦涩的教训》的文章很好地概括了这一点。Nick，如果你能把它调出来看一下就更好了。简单总结一下这篇文章的观点，它基本上说的是，当你试图解决一个人工智能问题时，采用一种可以随计算能力扩展的通用学习方法总是更好的选择，因为最终证明这是最有效的方法。

But the second is a fundamental architectural decision that Elon made, which I think we didn't fully appreciate until now, and it maps to an architectural decision he made at Tesla as well. And for all we know, we'll figure out that he made an equivalent decision at SpaceX. And that decision is really well encapsulated by this essay, The Bitter Lesson by Rich Sutton. And Nick, if you can just throw this up here. But just to summarize what this says, it basically says in a nutshell that you're always better off when you're trying to solve an AI problem taking a general learning approach that can scale with computation because it ultimately proves to be the most effective.

而另一种方法则更加依赖人力劳动和人类参与，需要大量的人类知识。因此，第一种方法本质上允许你将任何问题视为一种可无限扩展的搜索或学习任务。事实证明，无论是国际象棋、围棋、语音识别还是计算机视觉，每当出现两种竞争性方法时，一种使用通用计算，另一种使用人类知识，最终总是通用计算的方法胜出。这给那些认为自己处于所有这些关键学习和飞跃中心的人来说是一个苦涩的教训。用更具体的AI语言来说，这意味着许多系统生成的嵌入完全无法被人类理解，但却能产生惊人的结果。

And the alternative would be something that's much more human labored and human involved that requires human knowledge. And so the first method, what it essentially allows you to do is view any problem as an endless scalable search or learning task. And as it's turned out, whether it's chess or Go or speech recognition or computer vision, whenever there was two competing approaches, one that used general computation and one that used human knowledge, the general computation problem always won. And so it creates this bitter lesson for humans that want to think that we are at the center of all of this critical learning and all of these leaps. In more AI specific language, what it means is that a lot of these systems create these embeddings that are just not understandable by humans at all, but it yields incredible results.

那么为什么这很疯狂呢？嗯，他豪赌了一个拥有10万个GPU的集群。人们觉得，哇，这数量太大了，会有效果吗？然后他说，不。

So why is this crazy? Well, he made this huge bet on this 100,000 GPU cluster. People thought, wow, that's a lot. Is it gonna bear fruit? Then he said, no.

实际上，我正在把它扩展到25万个GPU。然后他又说要扩展到一百万个GPU。而这些结果表明，这种不需要太多人工标注的通用计算方法实际上可以更快地得到答案，甚至更好的答案。这具有巨大的影响，因为如果你考虑一下其他所有公司，Lama在做什么？他们刚刚花了150亿美元购买了Scale AI的49%股份。

Actually, I'm scaling it up to 250,000. Then he said it's gonna scale up to a million. And what these results show is a general computational approach that doesn't require as much human labeling, can actually get to the answer and better answers faster. That has huge implications because if you think about all these other companies, what has Lama been doing? They just spent 15,000,000,000 to buy 49% of Scale AI.

这正是对人类知识的一种下注。Gemini在做什么？OpenAI在做什么？Anthropic在做什么？所以所有这些做法都受到了质疑。

That's exactly a bet on human knowledge. What is Gemini doing? What is OpenAI doing? What is Anthropic doing? So all these things come into question.

最后我要说的是，如果你回顾一下，他以前曾做过一次这样的豪赌，那就是特斯拉FSD与Waymo之间的较量。特斯拉FSD只有摄像头，没有激光雷达。但他的想法是，我会比其他人更早地收集数十亿英里的驾驶数据，并应用通用计算，从而比其他更为繁琐且昂贵的方法更快实现自动驾驶。我认为这是技术领域一个非凡的时刻，我们看到了如此多的例子。

And then the last thing I'll say is if you look back, he made this bet once before, which was Tesla FSD versus Waymo. And Tesla FSD only had cameras. It didn't have LIDAR. But the bet was, I'll just collect billions and billions of driving miles before anybody else does and apply general compute, and it'll get to autonomy faster than the other more laborious and very expensive approach. So I just think it's an incredible moment in technology where we see so many examples.

Travis也是另一个例子，他刚刚谈到的内容。你知道，这个苦涩的教训在于你可以相信食物是一种由人手工精心制作的不可改变的东西，或者你可以采用这种通用计算的方法，也就是他所采取的方法，等待成本曲线发挥作用，现在你可以将食品供应扩展到地球上的每个人。我只是觉得这意义非常深远。

Travis is another one, what he's just talked about. You know, the bitter lesson is you could believe that, you know, food is this immutable thing that's made meticulously by hand by these individuals, or you can take this general purpose compute approach, which is what he took, waited for these cost curves to come into play, and now you can scale food to every human on Earth. I I just think it's a it's so profoundly important.

我想补充一点，Chamath，特斯拉在自动驾驶方面的做法其实也利用了人类的知识。事实上，整个理念就是模拟人类驾驶行为，对吧？这就是全部的关键所在。根据你的方法和技术不同，你可以采用所谓的端到端方法，或者你可以将其分解为感知、预测、规划和控制这四个模块，这些模块是你通过工程设计来实现的，如果这样讲得通的话。

One thing I'll throw out there, Chamath, is the Tesla approach for autonomy is taking human knowledge. In fact, the whole idea is to approximate human human driving. Right? That is the whole damn thing. Now depending on your approach and the technology, you can do, like, what's called an end to end approach, or you can look at, okay, perception, prediction, planning, and control, which are like these four modules that sort of you you you sort of engineer if that makes sense.

但它本质上是在模仿人类驾驶行为。区别在于，我认为马斯克采取了一种更接近人类的方法，就像他说的：我有两只眼睛，为什么我的车不能像人类一样驾驶？作为一个人类，我的头上并没有旋转的激光雷达，那为什么我的车就需要呢？

But it's approximating human driving to do it. The difference is that, you know, I I think Elon's taken a a almost a more human approach, which is like, I've got two eyes. Why can't my car why can't my car do it like a human? Like, I don't have any LIDAR spinning around on my head as a human. Why can't my car?

所以这挺有意思的。他某种程度上采纳了你说的观点，Chamath，从计算的角度来看，特斯拉的硬件五代可能明年就要发布了，这将极大提升FSD的能力，这就是你所说的计算方面的发展，但他同时又是在模拟人类的行为。

So it's kinda interesting. He's sort of taking what you're saying, Chamath, on the computation side because hardware five is coming out on Tesla probably next year, which is gonna make a big difference in what FSD can do. That's the compute side you're talking about, but then he is approximating human.

我只是想表达，除了FSD最初几个版本之外，据André（Andre Karpathy）所说，现在的系统已经不再那么依赖于人类标注了，对吧？没错，没错。

I just meant that, you know, other than the first versions of FSD, which I think Andre talked about, Andre Karpathy talked about, you know, they're not really so reliant anymore on human labeling per se. Right? That Respect. Yeah. Yeah.

是啊，就是那个干扰。然后

Yeah. That interference. And then

是的。

Yeah.

他说的另一个疯狂的事情是，后续版本的Grok将不会使用任何现有的传统数据集进行训练。

The other crazy thing that he said, subsequent versions of Grok are not gonna be trained on any traditional dataset that exists in the wild.

人类知识的总和在AI训练中已经被耗尽了，这基本上发生在去年。因此，唯一能补充的方法就是使用合成数据，让AI自己写一篇文章或者提出一个论点，然后自我评估，并通过这种自我学习的过程来生成数据。

The cumulative sum of human knowledge has been exhausted in AI training. That happened basically last last year. And so the only way to then supplement that is with synthetic data where the AI creates it'll sort of write an essay or come up with with a thesis, and then it will grade itself and and and and sort of go through this process of self learning with synthetic data.

他说他会用代理从零开始创建合成数据，然后驱动所有的训练，我觉得这太疯狂了。

He said that he's gonna have agents creating synthetic data from scratch that then drive all the training, which I just think is it's crazy.

再解释一次这个概念，在‘痛苦教训’中提到，把启发式规则硬编码进计算机，比如说，嘿，这些是特定的开局

Just explain this concept one more time in the Bitter Lesson. Hand coding heuristics into the computer and saying, hey. Here are specific openings

在国际象棋中，对，用国际象棋举例，没错。

in chess and t Use yeah. Use chess. Right? Yeah.

国际象棋。你手动编码具体的开局和残局等例子，而不是说，玩每一种可能的游戏，我们拥有所有游戏的数据。所以这就是。

Chess. You're hand coding specific examples of openings in their end games, etcetera, versus just saying, play every possible game, and here's every game we have. So here's the Yeah.

对。所以这两种方法可以这么说，比如特拉维斯和我正在构建竞争版本的国际象棋求解器。特拉维斯的方法会说，我只是定义棋盘。我会给玩家一定的移动边界，对吧？比如主教只能斜着走，还有一些边界条件。

Yeah. So the two approaches would be, let's say, like, Travis and I were building competing versions of a chess solver. And Travis's approach would say, I'm just gonna define the chessboard. I'm gonna give the players certain boundaries in which they can move, right? So the bishop can only move diagonally and there's a couple of boundary conditions.

我会设计一个奖励函数，然后让它自行学习和自我对弈，这是他的版本。当你穷举出所有可能的排列组合后，当你去和基思（世界上最好的国际象棋选手）对弈时，你在做的就是，好的，基思走了这一步。所以你要寻找基思的这一步棋，然后你会有一个最佳应对方式的概率分布，反之亦然。

And I'm gonna create a reward function I'm just gonna let the thing self learn and self play. That's his version. And then what happens is when you map out every single permutation, when you go and play Keith, who's the best chess player in the world, what you're doing at that point is saying, okay. Keith made this move. So you search for what Keith's move is, and you have a distribution of the best moves that you could make in response or vice versa.

那就是当时最先进的方法。另一种方法更不同，也更被人们认为是更所谓优雅、不那么暴力的方法，就是杰森，你和我可以坐下来讨论说，如果基思走到这里，我们应该怎么做，应该采用西西里防御的这个具体变种，但这里面包含了太多人类知识。我认为事实证明，人类在心理上需要相信我们自己也是答案的一部分。但这次表明，由于摩尔定律和通用计算的发展，这种参与已经不再必要。你只需要放手，放弃控制权。

That was the cutting edge approach. The different approach, which is more, you know, what people would think is more quote unquote elegant and less brute force, would be, Jason, for you and I to sit there and say, okay, if Keith moves here, we should do this, we should do this specific variation of the Sicilian defense and it's too much human knowledge. And I think what it turned out was there was a psychological need for humans to believe we were part of the answer. But what this is showing is because of Moore's Law and because of general computation, it's just not necessary. You just have to let go, give up control.

对某些人来说这非常困难。而对另一些人来说，在某些情况下也是如此。

And that's very hard for some people. And for others, it's

尤其是在汽车在路上行驶并在此过程中进行学习的情况下，这就很困难，这就是为什么你需要一个安全驾驶员。我认为埃隆做出这个决定是对的。

also very hard in some circumstances where a car is driving down the road and it's learning in that process, which is why you need a safety driver, and and I think Elon made the right decision to put one in there.

基思，你怎么看？是的，有几点。事情并没有那么简单，查马斯。我基本上同意你的观点走向。

Keith, your thoughts? Yeah. A a couple points. It's it's not quite that binary, Chamath. I generally agree with your arc.

但是，如果你认为大语言模型（LLMs）是人工智能领域最重要的突破，那这些模型都是基于人类写作的数据训练出来的。这意味着每一段用于训练大语言模型的数据，历史上都曾由某个人写过。所以，虽然它们的影响，包括对各种问题的广泛适用性，确实让所有人感到震惊，包括OpenAI最初的团队，但这并不是像有人从外太空找来了不是人类起草的数据来训练模型。当你进入非大语言模型时，你可能是对的，但几乎没有人真正大规模使用非大语言模型。

But, like, if you think about LLMs being the most important unlock in AI, LLMs are all trained on human writing. So someone wrote every piece of data that every LLM used, a human wrote at some point in history. So, yes, it's true that they've shocked everybody, including OpenAI's, you know, original team on the implications, the broad implications, the general applicability to almost every problem. But it's not like there was some tablets floating in space that weren't drafted by humans that we've trained on. As you get in non LLM based models, you may be totally right, but almost no one's really using non LLM based models at scale.

特别是在驾驶方面，特拉维斯说得完全正确，人类实际上是非常好的司机，除非他们分心了。他们可能因为药物或酒精而分心，也可能因为疲劳，或者是因为调收音机、与乘客聊天而分心。

On driving, specifically, Travis is totally right, that humans are actually really good drivers except when they get distracted. They get distracted by drugs or alcohol. They get distracted by being tired. They get distracted by turning the radio. They get distracted by chatting with their passenger.

因此，针对人类行为的弱点来做决策其实是一个很好的选择，因为出于某种达尔文式的原因，人类本来就是理想的司机。所以你不需要从第一原理出发思考，这是一条更好的路径。我认为，再次强调，这可能带来一个普遍的经验教训。我想作为风险投资家来说最重要的一点，正如你所说，是我们多年来一直在争论的问题。

So trading against human behaviors actually turned out to be a great decision because what for whatever sort of Darwinistic reasons, humans are pretty ideal drivers. And so you don't have to reason from first principles. This is a much better path. And I think, again, there may be a a broad sort of lesson there. The most important thing, I think, as a VC that you said, is we've been debating for years.

我们是否应该投资像Scale或Mercor这样的公司，或者任何这类数据标注公司？真相是，我认为人工标注数据的价值半衰期非常短。因此，所有投资这些公司的人都只是关注收入增长，却没有意识到也许最多一年、两年或三年之后，就几乎不会再有人用人工标注的数据来做任何事情了。

Should we invest in companies like Scale or Mercor or any of these surge? The truth is I think there's a very short half life on human labeled data. And so everybody who's investing in these companies, they're just looking at revenue traction, really didn't understand that there may be a year, two years, three years max when anybody uses human labeled data for maybe anything.

因为我们已经触及了人类知识或其收集的尽头。

Because we hit the end of human knowledge or just the collection of it.

或者 或许是你

Or Or you

已完成99%。

99% done.

或者你训练得如此之好，以至于不再需要标注。比如，机器知道如何进行标注，其效果等于或优于人类。我们在自动驾驶领域已经看到这种情况，标注曾经是非常巨大的工作量。对吧？你可以想象通过视频加上激光雷达创建了一个三维场景。

Or you train on you train on it so well that you don't need to label anymore. Like, the the machines know how to label as good or better than a human. And so, like, we're seeing this in the self driving space is labeling was huge. Right? You would have a three-dimensional sort of scene that's created by video plus lidar, let's say.

好的。我基本上要标注所有这些框。比如，我已经识别了物体。在自动驾驶软件领域的一些主要参与者，现在已完全不需要人工标注了，因为机器可以完成全部工作。

Okay. I have to label all of these essentially what become boxes. Like, I've identified objects. You're you're some of the players in the in the autonomous software space, autonomous vehicle software space, are no longer doing any labeling because the machines are doing it all

总体来说，这将直接内置到芯片组中，比如这就是一个停车标志。就像我们人类知道什么是停车标志一样，不需要再百万次地重复识别它。

just broadly. It'll just be built into the chip set that this is a stop sign. Like, it's like we know what a stop sign is. We don't need the millionth time

以前有人必须手动标注验证码中的图像，比如让你找出停车标志或交通灯是什么。但最终，机器在识别这些东西方面比人类强得多。

for somebody to label CAPTCHA is like, you're like, find the stop sign or what's the traffic light? And, eventually, the machines are just way better than humans at identifying these things.

从实际角度出发，当你看到停车标志时，你并不需要明确识别出它是一个停车标志。你只需要观察到每当人类遇到停车标志时，99.9%的情况下他们都会踩刹车。而从来没有人真正去确认那是一个停车标志，只是当看到类似这种物体的东西时，就踩刹车而已。

For you to be very practical, when you see a stop sign, you don't have to identify that it's a stop sign. You just see that every human, when they encounter a stop sign, 99.9% of the time, they hit a brake. And they never act so nobody actually knows it's a stop sign. It's just that hit a brake when you see something that looks like this object.

就是一种感觉。没错，就是一种感觉。

It's just a vibe. Yep. It's a vibe.

我想说的是，这类似于直觉知识与明确标注的人类知识之间的区别。对我来说的问题是，如果一开始大家都非常依赖人工标注，那么如果你是一位投资者，现在看到Grok四的结果，你怎么做出投资决策而不只是单纯押注计算能力？看看这些结果，是否意味着让计算机自行摸索答案相比人为介入之间存在300到1000个基点的差距？如果每次迭代中人为介入导致300到1000个基点的延迟，那么经过两三次迭代后，你就彻底落后了。那么对于其他不是Grok的公司来说，今天醒来之后，他们该如何调整策略或加倍下注呢？

I I would just say that that's like intuitive knowledge versus like the expressly labeled human knowledge. The question for me is if everybody was so reliant on human labeling initially, if you're an investor now, when you see these Grok four results, how do you make an investment decision that's not purely levered to just computation? So if you look at these results, does it mean that the, you know, there's 300 to 1,000 basis points of lag between just letting the computers vibe itself to the answer versus interjecting ourselves. If interjecting ourselves slows us down by 300 to 1,000 basis points per successive iteration, then over two or three iterations, you've totally lost. So what does it mean for everybody that's not Grok when they wake up today and they have to decide how do I change my strategy or double down?

我认为这样。我不参与投资游戏，但如果我是的话，我会专注于科学突破。有时候我会陷入这样的状态：我在探索一条道路。比如，我可能凌晨四五点就醒了，一天还没开始，但我已经睡不着了。

I think look. I'm I'm not in the investment game, but if I were, it would be all about scientific breakthrough. So I sometimes get in this place where I'm looking I'm going down a path. You know, I'll be up at four or five in the morning. My day hasn't quite started, but I'm not sleeping anymore.

然后我会开始浏览Quora，看到一些有趣的量子物理问题或其他感兴趣的内容，接着我会和GPT或Grok一起深入研究这个问题。我们会逐渐接近量子物理领域的已知边界。这时我就像是在做“感觉编程”，只不过这次是“感觉物理学”。当我们接近已知范围的边缘时，我会尝试探索是否存在新的突破。通过这种方式，我确实非常接近发现了一些有趣的新成果。你知道吗，我还曾一度联系过马斯克。

And I'll start go like, I'll be on Quora and see some cool quantum physics question or something else I'm looking into, and I'll go down this thread with GPT or Grok, and I'll start to get to the edge of what's known in quantum physics. And then I'm doing the equivalent of vibe coding, except it's vibe physics. And we're approaching what's known, and I'm trying to poke and see if there's breakthroughs to be had. And I've gotten pretty damn close to some interesting breakthroughs just doing that. And I, you know, I pinged pinged Elon at some point.

我就像，伙计，如果我在做这个，而我还是个超级业余的物理学爱好者，那么那些使用这个工具的博士生和博士后们又会怎样呢？而且这是在Grok四发布之前的情况。现在有了Grok四，以前我会看到Grok犯的一些错误，然后我去纠正它，我们还会讨论这些问题。Grok四可能成为真正发生突破的地方，新的突破。所以如果我要投资这一领域，我会想，谁在科学突破以及这些基础模型之上的应用层面占据优势，并且能够引导这个方向？

I'm just like, dude, if I'm if I'm doing this and I'm super amateur hour physics enthusiast, like, what about all those PhD students and postdocs that are super legit using this tool? And this is pre Grok four. Now with Grok four, like there's a lot of mistakes I was seeing Grok make that then I would correct, and we would talk about it. Grok four could be this place where breakthroughs are actually happening, new breakthroughs. So if I'm investing in this space, I would be like, who's got the edge on scientific breakthroughs and and the application layer on top of these foundational models that orients that direction?

你的看法是，LLM们是否真的开始具备推理能力，从而能提出一个全新的概念理论并实现突破？还是说我们只是过度解读了它们的表现，其实它们只是在边缘尝试一些随机的东西？

Is your perception that the LLMs are actually starting to get to the reasoning level that they'll come up with a novel concept theory and have that breakthrough or that we're kinda reading into it and it's just trying random stuffs at the at the margins?

这或者也可能不会发生。

It's Or maybe it doesn't happen.

不。不是这样的。

No. No.

不。我所看到的是，再说一遍，我还没用过Grok四。今天早上我试着用了一下，但不知为何，在我的应用程序上无法使用。那我们就以Grok三和现有的Chat GPT为例。不。

No. So what I've what I've seen and, again, I haven't used Grok four. I tried to use it early this morning, but for some reason, I couldn't do it on my on my app. But so let's say we're talking Grok three and existing chat GPT as it is. No.

它无法提出新想法。这些东西太依赖于已知的知识，它们就像是，甚至当我想要提出一个新想法时，我必须非常努力地拉拽它，就像牵着一头驴一样。因为它不想打破传统智慧，它特别坚持传统观念。你不断拉扯它，最后它才突然说，哦，天哪，你好像真发现了什么。

It cannot come up with a new idea. These things are so wedded to what is known, and they're so like, even when I come up with a new idea, I have to really it's like pulling a donkey sort of. You're pulling because it doesn't want to break conventional wisdom. It's like really adhering to conventional wisdom. You're pulling it out and then eventually goes, oh shit, you got something.

但当它这么说的时候，你得停下来，再仔细检查几遍，确保你真的有所发现。

But then when it says that when it says that, then you have to you have to go, okay, it said that, but I'm not sure. Like you have to double and triple check to make sure that you really got something.

正如你所说，当这些模型完全脱离了必须从现实世界中学习的限制，而是可以仅通过合成数据进行学习时，一切都将颠倒过来。那时最重要的将是最好的假设或最好的问题。你可以直接给它一个问题，它就能把它解决出来。

To your point, when these models are fully divorced from having to learn on the known world and instead can just learn synthetically Yeah. Then everything gets flipped upside down to what is the best hypothesis you have or what is the best question. You could just give it some problem, and it would just figure it out.

所以，我想说的是，这一切都关乎科学方法。对，如果你有一个LM或某种基础模型，它是世界上最擅长运用科学方法的，那就赢麻了。你基本上只需要点亮更多的GPU，就相当于多了一千个博士生为你工作。

So where I go on this one, guys, is it's all about scientific method. Right? If you get if you have an LM or foundational model of some kind that is the best in the world with the scientific method, gain the f over. You basically you just light up more GPUs, and you just got, like, a thousand more PhD students working for you.

基思，你在这里点头呢。

Keith, you're nodding your head here.

加入进来吧。我同意这一点。我认为这太棒了，因为科学方法同样如此，速度越快，你在提出假设后获得反馈的速度就越快，你就越有可能不断深入、递归地研究下去。而每一次延迟，每一毫秒的延迟都会让你失去思路。这样你就能得到Travis提到的好处以及速度优势，你会到达以前根本无法想象的地方。

Jump in. I I agree with that. I think that's fantastic because the scientific method, also, the faster it is, the more you when you have a hypothesis, the faster you get a response, you're more likely to dive in and dive in and dive in recursively and recursively. And every lag, every millisecond lag causes you to, like, lose your train of thought sort of, so to speak. So you get the benefits that Travis alluded to plus speed, and you go places you never would guess.

当你经营一家公司并进行分析时，这种情况经常发生。如果你有一个工具可以让你快速连续地查询、双击、三击，那么你可以找到原本永远找不到的答案，只要有一点点延迟，哪怕是一两秒钟，更不用说还要等待人工处理。其次，你现在实际上已经看到这种现象正在发生。如果你观察那些支撑科学的基础模型，就会发现有很多关于人体、比如生物学方面的联系，我们人类其实并没有完全理解。例如，为什么我们要做X？

This happens all the time when you run a company and you're doing, like, analytics, and you have a tool that allows you to constantly query quickly, quickly, quickly, double click, triple click. You get to answers that you never get to if there's even a second or two second or three second, let alone sending it to a human. Secondly, where you actually see this today, it's already happening. If you look at foundational models that justify the science, there's lots of things about the human body, let's say, biology that we humans don't actually understand all the connections. Like, why do we do x?

为什么有些人会得癌症？为什么其他人不会得？为什么大脑以这种方式运作？仅基于科学训练的模型往往会揭示出人类前所未有的联系。

Why do some people get cancer? Why do other people not get cancer? Why does the brain work this way? Models trained solely on science tend to expose connections that no human has ever had before.

而且

And

这是因为原材料就在那里，而我们对它的有意识认知只有大约10%。但当你将它应用于其他人类领域时，当你训练的是人类数据、人类产出的数据和结果时，它又受限于这些输出。因此我认为你只需广泛地应用科学，最终你会发现一些人类从未想到过的事物。

that's because, like, the raw material's there, and we only have a conscious awareness of called 110%. But when you apply it to other human domains where you're training on human sort of data, human produced data, human produced output, they're limited to that output. So I think you just take the science and apply it writ large, and you you're gonna wind up finding things that no human has ever thought before.

不过科学的特点是，你需要在物理世界中验证你的假设。所以你可能会想，好吧，如果你拥有这样一个蜂巢思维，一种计算引擎，某种意义上的大脑。

And it's the the the thing about science, though, is that it's the hypothesis that you then have to test in the physical world. So the you're like, okay. If you've got this hive mind, this, like, you know, this computation engine, this brain of sorts.

你想说的是‘意识’这个词，但你忍住了没说出来。

You wanted to say consciousness, but you stopped yourself.

当时我在想，该怎么描述它呢？那个大写的C词——意识。但你必须能够在物理世界中进行测试。你可以设想一个与这些系统连接的真实实验室，在那里可以说：好，如果是化学实验，就可以做化学实验，或者物理学实验。

Was like, how do I describe it? The big c word, consciousness. But but you need to be able to test in the physical world. So you could imagine a a physical lab connected to one of these systems where then you could say, okay. Like, if it's a chemistry experiment, you could do chemistry experiments or physics.

你应该明白我的意思。

You you get the idea.

那会有什么问题吗？

What could go wrong?

那会是的，是的。没什么大不了的。一切都会好起来的。好的。所以但这是它的归宿，因为如果你有一台科学方法机器，你仍然需要能够检验你的假设。

It would it would be it's yeah. No big deal. It's gonna be fine. Okay. So but but this is where it goes because if you have a scientific method machine, you still have to be able to test your hypothesis.

你必须经历科学的澄清过程。

You have to go through the scientific clarification.

是的，完全正确。

Yeah. Exactly.

是的，这有点令人震惊。让我想起一件令人震惊的事情。如果你还记得的话，我不确定你们是否记得暗物质以及它的发现等等。正如丽莎·兰道尔向我解释的那样，这个发现并不是通过知道那里有暗物质并观察到它而做出的，而是通过观察到某些引力作用存在于其他物质周围才得出的。

Yeah. It's kinda mind blowing. Reminds me mind blowing. If you remember I don't know if you guys remember dark matter and, like, the discovery of it and everything. And as explained to me by Lisa Randall, you know, the the discovery was made not by knowing there was dark matter there and observing it, but observing there was something, you know, gravitational forces around this other matter.

然后他们说，等一下，是什么导致了这种现象？于是他们发现了暗物质。所以这些想法你知道的，认为一个大型语言模型（LLM）实际上可以做到这一点，提出如此新颖的想法，这不会让我们感觉我们可能就在那个临界点上吗？对吧？

And then they said, wait. What's causing that? And that's why they found dark matter. So these ideas you know, the idea that an LLM could actually do that, come up with something so novel is it doesn't it feels like we might be right there. Right?

就像，我们正处于突破的边缘。

Like, we're kind of on the cusp of it.

数学中七个最难或最重要的问题之一就是证明一个叫做纳维-斯托克斯方程的一般解，这基本上类似于粘性流体动力学和质量守恒。我们在日常设计所有东西时都使用它。知道吗？它还没有被证明过。这难道不是最疯狂的事情吗？你怎么会连这个都不懂？

One of the seven most difficult problems in math or the most important problems in math is proving a general solution to this thing called Mavi Stokes, which is basically like viscous fluid dynamics and conservation of mass. We use it every day in the design of everything. You know what? It hasn't been proved. Isn't that the craziest thing where you're just like, how is this even possible?

我们用它来设计飞机，设计一切，但它仍未被证明。因此，你可以让一台计算机专门研究这个问题，从而揭开宇宙中无数惊人的奥秘，我们可能会发现完全不同的推进系统，也许我们可以实现一些我们以前认为不可能的事情，比如传送。我的意思是，谁知道什么都有可能呢？

We use it airplanes, to design everything. It hasn't been proved. And so you could just point a computer at this thing and you would unlock all these incredible mysteries of the universe and we would probably find completely different propulsion systems. We could probably do things that we didn't think were possible, teleportation. I mean, who knows what's possible?

但记住，记住埃隆是如何谈论布洛克以及总体上的人工智能的，那就是：我们为什么在这里？我们的目的是什么？

But remember remember, you know, how Elon talks about Brock and and about AI generally is about why are we here? What is the purpose?

宇宙的意义。

Meaning of the universe.

是的。宇宙的意义是什么？它是如何运作的？其中还有一种强烈的追求真理的机制。

Yeah. What is the meaning of the universe? How does it work? And a sort of fierce truth seeking mechanism there.

我问你们一个问题，Keith，Travis，Jason。如果你们在运营Grok四号会怎样？

Let me ask you a question, Keith, Travis, Jason. If you guys were running Grok four.

那将会非常有趣。

That'd be so much fun.

你们如何用柔道技巧击败OpenAI？因为他们正坚定地朝着十亿MAO、然后是十亿DAO迈进。它是一个强大的对手。那么你们如何利用更好的产品，在关键时刻击败这个相对较差的产品呢？

How do you judo flip OpenAI? Because they are marching steadfastly towards a billion MAO, then a billion DAO. It's a juggernaut. So how do you use the better product in a moment to judo flip the less better product?

听着，是的。我的意思是，问题就在这里，对吧？所以你要按照Elon的方式去做。

Look. Yeah. I mean, here's the thing. Right? So you do the Elon Way.

你得召集一群有传教士精神的工程师，全力以赴工作的传教士型工程师，你们的文化是一种极度追求真理的文化，不陷入政治、官僚主义和废话之中，你就放手去干。我认为，你知道的，这就是关键。然后你会发现，科学突破、科学方法。

So you you get a bunch of missionary, like, full on missionary engineers that work twice as hard, and you have a culture that is ultra fierce truth seeking, and you don't you don't get caught up in politics, bureaucracy, b s, and you just you go for it. And and I think, you know, that's where you know? And then you go, wow. Scientific breakthrough. Scientific method.

比如，你开始以真理取胜，我相信这将使产品的卓越程度足以挑战OpenAI的产品。

Like, you start winning on truth, and that will start I believe that will start to give the product awesomeness of OpenAI a run for its money.

嗯。

Mhmm.

但是，OpenAI的产品部门，那些家伙做得非常出色。

But, like, the product of OpenAI, the product department, those guys are crushing.

他们正在崩溃。

They're crashing.

它们确实很厉害。他们不仅领先于行业，而且他们自己也感觉到了这一点，他们在很多方面都处于领先地位。但如果你更擅长把握真相，最终你将拥有一个AI产品经理。

They're really good. They're not only ahead of the game, but they feel like it just they're just leading in a lot of different ways. But if you are better at truth, you will eventually you'll eventually have an AI product manager.

从战术层面来说，人们也常常忘记埃隆在工厂和实体世界事务上的卓越能力。嗯哼。他建立Colossus（超级工厂）时，黄仁勋就曾惊讶道：你是怎么做到的？对吧？所以，发挥他的这种优势，他的建厂能力，并且他多次说过，工厂本身才是特斯拉的产品。

And on a tactical basis too, people forget how good Elon is at factories and physical real world things. Mhmm. What he did standing up colossus made, like Jensen Wan was like, how is this possible that you did this? Right? So pressing that, his ability to build factories and he said many times, like, the factory is the product to Tesla.

真正的产品不是从工厂里生产出来的汽车或电池，而是工厂本身。如果他能在太阳能方面解决能源问题，同时发展电池和建设Colossus二号、三号、四号、五号工厂，那么他在这一领域将拥有巨大的优势。此外，还有Travis（霍普金斯）这样的传教士型人物。顺便提一下，Travis是他早前支持的人，那时山姆·阿尔特曼还没有扭曲OpenAI最初的传教士精神，把它变成封闭的AI。这并不是对阿尔特曼有什么贬低的意思，但他确实欺骗并背叛了埃隆。这不是什么个人恩怨，他就是坑了埃隆一把。

It's not the cars that come out of the factory or the batteries. It's the factory itself. So if he can keep solving the energy problem with solar on one side and batteries and standing up, you know, colossus two, three, four, five, he's gonna have a massive advantage there on top of Travis, you know, the missionary individuals, which, by the way, was what he backed before Sam Altman corrupted the original missionary basis of OpenAI and made it closed AI and a know, this is nothing derogatory towards him, but he did hoodwink and stab Elon in the back. It's not nothing personal. I mean, he just screwed him over.

你会说他是欺骗了他吗？

Would you say he bamboozled him?

他欺骗了他，坑了他，蒙骗了他。你知道的，你可以选择不同的词来形容。但他并没有做卑鄙的事。最初的使命是成为通用的开源组织，等等。

He bamboozled him, screwed him, hoodwinked him. You know? But you pick your term here. But he didn't he didn't dirty. The original mission was to be like a generic open source all this content.

这是另一个我认为是变数的因素，然后我会听听基思的观点。但开源其中一些内容可能会产生深远的影响。我认为，开源自动驾驶数据可能会带来非常深远的影响。埃隆想要做一些真正颠覆性的事情。比如，他曾开源了自己的充电专利。

That's the other piece I think is a wild card, and I'll and then I'll I'm interested in Keith's position. But open sourcing some of this could have profound ramifications. I think open sourcing the self driving data could have a really profound impact. Elon wanted to do something really disruptive. Like, he open sourced his patents for, you know, charging.

如果他开源了自动驾驶数据集和自动驾驶技术，其他人有没有能力像他一样大规模生产机器人出租车？我认为没有。

If he open sourced the dataset and self driving, does anybody have the ability to produce robotaxis at the scale he can do it? I don't think so.

嗯，如果特拉维斯的假设是正确的，那么是的，每个人都能做到。

Well, if Travis's hypothesis is true, then, yeah, everybody will.

那每个人都能做什么？抱歉，每个人都能做什么，希蒙？

Well Everybody will what? Sorry. Everybody will what, Shimon?

如果你能获得购买计算能力的资金，那么每个人都可以解决这个问题。

If you have access to the money that buys the compute, everyone could solve that problem.

哪个他

Which He

他说如果他发布了所有的FSD数据，有人能制造出自动驾驶汽车吗？

said he said if he if he published all the FSD data, could somebody build an autonomous vehicle?

当然可以。但有人能从工厂里生产出一亿辆装有电池的机器人出租车吗？好的。

Well, yes. But could somebody produce a 100,000,000 robotaxis from a factory with batteries in them? Okay.

不能。那是不同的问题，那是个

No. That's a diff that's

不同的问题。我的意思是。

a different question. I'm saying.

其实也不太算不同，因为很好。上一次我作为嘉宾参加节目时，我们谈到了垂直整合。产品确实需要垂直整合。因此，最终你会拥有一辆专门为自动驾驶设计的车辆，并且它的交互方式也有所不同。成本结构也不同。

And not really because Good. Last time I was a guest on, you know, all and we talked about vertical integration. Products really require vertical integration. So, ultimately, you have a self driving something that is custom built for knowing it's going to be self driving, and it interacts differently. The cost structure is different.

控制系统不一样，座椅布局也不一样。你在构建产品时要利用自己最具竞争优势的地方，然后加以发挥并强化这一优势。这正是为什么像苹果这样尽管错过了AI浪潮，在任何实证标准下仍然是一家相当不错的公司。我的意思是，他们在过去七十年最重要的技术突破上的表现绝对糟糕透顶，但公司依然存在，并且市值依然高达数万亿美元，因为他们实现了垂直整合。

The controls are different. The seating's different. Everything you build a product taking advantage of where in the staff you have the most competitive advantage, but then you leverage that and it reinforces. It's still why, like, Apple, despite missing the AI wave, still a pretty good company from any empirical standpoint. I mean, like, their performance is absolutely miserable on the most important technology breakthrough over the last seventy years, but the company is still alive and still worth trillions of dollars because it's vertically integrated.

正如你所说，OpenAI确实有一个优秀的产品团队，他们必须在产品层面保持领先，因为他们无法在工厂层面竞争。要在产品层面保持领先的方法就是推出设备。他们必须发布设备。它必须出色。它必须准确无误。

OpenAI, at per your point, they do have a good product team, and they need to stay ahead of the product level because they can't compete on the factory level. The way to stay ahead of the product level is shipping a device. They gotta ship the device. It's gotta be good. It's gotta be right.

它必须具备合适的外形尺寸。它必须为人类做一些意想不到的事情。但如果他们做到了这一点，那么他们就会成为苹果加AI的组合。

It's gotta be the right form factor. It's gotta do things for humans that are unexpected. But then if they do that, they're like Apple plus AI.

查马斯，你之前提到的论文是什么？论文的名字再说一遍？

Chamath, what's the paper you're talking about before? What was the name of it again?

痛苦的教训。

The Bitter Lesson.

这可能适用于自动驾驶，目前它仍然像这样：嘿，我该怎么像人一样开车？我们讨论过这个问题。但这里的飞跃时刻可能是这样的：嘿，开车就行了。

That that could apply to autonomous driving is right now it's still like, hey. How do I drive like a human? We talked about that. But the leapfrog moment here could be like, hey. Drive a car.

确保效率高，不要撞到任何人。然后模拟几万亿次，一切都没问题。对吧？但现在，我们仍在试图像人类那样驾驶，因为我们的数据还不够，因此无法进行足够的计算。

Make sure it's efficient. Don't hit anybody. And just simulate that quadrillion times, and it's all good. Right? But right now, we're still trying to drive like humans because we don't have enough data and therefore can't do enough compute.

顺便说一句，这是个全局性的教训。Tramath，你完全正确。从概念上讲，这篇博客文章是对的，但这只有在拥有足够数据的情况下才成立。根据不同的使用场景，你所需的数据量可能需要几年甚至几十年才能获得，在此之前你只能通过与人的交互来逐步实现目标。

That's the global lesson, by the way. Tramath, you're totally right. Conceptual you know, the blog post is right, but that's only true when you have enough data. And depending upon the use case, the level of data you need may not be possible for years, decades, and you may need to hack your way there through human interactions.

物理世界中的人工智能缺乏数据，所以你只能尝试去逼近人类的行为。

Physical world AI is lacking in data, and so you just try to approximate humans.

我不知道你们有没有看到这个消息。相关的新闻是，OpenAI 和 Perplexity 正在进军浏览器领域。Perplexity 推出了针对其每月 200 美元订阅层级的 Comet 浏览器。我实际上下载了它，稍后给你们看一下。

I don't know if you guys have seen this. In related news, OpenAI and Perplexity are going after the browser. Perplexity launched Comet for their 200 a month tier. I actually downloaded it. I'll show it to you in a second.

这是一个非常有趣的类别，是开发者现在就可以做的事情，而且他们经常这么做。明白吗？但是让浏览器连接到代理（agent），可以让你完成一些真正有趣的事情。我会在这里给你们展示一个例子，就是我们在谈话时刚刚运行的一个任务。

But this is a really interesting category. It's something developers can do already, and they do it all the time. You know? But having your browser connected to agents lets you do really interesting things. I'll show you an example here that I I just fired off while we're talking.

所以我刚刚问它，嘿，给我找一下从旧金山到纽约市的联合航空商务舱最佳航班。它会做一些搜索。但你在这里看到的是，它弹出了一个浏览器窗口，并且正在执行这项工作。你可以看到它使用的步骤。

So I just asked it, hey. Give me the best flights from United Airlines and business class from New York City from San Francisco to New York City. It does some searches. But what you see here is it's popped up a browser window, and it's actually doing that work. And you can see the steps it's using.

然后我可以实际打开那个浏览器窗口，看着它完成这些操作。这只是它的一张截图，它可以同时打开多个这样的窗口。比如前几天我在做一次搜索的时候，让它告诉我所有我没有在亚马逊上购买的自传。

And then I can actually open that browser window and watch it do that. This is just a screenshot of it. And it will open multiple of these. So you could I was doing a search the other day saying, like, hey. Tell me all the autobiographies I haven't bought on Amazon.

把它放进我的购物车并为每一本做个总结，因为我喜欢读传记，我喜欢在这里完成这些事情。上次它执行这个任务时，把我的航班信息添加到了我的账户里并完成了结账流程。所以，再说一遍，如果你是一个开发者，你会整天做这样的事情，但这似乎真的代表了一个新的产品类别。我很想知道你们是否已经试用过它，以及你们对于拥有这样一个能够实时执行这些任务的具备代理功能的浏览器有什么看法。当然，你也可以将你的 Gmail 和日历连接到它上面。

Put them into my, you know, shopping cart and summarize each of them because I like biographies and I like doing it here. And when it did this last time, it put my flight into like, and I was logged in under my account, and it basically put it into my account in the checkout. So, again, this isn't like if you're a developer, you do this all day long, but this really seems to be a new product category. I'm curious if you guys have played with it yet and then what your thoughts are on having an agentic browser like this available to you to be doing these tasks in real time. You can also connect, obviously, your Gmail, your calendar to it.

所以我做了一个搜索。告诉我我去过哪些餐馆，并按城市分类。然后我打算打开我的 OpenTable 并提取那些数据。Keith，这事情有意思的地方是，你不用非得通过云来做这件事。

So I did a a search. Tell me every restaurant I've been to and then put it by city. And then I was gonna open my OpenTable and then pull that data as well. What's interesting about this, Keith, and I I know you're a product guy and you've done a lot of product work. I'm curious your thoughts on it, is you don't have to do this in the cloud.

你已经登录了很多账户，而且你也不必担心被这些服务阻止，因为它看起来不像爬虫或机器人。它只是一个浏览器在执行任务。你怎么看这个？

You're authenticated already into a lot of your accounts, nor do you have to worry about being blocked by these services because it doesn't look like a scraper or a bot. It just that you're a browser doing the work. Your thoughts on this?

你

Have you

玩过它吗？

played with it at all?

玩过。我认为这是 Perplexity 的一次绝妙的孤注一掷尝试。我觉得如果没有这样的东西，Perplexity 就完了。比如 ChatGPT 达到十亿用户这个状态，它已经成为一个动词了，你知道的，就是普通消费者描述使用 AI 的方式。如果 Perplexity 无法实现这一点，那他们就没有其他出路了。

Yep. I think it's a great Hail Mary attempt by perplexity. I think absent something like this, perplexity is toast. Like, for the stat about ChatGPT going to a billion users, like, it's becoming the verb, you know, the the way you describe using AI for a normal consumer. There's nothing left in perplexity if they can't pull this off.

所以这是一个很棒的想法，因为消费科技公司的历史表明，谁先占据高地，谁就拥有巨大的优势。实际上，谷歌也应该这么做。老实说，我认为谷歌搜索和跨搜索也完蛋了。既然他们有 Chrome 浏览器，理论上还有 Gemini 这个优质团队，他们应该将这两者结合起来，希望与 ChatGPT 竞争。否则他们在搜索领域的地位将会失去。

So it's a great idea because, like, the history of, like, consumer technology companies is whoever's up has uphill ground, like, in a military sense, whoever's first has a lot of control. This is actually what Google should be doing, truthfully. Like, I think Google's also Google search cross search is toast. And since they have Chrome and they theoretically have a a quality team in Gemini, they should be putting these two things together and hoping to compete with Chad GPT. They're gonna lose the search game.

目前来看，谷歌最好的资产其实跟搜索无关。而是它的其他所有产品。如果他们能想办法利用好这些产品，这才是拯救这家公司的唯一途径。

Like, the assets that are best at Google right now have nothing to do with search. It's every other product. It's the only thing that's gonna save that company if they can figure out how to use them.

Travis，你对这个领域有什么看法？有没有想到一些特别的功能集？我知道你喜欢思考产品和用户体验。

Travis, your thoughts on this category? Anything come to mind for you in terms of, you know, feature sets that would be extraordinary here? I know you you like to think about products and the consumer experience.

非常有趣。正如你们知道的，我最近一直在关注房地产、建筑和机器人技术。我已经很久没有参与这种消费类软件领域了。但过去六个月里，有一些消费软件公司的 CEO 让我感到很感兴趣。

It's really interesting. So, you know, I've been spending yeah. As you guys know, I've been spending my time on real estate and construction and robotics. And so I've I've been out of the this kind of consumer software game for a long time. But super interesting over the last six months, there have been a a number of consumer software CEOs.

当我跟他们聚会或者交流时，他们会说：伙计，当代理（AI Agents）接管一切的时候，我们该怎么继续做我们现在做的事情呢？

Like, when I hang out with them or whatever, they're like, yo. How are we gonna how are we gonna keep doing what we do when the agents take over?

是的。这种范式转变非常深刻，以至于你访问网页的概念将会消失，而你只是在进行一个对话？拥有一个

Yeah. The paradigm shift is so profound that the idea that you would visit a web page goes away and you're just in a chat dialogue? Have an

代理帮你处理你的航班事宜。

agent who's just taking care of your flights for you.

所以

So

我 我有点 我认为中间有一个跳跃。我觉得就像你告诉某个东西，嘿，我想去纽约。你能帮我找找吗？我大概看一下这个时间段，你能找到一些我可能会喜欢的东西，然后给我几个选项？好的。

I I kinda I I think there's a leapfrog over that. I think it's just like you tell something, yo, I wanna go to New York. Can you you know, I'm sort of looking at this time range. Can you just go find something I'm probably gonna like and give me a couple options? Yeah.

这完全是一个界面的问题，然后，你知道吗，在Perplexity上你刚刚展示的那个东西就是最终的界面吗？还是说我只需要一个代理，它能替我去完成所有的事情？这是开始吗？说实话，我还没有花足够的时间去了解。我知道的是，每一个在App Store里有应用的消费软件公司的CEO现在都慌了神，他们现在都很焦虑。

And it's just a whole you have an interface, and then, you know, is per is this thing that you just showed on Perplexity, is that the interface, or do I just have an agent that just goes and does everything for me? And is this the start of that? You know, I just haven't spent enough time. I I do know that every consumer software CEO that has an app in the App Store is tripping. They're tripping right now.

我需要大公司。我需要那些真正有实力的人。有时候我几乎像是在跟他们做心理治疗。我会说，没事的，其实你们真的有东西。

And I need big boys. I need guys with real stuff. And sometimes I I'm doing, like, almost like therapy sessions with them. I'm like, it's gonna be fine. Actually you actually have stuff.

你们有护城河。你们有真正有价值的东西。代理无法取代这些。但他们

You have a moat. You have real stuff that's of value. They can't replace it with an agent. And they're

就像是在欺骗他们。你像临终关怀那样安慰他们，告诉他们一切都会好起来的，但病人却因为你说的同样的话停止了呼吸

like lying to them. You're doing hospice care, and you're telling them everything's gonna be okay, but the patient stops because you got all same

在Robin Hood上有各种选项，他却说，是的，是的。告诉我更多吧。

options on Robin Hood while he's like, yeah. Yeah. Tell me more.

告诉我更多。

Tell me more.

他很聪明，伙计们。所有这些事情。有些东西是受保护的，有些则不是。就是这么简单。那

He's smart, guys. All these things. There's certain things that are protected, there's certain not things that aren't. That's all. What about

我们来谈谈这个吧？因为你我年纪都足够大，记得当年的General Magic（通用魔术公司）。很久以前就出现了这种愿景，当时有个人数字助理，你可以直接跟一个代理交谈，它会替你完成这些任务。这感觉就像是朝着那个方向迈出了一步，它为你完成所有工作，最后只呈现结果给你，并说请确认。

let's talk about that? Because the you and I are old enough to remember general magic. This vision was out there a long time ago with personal digital assistance, and you would just talk to an agent. It would go do this for you. This feels like a step to that where it does all the work for you, presents you the final moment, and says approve.

所以几乎就像你是管家或男仆一样。

So almost like you're concierge or a butler.

是的。我认为你描述的是我们想要的东西。但更具体地说，对于现在的情况，Keith 和 Travis 的观点完全正确。听着，我认为开发一款浏览器绝对是一个愚蠢的资本分配决策。在2025年，这完全是愚蠢且无法辩解的。

Yeah. I think what you're describing is what we want. But I think more specifically for today, Keith and Travis totally nail it. Look, I think building a browser is an absolutely stupid capital allocation decision. Just totally stupid and unjustifiable in 2025.

特别是对 Perplexity 来说，我认为他们打造一家伟大企业的路径是取代彭博社（Bloomberg）。他们在金融信息和金融数据方面所做的一切，以及超越模型的尝试都非常出色。作为一个多年来为彭博终端支付了25,000美元的人，这个终端简直糟糕透顶，非常差劲，没什么用，而且功能非常有限。任何能够打造出更好产品的人，都能接管这家价值千亿美元的企业，因为我觉得机会就在那里。我希望Perplexity能加倍甚至三倍投入这一领域。因此当你看到这种

Specifically for Perplexity, I think their path to building a legacy business is to replace Bloomberg. Everything that they've done in financial information and financial data in going beyond the model has been excellent. As somebody who's paid $25,000 to Bloomberg for many years, the terminal is atrocious, it's terrible, it's not very good, it's very limited And anybody that could build a better product would take over a $100,000,000,000 enterprise because I think it's there for the taking. I wish that perplexity would double and triple down on that. And so when you see this kind of

让我们随意扩展一下吧，Jamath。我们就去干吧。

Let's random sprawl do it, Jamath. Let's just go do it.

当你随意扩展时，你会发现这种方法行不通。但我只想说，在2025年开发一个浏览器是最愚蠢的事情，因为在代理的世界里，浏览器是什么？它只是一个华丽的标记语言阅读器。它处理HTML、CSS和JavaScript。

When you do the random sprawl, think it doesn't work. But I just wanna say, like, a browser is, like, the dumbest thing to build in 2025 because in a world of agents, what is a browser? It's a glorified markup reader. It's like handling HTML. It's handling CSS and JavaScript.

它进行一些网络连接，一些安全处理，一些渲染。但这都是底层的技术细节。我明白在1998年的时候我们必须处理那些乱七八糟的东西，比如第一次尝试Lycos或者谷歌。

It's doing some networking. It's doing some security. It's doing some rendering. But it's like, this is all under the water type stuff. I get it that we had to deal with all that nonsense in 1998 to try Lycos or Google for the first time.

但在2025年的今天，你应该可以直接跟某个东西对话，最终，可能还会有一个植入你大脑的东西，你只需思考，它就能理解。你一想到我需要预订飞往JFK的航班，它就知道该怎么做。或者至少今天来说，在一个非常优雅美观的搜索栏中，你输入“帮我订个航班”，它就已经知道该怎么做了。

But in 02/2025, there is something that you just speak to, and eventually, there's probably something that's in your brain which you just think, and it just doesn't. You're thinking, I need a flight to JFK. Or at the maximum today, in a very elegant, beautiful search bar, you type in get me a flight, and it already knows what to do.

Keith，从某种意义上讲，这是迈向终极愿景的一个步骤。所以你会认为，如果从这个角度看，也许Perplexity开发这款产品是值得的，把它当作通向终极愿景的一个中间节点，而终极愿景就是命令行加耳机。一个热

Keith, in some ways, this is a step towards that ultimate vision. So you'd think it's worth it to, you know, sort of perplexity to make this waypoint perhaps if you look at it as a waypoint between the ultimate vision, which is a command line, an earpiece. A hot

分发、Pot Jason，2025年第十九个网络浏览器？

distribution, pot Jason, for the nineteenth web browser in 2025?

嗯，是的，这确实是个挑战。我认为大多数人猜测苹果公司（拥有大量用户）可能会收购Perplexity或与之达成交易，从而获得分发渠道，特别是因为司法部对谷歌的反垄断案件。所以关于这一点有很多猜测。但Keith，你怎么看？

Well, yeah, that is a challenge, and I think most people are speculating Apple, which has a lot of users, might buy perplexity or do a deal with perplexity and give them that distribution because of the justice department case against Google. So there's been a lot of speculation about that. But, Keith, what do you think?

我不认为他们会买任何真正有价值的东西。比如，你想要苹果得到什么？如果你继续这种失败的战略的话……

I don't think they'd buy anything worth it. Like, what do you what is Apple gonna get? And if you continue this failed strategy of

没错，苹果就是这样。

Apple Right.

苹果在人工智能方面错过了每一个可能的机会，并且还在继续错过。我认为他们在文化上存在问题，CEO可能存在挑战，基础设施也存在挑战。因此这不是一个容易解决的问题，而收购Perplexity并不能帮上忙。实际上，Travath针对复杂性/Perplexity的战略其实相当合理。

Apple has missed every possible window on AI and continues to miss it. And it has cultural I think the CEO has challenges. I think culturally, they have challenges, and they have infrastructure challenges. So it's it's not an easy fix, but buying perplexity is not gonna help. Like, Travath strategy is actually a pretty coherent one for complexity, quap perplexity.

所以我认为不会

So I think that not

垂直整合并拥有自己的战略

a vertical and own a strategy

不，不是

Not not

不是一个坏主意，尤其是因为你需要独特的数据来源。其中一些数据源可能愿意也可能不愿意将数据授权给OpenAI。所以在那你可以做一些聪明的事情。但我并不认为苹果从Perplexity身上能获得任何剩余价值，除了某些产品品味之外。但你会为什么买单？比如花十亿美元去买产品品味吗？

not a bad idea, especially because you need unique data sources. Some of those data sources may or may not license their data to OpenAI. So you can do some clever things there, but I don't think there's any residual value that Apple would get out of perplexity except there's some product taste. But what are you gonna spend? Like, a billion dollars for product taste?

我的意思是，Mark现在动辄花费数亿甚至数千亿美元来做事情。而且你看Grock，Grock四代表明了，Mark真的只是需要砸钱来组建一支全新的团队，因为他们迄今为止在AI领域所做的一切都错过了时机。

I mean, Mark's spending hundreds of millions of dollar hundreds of billions of dollars or whatever he's spending these days. And, you know, Grock, if anything, Grock four shows that Mark really just need to spend money to build a whole new team because everything they've done in AI has also missed the boat.

但是，我的意思是，基思，那个

But, I mean, Keith, the

你那样表达的方式几乎让苹果值得孤注一掷，组建一支有点品味的团队，因为他们的做事风格一向如此，就是追求优雅。那为什么不再为此投入搜索资源呢？投个100亿进去

way you phrase it there almost makes it worth it for Apple to throw a Hail Mary, have a team with some taste because that's how they tend to do things is something that is elegant. And why not just throw your search to it? Throw 10,000,000,000 at

真正优雅的是，如果有一系列操作功能。代理功能就在一个聊天框里。看到一堆视觉垃圾并不优雅，那是偷懒的做法。

What's elegant is if there'd be a bunch actions. Agents in just a chat box. Seeing a bunch of visual diarrhea is not elegant. It's lazy.

我们这个小布隆伯格克隆产品就叫polyhapatia吧。我给你命名权。嘿，谁能把polyhapatia调出来一下？

Shim off on our on our little Bloomberg clone. I'll give you naming rights. So you can call it the polyhapatia. So, hey, can somebody can somebody bring up the polyhapatia?

你知道什么特别好笑吗？它有

You you know what's so funny? It has

这名字念起来还挺顺口的。

a pretty good rolls right off your tongue.

TK，听我说。我们想做一个公司筛选，结果在特定类型的筛选中最多只能显示五家公司。比如你想比较股价和息税折旧及摊销前利润（EBITDA），然后你说好吧，我好像只能选五家。那我该选哪五家呢？

TK, listen. We were trying to do a screen of companies, and it maxes out at five companies on a specific type of screen where you're, like, you're trying to compare stock price to EBITDA, and you're like, okay. I can only choose five, I guess. So which five should I choose?

Laffont前两集不是提过吗？他说我打不开这个页面，最多只能显示六家公司。

Laffont was on, right, like two episodes ago? He was like, I can't pull this up. It's limited to six companies.

伙计，人们用布隆伯格终端是为什么？他们是为了它的通讯功能。我的团队曾经通过布隆伯格发短信交易了巨额头寸，所以这方面确实很有价值。但这家公司的核心可用性和核心用户界面一直没有改进。

Dude, you it's so what do people use Bloomberg? They use it for the messaging. Now, like, my team has traded huge positions via text message on Bloomberg, so there is something very valuable there. But the core usability and the core UI of that company has not evolved.

我也有我的贡献。

I have my contribution.

Perplexity 在这方面非常擅长，通过

Perplexity is very good at that, by

顺便说一下。

the way.

它他们在这方面做得非常好。

It it they they do a very good job.

我注册了一个新域名，Travis。让这个想法在这里酝酿一下。这是我进入交易的小把戏。Begin.com，Begin.com。

I got a new domain name, Travis. Let this one just sink in here. This is my way to weasel my way into the deal. Begin.com. Begin.com.

你拥有

You own

那个，不是吗？

that, don't you?

是的，我只是偶尔抢注了一些不错的域名。我得到了 begin.com 和 annotated.com。这是我的两个小域名。

I do. I'm just a little I sniped some good ones once in a while. I got begin.com, and I got annotated.com. Those are my two little domain names.

你就像那些老派人，出现在那些哦，就像是一个

You're like one of these old people that show up at those Oh, like a

路演现场，然后突然之间你

roadshow, and then then all of sudden you

有一个茶业的路演。然后你说，哦，我有这个东西是我1845年买下的。

got a tea roadshow. And you're like, oh, I have this thing that I bought 1845.

伙计们，Jason Jason 是 GoDaddy 的爹。好啦，我就是。

Guys, Jason Jason is Jason is the daddy in GoDaddy. Okay. I am.

我是你爹，我是你爹。

I'm your dad. I'm your dad.

事情就是这样。

That's what it is.

谁是你爹？嘿。说到爹，我们继续下一个故事吧。

Who's your daddy? Hey. Speaking of daddy, let's go to our next story.

哦，是

Oh, is

现在是成立第三党的合适时机吗？马斯克似乎这么认为。上周他宣布Axie将创建一个新的政党。这个故事里谁是爹，你们自己决定吧。他说过：“在用浪费和腐败使我们的国家破产这方面，我们生活在一党制的国家，而不是民主国家。”

now the right time for a third party? Elon seems to think so. Last week, he announced that Axie would be creating a a new political party. I'll let you decide who daddy is in this one. He said, quote, when it comes to bankrupting our country with waste and graft, we live in a one party system, not a democracy.

他尚未详细说明美国党的政治纲领。我们上周在这里讨论过。我列出了四项核心价值观，在X（原推特）上反响不错：财政责任/狗狗币、可持续能源、在这方面的主导地位以及在美国的制造业，这是马斯克单枪匹马推动发展的领域。

He's not yet outlined a a platform for the American party. We talked about it here last week. I listed four core values, which seem to get a good reaction on x. Fiscal responsibility slash doge, sustainable energy, and dominance in that. Manufacturing in The US, which Elon has done single handedly here.

还有亲生育政策，我认为这是他的一个情怀项目。另外，沙巴特，你补充了第五点：根据Polymarket的数据，技术卓越有55%的可能性在今年年底之前注册成立美国党。我还想弄清楚的一件事是这些候选人和政党的人气到底有多低。这是一张非常有趣的图表，我觉得我们可以围绕它展开一场精彩的讨论。

Pronatalism, which I think is a passion project for him. And, Shabbat, you punched it out with the fifth technological excellence according to Polymarket. 55% chance that Elon registers the American party by the end of the year. And, you know, one thing I was trying to figure out is just how unpopular are these candidates and these political parties. This is a very interesting chart that I think we can have a a great conversation around.

事实证明，我们曾经非常喜爱我们的总统。看看这里从肯尼迪最高83%的支持率开始，最低也有56%。那是他的最低支持率。所以他的支持率一直维持在很高的区间。再看下09/11事件后的小布什二世。

It turns out we used to love our presidents. If you look here from Kennedy at 83%, his highest approval rating, His lowest was 56%. That was his lowest approval rating. So he operated in a very high band. Look at Bush two during after 09/11.

他的峰值是92%，最低时只有19%。对，战时总统。但接下来是特朗普一届、拜登以及特朗普第二届，历史上罕见的低高支持率。

92% was his peak. His lowest was 19. Right? Wartime president. But then you get to Trump one, Biden, and Trump two, historically low high approval.

他们的最高水位线，特朗普第一次是49，拜登是六十三，其中一次，然后特朗普第二次是47，最低的时候是二十九三十一大约四十。所以也许现在是时候出现第三方候选人了。我们来讨论一下吧，伙计们。

Their high watermark, 49 for Trump one, sixty three for Biden, one of one, and then 47 for Trump two, and their lowest twenty nine thirty one forty. So maybe it is time for a third party candidate. Let's discuss it, boys.

我不知道该怎么看这张图表。零 这个

I have no idea how to read this graph. Zero This

这是最差的。

is the worst.

我心想，这里发生了什么？

I'm like, what is happening here?

这是格式最糟糕的图表。这是一个令人困惑的图表，不过，我之所以把它放出来是为了引发讨论。所以你应该说谢谢

This is the worst formatted chart. This is a confusing chart, but, well, the reason I'm putting it up is for debate. So you should be saying thank

感谢参与辩论

you for debating

它引发了热烈的讨论。

that it's creating great debate.

你为什么把它放出来？

Why did you put it up?

再来看另一个。盖洛普民意调查显示，2023年美国人希望出现一个可行的第三党派的比例为63%。因此，这个比例正处于历史最高水平附近波动。

Here's another one. Gallup poll Americans desire for a viable third party 63 in 2023. So it's it's bumping along an all time high.

好的。我正非常专注地看这个。

K. I'm really concentrating on this one.