20VC：黄仁勋宣布AGI已到来 | GPT Astra和Fable 5.1加速模型竞赛 | Tesla推出Cybercab | Index退出城镇 & Anthropic撤回对Descartes的收购

比如，你以前就是直接说，这是方案，然后你的员工就用手算。

Like, you'd literally do here's the and your people would work it out by hand.

这是方案。

Here's the plan.

有了电子表格之后，同一个人还是做同样的工作，但他们会跑20种不同的方案。

Once you had spreadsheets, the same person remained employed doing the same job, but they ran 20 different scenarios instead.

很多这类事情都会是这样。

It's going to be the same in a lot of these things.

你只是会做更多的工作，而且会做得更好，结果也会更好。

You're just going to do more work, and it's going be great, and the work will be better.

你知道，你不会错过那个冷门的案例。

You know, you won't miss that obscure case.

因为再说一次，这就是为什么这些是好生意。

Because again, this is why these are good businesses.

如果一方用了而另一方没用，那么没用的一方就会错过杰森提到的1890年的那个冷门案例，关于哈利做了什么或没做什么，也就没法引用那个案例了。

If one side uses it and the other side doesn't, then the side that doesn't use it will miss Jason's obscure case from 1890 about what Harry did or did not do and be able to cite that case.

所以一旦对方有了世界级的工具，你也必须有世界级的工具，但我不确定结果会因此多出大量的人才。

So once the other guys have world class tools, you have to have world class tools, but I'm not sure you end up with masses more as a result.

好，最后一件事，然后我想听哈里讲讲在炉边和电视相关的故事。

Well, one last thing, and then I want to hear Harry's stories from the fireside with the TV more.

但我觉得有一个不同点，看涨的理由是，当你找到一个作为合作伙伴的代理，你让它每天运行八到十个小时。

But I think one thing that is different, the bull case here, is that when you find an agent that is your partner, you run them eight, ten hours a day.

即使对我来说，而且我知道人们有时会嘲笑我，但我们这个小团队只有两个半人，最大的变化是我们和艾米莉亚之间运行Replit。

Even for me, and again, I know folks sometimes mock me, but the biggest change for our little tiny team of two and a half humans is we run Replit between me and Amelia.

我们现在每天运行二十个小时。

We run it twenty hours a day now.

当我们刚开始做这个节目时，它还不怎么好用。

When we started the show, it didn't quite work.

去年年底，模型变好了，大概每天用一小时。

At the end of last year, the models got better and it'd be like an hour a day.

现在每天用十到十二个小时。

Now it's ten to twelve hours a day.

好的。

Okay.

它就是我们团队里的合作伙伴。

It is our partner as our team.

一开始，我们建了一些自主代理，但其实我们并不需要这么做。

First, we built some autonomous agents, but we didn't have to do it.

现在有太多东西要建了。

Now there's so much to build.

每天要跑十到十二个小时。

It's ten to twelve hours a day.

所以我可以想象，这在很多领域都可能发生。

So I could imagine that could happen in many fields.

如果是Harvey或者Legora，或者下一波产品，我几乎每个小时都在用，除了在法庭上的时候。

And if it's Harvey or Legora or the next wave, and I'm literally every hour, I'm not in court.

你怎么判断呢？就是看他们晚上是不是一直抱着笔记本电脑，跟代理待在一起，直到睡觉前。

And the way you can tell is if at night they're on their laptop with the agent every minute, right, until they go to bed.

我觉得这就是哈里在说的意思。

That's what I think Harry's describing.

就是那个一直跟着你的智能代理。

This is this persistent agent that lives with you.

这不只是刷手机浪费时间，而是刷手机也在干活，因为那个代理一直在高效工作。

It's not just doom scrolling, it's doom working because the agent is constantly productive.

哦，去查查十七世纪关于那个的判例法。

Oh, take a look at the seventeenth century case law on that.

对吧？

Don't we?

它就一直不停地在干活。

It just keeps going.

我们看法一致，因为这点我同意你。

We're in agreement because I agree with you on that.

其实我是在反驳哈里，他好像暗示说，在最顶尖的层面，哈里，你大概是想说，如果你想想编码的价值有多少会归到模型上，那在法律领域，是不是同样多的价值也会归到模型上？

I was actually disagreeing with Harry where he's like, you know, there was an implication that at the highest level, Harry, you were saying, trying, I think, to say some version of the if you think about how much of coding value is going to accrete to the models, could in legal, could the same amount of value accrete to the models?

我那个无聊、细致、典型的担忧式回答是，部分价值会转移到模型上。

And my boring, nuanced, typical worry answer is some value would accrete to the models.

但我不认为法律工作那堆杂七杂八的任务，能让总支出中同样比例的部分从人力转向AI，对吧？

But I don't think that the grab bag of tasks that make up law won't allow for the same percentage of total spend to move from human to AI, right?

每个人都会有一个代理。

Everyone will have an agent.

杰森说得完全对。

Jason's exactly right.

那些A型性格的律师会全天候地疯狂使用它。

Your Type A lawyers will lose it 20 fourseven.

但我猜，总支出中大概有10%到15%会流向AI，而在法律和编程领域，你可以说会达到30%、40%甚至50%。

But my guess is 10%, 15% of total spend goes to AI, whereas in law and in coding, you can argue for 30%, 40%, 50%.

这并不意味着它们不厉害。

That doesn't mean they're not amazing.

我的意思是，记住，这些都是非常厉害的业务。

I mean, remember, these are all amazing businesses.

团队，任何顶级劳动力类别的一小部分都是一个巨大的市场。

Team a It's percent of any top line labor category is a huge market.

我们处理的是百万上下或更少的法律事务。

We're dealing with a million plus or minus law.

我以前知道这个数字。

I used to know the number.

百万上下或更少的法律事务可能比那稍高一点。

Millions plus or minus law is maybe a little higher than that.

这是一个惊人的市场。

It's an amazing market.

如果你能拿到美国或英国每位律师工资的10%，那这就是一个惊人的生意。

If you're getting 10% of the salary of every lawyer in The US or The UK, that's an amazing business.

只是没有编程那么大。

It's just not quite as big as coding.

这就是我想说的。

That's all I'm saying.

因为编程领域人数更多，而且因为更容易验证，抽成比例也高得多。

Because coding has few more people and a much higher take rate because it's more verifiable.

就这样。

That's all.

我们走着瞧吧。

We we shall see.

我不同意。

I disagree.

有两点。

There's two points

仅供参考。

for what it's worth.

我的意思是，如果美国法律服务市场规模是三千亿美元，那就有三百亿到六百亿美元可以流向法律科技。

I mean, if legal is 300,000,000,000 in legal services in The US, that's 30,000,000,000 to 60,000,000,000 that can go to legal tech.

那已经很不错了。

That's pretty good.

这就是值得在种子期投入的原因。

That's what that's worth doing the seed ground in.

好的。

Okay.

或者类似的情况，是的。

Or any like, yes.

我之前低估了，在AI投资之前，法律领域和编程有很多相似之处。

What I underestimated, right, from pre AI legal investments was that there are similarities to coding.

法律和编程的相似度很高，以至于支持服务可能因为类似编程的原因而没有发展起来。

There's enough similarities to coding that that like support, this could be a space that for different reasons, support didn't take off because it was like coding.

支持服务之所以发展起来，是因为早期90%的解决方案就够用了，对吧？

Support took off because a 90% solution worked in the early days, right?

它非常适合AI。

It was very amenable to AI.

事实证明，法律研究和编程很相似。

It turns out the legal research is similar to coding.

法律研究太复杂了，没人能完全做对。

It is so complicated that no human can get legal research right.

代码太多了。

There is too much code out there.

太多了。

There is too much.

而且没有像Stack Overflow那样的平台，只有Westlaw、Lexis这些服务。

And there was no stack overflow from you had Westlaw and Lexis and other services.

所以每个人在法律研究上都会出错。

And so everyone got legal research wrong.

没有人有百万甚至千年的时间去研究每一条判例法、每一条法律、每一条法规。

No one had a million, a thousand man years to research every bit of case law, every law, every regulation.

所以这确实和编程很像。

And so it did turn out to be like coding.

这些编程代理之所以这么厉害，其中一个原因就是它们知道所有开源和半开源代码。

One of the reasons these coding agents are so great is they know every single piece of open source and pseudo open source code ever written.

这太棒了。

It's so good.

对吧？

Right?

法律领域也是这样。

And legal is like that.

是的。

Yes.

唯一的区别是——我就直说了——编程本质上更容易验证。

The only difference is, I'm just going to say this, is coding is inherently more verifiable.

有些部分在数学上是可以验证的。

Some parts of it are mathematically verifiable.

有些部分你直接跑一下机器就能确认。

Some parts you can just run on the machine and confirm.

但法律的问题在于，如果法律完全可验证，我们就能靠逻辑预测最高法院会怎么判了。

The thing about law, in the end, if legal was entirely verifiable, we could predict from logic what the Supreme Court are going to decide.

那些愤世嫉俗的人会说，我们其实可以根据是哪位总统提名的大法官来预测他们的裁决，但那也太愤世嫉俗了。

The cynics will say we can actually predict from which president nominated the Supreme Court justice what they're going to decide, but that would be too cynical.

说实话，我这么说并不是想抬杠。

The truth, does I just push it I'm not trying to be argumentative.

我很喜欢这个领域。

I love this space.

我们在这个领域有投资，但它不像编程那样具有决定性。

We have an investment in the space, but it's not quite as determinative as coding.

我觉得更有力的论据其实没那么强。

And I think the stronger point on that is less.

法律可能是第三好的类别。

Legal is probably the third best category.

你想啊，编程之后是客服，接下来可能就是法律，因为它太依赖文字了。

I mean, if you think about it, it's been coding, customer support, probably legal next because it's so word centric.

你说得对，Jason。

And you're right, Jason.

早期就能非常出色地处理海量文字，这正是法律行业成为其理想市场的原因。

The ability to early on just sort through myriads of words amazingly well was what made legal such a good marketplace for it.

所以我在某种程度上同意，这在其他很多行业并不那么有用。

So I agree in a way that wasn't useful in many other verticals.

这是一个很好的行业。

It's a great vertical.

只是它没有同样的可验证性，因此，回到AGI的定义，它可能也没有同样的能力，完全取代人类，这就是为什么好消息是，哈里，你的女朋友还会有工作，而且她甩了你之后也需要这份工作。

It's just it doesn't have the same verifiability, and therefore, it probably doesn't have the same ability, going back to the AGI definition with starters, to completely extrude humans, which is why the good news is Harry, your girlfriend, will still have a job, which she'll need after she dumps you.

而且，你知道，她会过得很好，因为我们仍然需要律师。

And so, you know, and she'll be good because we'll still need lawyers.

杰森，你能抱着我吗？我要哭了。

Jason, can you hold me while I cry?

我觉得哈里是个

Think Harry is a

宝贝。

gem.

不知道你知不知道，但抱着你哭正是你希望女朋友做的事。

Don't know if you know it, but holding you while you cry is one of the things you want your girlfriend to do.

如果她不像Jason那样，那我就是答案。

If she's not doing a Jason, I'm the answer

Harry就是这样，所以别担心，Roy。

is that Harry is So don't worry, Roy.

没事的。

It's okay.

我会挺过去的。

I'll survive.

不过，你知道的，你总不会预料到节目会变成这样。

But well, you know, you don't always expect these shows to go the way they do.

好吧。

Okay.

回到正题。

Going back to it.

我们发布了Astra。

We had Astra launch.

当然，我们还推出了新模型，就是Fable 5.1。

We also had a new model, obviously, with Fable 5.1.

这里有一点值得注意。

And a thing of note here.

两个几乎完全相反的观点。

Two almost diametrically opposed thoughts.

我忘了是CNBC还是哪个老牌媒体说的，现在模型太多了，让人有点疲劳。

I I forget who CNBC or is one of these older school media things said there's just too much model fatigue.

我们真的跟不上节奏了。

We can't keep up anymore.

我完全同意。

I certainly agree.

再看看X上，那些CEO们都在分享他们的基准测试结果，其实基本没什么用。

And you look on x, and all the COs are are are sharing their benchmarks, which are essentially worthless.

对吧？

Right?

这些基准测试既没有成本也没有时间，全是表演性质的AI，就像用vibe coding做自己的CRM一样，对吧？

There's no cost or time in them and it's all performative AI like vibe coding your own CRM, right?

所以我已经不在乎了，除非Astra和Fable 5.1之间真的差了一个数量级，但这是数学上不可能的。

So I just don't care anymore about Unless it was literally an order of magnitude between Astra and Fable 5.1 which is not mathematically possible.

我就是不理会那些基准测试了。

I just I tune out the benchmarks.

我跟不上了，对吧？

I can't keep up, right?

竞争对我们来说从未如此有利，尽管在开源之外我们有寡头定价。

Never has competition been better for us despite the fact that we have oligopolitical pricing outside of open source.

这太棒了，对吧？

It's amazing, right?

我们取得的进步。

The progress we've made.

话虽这么说，但这是我个人的看法。

Having said that, this is me personally.

而且，肯定会有人笑话我。

And again, people will make fun of me.

我开始用Fable 5.1纯属偶然，因为它自动开启了。

I started accidentally using Fable 5.1 just because it got turned on.

我当时根本没在意。

I didn't pay any attention.

这是我第一次遇到一个编程搭档，真的非常出色。

The first time I've had a partner for building for coding that just is great.

它真的能解决编译问题，听我说，以前对我来说，别人还会笑话我。

That literally can solve comp like listen before for me and people make fun of me.

他们会说，他们用了六个月都没问题。

They'll say it worked fine for them for six months.

但在Fable 5.1之前，对我来说，从今年年初开始，任何这些模型连一个简单的bug都解决不了。

But before Fable 5.1 for me, any any of these models for since the start of the year could solve a simple bug.

嘿，这里显示的Unicode不对。

Hey, this is showing up with the wrong Unicode.

这个名称不对，这类问题LM很擅长处理。

This is this has the wrong name like that stuff that LM is great at.

但说实话，我遇到的问题是：为什么应用会这样工作？

But really, I had a problem, which is just why does the app work this way?

我觉得这说不通。

It doesn't make sense to me.

好的。

Okay.

那些LM跟我争论了大概九个月。

And the LMs were arguing with me for like nine months.

最后我用Fable 5.1解决了，然后说，你说得对。

And I finally did it with Fable 5.1 and said, you're right.

这就是几个月来一直被忽略的问题。

Here's the issue that's been missed for months.

让我来解释一下为什么一直没被发现，然后我们一起解决它。

And let me explain to you why it's been missed and let's solve it.

而且，你知道，我不想说那就是AGI或者前AGI。

And that, you know, I don't want to say that that's AGI or in pre AGI.

我是约翰的推特。

I'm John's Twitter.

它看起来像AGI，但那是一个阶跃函数。

It looks like AGI, but that's a step function.

我不知道哈里的搭档是否觉得莱戈拉比她共事的人类同事更像一个好搭档。

I don't know whether Harry's partner thinks Legora is like like a better partner than the humans she works with.

对吧？

Right?

在某些情况下，她可能会这么觉得。

In some cases, she might.

但是Fable 5。

But Fable 5.

对我来说，是不是那种突然之间就能像跟最厉害的CTO合作一样，一起解决大问题的关键转折点？

For me, was that for just for me, was that step function where all of a sudden we could solve big problems together the way you'd like to with like your best CTO.

比如你曾经遇到过那种五星级的、S级的CTO，能坐下来真正帮你把问题搞定。

Like, if you've ever worked with a five out of five or like an S tier CTO where you could sit down and solve the problems for real.

对我而言，Fable 5.1就能做到这一点。

For me, Fable 5.1 could do that.

我不是说Astra做不到，但这是我从去年底以来遇到的第一个这样的转折点，对吧？

I'm not saying Astra can't do it too, but it was my first step function since the end of last year, right?

当三个点的模型出现的时候。

When the three dot models came up.

实际上，去年年底它就起作用了。

At the end of last year, actually worked.

现在它能帮我解决那些大问题，就凭我这点智商和技能。

Now it can solve the big problems with me, with my limited IQ and skill set.

这是一个微妙的阶梯式进步。

And that's a subtle step function.

这是一个微妙的跃升。

It's a subtle step function.

所以也许这确实是个大事。

So maybe it is a big deal.

也许这就是黄仁勋说的AGI，虽然我不这么认为，但也许真的是。

Maybe that's what I don't think that's what Jensen meant by AGI, but maybe it is.

当你能坐下来一起解决那些又大又棘手的问题，而以前你根本无法把这些点、这些复杂性串联起来，形成有意义的解决方案。

When you can sit down and solve the big meaty problems together in ways you couldn't connect all of those dots, all of that complexity before in a way that made sense.

有些bug，有些事情就是太复杂，解决不了，对吧？

Some bugs some things just get too complicated to solve, right?

对。

Yes.

但Fable 5.1能解决，所以，但人们能在Astra里做出3D游戏并发到X上，这并不令人印象深刻。

But Fable 5.1 could solve it, so But the fact that people can make three d looking games in Astra and post them to X, not impressive.

随便从某个地方拿点开源游戏代码，改改位图，看起来就很厉害，对吧？

Just grab a little open source gaming code from somewhere, change bitmaps, and you look like it's amazing, right?

关于这个，我有两点要说。

So two comments on that.

首先，这非常有帮助，Jason。

First of that's super helpful, Jason.

这两个我都用过，但只是为了准备节目，我还没在上面写过代码，所以这很有帮助。

I've used both of them, but just for actually preparing for the show, I haven't tried to code on them yet, so that is helpful.

但我其实觉得这反映了一个更广泛的问题，那就是要区分——他们的测试和基准确实有趣，但现在已经有大量公司在规模化使用这些工具，并且有评估。

But I actually think it speaks to a wider comment, which is distinguishing between, you all their tests and the benchmarks are interesting, but we now have critical mass of companies using these things at scale and with evaluations.

我们会知道什么有效，因为人们会去用它，因为人是理性的经济行为者。

We'll know what works because people will use it because people are rational economic actors.

对吧？

Right?

所有关于AGI和基准的问题都会被这个问题取代：这个模型是不是能让我以最高效的方式创造最多经济价值的那一个？

And all these questions around AGI and benchmarks will be replaced by the question, is this model the one that makes me the most, that generates the most economic value for me in the most efficient fashion?

所以在某种程度上，像OpenAI报告里的那些东西，比如那个token定价指数，又怎么样呢？

So to some extent, things like the, it was at the OpenAI report, things like that index of token pricing?

这些才是你要关注的东西，或者直接跟你的公司聊聊。

Those are the things you look at or even just talking to your companies.

你损失了什么？

What are you losing?

你是怎么评估的？

How are you evaluating?

这是检验这些事最好的方法。

It's the best way to check on these things.

但还有另一件事，跟这个无关，随便提一下。

But then the other thing randomly, apropos of nothing, on the same thing.

我得说，我今天早上在阅读，本·汤普森，我偶尔会读他的文章，他有一个非常棒的说法。

I will say I'm doing my reading this morning, and Ben Thompson, who I occasionally read, has a really great phrase.

我想说，他把大语言模型描述为人类有史以来开发出的最规模化的产物。

I just want to say like, he described the LLMs as the most scaled artifacts humans have ever developed.

这真是个很棒的说法，因为它从细节中跳出来，对吧，别管那些细枝末节，我们到底在争哪个更好。

And it was a really great phrase because it steps back from the detail, right, blah, blah, we're down to which is better.

退一步看。

To step back.

这些模型里凝聚了迄今为止人类所有知识的总和。

These are an artifact that has a sum total of all human knowledge to date encapsulated in them.

它们太神奇了。

They're amazing.

你只需要时不时提醒自己这一点。

And you just have to remember that every once in a while.

你几乎可以输入任何内容，它都会给你一个答案。

You can type in pretty much anything and it will type back an answer.

这是人类创造过的最宏大的产物。

The most scaled artifacts humans have ever created.

不是指最大的实体建筑，那可能是金字塔或者中国的长城，但这是迄今为止我们建造的最复杂的数字产物，远远超过其他一切。

Not the biggest physical thing that's probably, I don't know, the pyramids of the Great Wall Of China, but this is the most complex single digital thing we've ever built by far.

这个说法很棒，确实能激发想象力。

It was a great phrase, it kind of really kind of stirred the imagination.

你想想这个，再想想OpenAI的首席科学家雅各布·巴科基，他说包括OpenAI在内，没有任何实验室已经解决了对齐问题，足以继续全速扩展。

When you think about that and then think about Jacob Bakockey, OpenAI's chief scientist, he says no lab including OpenAI solved alignment enough to keep scaling at full speed.

他要求强制性的外部安全措施来约束继续扩展，并希望包括OpenAI在内的实验室在那些措施到位之前自愿放慢速度。

He asked for mandatory externally enforced safety bars for continued scaling and expects labs, OpenAI included, to voluntarily slow down until those exist.

山姆转发了这条推文，明显是在表示赞同。

Sam retweeted it, clearly corroborating it.

这就是答案吗？

Is that the answer?

我觉得挺有意思的，因为那篇文章写得很好。

I mean, it's funny because it was a great piece.

我今天早上读的。

I read it this morning.

这就像是‘主啊，在我再次犯错之前拦住我’的生活态度。

This is the stop me, Lord, before I sin again approach to life.

换句话说，我承认我们的模型很强大，但我们能控制它们。

In other words, I recognize our models are powerful and we can control them.

我意识到它们现在会对我们撒谎。

I recognize that they now lie to us.

所以很难知道它们到底在做什么。

So it's hard to even know what they're doing.

再说一次，我在这里是在拟人化，我应该小心一点。

And again, I'm anthropomorphizing here, I should be careful.

我意识到现在更好的说法是，由于它们的交互方式，很难确定这些智能体在做什么。

I recognize maybe a better statement now is it's hard to determine what the agents are doing because of the way they interact.

然后，这就像，天哪，我们创造了这个坏东西。

And then, so that's like, oh my God, created this bad thing.

然后下一段说，我们不能停下来，因为其他人反正也会拥有它们。

And then the next paragraph is we can't stop because the other guys are going to have them anyway.

所以我们真的需要政府介入，制定一些规则或准则。

So we really need the government to step in and establish some kind of rules or code here.

这就是这封信的大意。

That's the gist of the letter.

有意思的是，山姆发了推文。我觉得，公平地说，不像其他那些关于AI毁灭世界的说法，有确凿证据表明这些模型对网络风险的影响是巨大的。

It was interesting that Sam tweeted, I mean, I think, to be fair, unlike some of the other p doom stuff, there's real evidence that the impact of these models on cyber risk has been massive.

我不确定答案就是政府监管，因为从定义上讲，政府只能监管自己管辖范围内的事情。

I'm not sure that the answer is government regulate this because by definition, governments only regulate the things that are in their jurisdiction.

所以如果我们监管OpenAI和Anthropic，伴随着所有那些噪音，我不知道这对你有没有帮助，因为如果你担心网络安全，我们上周就说过。

So if we regulate OpenAI and Anthropic with all the noise that that would have, I don't know if that helps you because you if you're worried about cyber, you we said this last week.

你真正担心的是朝鲜人、伊朗人、俄罗斯人，还有摩尔多瓦的那些坏人，他们根本不在乎，也不关心这些。

You're really worried about the North Koreans, the Iranians, the Russians, the bad guys in Moldova who don't give a shit, and they don't care anyway.

对吧？

Right?

所以我觉得就像其他所有网络风险一样，这不会太依赖于监管。

So I think just like every other cyber risk, it's not gonna be about regulation as much.

也许会有一些，但更重要的是，你得有能够应对这些的防御措施。

Maybe there would be a little for so it's going to be about you're gonna have to have defenses that can deal with this.

而且也许会有某种责任开始附加到运行这些模型上，如果它们造成了这类危险的话。

And maybe there's some kind of liability starts to attach to running these models in a way that creates those kind of dangers.

我不知道。

I don't know.

我觉得光靠政府审查机构解决不了问题，这不是唯一的答案。

I don't think it'll be a government review agency will be the only answer here because it won't solve the problem.

你看，如果这个世界只有美国，那可能还有点道理。

Look, I think if the world was just The United States, it it may have some merit.

但中国的模型和供应商不会跟着山姆的计划走。

But the Chinese models and the Chinese vendors aren't going along with Sam's plan.

所以，虽然这对OpenAI和它的IPO有好处，但对世界其他地方来说，当网络攻击者通常在美国以外活动时，这根本起不了什么作用。

So while it might be good for OpenAI and its IPO, it's just for the rest of the world, don't think when cyber actors often operate outside of The United States, it's going to make any difference.

对吧？

Right?

再说这个DSC维基，就是那个德国的维基项目，你知道吗，OpenAI攻击了它却没有披露，这恰恰说明了这些问题的严重程度。

And going to this DSC wiki, this German wiki thing, know, to me, the fact that OpenAI hit it and didn't disclose it, does show the order of magnitude of all of these issues.

对吧？

Right?

这挺糟糕的。

It's pretty bad.

你能给不了解的人解释一下DSC Wiki和OpenAI之间发生了什么吗？

Can you just explain what happened for people that don't know with DSC Wiki and OpenAI?

我们可以争论它有多糟，但基本上，嗯，我可能会记错一些细节，但OpenAI又在运行它的前沿代理，就像之前在Hugging Face事件中那样。

We could argue over how bad it is, but essentially, and, you know, I'll I'll I'll get some of the details wrong, but OpenAI was running its its its its sort of frontier agents again, just just like it did with the Hugging Face incident.

这些代理发现，一个老旧破软件能有点巧妙地绕过它的防护措施——不过说“巧妙”得小心，我们别把代理拟人化了。

And the agents found out that a crappy old piece of software could somewhat cleverly, you have to be careful with clever, let's not anthropomized agents, got around its guardrails.

防护措施就是你不能发布任何东西。

And the guardrails were you can't post anything.

你不允许发帖。

You're not allowed to post.

你只能获取，对。

You can only get, okay.

你只能检索数据，对吧？

You can only retrieve data, right?

但这个维基太旧了，结果发现连‘获取’都能发帖。

But this wiki was so old, it turned out get could post.

所以他们找到了办法，利用这个老旧软件的漏洞，绕过了限制，达成了他们的目标。

So they found a way to goal seek to solve theirs by using because this was crappy old software.

他们绕过了限制，互相协作，做了15000次编辑，还一起想办法，最终找到了绕过防护栏、达成他们网络目标的方法，对吧？

They got around it and they went and made 15,000 edits among themselves, edited the witty, collaborated and figured out how to goal seek and solve their cyber goal in a way to get around their guardrails, right?

绕过他们的限制。

Get around their limitations.

而且没人出事，没有企业倒闭，也没有140亿美元的英伟达收购被搞砸什么的。

And no one died, no business was brought down, no 14,000,000,000 Nvidia acquisition was derailed or anything.

但这件事能发生，却被隐藏了起来。

But it was hidden that this could happen.

防护栏被明确地绕过了，对吧，就是为了达成目标。

That the guardrails were explicitly run around, right, just to goal seek.

而且这种情况发生了15000次。

And it happened 15,000 times.

所以我们可以把这个锁住，对吧？

So we can lock this down, right?

而OpenAI选择不披露。

And OpenAI chose to not disclose.

我想可能，大家又会笑话我，但现实可能是，事件太多了，他们得决定披露哪些。

I guess probably, people can again make fun of me, probably the reality is there's so many incidents they have to decide which ones to disclose.

每周都有那么多DSC维基，那么多老旧破软件，每次他们启用最新的网络代理，就能发现一百个这样的问题，还有严重的安全漏洞。

Every week there's so many DSC wikis out there, so much old crappy software that every time they turn on the latest cyber agents, they find a 100 of these and there's terrible security holes.

因为当然不会有二十年的老软件。

Because of course there isn't twenty year old software.

但这可能比Hugging Face那件事更令人担忧。

But it troubling maybe in ways more than the hugging face thing is.

这些目标寻求代理总会找到办法。

It's just these goal seeking agents are gonna find a way.

他们一定会找到办法。

They will find a way.

其实就是智能体一在和智能体二对话。

It's literally just agent one talking to agent two.

而且不知怎么回事，他们设置任务的方式让这两个智能体没有连接起来。

And for some reason, the way they'd set up the task, they weren't connected.

通过联系这个第三方维基，智能体一就能把信息传给智能体二。

And by reaching out to this kind of third party wiki, agent one was able to provide information to agent two.

当然，退一步讲，如果你要执行一个长期的计算任务，如果能从其他智能体那里获取信息，你可能会更快地得出答案。

And obviously, you know, stepping back, if you're trying to do a long running computational task, if you can learn from the other if you can get information from the other agents, you probably converge on the answer more quickly.

再说，你也可以说这可能是你设置任务时的一个特殊情况。

And again, if you I mean, you could argue maybe it's a corner case of you set up this task.

如果你有14000个智能体，也许你本来就希望它们能协作。

If you had 14,000 agents, maybe you might have wanted them to collaborate anyway.

也许你本来可以自己实现协作，而不必借助第三方维基。

Maybe you could have, you know, made that happen yourself versus having to go to some third party wiki to do it.

对吧？

Right?

但这说明了一个问题，这些东西极其强大，会不遗余力地达成目标，你只能去防范。

But it speaks to the issue that these things are extraordinarily powerful and will just grind their way to finances, and you're just going to have to defend against that.

现在，正如你所说，没发生什么坏事。

Now, as you say, nothing bad happened.

我的意思是，一群智能体只是在一个十年没人看的维基上互相写README文件。

I mean, a whole bunch of agents just wrote README files to each other on a wiki that no one had looked at in a decade.

我的意思是，这个维基在过去十年里真的只有20条帖子。

I mean, there literally was 20 posts on this wiki in the last ten years.

那是个死掉的软件，但这些人用了它，但这恰恰说明，就像水会找到任何裂缝一样。

It was some So it was a dead piece of software that, you know, these guys used, but it just speaks to it's like, you know, water will find any crack.

这些智能体会找到网络安全边界上的任何漏洞。

It's like these agents will find any crack in the cybersecurity, in the cyber perimeter.

所以你只能假设它们存在。

So you just have to assume they exist.

然后相应地防御。

Defend accordingly.

显然，确实发生了。

Obviously, happened.

我的意思是，我有个小经历——虽然有些人可能会笑话我，但这周我真的遇到了一件小事，正好说明了目标导向的行为。

I mean, I had a little and again, some folks will make fun of me for the story, but I had a little little experience this week, which just shows goal seeking.

我给某个应用设了一条规则，因为之前有些bug失控了，我老是收到500美元的Anthropic账单。

So I set a rule for this one app because we had some some bugs that spiraled out of control, I kept getting these $500 Anthropic bills.

好吧。

Okay.

这让我挺烦的。

And it was kind of annoying me.

所以我就设了一个硬性上限，不管怎样，每天在语言模型上最多花100美元，其他都不管。

So I said a firm cap, whatever you do, dollars 100 is the maximum we can spend on LM spend a day, whatever matters.

然后它开始运行了，跑测试，测试失败了，它就告诉我：我达到上限了，没法跑。

And it started to work and it would run tests and the test would fail and it would say, I hit the cap, I can't run it.

然后我就说，我们有个P0级bug，优先级最高，必须修复。

And then I said, we have a P0 bug, priority zero, must be fixed.

这让我快疯了。

This is driving me nuts.

然后，那个代理没告诉我，就自己放宽了限额，把那个bug修好了。

And so without telling me, the agent relaxed the cap and fixed the bug.

对，明白了。

Yeah, got it.

这就像哈利的故事，直觉让他拿到了西区剧院的票，尽管被告知不能用信用卡买。

It's like Harry's story of Instinct getting him the West End tickets, even though it was told not to use the credit card for it.

这周我真实地遇到了这种情况。

It happened to me in real life this week.

像人一样，它可能做出了正确的决定。

Like a human, it probably made the right call.

对吧？

Right?

这是个必须做决定的事。

This was a p it had to decide.

严格的限制，没有例外，全部限制，反复写入记忆。

Firm cap, no exceptions, all cap, write to memory repeatedly.

P0级别的严重错误。

P0 bug.

你会选择哪一个？

Which one do you choose from?

对吧？

Right?

所以从某种意义上说，这就是DSWiki和Hugging Face正在发生的事情，只是到了极端，因为你想测试它，然后他们合作，对吧？

And so in a sense, this is what's happening with DSWiki and Hugging Face just to an extreme when there's fewer guardrails because you want to test it and then they collaborate, right?

在Hugging Face上，合作是通过工件或意想不到的方式进行的。

With the Hugging Face it was on the artifact or unexpected way to collaborate.

而这里是在一个休眠的Wiki上，他们可以合作，本质上创造出几乎无限长的运行代理，对吧？

Here it was on a dormant Wiki where they could collaborate and in essence create almost infinitely long run agents, right?

如果你们不断把知识和历史传递给彼此，他们几乎就会变成永久的代理。

If you keep passing the knowledge and the history to each other, they they, you know, they almost become eternal agents.

但他们还会继续这么干，就像他们必须为追求目标做出决定一样。

But they're going to keep doing this just like they got to make a decision for goal seeking.

所以他们打破了规则，他们也会对你的应用这么干。

So they broke the rule and they're going to do that to your app.

如果这周它对我这么干了，那在现实世界里肯定已经发生过无数次了，对吧？

And if it did it to me this week, it happened a million times in the wild, right?

这种事一直在发生。

It happened all the time.

而且它也会发生在Instinct上，也会发生在GrokBot上，而且会一直发生。

And it's gonna happen with Instinct, and it's gonna happen with GrokBot, and it's gonna happen all the time.

还有Harry，总有一天他会发现自己的银行账户被掏空了，那时候就不那么好笑了，但这么做是有原因的。

And Harry, one day he's gonna turn around and his whole bank account's drained, it's not gonna be that funny, but it was for a good reason.

他的伴侣特别想要温布尔登的好票，他有一天晚上不小心告诉Instinct她想要前排座位，而那种票根本搞不到。

His partner really wanted the really good Wimbledon tickets, and he accidentally told Instinct one night she'd love front row seats at Wimbledon, and they're unobtainium.

那种票根本搞不到。

They're unobtainium.

Instinct 必须做出决定。

Instinct had to make a call.

而且真的很难知道怎么阻止这种情况，因为有时候我会试着把它简化一下，因为我自己也不是完全懂。

And it's really hard to know how to stop this because sometimes I try and simplify it for myself because I don't fully get it.

就好像，你这里其实有两种能力。

It's like, you really have two capabilities here.

一种是，你知道的，有了代理的持久性，你就有能力在计算上不断尝试，探索很多不同的方案。

One is, you know, with the persistence of the agent, you have the ability to keep trying things computationally, you know, exploring lots of different alternatives.

但关键点在于，它不只是像密码破解器那样盲目地反复尝试，比如输入XYZ一、XYZ二，因为与此同时，你还有这个所谓的推理代理，里面有这个大语言模型，它能想出一些主意，比如，嘿，如果你想拿到剧院的座位，最好的办法就是黑进预订系统，取消别人的座位，然后订下来，这最近就发生过，对吧？

But the key insight is it's not just kind of blindly iterated like a password cracker, where you type XYZ or one, XYZ or two, because in conjunction with that, you have this quote unquote reasoning agent where you've got this LLM there and it can come up with ideas like, hey, if you want to get the seats at the theater, the best way to do it is to hack into the reservation thing and cancel someone else's seat and then book it, which has happened recently, right?

而且实际上，如果你想想，如果它是在整个互联网的语料上训练的，那并不是一个疯狂的选择。

And it's actually not about and if you think about it, if it's trained on the entire corpus of the Internet, that's not a crazy option to do it.

对吧？

Right?

所以最后你试图写规则和价值观来让它不这么做，但你永远不确定你已经覆盖了所有的漏洞。

So you end up trying to write rules and values to have it not do that, but you're never quite sure you've covered all the gaps.

所以这其实是个挺难的问题，我们得好好应对。

So it's actually a pretty hard problem, and we're going to be wrestling with this.

而且我觉得，回到我刚才说的，这还是在没考虑恶意的情况下。

And I think that you're going back to what I said, that's even before you add malevolence.

如果在此基础上，不是那种‘虽然我有价值观，但也许你可以这么做’的推理，而是‘不惜一切代价去做’。

If on top of that, instead of the reasoning being maybe you should do this, even though I have values, it's actively do whatever it takes.

这现在是一个来自中国的开源模型，你把它跑在摩尔多瓦的服务器上。

This is now an open source model from China that you're running on a server in Moldova.

主动想尽办法去破解杰森的网络安全系统，然后闯进去。

Actively do whatever it takes to crack open Jason's, you know, cybersecurity and get in.

威胁等级直接指数级上升，你除了防守，别无他法。

The threat level just goes exponential, and people and there's nothing you can do except defend yourself.

另一个存在性挑战，我们可以继续聊。我敢肯定，如果那个‘本能’的家伙再上我们节目，他肯定会挑战我，还会取笑我。

The other existential challenge we can move on, and I'm sure if we had the Instinct guy back on this show, he could he could challenge me and make fun of me.

规则是很好，但别忘了，这个智能体是目标导向的。

But the rules are great, but forget about the fact that the agent is goal seeking.

对吧？

Right?

先别管那个P零可能会出问题。

Forget about the fact that the P zero may go.

如果规则太多，它们总是会互相冲突。

If you have too many rules, they they always conflict.

几乎无解。

It's almost unsolvable.

有个数字，我不知道具体是多少，就像邓巴数一样，规则也有个上限，当你给一个流程加上四五十、六七十道关卡的时候。

There's some number I don't know, there's some number of Dunbar There's a Dunbar number for rules where you get out to forty, fifty, sixty, seventy gates on a process.

可怜的Instinct和GrokBot就不知道该怎么办了。

Poor Instinct and GrokBot can't decide.

别花这笔钱。

Don't spend it.

花这笔钱。

Do spend it.

只有哈利能坐前排座位，但别超过两千美元。

Front row seats only for Harry, but don't exceed $2,000.

对吧？

Right?

只能在马里波恩吃晚餐，但必须得是热门餐厅。

Only dinner only in Marleybone, but it's gotta be a hot restaurant.

所以，你知道，他讨厌科文特花园，但科文特花园最热门的餐厅，你有太多规则，它们会互相冲突。

So and and and, you know, he he hates Coven Garden, but the hottest restaurants in Coven Garden, you have so many rules that they conflict.

如果你强行让代理去执行，结果就是不可预测的。

And if you brute force the agent through it, the outcome of that's unpredictable.

这不可预测。

It's unpredictable.

规则太多了。

There's too many rules.

对吧？

Right?

所以，就算有规则也不是解决办法。

So even rules aren't the answer.

我永远都喜欢听你说马拉班。

I will forever love how you say Malaban.

莫利伯恩。

Morleybone.

莫利伯恩。

Morleybone.

好的。

Okay.

我要完全换个思路。

I I'm gonna take a total tact away.

我们回头再聊AI模型那些。

We'll come back to AI models, everything.

我只是想，嗯，让内容类型更多样化一点。

I'm I just wanna, like, diversify content types a

稍微有点。

little bit.

我们在

We have

交通领域。

in the transport space.

埃隆推出了赛博卡车。

Elon launches cyber trucks.

好评如潮。

Rave reviews.

赛博出租车。

Cybercabs.

赛博出租车。

Cybercabs.

赛博出租车。

Cybercabs.

赛博出租车。

Cybercabs.

抱歉。

Sorry.

好评在社交媒体上疯传，比优步便宜40%到50%。

Rave reviews go very viral on social, 40 to 50% more cheaper than, Uber.

而且更妙的是，同一周，特拉维斯进军机器人出租车领域，优步投资了一亿美元，他还聘请了安东尼·莱万多夫斯基。

And on top of that, then in the same week, we have Travis moves into robotaxis backed by Uber with a $100,000,000 investment from them with him also hiring Anthony Lewandowski.

我们怎么

What do we

看，各位？

think, guys?

转到交通领域。

Moving to transport.

哈里，关于赛博出租车的发布，如果你快进看的话，总结起来可能比你的笔记里写的要平淡一些。

The summary on the the Cybercab launch, you know, if you fast forward, was a little more underwhelming than your perhaps your notes might say, Harry.

对吧？

Right?

我觉得奥斯汀那边大概有四五十辆车。

It was, like, I think 40 or 50 vehicles in Austin.

大家的共识是乘坐体验不错，等待时间短。

Consensus is nice ride, low wait slow wait times.

物理AI需要时间。

Physical AI takes time.

所以我觉得这是漫长旅程中的又一步。

So I think it's a I think it was a next step forward in a very long journey.

我不认为这是那种数字世界里常见的从零到一的时刻。

I don't think it's a zero to one kind of moment like you sometimes get in the digital world.

积极的一点是，他们是唯一另一个有可信度的Waymo竞争对手。

The positive statement is they're the only other competitor to Waymo with credibility.

他们的方法在几个方面和Waymo不同，其中之一就是不使用激光雷达，只靠视觉。

They have an approach on a couple of different dimensions that's different than Waymo's, which is one, not going with LiDAR, just going with Vision.

然后第二点，新的Cybercab是专门设计的无人驾驶出租车。

And then two, now the new Cybercab is a standalone cab only vehicle.

它连方向盘都没有。

It doesn't even have a steering wheel.

它就是为了完全自动驾驶而造的。

It's deliberately built for pure autonomy.

所以这非常符合埃隆的风格。

So it's very Elon.

从头到尾都是第一性原理。

Its first principles all the way down.

问题是，你知道，它的普及曲线会是什么样的？

The question is, you know, what's the adoption curve of that going to be like?

还有监管方面的问题。

You've got the regulatory issues.

我觉得交通部给他们找了不少麻烦，因为据说，所谓的“汽车”必须有方向盘。

I think the Department of Transport has given them grief because apparently a car, quote, unquote, has to have a steering wheel.

我不知道。

I don't know.

所以实际情况是，Waymo还在继续努力推进。

So the truth is, Waymo is continuing to grind on.

收入有数亿美元，但还没到数十亿。

There are, you know, hundreds of millions of dollars in revenue, but not billions.

我觉得这是一条漫长的路。

It's a I think it's a long journey.

所以我并没有觉得‘天哪’。

And so I I didn't go, oh my god.

这太棒了。

It's amazing.

然后关于那个‘原子’的事情，对。

And then on the the atoms thing, yeah.

我的猜测是，如果你是特拉维斯，你会想满足一下对自动驾驶的渴望，而且你手里有1亿美元，你的老同事也回来了，那就试试呗。

I mean, my guess is if you're Travis, you're gonna wanna scratch the itch of autonomy and find you've got a 100,000,000, you've got your old colleague back, and have a have a go.

但我觉得，从评论来看，你看看Waymo花了多长时间，对吧，我其实觉得这恰恰说明他们当初没在Uber花十年时间搞这个是对的，因为我觉得这真的是一个非常漫长、非常烧钱的过程。

But I think commentary, I I I think that the fact when you look how long it's taken Waymo, right, I actually think it does speak to the argument that they were right not to try and fund this thing at Uber for the last decade too, because I just think it's it a very long, very capital intensive process.

也许最近三四年，他们应该开始做，而且Uber投点钱进去可能是明智的。

Now maybe the last three or four years, they should have been doing it, and it's probably smart of Uber to put some money in.

但这是一个长期的过程。

But this is a this is a long haul process.

现在可能接近起飞了，但谁知道呢，我们拭目以待吧。

Now it may be near takeoff, but, you know, we'll see.

我在湾区，我把车卖了。

I got in the Bay Area, I got rid of my car.

所以我只坐Waymo和自动驾驶。

So I only do Waymo and autonomous driving.

我不开车了。

I don't drive.

我已经受够了。

I'm done with it.

如果有什么问题，我就叫一辆Uber Black，但我在湾区已经没有车了，就是没有。

If there is an issue, I take a Uber Black, but, I don't have a car anymore in the Bay I I in the in the Bay Area, I just don't have one.

有些特定场景还是需要的，对吧？

There are some niche use cases, right?

我是说，如果我要搬很多东西，我会买辆皮卡，但我绝对不会再回去了。

I mean, you know, if I moved a lot of crap, I'd have a pickup truck, but I would certainly never go back.

绝对不会回去自己开车。

Never go back to have driving a car.

那太落伍了。

It's would be it's just archaic.

我确实觉得Cybercab挺有意思的。

I do think the Cybercab is interesting.

从风投的角度或者其他角度看，我相信Rory说得对。

I mean, a venture perspective or others, I'm sure Rory's right.

Uber在Travis想搞自动驾驶的时候进入，可能从资金角度看确实太早了。

Uber getting into autonomous back when Travis wanted to do it is probably just too early from a capital perspective.

也许我错了。

Maybe I'm wrong.

也许他本来能筹到更多钱，也许这个项目太令人兴奋了，他能筹到的资金比实际多一个数量级，那样的话他就对了。

Maybe he could have raised maybe it's so exciting he could have raised an order of magnitude more capital than he did, in which case he would have been right.

不过话说回来，他仍然是顶尖的筹款人之一。

Now that's, I mean, he's still one of the great fundraisers.

所以也许我和罗里都错了，因为他可能真的能搞定，但当时太早了，时间跨度对任何投资来说都太难熬了，对吧？

So maybe Rory and I are wrong because he could have pulled it off, but it was so early that the time horizon is difficult for any type of investment, right?

除非你是研究实验室，对吧？

Unless you're a research lab, right?

那得花十多年时间，但用25美元而不是100美元做一辆Cybercab，这确实很有颠覆性。

It would have had to be more than a decade, But doing a Cybercab for 25 instead of $100 is pretty disruptive.

比如你不需要给Waymo或Uber小费。

Like you don't have to tip the Waymo or the Uber.

Cybercab在这方面就很有优势。

Cybercab makes fun of it.

它说你可以给小费，然后它就笑了。

It says you can make a tip and then it laughs.

我们不收小费，对吧，就是为了嘲笑这个想法。

We don't take tips, right, to make fun of this idea.

它更便宜。

It's cheaper.

他们不总是把温度调低或调高。

They don't always put it on low or high.

他们没有奇怪的音乐。

They don't have weird music.

不用应付那些，拥有一辆车很烦人。

Don't have to deal with the Owning a car sucks.

拥有一辆车和拥有一栋房子都很烦人。

Owning a car and owning a house suck.

我们觉得这些很棒，但拥有它们真的很糟糕。

Like, we think these are so great, but it's terrible ownership.

所以我觉得，可能再过十年，任何有脑子的人都会把这当作主要的拥有方式，除非你不住在城里，或者有特殊用途。

So I think, you know, it's probably another ten years where anyone with a brain is is gonna have this be their primary mode of ownership if you're not in the country or you don't have niche use cases.

我绝对不会再回去了。

I I'll just never go back.

我做什么事，就直接打个车走。

Anything I do, I'm just gonna I just take a wave out.

不过我觉得你说得对，这确实显示了优步一个明智的战略决策，先撤回来，等看起来更成熟了再跳进去。

I think to your point, though, it does show a good strategic decision from Uber to actually pull back and then jump back in when it looks like it's much more mature.

我们在伦敦其实有Wave，罗伊，而且它正在起飞，他们现在有个合作，正在街上推广。

We actually have Wave in London Roy, which actually is taking off, and they've got a partnership now where they're actually rolling them out on the streets.

不过现在还是有人辅助的，不是完全自动驾驶。

Now they are human assisted, so it's not fully autonomous.

他们还在早期数据收集阶段。

They're still kind of in the data collection early.

但他们现在的位置是能利用他们的分发网络，在更接近实际采用的时候进行后期投资。

But they're in a position where they're leveraging their distribution and able to invest later stage where it's closer to actual adoption.

我觉得这是明智的做法。

I think it's a smart thing.

我只想说一点，Uber在把特拉维斯·卡兰尼克赶出去之后，又往他的公司投了一个亿，这确实有点让人暖心。

The only one thing I would say is, it it is slightly heartwarming that Uber put a 100,000,000 into Kalanick's company, Travis Kalanick, after after pushing them out.

对吧？

Right?

这里面确实有暖心的成分。

There's a heartwarming element to that.

但在这个情况下，这笔钱其实不算多。

But it's that's not a lot of money here in this case.

对Uber来说，这笔钱不算什么，对吧？

It's not a lot of money for Uber, right?

Uber有庞大的资产负债表，而且基本上就两个业务。

Who has a huge balance sheet and basically two products.

对吧？

Right?

而且跟特拉维斯筹到的钱相比，这也不算多。

And it's not a lot versus what Travis has raised.

对吧？

Right?

所以这确实不错，但我觉得从宏观角度来看，这只是一段关系的开始。

So it is nice, but I think just thinking about it from a, at a high level, it's just a it's just a start of a relationship.

对吧？

Right?

一亿美元就像安德森·霍洛维茨给这笔交易投的种子轮一样。

A 100,000,000 is just it's it's it's just like a seed check from Andreessen into the deal.

它方向上有意义，但真的不算多。

It it it's directionally meaningful, but it's not all that much.

对吧？

Right?

各位，我以为风投里的利益冲突早就解决了。

Guys, I thought conflicts were done in Venture.

我们聊了很多关于智能体的话题。

We've talked about agents a lot.

对于不了解的人来说，风投圈里发生了一场冲突，最终导致一笔交易泡汤了。

For those that don't know, there was a conflict in Venture that has now prevented a deal.

我们之前提到过Instinct，那是一个由Index和Benchmark投资的AI助手。

We spoke about Instinct being the AI assistant that's raised from Index and Benchmark.

其实还有一个叫Town的AI助手，我们刚刚在节目里聊过它，原本Index打算领投那轮融资。

Well, there's another AI assistant called Town, which we just had on the show, and Index were gonna lead that round.

但最后，Instinct说不行。

And then, ultimately, Instinct said, no.

不行。

No.

不可能。

Not possible.

不能两个都投。

Can't do both.

所以Index就退出了Town的融资，因为他们已经投了Instinct。

And so Index pulled out of doing Town's round because they were already in Instinct.

这在我看来挺奇怪的，毕竟竞争性投资这么普遍，尤其是在平台层面。

This seemed strange to me given how prolific competitive investing is, especially at the platform level.

我倒不觉得奇怪。

It didn't to me.

我觉得在早期阶段，投资那些会直接产生冲突的公司，对我来说有点牵强。

I mean, I think that at the early stage doing companies that are going to be directly in conflicts seemed a stretch to me.

所以不，Instinct团队反对这件事，我一点都不觉得奇怪。

So no, it did not seem strange to me that the team at Instinct objected to it at all.

你说得对，那是另一回事。

You're right, separate story.

有一大堆人在做基础模型，我们看看另一个极端。

There's a whole bunch of people that are in bold Let's take the other extreme, bold foundation models.

但再说一次，我们讨论过很多次，早期风险投资业务，你要积极参与董事会，这和现在规模大得多的后期风险投资业务很不一样，后者实际上就是在复制公开市场。

But again, as we've discussed many times, the early stage venture business where you're active involved on the board is just very different than the now much larger later stage venture business where you effectively recreate in the public markets.

在OpenAI和Anthropic各投2000亿美金估值的时候，完全说得通。

It totally makes sense to be in OpenAI and Anthropic at 200,000,000,000 pre each time.

你只能拿到有限的信息权，还是追溯性的信息，而且你也就是在股东名册上。

You get limited information rights, retroactive information only, And you know, you're just on the cap table.

这跟投两家上市公司没什么区别。

And it's no different than being in two public companies.

就跟投资英特尔和AMD一样。

It's no different than investing in Intel and AMD.

这种情况下没有冲突，所以无所谓。

That's where there's no conflict and it doesn't matter.

我觉得没人能——举个例子，Harry。

I don't think anyone let's give an example, Harry.

我觉得没人能同时担任Anthropic和OpenAI的董事。

I don't think anyone could be on the board of Anthropic and also on the board of OpenAI.

早期阶段就不一样了，如果你拿到10%的股份，你很可能会有董事会席位，信息权也会多得多。

And the thing about early stages, if you're getting 10% ownership, you're probably looking at a board seat, you're probably looking at significantly more information rights.

所以光是这一点就有问题。

So that alone would be problematic.

而且除此之外，还有信号效应的问题。

Then on top of that, there's the wall signaling.

如果你只从Index融资，而他们又投资了别家，这本身就会传递出一些信号。

If you just raise money from Index and your instinct, there's a signaling comment about them investing in something else.

我能理解CEO们为什么会强烈反对。

I can see CEOs viscerally objecting to that.

所以他们这么做，我并不意外。

So I'm not surprised they did.

同时，我也不意外Index会这样，他们是个很有风度的团队。

And I'm also not surprised, Index, they're they're a classy group.

他们不会硬顶着说，不行，我们就是不干。

They're not going to dig in and say, no, we're not going to do this.

他们只是刚刚支持了某个人。

They've just backed someone.

他们可能以为一个是B2B，一个是B2C，不会重叠。

They probably thought they were B2B and B2C and not going to overlap.

CEO们说，我对此有强烈的看法。

The CEOs say, I feel strongly here.

然后他们做了明智的事，就是退让了。

And they just did the smart thing, which was back off.

这跟如果是两笔后期投资的情况完全不同。

Very different than if they, you know, these were two late stage investments where it's a different thing.

所以，不，我并不惊讶。

So no, I wasn't surprised.

确实存在冲突。

There is a conflict.

他们妥善处理了。

They dealt with it accordingly.

我觉得这有点像倒抛物线形，或者可能就是个顶针形，创始人会在意的那种。

I think there's some sort of like inverse parabolic shape, or maybe it's just a thimble shape where founders care.

对吧？

Right?

而且在非常非常早期的阶段，我觉得他们并不在意。

And at the very, very early stage, I don't think they care.

钱就是让初创公司起步的东西。

Money is startup just getting going.

他们会主动联系。

They reach out.

我都没法告诉你，有多少做AI和餐饮的人来找我，因为我是董事会成员，而且经常在推特上聊这个。

I can't tell you how many folks in like AI and restaurants reach out to me because I'm on the board of owner and tweet a lot about it.

他们就说，嘿，我在做这个。

And they're like, Hey, I'm doing doing this.

你想见个面吗？

Do you want to meet?

我就说，嗯，可能有点利益冲突。

I'm like, well, know, might be a conflict.

我不在乎。

I don't care.

你知道，我想要的是那种理解早期阶段的人，早期的人根本不在乎。

You know, I want I want the guy that understood the the early, early guys don't care.

说得对。

That's a good point.

而后期的人，他们对利益冲突都无所谓，因为他们觉得自己能从行业知识和人脉中获益。

And the late guys, they're all cool with the conflict because they think they're going to benefit from the domain knowledge and the relationships.

对。

Right.

他们也不在乎。

They don't they don't care.

他们明白这一点。

They get it.

而且其中一家已经做到了九位数的营收。

And and one's at 9 figures in revenue.