20VC：为什么机器人技术将比编程智能体大100倍 | 推理将成为一个价值10万亿美元的市场：无人察觉的真相 | 软件投资者为何因Naveen Rao而陷入困境， unconventional

这就像卡尔达舍夫等级一样。

This is like the Kardashev levels.

我们在打某种星际战争，不管那是什么。

And we're fighting some intergalactic battle, whatever that is.

要做到这一点，你需要智慧。

You need intelligence to do that.

智慧是这些对抗性战斗的燃料和经济。

Intelligence is the fuel, the economy of these adversarial battles.

所以它的需求是无限的。

So it has infinite demand.

而要创造智慧，你需要能量。

And in order to create intelligence, you need energy.

把能量转化为智慧的就是硬件。

The conversion of energy to intelligence is hardware.

我说的硬件，是指整个技术栈。

And when I say hardware, I mean the whole stack.

我指的是芯片、系统、云、网络，所有这些。

I mean the chips, the systems, the cloud, the networks, all of those things.

所以对我来说，如果纯粹从经济角度看，我要卖的是把能量转化为智能的能力，而我做得越好，积累的价值就越多。

So to me, if I look at it just from an economy, I'm going to be selling the capability to convert energy to intelligence, and the better I do that, the more value I accrue.

所以我不认为这个市场短期内会变成大宗商品化。

So I don't see it commoditizing anytime soon, as far as the market goes.

你担心过自己贪多嚼不烂，导致产品质量下降吗？

Do you worry that you bite off more than you can chew and then deprecate product quality?

在从能量到智能的转化过程中，中间能吃掉的东西越多，赚的钱就越多。

In that transition from energy to intelligence, the more that you can eat in between, the more money you're going to make.

但昨天我请了Modal的创始人Eric上节目，我很喜欢他，但那家伙就是什么都想干。

But I had the founder of Modal on the show yesterday, Eric, and I love him, but the dude just wants to do everything.

我是说，什么都想干。

I mean everything.

我当时就想，产品广度到什么程度会损害深度呢？

And I was like, at what point does product breadth damage depth?

你有想过这个问题吗？

Do you think about that?

我确实觉得现在世界因为AI在发生变化。

I do think the world is changing now with AI.

我们可以承担更多的广度。

We can take on more breadth.

我们能探索的领域比以前更广了。

Our surface area that we can explore is greater than it was before.

就是因为开发速度更快了吗？

Just because of speed of dev is much faster?

为什么现在每家公司都像是复合型初创公司？

Why is that now every company is a compound startup, so to speak?

我可以给你一个很具体的例子。

I can give you almost a pretty concrete example.

在我们的第一颗芯片里，我们做了一组耦合振荡器。

In our first chip, we built a set of coupled oscillators.

好的。

Okay.

它们有一个耦合矩阵，还有一堆互相影响的振荡器。

They have a coupling matrix, they have a bunch of oscillators that sort of influence each other.

所以当我们做这个的时候，在以前，也就是AI出现之前，我可能会问一个问题，比如：“好的，如果我们不做全连接，噪声会怎么传播？”

And so when we're doing this, I would ask in the old world, say before AI, I might have a question like, "Okay, well, how does noise propagate if we don't do all-to-all connectivity?"

也就是说，我有一个菊花链式的振荡器，开头有一些噪声特性，我想看看经过四到六跳之后，信息会怎么退化。

So meaning I have a daisy chain of oscillators, and I have some noise characteristic at the beginning, and I want to see how that degrades information over, say, four or six hops.

在过去，我得跟我的研究团队说：“嘿，你能去建模一下这个吗？”

In the old world, I'd have to ask my research team, "Hey, can you go model this?

“加个噪声模型进去。”

Put some sort of noise model in."

我只是想要一个初步的理解，用来指导我自己的想法。

And I just want a zeroth-order understanding here, just to guide my own thoughts.

那可能要花一个星期左右，因为他们会说：“哎呀，我在忙别的事情。”

That might take a week or so because they're like, "Ah, I'm working on this other stuff.

我下周再回来处理。

I'll come back to it next week."

可能得花一周时间写代码、验证，然后给我结果。

Maybe it takes a week to write the code and verify it and give me an output.

所以我要等几周才能得到答案。

So I got a couple of weeks to get an answer.

我现在可以直接问模型这个问题。

I can literally ask this question to a model.

它直接生成代码。

It spits out the code.

我可以检查代码。

I can introspect the code.

我可以和它来回沟通几次。

I can go back and forth with it a few times.

我一小时就能得到答案。

I get an answer in an hour.

所以以前需要两周才能想清楚的事情，现在一个小时就搞定了。

So something that would've taken me two weeks to think through has now taken me an hour.

因为有了AI，我能在单位时间里思考更多的事情。

So I can think through more things per unit time because of AI.

不过，我觉得我们还是有个问题，而且我昨天刚发了一篇帖子，说我们有个AI垃圾内容的问题。

Now, I think we still have a problem, and I actually just posted on this yesterday that we have an AI slop problem.

这些模型产出了太多垃圾，而且很难去验证它们。

We have too much crap coming out of these models, and it's very hard to verify it.

有些是好的，有些不是，我们需要想办法大规模地处理这个问题。

Some of it's good, some of it's not, and we need to figure out how to do that at scale.

但我能更快了。

But I can go faster.

我能覆盖比以前更广的范围。

I can cover more surface area than I could before.

你为什么说我们有太多垃圾内容？

What makes you say we have too much crap?

是不是就像在X上看到那些蠢话，现在我的时间线上全是这些？

Is it just like looking on X and seeing the dumb takes on the thing that is just covering my timeline right now?

没有不尊重美国的意思，但你们的政客脸长得挺有特色的，用英语说就是那样。

No disrespect to the US, but it is your politicians who are rather interestingly structured faces, would be an English way of saying it.

这些我都记着，但他们的AI生成版本有好几代。

That's all in mind, but there's different AI generations of them.

还有外星人，还有从国会出来的。

There's aliens and coming out of Congress.

太疯狂了。

It's nuts.

你说的AI垃圾就是指这个吗？

Is that what you mean by AI slop?

不是。

It's not.

那是娱乐。

That's entertainment.

对，那更像是娱乐。

Yeah, that's more entertainment.

顺便说一句，我得说，英国人把自嘲式幽默玩到了极致。

I will say, by the way, the English have cornered the market on self-deprecating humor.

比我们美国人能做的要好太多了。

It is so much better than we can ever do in the US.

说实话，我们在英国看着这些，都在想，'哇，我们一直在抱怨自己的政客。'

Honestly, by the way, we're all looking at this from the UK being like, "Wow, we've been moaning about our own politicians.

'妈的。'

Fuck."

'这些人。'

"These guys."

是啊，我知道。

Yeah, I know.

不，我们已经把它提升到了新高度。

No, we've taken it to new levels.

我跟你说过我们是创新者，对吧？

I told you we're innovators, all right?

所以这其实是个很好的例子，你看到一个结果，然后你会想，这符合你的直觉吗？

And so actually, this is a perfect example of you see an output, and you say, okay, does it meet your gut check first?

然后你会觉得，可能不太符合。

And you're like, "Maybe not."

所以真正有趣的是，现在年纪大的人看到图片，他们不会有那种直觉判断。

So what's really interesting is older people today see images, and they don't have that gut check.

他们会觉得，哇，这是真的。

They're like, "Wow, that's real."

我爸妈就说过，给我看看那些AI生成的特朗普弹吉他的视频。

I've had my parents say, "Show me these AI-generated videos of Trump playing the guitar."

然后他们就说，哇，他真的会弹吉他。

And they're like, "Wow, he plays the guitar."

我就说，妈，他没在弹吉他。

I'm like, "Mom, he's not playing that guitar."

那是AI生成的东西。

That's an AI-generated thing.

她就说：“哦，好吧。”

She's like, "Oh, okay."

所以对我来说，很明显那是AI生成的。

So for me, it's quite obvious that it's AI-generated.

所以我们在公司里看到的那个验证问题。

So that verification problem we see in companies.

我在我们公司也看到了。

I see it in our company.

我们可以制造各种东西。

We can build all kinds of artifacts.

我们可以写代码、建模型、做各种事情，但你必须闭环，确保每一样都是正确的，才能真正信任结果。

We can build code, we can build models, we can do all kinds of things, but you have to close the loop and make sure each one of those things is correct to actually trust the result.

而这样做实际上开始成为一个大瓶颈，因为你产出了太多东西。

And doing that actually starts to become a big bottleneck because you're producing so many artifacts.

我觉得问题已经从生产转移到了验证上。

I think we've moved the problem from producing to verification.

验证的价值在于，如果你掌握了验证层，你就掌握了自动化层，因为你有可验证性，可以在此基础上继续推进。

And the value of verification is if you own the verification layer, then you own the automation layer because you have verifiability that you can then progress thereon.

是这样吗？

Is that it?

对，我想是的，而且我认为这是用AI开发东西的必要部分。

Yeah, I suppose, and I think it's a necessary part of using AI to develop things.

但如果你掌握了验证层，你就掌握了生产层吗？

But if you own the verification layer, do you own the production layer?

我不太确定。

I'm not entirely sure.

我觉得你还得掌握规格说明。

I think you have to own the spec as well.

要产出有用的成果，比如造一辆车或一座桥，你得能制定规格，搞清楚，我是不是在解决一个有用的问题？

Producing a useful output, like building a car or a bridge or something, you have to be able to spec it, figure out, am I solving a problem that's useful?

然后每一个产出的工件都需要验证是否正确，并且朝着那个规格前进。

Then each one of these artifacts that's produced has to be verified to be correct, and then marching towards that spec.

最终，你必须说：“是的，我确实按照规格实现了。”

And eventually, you have to say, "Yes, I did indeed implement the spec as intended."

所以这是一个端到端的过程，不仅仅是验证。

So it's an end-to-end process, not just verification.

验证只是其中的一部分。

So verification is a piece of it.

我不知道该怎么给这个估值。

I don't know how one would value that.

我觉得我会把它看作整体的一部分，就像达到最终结果一样。

I think I value it as part of the whole, like getting to the end result.

我们请了Atlassian的Mike上节目，他非常棒，他谈到了思维链的重要性，以及理解公司不同职能中由代理主导的决策与由人主导的决策之间的区别。

We had Mike on the show from Atlassian, awesome guy, and he was talking about the importance of chain of thought and understanding agent-led decision-making versus human-led decision-making across different functions in companies.

我有点觉得，我完全理解这一点，这非常有道理。

And part of me was like, I totally get that, and that makes absolute sense.

我也有点觉得，这么想有点过时了，那到底是什么意思呢？

And part of me was like, that feels kind of old to think of it like that, and what does that mean?

我们对智能体决策的信任程度不如对人类决策的信任。

We don't trust agent decisions as much as we trust humans.

这也有点奇怪。

That feels kind of weird, too.

确实不信任。

We don't.

听着，我同意你的看法。

Look, I agree with you.

我觉得在很多情况下，我们还不应该像信任人类那样信任智能体的决策。

I don't think we should trust agent decisions as much as humans in a lot of cases just yet.

就像我们为了培养判断力，要花很多年去训练某些技能。

It's like we train for certain things for many years to build judgment.

比如成为神经外科医生。

Becoming a neurosurgeon.

医学院毕业后还要培训七年。

It's a seven-year training after med school.

这是有原因的。

There's a reason for that.

并不是说学缝合有多难。

It's not that learning how to throw a stitch is that hard.

我在研究生阶段就做过这个。

I've done that in grad school.

当外科医生，问题不在于身体操作或机械性的部分。

It's not the physical, mechanical part of being a surgeon that's the problem.

难的是判断该往哪个方向走。

It's having the judgment on which direction to go.

你得看过足够多的病例，看过结果，看过人们怎么思考这些问题，才能建立起那种判断力。

You have to have seen enough cases, seen the outputs, seen how people think through these problems to then build that judgment up.

我们现在对智能的定义是，我们用一个随机模型。

And the way we're defining intelligence right now is we take a model that's stochastic.

我们可以多次提示它，然后会得到不同类型的输出。

We can prompt it several times, and we'll get different sorts of outputs.

你实际上是在探索可能解决方案的分布。

And what you're doing is you're exploring the distribution of possible solutions.

然后我们需要挑选出哪些解决方案是好的。

And then we need to pick which of those solutions is good.

你会得到很多不同的输出，但大多数都不好。

You'll get a lot of different outputs, and most of them are not good.

你可以说人类也是这样做的。

And you could argue that a human does this, too.

我们会说，好吧，我该怎么做出决定？

We say, okay, how do I make a decision?

嗯，我可能会思考所有的输出，然后有一些判断好坏的经验法则。

Well, I might think through all the outputs, and I have some kind of heuristic of deciding what's good.

我们在内部这样做，也许是通过内心独白或某种内部过程，但模型会吐出一连串的思考，然后我们有很多这样的思考链，我们判断哪些是好的，然后走那条路。

We do this internally, maybe as an internal monologue or some sort of internal process, but a model spits out a chain of thought, and then we have many chains of thought, and we decide what's good, and we take that path.

那个决定性的环节，其实并不是每个任务都搞得很清楚，所以我觉得我们不应该完全信任这些东西。

That deciding element is actually not well understood for every task, and so I don't think we should trust these things.

我觉得你构建的决策链越长，一环扣一环，就越容易出问题。

I think the longer and longer decision upon decision upon decision chain you build, the more off the rails it can go.

而且你可能会得到看起来非常可靠的结果，但实际上完全错误。

And you can get things that look very robust, but they're completely wrong.

所以这就是我们现在的状况。

So that's where we are today.

未来这种情况可能会改变。

That will probably change in the future.

我觉得要达到那个水平，还需要一些突破，但这就是我们现在的状况。

I think we need a few more breakthroughs to get there, but that's where we are today.

我喜欢这样的问题：你今天相信什么，看起来古怪、疯狂、离谱，但五年后会变得非常普遍和公认？

I love a question like, what do you believe that today looks kooky, weird, nuts, that in five years' time will be very commonplace and agreed?

我记得我把信用卡绑定到PayPal时，我父母还骂我，还有记得OkCupid，当时大家觉得你在网上找老公或老婆？

I remember putting my credit card into PayPal and my parents telling me off, and remember OkCupid and being like, you find your husband or wife online?

这想法太离谱了。

What a ridiculous idea.

当时确实很疯狂。

Nuts at the time.

我觉得我们会更信任我们的智能体，而不是我们自己。

I think that we will trust our agents more than we trust ourselves.

我觉得在某个时候确实会这样。

I think at some point that's true.

但你觉得我们现在离那一步还很远吗？

But you think that we are far from that today?

我们现在还没到那一步。

We're not there today.

而且专业知识其实不会过时。

And expertise doesn't really go out of style.

只是会以不同的形式出现。

It's just going to take on different forms.

我觉得未来的样子是，我们开始构建所谓的模型，但实际上我不认为那只是一个模型。

And what I think a future will look like is that we start building, we'll call it models, but I actually don't think it's a model.

我认为那是整个机器。

I think it's the whole machine.

我认为智能有物理层面，而智能层本身也是一体的。

I think of intelligence as having a physical and having the intelligence layer as one.

它们是不可分割的。

They're not separable.

我有一台机器。

I have a machine.

它几乎像一个有机体，我可以教它成为各种任务的专家，而价值就在于培养这种专业能力。

It's almost like an organism that I can teach to be an expert at various tasks, and the value will come from building that expertise up.

就像，你上的是哪所学校？

It's like, which school did you go to?

你是怎么学会当神经外科医生或者厨师之类的？

How did you learn how to be a neurosurgeon or a chef or whatever?

这个出身背景很重要。

That provenance matters.

你跟着这个人学习，学到了他们的方法和思维方式。

You went and studied under this person, and you learned their methodology, their thinking.

这就是价值的来源。

That's where the value came from.

你或者任何人从头到尾都是一台智能机器，但真正让你有价值的是你被教导的方式。

You or any human is an intelligent machine top to bottom, but it was the way you were taught that actually made you valuable.

我觉得我们会在机器上开始看到类似的情况。

And I think we're going to start to see something similar in machines.

我们可以复制那种智能和所有的一切，但真正重要的其实是你被教导的方式。

And we can replicate that intelligence and all of that, but the way you were taught is actually going to be the important thing.

所以，其中一部分就是知道如何得到一个真实、可行、经过验证的最终结果。

And so part of that is actually knowing how to get to an end result that is real, that will work, that is verified.

当你思考那种学习中的价值时，很多价值在于你多年来积累的数据。

When you think about the value in that learning, a lot of that is in the data that one accrues over time in the years that you're doing something.

我们现在看到的数据公司，比如我在的Macaw，营收已经达到了天文数字，以前这个数字大得惊人。

We see data companies today, I'm in Macaw, but we see tens hit astronomical levels of revenue that previously were astonishingly large.

现在依然大得惊人，但时间却短了很多。

They're still astonishingly large, but in such a short time.

Naveen，老兄，你真是个聪明人。

Naveen, dude, you're such a smart dude.

跟你聊天真是太开心了。

It's such a joy talking to you.

我特别想问你很多关于这个世界的问题。

I just want to ask you lots of questions about the world.

这个市场会怎么发展？

What happens to this market?

是会随着时间慢慢整合吗？

Does it just consolidate over time?

还是会越来越分散，然后你建立更强的护城河？

Does it unbundle more and more and you build more defensibility?

那会怎么样？

What happens?

说实话，我觉得会专业化。而且我认为，如果你专注于解决重要问题，你就能拥有一个可防御的大市场。对。

I think it specializes, to be honest, and I think if you specialize and actually start to work on important problems, you have a big market that's defensible Yeah.

再说，技能也遵循幂律分布。

Again, it's a power law for skills.

你可以说，所有人类都有能力学习很多技能，但某些技能比其他技能更有价值。

You could argue that all humans are capable of learning many skills, but certain skills are more valuable than others.

对于你构建的数据和课程来说，也是一样的道理，对吧？

And the same thing would be true with data and curricula that you build, right?

所以我们看到数据公司正在专业化。

And so we see the specialization of data companies.

所以你不认为它们是商品化的？

So you don't think they're commodities?

对，我不认为它们是商品化的。

Yeah, I don't think it's commodity.

我觉得一开始它确实像商品，因为那时候就是大规模收集观察数据，对吧？

I think it started as a commodity because it was sort of like at volume gathering of observations, right?

但当你开始构建更高质量的数据集时，就像我们在AI领域说的，比如针对编程的高质量数据，或者针对思维链的高质量数据，这些数据其实就变得非常有价值了。

But as you start to construct higher quality data sets, that's what we call it in the AI world, like higher quality for coding, higher quality for chain of thought or whatever, you actually start to have a lot of value in the data.

从某种意义上说，模型几乎只是它所学到的东西的体现。

The model is almost just a manifestation of what it learned, in a sense.

所以我觉得数据层在未来会变得非常重要。

So the data layer, I think, is going to be pretty important going forward.

我不觉得这个会变成商品化。

I don't see that commoditizing.

我觉得现在真正的关键就是构建模型，并且知道怎么选择那些数据。

I think building models and knowing how to choose that data is actually the game right now.

那些大型模型公司，其实很大程度上是数据问题。

The big model companies, it's largely a data problem.

他们在模型本身也在进步，但大部分能力并不是来自那里。

They are making advances on the model itself, but that's not been where most of the capabilities come from.

这其实一直都是在整理数据，用不同的损失函数。

It's been from curating data, using different sorts of loss functions.

比如我们之前有预训练，那是一种很特别的损失，然后现在我们进入了所谓的中间训练，就是强化学习、监督微调，各种不同的组合。

Like we had pre-training before, which was a very particular kind of loss, then we moved to what we call mid-training now, which is reinforcement learning, supervised fine-tuning, all of these different sort of mixes.

所以这个课程安排，从预训练到各种不同的训练来调整模型行为，这才是真正的秘诀。

So that curriculum, how you went from pre-training to all these different sorts of trainings to modify the behavior of the model, is actually the secret sauce.

我能问一下吗？我们注意到你提到过预训练、强化学习，然后现在明显重心已经转向了后训练。

Can I ask you, we've seen you mention that kind of pre-training, RL, and then obviously a real shift has moved to post-training.

随着重心从预训练转向后训练，你觉得这种转变是持久的吗？在模型改进的效率和我们达成目标的方式上，这是永久性的变化吗？

With that shift of emphasis from pre-training to post-training, do you think that is an enduring, permanent shift in the efficiency of model improvements and how we get there?

还是说我们会回到以强化学习为主，或者以预训练为主？

Or will we move back to RL being the primary or pre being the primary?

从长期来看，这些的重要性会怎么变化？

How does that look in terms of importance over time?

嗯，我不确定这是不是更高效。

Well, I don't know that it's more efficient.

其实我觉得强化学习效率更低。

Actually, I think RL is more inefficient.

如果按单位算力或单位成本来看模型提升的幅度，它其实更不划算。

If you look at it in terms of how much better the model gets per unit compute or per unit dollar, it's actually more inefficient.

但它确实在持续进步，对吧？

But it does keep moving forward, right?

我觉得我们的想法是让模型在重要任务上表现更好。

I think the idea is that we want to make the models better at the tasks that are important.

所以我认为在未来几年里，它确实非常重要。

And so I think it's probably for the next several years, yes, very important.

我想我们会开始在这些中间层，甚至预训练层中引入专门化。

I think we'll start to build specialization in those middle layers, maybe even in the pre-training layer.

所以我觉得模型公司不会在性能上考虑得太长远，比如“哦，我要在AI系统的思考方式上取得重大突破，从而获得优势。”

So I don't think the model companies are thinking very long term in terms of performance, like, "Oh, I'm going to make some massive breakthrough on the way I think about an AI system that's going to give me an advantage."

我觉得大实验室在这方面做得不多。

I would say the big labs are not doing that as much.

你觉得在当前这种日复一日的竞争中，他们能承受得起吗？

Do you think they can afford to in the day-by-day race that is this current ecosystem?

这就是问题所在。

That's the problem.

节奏太快了，你必须时刻保持领先。

It's really fast, and you've got to be on top of it.

如果你在模型发布上晚了一个月，对OpenAI、Anthropic、Grok或谷歌来说，那就是灾难。

If you're one month late on a model, that's disaster if you're OpenAI or Anthropic or Grok or Google.

他们现在还不敢冒那么大的风险。

They can't afford to take those big risks just yet.

你看到了，比如马斯克创立xAI时，他基本上就是，"好吧，这就是范式。"

You're seeing this, like when Elon spun up xAI, he basically took, "Okay, here's the paradigm.

然后就去招人，砰的一下，他就很快达到了前沿水平，因为他有资源。

Let's go hire the people" and boom, he got to the frontier pretty fast because he had the resources.

所以这让我觉得，其实并没有太多深层次的创新。

So what that points out to me was that there wasn't a ton of very deep innovation.

就是把现成的零件拼起来，找一支懂行的团队去组装，然后快速执行。

It was taking the pieces, putting them together with a team that knows how to build them, and executing very fast.

我觉得这和尝试开创一个全新范式是完全不同的。

I think that's very different than trying to come up with a new paradigm.

再看看杨立昆的公司和李飞飞的公司，他们是在做更深层的创新。

And you look at Yann LeCun's company and Fei-Fei Li's company, they were trying to do deeper innovations.

我觉得那可能更适合初创公司来做，而且看起来有点像药物研发，对吧？

And I think that might be in the realm of startups, and it starts to look like drug development, right?

所有的大风险都由初创公司承担，而所有渐进式的改进都由大公司完成。

All the big risks are taken by startups, and all the incremental stuff is done by big companies.

大公司会在初创公司展现出潜力的时候把它们收购。

The big companies buy the startups when they show promise.

我有时候会问一些简单的问题，但我不知道答案，而你非常聪明。

I sometimes ask simple questions, but I don't know the answers, and you're very smart.

为什么现在大家对强化学习这么兴奋？

Why does everyone get so excited by RL today?

你提到了，就是那种中间学习的过程，然后我身边好多人都在说，"啊，关键就是强化学习。"

You mentioned it, kind of the mid-learning there, and I just have so many people, "Ah, it's all about RL."

我就说，"哦，我同意。

I'm like, "Oh, I agree.

我同意。"

I agree."

有几个原因。

There's several reasons.

你想想我们是怎么学习的，主要就是强化学习。

If you think about the way we learn, it's primarily reinforcement learning.

如果我要教你一项新技能，比如学发网球，那其实就是一个强化学习的循环。

If I were to teach you a new skill, learn how to serve a tennis ball or something, it's really an RL loop.

我会给你一个示范，告诉你应该是什么样。

I would give you a paradigm of what it should look like.

你的发球应该这样。

Here's how your serve should look.

也许我会演示一下，或者给你看个视频之类的，这样你就能在心里建立一个对‘好’或‘优秀’的标准模型，然后你不断拿这个标准来对照自己，说：‘对，我快做到了，我在进步，或者我还没做到，我需要寻求帮助。’

Maybe demonstrate that or show you a video or something, and it will allow you to build an internal model of what good is or great is, and then you keep judging yourself against that internal model and say, "Yes, I'm getting there, I'm getting better, or I'm not, and I need to ask for help."

所以这就是强化学习，我觉得它重要的原因是我们无法收集足够的数据来解决世界上所有的任务。

So that is RL, and I think the reason it's important is we cannot gather enough data to solve every task in the world.

我们根本做不到。

We simply can't.

你不能光靠收集数据就指望预训练能解决一切问题。

You can't pre-train your way out of this by just gathering data.

那太难了。

It's just too hard.

而且说实话，模型本身也不够高效，没法充分利用这些数据。

And the models, frankly, aren't efficient enough to make use of it.

所以你需要一个范式，让你能开始说：‘好，这些是最重要的10个任务。’

So you need to have a paradigm where you can start to say, "Okay, these are the 10 most important tasks."

然后我有某种我们称之为可验证奖励的东西。

And then I have some sort of what we call a verifiable reward.

我有了那个内部模型。

I have something that's that internal model.

比如一个好的发球，就是我用力打过去，而且球进了。

Like a good serve is, I hit it really hard and it went in.

这是可以验证的。

It's verifiable.

我可以测量速度。

I can measure the speed.

我可以判断球是否出界。

I can measure whether it was in or out of bounds.

当我有这种可验证的奖励时，我就可以开始说："好，我采取的行动，那些行动是否带来了好的结果？"

When I have that verifiable reward, then I can start to say, "Okay, did the actions I took, did those actions lead to that good outcome or not?"

所以现在我有了一个结构，让我能够解决任何一组通用任务。

So now I have a structure that allows me to solve- Any set of general tasks.

这就是强化学习如此令人兴奋的原因。

And that's why RL is so exciting.

你认同杨立昆的后Transformer世界理论吗？这会怎么改变一切？

Do you buy Yann LeCun's post-transformer world theory, and how does that change everything?

我确实认为我们会进入一个后Transformer的世界。

I do think we will be in a post-transformer world.

我不觉得Transformer就是终极答案。

I don't think the transformer is the be-all end-all.

我可能直到几年前才真正明白这一点，因为Transformer把很多东西混在一起了。

I don't think I truly understood this until maybe a couple of years ago, because transformers conflate a lot of things.

但我们真正关心的是推理能力，就是能一步步推理解决问题。

But what we care about is reasoning, being able to reason through a problem.

我不知道是谁说的，但智能就是你在没有信息时做的事情，而不是在有信息时做的事情。

I don't know who said it, but it's like intelligence is what you do when you don't have the information, not what you do when you do have the information.

所以，能基于某种机制或基本原理的理解来推理问题，这才是智能。

So being able to reason through a problem from some sort of mechanistic or first principles understanding is what intelligence is.

所以Transformer某种程度上具备这一点，但实际上是靠这种代理式的流程，通过多次采样然后判断哪个好来实现的。

So transformers kind of sort of have that, but they really do it through this agentic kind of flow where you're taking multiple samples and you're deciding what's good.

所以我拿一条轨迹和另一条轨迹比较，然后说：‘好，这条轨迹更好。’

So I take one trace and then compare it to another trace, and I say, "Okay, this trace is better."

所以我就把那条叫做我的推理轨迹。

So I'm going to call that my reasoning trace.

而我们实现这种推理的方式，就是通过一种检索机制。

And so the way we implement this reasoning thing is through a retrieval mechanism.

所以注意力机制基本上就是有这个上下文，这其实是一种检索，然后我们可以基于它进行采样，最终得到看起来像推理的东西。

So attention basically has this context, which is a retrieval thing, and then we can sample off of that and eventually get to something that looks like reasoning.

我觉得用这种方式做推理效率太低了。

I think this is extremely inefficient for how we do reasoning.

我们不需要用这种方式来构建推理。

We don't need to build reasoning this way.

我们的大脑不是这样运作的。

Our brains do not work in this way.

我们不会构建一个巨大的轨迹，把它输入到另一个模型里，然后再产生一条轨迹。

We don't build a huge trace, pump it into the input of another model, and then produce another trace.

我们的大脑不是这样运作的。

That's not how our brains work.

所以从算法的角度来看，有更高效、更高性能的方法来实现这一点。

So there are way more efficient ways to do this and way more high-performance, from an algorithmic standpoint, ways to do this.

因此，我们将超越Transformer。

And so we will move past the transformer.

你觉得成为一位伟大的创始人所需的条件有没有改变？

Do you think what it takes to be a great founder has changed or not?

我觉得没有变。

I don't think it's changed.

我认为基本原则还是一样的。

I think the basic principles are still the same.

我觉得对未来有看法，你的愿景现在可能比以前更重要。

I think having a viewpoint on the future, your vision actually may matter more now than it did before.

我认为在人工智能出现之前，你只要执行力强就能成为一个相当不错的创始人。

I think in the pre-AI world, you could be a pretty good founder by just being able to execute.

你不需要有很好的远见。

You don't have to have very good vision.

你只需要知道怎么执行。

You just have to know how to execute.

如果你能比大公司执行得更快，其实这并不难，因为大公司又慢又笨重，我觉得你确实能创立一家初创公司，然后卖掉它或者把它发展成一家大公司。

And if you just go out-execute a big company, which isn't that hard because big companies are slow and lumbering, I think you could actually build a startup, and you could sell it or build a company out of it.

我觉得现在，如果只靠执行，你不会走得太远。

I think these days, if you're just based on execution, I don't think you're going to get very far.

我们之前谈到了不同的层级，你说过：“我不认为数据会被商品化。

We spoke about the different layers, and you said, "I don't think this is going to get commoditized in data.

我认为专业化，那里有价值。”

I think specialization, I think there is value there."

哪个层级会随着时间被商品化？

Which layer does get commoditized over time?

对，这是个好问题。

Yeah, it's a good question.

可以说，能源已经是商品了。

Arguably, energy's already a commodity.

所以这已经商品化了。

So this is already commoditized.

我认为一定程度的智能确实开始变得商品化。

I think a certain amount of intelligence does actually start to become commoditized.

说实话，我们已经看到了这一点。

We're already seeing it, to be honest.

六个月前的模型，比如中国的开源模型，那个水平的，已经相当有用了。

Models that are six months old, let's say the Chinese open models, that level, they're pretty useful.

六个月前的东西仍然相当不错。

Something that's six months old is still pretty darn good.

所以我认为这就是商品化的标志。

So I would argue that's a hallmark of commoditization.

前沿并不是唯一重要的东西。

The frontier is not the only thing that matters.

对于某些问题确实如此，但六个月后，那个前沿就会出现在开源模型里。

For certain problems it is, but in six months, that frontier will be in the open models.

那帮我理理思路。

So then help me out.

如果交付机制或所有基础设施的输出都被商品化了，会发生什么？

What happens if the delivery mechanism or the output of all the infrastructure is then commoditized?

我觉得商品化不一定就是坏事。

Well, I think commoditization doesn't necessarily mean it's bad.

它只是意味着利润率会大幅下降。

It just means that margin, yeah, falls quite a lot.

但同时也意味着市场非常大。

But it also means that the market is very big.

你可以说西红柿是商品化的，但我们还是吃它们。

So you could argue that tomatoes are commoditized, but we still eat them.

我们还是需要它们。

We still need them .

所以我们会有一个巨大的智能市场，而且这个市场会继续增长，但利润率可能不会很高。

So we will have a huge market for intelligence, and that market will continue to grow, but the margins may not be huge.

我其实不认为模型公司会占据整个产业链的利润大头。

I actually don't think the model companies will be the ones who own the margin of the whole stack.

现在就已经不是这样了。

It's not even true today.

现在产业链的利润大头在英伟达手里。

NVIDIA owns the margin of the stack today.

我觉得他们可能拿走了利润的最大份额。

I would say they're probably the one who take the lion's share of the margin.

很多人跟我说，未来几年推理成本会大幅下降，推理市场可能会被商品化。

A lot of people say to me the cost for inference will actually go massively down over the next few years, and we'll see the commoditization, potentially, of the inference market.

你同意这个观点吗？

Would you agree with that or not agree with that?

好，我们来分析一下。

Well, let's break it down.

推理市场的商品化意味着什么呢？

What does commoditization of the inference market mean?

我想，这意味着运营推理服务的那些公司。

What that means is that the operator who's running inference, I assume.

对，完全正确。

Yeah, absolutely.

比如Fireworks、Modal、Base10这些公司以及他们提供的服务。

Your Fireworks, your Modal, your Base10s, and the services that they provide.

因为如果你看他们的价值主张，他们做得非常好的一点是，拿一个现有的模型，比如开源模型，然后让它运行稳定，保证可用性，并且在基础设施上运行得相对成本高效。

Because if you look at their value prop, what they've done, and they've done very well, is take a model that exists, an open model, say, and then make it work well with uptime guarantees and make it relatively cost-efficient to run on infrastructure.

所以他们本质上是在做基础设施。

So they're an infra play.

他们基本上就是把模型映射到物理硬件上，并且做得很好。

They're basically mapping the model onto physical hardware and doing that well.

我认为未来这里面还是有价值的，但是推理的价格，也就是人们愿意为每个token支付的费用，会下降。

Now, I think there will be value there in the future, but the price of inference, like what people are willing to pay per token, will drop.

所以他们得从硬件能力中挤出更多利润来交付那个token。

And so they have to squeeze more margin out of the hardware's capabilities to the delivery of that token.

我觉得这一层的利润空间会越来越小。

And the amount of juice in that layer, I think, will decrease.

所以我认为他们的利润率会下降。

So I think it'll result in them having lower margins.

而且我觉得模型本身已经商品化了。

And I think the model itself is commoditized.

现在有好几种方式可以获得它。

You have several different ways to get it now.

但硬件那边可能不会崩。

But the hardware side of it probably won't collapse.

我觉得那部分还是会有优势，因为你还能从里面榨出更多价值，比如，我能更高效地把能量转化为智能。

I think that part of it is still going to have an advantage because you can still squeeze more juice out of like, hey, I convert energy to intelligence more efficiently.

在我看来，那里还有很大的榨取空间，而基础设施层就没那么多了。

We have a lot more juice to squeeze there versus the infrastructure layer, in my view.

我其实觉得未来十年，硬件才是最有前景的领域。

I actually think the hardware is the place to be over the next 10 years.

兄弟，做硬件投资真的挺难的。

Dude, it's quite fucking hard doing hardware investing.

这完全是另一种思维方式。

It's a very different mindset.

这一代投资人具备做硬件投资的能力吗？

Is this generation of investors equipped to be hardware investors?

不具备。说实话，如果你纯粹做软件投资，那基本就没什么戏了。

No. I would say if you're a pure software investor, you're kind of fucked, honestly.

我可以很肯定地说，因为我觉得SaaS已经不再是能获得高利润的地方了。

And I'll say that pretty confidently because SaaS, I don't think is going to be where you gain margin anymore.

而且我们这种兴奋状态从什么时候开始的？

And we had this sugar high since what?

大概从2010年、2011年到现在，一直觉得可以很快在AWS云服务上搭建东西，快速测试市场，尝试各种想法，然后去追逐一个一亿或两亿美元的市场。

2011, 2010 timeframe till now of, I can very quickly spin something up on an AWS cloud service, test the market quickly, try things out, and then go after a $100 million or $200 million market.

SaaS导致了市场的极度碎片化，因为交付方式变得非常容易和无障碍。

SaaS was massive fragmentation of the market because the delivery mechanism became very easy and frictionless.

现在我们处于应用开始更注重智能层的阶段，我真的看不到20年后软件还能成为巨大的市场。

Now we're in this place where applications start to become more about the intelligence layer, and I just don't see software being a huge market in 20 years.

你不觉得SaaS会变成Lawora吗？

Do you not just think SaaS becomes Lawora?

我的意思是，Lawora算SaaS吗？

And what I mean by that is, is Lawora SaaS?

嗯，算是吧，企业级SaaS，但又不完全是。所以语义不就是在变吗？

Well, kind of, yeah, enterprise SaaS, but kind of not, no. So does the semantics not just change?

我觉得没人会真的从零开始，然后说‘哦，我们是一家SaaS公司’。

I don't think anyone really is starting fresh, being like, "Oh, we're a SaaS company."

完全同意。

So exactly.

我觉得你说得对。

I think you're right.

所以我当时用的是SaaS的经典定义，我想。

So I was taking the classic definition of SaaS, I suppose.

但我认为，那些通过软件和服务结合来交付结果的公司，绝对是。

But I think companies where you deliver an outcome, and you do so through a combination of software and services, absolutely.

我们在好几家公司都看到了这种情况，而且我通常投资种子期的初创公司，挺活跃的。

We're seeing that in several companies, and I invest pretty actively in startups, generally at the seed stage.

我看到一批公司冒出来，它们接手那些长期存在、传统方式完成且必须做对的工程任务，通过AI加人工，让这些任务变得更快、更便宜。

And I'm seeing a crop of companies come where they're taking engineering tasks that have existed for a long time, that were done in a traditional way, and have to be done correctly, and they're making those tasks much faster, much cheaper, by using AI plus humans.

这很棒。

It's great.

我觉得Harvey就是律师事务所的第一个例子。传统方式是，我有一群律师去审合同、写合同之类的，他们基本上说：“我可以利用AI，用人工来验证输出，用更少的人交付同样的东西，而且更快，质量相同或接近相同，我要颠覆这个市场。”

I would say Harvey was a first example of this for law firms, where the old way was, I have a set of lawyers who go and look at contracts and write contracts, that kind of thing, and they basically said, "Well, I can use AI and verify the outputs with humans and deliver the same thing with fewer humans and deliver it faster at the same quality or near the same quality, and I will disrupt that market."

所以我觉得我们会看到这种情况。

So I think we're going to see it.

这算是新的SaaS。

This is sort of the new SaaS.

我能问一下吗？

Can I ask?

我们看到这种情况，我称之为“词汇游戏”，就是风投们编造一套世界理论，比如那个“万亿美元服务”的说法。

We see it, I call it word wank, which is where VCs come up with a coined theory of the world, and it's like the trillion-dollar services thing.

我觉得红杉资本做了一篇很棒的文章，我很喜欢写这篇文章的朱利安。

I think Sequoia did a brilliant piece on, and I love Julian, who did it.

写得非常好。

It's very good.

但我们是不是只是把它稍微美化了一下？

But are we not just slightly jazzing it up?

那其实就是未来，人们用AI高效做事而已。

That's just the future, people using AI to do things efficiently.

这种过度美化，有点过头了，不是吗？

This like whole glamorization of it, bit overkill, isn't it?

你那样做，不就是投资一家律师事务所吗？

Aren't you just investing in a law firm when you do that?

确实是这样，但它是律师事务所的升级版。

Effectively, yeah, but it's a better version of a law firm.

就这么简单。

That's all.

但你这么做没问题，因为我遇到很多公司正是你说的那样，我就想，'我可不想开律师事务所。'

But you'd be okay to do that, because I meet a lot of companies doing exactly what you say, and I'm like, "I don't want to do a law firm."

我很乐意做像Lawora或Harvey这样的公司，它们卖给律师事务所，但自己并不是律师事务所。

I'm happy doing a Lawora or a Harvey, which sell to law firms, but they're not actually a law firm.

嗯，但我们确实看到了这种情况。

Well, but we are seeing that.

一些大型律师事务所现在直接投资AI。

Some of the large law firms are now investing in AI directly.

所以我觉得在某种程度上，你会看到Harvey基本上卖给律师事务所，因为律师事务所有渠道。

So I think in its limits, you'll see basically Harvey sells to the law firms because the law firms have the channel.

这个渠道能长久吗？

Is that channel going to be enduring?

我不知道。

I don't know.

但我也能想象他们会说，"你知道吗？

But I could also see them saying, "Well, you know what?

我们可以直接卖给公司。"

We could just sell directly to companies."

就我个人而言，我会买。

Personally, I would buy that.

如果Harvey说，"嘿，我有个解决方案，能处理你所有的保密协议。

If Harvey said, "Hey, I have a solution that does all your NDAs.

它能审查你的合同。"

It reviews your contracts."

像所有这些事情，而且我能确信它有效，我会把80%的工作量，我的法律工作量，交给他们。

Like all these things, and I can gain confidence that it works, I would send 80% of my workloads, my legal workloads to them.

不过我很好奇，因为你提到有些律所在自己开发。

I am intrigued, though, because you spoke about some firms that are building their own.

我觉得你看到了法律界的K&E投入了大约5亿美元在这上面。

I think you saw K&E in the legal space commit like $500 million to it.

是的。

Yep.

Lynn在Fireworks，我是那里的投资人，他非常公开地谈到了拥有自己的智能，以及公司利用自己的数据来构建自己的模型的重要性。

Lynn at Fireworks, where I'm an investor, spoken very openly about owning your own intelligence and the importance of companies leveraging their own data to build their own models.

你觉得这是一个非常可行的未来，还是从一个非常特定的角度出发的稍微有点技术乐观主义的预测？

Do you think that is a very viable future or actually a slightly techno-optimist prediction from a very specific standpoint?

我觉得这有点技术乐观主义，因为我认为你必须有一定的规模才能构建并拥有自己的智能。

I think it's a bit techno-optimist because I think you have to have a certain amount of scale in order to build your own intelligence and own it.

就拿律师事务所来说吧。

Keep going with law firms, for instance.

如果你是Cooley或者那些大所之一，是的，你可以投资这些，并且能够实现。

If you're Cooley or one of these big firms, yeah, you can invest in these things and you can make it happen.

但如果你在内布拉斯加中部的一个夫妻店，你就做不到。

If you're a mom-and-pop shop in the middle of Nebraska, you can't do that.

所以到时候你会用某种服务。

So what you're going to use then is a service of some sort.

你可以用那个服务来增强你的律师团队，这样你就能为客户提供服务了。

You may augment your attorneys with that service, and that's how you'll provide service to your clients.

所以我不认为每家公司都会用AI来利用自己的历史数据去优化自己的模型。

So I don't see this as every company will use AI in a way of leveraging their past data to make their own models better.

他们不会这么做的。

They're not going to do that.

这对他们来说太难了。

It's too hard for them.

他们想要的是开箱即用的解决方案。

They want an out-of-the-box solution.

所以我觉得卖这种开箱即用的解决方案是有市场的。

So I think there is a market for selling those out-of-the-box solutions.

但这会是一个很长的尾巴。

But it's going to be this long tail.

大公司会开始定制化，因为他们有规模来支撑这个投入。

The big players will start to customize because they have the scale to justify it.

这很合理。

It makes sense.

这又回到了我们世界处处充满对抗性的观点。

And it goes back to this idea that every part of our world's adversarial.

这取决于你在和谁竞争，对吧？

It depends on who you're fighting, right?

如果你在对抗另一家大型律师事务所，而他们用AI来大幅提升利润，那你也不得不这么做。

If you're fighting another big law firm and that big law firm goes and uses AI to supercharge their margins, well, you kind of have to do the same.

小作坊式的律所可能不会这么做，而他们的对手也不会这么做。

The mom-and-pop shop probably isn't doing that, and the people they're fighting against are not doing that.

所以他们会在另一个层面待上一段时间。

So they're going to stay in this other realm for some time.

我觉得很多法律编码背后有一个关键点，就是我们会把多少比例的薪酬转化为技术投入。

I think one thing that underpins a lot of law coding is the percent of salary that we will then transition to technology spend.

我们看看现在的律师行业，美国法律市场大概有三千亿美元，你看着这个数字会想，这些服务提供商会吃掉多少份额呢？

When we look at lawyers today, I think it's about a $300 billion legal market in the US, and you're looking at that going, well, how much of that will be eaten by these providers?

我觉得编程领域也会面临同样的问题。

I think you see that same question there with coding.

你看到贝尼奥夫说过，他们会在六十亿美元的开发者薪资预算中，拿出三亿美元用于代码和编程工具，这已经是很显著的比例了。

You saw Benioff say that they'll spend $300 million of a $6 billion developer salary budget on claw code and coding tools significantly.

我想问你的问题是，这个比例对开发者来说会一直保持在5%吗？

My question to you is, will that stay at 5% for developers?

它会涨到30%或50%吗？

Will that go up to 30%, 50%?

还是会降到1%？

Or will it come down to 1%?

因为这才是最关键的问题。

Because that is the motherload question.

我其实觉得这个比例可能会保持稳定，因为我们最终会遇到验证的问题。

I actually think the percentages will remain maybe sort of stable because we end up with this verification problem.

这很难，因为训练人类去判断什么是好的，其实才是最难的。

It's hard because training humans to know what's good is actually the hard part.

对于开发者来说，关键不在于你当天直接产出了什么。

Developers, it's actually not about what your direct output is on this day.

不是看代码行数。

It's not like lines of code.

真正重要的是代码的质量。

It's actually quality of code that matters.

所以那些经验丰富、顶尖的开发者，那些所谓的“十倍效率”开发者，他们培养出了一种对什么是好、什么看起来好、什么在美学上正确的直觉，而且他们能让事情快速运转起来。

So very experienced, very top-notch developers, those 10X developers, whatever, they developed a sense of what is good, what looks good, what's aesthetically right, and they can make things work fast.

在这个新世界里，那些知道什么是好的程序员，实际上可以验证模型的输出，从而变得更高效。

In-In this new world, those coders, those people who know what's good, can actually verify the output of models and become more efficient.

所以我认为，结果就是我们产出更多，而市场规模——比如说我们为人类加AI支付的费用——不会大幅改变，只是他们的产出会大大不同。

So I think what happens is we just produce more, and the market size, let's say that we're paying for humans plus AI doesn't change drastically, just their output greatly changes.

我是这么看的，因为稍微推演一下。

That's the way I look at it, because play it out a little bit.

所以我在一家公司，每年花一千万美元来开发我的软件。

So I'm spending in a company, let's say $10 million a year to develop my software.

现在突然之间，我的软件客户开始有不一样的需求了。

Now all of a sudden, my customers of that software start having different demands.

十年前，他们觉得一年发布一次就很满意了。

10 years ago, they were happy with one-year release cycles.

现在他们却说：“嘿，我希望我的bug能在一周内修好。”

Now they're saying, "Hey, I want my bug fixed in a week."

所以唯一能做到的办法，就是让AI和人类一起合作。

And so the only way to do that is actually to use AI with the humans.

我并没有做得更少。

I'm not doing less.

实际上我做得更多了。

I'm actually doing way more.

我一周内修好那些bug，最后甚至能在一小时内修好。

I'm fixing those bugs in a week, and eventually fixing them within an hour.

但我还是每年花一千万美元来继续开发整个系统。

But I'm still paying $10 million to keep developing this whole thing.

而且我需要人来验证那些输出，并且真正朝着有用的目标前进，比如修复漏洞、添加功能。

And I need the humans that can verify those outputs and actually move to a goal that's useful, fixing the bug, adding features.

所以我觉得交付速度会加快，因为需求方开始要求更快了。

So I think the pace of delivery goes up because the demand side starts to demand it.

我觉得这里被忽略的关键点是，并不是每个人的工作都会消失，因为同样的任务用AI加人类可以花更少的钱完成。

And I think that is the key thing here that's being left out, is that it's not that everyone's job goes away because the same task could be done for less with AI plus humans.

而是说，确实如此，但现在我会做更多这样的任务，因为需求增加了。

It's that now, yes, that is true, but I will now do way more of those tasks because the demand side has increased.

这两者是相辅相成的。

You don't get one without the other.

当供给变得更便宜时，需求方也会跟着变化。

When the supply gets cheaper, it changes the demand side as well.

所以我觉得接下来会发生的事情就是，我们最终会做更多的工作。

So I think that's what's going to happen, is that we just end up doing more.

我们的发布周期会更频繁，变更也会更多。

We have more release cycles, more changes.

我举个例子，大家应该能理解。

Here's a physical example I think that people can get their heads around.

比如说，现在我们建一栋楼的速度是这样的。

Let's say today we have a rate of building a building.

如果我想建一座摩天大楼，大概需要两年的施工时间。

If I want to build a skyscraper, it takes two years of construction, say.

这取决于一定数量的人，以及他们具备的各种技能。

And that's based upon X number of humans with Y sets of skills.

假设突然之间，机器人能完成其中80%的工作。

So let's say all of a sudden, I have robots that can do 80% of that.

那会发生什么呢？

What's going to happen?

好吧，短期内体力劳动可能会受到冲击，但接下来你会说，你知道吗？

Well, okay, there might be a disruption in physical labor for a moment, but then you'll say, you know what?

突然之间，我建同一座摩天大楼的成本和时间，以前要两年，但现在80%的工作由机器人完成。

All of a sudden, my cost and time to build that same skyscraper, it was two years, but now 80% of those tasks are done by robots.

我可以把整个工期缩短到八个月。

I can bring the whole thing down to eight months.

所以不是说你就建同样数量的摩天大楼，只是用八个月完成。

So it's not that you're just going to build the same number of skyscrapers and just do it in eight months.

你会建两倍的摩天大楼，对吧？

You're going to build twice as many skyscrapers right?

会有更多的事情在运转，更多的建筑被拆除，更多的劳动力进入那个市场。

There's going to be more things churning, more buildings being torn down, more labor going into that market.

所以项目总数会增加。

So the total number of projects will go up.

投入所有这些项目的资金可能相同，但建造它们将是人和机器人的结合。

The amount of dollars going into all of those projects may be the same, but it'll be a combination of humans and robots to build them.

纳文，你说过你显然有天使投资。

Naveen, you said that you obviously angel invest.

你的天使投资非常成功。

You angel invest very successfully.

你有点轻描淡写了。

You kind of downplayed it.

我觉得我现在投资的方向，一个反直觉的想法是，我不喜欢投资机器人领域。

I think where I invest today, a contrarian thought for me is I do not like robotics investments.

我觉得我们离目标还差得很远，而且我确实认为时机很重要。

I think we are so far from where we need to be, and I do think that timing is important.

不过我不是说那些支持机器人的基础设施，我觉得围绕它们的数据集很有意思。

Now, I'm not saying the infrastructure to power robotics, I think the data sets around them are very interesting.

我们看到很多像Makra或Micro AGI这样的公司，实际上在做这方面的基础设施建设。

We see a lot of people like Makra or Micro AGI that are actually in the infrastructure build-out for it.

但真正的应用层面，比如机器人清洁工、机器人建筑工，我们离目标还差得很远。

But the actual application layer where we see robot cleaners, robot builders, we are so far away from where we need to be.

我不想涉足那个领域，也不想承担时机上的风险。

I don't want to be there, and I don't want to take the timing risk.

我说错了吗？

Am I wrong?

我是不是太天真了？

Am I naive?

你比我聪明多了。

You're much smarter than me.

嗯，不过在这个问题上我挺执着的，所以我得强烈反对一下。

Well, again, I'm pretty pilled on this one, so I am going to disagree strongly on it.

好，你说我们离目标还很远，那具体是多久？

Okay, when you say we're so far away, put that in terms of time.

你觉得时间线是怎样的？

What do you think is the timeline?

五年以上。

Five-plus years.

好。

Okay.

我们在这点上分歧不大。

We're not so far apart on this.

我觉得未来两到三年内，我们就能看到一些这类任务的可行原型，而市场会在五到七年内出现。

I think we'll have working prototypes for some of these tasks in the next two to three years, and the market will emerge in five to seven years.

我觉得没错。

I think that's right.

我觉得人形机器人和非人形机器人之间也有很明显的分界线。

I think there's also a clear line of delineation between humanoids and non-humanoids.

非人形机器人显然有更清晰的路径。

Non-humanoids absolutely see that clearer pathway.

至于人形机器人，每当我以为某件事很明显的时候，我总会想，"我肯定大错特错了。"

The humanoids, whenever I think something's really obvious, I'm always like, "I must be very wrong."

因为人形机器人显然还差得远。

Because the humanoids are so obviously so far off.

我觉得我们会生活在一个指数级发展周期的奇怪世界里。

I think we're going to be in this weird world where we live in exponential development cycles.

我们很难再用过去的模式来思考了。

It's really hard to apply our old models.

有些东西今天看起来遥不可及，但按照指数曲线发展，最终会比你想的快得多。

Something that looks so far away today with an exponential curve ends up being much faster than you thought.

就像TAM，市场规模——我觉得是推理市场。

And like TAMs, a market size is-- I think the inference market.

我现在知道什么，是希望24个月前就知道的？

What do I know now that I wish I'd known 24 months ago?

就是推理市场其实没有上限。

That the inference market was just unbounded.

确实没有上限。

It's unbounded.

那是数万亿美元的市场。

It's multi-trillions.

如果你10年前问我，哪些是万亿美元的市场？

If you asked me 10 years ago, what were trillion-dollar markets?

这是基本需求。

It's basic needs.

住房、食物、交通，对吧？

It's housing, food, transportation, right?

你绝不会想到科技领域能出现万亿美元级别的市场。

You would never have thought something in tech would be a trillion-dollar market.

而且我认为，仅AI模型本身在未来几年就会成为万亿美元市场，而推理市场在十年内将达到10万亿美元。

And I think just AI models themselves will be a trillion-dollar market in the next few years, and the inference market is going to be a $10 trillion market within 10 years.

所以这想想都觉得不可思议，对吧？

So this is insane to think about, right?

而且正如你所说，我们正在把一部分劳动力和价值创造环节转移到技术上。

And to your point, we're moving some of that labor market and part of the value generation stack to technology.

所以这就是我们获得这些巨大TAM的原因。

So that's where we're getting these giant TAMs.

所以我看待Unconventional的方式是，我认为它是一个绝佳市场的原因是，如果我们在token交付效率上提高一个数量级，那实际上会在某些方面让市场扩大10倍。

So the way I look at Unconventional, the reason that I think it's a fantastic market to be in is that if we knock off a zero, one order of magnitude of efficiency on delivery of a token, that effectively 10x's the market in some ways.

所以如果某个市场原本是万亿美元级别，现在突然有了更多的应用场景。

So if something was a trillion-dollar market, now all of a sudden you have way more applicability.

潜在用途的基础增长不止十倍。

The base of potential uses just increases more than 10x.

这就是杰文斯悖论。

This is Jevons paradox.

所以我觉得我们现在处于一个可以构建这些指数级增长的位置，这能释放巨大的价值和市场。

So I think we're in this place where we can build these exponentials and it just unlocks massive value and market.

所以我实际上认为机器人技术会比智能体编程大一百倍。

So I actually think robotics is going to be 100 times larger than agentic coding.

哇，这跟我的观点有点冲突啊，Naveen。

Whoa, well, that is slightly disagreeing with me, Naveen.

在我们快速问答之前，能不能先展开讲讲？

Before we do a quick fire, can we just unpack that?

因为我不想显得太蠢。

Because I don't want to be completely moronic.

你为什么这么说？

Why do you say that?

我觉得我们对智能输出的理解还只是皮毛。

I think we're still scratching the surface of how to see the output of intelligence.

所以物理世界是我们智能从未真正涉足的领域。

So the physical world is something we've just not been in at all for intelligence.

我认为到目前为止的机器人，比如工业机器人，并不真正智能。

I would argue robotics to date, like industrial robotics, are not really intelligent.

它们只是机械重复。

They're rote.

我们把人类的智能，比如如何完成任务，灌输给它们，然后它们只是机械地执行。

We're imbuing them with intelligence that we have, like how to do a task, and we're just doing it rote.

那并不是真正的新问题解决能力。

That's not really novel problem-solving.

在不久的将来，我们将进入物理问题解决的领域。

We will be in the realm of physical problem-solving in the not-distant future.