20VC：萨提亚·纳德拉谈开放与封闭模型之争：谁将胜出 | 是否应信任前沿模型提供商 | 为何企业需掌握自身智能 | 垂直化的重要性：从芯片到模型，以及更多

达里奥、伊利亚、萨姆、格雷格还有埃隆，他们那时候差不多就是一家人，关系特别好。

Dario and Ilya and Sam and Greg and Elon were all, I guess, one happy family.

他们真的是规模法则的坚定信徒，而我在2018年、2019年的时候也相信这一点。

They were really the believers in scaling laws, and that's what I believed in back in 2018, 2019.

我觉得现在最大的变化是，我以前总觉得对齐问题会自动解决。

I think the biggest change now I have is I used to always think that alignment will just fall out.

但现在我不这么认为了。

I no longer believe that.

这里是20VC，我是哈里·斯特宾斯。

This is 20VC with me, Harry Stebbings.

说实话，这一期节目对我来说挺感动的。

Now, this show honestly is quite an emotional one for me.

我是在伦敦的卧室里创办了20VC。

I started 20VC from my bedroom in London.

那时候我才18岁。

I was 18 years old.

我那时身无分文。

I had no money.

我手机通讯录里一个联系人都没有。

I had no contacts in my phonebook.

而今天，我有机会与微软CEO萨蒂亚·纳德拉坐在一起，微软是过去三四十年里最伟大的公司之一。

And today, I have the chance to sit down with Satya Nadella, CEO of Microsoft, one of the greatest companies of the last three or four decades.

这个节目花了12年才做到这一步。

It was 12 years in the making with this show.

我们做了3000期节目，才有机会和萨蒂亚坐进那个房间。

We've done 3,000 shows to build up to sitting in that room with Satya.

真的没有一蹴而就的成功。

There really is no overnight success.

努力工作，与人为善，永不言弃。

Work hard, be kind, never give up.

我只想说，关于萨蒂亚，他在如此短的时间里为这次访谈注入的那种温暖，几乎是我前所未见的。

And I will just say on Satya, the warmth that he was able to imbue in this interview in such a short space of time was almost unlike anything I've seen before.

他论证的清晰度和他的想法，以及他快速深入问题的能力。

The clarity of his arguments and his ideas, his ability to go very deep very quickly.

这几乎是我从没见过的那种表现。

It was almost unlike anything I've ever seen before.

这是我会永远记得的一期节目，我真的希望你们喜欢。

It was a show I'll always remember, and I really hope you enjoy it.

但在我们开始今天的节目之前，Base44是我用过的一个产品，我当时立刻就想：“天哪，如果我在用50美元和风投圈没人脉的情况下创办20VC的时候就能用上这个，该多好。”

But before we dive into the show today, Base44 is one of those products I used, and I just thought straight away, "Man, I wish I'd had this when I started 20VC with $50 and no contacts in venture."

我想看看它能做什么，所以我为这个节目建了一个嘉宾研究中心。

I wanted to see what it could do, so I built a guest research hub for the show.

即将到来的嘉宾、公司简介、以往的采访、赞助商联系方式、后续跟进，全部集中在一个很棒的地方。

Upcoming guests, company notes, previous interviews, sponsor contacts, follow-ups, all in one amazing location.

我用通俗易懂的英语描述了需求。

I described it in plain English.

几分钟后，我就看到产品在我面前打开了。

A few minutes later, I had the product open in front of me.

数据库、登录、后端和托管都已经处理好了，全都整合在一起。

Database, login, backend, and hosting already handled, and just all in one.

这让一切变得特别简单，这就是我喜欢它的原因。

It made it so easy, and that's what I loved.

我没有写长长的需求文档。

I didn't write a long spec.

也没有把五个不同的工具拼在一起。

I didn't stitch five different tools together.

我直接做出来，然后就开始用了。

I built the thing, and I started using it.

任何正在创业的朋友，重点就在这里。

For any founder listening, that's the point.

你越快把想法变成现实，就能越早推出产品。

The faster you turn an idea into something real, the faster you launch.

应用、仪表盘、内部系统、AI智能体，不管你要做什么，Base44都能帮你搞定。

Apps, dashboards, internal systems, AI agents, whatever you need to build, Base44 gets you there.

去 base44.com/20VC。

Go to base44.com/20VC.

就是 base44.com/20VC。

That's base44.com/20VC.

Base44 帮你更快地搭建产品，而 Laridin 帮你衡量 AI 的影响。

While Base44 helps you build faster, Laridin helps you measure AI impact.

现在每家公司都同时拥有两类劳动力：人类和 AI 代理。

Every company now runs two workforces, humans and AI agents.

Laridin 是第一个能统一监测这两类劳动力的仪表盘，它能衡量他们的工作成果，并优化他们的影响力。

Laridin is the first unified dashboard across both workforces that measures their work and optimizes their impact.

Laridin 回答四个问题：谁在使用 AI？

Laridin answers four questions: Who's using AI?

用了哪些工具？

Which tools?

成本是多少？

What is it costing?

而且每个团队、每个供应商实际上到底产出了什么？

And what's it actually being produced for every team across every vendor?

AI的投资回报率不再靠猜，而是变成了一个你能持续并自信衡量的数字。

AI ROI stops being a guess and becomes a number that you can consistently and confidently measure.

Gainsight、Vertiv、SurveyMonkey，以及许多其他企业级组织都在Laridin上运行他们的AI项目。

Gainsight, Vertiv, SurveyMonkey, and many other enterprise organizations run their AI programs on Laridin.

数千种AI工具和数百万美元的AI支出都被衡量了。

Thousands of AI tools and millions of dollars of AI spend are measured.

你押注AI，Laridin来证明它正在起作用。

You bet on AI, Laridin proves that it's working.

访问laridin.com了解更多。

Visit laridin.com to learn more.

那是L-A-R-R-I-D-I-N.com。

That's L-A-R-R-I-D-I-N.com.

你已经到达目的地了。

You have now arrived at your destination.

萨蒂亚，非常感谢你能和我们一起做这个节目。

Satya, thank you so much for doing this with us.

我本来想说谢谢你来演播室，但实际上是你在招待我们，所以真的很感谢你抽时间。

I would say thank you for coming to the studio, but you're actually hosting us here, so really appreciate you taking the time.

不，哈里，谢谢你邀请我上节目。

No, thank you so much for having me on the show, Harry.

这真是期待已久，但我很高兴我们终于做到了。

It's been a long time coming, but I'm faintly glad we were able to do it.

你大概在10天前写了一篇重要的文章，你说我们应该假设一个模型已经被攻破了。

Now, you wrote a seminal piece 10 days ago or so, and you said that we should assume a model is compromised.

我想就此展开聊一聊。

And I just wanted to expand on that.

这是否意味着我们不能或不应该依赖模型提供商公司的披露？

Does that mean that we can't or shouldn't rely on the disclosures of the model provider companies?

我觉得现在的挑战，在我看来，是模型变得越来越强大了，对吧？

I think the challenge, the way I see it right now, is the models are becoming more and more capable, right?

扩展法则正在起作用。

Scaling laws are working.

不过我原本的一个假设是，对齐问题会随着扩展自然解决，但现在看来并非如此，那里还有更多的科学挑战需要攻克，对吧？

Except one of the assumptions I had was that you'll have alignment just fall out of scaling, and it doesn't seem to be, and there's more science challenges there that need to be tackled, right?

而机制层面的理解也仍然难以捉摸。

And mechanistic understanding is also elusive.

所以在这种情况下，如果你想要继续使用这些先进的模型，我认为最好的姿态就是把它们放在一个信任架构中来对待。

And so given that, I think the best posture to have if you want to continue to use these sophisticated models is to treat these models in a trust architecture.

毕竟，零信任并不是什么新鲜事，对吧？

I mean, after all, zero trust is not something new, right?

我们早就有了零信任架构的方法。

We've had zero trust architectural approaches.

所以我要强调的观点是，让我们构建一个工程化的零信任环境，能够在任何失控的代理造成伤害之前将其控制住，对吧？

So the claim at least I'm making is let's create an engineered zero trust environment that can contain any malfunctioning agent before it does any harm, right?

这就是我的要点。

So that's the point.

因为如果你真的要让那些能力更强但没那么对齐的智能体运作起来，你就需要一个可信赖的信任系统。

Because if you're really going to have agents that are more capable and yet less aligned, then you need to have a trust system that you can rely on.

我能问一下，那个信任系统具体是什么样子的？

Can I ask, what does that trust system look like?

它的架构是谁来设计的？

Who's the architecture of it?

谁来做它的仲裁者？

Who's the arbiter of it?

这个问题问得好。

It's a great point.

所以这就是为什么我看待这件事时，会把AI安全排除在外，对吧？

And so that's why I look at it and say you leave AI safety outside of this, right?

因为那个对齐问题，至少在我看来，它基本上必须在训练阶段就体现出来，对吧？

Because that alignment, for example, at least my take is that it really mostly has to show up in training, right?

实际上，也许甚至在预训练阶段就开始了。

In fact, maybe even in pre-training.

我觉得你甚至没法通过强化学习来达到对齐。

I don't think you can even RL your way to alignment.

你必须真正让对齐从一开始就深深嵌入到权重里。

You have to truly get alignment to be something that truly gets embedded in the weights right at the very beginning of it.

所以，模型本身，希望有一天能够变得对齐。

And so that's the model itself, hopefully at some point will get aligned.

所以如果你把AI安全放在一边，那么信任架构就只能靠像微软这样的公司，或者所有网络安全人员来做了。

So if you leave AI safety, then the trust architecture is going to be just what a company like Microsoft does or all the cyber folks do.

比如，我们做了两样东西，对吧？

So what we have done, for example, is we built these two things, right?

Agent 365和Perception，对吧？

Agent 365 and Perception, right?

它是干什么的呢？

What does it do?

它给每个智能体一个身份。

It gives every agent an identity.

它会记录所有的活动，包括所有的工具调用，实际上，每一次模型调用和其中的思维链都会记录。

It then logs all of its activity, all its tool calls, in fact, every model call and the COT in it.

实际上，我还要求所有的思维链最好用英语写，而不是用新版本语言。

In fact, that's where I'm also asking that all the COT better be in English and not in new release.

然后你可以用多个模型来检查。

And then you can check using multiple models.

这最好是防篡改的，对吧？

It better be tamper-proof, right?

你不能让智能体自己去修改它的日志。

You can't have the agent go modify its own logs.

然后在此基础上，你设定策略，然后你再行为上监控智能体的活动，确保它们符合策略。

And then on that, you set policy, and then you behaviorally monitor the activity of the agents so they're conformed to the policy.

这就是那个架构。

So that's the architecture.

我们今天还没有这样的架构。

We don't have that architecture today.

这只是在反映我们目前所处的阶段吗？

Is that just symbolic of where we are in the progression?

不，那个架构已经存在了。

No, that architecture exists.

所以有两种——我觉得人们把所谓的工程化安全、零信任架构系统用于隔离，和AI安全混为一谈了。

So there are two-- I think people conflate what I would call an engineered security, zero trust architectural system for containment-

……对吧？

... and AI safety, right?

因为我觉得人们会说：“嗯，AI安全很难，而且还没实现，所以我们在安全方面肯定什么都做不了。”

Because I think people say, "Well, AI safety is hard and elusive and not yet achieved, so therefore, there must be nothing we can do on security."

我并不是说这两件事必须同时实现。

And I'm not saying that these two things all got to come together at the same time.

实际上，我想说的是，嘿，我们必须面对现实，所以让我们在模型外部建立一个信任系统，然后在制定政策和监控政策偏差方面采取非常积极的态度。

In fact, I would just say, hey, we have to deal with the situation as it exists, so let's build a trust system outside, in fact, of the model, and then just be super aggressive in setting policy and monitoring for policy deviation.

在监控方面——

On the monitoring side-

如果考虑思维链，它的可靠性如何？

... if we think about chain of thought, how reliable is that?

它并不可靠，是吧。

It's not, right.

我们知道思维链本身可能不忠实，而且事实上，当模型公司说“它不忠实，我要把它藏起来”时，情况会变得更糟。

We know that the chain of thought itself can be unfaithful, and in fact, it can be made worse when the model company says, "Well, it's unfaithful, and I'm going to hide it."

所以这就变得更麻烦了。

And, and so it becomes even more problematic.

这就是我觉得需要推动标准化和思维链公开的地方，至少对可信伙伴来说，要让它可检查、可审计，我可以用多种模型来验证等等。

And that's where I think a little bit of the push for standardization and the CoT, at least for trusted partners being available as is, so that it's inspectable, it's auditable, and I can check using multiple models and what have you.

这将是其中之一……事实上，这甚至可能推动更多开放权重，老实说，因为如果你不能信任——有趣的是，闭源模型更强大，但如果它们没有意识到需要更透明，尤其是当可信伙伴在使用时，那我们可能就有问题了。

So that's going to be one of the... In fact, that's one of the things that may even drive more open weights, quite frankly, because if you can't trust-- It's funny, closed weight models are more powerful, but if they kind of don't recognize the need to be more transparent, especially as their trusted partners use it, then we may have a problem.

我能问一下，当我们谈到开源崛起时，我完全同意你的看法，闭源确实更强大。

Can I ask you, when we think about that rise of open source, I completely agree with you in terms of closed being more powerful.

开源主要还是以中国模型为主。

Open source stays predominantly Chinese models.

那当我们考虑安全要融入预训练过程，以及像我们刚才说的那些披露机制，再加上开源模型大多数是中国的，这是不是就成了一个让人担心的问题呢？

Is that a concern then, when we think about safety being baked into pre-training, when we think about disclosures, as we said there, and then juxtaposed with the majority of open source being Chinese?

是的，我觉得这里有两方面的事情。

Yeah, I think there are two sets of things.

我相信，如果你退一步看，AI安全缺失对中国、对美国和英国都不是好事。

I do believe if you step back somehow, AI safety not being there is not a good thing for China, is not a good thing for the United States, is not a good thing for the UK.

所以我觉得各国都会有这个共识，因为一个不受约束的开源中国模型很可能反过来攻击中国自己。

So I think that there's a part of everyone, because this is not something that somehow an unaligned open weight Chinese model will probably attack even China.

所以你不能想着它只会被用来当作攻击手段。

So it's not like you can rest think that this will only be used as an attack vector.

除了这一点，我对西方一些开源模型取得的进展还是挺满意的，当然，Jensen和其他人也在努力。

So setting that aside, I'm pretty pleased with some of the progress even the open weight models in the West are making, and obviously, Jensen's working on it, the other folks are working on it.

我们甚至会有MAI模型，它们不会开源权重。

We will have even the MAI models that will be not open weight.

我更多的是把它们看作授权权重。

I think of it more as licensed weights.

你可以这么理解。

Think of it that way.

我想把它变成一个简单直接的软件业务，而不是一个依赖网络效应的智能业务。

I want to turn this into just a simple, straightforward software business as opposed to a network effects intelligence business.

所以我认为会有不同的商业模式，最终我相信，你使用这些模型和控制它们的能力，才是决定哪些模型会被使用的关键。

So I think there will be different business models, and I ultimately trust in your ability to use those models and control those models is what's going to determine what gets used.

作为风险投资人，我必须预测今天流量会在封闭和开放模型之间如何分配。

As a venture investor, I have to predict where traffic will go today between closed, open.

你想想这两个不同的维度，一是流量，二是收入。如果看未来24个月，你觉得大概有多少比例的流量会流向开放模型，多少流向封闭模型？

When you think about the two different dimensions, which is traffic and dollars, if we look forward, say 24 months, what percent of traffic goes through open versus closed, do you think?

这个问题问得好。

It's a great question.

我觉得这取决于前沿模型的情况，对吧？

I think it depends on what the frontier looks like, right?

也就是说，我觉得大多数东西最终都会趋于饱和，一旦饱和了，开放权重模型，或者即使是封闭权重的更便宜版本，似乎就够用了。

Which is, I think most things seem to get saturated, and when it gets saturated, open weights or even just cheaper models of even closed weights seems to work.

因为特别重要的一点是，如果你能用自己的数据来训练，那问题没变的话，你其实不需要最新模型就能解决同样的问题，对吧？

Because especially one of the things is if you can train on your own traces, then you really don't need the latest model to be able to solve that same problem if the problem has not changed, right?

所以，如果你有一个评估标准想持续提升，那你可能就愿意花钱。

So if you have an eval that you want to continuously keep moving forward, then you may want to pay.

但即便是那样，顺便说一句，我对我们现在GitHub上这个HydroFusion的东西非常兴奋，因为它——

But even there, by the way, I'm very thrilled about this HydroFusion stuff we now have in GitHub where-

我刚看到了。

I just saw this

对，因为那里酷的地方在于它是一个学习模型。

Yeah, because the cool thing there is it's a learned model.

总之，我们现在正进入这些代理集群和多个子代理的阶段。

Anyway, we are now going into these agent swarms and multiple sub-agents.

所以我可以调用很多子代理，甚至这些子代理可能是本地模型。

So I can call a lot of sub-agents that even happen to be local models.

所以我怀疑，前沿模型总是会有市场的。

And so I think my suspicion is that there will always be a market for frontier.

总会有些新东西出现。

There will always be some new thing.

科学可能是下一个前沿领域的突破点，但我认为主要的令牌使用量会是商品化的令牌使用。

Science may be the next thing that frontier is being set at, but I think predominantly token usage will be commodity token usage.

那是不是意味着主要是开源的？

Does that mean it's open source predominantly?

可能是开放权重，或者像我说的，也可以是便宜的封闭权重，对吧？

It may be open weights or as I said, it can also be cheap closed weights, right?

因为会有很多价格竞争，我这么说吧，对吧？

Because there will be lots of price competition, let me put it that way, right?

因为你看我在copilot里的使用情况，我们用了很多闭源但便宜的模型，还有开放权重的模型。

Because if you look at my own usage inside a copilot, we use a lot of closed source, cheaper models along with open weight models.

你提到了拥有自己智能的重要性，

You said about the importance of owning your own intelligence-

……对你那些出色的客户来说。

... for some of your amazing customers that you have.

你之前提到英国的一些情况。

You were talking about some in the UK.

那是不是就是我们未来要面对的模式，公司用自己的数据来实现专业化智能？

Is that the model environment that we have moving forward, specialized intelligence for companies with their own data?

所以我觉得，这对我来说是最重要的事情之一，对吧？

So I think this is, to me, one of the most important things, right?

也就是，公司的未来是什么——

Which is what's the future of a firm-

在这个智能时代，对吧？

... in an era of intelligence, right?

因为即使追溯到哈耶克，他对经济的基本描述就是本地智能。

Because if you go back even to Hayek, his fundamental description of economy is local intelligence.

隐性知识分散在整个经济体中，对吧？

Tacit knowledge is distributed throughout the economy, right?

没有这个，就不可能只有一家公司，也不可能只有一种智能。

Without that, there can't be only one firm, and it can't be one intelligence.

所以这基本上就是最根本的东西。

And so that's sort of the foundational thing.

实际上，资本主义之所以能运转，靠的就是这个。

In fact, capitalism works because of that.

价格发现能起作用，或者说定价机制能起作用，靠的也是这个。

Price discovery works because of that, or pricing mechanism works because of that.

所以，我认为每家公司都要能够利用AI大规模积累知识，而不是让知识流失。

So therefore, I think it'll be important for every firm to be able to accumulate knowledge at increasing scales using AI versus leak.

这就是为什么我把它叫做“反向信息悖论”，对吧？

That's why I called it the reverse information paradox, right?

肯尼斯·阿罗描述过这个信息悖论：如果你是卖家，在没有专利、商业秘密保护或版权保护的情况下透露秘密，你就会泄露价值。

Which is Kenneth Arrow kind of described this paradox of information where if you're a seller and you disclose the secret without a patent or a trade secret protection or copyright protection, you leak the value.

而现在是第一次，买方需要保护，因为买方——即使大多数人觉得，哦，那是我的数据。

Now, for the first time, the buyer needs the protection because the buyer, even most people think of it as, oh, my data.

甚至是我使用过程中产生的数据，对吧？

It's even my exhaust of using it, right?

点赞或者点踩的信号。

Thumbs up, thumbs down signal.

你在为模型公司创造价值，而不是把这些价值留给自己。

You're creating value for the model company versus capturing that value for yourself.

所以我认为，至少我自己觉得，要保持稳定，就需要让更多由智能使用产生的剩余价值留在企业手里。

So I think one of the things that at least I ascribe is that we need to have, for this to be stable, more of that surplus that gets created by use of intelligence to remain at the firm.

顺便说一句，这不一定非得通过把你的知识嵌入权重来实现。

And by the way, it doesn't need to be only by embedding your knowledge in weight.

也可以放在上下文里。

It can be in context.

也可以放在你的评估里，对吧？

It can be in your evals, right?

你可能会有一个基准测试。

You may have a benchmark.

实际上，我很热衷的一件事就是每家公司都会有一个私有的评估基准，这是它们的核心知识产权。

In fact, one of the things that I'm very gung-ho about is every firm will have a private eval benchmark, which is their core IP.

每个模型进来都会试图在那个评估上往上爬。

And every model comes in and tries to sort of hill climb on that eval.

对我来说，这就是公司运作的方式。

And that, to me, is how a firm works.

我喜欢这个。

I love this.

每家公司的私人和评估基准。

Private evals, benchmarks for each company.

这听起来像是少数精英公司的专属，比如Harvey或Agora，或者那些非常前瞻的AI公司。

This is for a rarefied air of companies, so it would seem, your Harvey's or your Agora's or the kind of very forward-thinking AI companies.

这会像你提到的，扩散到世界上的每家公司吗，用你的——

Will that disperse to every company in the world as you mentioned with your

是的，它会——

Yeah, it happens-

...合作伙伴？

... partners?

对，哈里，这种事一直都会发生，对吧？

Yeah, it happens, Harry, always, right?

我记得在80年代末、90年代初的时候，那时候只有少数公司部署了个人电脑，并且在做预测分析，那感觉就像魔法一样。

Which is, there was a day when I remember in the late '80s and the early '90s that the few companies that had deployed PCs and were doing- ... forecasting, which felt magical.

他们能把数据放进Excel表格里，通过电子邮件发送，不用再用传真机和公司内部备忘录，那时候人们会觉得：‘哇，这什么时候才能普及啊？’

They were able to put it in Excel spreadsheets, send it in email, and avoid the fax machine and interoffice memos, and it felt like, "Oh, wow, when will this happen?"

可能花了10年时间，这项技术才慢慢普及开。

And it took 10 years maybe, and it spread out throughout.

现在被认为是‘我们是AI原生公司，我们做这些’的事情，我觉得五年之后就会变得很普通了。

This time around, what is considered a, "Oh, I'm an AI native company, and I do this," I think, five years from now, it'll be pretty passe, I would imagine.

然后那就成了……

And then that's what...

这些科技周期总是这样开始的，对吧？

These tech cycles always start like that, right?

就连云计算也是一样。

Which is even cloud.

我记得云计算刚开始的时候，大家最关心的就是谁是不是云原生。

I remember in the beginning of cloud, it was like the big deal was, oh, who's cloud native and who's not?

如果你不是云原生，那现在大家都是了，对吧？

And if you're not cloud native, and now everybody is, right?

就连Lloyds，那家大概有500年历史的公司，现在也是云原生。

Even Lloyds, which is, what, 500 years old, is cloud native.

所以事情就是这样发展的。

And so it happens.

那我们是不是要进入一个实施经济了？

Do we move into an implementation economy then?

因为想想这些客户，比如说你提到的一些大客户，他们会需要很重的FDE支持，比过去十年里传统的支持要多得多。

Because when we think about these customers, say, as you said, some of your very big customers, they will need heavy FDE support, much more so than they have done in traditional years, within the last 10 years.

我们会不会有一个非常大规模的实施经济，来把现有客户转到这个上面来？

Will we have a very heavy implementation economy to transition existing customers into this?

对，这确实很有意思，对吧？

Yeah, it's interesting, right?

实施经济嘛，系统集成商其实一直都存在的，对吧？

The implementation economy, like systems integrators have always existed, right?

我们一直依赖实施经济来推广其他很多技术。

We've depended on implementation economy for lots of other technology diffusion.

所以这一次，我觉得情况有点不一样，对吧？

So this one, I think, has different contours, right?

因为——

Because-

你觉得这次有什么不同呢？

What do you think is different about this one?

可以说是技术和业务的结合吧。

It's the combination of tech and business, if I could say.

但它既不是传统的管理咨询，也不是传统的系统集成，也不是传统意义上的技术支持。

But it's not classic management consulting, it's not classic systems integration, and it's not classic sort of technology support.

它有点像——比如说，你想想看，对吧？

It's sort of a-- For example, think about this, right?

你想开始。

You want to start.

看看我的FDE项目，就拿我今天在主题演讲中提到的联合利华例子来说。

When I look at my FDE engagements, right, even take the Unilever example I talked about at the keynote today.

真是让人着迷。

It's fascinating.

他们建了一个代理系统，说：‘嘿，我们真的想用代理系统实现订单到现金的自动化，’再加上他们的贸易索赔。

They've built an agentic system to say, "Hey, we want to really have order-to-cash automated using an agentic system," plus their trade claims.

而且，他们一开始用的是前沿模型，然后他们换了，实际上，我觉得他们现在用的是MAI模型来处理所有贸易索赔，因为这样成本更低。

And by the way, they started using a frontier model, and then they substitute for, in fact, I think they're using currently an MAI model for all the trade claims because it's sort of a lower cost.

顺便说一下，他们控制着这一切，对吧？

And it's, by the way, they control it, right?

因为他们可以在自己的追踪数据上训练。

Because they're able to train on their traces.

所以我们有一些FDE，但大部分部署的其实很多，现在他们的IT团队自己管理这一切了。

And so we had some FDEs, but a lot of who got deployed, and now their IT team itself is managing it all.

但它一开始关注的是业务是什么，评估是什么，以及如何构思智能体架构。

But it started with what's the business, what's the eval, what's the way to conceptualize the agentic architecture?

所以这有点像商业咨询和技术咨询的结合。

So it's sort of business consulting combined with technical consulting.

而且我觉得，是的，你说得对，生态系统中会培养出一套新的技能。

And I think, yeah, you're right in saying that there will be a new set of skills that'll get built in the ecosystem.

今天，我想我们都有所谓的“自己的FDE”，但FDE会变得像传统的系统集成一样，成为新的FDE。

Today, I think all of us have, quote-unquote, "our own FDEs," but FDEs will become like traditional systems integration will become new FDE.

你提到了，显然，MAI。

You've mentioned, obviously, MAI.

但当我们看很多不同的玩家时，大家逐渐意识到你需要成为全栈的，从芯片到模型再到应用。

But when we look at a lot of different players, it's kind of this realization that you need to be full stack, from chips to models to application.

这是未来吗？

Is that the future?

你必须拥有垂直的所有权吗？

You have to have vertical ownership?

对，这完全取决于行业结构，对吧？

Yeah, it all depends on the industry structure, right?

我觉得今天可能会有五六个闭源模型在竞争，也许还有五六个开源权重模型在竞争。

I think that I would say today, there'll be, who knows, five, six closed models all competing, maybe five, six open weight models all competing.

所以某种程度上，如果你在构建像代理系统这样的应用，你甚至可以用我之前说的那种方式，就是带内置决策引擎的路由器，然后你只需要根据具体任务路由到最好的模型。

And so therefore, at some level, if you're building an application, like agentic system, you could even have something like what I said, which is a router with even a decision engine built into it, and then you just route to the best model for a given task.

所以那是一种可能性。

So that's a possibility.

当然，也总是存在一种风险，就是可能会出现一个从芯片到应用层完全垂直整合的单一系统。

Of course, there's always the risk that there could be one vertically integrated model and app tier and all the way to the silicon end system.

但我觉得这不太可能，对吧？

It's unlikely to me, right?

我一度以为可能会出现这种情况，就是只有一个这样的系统。

I thought at some point that that might be the case, that there will be one.

但如果有的话，趋势反而是相反的，每天你醒来都会发现，每一层都在出现更多竞争，不管是芯片层面、芯片管理系统的层面、模型层面，还是应用层面。

If anything, it's going the other way, which is every day you wake up and you say, "Wow, there's more competition in every layer of the stack," whether it's in silicon, whether it's in systems to manage the silicon, in the model layer, in the app layer.

所以我觉得这是一个健康的发展方向，但谁知道呢，对吧？

And so I think that's a healthy direction of travel, but you never know, right?

因为这种事情，一切都取决于哪个设计会成为主流。

Because this is one of those things where it all depends on what becomes the dominant design.

但我希望这能变成一种横向的细分，每一层都有多个赢家，然后人们能够自由组合。

But my hope is that this becomes a little bit of a horizontal specialization with multiple winners in each layer, and then the ability for people to compose.

我没想到的是每一层的深度。

The thing I didn't see is the depth of each layer.

我现在知道而十二个月前希望知道的是，推理的需求是无穷无尽的。

What I know now that I wish I'd known 12 months ago is just the infinite demand for inference.

说得好。

That's a great point.

如果十二个月前我知道这一点，我的投资决策会非常不同。

And I wish I'd known it 12 months ago, I would've done some very different investing.

说到这个，穆斯塔法·苏莱曼说过，千亿规模的训练即将到来，而只有少数实验室能负担得起。

Speaking of that, we've had Mustafa Suleyman saying training rounds of 100 billion are coming, and only a handful of labs can pay for that.

1000亿美元的训练运行。

$100 billion training runs.

我们有没有足够的能源、基础设施和电力，来支撑这种规模的整体堆栈，特别是当我们——

Do we have the energy, infrastructure, and power to enable this scale across the stack, but especially when we-

我的看法是，最终这得靠全球GDP的增长，然后让这些增长的一部分用于研发，因为当人们提到训练运行，它只是研发投入的一部分。

The way I come at it is, ultimately, it has to come from the world GDP inflecting, and then that allowing someone to then have as a percentage of R&D, because when somebody says training run, it's just a percentage of R&D.

那些研发投入大的公司，在研发上花多少钱？

Companies that are heavy R&D spend what?

15%到20%吧。

15%, 20% on R&D.

所以，是的，如果收入足够，你就能在训练上投入足够多的钱。

And so therefore, yeah, at some point, if you have enough revenue, you can spend enough on training.

然后，当然，如果你为世界经济增长贡献了力量，你就会获得相应的收入，也就是说，你能分到一杯羹。

And then, of course, you'll have enough revenue if you're contributing to the world economy growing so that you can get your fair share, so to speak.

然后你还提到了另一面，也就是供应端，我们目前面临电力短缺、芯片短缺，还有各种各样的问题。

And then you also bring up the other side, which is the supply side, which is we have power shortages, we have silicon shortages, we have all kinds of things today.

这些东西的交货周期很长。

They're long lead items.

没法打个响指就变出发电设施，也没法凭空冒出一座新工厂。

You just can't snap your finger and have power generation, or you can't have a new fab.

所以现在有趣的是，因为需求大，大家都在建。

And so now, interestingly enough, given the demand, everyone's building.

问题是，这些什么时候能全部上线？

And so the question is, when will all of this come online?

供需会变成什么样？

What will be the demand supply?

没人知道。

No one knows.

说实话，今天在座各位都只能靠猜。

Quite frankly, any one of us who are sitting here today will just be speculating.

但我的感觉是，如果你先假设最终稳定下来的状态跟今天差不多。

But my feeling is that if you start by thinking of the stable state would be just like today.

公司的损益表不会突然变得不一样。

Income statements are suddenly not going to be different for a company.

也就是说，嘿，我有一些收入，我也有一些成本，而且你得有点利润。

It's going to be, hey, I have some revenue, I have some cost of revenue, and you have to have some profit.

如果真是这样，那么问题就是，全球GDP增长率是多少？

And if that is going to be true, then the question is, what is the world GDP growth rate?

这对我来说才是更重要的事情。

That's the bigger thing for me.

关于GDP增长率的膨胀，世界在很长一段时间内都稳定在每年2%左右。

In terms of the inflation of GDP growth rate, the world has kind of stabilized at 2% a year for a very large amount of time.

人工智能有没有可能在短时间内大幅推高这个增长率？

Does AI have the ability to seriously inflate that in a short amount of time?

我觉得会，而且某种程度上我希望如此，因为这才是重点。

I think so, and I somewhat hope so because that is the point.

就拿我们各国GDP中医疗保健支出来说，哪怕只是作为一个比例来看。

Let's take even as a percentage of any of our GDPs, the amount of money we spend in our healthcare.

所以这就是为什么我看英国国家医疗服务体系部署Copilot的时候会觉得，哇，如果它能在医疗领域改善病人治疗结果并降低成本，那本质上就是在帮助英国GDP增长。

So that's why I look at even the deployment of Copilot in NHS and say, wow, if it can improve patient outcomes and reduce cost in the healthcare, that is by definition helping the GDP of the United Kingdom.

所以我觉得经济中有这些适合采用人工智能的领域，而且实际上，这些领域人员流动最少，收益却最大。

And so I think that there are these pockets of the economy which are ripe for adoption of AI, and really, in fact, there's the least amount of displacement, the most amount of benefit.

所以我认为正是在这些领域，我们需要找到最佳的应用点。

And so that I think is where I think we'll have to find the best fit.

软件开发是另一个例子。

Software development is another one.

让人着迷的是，软件开发的需求弹性竟然这么大。

It's fascinating to see how much the demand is elastic for software development.

这意味着IT积压的需求量非常大，我们甚至都没意识到这一点。

That means the IT backlogs were huge, and none of us even recognized it.

萨提亚，我以前学到的是，要在风投领域获胜，你必须成为市场中的前两名。

Satya, I was taught that you had to be the top one or two in a market to win in venture.

而我们在Factory，排名第五，但我们正在赢。

We're in Factory, which is fifth, and it's winning.

第五名。

Fifth.

这完全打破了风投的规则。

This breaks all rules of venture.

我有一个问题，还是有点不按套路，不在计划之内，但就是想知道，我们会不会在替代职能方面看到能力提升的拐点，比如财务、销售、市场，就像我们在编程领域看到的那样？

A question that I have, again, slightly naughty, not on schedule, but it's just like, will we see the tipping in terms of capability advancement in alternative functions, finance, sales, marketing, that we've seen in coding?

因为编程一直是最大的金矿。

Because coding's been the mother lode.

我觉得这是个好问题。

I think that that's a good point.

人们经常讨论，你有没有可验证的回报，而做这些很容易，因为很多神奇之处在于预训练模型之上的强化学习。

People talk a lot about, do you have a verifiable reward and it's easy to do these because a lot of the magic has been RL on top of these pre-trained models.

正是这个推动了编程的发展。

That is what has helped coding.

但在知识工作中能否突破这一点，似乎要难得多。

Can we crack that in knowledge work seems much harder.

但再说回来，我之前提到的那些例子，像呼叫中心、供应链系统、还有营销和现金到订单这些系统，我觉得都是有可能的，而且我们已经看到了。

But again, all the examples I talked about, sort of in contact centers, in these supply chain systems, in these marketing/cash to order systems, I think it's possible, and we're seeing it.

但需求是不是像编程那样有弹性呢？

Is the demand as elastic as coding?

我觉得也许科学领域也会有同样的弹性。

I think maybe science may be as elastic.

或者整个工程领域。

Or engineering in general.

顺便说一句，我看看内部正在发生的变化，不只是软件开发，就拿芯片设计来说吧。

By the way, if I look at even what's happening inside, not just in software development, let's take silicon development.

那是另一个正在经历剧变的领域。

That's another thing that's going through a radical shift.

你流片的速度快多了，在计算机上模拟一切的能力也大大提升了。

Your ability to tape out much faster, simulate everything in silico is much better.

你的第一次流片就能完全顺利出来，

Your first spin can completely come out-

但即使只是维护的优化——

But even just optimization of maintenance-

哦，对

Oh, yeah

……就像你提到的罗尔斯·罗伊斯的报告，我从来没想过——

... reports like you said with Rolls-Royce, which I would never-

真是难以置信。

It's unbelievable.

他们一直要处理这个问题，就是安排维护，我之前没意识到，哇，那是个很大的优化问题，现在可以运行了。

It's a thing that they've been dealing with forever, just schedule maintenance, and I had not realized that, wow, that's a big optimization problem that you can go now run.

我看过很多你的节目，我们来个快速问答环节吧。

So I've watched many of your shows, and we're going to do a quick fire round.

好的。

Yep.

有些问题别人没问过，但我一直想知道，我觉得大家都想知道，但还没人问过。

And people haven't asked some of these questions, which I've always wanted to know, and I think everyone wants to know, but people haven't asked them.

萨蒂亚，你经常出差，怎么保持锻炼的习惯？

What's the exercise routine, given the travel schedule, Satya?

这个嘛——

Well-

你身材保持得挺好的。

You're in good shape.

嗯，我尽量吧。

Well, I try to be.

我也不确定是不是藏得好。

I don't know whether it's all well hidden.

不过有件事我坚持了大概25年了，就是每天早上起来，我以前还能跑步。

But the one thing I've gotten into the habit, this has been now maybe 25 years even on, which is every morning I get up and I used to be able to run.

可惜膝盖不行了，所以现在至少我会走坡道。

Unfortunately, my knees have gone away, and so now at least I walk with incline.

就算今天早上，我也抽了半小时锻炼，所以不管在哪里，不管时差多严重，我每天都去健身房。

Even this morning, I was able to get a half hour in, and so that's what, in spite of any jet lag, wherever I am, I hit the gym every day.

我的膝盖也不太好。

I too have bad knees.

也是因为年纪大了。

It's my age as well.

过去12个月里，你改变看法的事情有哪些？

What have you changed your mind on in the last 12 months?

可能是因为，毕竟我是最早的那批人之一，当时达里奥、伊利亚、山姆、格雷格和埃隆都在，我想想，也算是一家人吧。

It might be because, after all, I was one of the first, obviously when Dario and Ilya and Sam and Greg and Elon were all, I guess, one happy family.

他们真的是规模定律的坚定信徒，而我在2018年、2019年也是这么相信的。

They were really the believers in scaling laws, and that's what I believed in back in 2018, 2019.

我觉得现在最大的变化是，我以前总觉得对齐问题会自然而然地解决。

And I think the biggest change now I have is I used to always think that alignment will just fall out.

我现在不再这么认为了。

I no longer believe that.

所以我觉得对齐这个问题可能更棘手，也更复杂。

So I think this alignment thing may be a harder, different problem.

所以，是的，这就是我改变想法的那件事。

And so, yeah, that's the one thing I've changed my mind on.

今天看起来疯狂，但五年后会变得非常普遍和被接受的事情是什么？

What looks crazy today, but will be incredibly common and accepted in five years?

可能就是刚才我们谈到的那个，那些评估。

It might be that thing that we just talked about, those evals.

什么是公司？

What is a firm?

我认为公司会是一个AI基准。

I think a firm would be an AI benchmark.

最后问你一个问题。

Final one for you.

未来十年有很多令人兴奋的事情。

There are many things to be excited about over the next 10 years.

对你来说特别的话，比如对我来说，我妈妈患有多发性硬化症。我对可能实现的药物开发感到非常兴奋。

For you specifically, like for me, it's my mother's got MS. I'm so excited for the drug development that could be capable.

你最期待的是什么？

What are you most excited for?

这个问题问得好。

That's a great one.

我觉得我一直期待的，就是一种通用技术，即使像现在这么强大，人们还是会说：“哦，这不只是一种通用技术，”不管它是什么。

I think one of the things that I've always been, I would say, been looking forward to is a general-purpose technology, even as powerful as this, and people say, "Oh, this is not just another general-purpose technology," whatever this is.

我希望它能比任何东西都更快地传播到全世界，渗透到经济的各个领域，不像其他任何东西。

I hope this spreads, diffuses Throughout the world faster than anything else, and throughout all sectors of the economy, unlike anything else.

而且终于有一次，全球增长的前沿真正发生了转变，不只是计算赢家和输家，而是好像所有人都被提升了。

And truly for once, there is a real shift in the frontier of that global growth that happens, where it's not about just counting winners and losers, but it's kind of like everyone gets lifted.

那一直是我的梦想。

That's been my dream.

其实，我父亲是一位发展经济学家，他总是关注收敛增长。

In fact, my father was a development economist, and it was always about convergence growth.

全球南方什么时候能得到某种能突然推动它的东西呢？

When will the Global South get something that just suddenly boosts that?

我希望就是它了。

I hope this is it.

萨蒂亚，能和你聊这些真是太高兴了。

Satya, this has been such a joy to do.

非常感谢你抽出时间。

Thank you so much for taking the time.

当然。

Absolutely.

这是我的荣幸，哈里。

It's a pleasure, Harry.

但在我们结束之前，Base44是我用过的产品之一，我当时立刻就想，‘天哪，真希望我当初创办20VC时就有这个，那时候我只有50美元，在风投圈也没什么人脉。’

But before we leave you today, Base44 is one of those products I used, and I just thought straight away, "Man, I wish I'd had this when I started 20VC with $50 and no contacts in venture."

我想看看它能做什么，所以我就为节目建了一个嘉宾研究中心。

I wanted to see what it could do, so I built a guest research hub for the show.

未来的嘉宾、公司简介、以前的采访、赞助商联系方式、后续跟进，全都集中在一个很棒的地方。

Upcoming guests, company notes, previous interviews, sponsor contacts, follow-ups, all in one amazing location.