第12集 - 人工智能行业现状

你好。

Hello.

我是安德鲁·梅恩，这是OpenAI的播客。

I'm Andrew Main, and this is the OpenAI podcast.

今天，我们的嘉宾是OpenAI的首席财务官莎拉·弗莱尔，以及知名投资者、Khosla Ventures的维诺德·科斯拉。

Today, our guests are Sarah Fryer, CFO of OpenAI, and legendary investor Vinod Khosla of Khosla Ventures.

在本次讨论中，我们将探讨人工智能生态系统的现状，我们是否处于泡沫之中，以及在人工智能发展过程中初创公司和投资者如何取得成功。

In this discussion, we're going to talk about the state of the AI ecosystem, whether or not we're in a bubble, and how startups and investors can succeed as AI progresses.

与Netflix这样的平台不同，它们每天播放大量内容，我认为人工智能更像基础设施，比如电力。

Unlike something like Netflix, where they're running so many hours in the day, I think of it much more like infrastructure, electricity.

今天，需求的限制因素不是别的，而是计算资源的可用性。

Demand is limited, not by anything other than availability of compute today.

我认为我们需要讨论的问题是：人们将会做什么？

I think the conversation we need to have is what will people do?

2025年将是关于智能代理和氛围编程的一年。

2025 was about agents and vibe coding.

现在是2026年。

Now it's 2026.

2026年的故事是什么？

What's the story of 2026?

我认为我们在2025年已经让氛围编码更加成熟了。

I think we matured in vibe coding in 2025.

我认为我们在智能体方面还没有成熟。

I don't think we've matured in agents.

所以智能体，尤其是多智能体系统，将会成熟到产生切实可见的影响。

So agents, especially multi agentic systems, will mature to the point of having real visible impact.

无论你是企业，拥有一个多功能智能体系统为你完成完整任务，比如每天为你运行ERP系统，处理所有对账、计提和合同跟踪。

Whether you're an enterprise and you have a multi agent systems doing full task, like running an ERP system for you, you know, doing all the reconciliation every day, accruals every day, tracking contracts every day.

我认为这在企业层面是如此。

I think that on the enterprise side.

但今天，在消费者层面，计划一次旅行仍然很麻烦。

But today, on the consumer side, you know, it's still a hassle to plan a trip.

这种多智能体系统会综合考虑你的饮食偏好、餐厅预订、航班时刻表以及个人日程等多个方面，我认为一年后就会开始成熟。

That's a multi agentic thing that looks across a lot of different things from your food preferences, to the restaurant preservation, to airline schedules, to your personal calendar, those will start to mature, I think a year from now.

所以我对这一点非常期待。

So I'm pretty excited about that.

我认为，机器人领域以及超越机器人本身的现实世界模型，比如通用直觉，都将在明年开始出现。

I think models in robotics and real world models that go beyond well beyond robotics, like general intuition will all start to happen in the next year.

所以我认为这些是值得关注的领域。

So I think that those are areas to look for.

像LLM中的记忆功能、持续学习、减少幻觉影响等，都是常见的功能。

There's usual functions like memory in LLMs, continual learning in LLMs, reduction of the impact of hallucinations.

这些都是AI目前表现不佳的领域，我还能列举更多，大约有六七个方面，未来将开始得到改善。

Those are all areas, I could go on, there's half a dozen areas in which AI doesn't do as well today that will start to be addressed.

是的。

Yeah.

我认为，从根本上说，维诺德的意思是：2026年将是缩小这一能力差距的起点。

And I think at its baseline, what Vinod is saying is '26 is the beginning of closing this capability gap.

所以我们知道，我们已经赋予了人们巨大的智能。

So what we know is we've handed people massive intelligence.

对吧？

Right?

我们把法拉利的钥匙交到了他们手里，但他们才刚开始学习如何上路驾驶。

We've handed them the keys to the Ferrari, but they are only learning how to take it out on the road for the first time.

我们需要为消费者提供越来越多简单易用的方式，让ChatGPT不再只是一个聊天问答工具。

We need to give consumers more and more easy ask, easy ways to go from ChatGPT is just a chatbot call and response.

今天大多数人只是用它来提问。

Most people use it today just to ask questions.

但我们要如何让它演变成真正的任务助手，帮他们预订旅行、为他们刚从医生那里听到的信息提供第二意见，或帮他们为患有糖尿病的孩子设计菜单呢？

But how do we take it towards being a true task worker that that books that trip for them or helps them get a second opinion on what they just heard from their doctor or enables them to create a menu for their diabetic child?

对吧？

Right?

我们该如何帮助他们从简单的提问，转向真正改善生活的实际成果？

How do we help them really move from simple questions into actual outcomes that make my life better?

在企业端，情况也是同样的连续过程。

And then on the enterprise side, it's that same continuum.

我们如何缩小能力差距？

How do we close the capability gap?

我们从首席经济学家去年年底发布的《企业人工智能现状报告》中了解到，前沿企业与普通企业之间的差距。

One of the things we know from our state of the enterprise AI and the enterprise report that our chief economist put out at the end of last year is on the frontier versus just even the median corporation.

平均消息量或中位数约为六倍，这表明前沿企业的使用量是普通企业的六倍。

The the average number of messages or the median is about six x, which will tell you that six x the usage from a company that's already on the frontier.

我们知道，前沿企业甚至还没有达到其最大潜力。

And we know that frontier isn't even pushed to its max.

因此，对我们而言，重点是如何帮助消费者沿着这一连续过程，迈向真正的智能任务执行？

So for us, it's this focus of how do we help consumers move along that continuum to true agentic task working?

而对于企业，我们如何创建更复杂、垂直专业化的成果，使他们能够从简单的ChatGPT应用，逐步发展到彻底改变其核心业务的系统？

And then for enterprises, how do we create a much more sophisticated, vertically specialized outcome for enterprises that allows them to go from maybe a very simple chat GPT implementation the whole way to something that's transforming the most important part of their business.

对于医疗提供商来说，这可能是他们的药物研发流程。

For a healthcare provider, it might be their drug discovery process.

对于一家医院来说，可能是缩短患者入院时间，帮助患者更快回归社区。

For a hospital, it might be the time to admit a patient to get that patient back into the community.

对于一家大型零售商来说，可能是增加购物车规模、提高转化率，以及让客户更满意。

For a really large retailer, it might be just larger basket sizes, higher conversion rates, and much happier customers.

因此，这本质上是缩小能力差距的基础。

So it's the basics of closing that capability gap.

我可能会补充另一个视角。

So I might add one other perspective.

我们已经讨论过技术将取得进展的多个领域，以及能力将如何提升。

We've talked about the number of areas in which the technology will advance and capability will advance.

我猜测，如今使用人工智能的人，无论是个人还是企业，只有个位数百分比的人甚至用到了人工智能30%的能力。

I would venture to guess today, of the people using AI, whether it's personal or enterprise, some single digit percentage are even using 30% of the capability of the AI.

因此，使用人工智能30%或50%能力，更不用说80%能力的人群比例将持续增长。

So this percentage of people who are using 30% or 50%, let alone 80% of the AI's capabilities will keep increasing.

我认为，人们要掌握如何使用人工智能，还需要十年时间。

I think that's a ten year journey before people learn to use AI.

我见过这种情况，有些人，一些评论家把采用曲线和能力曲线搞混了。

I've seen this, some people kind of pundits confuse adoption curves for capability curves.

这确实出现过，你已经看到了。

And that's that's come up where you've seen.

这正是我想表达的观点。

So that's the point I'm making.

这是一个倍增器，因为今天已经有超过8亿人每周在使用ChatGPT。

And it's a force multiplier because today we have over 800,000,000 using ChatGPT today, 800,000,000 consumers weekly using.

但你知道，这个数字应该达到数十亿。

But, you know, that number should be in the billions.

那么，他们使用它的比例是多少呢？

And then what percentage use are they using it for?

这就像我们刚刚在家里通上了电。

It's like we've just turned electricity on in the home.

我们已经布好了电线，人们打开了灯，但他们根本不知道现在还能用它来给家里供暖。

We've wired up the home and they've turned on the lights, but they have no idea that they could now heat their home.

他们可以用它做饭，也可以用它卷头发，对吧？

They could cook, they could curl their hair, right?

现在你能做的事情太多了。

There's so many things you now can do.

我用过一个类比：从1990年到2000年，电子邮件并没有明显改进，手机也是如此，但使用人数却大幅增长，问题不在于‘我们需要更好的电子邮件，需要更先进的手机’。

An analogy I've used is that email didn't really get much better between 1990 and the year 2000, neither did mobile, but usage went way up and the problem wasn't like, well, we need better email, we need more better mobile.

而是人们需要学会所有能用它来做的事情。

It's like people need to learn all the things they could use it for.

没错。

Right.

是的。

Yeah.

更复杂一点来说，手机就是一个很有趣的例子，当手机普及时，人们只是把桌面网站直接搬到手机上，结果页面很难滚动。

And in a more sophisticated way, like mobile is always one that's interesting to me because when mobile took off, people just took their desktop websites and turned them into mobile and they were really hard to scroll.

但至少你把它放进了口袋里。

But I guess you at least had them in your pocket.

但后来你意识到你有了GPS。

But then you realized you had a GPS.

所以现在你可以使用Uber，也可以利用位置功能，或者随时用上相机。

So now you could have Uber, and now you could do things with location or you had a camera at your fingertips.

好吧。

Okay.

所以现在，是的，我可以给所有朋友拍照，但我也能拍下支票并存入银行账户，尽管我们该解决一下纸质支票的问题。

So now, yeah, I can take photographs of all my friends, but I can also snap, you know, a check and deposit it into my bank account, although we should fix the whole paper check thing.

但那是题外话。

But that's on the side.

它看起来仍然如此。

It still seems like it.

但我可以

But I can

只需拍下这张照片

just take a photo of this

现在我的银行账户里有钱了。

and now I get money in my bank account.

是的。

Yeah.

但你知道，这些功能在手机普及的那一刻就已经存在了。

But, you know, those those that all existed in in the minute mobile was available to us.

但更重要的是，人类的创造力得以发挥作用。

But just the, you know, the the the ability for human ingenuity to come to work on it.

所以我认为你是对的。

So I think you're right.

我甚至不知道，我们是否需要比今天更多的智能，就能大幅提高成果。

I don't even know if we need more intelligence than we have today to vastly increase outcomes.

但当然，这些模型也会继续变得更加智能。

But of course, the models are going to keep getting more intelligent as well.

你提到了健康，这是我们思考时最看重的领域之一，可能也是最重要的事情。

You mentioned health and that's one of the really kind of high stakes things we think about when it comes to just probably the most important thing.

想想几年前我们刚得到ChatGPT时，只用它来做一些非常简单的应用，而现在我们却开始让它处理符合HIPAA标准的敏感数据，这确实很有趣。

And it's kind of fascinating to think about the gist, you know, a few years ago we got ChatGPT and we're using it for very simple applications and now we're trusting with HIPAA compliant data.

你是否认为这是衡量技术加速发展速度或程度的一个标志？

Do you look at that as sort of a marker of how fast or how well things have been accelerating?

你还能想到其他类似的例子，来说明我们已经进入了一个全新的阶段吗？

Are there other ones like that you think about to say, okay, now we know we're some new level.

健康显然是其中一个领域，我长期以来一直相信，AI将通过让专业知识变得普及来彻底改变医疗行业。

Health is clearly one of those areas I've long believed it'll revolutionize health by making expertise be a commodity in all areas of health.

但医疗领域的问题在于监管。

The problem with health is regulatory.

首先，AI能做的事情受到诸多限制。

So first, there's constraints on what AI can do.

即使AI开处方的能力比人类更强，它也无权合法开具处方。

And AI can't legally write a prescription, even if it's better than human beings at writing a prescription.

这不仅是FDA的限制，更是由美国医学会等机构在制度层面上控制着这一职能。

That is not only the FDA, but it's actually beyond the FDA into the American Medical Association institutionally controls that function.

因此，在许多领域都会存在既得利益者的抵制。

So they will be incumbent resistance in a lot of areas.

如果你愿意，我们可以聊聊这个。

I think we can talk about it if you like.

但诊断仍然是一个限制，因为FDA对此有控制权。

But diagnosing is still a constraint because the FDA controls that.

目前还没有任何AI被批准为医疗设备。

There's no AI approved as a medical device yet.

幸运的是，这一届政府正在迅速行动，并采取了适当的风险水平。

That all, fortunately, this administration is doing a very good job of moving quickly and taking the appropriate level of risk.

因此，我对那里发生的事情感到非常满意。

So I'm pretty pleased to see what's happening there.

在医疗领域，我们的数据显示，每周有两亿三千万人向ChatGPT询问健康问题。

On the health front, we see in our data, two thirty million people every week ask CHAT GPT a health question.

66%的U。

Sixty six percent of U.

S.

S.

医生表示他们在日常工作中使用CHAT GPT。

Physicians say they use CHAT GPT in their daily work.

我跟你讲个个人经历，我哥哥是英国重症监护室的医生，他的工作就是你急诊室一到，没人知道怎么分诊你，就把你推给他。

I'll tell you at a personal level, my brother is an HDU doctor in The UK, so his job is you hit the ER, They don't know how to triage you.

所以他们把你送到他那里。

So they send you to him.

你其实不太想去找他。

You kind of don't want to show up to him.

但他被期望要

He's expected to

不过他非常出色。

have He's very good, though.

他做他这份工作非常在行

He's very good at what he

但他意味着你情况不太妙。

does, but it means you're not in good shape.

但他被期望拥有几乎百科全书般的知识，了解所有存在过的疾病。

But he's expected to have an almost an encyclopedic knowledge of every disease that ever existed.

所以我总是举这个例子。

So I always give the example.

他在苏格兰的阿伯丁工作。

He works in Aberdeen in Scotland.

如果你得了疟疾去找他，他根本不会往那方面想。

If you showed up with malaria, he will not think of that.

这不在他的模式识别范围内。

That is not in his pattern recognition.

然而这种情况是有可能发生的。

And yet that could have happened.

我不知道。

I don't know.

你夏天去度假，被蚊子叮了。

You went on vacation summer, you got bitten by mosquito.

突然间，你出现在阿伯丁的急诊室。

Boom, you're showing up in an ER room in Aberdeen.

ChatGPT 或模型能做的，其实是为医生提供强大的辅助，这就是为什么我认为有百分之六十六的医生在使用它。

What chat GPT can do or what the model can do is really act as a great augmentation to the doctor, which is why I think sixty six percent of them are using it.

而且这个数字还在持续增长。

And that number is only growing.

你知道，实际比例可能已经高得多。

You know, it's probably already, much higher.

因此，我认为这很好地说明了在医疗领域，我们的医生能够随时接触到最新的研究成果，了解最新的药物方案与患者个体生活和体验之间的相互作用。

And so I think it's just a great example of where something like health, we're getting the benefit of our doctors being able to have always the latest research in front of them, always the latest known, interactions, say between someone's drug regime and what they're living through and experiencing as individuals.

但它也把一部分自主权重新交到了消费者手中。

But it also puts some independence back into consumers' hands.

因此，我现在有机会提前研究一下我的症状可能意味着什么，以便能与医生进行更有依据的对话。

So now I get the opportunity to, ahead of time, do some research on what my symptoms might be saying so I can have a much more educated conversation with my doctor.

它让我能够获得第二意见，或者知道我应该去寻求第二意见。

It allows me to maybe get a second opinion or know that I want to go ask for a second opinion.

我们虽然很容易走向极端，但即使是像‘我每天只有二十分钟锻炼’这样的小事也一样。

It also we go very fast to these extreme places, but just to even things like, hey, I've got twenty minutes a day to exercise.

我知道我患有1型糖尿病。

I know I'm suffering from type one diabetes.

那我二十分钟能做些什么呢？

What what could I do in twenty minutes?

或者我女儿在饮食上有一些特殊的问题。

Or my daughter has an interesting issue with the food she eats.

以前去餐厅是件特别令人沮丧的事，因为我们得向服务员问一大堆问题。

And so it used to be a super frustrating thing to go to a restaurant even because we'd have to almost ask the server so many questions.

现在我们可以拍下菜单，聊天助手会推荐最适合她点的菜品。

And now we can photograph a menu, chat suggests what are likely the best dishes for her to order.

然后我们可以进行更有成效的对话，更好地决定哪些食物适合她。

And then we can have a bit more of a of a tercer conversation, but a bit more productive on what's going to work.

它彻底改变了我们对饮食的看法，让我们不再只关注食物本身，而是更注重为什么我们要一起外出用餐。

And it has just changed how we think about just eating, takes it away from all about the food to why we're going out for dinner together.

所以我认为，这些都是健康领域的一些典型案例。

And so I think they are all these just examples of something like health.

这种事情已经在发生，并且会变得越来越好。

It's already happening and it's going to keep getting better and better.

至于法诺瓦特的观点，我认为监管环境必须跟上步伐。

And then to Fanowat's point, I think regulatory environment is going to have to catch up.

无论你处于什么样的医疗体系下，医疗成本的增长速度都已超过每个国家的GDP。

It's no matter what kind of system you're under, the cost of medical care is exceeding the GDP of every country, the rate at which increases.

我们似乎急需人工智能，现在就需要它，而且它确实能提供帮助，正如你所指出的，这是医疗智能成本首次出现逐年下降的情况。

And it seems like we needed AI, we needed it now and, you know, it can be helpful and as you pointed out, it's the first time the cost of medical intelligence has dropped year over year.

但这带来了对计算资源的巨大需求，我们还有更多问题希望得到解答。

But that comes with a lot of demand for compute and we have a lot more questions, you know, that we want to have answered.

当然，人们都能看到对更多计算资源的需求，但OpenAI在计算资源上的投入规模和范围实在惊人。

And certainly people can see the need for more compute but the scale and scope at which OpenAI is investing in compute is incredibly huge.

你知道吗，我们谈论的这些数字真的很难想象。

You know, we're talking, you know, numbers that are just really hard to fathom.

OpenAI是如何确定这种需求的？

How does OpenAI determine that need?

你们会看哪些指标来判断，是的，我们需要花这么多钱？

You know, what are the metrics you're looking at to think that like, yes, we need to spend this much?

首先，我们努力确保在算力上的投入与我们的收入增速保持一致。

So first of all, we are trying to make sure we stay investing in compute to match the pace of our revenue.

我们发现，当期算力与当期收入之间存在很强的相关性。

And we've seen a really strong correlation between in period compute, in period revenue.

我举个例子。

I'll give you an example.

如果你回看2023年、2024年和2025年，我们的算力从200兆瓦增长到去年底的2吉瓦。

If you just go back in '23, '24 and '25, our compute was 200 megawatts, 600 megawatts when it ended last year at two gigawatts.

与此对应的是，数字非常吻合，我们在2023年底的年经常性收入（ARR）为20亿美元。

Against that, and it's really easy because the numbers match up, we exited '23 at 2,000,000,000 in ARR.

所以是200兆瓦，20亿美元。

So 200 megawatts, 2,000,000,000.

我们在2024年结束时达到了60亿美元。

We exited '24 at 6,000,000,000.

所以是60亿美元，600兆瓦。

So 6,000,000,000, 600 megawatts.

去年我们结束时营收略高于200亿美元，200亿美元，对应两吉瓦。

And we exited last year a little over 20,000,000,000, 20,000,000,000, two gigawatts.

实际上，增长速度正在加快。

Actually, it's been accelerating.

所以即使你看这条线的斜率，也能看出更多的算力带来更多的收入。

So that's just even if you look at the slope of the line, it says more compute, more revenue.

当然，这里存在时间上的不匹配，因为我今天就必须做出决策，确保我们不仅为2026年或2027年，而是为2028、2029、2030年准备好算力。

Now there is definitely a timing mismatch because I have to make decisions today about making sure we have compute in not even '26 or '27, but '28, 2930.

因为如果今天我不下订单，不发出建设数据中心的信号，到时候就来不及了。

Because if I don't put in orders today and don't give the signal to create data centers, it won't be there.

今天，我们在算力方面感到绝对受限。

Today we feel absolutely constrained on compute.

如果我们今天拥有更多算力，我们可以推出更多产品，训练更多模型，探索更多多模态应用。

There are many more products that we could launch, many more models that we would train, many more multimodality things we would explore if we had more compute today.

所以，例如在去年，全球整体硬件投资增长了大约2200亿美元，这只是实际支出的增长幅度。

So for example, even in the last year, I think the overall hardware investments globally has gone up by something like $220,000,000,000 That's just how much actual spending has gone up.

如果你看芯片领域，芯片预测增长也类似，达到了约3340亿美元。

If you look at chips, chip forecasts have gone up similarly about $334,000,000,000.

这不仅仅是OpenAI的问题。

So it's not just OpenAI.

整个环境发出的信号是：人工智能是真实的。

The signal from the whole environment is AI is real.

我们正处于一场范式转变之中。

We are in a paradigm shift.

我们需要投资，以提供人们实现我们刚才所谈所有事情所需的智能。

We need to invest to give people the intelligence they need to do all the things we just talked about, for example.

因此，在OpenAI内部，我们花了很多时间深入研究消费者、企业及开发者层面的需求信号。

So back inside of OpenAI, we do spend a lot of time going very deep on what is our demand signal in consumer and enterprise and developers.

我们首先从基础设施层出发，思考如何构建一个最基础的拼图，以实现最大的灵活性。

We think about what's the mosaic first at the base, like on an infrastructure layer, how do we create max optionality?

因此，我们希望实现多云、多芯片架构，这为我们提供了基础设施层面的有趣优势。

So we want to be multi cloud, multi chip, and that gives us an interesting layer at the infrastructure layer.

再往上一层，是产品层面。

One tick up at the product layer.

我们还希望变得更加多维。

We also want to become more multidimensional.

过去我们只有一个产品：ChatGPT。

So we used to just be one product chat GPT.

如今，我们为消费者提供了集成了各种功能的ChatGPT，涵盖医疗健康等领域。

Today we are chat GPT for consumer with all of the blades inside it, health care and so on.

面向工作的ChatGPT。

Chat GPT for work.

但我们还有一个新的平台Sora。

But we also have Sora as a new platform.

我们还有一些变革性的研究项目。

We have some of our transformational research projects.

再往上一层。

One tick up.

我们还有一个正在变得越来越多元化的商业模式生态系统。

We also then have a business model ecosystem that's becoming much more multidimensional.

最初我们只有一个订阅服务，因为我们推出了ChatGPT，需要一种方式来支付计算成本。

Began with a single subscription because we'd launched ChatGPT and we needed a way to pay for the compute.

我们现在有多个价格

We now have multiple price

顺便说一下，ChatGPT的订阅用户。

ChatGPT subscriber, by the way.

我

I

谢谢你这么做。

love you for that.

订阅服务。

Subscriptions.

我们进入了企业市场，采用了基于SaaS的定价模式。

We went to the enterprise and had SaaS based pricing.

现在，对于那些高价值场景，我们采用了基于积分的定价方式。

We have credit based pricing now for places where the high value is being, found.

人们愿意多付钱来获得更多的服务。

People want to pay more to get more.

我们开始考虑诸如电商和广告之类的事情。

We're beginning to think about things like commerce and ads.

当然，从长远来看，我喜欢像许可模式这样的方式——比如在药物研发中，如果我们授权我们的技术，而你取得了突破，药物大获成功，我们就能从其全部销售额中获得一部分许可收益。

And then of course, longer term, I like models like, for example, would we do, licensing models to really align let's say in drug discovery, if we licensed our technology, you have a breakthrough, that drug takes off and we get a licensed portion of all its sales.

这对我们和客户来说都是绝佳的协同方式。

It's great alignment for us with our customer.

所以，如果你思考一下这三个层级，我实际上把它想象成一个魔方。

So kind of if you think about those three tiers, I actually think of it like a Rubik's cube.

好的。

Okay.

我们从单一的模块——比如一个云服务提供商、一个芯片、一个产品、一种商业模式——发展到了现在这个三维的立方体。

So we went from a single block, you know, one CSP, Microsoft, one chip, one product, one business model to now a whole three-dimensional cube.

我喜欢魔方的一点是，虽然我不确定数字是否完全准确，但我记得它有大约430亿亿种不同的状态。

And one of the things I love about a Rubik's cube, I'm probably not getting the number exactly right, but I think it has 43 quintillion different states it can be in.

当我

It always blew my mind when I

在

was in

现在就想象一下这个立方体在旋转。

so now just think about that cube spinning.

因此，当我们选择一个低延迟芯片，配合编码速度达到人们预期五倍的功能时，我们可以为此收取高端订阅费用。

So we pick a low latency chip going alongside something like coding that's five x the pace that people expect, we can charge a high end subscription for that.

所以这就像你把魔方对齐，让一面呈现出三种颜色。

So it's almost like you line up the cube and you get three colors on one side.

我们可以再次转动魔方，比如使用低延迟芯片、更快的图像生成，吸引更多免费用户，但这会为最终可能的广告平台创造更多库存。

We could spin the cube again and say low latency chip, faster image gen, more free users come in, but that creates more inventory for ultimately perhaps an ads platform.

因此，你可以看到过去十二个月的目标一直是创造越来越多的战略选择，让我能够持续支付所需的计算资源，以真正实现造福人类的AGI使命。

So you can start to see how the goal in the last twelve months has been creating more and more strategic options that allow me to keep paying for the compute we need to really achieve our mission, AGI for the benefit of humanity.

所以，简化来说，需求的限制并非来自其他任何因素，而是取决于计算资源的可用性，无论是Sora还是更广泛的意义上。

So, you know, the way to simplify that is demand is limited not by anything other than availability of compute Whether it's Sora or more broadly.

然后是价格弹性，计算资源的需求是无限的。

And then there's price elasticity where demand is infinite for compute.

所以我认为这就是看待这个问题的方式。

So I think that's the way to think about it.

是的。

Yeah.

我们甚至还没有开始利用价格弹性的杠杆。

It just, we haven't even started to exercise the price elasticity lever.

我们根本无法满足需求。

It's just we can't fulfill demand.

没错。

Right.

这受到算力的限制。

And it's limited by compute.

所以所有谈论泡沫之类的人，我认为都走错了方向。

So all the people talking about bubbles and things, I think are on the wrong track.

他们完全不了解这一变革的规模，以及对API调用需求弹性的巨大需求。

They have no sense of how large this changes and how much more demand elasticity there's a need for API calls.

作为OpenEye最早的投资者之一，你早早地押注了，你看到了这个方向的发展，但你也经历过互联网泡沫。

As one of OpenEye's earliest investors, you made a bet early on, you saw where this was headed, but you've saw the .com bubble.

你目睹了当时发生的一切，但也见证了移动革命。

You watched what happened there, but you've also seen other things, the mobile revolution.

你已经看到其他领域也发生过类似的情况。

You've seen this happen with other areas.

你提到了‘广泛’这个词，这是否就是你信心的来源？

And you mentioned the term broad and is that sort of where your conviction comes from?

只是因为它涉及了这么多不同的领域吗？

It's just how many different areas it touches?

是的。

Yeah.

当我们投资时，我们只有一个简单的衡量标准。

Look, when we invested, we had one simple metric.

当时没有可参考的预测，没有产品计划，也没有ChatGPT可以参考。

There was no projections to look at, no product plans to look at, no chat GPT to look at.

很简单，如果我们能开发出接近甚至超越人类智能的东西，其影响将是巨大的。

It was very simply the idea, if we develop anywhere near close to human intelligence, let alone supersede human intelligence, its impact is going to be huge.

所以我们采取了一种‘婴儿学步’的方式：如果成功，后果将极其深远，那为什么不试试呢？

So it was this hand baby approach, like the consequences of success are really going to be consequential, so why not try that?

还有一种有趣的说法叫‘泡沫’。

There's also this funny notion of bubble.

人们把泡沫等同于股价，而这与投资者的恐惧和贪婪毫无关系。

People equate bubble to stock prices, which has nothing to do with anything other than fear and greed among investors.

所以我总是认为，泡沫应该通过API调用次数来衡量。

So I always look at bubbles should be measured by the number of API calls.

或者在人们所提及的互联网泡沫中，应该看的是互联网流量，而不是股价的涨跌，因为有人过度兴奋或过度悲观。

Or in the .com bubble, which people refer to, it should be amount of internet traffic, not by what happened to stock prices because somebody got overexcited or underexcited.

而在一天之内，他们可能从热爱英伟达突然转为憎恨英伟达，只因为觉得它被高估了。

And in one day they can go from loving Nvidia to hating NVIDIA because it's overvalued.

这些剧烈波动并不是现实。

Those gyrations aren't reality.

真正的现实是底层的API调用次数。

The reality is the underlying number of API calls.

如果你查看互联网泡沫时期的互联网流量，股价可能剧烈上涨又剧烈下跌。

If you look at internet traffic during the .com bubble, prices may have gone up violently and gone down violently.

但在互联网流量中，并未发现泡沫的迹象。

There's no bubble detected in internet traffic.

我几乎可以肯定，你在API调用数量上看不到泡沫。

I would almost guarantee you, you won't see the bubble in number of API calls.

如果你把API调用作为衡量AI真正用途、AI价值和AI需求的基本指标，那么你在API调用上根本看不到泡沫。

And if that's your fundamental metric of what's the real use of your AI, usefulness of AI, demand for AI, you're not going to see a bubble in API calls.

华尔街通常怎么对待它，我其实并不关心。

What Wall Street tends to do with it, I don't really care.

我觉得这基本无关紧要。

I think it's mostly irrelevant.

这对新闻报道来说很有用，因为媒体需要填满版面，但这并不是现实。

Great for press articles because press has to fill their column inches, but it's not reality.

所以事物的价格，或者股票价格、私营公司估值，都不是现实。

So prices of things aren't reality, or stock prices, private company valuations.

现实是AI的实际需求，也就是API调用的数量。

The reality is what's the actual demand for AI, which is the number of API calls.

对。

Right.

如果我回想起1999年那个时刻，当时人们从互联网中获得的价值其实非常有限，因为互联网还太年轻、太初期，你根本看不出它如何真正改变人们的生活。

And I think if I hark back to that moment where you were looking at 1999, the value people were getting from the Internet at the time was actually very it was so young, so nascent that you couldn't really see how it was changing their lives.

我认为，对于人工智能来说，这种变化发生得实在太快了。

I do think that with AI, it's happened so fast.

是的。

Yeah.

这种变化是真实存在的。

That change, it's very real.

比如，作为一个CFO，别说是OpenAI的CFO，单就作为一个CFO而言，我在我的组织里看到的变化，确实是把那些过去我不得不不断增派人手来处理的、相当枯燥的任务取代了。

Like, as a CFO, forget about being the CFO of OpenAI, but as a CFO, what I see happening in my organization is truly taking tasks that previously I would have kept having to add more and more people doing fairly mundane things.

比如，拿收入管理来说。

Like, let's take something like revenue management.

在负责收入管理的团队中，他们每天的一项工作就是下载前一天或本周内签署的所有合同，并逐份阅读这些合同，确保其中没有隐藏任何意外的、非标准的条款。

So in in a team that does revenue management, they have one of the things they do every day is they have to download all the contracts that we signed the day before or through the week, and they have to read all of those contracts to make sure there's no terms sitting in it that are unexpected, that are effectively nonstandard terms.

因为非标准条款意味着可能需要调整收入确认方式。

Because a nonstandard term means that there could be a revenue recognition change that has to happen.

这对财务团队来说是非常重要的事情。

And that's a very big deal for a finance team.

这通常是审计师来审计你的首要事项。

That's the number one thing usually your auditors come in to audit you on.

我们增长的速度，没错，每天的合同数量正在成倍增加。

The pace at which we are growing, right, that number of contracts every day is going up in multiples.

所以在没有AI的时代，我唯一的选择就是雇更多人。

So my only choice in a pre AI world would have been hire more people.

想象一下那些人的工作是什么样子。

Imagine what those people's jobs are like.

你每天上班，读一份合同，然后读下一份，再下一份。

You come to work every day and you read a contract and then you read the next one and the next one.

这太枯燥乏味了。

It is so mundane and such drudgery.

这并不是人们上学学习会计或立志成为财务专业人士的原因，但我们却把这种工作当作入门级职位交给他们。

It's not why people went to school and learned about the accounting field or thought about being a finance professional, but that's kind of the job we hand them as an entry level job.

今天，借助我们在OpenAI内部开发的工具，我可以在一夜之间将所有这些合同从系统中提取出来。

Today, using our own tools here at OpenAI, I now have overnight, all of those contracts are pulled out of a system.

它们被导入到一个表格数据库中，对我们来说就是Databricks数据库。

They are put into a tabular database, the Databricks database in our case.

这个智能代理能够逐一分析这些合同。

The the agent or the intelligence is able to go through.

它能准确告诉我哪些条款是非标准的，以及原因是什么。

It shows me exactly what is nonstandard and why.

它不仅会建议相应的收入确认方式，还会提供深入洞察，比如：这个条款真的有必要存在吗？

It suggests what therefore the rev rec is, but it also suggests the insight, which is, you know, should this term even be here?

销售代表是不是擅自让步了本不该让步的内容？

Did the salesperson just give away something they shouldn't have?

如果是这样，我就会去指导他们。

In which case, you know, I go and I coach them.

它是否实际上在向我揭示我们业务中正在发生的变化？

Is it actually telling me something about my business that's starting to shift?

在这种情况下，这个非标准条款实际上应该成为标准条款。

In which case, this nonstandard term is actually should become a standard term.

而我实际感受到的是我的商业模式正在发生变化，这可能是一件好事。

And I'm actually what I'm experiencing is a shift in my business model, which might actually be a good thing.

或者，也许我想找到一种不同的方式，既帮助客户获得他们想要的，也帮助销售人员实现他们的目标，同时保持我的收入确认和现有商业模式。

Or perhaps I wanna find a different way to help get the customer what they're looking for, the salesperson what they're looking for, but maintain my revenue recognition, my current business model.

对吧？

Right?

我知道我那些资历较浅的初级人员在讨论的右侧，他们正在重新找到自己热爱的工作。

So I know my more junior entry level people are over on the right of that discussion, and they're kind of refinding the job they loved.

对我来说，这正是它不是一个泡沫的原因，因为价值是真实且切实的。

That to me is why it's not a bubble because the value is real and tangible.

这意味着我可能只需要一个更小的团队。

I get also means I probably can have a smaller team.

我可以拥有一个绩效更高、士气更旺盛、留存率更高的团队。

I can have a much more high performing team, a much higher morale on my team, better retention rates.

对吧？

Right?

所有这些我都可以量化，说明我的业务现在更健康了。

All of these I can put into like numbers to say my business is now healthier.

我认为，当媒体试图主导关于泡沫的讨论时，他们忽略了这一点。

I think that's the piece when the press is trying to lead with the bubble conversation or whatever.

他们只是没注意到，我们目前的投资是基于需求的，甚至可以说是跟随需求。

They just miss that we are investing with demand, if anything behind demand at the moment.

对我来说，泡沫意味着你是在需求之前投资，最终会出现缺口。

A bubble to me suggests you're investing ahead of demand and there's going to be a gap.

如果你看一下生产率数据，那些正在采用人工智能的公司，尤其是新兴的科技公司，生产率正在上升。

And you look at productivity numbers, they're going up in the companies that are adapting AI, especially the newer set of tech oriented companies.

这些数据简直惊人。

The numbers are just absolutely amazing.

我最喜欢的一个例子是一家叫Slash的小公司，年经常性收入约为1.5亿美元。

So one of my favorites is a little company called Slash, about 150,000,000 ARR.

他们只有一个会计人员，仅有一位财务主管，因为他们采用了人工智能导向的ERP系统。

They have one person in accounting, only a controller, because they adapted an AI oriented ERP system.

他们用这个系统替换了NetSuite。

They replaced NetSuite with it.

他们能做到的事情简直令人惊叹。

And it's just amazing what they can do.

公司的首席执行官向我道歉，说他可能不得不雇用第二个人。

And the CEO was apologizing to me he might have to hire a second person.

而且他们的进展非常迅速。

And they're moving really rapidly.

我刚看到一个故事，有人用一个SDR加上AI取代了十个SDR，剩下的那名SDR主要负责监督AI的工作。

I just saw a story, somebody replaced 10 SDRs with one SDR and AI essentially that the one SDR remaining supervises.

我听到了两个故事，说的是公司不再招聘那些不直接推动增长的岗位人员，现在招聘时，他们更倾向于雇佣那些曾为公司创造过显著增长的人。

I've been hearing two stories about where instead of hiring somebody that's in an area that doesn't create growth, they can now then when they hire, hire people that have created a lot more growth for the company.

这就是为什么你看到这么多科技公司能够如此迅速地扩张。

And that's why you're seeing a lot of these tech companies just build so fast.

你知道，那句老话，未来已经到来，只是分布不均。

You know, that old phrase, the future is here now, but it's not evenly distributed.

是的。

Yes.

我看到所有这些单一环节带来了巨大的生产力提升、效率提升或敏捷性提升，能够更快地行动。

I see all these single points of huge productivity gains and efficiency gains or agility gains, the ability to move faster.

但世界上，无论是美国还是全球，只有极少数人采用了这些技术，甚至都不知道它们的存在。

But very small percentage of the people in the world, in The US or worldwide, have adapted these or even know they exist.

对。

Right.

所以回到需求问题，我认为这些理念，这些例子，最终会随着时间推移普及到每个人。

And so this issue back to demand, I think these ideas, some of these examples will spread to everybody over time.

你会看到这些技术的采用呈指数级增长。

And you'll see an exponential growth of adoption of these technologies.

这就是为什么我不认为需求是个问题。

That's why I don't think demand is the question.

是的，这个笔记完全切中要点。

Yeah, the note is absolutely spot on.

我认为麦肯锡做过一项研究，显示那些处于顶尖四分位的公司，其生产力——无论用哪种财务指标衡量——都提升了百分之二十七到三十三。

I think McKinsey did a study that showed for companies that are more in the top quartile, their productivity as measured by any kind of financial metric you would pull is up you know, in the twenty seven to thirty three percent.

这可是个非常显著的提升。

Like, that's a really meaningful jump.

我想你想要表达的是，这并不意味着整体员工数量减少。

I think where you were going is it doesn't just mean fewer employees overall.

确实存在将人员转向更具增长性岗位的空间。

There's definitely a place to kind of shift people over into more growth oriented jobs.

这个周末我徒步时遇到一个人，他经营着一家你们都熟知的大型咨询公司。

I was hiking this weekend with someone who runs a very large consulting company that you all would know of.

他谈到自己公司，或者说他更关注的后台系统。

And he was talking about how his and his what he thinks of more his back end systems.

那里的负责人现在把她的团队描述为‘人加智能代理’。

The leader there is now talking about her organization as people plus agents.

她的人机比例是1比5，即一个人对应五个智能代理。

And she has a one to five ratio, one person to five agents.

但在前端，他们实际上正在重新招聘以扩大规模，因为客户现在需要更多帮助来思考如何部署人工智能。

But on the front end, they're actually back out rehiring to grow because clients need more help now to think about deploying AI.

所以，我认为这实际上是在回归到人们真正想做的工作，而不是那些仅仅因为世界信息量过大、人们不得不处理这些信息而被迫接受的工作。

So it's actually shifting back, I would say, to the jobs people want to do, not the jobs that maybe were just open to them because more and more of the world had become this kind of, you know, so much information that people were parsing it.

现在，我们终于回到了机器和智能代理来处理这些信息的时代。

Now we're finally back to a machine and agent intelligence parsing it.

我想再回到消费者层面谈谈。

I want to I want to touch back on the consumer side.

你提到了广告，而我们之前提出的观点是，通过广告可以增加对用户的益处。

You mentioned ads and certainly the argument we made that with ads you can increase the benefits to people.

你可以提供更多的服务、更多的AI功能，帮助支付计算成本，用户在这些高级套餐中也能获得更多的价值。

You can provide more services, more AI, you can help pay for the compute and people get more out of those tiers with that.

但这引出了一个信任问题：当人们最初想到人工智能并提出问题时，他们会担心聊天机器人会如何处理他们的信息。

But that brings up the question though of trust and when people think about AI initially even asking questions, people worried about what does chat should be to do with my information.

一旦引入广告，人们就会担心，因为这通常会引发一个大问题：广告如何影响产品的其他部分和整个组织？

Once you have ads in play, people worry about that because it's often just a big question of how does that affect the rest of the product and the org?

是的。

Yeah.

所以我认为你从正确的角度切入了：如今，我们95%的用户在消费者端免费使用我们的平台。

So I think you started in the right place, which is today 95% of our users use our platform for free on the consumer side.

而这正是我们的使命所在。

And that's absolutely where our mission is.

对吧？

Right?

AJI的目标是为全人类服务，而不是只为付得起钱的人服务。

AJI, for the benefit of humanity, not the benefit of humanity who can pay.

对吧？

Right?

因此，可及性非常重要。

So access is very important.

从广告的角度来看，我认为最重要的是，我们必须确保每个人都明白，你始终会得到模型能提供的最佳答案，而不是付费得到的答案。

From an ads perspective, I think number one, we have to just make sure everyone understands you're always going to get the best answer the model can provide you, not the paid for answer.

我认为其他平台在这方面已经倒退了，你无法确定这个链接是赞助的，还是真正最好的结果。

And I think other platforms have fallen back into that where you're not sure, this is a sponsored link or is this truly the best outcome.

我们有一个北极星目标，那就是模型始终会给你最佳答案。

We have a North Star, which is that the model will always give you the best answer.

我认为第二点需要理解的是，广告其实可以带来很多价值。

I think the second thing to understand is that there can be a lot of utility in ads.

所以我们希望确保人们清楚地知道什么时候他们正在与广告互动。

So we want to make sure people know when it is an ad that they're working with.

但例如，如果我搜索一个周末度假目的地，比如我最喜欢的圣地亚哥，Airbnb的广告实际上可能会非常有帮助。

But for example, if I do a search for a weekend getaway to pick your favorite city, I don't know, San Diego, an ad for Airbnb might actually be very helpful.

在这种情况下，你甚至可能想在ChatGPT这样的丰富对话环境中与广告或广告主进行交流，但你必须清楚这属于广告场景。

And you might even want to have a discussion with the ad or with the advertiser in that case in a chat GPT setting that's very rich, but you're clear that it's in an advertising setting.

我认为，这正是需要更多创新的地方——要创造出真正契合平台特性的广告形式，而不是简单照搬传统的横幅广告等旧模式。

And I think this is where there's there has to be more innovation on what feels endemic to the platform, not just kind of the old world of stick, you know, banner ads on things.