Let's just say that the total sort of compute capacity of all GPUs in the world today is X.

比方說，當今世界上所有 GPU 的總計算能力排序為 X。

What do you think, what do you think, what multiple of X will we be at in five years?

你覺得，你覺得，五年後我們的 X 倍率會是多少？

First of all, you know that I'm going to regret saying this.

首先，你知道我說了這些話會後悔的。

And this is, I'm a public company, you crazy person.

我是一家上市公司，你這個瘋子。

See this is, here's, this is, how nice is it to be private? Safe to say considerably more?

你看，這就是，這就是，私密性有多好？ 可以說是相當不錯吧？

Well let's reason about it, shall we?

讓我們來推理一下，好嗎？

Okay so let's reason about it, let's reason our way through, okay?

好吧，讓我們來推理一下，讓我們推理一下，好嗎？

So first of all, it goes like this.

首先，事情是這樣的。

The world has installed about a trillion dollars worth of data centers.

全球已安裝了價值約一萬億美元的數據中心。

Those trillion dollars worth of data centers use as general purpose computing.

這些價值萬億美元的數據中心用作通用計算。

General purpose computing is run as course.

通用計算按課程運行。

We cannot continue to process that way.

我們不能繼續這樣處理。

And so the world is going to accelerate everything.

是以，世界將加速一切。

Data processing, you name it, okay?

數據處理，你說了算，好嗎？

And so we're going to accelerate everything.

是以，我們將加快一切進程。

When we accelerate everything, every single data center, every single computer will be an accelerated server.

當我們加速一切時，每一個數據中心、每一臺計算機都將成為加速服務器。

Well there's about a trillion dollars worth of computers if we don't grow at all over the next, call it four years that we have to go replace.

如果我們在接下來的四年裡不增長，那麼就有價值一萬億美元的電腦需要更換。

Four years, six years, pick your number of years.

四年、六年，隨便你選。

But if the computer industry continues to grow at some 20% or so, we'll probably have to replace over the course of next, you know, pick your number of years, about two trillion dollars worth of computers with accelerated computing.

但是，如果計算機行業繼續以 20% 左右的速度增長，那麼在接下來的幾年裡，我們可能需要更換價值約兩萬億美元的計算機，並採用加速計算技術。

So just make that GPUs, okay?

所以就用 GPU，好嗎？

That's number one.

這是第一位的。

And this is the second part, this is the reason why all of you, Stripe, you're onto something just absolutely monumental.

這是第二部分，這也是為什麼 Stripe 所有的人都在做一件絕對不朽的事情。

This idea, this idea called, you know, you've heard me say an industrial revolution.

這個想法，這個叫做 "工業革命 "的想法，你知道，你聽我說過。

Let me tell you why.

我來告訴你為什麼。

We are producing something for the very first time that has never been produced before.

我們將首次製作從未製作過的作品。

And we're producing it in extremely high volume.

我們的產量非常高。

And the production of this thing requires a new instrument that never existed before.

而生產這種東西需要一種前所未有的新工具。

It's a GPU.

這是 GPU。

And the thing that we're producing for the very first time, for the mathematicians and all the computer scientists in the room, for all of you, you know that we're producing tokens.

我們第一次製作的東西，對於在座的數學家和計算機科學家來說，對於你們所有人來說，你們知道我們正在製作代幣。

We're producing floating point numbers at high volume for the first time in history.

我們有史以來第一次大批量生產浮點數。

And the floating point numbers have value.

而浮點數是有價值的。

The reason why they have value is because it's intelligence.

它們之所以有價值，是因為它是智慧。

It's artificial intelligence.

這是人工智能。

You can take these floating point numbers, you reformulate it in such a way that it turns into English, French, proteins, chemicals, graphics, images, videos, robotic articulation, steering wheel articulation.

你可以把這些浮點數重新組合，變成英語、法語、蛋白質、化學物質、圖形、影像、視頻、機器人銜接、方向盤銜接。

We're producing tokens at extraordinary scale.

我們正在以超常規的規模生產代幣。

Now we've discovered a way through all of the work that we do with artificial intelligence to produce tokens of almost any kind.

現在，我們發現了一種方法，通過人工智能的所有工作，幾乎可以生產出任何種類的代幣。

So now the world is going to produce an enormous amount of tokens.

是以，現在全世界將生產大量代幣。

Now these tokens are going to be produced in new types of data centers.

現在，這些代幣將在新型數據中心生產。

We call them AI factories.

我們稱之為人工智能工廠。

Back in the last industrial revolution, water comes into a machine.

在上一次工業革命中，水進入了一臺機器。

You light the water on fire, turn it into steam, and then it turns into electrons.

把水點燃，變成蒸汽，然後變成電子。

Atoms come in, electrons go out.

原子進來，電子出去。

In this new industrial revolution, electrons come in and floating point numbers come out.

在這場新的工業革命中，電子進來了，浮點數出來了。

And just like the last industrial revolution, nobody understood why this electricity is so valuable and is now sold, marketed, kilowatt hours per dollar.

就像上一次工業革命一樣，沒有人明白為什麼這些電力如此寶貴，而且現在還以每美元千瓦時的價格出售、銷售。

And so now we have a million tokens per dollar.

是以，我們現在每美元有一百萬代幣。

And so that same logic is as incomprehensible to a lot of people as the last industrial revolution, but it's going to be completely normal in the next 10 years.

是以，對很多人來說，這種邏輯就像上一次工業革命一樣難以理解，但在未來 10 年內，它將變得完全正常。

Well, these tokens are going to create new products, new services, enhance productivity on a whole slew of industries, $100 trillion worth of industries on top of us.

這些代幣將創造新的產品、新的服務，提高各行各業的生產力，其中包括價值 100 萬億美元的行業。

And so this industry is going to be gigantic.

是以，這個產業將是一個巨大的產業。

And in order to monetize that, transact that, you're going to need Stripe.

為了實現貨幣化和交易，你需要 Stripe。

And so...

於是......

I got to tell you, this is one of my favorite companies.

我得告訴你，這是我最喜歡的公司之一。

The first time I met Patrick, he had to explain Stripe to me.

我第一次見到帕特里克時，他不得不向我解釋 Stripe。

I was...

我...

First of all, it was so complicated.

首先，它太複雜了。

Because it's a complicated...

因為這是一個複雜的...

We tried to refine the descriptions over time, but you had an early version.

隨著時間的推移，我們試圖完善這些描述，但你們有一個早期版本。

No, no.

不，不

First of all, you're in a complicated business no matter what.

首先，無論如何，你從事的都是一項複雜的工作。

But nonetheless, I was so inspired by it.

儘管如此，我還是深受啟發。

It's incredible what you guys have built.

你們的成果令人難以置信。

Are we going to get you migrated to Stripe billing now that we have usage-based billing?

既然我們有了基於使用量的計費方式，我們是否會讓你們遷移到 Stripe 計費方式？

I wish I had a business that required billing.

我真希望我的生意也需要計費。

I think your public filing suggests you're doing a lot of billing.

我認為你的公開申報表明你在做大量的賬單。

We'll follow up on it.

我們會繼續跟進。

All right, so...

好吧，那麼...

It's only 10 transactions, just so you know.

只有 10 筆交易，只是讓你知道一下。

Your economics serving us is like nothing.

你們為我們提供的經濟服務真是無與倫比。

It's like 10 transactions.

好像是 10 筆交易。

We'd happily take the 2.9%, but...

我們很樂意接受 2.9%，但...

We can discuss that separately.

我們可以另行討論。

So...

所以...

Done.

完成。

Thinking about this token...

想想這塊令牌...

You can't say that.

你不能這麼說。

You're a public company.

你是一家上市公司。

So thinking about these token factories, I feel like a big question right now is whether the models saturate.

是以，考慮到這些代幣工廠，我覺得現在的一個大問題是模型是否飽和。

In the sense that we demoed the Sigma Assistant on stage earlier, and you can write some natural language, and we convert that to SQL.

我們之前在臺上演示了西格瑪助手，你可以編寫一些自然語言，然後我們將其轉換為 SQL。

And going from a 7 billion parameter model to a 70 billion parameter model or something like that, there might be a significant kind of consequential improvement in query accuracy for the user for the typical kinds of queries that people tend to construct.

從 70 億個參數模型到 700 億個參數模型或類似的模型，對於人們傾向於構建的典型查詢類型，用戶的查詢準確性可能會有顯著的提高。

But maybe going to a model that's 10x larger than that is sort of unnecessary.

但是，也許使用比它大 10 倍的機型是沒有必要的。

At some point, you get to good enough, you can reliably convert the natural language to SQL.

到了一定程度，你就可以將自然語言可靠地轉換成 SQL。

I think there's a question for the use cases for which LLMs are being deployed.

我認為這與部署 LLM 的用例有關。

What does that saturation curve look like?

飽和度曲線是什麼樣的？

And for how many use cases does one need a trillion parameter model or a 10 trillion parameter model?

有多少用例需要萬億參數模型或 10 萬億參數模型？

Or do we simply reach a point where some number that is, say, less than 100 billion is sufficient?

還是說，我們只需達到一個點，比如說某個小於 1000 億的數字就足夠了？

Do you have any point of view on that?

你對此有什麼看法嗎？

Or is that even a reasonable way to look at the question in the first place?

或者說，從一開始就這樣看待這個問題是否合理？

Okay.

好的

Let's break it down.

讓我們來分析一下。

Let's reason about it.

讓我們來推理一下。

In public.

在公共場合

Almost everything.

幾乎無所不能

Every question I get.

我收到的每個問題

Okay.

好的

Let's break it down.

讓我們來分析一下。

Let's reason about it.

讓我們來推理一下。

So let's start with an example.

讓我們從一個例子開始。

In 2012, AlexNet was computer vision, ImageNet, image recognition, 82% or something like that Over the next almost not quite 10 years, I think it was like seven years, every single year, the accuracy error reduced in half, right?

2012 年，AlexNet 是計算機視覺，ImageNet 是圖像識別，準確率為 82%，差不多是這個水準 在接下來的差不多 10 年裡，我想大概是 7 年，每一年的準確率誤差都減少了一半，對嗎？

Every year, the error reduced in half or otherwise known as Moore's law. So you double the performance, you double the accuracy, and you double its believability every single year.

每一年，誤差都會減半，也就是所謂的摩爾定律。 是以，性能每年翻一番，準確性每年翻一番，可信度每年翻一番。

Over the course of seven years, it's now superhuman.

在七年的時間裡，它現在已經是超人了。

Same thing with speech recognition.

語音識別也是如此。

Same things with natural language understanding.

自然語言理解也是如此。

We want to know.

我們想知道

We want to believe.

我們願意相信。

Not know.

不知道。

We want to believe that the answer that's being predicted to us is accurate.

我們希望相信預測給我們的答案是準確的。

We want to believe that.

我們願意相信這一點。

And so the industry is going to chase that believability or that accuracy and double its accuracy two X every year.

是以，該行業將追逐可信度或準確度，並每年將準確度提高兩倍。

I believe that's going to be the same thing with natural language understanding.

我相信，自然語言理解也會是同樣的情況。

And of course, the problem space is a lot more complicated.

當然，問題空間要複雜得多。

But I have every certainty that we're going to double its accuracy every single year to the point where it is so accurate.

但我堅信，我們的精確度會每年翻一番，達到非常精確的程度。

And we've largely tested across many of your examples when you interact with it that you go, you know what?

當你與它互動時，我們已經對許多例子進行了測試，你會發現，你知道嗎？

This is really, really good.

這真的非常非常好。

I believe the answer that it's producing for me.

我相信它給我的答案。

That condition is very important.

這個條件非常重要。

The second thing is this.

第二件事是這樣的。

Today's language models, today's AI and everything that we've shown are one shot.

今天的語言模型、今天的人工智能以及我們所展示的一切，都是一次性的。

And yet, you and I both know that there are many things that we think about that are not one shot.

然而，你我都知道，我們思考的很多事情都不是一蹴而就的。

You have to iterate.

你必須反覆推敲。

And so how do you come up?

那你是怎麼上來的？

How do you reason about a plan?

如何推理計劃？

How do you come up with a strategy to solve a problem?

如何制定解決問題的策略？

Maybe you need to use tools.

也許你需要使用工具。

Maybe you have to look up some proprietary data.

也許你需要查閱一些專有數據。

Maybe you have to do some research, in fact.

事實上，也許你需要做一些研究。

Maybe you have to ask another agent.

也許你得問問其他經紀人。

Maybe ask another AI.

也許可以問問其他人工智能。

Maybe you have to be human in a loop.

也許你必須在循環中成為人類。

Ask a human.

問問人類。

Trigger an event.

觸發事件。

Send an email to somebody or text to somebody.

給某人發送電子郵件或簡訊。

Get a response before you can move on to the next step of that plan.

得到答覆後，才能進入計劃的下一步。

And so a large language model has to iterate and think of a plan.

是以，一個大型語言模型必須不斷迭代，並制定計劃。

That's not a one shot thing.

這不是一朝一夕的事。

And once it comes up with a plan, as it traverses that graph, there's a whole bunch of language models that are going to get instantiated and initiated.

一旦它提出一個計劃，在遍歷該圖的過程中，就會有一大堆語言模型被實例化並啟動。

And so I think your future models are going to iterate.

是以，我認為你們未來的模式會不斷迭代。

And so instead of just a one shot model, it's going to be a planning model with a whole bunch of other models around it that get particular skills.

是以，它將不再是一個單一的模型，而是一個規劃模型，周圍還有一大堆其他模型，這些模型都會獲得特定的技能。

And so I think we have long ways to go.

是以，我認為我們還有很長的路要走。

And meta garnered a lot of attention last week for the release of Llama 3, which seems to be the most impressive open source model thus far.

Meta 上週發佈的 Llama 3 引起了廣泛關注，它似乎是迄今為止最令人印象深刻的開源模型。

Any thoughts on open source models?

對開源模型有什麼想法？

If you ask me what are the top most important events in the last couple of years, I would tell you, of course, ChatGPT, reinforcement learning, human feedback, grounding it to human values and having the technology necessary to do that.

如果你問我過去幾年中最重要的事件是什麼，我會告訴你，當然是 ChatGPT、強化學習、人類反饋、以人類價值觀為基礎以及實現這些目標所需的技術。

Obviously, a breakthrough and democratized computing.

顯然，這是一項突破，也是計算的民主化。

It made it possible for everybody to be a programmer.

它讓每個人都有可能成為程序員。

Everybody's now doing amazing things with it.

現在，每個人都在用它做著令人驚歎的事情。

ChatGPT, the work that OpenAI did, Greg and Sam and the team, really proud of them.

ChatGPT、OpenAI 所做的工作、格雷格和山姆以及團隊，真的為他們感到驕傲。

The second thing that I would say that is just as important, I would say, is Llama, not Llama 1, but Llama 2.

我要說的第二件事同樣重要，那就是 "拉瑪"，不是 "拉瑪 1 號"，而是 "拉瑪 2 號"。

Llama 2 activated just about every industry to jump into working on generative AI.

Llama 2》激活了幾乎所有行業，使其紛紛投入到生成式人工智能的研究中。

And it opened the floodgates of every industry being able to access this technology.

它打開了各行各業獲取這項技術的閘門。

Healthcare, financial services, you name it, manufacturing, you name it, customer service, retail, all kinds.

醫療保健、金融服務、製造業、客戶服務、零售業，應有盡有。

I think Llama 2 and Llama 3, because it's open sourced, it engaged research, it engaged startups, engaged industry, it made generative AI accessible.

我認為，由於 Llama 2 和 Llama 3 是開源的，它吸引了研究人員、初創企業和行業的參與，使生成式人工智能變得容易獲得。

I think that's a very big deal.

我認為這是一件大事。

And so I think ChatGPT democratized computing.

是以，我認為 ChatGPT 實現了計算的民主化。

I think Llama democratized generative AI.

我認為 Llama 實現了生成式人工智能的民主化。

Does it make sense?

有道理嗎？

And I think without it, it's very hard to have activated all of the research on safety and all of the different ways of change of thoughts and all the reasoning technology that's now being developed and all the reinforcement learning stuff.

我認為，如果沒有它，就很難激活所有關於安全的研究、所有改變思想的不同方法、所有現在正在開發的推理技術以及所有強化學習的東西。

And that stuff would have been very hard to have activated without Llama.

如果沒有 Llama，這些東西很難被激活。

And Dario Amode was on Ezra Klein.

達里奧-阿莫德（Dario Amode）也上了埃茲拉-克萊因（Ezra Klein）的節目。

Oh, wow.

哦，哇

I haven't thought about this in a long time.

我已經很久沒有想過這個問題了。

I guess in a lot of ways, probably...

我想在很多方面，可能...

Okay, here's one.

好吧，這裡有一個。

When I first started NVIDIA, I was 29 years old and I was 29 years old with acne.

剛加入英偉達時，我 29 歲，29 歲的我滿臉痘痘。

And you go recruit law firms and VCs and I got a big zit on my forehead.

你去招聘律師事務所和風險投資公司，我額頭上就長了個大痘痘。

And I don't have one today, so I feel comfortable talking about it.

而我今天沒有，所以我覺得談起來很舒服。

But it could happen.

但這是有可能發生的。

And so anyways, you feel rather insecure because most of CEOs back then wore suits and they're quite accomplished and they sound like adults and they use big words and they talk about business and things like that.

總之，你會感到相當沒有安全感，因為那時的首席執行官大多西裝革履，頗有成就，聽起來像個成年人，他們會使用大詞，會談論生意和諸如此類的事情。

And so when you're young, you feel rather intimidated.

是以，當你年輕的時候，你會感到相當害怕。

You're surrounded by a bunch of adults.

你周圍都是成年人。

Well, now, if you don't have acne, I don't think you deserve to start a company.

現在，如果你沒有痤瘡，我認為你就不配開公司。

That's one big difference.

這是一個很大的區別。

Acne.

痤瘡

The takeaway from Jensen's speech.

詹森演講的啟示

What it means is really, we've enabled younger people to be extraordinary.

這實際上意味著，我們已經讓年輕人變得非凡。

I think that the young generation of CEOs, the type of things that you guys know at such a young age is really quite extraordinary.

我認為，年輕一代的首席執行官們，你們年紀輕輕就懂得這麼多東西，真的非常了不起。

I mean, it took me decades to learn it.

我是說，我花了幾十年才學會。

And so...

於是......

Last question.

最後一個問題。

And that was a compliment.

這是一種讚美。

See how he quickly changed it?

看看他是如何迅速改變的？

I wasn't saying you have acne.

我不是說你有痤瘡。

I was just saying you were smart.

我只是說你很聰明。

NVIDIA has a market cap of roughly $2 trillion.

英偉達的市值約為 2 萬億美元。

And you're now within spitting distance of Apple and Microsoft.

現在，你與蘋果和微軟的距離近在咫尺。

And I just checked and they have 220,000 and 160,000 employees respectively.

我剛剛查了一下，它們分別有 22 萬和 16 萬名員工。

NVIDIA has 28,000 employees.

英偉達擁有 28000 名員工。

So less than a fifth of the smaller of the two there.

是以，還不到這兩個小的五分之一。

And then you just said when we were chatting backstage, and I jotted this down, you can achieve operational excellence through process, but craft can only be achieved with tenure.

你剛才在後臺哈拉時說，我把這句話記下來了，你可以通過流程實現卓越營運，但工藝只能通過任期來實現。

And so NVIDIA is considerably smaller than any of the other giants.

是以，英偉達比其他巨頭都要小得多。

And you seem to think that tenure really matters.

而你似乎認為任期真的很重要。

And I guess the craft really matters.

我想，工藝真的很重要。

Do you want to say a little bit more there?

你想多說一點嗎？

I think a lot of good things could be made, good things are made with operational excellence.

我認為可以做出很多好東西，卓越的營運可以做出好東西。

But you can't make extraordinary things through just operational excellence.

但是，僅靠卓越營運是無法創造非凡業績的。

And the reason for that is because a lot of the great things in your body of work and the products that you make, the company you created, the organizations you've nurtured, it takes loving care.