It could be electricity.

可能是電。

It could be an airline ticket.

可能是一張機票。

It could be anything you choose.

你可以選擇任何東西。

We reduced it in the last 10 years by 1 million times.

我們在過去 10 年中將其減少了 100 萬倍。

Well, when something happens, when the cost of something reduces by a million times, your habits fundamentally change.

好吧，當事情發生時，當某樣東西的成本降低了一百萬倍時，你的習慣就會發生根本性的改變。

How you think about computing fundamentally changed.

你對計算的思考方式發生了根本性的改變。

That is the single greatest contribution NVIDIA ever made.

這是英偉達有史以來做出的最大貢獻。

I totally see your point.

我完全明白你的意思。

Some of our professors here may slightly disagree because you still need a lot of money to buy your GPUs.

我們在座的一些教授可能會略有異議，因為你仍然需要很多錢來購買 GPU。

But I come back to this point later.

但我稍後會再談這一點。

Imagine a million times higher.

想象一下，再高一百萬倍。

That's right.

這就對了。

But Jensen, you know.

但是詹森，你知道的

I gave you a million times discount in the last 10 years.

過去 10 年裡，我給你打了無數次折扣。

It's practically free.

這實際上是免費的。

Ha, ha, ha, ha, ha, ha, ha, ha, ha, ha, ha, ha, ha, ha, ha, ha.

哈，哈，哈，哈，哈，哈，哈，哈，哈，哈，哈，哈，哈，哈，哈，哈，哈。

Jensen, should we turn around?

詹森，我們要掉頭嗎？

Should we turn around?

我們該掉頭嗎？

Turn around, please.

請轉過身去

Oh!

哦！

I am alumni of HKUST.

我是科大校友。

Yes!

是的！

Yes!

是的！

Jensen, it's just so nice to have you here at HKUST and in Hong Kong.

詹森，很高興你能來到科大和香港。

Let me just say, can we just please show how much we love Jensen?

我只想說，我們能不能表示一下我們有多愛詹森？

Yeah!

是啊

Thank you.

謝謝。

Thank you.

謝謝。

I love you, too.

我也愛你

Jensen, you know, I have been preparing for this moment.

詹森，你知道，我一直在為這一刻做準備。

I thought it's definitely a highlight of my HKUST career, for sure.

我想這絕對是我科大生涯中的一個亮點。

I just couldn't sleep last night for one very important reason.

我昨晚睡不著，有一個非常重要的原因。

Because I am going to introduce you as the number one CEO in this universe.

因為我要把你介紹成這個宇宙的頭號首席執行官。

Maybe entire universe, at least in this universe.

也許是整個宇宙，至少是這個宇宙。

But I was so worried.

但我很擔心。

This company, Apple's stock was going up last night.

這家公司，蘋果公司的股票昨晚上漲了。

And yours wasn't doing that particularly well.

而你的情況並非如此。

And I couldn't wait until the market was closed.

我不能等到市場關門。

And this morning, I woke up.

今天早上，我一覺醒來。

I asked my wife, I said, did Nvidia hang on there?

我問我的妻子，我說，Nvidia 堅持住了嗎？

No kidding.

別開玩笑了

Because we asked.

因為我們問過了。

So you are number one, Jensen.

所以你是第一，詹森。

So we have less than an hour this afternoon.

今天下午我們只有不到一個小時的時間。

I'm just diving, Jensen, with some tough questions for you.

詹森，我正在潛水，有些棘手的問題要問你。

So you have been leading this technology field, AI field, for a long period of time.

所以，你在這個技術領域，即人工智能領域，已經領跑了很長時間。

Just for the sake of our audience, just tell us how you think about AI, more specifically, recently, artificial general intelligence, the impact AI is having and will have for all the society and the industry.

為了聽眾著想，請談談您對人工智能的看法，更具體地說，最近的人工通用智能，以及人工智能正在和將要對整個社會和行業產生的影響。

First of all, thank you.

首先，謝謝你。

Thank you for the opportunity to spend time with you.

感謝您給我這個機會與您共度時光。

Harry is one of the most consequential computer scientists of our time.

哈里是當代最具影響力的計算機科學家之一。

And he's been a hero of mine and many others around the world for a very long time.

長期以來，他一直是我和世界各地許多人心目中的英雄。

And so it's a great pleasure to be here.

很高興來到這裡。

Harry, as you know, the transformative, the groundbreaking capability happened when an artificial intelligence network can learn to understand data of all kinds.

哈里，正如你所知道的，當人工智能網絡能夠學會理解各種數據時，就會產生變革性的、開創性的能力。

Of course, language, and images, and sequences of proteins, and sequences of amino acids, and sequences of chemicals.

當然，還有語言、影像、蛋白質序列、氨基酸序列和化學物質序列。

And all of a sudden, we now have computers that can understand the meaning of words and such.

突然之間，我們有了能理解文字含義的電腦。

And because of generative AI, we can translate from one modality of information to another modality of information.

由於人工智能的生成性，我們可以將一種資訊模式轉化為另一種資訊模式。

For example, from text to images, from text to text, from protein to text, from text to protein, from text to chemicals.

例如，從文本到影像、從文本到文本、從蛋白質到文本、從文本到蛋白質、從文本到化學物質。

This universal, initially a universal function approximator, evolved into a universal language translator of every kind.

這種通用性最初是一種通用函數近似器，後來發展成為各種通用語言翻譯器。

And so the question then is, what can we do with that?

那麼問題來了，我們該怎麼做呢？

And you see the number of startups around the world and a number of capabilities with the combination of all these different modalities and capabilities.

你可以看到全球初創企業的數量，以及將所有這些不同模式和能力結合在一起的能力。

And so I think the really amazing breakthrough is that we can now understand the meaning of information.

是以，我認為真正令人驚歎的突破在於，我們現在可以理解資訊的意義了。

Incredibly difficult information.

令人難以置信的困難資訊。

And so what does that mean to you if you are a digital biologist, so that you can understand the meaning of the data you're looking at, so that you could find a needle in a haystack?

那麼，如果你是一名數字生物學家，這對你意味著什麼呢？這樣你就能理解你所查看的數據的意義，從而大海撈針？

What does that mean if you are, in the case of NVIDIA, a chip designer, system designer?

如果你是英偉達公司的芯片設計師、系統設計師，這意味著什麼？

What does that mean to you if you're an ag tech, or climate science, or climate tech, or energy looking for a new material?

如果您是農業技術、氣候科學、氣候技術或能源方面的專家，這對正在尋找新材料的您來說意味著什麼？

So this is really the groundbreaking thing, is that we now have the concept of a universal translator.

是以，我們現在有了通用翻譯機的概念，這才是真正的突破。

You can understand anything you like.

你可以理解任何你喜歡的東西。

Yeah, Jensen, in May, we were at a Microsoft CEO summit.

是的，詹森，五月份，我們參加了微軟 CEO 峰會。

So many of us listened to your really description about the amazing impacts of AI on the society.

我們很多人都聽了您關於人工智能對社會的巨大影響的描述。

I thought that what you said really resonated with me.

我覺得你說的話讓我很有共鳴。

And to some extent, even shocked me.

在某種程度上，甚至讓我感到震驚。

So you kind of took us back to the entire history of a human being.

所以，你算是把我們帶回了人類的整個歷史。

You said, well, there's this agricultural revolution, so we actually manufacture more food.

你說，有了這場農業革命，我們實際上製造了更多的食物。

Then we have industrial revolution, we actually manufacture more iron steels.

工業革命之後，我們生產了更多的鋼鐵。

Then we have information technology, we actually have more information.

有了信息技術，我們實際上掌握了更多的資訊。

Now, in this intelligence, this era, so what you are doing at NVIDIA, what AI is doing is actually manufacturing intelligence.

現在，在這個智能時代，你們在英偉達所做的事情，人工智能所做的事情實際上就是製造智能。

Can you elaborate a little bit more on why this thing is just so enormously important?

你能再詳細解釋一下為什麼這件事如此重要嗎？

Yeah, when you look at what we've done together, Harry, and you were in the middle of all of this, from the perspective of computer science, we've reinvented the whole stack.

是的，當你看看我們一起做的事情，哈里，從計算機科學的角度來看，我們重塑了整個堆棧，而你就在這中間。

Meaning, the way we used to develop software, and when you think about computer science, you have to think about software development, how software is done.

也就是說，我們過去開發軟件的方式，當你考慮計算機科學時，你必須考慮軟件開發，考慮軟件是如何完成的。

And so we used to code software with our hands.

是以，我們習慣用雙手編寫軟件。

We imagined what the function it is that we would like to implement, whatever algorithm we'd like to implement, and we use our own creativity, we type it into the computer.

我們想象我們想要實現的功能是什麼，我們想要實現的算法是什麼，然後我們發揮自己的創造力，把它輸入電腦。

I started with Fortran, and I learned Pascal, and then C and C++, and of course, each one of these languages allow us to express our thoughts into code.

我從 Fortran 開始，然後學習 Pascal、C 和 C++，當然，每一種語言都能讓我們用代碼表達自己的想法。

And that code runs great on CPUs.

這些代碼在 CPU 上運行得非常好。

All of a sudden, now we use observed data, and we give this observed data to a computer, and we say, what is the function that you see inside this code?

突然間，我們使用了觀察到的數據，並將這些觀察到的數據交給計算機，然後我們說，你在這段代碼中看到的函數是什麼？

What are the patterns and relationships that you observe by studying all of the data that we presented to you?

通過研究我們向您提供的所有數據，您觀察到了哪些模式和關係？

And instead of using code, coding, we now use machine learning.

現在，我們不再使用代碼、編碼，而是使用機器學習。

And the machine generates not software, it generates neural networks that are processed on GPUs.

機器生成的不是軟件，而是在 GPU 上處理的神經網絡。

And so from coding to machine learning, from CPUs to GPUs, and because GPUs are so much more powerful, the type of software we can now develop is extraordinary.

是以，從編碼到機器學習，從 CPU 到 GPU，由於 GPU 的功能更加強大，我們現在可以開發的軟件類型非比尋常。

And what sits on top of it is artificial intelligence.

而人工智能則是其基礎。

That's what emerged.

這就是出現的情況。

And so computer science has really been transformed pretty much at the description I just said.

是以，計算機科學確實發生了很大的變化，就像我剛才所說的那樣。

Now the question is, what happens to our industry?

現在的問題是，我們的產業會發生什麼變化？

Of course, we're all racing to use machine learning to go discover new AIs.

當然，我們都在爭先恐後地利用機器學習去發現新的人工智能。

And what is AI?

什麼是人工智能？

And maybe that's, of course, one of the things about AI that you know very well is the automation of cognition.

當然，也許這也是你所熟知的人工智能的一個方面，那就是認知的自動化。

Automation of problem solving.

問題解決自動化。

And problem solving could be distilled down to, if I could, three basic ideas that you observe and perceive the environment, understand it, reason about it, and then come up with a plan to interact with it, whatever you decide your goals are.

如果可以的話，解決問題可以提煉為三個基本理念，即觀察和感知環境、理解環境、推理環境，然後制定與環境互動的計劃，無論你的目標是什麼。

And so perception, reasoning, and planning.

感知、推理和規劃也是如此。

The three fundamental steps of problem solving.

解決問題的三個基本步驟。

Well, perception, reasoning, and planning could be broken down into, for example, perceiving the environment around your car, reasoning about the location that you are and the location of all the other cars around you, planning how to drive.

例如，感知、推理和計劃可以細分為感知汽車周圍的環境、推理自己所在的位置和周圍所有其他汽車的位置、計劃如何駕駛。

So I just described self-driving cars.

所以，我剛才描述了自動駕駛汽車。

That self-driving car, in one manifestation, would be called a digital chauffeur.

這種自動駕駛汽車的一種表現形式被稱為數字司機。

And then you could do the same thing with you observe a CT scan, you understand it, you reason about everything that you see and you come to the conclusion there might be some anomaly that might be a tumor or something, and then you might decide to highlight it and describe it to the radiologist.

然後你可以做同樣的事情，觀察 CT 掃描，理解它，推理你所看到的一切，然後得出結論，可能有一些異常，可能是腫瘤或其他東西，然後你可能會決定突出它，並向放射科醫生描述它。

Now you're a digital radiologist.

現在，你是一名數字放射科醫生了。

In almost everything that we do, you can come up with some expression that artificial intelligence could then perform a particular task.

在我們所做的幾乎每一件事中，你都可以想出一些人工智能可以執行特定任務的表達方式。

Well, what happens is if we have enough of those digital agents, and those digital agents are interacting with the computer that's generating these digital artificial intelligence, digital intelligence, the total consumption of all of us into a data center makes the data center look like it's producing this thing called tokens or what we call tokens, but otherwise digital intelligence.

如果我們有足夠多的數字代理，而這些數字代理又在與產生這些數字人工智能、數字智能的計算機進行交互，那麼我們所有人對數據中心的總消耗就會使數據中心看起來像是在生產一種叫做代幣的東西，或者我們稱之為代幣，但其實是數字智能。

And so now let me now describe it a little bit differently. 300 years ago, as you know, General Electric and Westinghouse came up with a new type of instrument.

現在，讓我用不同的方式來描述它。大家都知道，300 年前，通用電氣和西屋公司發明了一種新型儀器。

In the beginning, a new type of machine that was called a dynamo and eventually became an AC generator.

起初，一種新型機器被稱為發電機，最終成為交流發電機。

And they were so smart to go and invent a consumer, a consumption of the electricity that they were able to produce.

他們很聰明，發明了一種消費者，一種他們能夠生產的電力的消耗者。

And that consumption, of course, would be things like light bulbs and toasters, right?

這種消費當然是燈泡和烤麵包機之類的東西，對嗎？

They created all kinds of digital appliances or electrical appliances that consumes the electricity that these plants would produce.

他們創造了各種數字家電或電器，消耗這些工廠生產的電力。

Well, look at what we're doing now.

看看我們現在在做什麼？

We're creating co-pilots and chat GPTs. We're creating all these different intelligence, basically light bulbs and toasters, and think of them as, right?

我們正在創造合作機器人和哈拉 GPT。 我們正在創造所有這些不同的智能，基本上就是燈泡和烤麵包機，把它們想象成，對嗎？

There are appliances that all of us would use, but you would connect it to a factory.

有些電器我們每個人都會使用，但你會把它連接到工廠。

It used to be an AC power generation factory, but this new factory is digital intelligence factory.

它過去是一家交流發電廠，而這家新工廠則是數字智能工廠。

And so what is just, from an industrial perspective, really what's happening here is we're now creating a new industry, and this new industry takes energy in and produces digital intelligence out.

是以，從工業的角度來看，我們現在正在創造一個新的產業，這個新產業吸收能源，生產數字智能產品。

And that digital intelligence would be used by all kinds of different applications.

數字智能將被各種不同的應用所使用。

And the consumption of it, we believe, is gonna be quite large.

我們相信，它的消費量會相當大。

And this entire industry never existed before, just like the AC generation industry never existed before that.

而這整個行業在此之前從未存在過，就像空調發電行業在此之前從未存在過一樣。

But that's really, truly amazing.

但這真的非常了不起。

Jensen, you are describing this really bright future for us.

詹森，你為我們描述了一個非常光明的未來。

Of course, we know this thing is going to happen.

當然，我們知道這件事一定會發生。

Really much of that's because of your efforts and Nvidia's contribution to the field, especially over the last 10 years, 12 years.

這在很大程度上要歸功於你的努力和 Nvidia 在這一領域的貢獻，尤其是在過去的 10 年或 12 年裡。

So one number just keeps coming back and people are talking about, in the name of scaling law and others, most recently, in your name, there's something called the Huang's Law, in comparison with to Moore's Law.

是以，有一個數字不斷出現，人們以縮放定律的名義，還有其他人，最近以你的名義，在與摩爾定律的比較中，都在談論一個叫黃氏定律的東西。

Of course, in the earlier, in computing industry, Intel came up with Moore's Law, basically meaning every 18 months, computing power will increase, will double.

當然，在早期的計算行業，英特爾提出了摩爾定律，基本意思是每 18 個月，計算能力就會增加一倍。

And then now, if we look at last 10, 12 years, under your leadership, it's not even, every year double, it's more than that.

而現在，如果我們看看過去的 10 年、12 年，在您的上司下，甚至不是每年翻一番，而是更多。

If we look at it from the consumption side, all those large language models over the last 12 years, every year, it's actually more than four times an increase of the computing needs.

如果我們從消費的角度來看，在過去的 12 年裡，所有這些大型語言模型的計算需求每年實際上都會增長四倍以上。

Now every year, it's four times.

現在每年都是四次。

Then in 10 years, it's an enormous number, it's actually a million.

再過 10 年，這是一個巨大的數字，實際上是一百萬。

So that's how, at least I explain to people why Jensen's stock went up 300x in 10 years.

這樣，至少我可以向人們解釋，為什麼詹森的股票在 10 年內上漲了 300 倍。

If you think about it, the computing needs is a million times more, so that's, then it explains the stock probably is not that expensive.

如果你仔細想想，計算需求要高出一百萬倍，所以，這就解釋了股票可能並沒有那麼貴。

My question for you then is that, as you look at, as you look into the future with your crystal ball, are we going to see that a million times more needs increase for the next 10 years?

那麼，我想問的是，當你用水晶球展望未來時，我們是否會看到未來 10 年的需求將增加一百萬倍？

So, Moore's Law depended on two concepts.

是以，摩爾定律取決於兩個概念。

One concept was VLSI scaling, and that was because of Carver Mead.

其中一個概念是超大規模集成電路（VLSI）擴展，這要歸功於卡佛-米德。

And the text by Mead and Conway really inspired my generation.

而米德和康威的文字確實激勵了我們這一代人。

The second is Dennard scaling.

第二種是 Dennard 縮放。

Constant current density scaling of transistors coupled with the shrinking of the transistors made it possible for us to double the performance, if you will, double the performance of semiconductors every couple of years.

半導體的恆定電流密度與半導體的縮小相結合，使我們有可能每隔幾年就將半導體的性能提高一倍。

And so every one and a half years, so that would be 10 times every five years, 100 times every 10 years.

是以，每隔一年半，也就是每五年 10 次，每十年 100 次。

And the other, what we're experiencing now is that the larger your neural network can become, and the more data that you train that neural network with, the more intelligent the AI seems to become.

另一方面，我們現在正在經歷的是，你的神經網絡越大，你訓練神經網絡的數據越多，人工智能似乎就變得越智能。

It's an empirical law, just like Moore's Law was.

這是一條經驗法則，就像摩爾定律一樣。

We call that the scaling law, and the scaling law appears to be continuing.

我們稱之為縮放定律，而縮放定律似乎仍在繼續。

But the one thing that we also know about intelligence is that pre-training, just taking all of the data in the world and discovering knowledge from it automatically, pre-training is not enough.

但我們也知道，關於智能的一件事是，預訓練，即僅僅獲取世界上的所有數據並自動從中發現知識，預訓練是不夠的。

Just as going to college and graduating from college is a very important milestone, but it's not enough.

正如上大學和大學畢業是一個非常重要的里程碑，但這還不夠。

There's post-training, which is learning a particular skill very deeply.

還有後期培訓，即深入學習某項技能。

And post-training requires reinforcement learning, human feedback, reinforcement learning, AI feedback, synthetic data generation, multi-path learning, reinforcement learning.

而後期訓練則需要強化學習、人工反饋、強化學習、人工智能反饋、合成數據生成、多路徑學習、強化學習。

There's a whole bunch of techniques, but basically, you're now going deep into a particular domain, and you're trying to learn something very, very deep about it.

有一大堆技巧，但基本上，你現在正在深入一個特定的領域，並試圖學習一些非常、非常深入的知識。

That's post-training.

那是訓練後的事情。

Once you select a particular career, you're gonna do tons and tons of learning again.

一旦你選擇了某一職業，你就又要進行成噸成噸的學習。

And then after that, of course, it's called thinking.

之後，當然就是思考。

And that's what we call test time scaling.

這就是我們所說的測試時間縮放。

Some things, you just know the answer to.

有些事情，你就是知道答案。

Some things, you have to break the problem down into step-by-step-by-step, into its first principled elements, and from first principles, try to find a solution for each one of them.

有些事情，你必須一步一步地把問題分解成最初的原則性要素，並從最初的原則出發，嘗試為每一個要素找到解決方案。

It might require you to iterate.

這可能需要你反覆推敲。

It might require you to simulate various outcomes because the answer is not predictive, and so on and so forth.

這可能需要你模擬各種結果，因為答案是無法預測的，諸如此類。

And so we call that thinking.

是以，我們稱之為 "思考"。

And the longer you think, maybe the higher quality the answer would become.

思考的時間越長，也許答案的品質就越高。

So notice, we now have three areas of artificial intelligence development where a great deal of computation would result in higher quality answers.

所以請注意，我們現在有三個人工智能發展領域，在這些領域中，大量的計算將帶來更高質量的答案。

Today, the answers that we have are the best that we can provide.

今天，我們所能提供的答案已經是最好的了。

But we need to get to a point where the answer that you get is not the best that we can provide, and somewhat, you still have to decide whether is this hallucinated or not hallucinated?

但我們需要達到這樣的程度，即你得到的答案並不是我們所能提供的最佳答案，在某種程度上，你仍然需要判斷這是否是幻覺？

Does this make sense?

這合理嗎？

Is it sensible or not sensible?

是合理還是不合理？

We have to get to a point where the answer that you get, you largely trust.

我們必須達到這樣的程度，即你得到的答案在很大程度上是可信的。

You largely trust.

你在很大程度上是信任的。

And so I think that we're several years away from being able to do that, and in the meantime, we have to keep increasing our computation.

是以，我認為我們還需要幾年時間才能做到這一點，與此同時，我們必須不斷提高計算能力。

Now, one of the things that you said earlier that I really appreciate is that in the last 10 years, we increased the performance by a million times.