The Inside Story of ChatGPT’s Astonishing Potential | Greg Brockman | TED

譯者: Lilian Chiu 審譯者: Helen Chang

We started OpenAI seven years ago

我們在七年前成立了 OpenAI，

because we felt like something really interesting was happening in AI

因為我們認為人工智慧 領域的發展很有意思，

and we wanted to help steer it in a positive direction.

而我們想要協助引導它 朝正面的方向前進。

It's honestly just really amazing to see

從那之後人工智慧領域的進展， 真的是很不可思議。

how far this whole field has come since then.

And it's really gratifying to hear from people like Raymond

也很高興能聽到像雷蒙這些人

who are using the technology we are building, and others,

把我們與其他人打造的這項技術

for so many wonderful things.

用來做許多美好的事。

We hear from people who are excited,

我們聽見一些興奮的聲音，

we hear from people who are concerned,

聽見一些擔憂的聲音，

we hear from people who feel both those emotions at once.

也聽到同時有這兩種情緒的聲音。

And honestly, that's how we feel.

老實說，那正是我們的感覺。

Above all, it feels like we're entering an historic period right now

尤其是，我們現在似乎正在 邁入一個歷史性的時期，

where we as a world are going to define a technology

我們這個世界要來定義一項技術，

that will be so important for our society going forward.

它對我們未來的社會相當重要。

And I believe that we can manage this for good.

我相信我們可以善加管理它。

So today, I want to show you the current state of that technology

今天我想告訴大家這項技術的現狀，

以及背後我們所堅守的一些原則。

and some of the underlying design principles that we hold dear.

So the first thing I'm going to show you

我首先要展示的是， 為人工智慧而非人類建立工具

is what it's like to build a tool for an AI

是怎麼一回事。

rather than building it for a human.

So we have a new DALL-E model, which generates images,

所以，

我們有新的 DALL-E 模型， 它會產生影像，

and we are exposing it as an app for ChatGPT to use on your behalf.

我們把它開發成應用程式， 讓 ChatGPT 可以代表你來使用。

And you can do things like ask, you know,

你可以做的事包括詢問

suggest a nice post-TED meal and draw a picture of it.

「建議聽完 TED 演說後吃什麼好，

然後把它畫出來。」

(Laughter)

（笑聲）

Now you get all of the, sort of, ideation and creative back-and-forth

接著你會得到所有的

想法和有創意的對話，

and taking care of the details for you that you get out of ChatGPT.

還會幫你顧到細節， 這些都是 ChatGPT 的功能。

And here we go, it's not just the idea for the meal,

來了，不只是關於吃什麼的想法，

but a very, very detailed spread.

還有非常非詳細的食物擺盤方式。

咱們來看看我們會得到什麼。

So let's see what we're going to get.

But ChatGPT doesn't just generate images in this case --

但在這個例子裡，ChatGPT 不只是產生影像——

sorry, it doesn't generate text, it also generates an image.

抱歉，它不只是產生文字， 也會產生影像。

And that is something that really expands the power

ChatGPT 代表你來實現 你的意圖的能力

of what it can do on your behalf in terms of carrying out your intent.

被大大擴展了。

And I'll point out, this is all a live demo.

我要強調，這是現場示範。

This is all generated by the AI as we speak.

我們說話的同時 人工智慧在產生這些資訊，

So I actually don't even know what we're going to see.

我甚至不知道我們會看到什麼。

This looks wonderful.

這看起來好棒。

(Applause)

（掌聲）

I'm getting hungry just looking at it.

我光看就餓了。

Now we've extended ChatGPT with other tools too,

我們也用其他工具讓 ChatGPT 更強大，比如，記憶。

for example, memory.

You can say "save this for later."

你可以說「記住這個，等下會用到。」

And the interesting thing about these tools

這些工具有趣的地方在於 你很容易就能檢視它們。

is they're very inspectable.

So you get this little pop up here that says "use the DALL-E app."

你會看到彈出訊息，寫著 「用 DALL-E 應用程式」。

順道一提，ChatGPT 使用者 在幾個月後就可以使用了。

And by the way, this is coming to you, all ChatGPT users, over upcoming months.

你可以揭開它的面紗，

And you can look under the hood and see that what it actually did

看到它所做的其實 就只是像人類一樣寫個指示。

was write a prompt just like a human could.

And so you sort of have this ability to inspect

所以你有能力可以去檢視

how the machine is using these tools,

機器如何使用這些工具，

which allows us to provide feedback to them.

這讓我們可以提供回饋給它們。

Now it's saved for later,

它被記著，等下再用， 讓我示範運用如何運用那些資訊，

and let me show you what it's like to use that information

and to integrate with other applications too.

以及和其他應用整合。

You can say,

你可以說：

“Now make a shopping list for the tasty thing

「幫我做一張採買清單，

我需要什麼才能做剛剛 提到的那個美味食物？」

I was suggesting earlier.”

And make it a little tricky for the AI.

給人工智慧出點難題。

"And tweet it out for all the TED viewers out there."

「再用推文的方式，

告訴所有的 TED 觀眾。」

(Laughter)

（笑聲）

So if you do make this wonderful, wonderful meal,

如果你真的做了這種美食， 我很想知道嚐起來如何。

I definitely want to know how it tastes.

But you can see that ChatGPT is selecting all these different tools

各位可以看見，ChatGPT 在選擇各種工具，

without me having to tell it explicitly which ones to use in any situation.

我都沒有明確告訴它 在什麼情況要用什麼工具。

And this, I think, shows a new way of thinking about the user interface.

我認為，這讓我們用新的方式 來看待使用者介面。

Like, we are so used to thinking of, well, we have these apps,

我們太習慣會認為 我們有這些應用程式，

we click between them, we copy/paste between them,

在它們之間點選、複製、貼上，

若你知道應用程式的選單和所有 選項，用它應該是很棒的體驗。

and usually it's a great experience within an app

as long as you kind of know the menus and know all the options.

是的，我希望你做。

Yes, I would like you to.

Yes, please.

是的，拜託。禮貌點總是好事。

Always good to be polite.

(Laughter)

（笑聲）

And by having this unified language interface on top of tools,

有這種統一的語言介面

建構在工具之上，

the AI is able to sort of take away all those details from you.

人工智慧就可以替你 接手所有這些細節。

So you don't have to be the one

就不用由你來

who spells out every single sort of little piece

解釋應該要如何做的每一個細節。

of what's supposed to happen.

And as I said, this is a live demo,

我剛說過，這是現場示範，

so sometimes the unexpected will happen to us.

有時會發生未預期的狀況。

But let's take a look at the Instacart shopping list while we're at it.

在此之際，我們就來看一下 配送服務的採購清單。

And you can see we sent a list of ingredients to Instacart.

可以看到我們發送了 原料清單到配送服務。

Here's everything you need.

你需要的都在這了。

And the thing that's really interesting

有趣的是，傳統使用者介面 仍然非常有價值。

is that the traditional UI is still very valuable, right?

If you look at this,

看看這個介面，

you still can click through it and sort of modify the actual quantities.

你還是可以在上面點選， 修改實際數量。

And that's something that I think shows

我認為這顯示 傳統使用者介面不會消失，

that they're not going away, traditional UIs.

It's just we have a new, augmented way to build them.

我們只是用擴充的 新方法來打造它們。

And now we have a tweet that's been drafted for our review,

現在有了一則推文的草稿， 等我們複審，

which is also a very important thing.

這也十分重要。

We can click “run,” and there we are, we’re the manager, we’re able to inspect,

我們可以點選「執行」， 由我們來主導，

如果想要，我們可以 檢視及改變人工智慧做的工作。

we're able to change the work of the AI if we want to.

And so after this talk, you will be able to access this yourself.

在這場演說之後，

各位自己可以用它。

And there we go.

好了。

Cool.

酷。

Thank you, everyone.

謝謝大家。

(Applause)

（掌聲）

So we’ll cut back to the slides.

我們可以切回投影片。

Now, the important thing about how we build this,

關於我們如何建造這些， 很重要的是，

it's not just about building these tools.

重點不只是建造這些工具，

It's about teaching the AI how to use them.

還要教人工智慧如何使用它們。

Like, what do we even want it to do

比如當我們問出非常高階的問題時 是希望它做什麼。

when we ask these very high-level questions?

And to do this, we use an old idea.

為此，我們用了一個老點子。

If you go back to Alan Turing's 1950 paper on the Turing test, he says,

回顧 1950 年艾倫‧圖靈的 圖靈測試論文，

you'll never program an answer to this.

他說，你永遠無法用程式 回答這個問題。

Instead, you can learn it.

但你可以用學的。

You could build a machine, like a human child,

你可以建造一台像人類兒童的機器，

and then teach it through feedback.

用回饋的方式教導它。

Have a human teacher who provides rewards and punishments

找個人類老師給它， 當它嘗試做好事或壞事時，

as it tries things out and does things that are either good or bad.

提供獎勵或懲罰。

And this is exactly how we train ChatGPT.

我們正是如此訓練 ChatGPT ，

It's a two-step process.

有兩個步驟：

First, we produce what Turing would have called a child machine

第一，我們製造出圖靈 所謂的兒童機器，

用非監督式學習來做到這一點。

through an unsupervised learning process.

我們只是讓它見識 整個網際網路世界，

We just show it the whole world, the whole internet

and say, “Predict what comes next in text you’ve never seen before.”

說：「在你從未見過的文字中 預測接下來會是什麼。」

這個過程賦予它各種很棒的技能。

And this process imbues it with all sorts of wonderful skills.

For example, if you're shown a math problem,

比如，你丟個數學題給它， 要解這題唯一的方式，

the only way to actually complete that math problem,

to say what comes next,

說出接下來會是什麼， 上面綠色的九，

that green nine up there,

is to actually solve the math problem.

就只能靠真正去解那個數學題。

But we actually have to do a second step, too,

但我們也得進行第二個步驟， 教人工智慧如何用那些技能。

which is to teach the AI what to do with those skills.

為此，我們要提供回饋。

And for this, we provide feedback.

We have the AI try out multiple things, give us multiple suggestions,

我們讓人工智慧做多種嘗試， 給我們多種建議，

由人類來對建議評分： 「這個比那個好。」

and then a human rates them, says “This one’s better than that one.”

這不僅能強化人工智慧 所給的單一回覆，

And this reinforces not just the specific thing that the AI said,

非常重要的是，也會強化 人工智慧用來產生該回覆的流程。

but very importantly, the whole process that the AI used to produce that answer.

這就讓它能做到一般化， 讓它能教導，

And this allows it to generalize.

It allows it to teach, to sort of infer your intent

能推論你的意圖，

and apply it in scenarios that it hasn't seen before,

並應用在它尚未見過 也沒得過相關回饋的情境中。

that it hasn't received feedback.

Now, sometimes the things we have to teach the AI

有時，你可能很難想像我們 必須要教人工智慧什麼。

are not what you'd expect.

For example, when we first showed GPT-4 to Khan Academy,

比如，我們初次展示 GPT-4 給可汗學院看時，

they said, "Wow, this is so great,

他們說：「哇，好讚， 我們可以教學生很棒的東西了。

We're going to be able to teach students wonderful things.

唯一的問題是 它不複查學生的數學。

Only one problem, it doesn't double-check students' math.

If there's some bad math in there,

如果有算錯，它會很高興地假裝 一加一等於三然後繼續做下去。」

it will happily pretend that one plus one equals three and run with it."

So we had to collect some feedback data.

所以，我們得收集一些回饋資料。

Sal Khan himself was very kind

薩爾曼‧可汗本人很好心地 投入二十小時自己的時間，

and offered 20 hours of his own time to provide feedback to the machine

和我們的團隊一起提供回饋給機器。

alongside our team.

And over the course of a couple of months we were able to teach the AI that,

經過幾個月的時間， 我們教會了人工智慧

"Hey, you really should push back on humans

「嘿，在這種特定的情況下 你應該要反駁人類。」

in this specific kind of scenario."

And we've actually made lots and lots of improvements to the models this way.

實際上我們用這種方式 對模型做了很多很多的改善。

And when you push that thumbs down in ChatGPT,

當你在 ChatGPT 中按了 大拇指向下的圖示，

that actually is kind of like sending up a bat signal to our team to say,

就等同向我們的團隊發出蝙蝠燈號，

「你在這個領域比較弱， 應該收集回饋。」

“Here’s an area of weakness where you should gather feedback.”

我們用這種按大拇指圖示的 方式來傾聽使用者

And so when you do that,

that's one way that we really listen to our users

and make sure we're building something that's more useful for everyone.

並確保我們的成品 對每個人都更有用。

Now, providing high-quality feedback is a hard thing.

提供高品質回饋是很困難的事。

如果你要孩子清理他們的房間， 而你卻只是檢查地板，

If you think about asking a kid to clean their room,

if all you're doing is inspecting the floor,

you don't know if you're just teaching them to stuff all the toys in the closet.

你不會知道你其實只是 教他們把玩具全塞到衣櫥裡。

This is a nice DALL-E-generated image, by the way.

順道一提，這張好看的圖 是 DALL-E 產生的。

And the same sort of reasoning applies to AI.

同樣的邏輯可套用到人工智慧。

As we move to harder tasks,

當我們邁向更困難的工作任務，

we will have to scale our ability to provide high-quality feedback.

我們就得提升我們 提供高品質回饋的能力。

But for this, the AI itself is happy to help.

但，人工智慧很樂意 提供這方面的協助。

It's happy to help us provide even better feedback

它很樂意協助我們提供更好的回饋，

and to scale our ability to supervise the machine as time goes on.

並逐步提升我們監督機器的能力。

And let me show you what I mean.

讓我用示範來說明。

For example, you can ask GPT-4 a question like this,

比如，

你可以問 GPT-4 這樣的問題：

of how much time passed between these two foundational blogs

這兩篇基礎的部落格文章 之間差了多少時間，

on unsupervised learning

一篇是談非監督式學習， 另一篇是談從人類回饋中學習。

and learning from human feedback.

And the model says two months passed.

模型說，間隔兩個月。

But is it true?

但，是真的嗎？這些模型 並非百分之百可靠，

Like, these models are not 100-percent reliable,

although they’re getting better every time we provide some feedback.

雖然每當我們提供回饋 就能讓它們改善一些。

But we can actually use the AI to fact-check.

但我們可以用人工智慧來查核事實。

And it can actually check its own work.

它可以檢查它自己的回覆。

You can say, fact-check this for me.

你可以說，幫我查核一下。

Now, in this case, I've actually given the AI a new tool.

在這個例子，我其實給了 人工智慧一項新工具，

This one is a browsing tool

一項瀏覽工具，

where the model can issue search queries and click into web pages.

讓模型可以發出搜尋查詢， 並點選去看網頁。

And it actually writes out its whole chain of thought as it does it.

這麼做的過程中， 它寫出了它的整串思路。

It says, I’m just going to search for this and it actually does the search.

它說，我要搜尋這個， 也實際去搜尋。

接著它找到了刊載日期和搜尋結果。

It then it finds the publication date and the search results.

接著它又發出另一個搜尋， 它要點進部落格文章。

It then is issuing another search query.

It's going to click into the blog post.

And all of this you could do, but it’s a very tedious task.

這些你都能做， 但都是很乏味的工作。

It's not a thing that humans really want to do.

那不是人類會想做的事， 坐在駕駛座上會有趣許多，

It's much more fun to be in the driver's seat,

坐在管理的位子上， 如果你想要再三複查都可以。

to be in this manager's position where you can, if you want,

triple-check the work.

最後出現引用文獻，

And out come citations

so you can actually go

這樣你真的非常容易就能去驗證

and very easily verify any piece of this whole chain of reasoning.

這整串推理的任何一部分。

And it actually turns out two months was wrong.

結果發現，不是兩個月。

Two months and one week,

兩個月又一週，那才是對的。

that was correct.

(Applause)

（掌聲）

And we'll cut back to the side.

我們再切回投影片。

And so thing that's so interesting to me about this whole process

這整個過程讓我覺得很有趣的部分

is that it’s this many-step collaboration between a human and an AI.

是人類和人工智慧之間 要做許多步驟的合作。

Because a human, using this fact-checking tool

因為人類使用查核事實的工具

is doing it in order to produce data

目的是要產生出資料

for another AI to become more useful to a human.

讓另一個人工智慧 變得對人類更有用。

And I think this really shows the shape of something

我認為這真的勾勒出了

that we should expect to be much more common in the future,

未來會更常見的樣貌，

where we have humans and machines kind of very carefully

未來的人類和機器 在設計上能很謹慎且微妙地

and delicately designed in how they fit into a problem

處理問題以及設法解決問題。

and how we want to solve that problem.

We make sure that the humans are providing the management, the oversight,

我們會確保由人類提供 管理、監督、回饋，

the feedback,

而機器運作的方式是 可以被檢視且能被信賴的。

and the machines are operating in a way that's inspectable

and trustworthy.

同心協力我們就能夠 創造出更能信賴的機器。

And together we're able to actually create even more trustworthy machines.

我認為假以時日， 如果我們做對這個流程，

And I think that over time, if we get this process right,

我們將能解決不可能解決的問題。

we will be able to solve impossible problems.

讓各位了解我說的是 什麼程度的不可能：

And to give you a sense of just how impossible I'm talking,

I think we're going to be able to rethink almost every aspect

我認為我們將來會能重新思考我們

of how we interact with computers.

和電腦互動的幾乎每個面向。

For example, think about spreadsheets.

比如，想想試算表。

They've been around in some form since, we'll say, 40 years ago with VisiCalc.

大約四十年前試算表就以 VisiCalc 的形式出現了。

I don't think they've really changed that much in that time.

在那個時期，它們沒有很大的變化。

And here is a specific spreadsheet of all the AI papers on the arXiv

這張試算表列出了 arXiv 上過去三十年間的

所有人工智慧論文。

for the past 30 years.

There's about 167,000 of them.

總共約十六萬七千篇。

And you can see there the data right here.

各位可以看到資料就在這裡。

But let me show you the ChatGPT take on how to analyze a data set like this.

讓我給各位看，ChatGPT 認為 要如何分析像這樣的資料集。

So we can give ChatGPT access to yet another tool,

我們讓 ChatGPT 使用另一項工具，

this one a Python interpreter,

這是 Python 直譯器， 可以跑程式碼，

so it’s able to run code, just like a data scientist would.

就如同資料科學家一樣。

And so you can just literally upload a file

你可以真的上傳一個檔案，

and ask questions about it.

並問相關的問題。

And very helpfully, you know, it knows the name of the file and it's like,

非常有幫助的是，

它看得懂檔名，會說：「喔， 這是 CSV 檔，逗號分隔值檔案，

"Oh, this is CSV," comma-separated value file,

"I'll parse it for you."

我幫你分析。」

The only information here is the name of the file,

這裡唯一的資訊是檔案的名稱、

the column names like you saw and then the actual data.

各位剛看到的欄位名稱， 接著是實際的資料。

And from that it's able to infer what these columns actually mean.

它可以從這些來推論出 這些欄位的意義是什麼。

Like, that semantic information wasn't in there.

這裡並沒有那些語義資訊。

它得要把它的世界知識 拼湊起來，要知道：

It has to sort of, put together its world knowledge of knowing that,

「喔，對，arXiv 是 大家上傳論文的網站，

“Oh yeah, arXiv is a site that people submit papers

and therefore that's what these things are and that these are integer values

所以這些是論文，

而這些是整數， 因此是論文的作者數目。」

and so therefore it's a number of authors in the paper,"

那些都是人會做的事，

like all of that, that’s work for a human to do,

而人工智慧很樂意幫忙。

and the AI is happy to help with it.

我甚至不知道我想要問什麼。

Now I don't even know what I want to ask.

幸運的是，你可以要機器這樣做：

So fortunately, you can ask the machine,

"Can you make some exploratory graphs?"

「你能做些探索性的圖表嗎？」

And once again, this is a super high-level instruction with lots of intent behind it.

這同樣也是個超高階的指示， 背後有許多意圖。

But I don't even know what I want.

但我甚至不知道我要什麼， 而人工智慧得推論我會想知道什麼。

And the AI kind of has to infer what I might be interested in.

我想它提出了一些好點子。

And so it comes up with some good ideas, I think.

每篇論文作者數目的直方圖、

So a histogram of the number of authors per paper,

每年論文數的時間序列、

time series of papers per year, word cloud of the paper titles.

論文標題的文字雲。

All of that, I think, will be pretty interesting to see.

我認為這些分析都挺有趣的。

很棒的是，它真的可以做這些分析。

And the great thing is, it can actually do it.

Here we go, a nice bell curve.

來了，很好的鐘形， 最常見的是三位作者。

You see that three is kind of the most common.

它正在製作這張很棒的圖，

It's going to then make this nice plot of the papers per year.

呈現每年的論文數。

Something crazy is happening in 2023, though.

不過，2023 年發生了很瘋狂的事。

Looks like we were on an exponential and it dropped off the cliff.

看來是在指數爬升後又落下懸崖。

是怎麼了？順道一提，這些都是 你可以檢視的 Python 程式碼。

What could be going on there?

By the way, all this is Python code, you can inspect.

And then we'll see word cloud.

接著是文字雲。

可以看到標題中會出現這些字詞。

So you can see all these wonderful things that appear in these titles.

But I'm pretty unhappy about this 2023 thing.

但這個 2023 年的狀況讓我 不開心，它讓這年看來很糟。

It makes this year look really bad.

Of course, the problem is that the year is not over.

當然，問題在於這年還沒結束。

So I'm going to push back on the machine.

所以我要對機器提出抗議。

[Waitttt that's not fair!!!

〔等等！那不公平， 2023 年還沒結束。

2023 isn't over.

What percentage of papers in 2022 were even posted by April 13?]

2022 年的論文有多少百分比 是在四月十三日前刊登的？〕

So April 13 was the cut-off date I believe.

我認為資料截止日是四月十三日。

Can you use that to make a fair projection?

你能用那資訊來做公平的預測嗎？

So we'll see, this is the kind of ambitious one.

等著看吧，這野心挺大的。

（笑聲）

(Laughter)

So you know,

所以，

again, I feel like there was more I wanted out of the machine here.

同樣的，在此我希望 機器能做的還要更多。

I really wanted it to notice this thing,

我真的希望它能注意到這件事，

maybe it's a little bit of an overreach for it

也許這樣要求它是有點超過了，

to have sort of, inferred magically that this is what I wanted.

希望它能很神奇地 推論出這就是我想要的。

But I inject my intent,

但我放入了我的意圖，

I provide this additional piece of, you know, guidance.

我提供了這項額外的指引。

And under the hood,

在背後，

the AI is just writing code again, so if you want to inspect what it's doing,

人工智慧又在寫程式碼了，

如果你想檢視它在做什麼， 是非常可行的。

it's very possible.

And now, it does the correct projection.

現在，

它做了正確的預測。

(Applause)

（掌聲）

If you noticed, it even updates the title.

注意看看，它甚至更新了標題。

I didn't ask for that, but it know what I want.

我沒要求這點， 但它知道我要什麼。

Now we'll cut back to the slide again.

現在我們再回到投影片。

This slide shows a parable of how I think we ...

這張投影片上是一個寓言， 說明我認為我們如何……

A vision of how we may end up using this technology in the future.

我們未來可能會如何 運用這項科技的預景。

A person brought his very sick dog to the vet,

一個人帶了他重病的狗去看獸醫，

and the veterinarian made a bad call to say, “Let’s just wait and see.”

獸醫做了很糟的判斷，說： 「我們先觀察就好。」

And the dog would not be here today had he listened.

如果他聽了獸醫的話， 狗今天就不在了。

In the meanwhile, he provided the blood test,

同時，他也把完整的血液檢測 醫療記錄提供給 GPT-4，

like, the full medical records, to GPT-4,

which said, "I am not a vet, you need to talk to a professional,

它說：「我不是獸醫， 你需要和專業人士談談，

here are some hypotheses."

以下是一些假設。」

He brought that information to a second vet

他把這些資訊帶給第二位獸醫，

who used it to save the dog's life.

獸醫用這些資訊救了這隻狗一命。

Now, these systems, they're not perfect.

這些系統並不完美。

You cannot overly rely on them.

你不能過度仰賴它們。

But this story, I think, shows

但我認為這個故事告訴我們，

that a human with a medical professional

有醫療專業的人類

and with ChatGPT as a brainstorming partner

加上 ChatGPT 當腦力激盪的夥伴，

was able to achieve an outcome that would not have happened otherwise.

就能夠達到本來達不到的結果。

I think this is something we should all reflect on,

我認為在考量如何將這些系統 整合到我們的世界中時

think about as we consider how to integrate these systems

我們應該反思這一點，

into our world.

And one thing I believe really deeply,

而我深深相信，

is that getting AI right is going to require participation from everyone.

要把人工智慧做對， 會需要每個人的參與。

And that's for deciding how we want it to slot in,

也就是決定我們希望它如何融入，

that's for setting the rules of the road,

就是為道路設立規則，

for what an AI will and won't do.

人工智慧能或不能做什麼。

And if there's one thing to take away from this talk,

如果說能從這場演說得到一個訊息，

it's that this technology just looks different.

那就是：這項科技看起來不同， 和大家所有的預期都不同。

Just different from anything people had anticipated.

And so we all have to become literate.

所以我們都得要有熟悉它。

And that's, honestly, one of the reasons we released ChatGPT.

老實說，這是我們推出 ChatGPT 的理由之一。

Together, I believe that we can achieve the OpenAI mission

我相信我們同心協力 就能達成 OpenAI 的使命：

of ensuring that artificial general intelligence

確保人工通用智慧能讓全人類受惠。

benefits all of humanity.

Thank you.

謝謝。

(Applause)

（掌聲）

(Applause ends)

（掌聲結束）

Chris Anderson: Greg.

主持人：葛雷格。

Wow.

哇。

I mean ...

我是說……

I suspect that within every mind out here

搞不好這裡每個人的腦中

there's a feeling of reeling.

都有種天旋地轉的感覺。

我猜想可能大部分 在看這場演說的觀眾

Like, I suspect that a very large number of people viewing this,

you look at that and you think, “Oh my goodness,

看著那些，心想：「喔，我的天，

pretty much every single thing about the way I work, I need to rethink."

我可能得重新思考 我工作方式的每個環節了。」

Like, there's just new possibilities there.

因為有新的可能性， 我說的對嗎？有誰認為

Am I right?

Who thinks that they're having to rethink the way that we do things?

我們得要重新思考我們做事的方式？

Yeah, I mean, it's amazing,

是啊，很不簡單，但也很駭人。

but it's also really scary.

So let's talk, Greg, let's talk.

所以，葛雷格，咱們來談談。

I mean, I guess my first question actually is just

我想我的第一個問題是： 你到底怎麼做到這些的？

how the hell have you done this?

（笑聲）

(Laughter)

OpenAI has a few hundred employees.

OpenAI 有幾百名員工。

Google has thousands of employees working on artificial intelligence.

Google 有數千名員工 投入在人工智慧上。

Why is it you who's come up with this technology

為什麼你們能創造出 這項震驚全世界的科技？

that shocked the world?

Greg Brockman: I mean, the truth is,

講者：毫無疑問， 我們都站在巨人的肩膀上。

we're all building on shoulders of giants, right, there's no question.

看看計算、演算法、資料的進步，

If you look at the compute progress,

都是整個產業的。

the algorithmic progress, the data progress,

all of those are really industry-wide.

但在 OpenAI 的早期，我們 做了許多非常刻意的選擇。

But I think within OpenAI,

we made a lot of very deliberate choices from the early days.

第一個就是要去面對現實。

And the first one was just to confront reality as it lays.

我們非常用力去思考：

And that we just thought really hard about like:

What is it going to take to make progress here?

要怎樣才能在此有所進展？

We tried a lot of things that didn't work, so you only see the things that did.

我們試了很多行不通的方法， 所以只會看到行得通的。

我認為最重要的一點是要

And I think that the most important thing has been to get teams of people

非常不同的人所組成的團隊 能夠和諧地合作。

who are very different from each other to work together harmoniously.

CA: Can we have the water, by the way, just brought here?

主持人：能不能把水拿過來？

我想我們會很需要， 這是會讓人口乾的主題。

I think we're going to need it, it's a dry-mouth topic.

But isn't there something also just about the fact

但，是不是也有是因為

that you saw something in these language models

你在這些語言模型中看到了什麼，

that meant that if you continue to invest in them and grow them,

意味著如果你持續 投資它們、發展它們，

that something at some point might emerge?

在某個時點可能會出現什麼？

GB: Yes.

講者：是的。

我認為，

And I think that, I mean, honestly,

I think the story there is pretty illustrative, right?

老實說，我認為 這個故事挺有說服力，

我認為高階、深度學習，

I think that high level, deep learning,

我們一直知道我們想要做深度 學習實驗室，但實際要如何做？

like we always knew that was what we wanted to be,

was a deep learning lab, and exactly how to do it?

在早期我們並不知道， 我們做了很多嘗試，

I think that in the early days, we didn't know.

We tried a lot of things,

有一個人投入在

and one person was working on training a model

訓練一個模型來預測亞馬遜 評論中的下一個字，

to predict the next character in Amazon reviews,

and he got a result where -- this is a syntactic process,

而他得到了這樣的結果：

這是個句法過程，所以你會預期

you expect, you know, the model will predict where the commas go,

模型會預測逗點在哪裡、 名詞和動詞在哪裡。

where the nouns and verbs are.

But he actually got a state-of-the-art sentiment analysis classifier out of it.

但他實際上卻從中得到了 一個先進的情感分析分類器。

This model could tell you if a review was positive or negative.

這個模型能告訴你 評論是正評或負評。

I mean, today we are just like, come on, anyone can do that.

現今我們會說，拜託， 任何人都做得到。

But this was the first time that you saw this emergence,

但這是你初次看到它出現，

this sort of semantics that emerged from this underlying syntactic process.

從這背後的句法過程中出現了語義。

And there we knew, you've got to scale this thing,

我們就知道我們得把它做大， 看看會如何發展下去。

you've got to see where it goes.

主持人：這也能協助解釋

CA: So I think this helps explain

讓每個人看到它都覺得困惑的謎團，

the riddle that baffles everyone looking at this,

因為這些東西被描述為預測機器。

because these things are described as prediction machines.

但我們看到它們的感覺是……

And yet, what we're seeing out of them feels ...

it just feels impossible that that could come from a prediction machine.

感覺這些事不可能是

預測機器會做的事， 你剛才展示的那些事。

Just the stuff you showed us just now.

And the key idea of emergence is that when you get more of a thing,

而「出現」的重點是，

有更多同樣的東西時， 不同的東西突然出現了。

suddenly different things emerge.

It happens all the time, ant colonies, single ants run around,

這很常見，螞蟻群， 單一隻螞蟻跑來跑去，

把夠多螞蟻放在一起，

when you bring enough of them together,

產生的螞蟻群會呈現出 全新的不同行為。

you get these ant colonies that show completely emergent, different behavior.

Or a city where a few houses together, it's just houses together.

或城市，幾間房子放在一起 就只是房子放在一起。

但當房子的數量增加，

But as you grow the number of houses,

things emerge, like suburbs and cultural centers and traffic jams.

就會有東西出現，

會有近郊、文化中心，和交通阻塞。

Give me one moment for you when you saw just something pop

告訴我，有哪個時刻是 你看到有什麼東西跳出來，

that just blew your mind

讓你很驚奇，你完全沒有預期到。

that you just did not see coming.

GB: Yeah, well,

講者：好，嗯，

在 ChatGPT 裡試著做 四十位數的加法——

so you can try this in ChatGPT, if you add 40-digit numbers --

CA: 40-digit?

主持人：四十位數？ 講者：四十位數，模型會做，

GB: 40-digit numbers, the model will do it,

這表示它真的學會了 一個內部迴路來做這件事。

which means it's really learned an internal circuit for how to do it.

And the really interesting thing is actually,

有趣的是，

如果你實際上要它 把四十位數加上三十五位數，

if you have it add like a 40-digit number plus a 35-digit number,

it'll often get it wrong.

它通常會算錯。

And so you can see that it's really learning the process,

可以看到它在學習這個過程， 但還沒有做到完全一般化。

but it hasn't fully generalized, right?

It's like you can't memorize the 40-digit addition table,

不可能記住四十位數的加法表， 那比宇宙中的原子還多。

that's more atoms than there are in the universe.

所以它得學會一般化的原則，

So it had to have learned something general,

但它還沒有完全學會：

but that it hasn't really fully yet learned that,

喔，我可以把這個給一般化， 來將任何長度的任何數字加總。

Oh, I can sort of generalize this to adding arbitrary numbers

of arbitrary lengths.

主持人：所以狀況就是 你讓它的規模擴大，

CA: So what's happened here

is that you've allowed it to scale up

and look at an incredible number of pieces of text.

去看超大量的文字。

And it is learning things

而它開始學會你本來 不知道它能學會的東西。

that you didn't know that it was going to be capable of learning.

GB Well, yeah, and it’s more nuanced, too.

講者：是的，且也還更微妙。

So one science that we’re starting to really get good at

我們漸漸變得很擅長一種科學，

is predicting some of these emergent capabilities.

就是去預測這些新產出的能力。

And to do that actually,

要做到這一點，我認為 工程品質在這領域裡被低估了。

one of the things I think is very undersung in this field

is sort of engineering quality.

Like, we had to rebuild our entire stack.

我們得要重新建構技術堆疊。

When you think about building a rocket,

在建造火箭時，每個公差 都必須要非常小。

every tolerance has to be incredibly tiny.

Same is true in machine learning.

機器學習也是，技術堆疊的 每一塊都要妥善設計，

You have to get every single piece of the stack engineered properly,

and then you can start doing these predictions.

接著你就可以開始做預測。

There are all these incredibly smooth scaling curves.

所有這些非常平滑的擴展曲線

They tell you something deeply fundamental about intelligence.

提供一些關於智慧的基礎資訊。

所有這些曲線都放在 我們的 GPT-4 部落格文章中。

If you look at our GPT-4 blog post,

you can see all of these curves in there.

現在我們可以開始做預測，比如，

And now we're starting to be able to predict.

So we were able to predict, for example, the performance on coding problems.

處理編碼問題的表現。

We basically look at some models

基本上我們看的是 小一萬倍或一千倍的模型。

that are 10,000 times or 1,000 times smaller.

And so there's something about this that is actually smooth scaling,

即使是在早期，也有些徵兆 可以看出很平順的擴展。

even though it's still early days.

CA: So here is, one of the big fears then,

主持人：來談談它 帶來的重大恐懼之一。

that arises from this.

If it’s fundamental to what’s happening here,

若這個現象是很基本的， 當你擴展規模，

that as you scale up,

things emerge that

會有新東西出現，

you can maybe predict in some level of confidence,

也許你可以預測到， 也有某種程度的信心，

but it's capable of surprising you.

但它有讓你驚奇的能力。

Why isn't there just a huge risk of something truly terrible emerging?

為什麼不用擔心 會有很可怕的東西出現？

GB: Well, I think all of these are questions of degree

講者：我認為這些都是程度、 規模，時機點的問題。

and scale and timing.

我認為大家忽略了一點： 和世界的整合也是非常新興的，

And I think one thing people miss, too,

is sort of the integration with the world is also this incredibly emergent,

也很強大。

sort of, very powerful thing too.

這也是我們認為逐步部署 十分重要的原因之一。

And so that's one of the reasons that we think it's so important

to deploy incrementally.

從看這場演說可以看出，我十分重視

And so I think that what we kind of see right now, if you look at this talk,

提供真正高品質的回饋。

a lot of what I focus on is providing really high-quality feedback.

現今我們做的工作任務 是可以被檢視的對吧？

Today, the tasks that we do, you can inspect them, right?

很容易可以去檢視 數學問題，接著說：

It's very easy to look at that math problem and be like, no, no, no,

不，不，不，機器， 正確答案應該是七。

machine, seven was the correct answer.

但，就連幫一本書做摘要，

But even summarizing a book, like, that's a hard thing to supervise.

那也很難監督，你要怎麼 知道這書摘做得好不好？

Like, how do you know if this book summary is any good?

你得讀完整本書，沒人想要這差事。

You have to read the whole book.

No one wants to do that.

（笑聲）

(Laughter)

And so I think that the important thing will be that we take this step by step.

所以我認為，重要的是 我們要按步就班。

And that we say, OK, as we move on to book summaries,

比如，好，當我們開始做書摘時，

we have to supervise this task properly.

我們得妥當監督這個工作任務， 要建立這些機器的追蹤記錄，

We have to build up a track record with these machines

that they're able to actually carry out our intent.

確保它們實際上有實現我們的意圖。

我認為我們得要創造出更好的、

And I think we're going to have to produce even better, more efficient,

更有效、更可靠的方法，

more reliable ways of scaling this,

在擴展規模時能 讓機器和你的觀念一致。

sort of like making the machine be aligned with you.

CA: So we're going to hear later in this session,

主持人：等一下的演說中 會聽到有些評論家說

there are critics who say that,

you know, there's no real understanding inside,

它的內部其實沒有真正的了解，

the system is going to always --

這個系統永遠都——

we're never going to know that it's not generating errors,

我們永遠不會知道它沒有產生錯誤、

that it doesn't have common sense and so forth.

它沒有常識等等。

Is it your belief, Greg, that it is true at any one moment,

葛雷格，你真的相信， 真的在任何時候，

but that the expansion of the scale and the human feedback

規模的擴大和人類的回饋，

that you talked about is basically going to take it on that journey

你剛才談到的那些，基本上真的可以

引領它走上那條路，得到比如

of actually getting to things like truth and wisdom and so forth,

真相、智慧等等， 你能信心滿滿地相信嗎？

with a high degree of confidence.

Can you be sure of that?

你能確定嗎？

GB: Yeah, well, I think that the OpenAI, I mean, the short answer is yes,

講者：對，我認為 OpenAI…… 簡短的回答是「是」，

I believe that is where we're headed.

我相信我們正朝那裡前進。

And I think that the OpenAI approach here has always been just like,

我認為 OpenAI 的方法一直都是

讓現實狠狠打在你臉上。

let reality hit you in the face, right?

就彷彿這個領域是打破承諾的領域，

It's like this field is the field of broken promises,

專家們都在說 X 將來會發生， 它的運作方式就是 Y。

of all these experts saying X is going to happen, Y is how it works.

有人說還要七十年 神經網路才可能可行。

People have been saying neural nets aren't going to work for 70 years.

他們還沒說對，他們可能會說對，

They haven't been right yet.

也許需要的是七十一年之類的。

They might be right maybe 70 years plus one

or something like that is what you need.

但我認為我們的方法一直是 你得要推向這項科技的極限，

But I think that our approach has always been,

you've got to push to the limits of this technology

看它實際運作，因為這樣你才會知道

to really see it in action,

喔，我們可以這樣轉向新的範式。

because that tells you then, oh, here's how we can move on to a new paradigm.

我們還沒把所有潛力開發出來。

And we just haven't exhausted the fruit here.

主持人：你採取的立場頗有爭議性，

CA: I mean, it's quite a controversial stance you've taken,

正確的做法是把它公開，

that the right way to do this is to put it out there in public

利用這一切，

and then harness all this, you know,

而不只是由你的團隊來給回饋，

instead of just your team giving feedback,

the world is now giving feedback.

全世界都在給回饋。

But ...

但……

If, you know, bad things are going to emerge,

如果有壞事會出現，它就在那裡了。

it is out there.

當你以非營利機構形式成立 OpenAI 時我聽到的原始故事是

So, you know, the original story that I heard on OpenAI

when you were founded as a nonprofit,

well you were there as the great sort of check on the big companies

你是個偉大的監督者

監管大型企業用人工智慧做些 未知且可能很邪惡的事情，

doing their unknown, possibly evil thing with AI.

And you were going to build models that sort of, you know,

而你要創造模型來

以某種方式要求他們 負責，且如果需要，

somehow held them accountable

and was capable of slowing the field down, if need be.

也能讓這個領域緩一下。

Or at least that's kind of what I heard.

至少那是我聽到的， 但實際發生的狀況卻相反，

And yet, what's happened, arguably, is the opposite.

你把 GPT 釋出，

That your release of GPT, especially ChatGPT,

特別是 ChatGPT，

在科技業引起巨大的震撼，

sent such shockwaves through the tech world

現在 Google 和 Meta 等等 都在拼命追趕，

that now Google and Meta and so forth are all scrambling to catch up.

And some of their criticisms have been,

有些批評是

你強迫我們公開這些，

you are forcing us to put this out here without proper guardrails or we die.

沒有妥當的保護措施， 但不做就是死路。

你要怎麼……

You know, how do you, like,

make the case that what you have done is responsible here and not reckless.

證明你所做的是負責的事， 而非魯莾行事？

GB: Yeah, we think about these questions all the time.

講者：我們一天到晚 在思索這些問題。

Like, seriously all the time.

真的是一天到晚。

And I don't think we're always going to get it right.

我不認為我們能一直不犯錯。

But one thing I think has been incredibly important,

但有一點我認為非常重要，

打從最初我們在思考要如何打造

from the very beginning, when we were thinking

about how to build artificial general intelligence,

能讓全人類受惠的人工通用智慧，

actually have it benefit all of humanity,

要怎麼做到啊？

預設的計畫是在背底裡秘密建造，

like, how are you supposed to do that, right?

And that default plan of being, well, you build in secret,

做出超強大的東西，

you get this super powerful thing,

想通如何處理安全性， 按下「啟動」，希望你做對了。

and then you figure out the safety of it and then you push “go,”

and you hope you got it right.

我不知道如何執行那個計畫。

I don't know how to execute that plan.

也許別人知道。

Maybe someone else does.

但我總覺得那很嚇人， 感覺就是不對。

But for me, that was always terrifying, it didn't feel right.

我認為這個替代方法，

And so I think that this alternative approach

是我唯一知道的其他途徑，

is the only other path that I see,

which is that you do let reality hit you in the face.

也就是讓現實狠狠打在你臉上。

And I think you do give people time to give input.

我認為你確實 給大家時間提供建議，

You do have, before these machines are perfect,

在這些機器達到完美之前，

before they are super powerful, that you actually have the ability

在它們超強大之前，

你能夠先看到它們實際運作。

to see them in action.

And we've seen it from GPT-3, right?

在 GPT-3 就可以看到， 我們那時很害怕大家

GPT-3, we really were afraid

that the number one thing people were going to do with it

拿它做的第一件事就是 產生假消息試圖影響選舉。

was generate misinformation, try to tip elections.

Instead, the number one thing was generating Viagra spam.

結果第一件事反而是 產生威而剛垃圾郵件。

(Laughter)

（笑聲）

CA: So Viagra spam is bad, but there are things that are much worse.

主持人：威而剛垃圾郵件不好， 但還有更糟糕的。

來做個思想實驗。

Here's a thought experiment for you.

Suppose you're sitting in a room,

假設你坐在房間裡，桌上有個盒子。

there's a box on the table.

You believe that in that box is something that,

你認為在盒子裡的東西有可能，

there's a very strong chance it's something absolutely glorious

非常有可能是很棒的東西，

that's going to give beautiful gifts to your family and to everyone.

是給你的家人和每個人的美好禮物。

But there's actually also a one percent thing in the small print there

但實際上也有 1% 非常小的

字樣寫著「潘朵拉」。

that says: “Pandora.”

And there's a chance

有可能

that this actually could unleash unimaginable evils on the world.

打開它就會將無法想像的 邪惡釋放到這個世界上。

Do you open that box?

你會打開盒子嗎？

GB: Well, so, absolutely not.

講者：絕對不會。

I think you don't do it that way.

我認為不該那樣做。

老實說，

And honestly, like, I'll tell you a story that I haven't actually told before,

讓我說個故事，我從來沒有講過：

which is that shortly after we started OpenAI,

OpenAI 剛成立不久，我記得

I remember I was in Puerto Rico for an AI conference.

我在波多黎各參加人工智慧 大會，我坐在房間裡，

I'm sitting in the hotel room just looking out over this wonderful water,

向外眺望美麗的水景， 大家都很開心。

all these people having a good time.

花點時間想想看，如果你能選擇

And you think about it for a moment,

if you could choose for basically that Pandora’s box

讓那個潘朵拉的盒子

to be five years away

出現在五年後，

or 500 years away,

或者五百年後，

which would you pick, right?

你會挑哪個？

On the one hand you're like, well, maybe for you personally,

一方面，你會想，就你 個人來說五年是比較好的。

it's better to have it be five years away.

But if it gets to be 500 years away and people get more time to get it right,

但如果是五百年後， 大家有更多時間把它做對，

你會挑哪個？

which do you pick?

And you know, I just really felt it in the moment.

在那一刻我真的很有感， 當然，五百年後再做。

I was like, of course you do the 500 years.

許多人……我兄弟那時還在軍中，

My brother was in the military at the time

and like, he puts his life on the line in a much more real way

他冒著生命危險，用著比 我們任何人都更真實的方式，

than any of us typing things in computers

我們那時只是用電腦打字， 開發這項科技。

and developing this technology at the time.

And so, yeah, I'm really sold on the you've got to approach this right.

是的，我真心認同必須要把它做對。

But I don't think that's quite playing the field as it truly lies.

但我不認為那符合實際情況。

Like, if you look at the whole history of computing,

想想整個計算史，

I really mean it when I say that this is an industry-wide

我是真的很認真的說，

這是整個產業的，或甚至

or even just almost like

幾乎是整個人類科技發展的轉變。

a human-development- of-technology-wide shift.

And the more that you sort of, don't put together the pieces

你越是不去把既有的拼圖給拼起來，

that are there, right,

we're still making faster computers,

我們仍然在發展更快的電腦， 仍然在改善演算法，

we're still improving the algorithms, all of these things, they are happening.

所有這些都在發生。

And if you don't put them together, you get an overhang,

如果你不把它們 拼湊起來，就會有危險，

意即，如果有人做到，一旦 有人設法把迴路連結起來，

which means that if someone does,

or the moment that someone does manage to connect to the circuit,

突然出現非常強大的東西， 沒有人有時間調整，

then you suddenly have this very powerful thing,

no one's had any time to adjust,

誰知道你能得到 什麼樣的安全預防措施？

who knows what kind of safety precautions you get.

And so I think that one thing I take away

我認為我得到的啟示是，

is like, even you think about development of other sort of technologies,

想想其他科技的發展 也是一樣，比如核武，

think about nuclear weapons,

people talk about being like a zero to one,

大家講得好像是從零到一， 好像是改變人類能做什麼。

sort of, change in what humans could do.

But I actually think that if you look at capability,

但如果去看「能力」， 隨時間的發展是很平順的。

it's been quite smooth over time.

And so the history, I think, of every technology we've developed

看看我們所發展出的 每一項科技的歷史，

has been, you've got to do it incrementally

都是你得按部就班， 必須要想辦法管理它，

and you've got to figure out how to manage it

強化它的每一刻都要能管理。

for each moment that you're increasing it.

CA: So what I'm hearing is that you ...

主持人：我聽到的是你……

the model you want us to have

你希望我們擁有的模型， 我們已經孕育出這個不凡的孩子

is that we have birthed this extraordinary child

that may have superpowers

可能有超能力

能帶人類進入全新的境界。

that take humanity to a whole new place.

It is our collective responsibility to provide the guardrails

我們所有人有共同的責任

為這個孩子提供護欄，

for this child

to collectively teach it to be wise and not to tear us all down.

一起教導它要明智，而不要

摧毀我們所有人， 模型基本上是這樣嗎？

Is that basically the model?

GB: I think it's true.

講者：我認為是。

And I think it's also important to say this may shift, right?

我還要說有一點也很重要： 這可能會轉變。

We've got to take each step as we encounter it.

我們必須要遇到一步就走一步。

And I think it's incredibly important today

現今，非常重要的是，

that we all do get literate in this technology,

我們都要熟悉這項科技，

figure out how to provide the feedback,

想辦法提供回饋， 決定我們想要它如何。

decide what we want from it.

And my hope is that that will continue to be the best path,

這希望這條路能一直是最佳的路，

but it's so good we're honestly having this debate

很高興我們能坦誠做這些辯論，

because we wouldn't otherwise if it weren't out there.

因為如果它沒有誕生， 也不會有這些辯論了。

CA: Greg Brockman, thank you so much for coming to TED and blowing our minds.

主持人：葛雷格‧布洛克曼， 謝謝你來 TED 讓我們大開眼界。

(Applause)

（掌聲）