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

翻译人员: Yip Yan Yeung 校对人员: Grace Man

We started OpenAI seven years ago

七年前，我们成立了 OpenAI，

because we felt like something really interesting was happening in AI

因为我们认为，AI 领域中 正在上演一些很有趣的事，

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

我们想助其走上正轨。

It's honestly just really amazing to see

这个领域自此取得的进步

how far this whole field has come since then.

着实令人咋舌。

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

我也十分欣慰看到像网友雷蒙德 （Raymond）这样的人们

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

使用我们正在打造的技术， 还有很多人

for so many wonderful things.

将其用于许多美妙的用途。

We hear from people who are excited,

我们看到有人很兴奋，

we hear from people who are concerned,

我们看到有人很忧虑，

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

我们看到有人两者兼有。

And honestly, that's how we feel.

说实话，这也是我们的心情。

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

关键是感觉我们 即将进入一个历史性时期，

where we as a world are going to define a technology

全球各地的人们将要 共同努力定义一项技术，

that will be so important for our society going forward.

而它对于我们社会 未来的走向至关重要。

And I believe that we can manage this for good.

我相信我们有能力做到。

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

今天，我想为大家介绍 这项技术的现况

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

和其背后我们所珍视的 一些设计原则。

So the first thing I'm going to show you

我要介绍的第一点是

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

为 AI （人工智能）， 而不是人类打造一个工具，

rather than building it for a human.

是什么样的。

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

我们现在有了一个新的 DALL-E 模型，可以生成图片，

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

它的呈现方式是 ChatGPT 直接替你调用的应用程序。

And you can do things like ask, you know,

你可以执行这样的操作， 比如问一问

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

这场 TED 结束了以后 吃点什么好呢？再画张图吧。

(Laughter)

（笑声）

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

你能看到 ChatGPT 为你提供的这种

构思和发挥创意的往复过程 以及为你处理好所有细节的能力。

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

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

看这些回答，已经不仅仅是 关于一顿饭的想法了，

but a very, very detailed spread.

而是非常、 非常详细的大餐细节。

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

我们来看看能得出什么结果。

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

但在这种情况下，ChatGPT 不止生成了图片，

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

它不仅生成了文本， 还生成了图片。

And that is something that really expands the power

这增强了它为了执行你的意图 而替你完成操作的能力。

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

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

我想提一点， 这都是现场演示。

This is all generated by the AI as we speak.

这都是在我们的言语之间 由 AI 生成的。

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

其实我都不知道 我们会看到什么样的结果。

This looks wonderful.

看起来非常棒。

(Applause)

（掌声）

I'm getting hungry just looking at it.

看着它我都饿了。

Now we've extended ChatGPT with other tools too,

我们也把其他工具 接入了 ChatGPT，

for example, memory.

比如记忆。

You can say "save this for later."

你可以输入 “保存，稍后再用”。

And the interesting thing about these tools

这种工具的有趣之处在于

is they're very inspectable.

检查它们是非常容易的。

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

你可以看到这个弹出插件 写着“调用 DALL-E 应用”。

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

顺带提一句， 你马上就能用上这个功能了，

每位 ChatGPT 用户， 在几个月内就能用上。

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

如果你看一下它背后的技术， 看看它到底进行了什么操作，

was write a prompt just like a human could.

其实就是像人那样 写了个提示（prompt）。

And so you sort of have this ability to inspect

你可以检查

how the machine is using these tools,

机器是如何使用这些工具的，

which allows us to provide feedback to them.

这样我们就可以为其提供反馈。

Now it's saved for later,

我们把它存了下来，之后再用，

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

我来给各位展示一下 如何利用这些信息，

and to integrate with other applications too.

调用其他应用。

You can say,

你可以这么说：

“Now make a shopping list for the tasty thing

“请你针对我之前提到的美食，

制作一份购物清单。”

I was suggesting earlier.”

And make it a little tricky for the AI.

我们来给 AI 增加一点难度。

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

“再发条推特 推给所有 TED 观众。”

(Laughter)

（笑声）

So if you do make this wonderful, wonderful meal,

如果你真的做了这顿美餐，

I definitely want to know how it tastes.

我很想知道味道如何。

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

你可以看到， ChatGPT 选用了各种工具，

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

而我并不需要明确告诉它 什么情况下该用什么工具。

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

我认为这体现了一个 看待用户界面的新方式。

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

我们很习惯这么思考： 我们有了一些应用，

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

在应用之间点击， 在应用之间复制、粘贴，

and usually it's a great experience within an app

通常在单个应用内体验良好，

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

只要你大概了解菜单 和所有选项就行。

Yes, I would like you to.

是的，就这么做。 （键入提示）

Yes, please.

是的，麻烦了。（键入提示）

Always good to be polite.

礼貌一点总是好的。

(Laughter)

（笑声）

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

有了统一的语言界面， 再加上工具，

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

AI 就能从你手中提取各种细节。

So you don't have to be the one

你不需要一字一句地说出

who spells out every single sort of little piece

of what's supposed to happen.

该产生什么结果。

And as I said, this is a live demo,

如我所说，这是个现场演示，

so sometimes the unexpected will happen to us.

所以，有时会发生意外。

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

我们来看一看 这份 Instracart 购物清单。

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

你可以看到，我们向 Instacart 发送了一系列商品。

Here's everything you need.

就是你所需的所有商品。

And the thing that's really interesting

很有趣的一点是

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

传统 UI （用户界面） 依旧是很有价值的，对吧？

If you look at this,

看一下这个界面，

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

你还是可以点上去， 修改实际数量。

And that's something that I think shows

我认为这就体现了

that they're not going away, traditional UIs.

传统 UI 并没有消失。

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

只是我们有了一个全新、 增强的方式来创建它。

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

我们有了一条草拟好的推特， 等待我们的审阅，

which is also a very important thing.

这也是一件很重要的事。

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

我们可以点击“运行”， 然后就成了这样，

我们是管理者，可以检查，

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

要是愿意，我们 也可以修改 AI 的成果。

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

在这场演讲之后， 你自己也可以试一下。

And there we go.

有了。

Cool.

酷。

Thank you, everyone.

谢谢大家。

(Applause)

（掌声）

So we’ll cut back to the slides.

我们再回到片子。

Now, the important thing about how we build this,

我们是怎么把它做出来的， 其中的关键之处

it's not just about building these tools.

不仅仅是 怎么做出这些工具的问题。

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

还要教会 AI 使用它们。

Like, what do we even want it to do

我们问了这些高层级的问题，

when we ask these very high-level questions?

到底想让它做些什么呢？

And to do this, we use an old idea.

要回答这个问题， 我们用了一个老方法。

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

你可以翻阅 艾伦·图灵（Alan Turing）

1950 年关于图灵测试的 论文，其中写道，

you'll never program an answer to this.

你是编不出一个程序 来回答这个问题的，

Instead, you can learn it.

而是通过学习。

You could build a machine, like a human child,

你可以做一台机器， 就像人类儿童那样，

and then teach it through feedback.

通过不断反馈指导它。

Have a human teacher who provides rewards and punishments

在它尝试、做出对的事 和不对的事的时候，

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

有那么一位人类老师 给它奖励和惩罚。

And this is exactly how we train ChatGPT.

这就是我们训练 ChatGPT 的方式。

It's a two-step process.

这个过程有两步。

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

首先，我们做了一台 图灵所称的“子计算机”，

采用了无监督学习过程。

through an unsupervised learning process.

We just show it the whole world, the whole internet

我们给它看了整个世界， 整个互联网，

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

然后要求：“这是你之前没见过的内容， 请预测接下来该输出什么文本。”

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

这个过程让它掌握了 各种神奇的技能。

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

比如，你给它一个数学问题，

the only way to actually complete that math problem,

要做出这道数学题的唯一方式，

to say what comes next,

输出接下来的答案，

that green nine up there,

也就是这个绿色的 9，

is to actually solve the math problem.

只有通过真正解出这个数学问题。

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

但其实我们也得进行第二步，

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

即指导 AI 使用这些技能。

And for this, we provide feedback.

为了达到这个目的， 我们会给予反馈。

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

我们让 AI 尝试做几件事， 给我们提几个建议，

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

然后人类会为其评分， 表示：“这个比那个好。”

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

这不仅强化了 AI 刚说的那个答案，

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

更重要的是，还强化了 AI 生成这个答案的整个过程。

And this allows it to generalize.

这样 AI 就可以泛化。

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

它可以指导、推测你的意图，

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

在没有出现过的情境下应用，

that it hasn't received feedback.

且是在没有接受过 反馈的情况下。

Now, sometimes the things we have to teach the AI

有时我们需要教给 AI 的东西

are not what you'd expect.

不是你意料之中的那些。

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

比如，我们第一次把 GPT-4 给可汗学院看的时候，

they said, "Wow, this is so great,

他们说：“哇，太棒了，

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

我们可以给学生教点好东西了。

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

只是有一个问题，它不会 检查学生的数学做得对不对。

If there's some bad math in there,

如果数学做得不对，

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

它会愉快地默认一加一等于三， 然后就这么运行下去。”

So we had to collect some feedback data.

于是我们就得收集一些反馈数据。

Sal Khan himself was very kind

萨尔·可汗（Sal Khan） 本人很不错，

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

他贡献了 20 个小时的私人时间 和我们的团队一起为计算机

alongside our team.

提供反馈。

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

在几个月的时间里， 我们已经可以这么教会 AI：

"Hey, you really should push back on humans

“嘿，你真的得 在这种特定的场合下，

in this specific kind of scenario."

反驳人类。”

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

我们已经以这种方式 对模型进行了大量改良。

And when you push that thumbs down in ChatGPT,

你在 ChatGPT 里 点击“踩”的时候，

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

其实是给我们的团队 发送了一个信号：

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

“这块儿有缺陷， 你们得收集一些反馈。”

And so when you do that,

你这么点击的时候，

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

这就会是我们真正倾听用户、

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

确保我们创造的东西 会对每个人更有用的方式。

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

提供高质量的反馈是件难事。

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

想象一下让孩子打扫房间，

if all you're doing is inspecting the floor,

如果你做的唯一一件事 只是检查地面，

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

就不会知道你是不是只教会了他们 把所有玩具塞进衣柜里。

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

顺带说一句，这是一张 由 DALL-E 生成的美图。

And the same sort of reasoning applies to AI.

同样的道理也适用于 AI。

As we move to harder tasks,

如果要执行更困难的任务，

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

我们就得提高 提供高质量反馈的能力。

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

但就这一点， AI 本身就乐意为之。

It's happy to help us provide even better feedback

它很乐意持续帮助我们 提供更好的反馈，

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

提高我们监督计算机的能力。

And let me show you what I mean.

我来给你看一下 我说的是什么意思。

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

比如，你可以问 GPT-4 这么一个问题：

of how much time passed between these two foundational blogs

这两篇关于无监督学习 和基于人类反馈学习的

on unsupervised learning

基础性博客之间

and learning from human feedback.

相差了多长时间，

And the model says two months passed.

模型说两个月。

But is it true?

但说的是真的吗？

Like, these models are not 100-percent reliable,

这些模型不是 100% 可靠的，

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

虽然每次我们提供了反馈以后， 它们都会优化。

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

但其实我们可以利用 AI 进行事实核查。

And it can actually check its own work.

它都可以核查自己的结果。

You can say, fact-check this for me.

你可以要求它， 帮我核对一下是否属实。

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

在这个例子中，我给了 AI 一个新工具。

This one is a browsing tool

这是个浏览工具，

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

模型可以搜索查询，点进网页。

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

在搜索的过程中， 它会写出完整的思考过程。

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

它会写出，我要搜这个， 然后它真的就去搜这个了。

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

然后它会找到 发表时间和搜索结果。

It then is issuing another search query.

然后它会再进行一次搜索。

It's going to click into the blog post.

点进博客。

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

你也可以做这些事， 但是这很冗长乏味。

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

这是人类不太想做的事。

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

掌控一切要有趣得多，

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

当管理者要有趣得多， 只要你想，你也可以

triple-check the work.

再查一遍结果。

And out come citations

你可以看到这些引用，

so you can actually go

这样你就可以

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

非常轻松地验证 整条逻辑中的任何环节。

And it actually turns out two months was wrong.

结果“两个月”是错误的。

Two months and one week,

两个月加一周，

that was correct.

这才是对的。

(Applause)

（掌声）

And we'll cut back to the side.

我们再回到片子。

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

我觉得整个过程中很有趣的是

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

这个过程是人类和 AI 之间 进行的多步合作。

Because a human, using this fact-checking tool

因为人类 利用这个事实核查工具，

is doing it in order to produce data

为的是产生数据，

for another AI to become more useful to a human.

让 AI 对人类更加有用。

And I think this really shows the shape of something

我认为这充分展示了

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

未来会越来越常见的一种情况：

where we have humans and machines kind of very carefully

在一个具体的问题和 我们设想的解决方法中，

and delicately designed in how they fit into a problem

人类和计算机的定位分工 都会被非常谨慎、精心地安排好。

and how we want to solve that problem.

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

我们得保证人类 提供管理、全局视角、

the feedback,

反馈，

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

计算机以一种可被检查、

and trustworthy.

值得信任的方式运行。

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

通过合作，我们可以 创造出更值得信任的计算机。

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

我认为随着时间的推移， 如果我们让这个过程走上正轨，

we will be able to solve impossible problems.

我们就能解决 不可能被解决的问题。

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

为了让大家理解我说的 “不可能被解决”是什么意思，

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

我认为，我们可以重新思考

of how we interact with computers.

我们与计算机互动的 几乎每一个方面。

For example, think about spreadsheets.

比如，表格。

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

它们从大约 40 年前 VisiCalc 时期就长这样了。

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

我觉得过了这么多年 也没发生什么变化。

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

这张表格记录了 arXiv 上过去 30 年里

for the past 30 years.

所有与 AI 有关的论文。

There's about 167,000 of them.

大约有 167000 篇。

And you can see there the data right here.

你可以看到这些数据。

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

但我想给你展示一下 ChatGPT 是如何分析这样的数据集的。

So we can give ChatGPT access to yet another tool,

我们再给 ChatGPT 加一个工具，

this one a Python interpreter,

这是个 Python 解释器，

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

它可以像数据科学家那样运行代码。

And so you can just literally upload a file

你只需要上传一个文件，

and ask questions about it.

问相关的问题。

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

它很有用，可以读取 文件名然后这样运行：

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

“哦，这是个 CSV 文件， 逗号分隔值文件（CSV），

"I'll parse it for you."

我来给你解析一下。”

The only information here is the name of the file,

然后留下的信息只有文件名、

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

你刚看到的列名和实际数据。

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

从这些信息中，它可以推测出 这些列都代表着什么意思。

Like, that semantic information wasn't in there.

这里并没有语义信息。

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

所以它就得搜罗 它知道的所有信息，

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

“哦，arXiv 是 人们提交论文的网站，

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

所以这些条目指的是论文， 而且这些数都是整数，

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

所以这个数值就是 论文作者的数量。”

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

等等操作， 这些都是人类要做的事，

and the AI is happy to help with it.

而 AI 乐意效劳。

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

我都不知道我该问什么。

So fortunately, you can ask the machine,

好就好在你可以问计算机：

"Can you make some exploratory graphs?"

“你可以帮我做一些 探索性图表吗？”

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

同样地，这是个非常高层级的指示， 背后还有很多深层的涵义。

But I don't even know what I want.

但我都不知道自己想要什么。

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

AI 就得推测我可能想要什么。

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

于是，它就给出了几个 我认为还不错的点子。

So a histogram of the number of authors per paper,

每篇论文作者数量的直方图、

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

每年论文的时间序列、 论文标题的词云。

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

我觉得这些都还挺有趣的。

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

妙就妙在它真的可以做出来。

Here we go, a nice bell curve.

来了，精美的正态分布曲线。

You see that three is kind of the most common.

你可以看到，3 是最常见的数量。

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

让它给每年的论文 都做这么一张漂亮的图。

Something crazy is happening in 2023, though.

2023 年有点奇葩。

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

看起来我们经历了一个 指数增长然后跳水。

What could be going on there?

发生了什么事？

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

顺带提一句，这些都是 Python 代码，你是可以检查的。

And then we'll see word cloud.

我们来看词云。

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

你可以看到这些标题中的奇妙内容。

But I'm pretty unhappy about this 2023 thing.

我对这 2023 年的现象 不太满意。

It makes this year look really bad.

这一年看起来太差了。

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

当然，今年还没过完。

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

所以我要把问题抛回给计算机。

[Waitttt that's not fair!!!

[等等！这不公平！

2023 isn't over.

2023 年还没过完呢。

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

2022 年有多大比例的论文是在 4 月 13 日以前发表的？]

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

我就把 4 月 13 日 当作截止日期。

Can you use that to make a fair projection?

[你可以用这个结论 生成一个公平的图表吗？]

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

我们来看看，这题有点难度。

(Laughter)

（笑声）

So you know,

同样地，我这么做是因为 我对计算机有更大的企图。

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

I really wanted it to notice this thing,

我想让它注意到这一点，

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

虽然对它来说，机智地猜出

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

我这样的意图可能有点太难了。

But I inject my intent,

但我表明了我的意图，

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

又为它提供了一些指导。

And under the hood,

如果看它底层的运作，

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

AI 只是又写了一遍代码， 所以如果你想检查一下它在干什么，

it's very possible.

是很容易做到的。

And now, it does the correct projection.

这次，它生成了正确的图表。

(Applause)

（掌声）

If you noticed, it even updates the title.

你仔细看的话， 它还更新了标题。

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

我可没要求它这么做， 但它知道我想要什么。

Now we'll cut back to the slide again.

我们再回到片子。

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

这张片子讲了一个故事， 表示我认为我们……

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

我们未来使用这项技术的愿景。

A person brought his very sick dog to the vet,

一个人带着他病重的狗狗去看兽医，

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

兽医做了个愚蠢的决定，说： “我们就等着观望吧。”

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

如果他当时就这么信了兽医， 狗狗现在已经走了。

In the meanwhile, he provided the blood test,

与此同时，他把验血结果、

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

完整病历提交给了 GPT-4，

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

它说：“我不是个兽医， 你得去找专业人士咨询，

here are some hypotheses."

以下是一些假设。”

He brought that information to a second vet

他带着这些信息 去找了第二位兽医，

who used it to save the dog's life.

兽医利用这些信息 拯救了狗狗的生命。

Now, these systems, they're not perfect.

这些系统不是完美的。

You cannot overly rely on them.

你不能过度依赖它们。

But this story, I think, shows

但我认为这个故事体现了

that a human with a medical professional

人类和医疗专家，

and with ChatGPT as a brainstorming partner

再加上 ChatGPT 作为头脑风暴的搭档，

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

可以创造出一些前所未有的成果。

I think this is something we should all reflect on,

我认为，这是我们所有人 在考虑将这些系统融入世界之时

think about as we consider how to integrate these systems

该反思、考虑的一点。

into our world.

And one thing I believe really deeply,

我坚信的一点是，

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

让 AI 走上正道 需要每个人的参与。

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

参与决定该如何 让 AI 加入这个世界，

that's for setting the rules of the road,

参与规划路该怎么走，

for what an AI will and won't do.

参与规定 AI 该做什么，不该做什么。

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

这场演讲能给你带来的 一个收获就是

it's that this technology just looks different.

这项技术就是与众不同。

Just different from anything people had anticipated.

与人们意料中的任何事物都不一样。

And so we all have to become literate.

所以，我们都得对它有所了解。

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

说实话，这也是我们 发布 ChatGPT 的原因之一。

Together, I believe that we can achieve the OpenAI mission

我相信，我们能一起努力 达成 OpenAI 的使命，

of ensuring that artificial general intelligence

让通用人工智能

benefits all of humanity.

造福全人类。

Thank you.

谢谢。

(Applause)

（掌声）

(Applause ends)

（掌声结束）

Chris Anderson: Greg.

克里斯·安德森（Chris Anderson）： 格雷格（Greg）。

Wow.

哇。

I mean ...

我……

I suspect that within every mind out here

我怀疑在场的每个人

there's a feeling of reeling.

都有天旋地转的感觉。

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

我怀疑观看这场演讲的很多观众，

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

看到这些内容，都会心想：“天哪，

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

我得重新考虑一下 我日常操作中的每一件事。”

Like, there's just new possibilities there.

这是新的可能啊。

Am I right?

对吧？

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

谁认为得重新思考一下 自己行事的方式？

Yeah, I mean, it's amazing,

太神奇了，

but it's also really scary.

但也太可怕了。

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

我们来谈谈吧，格雷格。

I mean, I guess my first question actually is just

我想问的第一个问题就是

how the hell have you done this?

你们到底是怎么走到这一步的？

(Laughter)

（笑声）

OpenAI has a few hundred employees.

OpenAI 有几百名员工。

Google has thousands of employees working on artificial intelligence.

谷歌有几千名员工 专攻人工智能。

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

为什么是你们想出了这项

that shocked the world?

震惊世界的技术？

Greg Brockman: I mean, the truth is,

格雷格·布罗克曼 （Greg Brockman）：其实，

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

我们是站在巨人的肩膀上前进， 毋庸置疑。

如果你看看计算的进步、

If you look at the compute progress,

算法的进步、数据的进步，

the algorithmic progress, the data progress,

这些都是覆盖整个行业的。

all of those are really industry-wide.

但我认为，在 OpenAI 内部，

我们汲取过往的经验 做出了很多慎重的决定。

But I think within OpenAI,

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

首先就是直面现实。

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

我们对此苦思冥想：

And that we just thought really hard about like:

What is it going to take to make progress here?

在这个领域前进， 要付出什么代价？

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

我们尝试了很多无效的东西， 而你们只会看到有效的东西。

我认为最重要的一点就是 召集一个团队，

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

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

互不相同却要和谐地合作。

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

CA: 我们准备了水，拿过来吧。

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

我觉得我们得喝点水， 这是个令人津津乐道的话题。

But isn't there something also just about the fact

但是，在这些语言模型中，

that you saw something in these language models

你有没有发现

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

如果你继续对其投入、研发，

that something at some point might emerge?

到了某个节点， 会出现一些情况？

GB: Yes.

GB: 有。

And I think that, I mean, honestly,

老实说，我认为

I think the story there is pretty illustrative, right?

我刚说的故事已经 很直观了，对吧？

I think that high level, deep learning,

我认为从高层级来看， 深度学习，

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

我们一直知道 这就是我们想要的结果，

就得有个深度学习实验室， 那我们该怎么把它做出来呢？

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

我觉得早期我们并不知道答案。

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

我们尝试了很多，

We tried a lot of things,

一位同事训练了一个模型，

and one person was working on training a model

用于预测 亚马逊评论中的下一个字母，

to predict the next character in Amazon reviews,

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

他得出了这样一个结果—— 这是个处理句法的过程，

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

你想要的结果是 模型预测出逗号在什么位置，

where the nouns and verbs are.

名词和动词在什么位置。

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

但他其实由此做出了一个 最先进的情感分析分类器。

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

这个模型可以判定 这条评论是正面还是负面的。

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

放到现在，我们会觉得， 人人都能做到啊。

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

但这是你第一次见到这样的东西，

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

由底层的句法处理进程 产生的语义。

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

由此，我们知道， 必须得扩大这项技术的规模，

you've got to see where it goes.

必须得见证它未来的发展。

CA: 我认为这有助于解开

CA: So I think this helps explain

让观众困惑的谜题，

the riddle that baffles everyone looking at this,

因为这些模型被称为预测机。

because these things are described as prediction machines.

但我们看到它们就感觉……

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

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

就感觉预测机怎么可能 可以产生这样的结果。

Just the stuff you showed us just now.

就是你刚才给我们看的东西。

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

新兴事物的关键就在于 一件东西出现得越多，

suddenly different things emerge.

突然间就会出现很多别的东西。

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

这是个普遍现象，蚁群， 一只蚂蚁跑来跑去，

when you bring enough of them together,

如果你把足够多的蚂蚁放在一起，

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

就会出现蚁群，呈现出完全 没有见过的、异于寻常的行为。

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

一座城市，几栋房子建在一起， 就只是几栋建在一起的房子而已。

But as you grow the number of houses,

但如果你增加房子的数量，

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

就会出现别的东西， 比如郊区、文化中心、堵车。

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

你有没有经历过这么一个瞬间， 眼前突然出现了什么东西，

that just blew your mind

让你震撼，

that you just did not see coming.

出乎意料？

GB: Yeah, well,

GB: 嗯，

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

你可以在 ChatGPT 里试一试， 输入 40 位数字……

CA: 40-digit?

CA: 40 位？

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

GB: 40 位数， 模型是可以运算的，

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

也就是说它已经学会了 运算这些数字的内部逻辑。

And the really interesting thing is actually,

很有趣的一点是

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

如果你试试将 40 位数 和 35 位数相加，

it'll often get it wrong.

它通常会算错。

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

你可以看出， 它在学习这个过程，

but it hasn't fully generalized, right?

但还没有完全泛化，对吧？

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

你不可能记住 40 位加法表，

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

这比宇宙中的原子数都多。

所以它得学习一些通用的东西，

So it had to have learned something general,

但它还没有完全学会，

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

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

把它泛化到可以算出任意长度的

of arbitrary lengths.

任意数字之和。

CA: So what's happened here

CA: 所以眼下的情况是

is that you've allowed it to scale up

你拓展了它的能力，

and look at an incredible number of pieces of text.

能处理很大量的文本。

And it is learning things

它在学习

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

你都不知道它能学会的东西。

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

GB: 对，也有点微妙。

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

我们逐渐掌握的一项技能是

is predicting some of these emergent capabilities.

预测这些新兴的性能。

And to do that actually,

要做到这一点，

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

我认为这个领域中 有件不太受到重视的事，

is sort of engineering quality.

那就是工程质量。

Like, we had to rebuild our entire stack.

我们要重建整个技术栈。

When you think about building a rocket,

你可以想象一下造火箭，

every tolerance has to be incredibly tiny.

每一项的容错率都得非常小。

Same is true in machine learning.

机器学习也是一样的。

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

堆栈里的每一个元素 都得正确地编写，

and then you can start doing these predictions.

之后才可以进行预测。

There are all these incredibly smooth scaling curves.

这些都是非常平滑的规模曲线。

They tell you something deeply fundamental about intelligence.

它们体现了关于智能 非常基础的特质。

如果你去读一读我们的 GPT-4 博文，

If you look at our GPT-4 blog post,

你就会看到这些曲线。

you can see all of these curves in there.

我们现在可以开始预测了。

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

比如，我们可以预测 代码问题的性能。

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

We basically look at some models

我们会去看一些

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

小一万倍或一千倍的模型。

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

虽然还处于早期，

even though it's still early days.

它们的增长还是平滑的。

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

CA: 最大的恐惧

that arises from this.

来源于此。

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

如果一定会发生这种情况：

that as you scale up,

随着规模的扩大，

things emerge that

会出现你可能有一定信心预测，

you can maybe predict in some level of confidence,

but it's capable of surprising you.

但依然会吓你一跳的事。

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

那不是会有很大风险 会发生糟糕透顶的事吗？

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

GB: 我认为这些问题 都是有关程度、

and scale and timing.

规模和时机的。

And I think one thing people miss, too,

我认为大家还忽略了一点，

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

那就是与世界接轨 也是一件新兴的、

sort of, very powerful thing too.

很强大的事。

这也就是我们认为逐步部署 如此重要的原因之一。

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

to deploy incrementally.

你可以通过这场演讲看出

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

我关注的一大重点就是 提供非常高质量的反馈。

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

我们如今进行的任务， 你都可以进行检查，对吧？

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

你可以很轻松地看看那道数学题， 然后发现不对， 不对，不对，

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

机器，7 才是正确答案。

machine, seven was the correct answer.

但就算是总结一本书 都是难以监管的事。

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

你如何得知这篇摘要 是一篇好摘要呢？

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

你得看完整本书。

You have to read the whole book.

没人想干这事儿。

No one wants to do that.

（笑声）

(Laughter)

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

所以，我认为重要的是 我们要一步一步来。

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

我们在处理书籍摘要时，

we have to supervise this task properly.

我们得适当地 对这个任务进行监管。

We have to build up a track record with these machines

我们得建立起 这些计算机的历史记录，

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

确认它们是否能 真正执行我们的意图。

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

我觉得我们得创造出 更好、更高效、

more reliable ways of scaling this,

更可靠的拓展方式，

sort of like making the machine be aligned with you.

让计算机与你保持一致。

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

CA: 这场演讲之后会提到，

there are critics who say that,

有人评论说

you know, there's no real understanding inside,

其内部根本没有理解，

the system is going to always --

系统一直会……

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

我们看不到 它不会产生任何差错、

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

它没有常识等等。

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

格雷格，你会不会 有那么一瞬间认为这说的是对的，

but that the expansion of the scale and the human feedback

但又十分自信你刚谈到的 规模的扩大和人类的反馈

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

会将其带上

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

真正触及真相、 智慧等等的道路？

with a high degree of confidence.

Can you be sure of that?

你确定吗？

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

GB: 我觉得 OpenAI， 简而言之答案是肯定的，

I believe that is where we're headed.

我相信这就是我们前进的方向。

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

OpenAI 的方式一直就是

let reality hit you in the face, right?

被现实打脸，对吧？

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

这个领域就像是个 “不讲信用”的领域，

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

专家们都说 X 要发生了， 就是按 Y 这么做的。

人们一直在说神经网络 运行不了 70 年。

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

他们说的还不一定对呢。

They haven't been right yet.

也许说得对， 也许 71 年是对的，

They might be right maybe 70 years plus one

或者你想说多少年就多少年。

or something like that is what you need.

但我认为我们的方式一直是

But I think that our approach has always been,

你得把这项技术推向极致，

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

见证这种情况的真正发生。

因为这才会告诉你，哦， 这就是我们走进新时代的方式。

to really see it in action,

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

我们还没有物尽其用呢。

And we just haven't exhausted the fruit here.

CA: 你们所处的立场 还挺有争议的，

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

正确的方式就是把它公之于众，

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

再乘势而上，

and then harness all this, you know,

不仅仅是你们团队提供反馈，

instead of just your team giving feedback,

the world is now giving feedback.

整个世界都在提供反馈。

But ...

但是……

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

如果发生了坏事，

it is out there.

就这么发生了。

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

我听说 OpenAI 的起源故事是

when you were founded as a nonprofit,

你们刚开始是作为 非营利性机构成立的，

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

你们是查大企业

doing their unknown, possibly evil thing with AI.

利用 AI 行不可告人的 不轨之事的利器。

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

你们做出了一些模型，

somehow held them accountable

让大企业负起责任，

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

也在必要的时候 降低这个领域的发展速度。

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

至少我是这么听说的。

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

但不得不说， 真正发生的是相反的情况。

That your release of GPT, especially ChatGPT,

你们发布了 GPT， 尤其是 ChatGPT，

sent such shockwaves through the tech world

掀起了科技界的巨浪，

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

让谷歌、Meta 等等公司 现在手忙脚乱地追赶。

And some of their criticisms have been,

有些批评是这么说的，

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

你在没有适当保护的情况下逼我们 要么接受它的存在，要么死。

You know, how do you, like,

那你怎么能

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

确保你的所作所为 是负责任的、不鲁莽的。

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

GB: 我们一直在想这些问题。

Like, seriously all the time.

真的就是一直在想。

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

我们不能保证 做的一直都是对的。

But one thing I think has been incredibly important,

但我认为有一个非常重要的点，

从一开始，在我们思考

from the very beginning, when we were thinking

如何创造通用人工智能，

about how to build artificial general intelligence,

让它造福全人类的时候，

actually have it benefit all of humanity,

你该怎么达到这个目标呢？对吧？

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

我们默认会采取这样的计划， 悄悄做，

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

you get this super powerful thing,

做出了一个超牛的东西，

然后考虑考虑它的安全性， 按下“运行”，

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

然后希望自己做成了。

and you hope you got it right.

我不知道如何执行这个计划。

I don't know how to execute that plan.

也许有人知道。

Maybe someone else does.

但对我来说，这太吓人了， 感觉不太对。

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

所以我们转而采用这种方式，

And so I think that this alternative approach

这就是我能看到的 唯一一条途径，

is the only other path that I see,

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

就让现实打你的脸。

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

你也给了人们时间， 让他们提供输入。

You do have, before these machines are perfect,

在计算机达到完美状态之前，

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

在它们变得超级强大之前， 你有机会

看到它们真正发生的情况。

to see them in action.

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

我们已经从 GPT-3 身上 看到了，对吧？

GPT-3, we really were afraid

GPT-3，我们真的很害怕

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

人们最想用它做的事

was generate misinformation, try to tip elections.

就是生成虚假信息， 试图左右选举。

Instead, the number one thing was generating Viagra spam.

其实他们最想做的是 生成壮阳药钓鱼邮件。

(Laughter)

（笑声）

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

CA: 壮阳药钓鱼邮件确实不太好， 但是还有更糟糕的事。

Here's a thought experiment for you.

我们来做一个思想实验。

Suppose you're sitting in a room,

假设你坐在一个房间里，

there's a box on the table.

桌子上有个盒子。

You believe that in that box is something that,

你认为那个盒子里的东西，

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

非常有可能是个 很让人喜欢的东西，

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

可以是送给家人、 送给各位的精美礼品。

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

但也有 1% 的可能性 上面写了个小小的

that says: “Pandora.”

“潘多拉”。

And there's a chance

有可能会向世界 释放出无法想象的灾厄。

that this actually could unleash unimaginable evils on the world.

Do you open that box?

你会打开这个盒子吗？

GB: Well, so, absolutely not.

GB: 当然不会。

I think you don't do it that way.

我觉得你不会这么做的。

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

老实说，我来给你讲一个 我之前没说过的故事，

which is that shortly after we started OpenAI,

我们创立 OpenAI 之后不久，

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

我记得当时我在波多黎各 参加一场 AI 会议。

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

我坐在酒店房间里看海景，

all these people having a good time.

所有人都玩得很开心。

就在此刻，思考一下：

And you think about it for a moment,

你现在可以打开 那个潘多拉魔盒，

if you could choose for basically that Pandora’s box

to be five years away

你可以选择在 5 年后

or 500 years away,

或者 500 年以后，

which would you pick, right?

你会怎么选？

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

一方面你可能觉得 对你自己来说，

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

5 年以后比较好。

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

但如果得在 500 年以后打开， 人们也有更多的时间让它走上正轨，

which do you pick?

你会选哪个？

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

当时我真的觉得该这么选。

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

我觉得你就得选 500 年以后。

My brother was in the military at the time

我的兄弟当时在服役，

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

他那么实在地冒着生命危险，

than any of us typing things in computers

比与此同时我们这些 在电脑上打打字、

and developing this technology at the time.

开发技术的人要实在得多。

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

我真的确信人们得 “正视这项技术”这种说法。

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

但我认为这与它所处的领域不符。

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

纵观整个计算机技术史，

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

我可以认真地说， 这是一场涉及整个行业，

or even just almost like

或更像是

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

“人类的技术进步”的变革。

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

你越不把这些现有的 线索、信息整合在一起，

that are there, right,

we're still making faster computers,

我们就还在制造更快的计算机、

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

改进算法， 如此种种都是进行时。

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

如果你不整合它们， 你就处在一种“悬空”的状态，

which means that if someone does,

意思就是如果有人动手做了，

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

或者在某一个瞬间， 有人真的打通了这条路，

then you suddenly have this very powerful thing,

你就突然间有了一个 非常强大的东西，

no one's had any time to adjust,

没有人有时间去调整，

who knows what kind of safety precautions you get.

谁知道会有什么样的 安全防护措施呢。

And so I think that one thing I take away

所以我认为 我学到的一点就是

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

就算你想的是开发其他技术，

think about nuclear weapons,

比如核武器，

people talk about being like a zero to one,

这都是人类力所能及的

sort of, change in what humans could do.

零到一的改变。

But I actually think that if you look at capability,

但如果你看看人类的能力，

it's been quite smooth over time.

其实一直以来都是很平滑的。

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

历史上我们研发出的每一项技术，

has been, you've got to do it incrementally

都是一步一步来的，

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

每一次你要增强这种能力的时候，

for each moment that you're increasing it.

你都得考虑如何掌控它。

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

CA: 我可以这么理解……

the model you want us to have

你想我们以后使用的模型

is that we have birthed this extraordinary child

是我们孕育出的神奇小孩，

that may have superpowers

它会拥有带领人类 走向新世界的超能力。

that take humanity to a whole new place.

It is our collective responsibility to provide the guardrails

我们都有义务为这个孩子

for this child

保驾护航，

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

齐心协力让它更智慧， 不要毁灭我们自己。

Is that basically the model?

这个模型是这样吗？

GB: I think it's true.

GB: 是的。

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

不得不说这种情况 也可能会改变，对吧？

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

我们要随机应变地向前迈进。

And I think it's incredibly important today

我认为如今很重要的一点是，

that we all do get literate in this technology,

我们都得对这项技术有所了解，

figure out how to provide the feedback,

搞明白如何提供反馈，

decide what we want from it.

搞清楚我们到底想从中获得什么。

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

我希望，我们依旧 走在最优的道路上，

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

但我也很高兴我们 坦诚地讨论这个话题，

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

因为如果这项技术不存在， 我们也不会有这样的讨论了。

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

CA: 格雷格·布罗克曼，

感谢你来到 TED， 让我们大开眼界。

(Applause)

（掌声）