Is AI Progress Stuck? | Jennifer Golbeck | TED

翻译人员: Yip Yan Yeung 校对人员: Jinnie Sun

We've built artificial intelligence already

我们已经打造出了在某些任务上 比人类表现更佳的人工智能。

that, on specific tasks, performs better than humans.

There's AI that can play chess and beat human grandmasters.

有会下国际象棋 并击败人类特级大师的 AI。

But since the introduction of generative AI

但自从几年前 生成式 AI 的问世，

to the general public a couple years ago,

there's been more talk about artificial general intelligence, or AGI,

人们更常讨论通用人工智能， 即 AGI，

and that describes the idea

这一概念指的是

that there's AI that can perform at or above human levels

有一种以匹敌或超越人类水平

on a wide variety of tasks, just like we humans are able to do.

完成一众任务的 AI， 就如同我们人类的能力。

And people who think about AGI are worried about what it means

思考 AGI 的人们在顾虑它意味着什么，

if we reach that level of performance in the technology.

当这项科技达到了那种能力水平。

Right now, there's people from the tech industry

如今，科技届的人们跑出来说：

coming out and saying

"The AI that we're building is so powerful and dangerous

“我们打造的 AI 太强大了，太危险了，

that it poses a threat to civilization.”

可能危害人类文明。”

And they’re going to government and saying,

他们跑去政府说：

"Maybe you need to regulate us."

“也许你得管管我们。”

Now normally when an industry makes a powerful new tool,

通常当一个行业做出了 一个强大的新工具，

they don't say it poses an existential threat to humanity

他们不会说它 置人类于危急存亡之秋中，

and that it needs to be limited, so why are we hearing that language?

也不会说它该受到限制， 那我们为什么会听到这种说法呢？

And I think there's two main reasons.

我认为有两个主要原因。

One is if your technology is so powerful

一是如果你的科技太强大了，

that it can destroy civilization, between now and then,

甚至有可能摧毁文明,

那么从现在开始到毁灭前，

there's an awful lot of money to be made with that.

其间能赚一笔大钱。

And what better way to convince your investors

还有什么比警告别人 你的工具太危险了

to put some money with you

更能说服投资人 给你多投点钱的呢？

than to warn that your tool is that dangerous?

The other is that the idea of AI overtaking humanity

另一个原因是 AI 超越人类这一想法

is truly a cinematic concept.

完全是个电影里的概念。

We’ve all seen those movies.

我们都看过这种电影。

And it’s kind of entertaining to think about what that would mean now

现在再去想想那些电影的意义 还是挺有意思的，

with tools that we're actually able to put our hands on.

当我们真的有了摸得着的工具，

In fact, it’s so entertaining that it’s a very effective distraction

有意思到它其实有效地让人们不去想

from the real problems already happening in the world because of AI.

世界上那些由 AI 带来的真实问题。

The more we think about these improbable futures,

我们对这种低可能的未来想得越多，

the less time we spend thinking about how do we correct deepfakes

我们就会花更少的时间 思考如何纠正深度伪造，

or the fact that there's AI right now being used to decide

或者纠正现在用 AI 决定

whether or not people are let out of prison,

囚犯该不该出狱，

and we know it’s racially biased.

即使我们知道 AI 带有种族歧视。

But are we anywhere close to actually achieving AGI?

但我们很快要真正实现 AGI 了吗？

Some people think so.

有些人认为是的。

Elon Musk said that we'll achieve it within a year.

埃隆·马斯克说 我们会在一年内实现。

I think he posted this a few weeks ago.

他应该几周前发了这个帖子。

But like at the same time

但与此同时，

Google put out their eye search tool that's supposed to give you the answer

谷歌推出了 AI 搜索工具， 不用点击链接即可得到答案，

so you don’t have to click on a link,

and it's not going super well.

效果不是特别好。

["How many rocks should I eat?"]

[“我应该吃多少石头？”]

["... at least a single serving of pebbles, geodes or gravel ..."]

[“……至少一份鹅卵石、 晶洞或砾石……”]

Please don't eat rocks.

不要去吃石头。

(Laughter)

（笑声）

Now of course these tools are going to get better.

当然这些工具会变好的。

But if we're going to achieve AGI

但如果我们要实现 AGI，

or if they're even going to fundamentally change the way we work,

或者要让它彻底改变我们工作的方式，

we need to be in a place where they are continuing

我们需要让它一直处于 能力飞速提升的趋势中。

on a sharp upward trajectory in terms of their abilities.

And that may be one path.

这可能是一种趋势。

But there's also the possibility that what we're seeing

但我们还看到了另一种可能性： 这些工具已经达到了

is that these tools have basically achieved

what they're capable of doing,

它们能够做到的水平，

and the future is incremental improvements in a plateau.

未来是基于稳定水平的逐步改良。

So to understand the AI future,

要理解 AI 未来，

we need to look at all the hype around it and get under there

我们得看穿关于它的 各种炒作，深入研究，

and see what's technically possible.

看看从技术上什么是有可能的。

And we also need to think about where are the areas that we need to worry

我们也要考虑 我们需要担心哪几块，

and where are the areas that we don't.

不需要担心哪几块。

So if we want to realize the hype around AI,

如果我们想实现 AI 相关的炒作，

the one main challenge that we have to solve is reliability.

我们需要解决的 一个主要问题就是可靠性。

These algorithms are wrong all the time, like we saw with Google.

这些算法一直是错的， 就像我们看到谷歌的例子。

And Google actually came out and said,

谷歌其实曾站出来表示过，

after these bad search results were popularized,

在这些糟糕的搜索结果广泛传播后，

that they don't know how to fix this problem.

称他们也不知道 该怎么解决这个问题。

I use ChatGPT every day.

我每天都用 ChatGPT。

I write a newsletter that summarizes discussions on far-right message boards,

我在写一份订阅新闻稿， 总结极右翼留言板上的讨论，

and so I download that data,

我下载了这些数据，

ChatGPT helps me write a summary.

ChatGPT 帮我写一份总结。

And it makes me much more efficient than if I had to do it by hand.

比起我手动去完成要高效多了。

But I have to correct it every day

但是我每天都得纠正它，

because it misunderstands something,

因为它会误解一些事，

it takes out the context.

会断章取义。

And so because of that,

因此，

I can't just rely on it to do the job for me.

我不能完全依赖它帮我完成工作。

And this reliability is really important.

这种可靠性非常重要。

Now a subpart of reliability in this space is AI hallucination,

这一领域中的可靠性 有一部分是 AI 幻觉，

a great technical term for the fact that AI just makes stuff up

这是一个很好的技术术语， 描述了 AI 很多时候都在乱编。

a lot of the time.

I did this in my newsletter.

我在我的新闻稿里这么做了。

I said, ChatGPT are there any people threatening violence?

我说：“ChatGPT， 有没有人威胁要使用暴力？

If so, give me the quotes.

有的话，给我原句。”

And it produced these three really clear threats of violence

然后它输出了这三条 显然是威胁使用暴力的信息，

that didn't sound anything like people talk

但是听起来并不像 人们会发在留言板上的内容。

on these message boards.

And I went back to the data, and nobody ever said it.

于是我回到数据中， 根本没人说过。

It just made it up out of thin air.

它就是凭空捏造出来的。

And you may have seen this if you've used an AI image generator.

如果你用过 AI 图片生成器， 你可能也见过这种情况。

I asked it to give me a close up of people holding hands.

我让它给我一张人们牵手的特写。

That's a hallucination and a disturbing one at that.

这就是一种 AI 幻觉， 让人看着挺不舒服的。

(Laughter)

（笑声）

We have to solve this hallucination problem

我们得解决幻觉问题，

if this AI is going to live up to the hype.

如果 AI 要匹配炒作的水平。

And I don't think it's a solvable problem.

我觉得这不是一个可以解决的问题，

With the way this technology works, there are people who say,

鉴于这项技术的工作原理，

有些人会说我们在几个月内 就能搞定这个问题，

we're going to have it taken care of in a few months,

but there’s no technical reason to think that’s the case.

但是这种言论并没有什么技术依据。

Because generative AI always makes stuff up.

因为生成式 AI 总是在瞎编。

When you ask it a question,

你问了一个问题，

it's creating that answer or creating that image from scratch

它就在你问的时候 凭空创造一个答案或者图片。

when you ask.

It's not like a search engine that goes and finds the right answer on a page.

而不是像搜索引擎那样 找出某个页面上的正确答案。

And so because its job is to make things up every time,

由于它的任务就是每次乱编，

I don't know that we're going to be able to get it to make up correct stuff

我都不知道我们能不能 让它编出正确的东西，

and then not make up other stuff.

不要编别的东西。

That's not what it's trained to do,

这不是它被训练做的事情。

and we're very far from achieving that.

我们距离实现这一点还有很远。

And in fact, there are spaces where they're trying really hard.

事实上有一些人们正在 为之努力的领域。

One space that there's a lot of enthusiasm for AI

热衷于使用 AI 的一个领域

is in the legal area

是法律领域，

where they hope it will help write legal briefs or do research.

人们希望 AI 可以帮忙 写法律摘要或者进行法律研究。

Some people have found out the hard way

有些人折腾了一番以后发现

that they should not write legal briefs right now with ChatGPT

现在还不能用 ChatGPT 写法律摘要、

and send them to federal court,

发给联邦法院，

because it just makes up cases that sound right.

因为它会编出一些 听起来没问题的案件。

And that's a really fast way to get a judge mad at you

这会轻轻松松激怒法官，

and to get your case thrown out.

拒绝受理你的案件。

Now there are legal research companies right now

现在有些法律研究公司

that advertise hallucination-free

号称他们在使用“无幻觉”生成式 AI。

generative AI.

And I was really dubious about this.

我对此非常怀疑。

And researchers at Stanford actually went in and checked it,

斯坦福的研究人员参与并进行了调查，

and they found the best-performing of these hallucination-free tools

他们发现这些“无幻觉”工具中 表现最好的一个

still hallucinates 17 percent of the time.

也有 17% 的时候会产生幻觉。

So like on one hand,

所以说，一方面，

it's a great scientific achievement that we have built a tool

这是一项伟大的科学成就， 我们打造了一个工具，

that we can pose basically any query to,

可以向它提出任何查询内容，

and 60 or 70 or maybe even 80 percent of the time

60%、70% 甚至是 80% 的时间

it gives us a reasonable answer.

它能给出合理的答案。

But if we're going to rely on using those tools

但如果我们要依赖这些工具，

and they're wrong 20 or 30 percent of the time,

而它们有 20% 或 30% 的概率是错的，

there's no model where that's really useful.

那就没有一个真的很有用的模型。

And that kind of leads us into

这就引出了一个问题，

how do we make these tools that useful?

我们怎么能让这些工具变得有用呢？

Because even if you don't believe me

因为即使你不相信我，

and think we're going to solve this hallucination problem,

认为我们可以解决这个幻觉问题，

we're going to solve the reliability problem,

我们可以解决可靠性问题，

the tools still need to get better than they are now.

这些工具仍然需要 比现在的水平更进一步。

And there's two things they need to do that.

为此需要两件事。

One is lots more data

第一，更多的数据，

and two is the technology itself has to improve.

第二，技术本身得进步。

So where are we going to get that data?

那么我们从哪儿获取这些数据呢？

Because they've kind of taken all the reliable stuff online already.

因为它们几乎已经获取了 线上所有可靠的内容了。

And if we were to find twice as much data as they've already had,

如果我们要寻找 比现在已有更多一倍的数据，

that doesn't mean they're going to be twice as smart.

并不代表它会变两倍的聪明。

I don't know if there's enough data out there,

我不知道有没有足够的数据，

and it's compounded by the fact

而且情况比较复杂，

that one way the generative AI has been very successful

因为生成式 AI 很擅长的一点就是

is at producing low-quality content online.

生成线上的低质量内容。

That's bots on social media, misinformation,

社交媒体上的机器人、虚假信息，

and these SEO pages that don't really say anything

还有那些搜索引擎优化的 推荐网页，言之无物，

but have a lot of ads and come up high in the search results.

但是有一大堆广告， 在搜索结果中遥遥领先。

And if the AI starts training on pages that it generated,

如果 AI 开始基于它生成的网页训练，

we know from decades of AI research that they just get progressively worse.

根据数十年来的 AI 研究， 我们都知道它们会逐渐退步。

It's like the digital version of mad cow disease.

那就像数字版本的疯牛病。

(Laughter)

（笑声）

Let's say we solve the data problem.

假设我们解决了数据问题。

You still have to get the technology better.

你还是得改良技术。

And we've seen 50 billion dollars in the last couple years

过去几年里， 我们看到了 500 亿美元

invested in improving generative AI.

被投入改良生成式 AI。

And that's resulted in three billion dollars in revenue.

得到了 30 亿美元的收入。

So that's not sustainable.

所以这是不可持续的。

But of course it's early, right?

但是当然它还在早期，对吧？

Companies may find ways to start using this technology.

企业可能会找到用这项技术的方式。

But is it going to be valuable enough

但是它的价值是否足以

to justify the tens and maybe hundreds of billions of dollars

证明数百亿，甚至数千亿美元

of hardware that needs to be bought

在优化这些模型所需购置的 硬件上的花费是值得的呢？

to make these models get better?

I don't think so.

我觉得不能。

And we can kind of start looking at practical examples to figure that out.

我们看看实际的例子就知道了。

And it leads us to think about where are the spaces we need to worry and not.

那能让我们思考我们 需要担心、不需要担心什么领域。

Because one place that everybody's worried with this

因为每个人对此担心的一点是

is that AI is going to take all of our jobs.

AI 会抢走我们所有的工作。

Lots of people are telling us that’s going to happen,

很多人告诉我们会是这样的，

and people are worried about it.

于是人们对此非常担心。

And I think there's a fundamental misunderstanding at the heart of that.

我认为其中有一个关键的误解。

So imagine this scenario.

想象这个场景。

We have a company,

有一家公司，

and they can afford to employ two software engineers.

他们请得起两位软件工程师。

And if we were to give those engineers some generative AI to help write code,

如果我们给这两位工程师 一些生成式 AI，帮助他们写代码，

which is something it's pretty good at,

这是 AI 很擅长的，

let's say they're twice as efficient.

假设他们的效率翻了一番。

That's a big overestimate, but it makes the math easy.

这是过高的估计了， 但是简单起见就这么算吧。

So in that case, the company has two choices.

这样这家公司就有了两个选择。

They could fire one of those software engineers

他们可以炒掉一位软件工程师，

because the other one can do the work of two people now,

因为另一位现在可以 一人干两个人的活，

or they already could afford two of them,

或者他们本来就请得起两位，

and now they're twice as efficient,

现在效率翻倍，

so they're bringing in more money.

于是能赚更多钱。

So why not keep both of them and take that extra profit?

为什么把两位都留着，多赚点钱呢？

The only way this math fails is if the AI is so expensive

唯一不成立的一点就是 如果 AI 太昂贵了，

that it's not worth it.

价值配不上成本。

But that would be like the AI is 100,000 dollars a year

但是这等同于 AI 每年花十万美元

to do one person's worth of work.

完成一人份的工作。

So that sounds really expensive.

听起来非常昂贵。

And practically,

实际上，

there are already open-source versions of these tools

这些工具有很多开源版本，

that are low-cost, that companies can install and run themselves.

价格低廉，企业可以 自行安装、运行。

Now they don’t perform as well as the flagship models,

虽然它们的性能不如旗舰模型，

but if they're half as good and really cheap,

但如果它们能达到一半的效果， 而且相当廉价，

wouldn't you take those over the one that costs 100,000 dollars a year

你难道会不选它们， 而去选每年花掉十万美元

to do one person's work?

完成一人份工作的 AI 吗？

Of course you would.

你当然会选它们。

And so even if we solve reliability, we solve the data problem,

即使我们解决了 可靠性问题、数据问题，

we make the models better,

我们提升了模型性能，

the fact that there are cheap versions of this available

有可用的廉价版本这一事实

suggests that companies aren't going to be spending

表明了企业并不会花费

hundreds of millions of dollars to replace their workforce with AI.

上亿美元用 AI 取代人力。

There are areas that we need to worry, though.

但还是有我们需要担心的领域。

Because if we look at AI now,

因为现在看看 AI，

there are lots of problems that we haven't been able to solve.

还是有很多无法解决的问题。

I've been building artificial intelligence for over 20 years,

我已经从事人工智能建设超过 20 年，

and one thing we know

我们知道有一点，

is that if we train AI on human data,

那就是如果我们用人类数据训练 AI，

the AI adopts human biases,

那么 AI 就会沿袭人类的偏见，

and we have not been able to fix that.

而我们对此束手无策。

We've seen those biases start showing up in generative AI,

我们已经在生成式 AI 中 开始看见这些偏见，

and the gut reaction is always, well,

我们的直觉反应一直都是

let's just put in some guardrails to stop the AI from doing the biased thing.

那就加上一些防护措施， 阻止 AI 做出一些带有偏见的事。

But one, that never fixes the bias because the AI finds a way around it.

但是其一，这无法消除偏见， 因为 AI 会见招拆招。

And two, the guardrails themselves can cause problems.

其二，防护措施本身就会出问题。

So Google has an AI image generator,

谷歌有一款 AI 图片生成器，

and they tried to put guardrails in place to stop the bias in the results.

他们试图加上防护措施， 避免输出中出现偏见。

And it turned out it made it wrong.

结果发现做错了。

This is a request for a picture

有人要求生成一张 签署《独立宣言》的图片。

of the signing of the Declaration of Independence.

And it's a great picture, but it is not factually correct.

图片不错，但是与事实不符。

And so in trying to stop the bias,

为了试图阻止偏见，

we end up creating more reliability problems.

结果创造了更多可靠性问题。

We haven't been able to solve this problem of bias.

我们还无法解决这种偏见问题。

And if we're thinking about deferring decision making,

如果我们在考虑延迟决策、

replacing human decision makers and relying on this technology

取代人类决策者、 依靠这项技术，

and we can't solve this problem,

还是解决不了这个问题，

that's a thing that we should worry about

那么这就是值得担心的一件事，

and demand solutions to

需要解决方法，

before it's just widely adopted and employed because it's sexy.

必须赶在它们因看似酷炫 而被广泛使用之前。

And I think there's one final thing that's missing here,

我觉得还有一点没讲到，

which is our human intelligence

那就是我们的人类智慧

is not defined by our productivity at work.

并不由我们的工作效率定义。

At its core, it's defined by our ability to connect with other people.

核心上，它是由我们 与他人的联结能力定义的。

Our ability to have emotional responses,

我们产生情绪反应的能力，

to take our past and integrate it with new information

将我们的过往与新信息结合，

and creatively come up with new things.

有创意地想到新点子。

And that’s something that artificial intelligence is not now

这是人工智能无论现在，

nor will it ever be capable of doing.

还是未来都无法做到的事。

It may be able to imitate it

它也许可以模仿，

and give us a cheap facsimile of genuine connection

拙劣地模仿真情实感的联结、

and empathy and creativity.

同理心和创造力。

But it can't do those core things to our humanity.

但是它无法做到人类的核心。

And that's why I'm not really worried about AGI taking over civilization.

这就是为什么我不是很担心 AGI 取代人类文明。

But if you come away from this disbelieving everything I have told you,

但是如果你结束时不相信我刚说的，

and right now you're worried

你现在在担心

about humanity being destroyed by AI overlords,

人类会被 AI 魔王毁灭，

the one thing to remember is,

那么请记住一点，

despite what the movies have told you,

无论电影告诉你什么，

if it gets really bad,

如果真的回天乏术，

we still can always just turn it off.

我们总还是可以关机的。

(Laughter)

（笑声）

Thank you.

谢谢。

(Applause)

（掌声）