推荐微信扫码登录以获得更好的体验!另:🎉建议更新到最新版,以体验新功能!🎉 版本更新记录→

TED演讲:chatgpt惊人潜力的内幕

TED演讲:chatgpt惊人潜力的内幕

TED演讲:chatgpt惊人潜力的内幕

发布于 2024-02-09 · 时长 30:10

字幕摘录

时间英文中文
0:04We started OpenAI seven years ago我们7年前就开始了OpenAI
0:06because we felt like something really interesting was happening in AI,因为我们觉得爱尔有件有趣的事情
0:10and we wanted to help steer it in a positive direction.我们想帮助引导它朝着一个积极的方向发展。
0:15It's honestly just really amazing老实说,真的很棒
0:17to see how far this whole field has come since then.看看这整个领域从那时起有多远
0:20And it's really gratifying to hear from people like Raymond,听到雷蒙德这样的人的声音 实在令人欣慰
0:24who are using the technology we are building and others,他们正在使用我们正在建设的技术和其他人,
0:27for so many wonderful things.这么多美好的事物。
0:29We hear from people who are excited,我们听到那些兴奋的人说
0:31we hear from people who are nervous,我们听到那些紧张的人说
展开字幕全文(627 条)
序号英文中文
1We started OpenAI seven years ago我们7年前就开始了OpenAI
2because we felt like something really interesting was happening in AI,因为我们觉得爱尔有件有趣的事情
3and we wanted to help steer it in a positive direction.我们想帮助引导它朝着一个积极的方向发展。
4It's honestly just really amazing老实说,真的很棒
5to see how far this whole field has come since then.看看这整个领域从那时起有多远
6And it's really gratifying to hear from people like Raymond,听到雷蒙德这样的人的声音 实在令人欣慰
7who are using the technology we are building and others,他们正在使用我们正在建设的技术和其他人,
8for so many wonderful things.这么多美好的事物。
9We hear from people who are excited,我们听到那些兴奋的人说
10we hear from people who are nervous,我们听到那些紧张的人说
11we hear from people who are concerned,我们从关心的人那里听到,
12we hear from people who feel both those emotions at once.我们从那些同时感受这些情绪的人那里听到
13And honestly, that's how we feel.老实说,这就是我们的感受。
14Above all, it feels like we're entering an historic period right now,最重要的是,感觉我们正进入一个历史时期
15where we as a world are going to define a technology我们的世界将定义一种技术
16that will be so important for our society going forward.这对于我们社会前进将非常重要。
17And I believe that we can manage this for good.我相信我们可以永远管理好这件事。
18So today I want to show you the current state of that technology所以今天我想向大家展示这个技术的现状
19and some of the underlying design principles that we hold dear.以及我们珍视的一些基本设计原则。
20So the first thing I'm going to show you所以首先我要告诉你
21is what it's like to build a tool for an AI做人工智能工具的感觉
22rather than building it for a human.而不是为人类建造它。
23So we have a new Dolly model, which generates images,因此,我们有一个新的多利模型, 生成图像,
24and we are exposing it as an app for chat GPT to use on your behalf.我们把它曝光为 聊天GPT 代表你使用的应用程序。
25And you can do things like ask, suggest a nice post-Ted meal你可以做一些事情,比如问, 建议一个漂亮的后Ted饭
26and draw a picture of it.并画其像.
27Now, you get all of the sort of ideation and creative back and forth现在,你得到了所有的 那种想法 和创造性前后
28and taking care of the details for you that you get out of chat GPT.并照顾你的细节 你从聊天GPT。
29And here we go, it's not just the idea for the meal,我们开始吧 不只是吃饭的主意
30but a very, very detailed spread.但非常,非常详细的传播。
31So let's see what we're going to get.故当观之.
32But chat GPT doesn't just generate images in this case.但聊天GPT并不只是在此情况下生成图像.
33It doesn't just generate text, it also generates an image.它不仅生成文字,还生成一个图像.
34And that is something that really expands the power这能真正扩大力量
35of what it can do on your behalf它能为你做什么
36in terms of carrying out your intent.就执行你的意图而言
37And I'll point out, this is all a live demo.我要指出,这都是现场演示。
38This is all generated by the AI as we speak,这都是我们发言时的人工智能造成的
39so I actually don't even know what we're going to see.所以我甚至不知道我们会看到什么。
40This looks wonderful.这看起来太棒了。
41Now I'm getting hungry just looking at it.现在,我饿了 只是看着它。
42Now, we've extended chat GPT with other tools, too.现在,我们已经扩展聊天GPT 其他工具,太。
43For example, memory.例如记忆.
44You can, say, save this for later.你可以,说,留这个以后。
45And the interesting thing about these tools is they're very inspectable.这些工具的有趣之处在于 它们非常可检查。
46So you get this little pop-up here that says, use the Dolly app.所以,你得到这个小弹出 在这里说,用多莉的应用程序。
47This is coming to all chat GPT users over upcoming months.未来数月所有聊天的GPT用户都会收到这个消息。
48And you can look under the hood and see that what it actually did你可以看看引擎盖下面 看看它到底做了什么
49was write a prompt, just like a human could.写一个快速的, 就像人类可以。
50And so you sort of have this ability to inspect所以你有这种能力检查
51how the machine is using these tools,机器是如何使用这些工具的,
52which allows us to provide feedback to them.这使得我们能向他们提供反馈。
53Now it's saved for later,现在它保存以后,
54and let me show you what it's like to use that information让我告诉你使用这个信息是什么感觉
55and to integrate with other applications, too.并与其他应用程序融合。
56You can say,你可以说,
57I now make a shopping list for the tasty thing I was suggesting earlier.我现在做了一个购物清单 上面有我之前建议的东西
58I'm going to make it a little tricky for the AI.我要让人工智能变得有点棘手
59And tweet it out for all the TED viewers out there.给所有TED观众发微博
60So if you do make this wonderful, wonderful meal,所以如果你做了这顿美妙的晚餐
61I definitely want to know how it tastes.我绝对想知道它的味道。
62But you can see that chat GPT is selecting all these different tools但你可以看到聊天的GPT正在选择所有这些不同的工具
63without me having to tell it explicitly which ones to use in any situation.无需我明确告诉它 在任何情况下要使用哪种。
64And this, I think, shows a new way of thinking about the user interface.我认为,这显示了一种对用户界面的新思维方式。
65Like, we are so used to thinking of,像,我们是如此习惯 想,
66well, we have these apps, we click between them,好了,我们有这些应用程序, 我们点击他们之间,
67we copy-paste between them,我们抄袭他们之间,
68and usually it's a great experience within an app,通常它是一个伟大的经验 在应用程序,
69as long as you kind of know the menus and know all the options.只要你知道菜单和所有选择
70Yes, I would like you to.是的,我希望你这样做。
71Yes, please.是的,谢谢。
72Always good to be polite.总是好礼貌。
73And by having this unified language interface通过这个统一的语言界面
74on top of tools,在工具之上,
75the AI is able to sort of take away all those details from you,人工智能能从你身上拿走所有细节
76so you don't have to be the one who spells out every single sort of little piece所以,你不必是一个 谁拼出每一个类型的小作品
77{\fn黑体\fs22\bord1\shad0\3aHBE\4aH00\fscx67\fscy66\2cHFFFFFF\3cH808080}什么应该发生的
78And as I said, this is a live demo,正如我所说 这是现场演示
79so sometimes the unexpected will happen to us.因此,有时意外会发生在我们身上。
80But let's take a look at the Instacart shopping list但让我们看看Instacart购物名单
81while we're at it.当我们在它。
82And you can see we sent a list of ingredients to Instacart.你可以看到我们把一份原料清单 送到Instacart。
83Here's everything you need.这是你需要的一切。
84And the thing that's really interesting有趣的东西
85is that the traditional UI is still very valuable.传统的UI仍然非常宝贵。
86Right, if you look at this,好吧,如果你看看这个,
87you still can click through it and sort of modify the actual quantities.您仍然可以点击并修改实际数量。
88And that's something that I think shows that they're not going away,我认为这说明他们不会离开
89传统的UIs.
90It's just we have a new, augmented way to build them.只是我们有了新的,强化的方法来建造它们.
91And now we have a tweet that's been drafted for our review,现在我们有一条微博 已经起草我们的审查,
92which is also a very important thing.这也是一件非常重要的事。
93We can click Run, and there we are.我们可以点击运行,我们到了。
94We're the manager, we're able to inspect,我们是经理 我们可以检查
95we're able to change the work of the AI if we want to.我们能改变AI的工作,如果我们想。
96And so after this talk, you will be able to access this yourself.所以,经过这次谈话之后, 你就可以自己访问这个。
97And there we go.就这样
98Cool.凉爽。
99So ...所以说...
100Thank you, everyone.谢谢大家
101So we'll cut back to the slides.故切为滑.
102Now, the important thing about how we build this,现在,重要的是我们如何建造这个,
103it's not just about building these tools.不仅仅是建造这些工具
104It's about teaching the AI how to use them.乃教阿难如何用.
105Like, what do we even want it to do像,我们甚至想要它做什么
106when we ask these very high-level questions?当我们问这些非常高级的问题时?
107And to do this, we use an old idea.为了做到这一点,我们用了一个老主意。
108If you go back to Alan Turing's 1950 paper on the Turing test,如果你回到艾伦·图灵1950年关于图灵测试的论文
109he says, look, you'll never program an answer to this.他说,你看,你永远不会计划一个答案。
110Instead, you can learn it.相反,你可以学习它。
111You could build a machine like a human child你可以像人一样造一台机器
112and then teach it through feedback,然后通过反馈传授它,
113have a human teacher who provides rewards and punishments有个能提供奖惩的老师
114as it tries things out and does things that are either good or bad.因为它会尝试一些事情 做一些好或坏的事情
115And this is exactly how we train ChatGPT.这就是我们如何训练ChatGPT的。
116It's a two-step process.两步过程.
117First, we produce what Turing would have called a child machine首先,我们生产图灵所谓的儿童机器
118through an unsupervised learning process.通过一个不受监督的学习过程。
119We just show it the whole world, the whole internet,我们只是展示它 整个世界,整个互联网,
120and say, predict what comes next in text you've never seen before.并说,预测接下来的文字 你从来没有见过。
121And this process imbues it with all sorts of wonderful skills.这个过程使它充满了各种绝妙的技巧。
122For example, if you're shown a math problem,例如,如果你被显示一个数学问题,
123the only way to actually complete that math problem,是真正解决数学问题的唯一办法
124say, what comes next, that green nine up there,接下来会发生什么 上面那个绿色的9号
125is to actually solve the math problem.也就是解决数学问题。
126But we actually have to do a second step, too,但实际上我们也要做第二步,
127which is to teach the AI what to do with those skills.教人工智能如何运用这些技能。
128And for this, we provide feedback.为此,我们提供反馈。
129We have the AI try out multiple things, give us multiple suggestions,我们有AI尝试 多种事情, 给我们许多建议,
130and then the human rates them, says, this one's better than that one.然后人类给它们评分 说这个比那个好
131And this reinforces not just the specific thing that the AI said,这不仅仅是强化了 大赦国际说,
132but very importantly, the whole process that the AI used to produce that answer.但非常重要的是,人工智能用来制作答案的整个过程。
133And this allows it to generalize,这使得它能够概括,
134it allows it to teach to sort of infer your intent它可以教它推断出你的意图
135and apply it in scenarios that it hasn't seen before,把它应用到以前从未见过的情景中
136that it hasn't received feedback.还没收到反馈
137Now, sometimes the things we have to teach the AI有时候我们得教人工智能
138are not what you'd expect.不是你想的那样
139For example, when we first showed GPT-4 to Khan Academy,例如,当我们第一次向汗学院展示GPT-4时,
140they said, wow, this is so great,他们说,哇,这太棒了,
141we're going to be able to teach students wonderful things.我们会教学生一些美妙的东西
142Only one problem, it doesn't double-check students' math.只有一个问题,它不重复检查学生的数学。
143If there's some bad math in there,如果有坏的数学在那里,
144it will happily pretend that one plus one equals three and run with it.它会很高兴地假装一加一等于三 并运行它。
145So we had to collect some feedback data.所以我们不得不收集一些反馈数据。
146Sal Khan himself was very kind沙尔汗本人很善良
147and offered 20 hours of his own time to provide feedback to the machine给了他20个小时的时间给机器提供反馈
148alongside our team.与我们的团队。
149And over the course of a couple months,几个月来
150we were able to teach the AI that, hey,我们可以教人工智能,嘿,
151you really should push back on humans in this specific kind of scenario.你真的应该 推回人类 在这种特定的情景。
152And we've actually made lots and lots of improvements我们做了很多改进
153to the models this way,对模型这样,
154and when you push that thumbs down in ChatGPT,当你把拇指往下推 在ChatGPT,
155that actually is kind of like sending up a bat signal to our team这其实有点像 向我们的团队发出蝙蝠信号
156to say, here's an area of weakness where you should gather feedback.可以说,这里是一个薄弱环节, 你应该收集反馈。
157And so when you do that,所以,当你这样做,
158that's one way that we really listen to our users这也是我们真正倾听用户意见的一种方式
159and make sure we're building something that's more useful for everyone.并且确保我们建造的东西 对每个人都更有用。
160Now, providing high-quality feedback is a hard thing.现在,提供高质量的反馈是件难事。
161If you think about asking a kid to clean their room,如果你考虑让孩子清理房间
162if all you're doing is inspecting the floor,如果你只是检查地板
163you don't know if you're just teaching them to stuff all the toys in the closet.你不知道你是否只是教他们 把玩具都塞进衣柜里
164This is a nice dolly-generated image, by the way.顺便说一句,这是个漂亮的多莉生成的图像。
165And the same sort of reasoning applies to AI.同样的推理也适用于AI.
166As we move to harder tasks,当我们转向更困难的任务时,
167we will have to scale our ability to provide high-quality feedback.我们必须扩大提供高质量反馈的能力。
168But for this, the AI itself is happy to help.但为此,AI本身乐于提供帮助.
169It's happy to help us provide even better feedback很高兴能帮我们提供更好的反馈
170and to scale our ability to supervise the machine as time goes on.随着时间的流逝,我们有能力监督机器
171And let me show you what I mean.让我告诉你我的意思。
172For example, you can ask for a GPD-4 question like this,例如,你可以问一个GPD-4这样的问题,
173of how much time passed between these two foundational logs这两个基础日志之间的时间间隔
174on supervised learning and learning from human feedback,关于从人类反馈中接受监督的学习和学习,
175and the model says, two months passed.模特儿说两个月过去了
176But is it true?但这是真的吗?
177Like, these models are not 100 percent reliable,这些模型不百分之百可靠
178although they're getting better every time we provide some feedback.虽然每一次我们提供反馈 他们都越来越好
179But we can actually use the AI to fact-check,但我们可以用人工智能来核实事实
180and it can actually check its own work.它可以检查自己的工作。
181You can say, fact-check this for me.你可以说,事实检查这个给我。
182Now, in this case, I've actually given the AI a new tool.在这种情况下,我给了AI一个新的工具。
183This one is a browsing tool,这个是浏览工具
184where the model can issue search queries and click into web pages.其中,模型可以发布搜索查询并点击网页。
185And it actually writes out its whole chain of thought as it does it.实际上,它写出了整个思想链 就像它一样。
186It says, I'm going to search for this, and it actually does the search.上面写着,我要去寻找这个, 它实际上做了搜索。
187It finds the publication date and the search results.它找到出版日期和搜索结果。
188It then is issuing another search query.然后它会发布另一个搜索查询。
189It's going to click into the blog post.将点击博客文章。
190And all of this you could do, but it's a very tedious task.所有这一切你可以做, 但这是一个非常乏味的任务。
191It's not a thing that humans really want to do.这不是人类真正想要做的事情。
192It's much more fun to be in the driver's seat,坐在驾驶座上更好玩
193to be in this manager's position,成为这个经理的职位,
194where you can, if you want, triple-check the work.在那里你可以,如果你想, 三倍检查工作。
195And out comes citations,并传出许多迹象,
196so you can actually go and very easily verify所以你可以去核实一下
197any piece of this whole chain of reasoning.整个推理链的任何部分
198And it actually turns out, two months was wrong,事实上,两个月是错的
199two months in one week.一周内两个月
200That was correct.这是正确的。
201And we'll cut back to the slide.然后我们再切回幻灯片
202And so, the thing that's so interesting to me about this whole process所以,我对这个过程很感兴趣
203is that it's this many-step collaboration就是这种多步骤的合作
204between a human and an AI,在人类和人工智能之间
205because a human using this fact-checking tool因为一个人使用这个事实检查工具
206is doing it in order to produce data for another AI这样做是为了为另一个大赦国际提供数据。
207to become more useful to a human.对人更有用
208And I think this really shows the shape of something我觉得这真的 显示的东西的形状
209that we should expect to be much more common in the future,我们期望将来会更加常见,
210where we have humans and machines kind of very carefully and delicately designed在那里,我们有人类和机器 一种非常仔细和微妙的设计
211in how they fit into a problem and how we want to solve that problem.如何融入一个问题 以及如何解决问题
212We make sure that the humans are providing the management,确保人类提供管理
213the oversight, the feedback,监督,反馈,
214and that the machines are operating in a way that's inspectable and trustworthy.而机器的操作方式 是可检查的和可信的。
215And together, we're able to actually create even more trustworthy machines.我们一起可以创造出更可靠的机器。
216And I think that over time, if we get this process right,我认为,随着时间的推移, 如果我们把这个过程正确,
217we will be able to solve impossible problems.我们能解决不可能的问题
218And to give you a sense of just how impossible I'm talking,为了让你明白我所说的不可能的事
219I think we're going to be able to rethink almost every aspect我想我们可以重新思考 几乎所有方面
220of how we interact with computers.我们如何与计算机互动。
221For example, think about spreadsheets.例如,考虑电子表格。
222They've been around in some form since,他们一直以某种形式出现,从那时起,
223let's say, 40 years ago with VisiCalc.假设说 40年前和维西卡尔克在一起
224And I don't think they've really changed that much in that time.我想他们当时并没有改变这么多
225And here is a specific spreadsheet of all the AI papers on the archive这是档案里所有AI文件的具体电子表格
226for the past 30 years.过去30年
227There's about 167,000 of them,大约有167,000个
228and you can see the data right here.你可以在这里看到数据。
229But let me show you the chat GPT take on how to analyze a data set like this.但是,让我告诉你聊天GPT 如何分析这样的数据集。
230So we can give chat GPT access to yet another tool,这样我们可以让聊天GPT访问另一个工具,
231这个是Python翻译
232so it's able to run code just like a data scientist would.所以它能像数据科学家一样运行代码
233And so you can just literally upload a file所以你可以直接上传一个文件
234and ask questions about it,你应当询问,
235and very helpfully, it knows the name of the file,并且非常有帮助, 它知道文件的名称,
236and it's like, oh, this is CSV, a comma-separated value file.这就像,哦,这是CSV, 一个逗号分隔的值文件。
237I'll parse it for you.我帮你分析一下
238The only information here is the name of the file,这里唯一的信息是文件的名字,
239the column names like you saw,你所看到的列名,
240and then the actual data.然后是实际数据。
241And from that, it's able to infer what these columns actually mean.由此可以推断出这些柱子的真正含义。
242That semantic information wasn't in there.语义信息不在那里。
243It has to put together its world knowledge它必须汇集世界知识
244of knowing that, oh, yeah, archive is the site that people submit papers,知道,哦,是的,档案 是人们提交文件的网站,
245and therefore, that's what these things are,因此,这就是这些东西,
246and that these are integer values,这些是整数,
247and so therefore, it's a number of authors in the paper.因此,论文中有不少作者。
248All of that, that's work for a human to do,所有这一切, 这是人类的工作,
249and the AI is happy to help with it.AI很乐意帮忙。
250Now, I don't even know what I want to ask,现在我甚至不知道我想问什么
251so fortunately, you can ask the machine,幸运的是 你可以问机器
252can you make some exploratory graphs?你能做一些探索性的图表吗?
253And once again, this is a super high-level instruction再说一遍,这是超高级的指令
254with lots of intent behind it,背后有许多意图
255but I don't even know what I want,但我甚至不知道我想怎样
256and the AI kind of has to infer what I might be interested in.人工智能必须推断出我感兴趣的是什么
257And so it comes up with some good ideas, I think.因此,我想,它产生了一些好主意。
258So a histogram of the number of authors per paper,所以每张论文作者人数的直观图
259time series of papers per year, word cloud of the paper titles.每年的时序论文,论文标题的词云.
260All of that, I think, will be pretty interesting to see.我认为,所有这些将是非常有趣的。
261And the great thing is, it can actually do it.伟大的事情是,它实际上可以做到这一点。
262Here we go, a nice bell curve.在这里,我们走了,一个漂亮的钟曲线。
263You see that three is kind of the most common.你看,这三 是一种最常见的。
264It's going to then write a ...然后写...
265It's going to make this nice plot of the papers per year.每年的报纸会是这么好的
266Something crazy is happening in 2023, though.2023年发生了一件疯狂的事
267Looks like we were on an exponential, and it dropped off a cliff.看起来像我们在指数, 它掉下悬崖。
268What could be going on there?怎么会这样?
269And by the way, all this is Python code you can expect.顺便说一下,这一切都是 Python 代码,你可以期待。
270And then we'll see the word cloud,后见字云.
271so you can see all these wonderful things that appear in these titles.这样你就能看到这些名词中出现的这些美妙的东西.
272But I'm pretty unhappy about this 2023 thing.但我对2023年的事情很不高兴
273It makes this year look really bad.这让今年看起来很糟糕。
274Of course, the problem is that the year's not over.当然,问题在于今年还没有结束。
275So I'm going to push back on the machine.所以我要推回机器。
276So April 13th is the cutoff date, I believe.因此,4月13日是截止日期,我相信。
277Can you ...你能...
278So we'll see, this is a kind of ambitious one.我们来看看,这是种雄心勃勃的。
279So, you know, again, I feel like there was more I wanted out of the machine here.所以,你知道,再次,我觉得有 更多的我想离开这里的机器。
280I really wanted it to, like, notice this thing.我真的希望它,喜欢,注意到这件事情。
281Maybe it's a little bit of an overreach for it to have sort of inferred magically可能有点太过份了 有点神奇
282that this is what I wanted.这就是我想要的。
283But I inject my intent.但是,我注入了我的意图。
284I provide this additional piece of, you know, sort of guidance.我提供额外的,你知道的, 那种指导。
285And under the hood, the AI is just writing code again,在引擎盖下,人工智能又在写代码
286so if you want to inspect what it's doing, it's very possible.所以,如果你想检查 它在做什么,这是非常可能的。
287And now, I'm going to push back on the machine.现在,我要推回机器。
288It does the correct projection.它做正确的投影。
289And if you notice, it even updates the title.如果你注意到,它甚至更新标题。
290I didn't ask for that, but it knows what I want.我没有要求,但它知道我想要什么。
291Now, we'll cut back to the slide again.现在,我们再切回幻灯片。
292This slide shows a parable of how I think we,这个幻灯片展示了一个例子,我怎么认为我们,
293you know, a vision of how we may end up using this technology in the future.未来我们如何使用这种技术
294A person brought his very sick dog to the vet,有人把他病重的狗带到兽医那里
295and the veterinarian made a bad call to say,兽医打了一个不好的电话说,
296let's just wait and see.让我们等着瞧。
297And the dog would not be here today had he listened.而狗今天不会在这里 如果他听。
298In the meanwhile, he provided the blood test,与此同时,他提供了血液测试,
299like the full medical records, to GPT-4,和完整的医疗记录一样,到GPT-4,
300which said, I am not a vet, you need to talk to a professional.也就是说 我不是兽医 你需要跟专业人士谈谈
301Here are some hypotheses.这里有一些假设。
302He brought that information to a second vet who used it to save the dog's life.他把这个信息带到第二个兽医,他用它来拯救狗的生命.
303Now, these systems, they're not perfect.这些系统并不完美。
304You cannot overly rely on them.你不能过分依赖他们。
305But this story, I think, shows但这个故事,我认为,显示
306that the human with a medical professional拥有医学专业的人
307and with CHAT GPT as a brainstorming partner作为集思广益的合作伙伴
308was able to achieve an outcome that would not have happened otherwise.能够取得否则就不会取得的结果。
309I think this is something we should all reflect on and think about我觉得我们都应该思考一下
310as we consider how to integrate these systems into our world.我们考虑如何将这些系统融入我们的世界。
311And one thing I believe really deeply还有一件事我深信不疑
312is that getting AI right is going to require participation from everyone.获得AI的权利需要每个人的参与。
313And that's for deciding how we want it to slot in,这是用来决定我们如何让它进入
314that's for setting the rules of the road for what an AI will and won't do.这是为AI会做和不会做的事 制定道路规则
315And if there's one thing to take away from this talk,如果有一件事能从这个演讲中消失
316it's that this technology just looks different,这种技术看起来不同
317just different from anything people had anticipated.和人们预料的不一样
318And so we all have to become literate.因此,我们都必须识字。
319And that's honestly one of the reasons we released CHAT GPT.老实说,这就是我们释放CHAT GPT的原因之一。
320Together, I believe that we can achieve the OpenAI mission我相信我们可以一起完成OpenAI任务
321of ensuring that artificial general intelligence确保人工一般情报
322benefits all of humanity.使全人类受益。
323Thank you.谢谢
324Greg.格莱格。
325Wow.珃
326I mean, I suspect that within every mind out here,我的意思是,我怀疑 在每一个头脑在这里,
327apart from there's a feeling of reeling,除了有回旋的感觉,
328I suspect that a very large number of people viewing this,我怀疑很多人在看这个
329you look at that and you think, oh my goodness,你看看这个,你觉得, 哦,我的上帝,
330pretty much every single thing about the way I work, I need to rethink.几乎每一件事 关于我的工作方式, 我需要重新考虑。
331There's just new possibilities there.这里只有新的可能性。
332I mean, right, who thinks that they're having to rethink the way that we do things?我的意思是,对的,谁认为 他们不得不重新思考 我们做事的方式?
333Yeah, I mean, it's amazing, but it's also really scary.是啊,我的意思是,这是惊人的,但它也是非常可怕的。
334So let's talk, Greg, let's talk.所以,让我们谈谈,格雷格,让我们谈谈。
335Absolutely.当然
336I guess my first question actually is just how the hell have you done this?我的第一个问题就是 你是怎么做到的?
337OpenAI有几百名员工
338Google has thousands of employees working on artificial intelligence.谷歌有数千名员工从事人工智能工作.
339Why is it you who's come up with this technology that's shocked the world?为什么你们想出这种技术来,让世界震惊?
340The truth is, we're all building on shoulders of giants.事实是,我们都在建设 在巨人的肩膀上。
341There's no question if you look at the compute progress,毫无疑问,如果你看看计算的进展,
342the algorithmic progress, the data progress,算法进步 数据进步
343all of those are really industry-wide.这些都是真正的工业。
344但我认为在OpenAI,
345we made a lot of very deliberate choices from the early days.我们从早期就做了很多非常周密的选择。
346And the first one was just to confront reality as it lays,第一个只是面对现实
347and that we just sort of like thought really hard而我们只是有点喜欢 深思熟虑
348about what is it going to take to make progress here?需要什么才能在这方面取得进展?
349We tried a lot of things that didn't work, so you only see the things that did.我们尝试了很多不起作用的事情,所以你只看到了那些事情。
350And I think that the most important thing我认为最重要的是
351{\fn黑体\fs22\bord1\shad0\3aHBE\4aH00\fscx67\fscy66\2cHFFFFFF\3cH808080}一直以来 {\fn黑体\fs22\bord1\shad0\3aHBE\4aH00\fscx67\fscy66\2cHFFFFFF\3cH808080}我一直想找一群 彼此迥然不同的人
352to work together harmoniously.和谐地合作
353Can we have the water, by the way, just brought here?顺便说一句 能把水拿过来吗?
354I think we're going to need it.我想我们需要它。
355It's a dry amount of topic.题数干.
356But isn't there something also,但是,是不是还有一样东西,
357just about the fact that you saw something in these language models你从这些语言模型中看到一些东西
358that meant that if you continue to invest in them and grow them,这意味着如果你继续投资 种植它们,
359that something at some point might emerge?某时某事可能会出现?
360Yes.对
361And I think that ...我觉得...
362I mean, honestly, I think the story there is pretty illustrative.我的意思是,说实话,我认为 故事是相当说明性的。
363I think that at a high level, deep learning,我认为,在高水平,深刻的学习,
364we always knew that was what we wanted to be, was a deep learning lab.我们一直知道这就是我们想要的 是一个深层的学习实验室
365And exactly how to do it, I think that in the early days, we didn't know.确切地说,如何做到这一点, 我认为在早期,我们不知道。
366We tried a lot of things,我们尝试了很多事情,
367and one person was working on training a model有一人正在训练一个模型
368to predict the next character in Amazon reviews,为了预测亚马逊评论中的下一个角色,
369and he got a result,他得到了一个结果,
370where this is a syntactic process.在其中这是一个协同过程。
371You expect the model will predict where the commas go,你预计模型会预测逗号的去向,
372where the nouns and verbs are,名词和动词在哪里,
373but he actually got a state-of-the-art sentiment analysis classifier out of it.但实际上他得到了一个最新的情绪分析 分类器出来。
374This model could tell you if a review was positive or negative.这个模型可以告诉你审查是正面还是负面.
375I mean, today, we are just like,我的意思是,今天,我们只是喜欢,
376oh, come on, anyone could do that.哦,来吧,任何人都可以做到这一点。
377But this was the first time that you saw this emergence,但这是你第一次看到这种出现
378this sort of semantics that emerged from this underlying syntactic process.这种语义 产生于这个基本的 合成过程。
379And there, we knew, you've got to scale this thing,在那里,我们知道, 你必须缩小这个东西,
380you've got to see where it goes.你得看看它会往哪里走
381CAOKEY So I think this helps explain the riddle that baffles everyone考凯 所以我觉得这能解释 谜语让每个人困惑
382looking at this,看这个
383because these things are described as prediction machines,因为这些东西被描述为预测机器,
384and yet what we're seeing out of them,然而我们所看到的,
385it just feels impossible that that could come from a prediction machine,感觉这不可能来自预测机,
386just the stuff you showed us just now.你刚才给我们看的东西
387And the key idea of emergence is that when you get more of a thing,出现的关键思想是 当你得到更多的东西,
388suddenly different things emerge.突然间发生了不同的事情。
389It happens all the time, ant colonies, single ants run around.它经常发生,蚂蚁殖民地, 单蚁到处跑.
390When you bring enough of them together, you get these ant colonies当你把足够多的蚂蚁聚在一起, 你就会得到这些蚂蚁殖民地
391that show completely emergent and different behavior.展现出全新的行为
392Or a city, where a few houses together is just houses together,或一个城市, 几个房子在一起 只是房子在一起,
393but as you grow the number of houses,但随着你增加房屋数量,
394things emerge, like suburbs and cultural centers and traffic jams.就像郊区和文化中心 交通堵塞
395Give me one moment for you,给我一分钟给你
396when you saw just something pop that just blew your mind,当你看到刚刚爆炸的东西 让你心碎,
397that you just did not see coming.你只是没想到会来
398CAOKEY Yeah, well, so you can try this in chat.jpt.考凯 是啊,所以你可以试试这个在聊天。
399If you add 40-digit numbers,如果你加上40位数字
400CAOKEY 40-digit?考基40位数?
401CAOKEY 40-digit numbers, the model will do it,CAOKEY的40位数字,模型会做到这一点,
402which means it's really learned a internal circuit for how to do it.也就是说它真的学会了 内部的电路如何做到这一点。
403And the really interesting thing is,而真正有趣的是,
404actually, if you have an ad, like a 40-digit number实际上,如果你有一个广告, 像40位数字
405plus a 35-digit number, it'll often get it wrong.加上一个35位数字,它往往会弄错了.
406And so you can see that it's really learning the process,所以,你可以看到, 它真的学习的过程,
407but it hasn't fully generalized, right?但它还没有完全普遍化,对不对?
408It's like, you can't memorize the 40-digit addition table.这就像,你不能记住 40 位加表。
409That's more add-ins than there are in the universe.这比宇宙中的加法还要多.
410So it had to have learned something general,所以它不得不学到一些东西 将军,
411but it hasn't really fully yet learned that,但它还没有完全知道,
412oh, I can generalize this to adding arbitrary numbers of arbitrary lengths.哦,我可以概括一下,增加任意的长度。
413CAOKEY So what's happened here is that you've allowed it to scale up考凯 所以这里发生的事情是 你允许它扩大
414and look at an incredible number of pieces of text,且看看许多不可思议的文字,
415and it is learning things that you didn't know学习你不知道的东西
416that it was going to be capable of learning.它将有能力学习。
417CAOKEY Well, yeah, and it's more nuanced, too,考凯 嗯,是的,它更细微,太,
418because one science that we're starting to really get good at因为有一个科学,我们开始 真正获得好
419is predicting some of these emerging capabilities.正在预测这些新兴能力。
420And to do that, actually, one of the things为了做到这一点,实际上,一件事
421that I think is very undersung in this field is sort of engineering quality.我认为这个领域非常低沉 是一种工程质量。
422We had to rebuild our entire stack and get ...我们不得不重建我们的整个堆,并得到...
423When you think about building a rocket,当你想到建造火箭时,
424you know every tolerance has to be incredibly tiny.你知道每一个容忍 必须是惊人的微小。
425Same is true in machine learning.机器学习也是如此
426You have to get every single piece of the stack engineered properly,你得把堆栈的每一块都妥善地设计好
427and then you can start doing these predictions.然后你可以开始做这些预测。
428There are all these incredibly smooth scaling curves这些都是非常平滑的曲线
429that tell you something deeply fundamental about intelligence.告诉你一些关于智慧的根本问题
430If you look at our GPT-4 blog post, you can see all of these curves in there.看看我们的GPT-4博客文章,
431And now we're starting to be able to predict.现在我们开始可以预测了。
432So we were able to predict, for example, the performance on coding problems所以我们可以预测,比如,编码问题的表现
433from, you know, we basically look at some models从,你知道, 我们基本上看一些模型
434that are 10,000 times or 1,000 times smaller.即1万倍或1000倍较小。
435And so there's something about this that is actually smooth scaling,所以有些东西是平滑的
436even though it's still early days.虽然现在还很早
437CAOKEY So here is one of the big fears, then, that arises from this.考凯 因此,这里有一个很大的恐惧, 由此而来。
438If it's fundamental to what's happening here,如果这是这里发生的一切的根本
439that as you scale up, things emerge当你放大时,事物就会浮现
440that you can maybe predict in some level of confidence,你可以在某种程度的自信下预测
441but it's still capable of surprising you.但它仍然可以让你惊讶。
442Why isn't there just a huge risk of something truly terrible emerging?为何不冒着巨大的风险 出现真正可怕的事情?
443CAOKEY Well, I think all these are questions of degree and scale and timing.考凯 我想这些都是程度、规模和时机的问题。
444And I think one thing people miss, too, is sort of the integration with the world.我认为,人们也忽略了一件事, 就是与世界融合。
445There's also this incredibly emergent, very powerful thing, too.还有一件令人难以置信的,非常强大的事情。
446And so that's one of the reasons所以这就是原因之一
447that we think it's so important to deploy incrementally.我们认为渐进部署非常重要
448And so I think that what we kind of see right now,所以我认为我们现在看到的,
449if you look at this talk, a lot of what I focus on如果你看这段话,我关注的很多
450is providing really high-quality feedback.正在提供真正高质量的反馈。
451Today, the tasks that we do, you can inspect them, right?今天,我们的任务, 你可以检查,对不对?
452It's very easy to look at that math problem看那个数学问题很容易
453and be like, no, no, no, machine seven was the correct answer.7号机是正确的答案
454But even summarizing a book, that's a hard thing to supervise.但是,即使总结一本书,这是很难监督的。
455How do you know if this book summary is any good?你怎么知道这本书的总结有什么用?
456You have to read the whole book.须读全书.
457No one wants to do that.没有人想这样做。
458And so I think that the important thing will be所以我认为重要的是
459that we take this step-by-step我们一步一步走
460and that we say, OK, as we move on to book summaries,而我们说,OK, 当我们开始书写摘要,
461we have to supervise this task properly.我们必须妥善监督这项任务。
462We have to build up a track record with these machines我们必须用这些机器建立记录
463that they're able to actually carry out our intent.他们能够真正实现我们的意图
464And I think we're going to have to produce even better, more efficient,我认为我们必须生产出更好的,更有效的,
465more reliable ways of scaling this,以更可靠的方式扩大范围,
466sort of like making the machine be aligned with you.有点像制造机器 与你一致。
467CAOKEY NAPIER-MURPHY-MARTINEZ So we're going to hear later in the session,考基·纳尼尔-穆里普西-马蒂内斯 所以,我们将听到 在会后,
468there are critics who say that there's no real understanding inside the system.有批评者说系统内部没有真正的理解.
469Is it going to always ...它会永远...
470We're never going to know that it's not generating errors,我们永远不会知道 它不会产生错误,
471that it doesn't have common sense and so forth.它没有常识等等。
472Is it your belief, Greg, that that is true at any one moment,格雷格,你相信 任何时刻都是真的吗?
473but that the expansion of the scale and the human feedback但规模的扩大和人类的反馈
474that you talked about你说过的
475{\fn黑体\fs20\shad2\2aH82\3aH20\4aH33\fscx95\3cH592001\be1}基本上要踏上那段旅程
476of actually getting to things like truth and wisdom and so forth获得真理 智慧等等
477with a high degree of confidence?信心十足吗?
478Can you be sure of that?你能确定吗?
479CAOKEY NAPIER-MURPHY-MARTINEZ Yeah, well, I think that the opening eye ...考基·纳尼尔-穆里普西-马蒂内斯 是啊,我想开眼...
480Yes, I believe that is where we're headed.是的,我相信这就是我们要去的地方。
481And I think that the opening-eye approach here我认为这里的开眼方法
482has always been just like, let reality hit you in the face.一直只是一样, 让现实打击你的脸。
483It's like this field is the field of broken promises这就像这个领域 是违背诺言的领域
484of all these experts saying X is going to happen, Y is how it works.这些专家说,X将会发生,Y就是这样的。
485People have been saying neural nets aren't going to work for 70 years.人们一直说神经网70年都没用
486They haven't been right yet.他们还没说对呢
487They might be right, maybe 70 years plus one or something like that他们也许是对的,也许70年 加上一个或者类似的东西
488is what you need.这是你需要的。
489But I think that our approach has always been,但我认为我们的方法一直是,
490you've got to push to the limits of this technology你得把技术推向极限
491to really see it in action,真正看到它的行动,
492because that tells you then,因为那告诉你,
493oh, here's how we can move on to a new paradigm.哦,这是我们如何 前进到一个新的范式。
494And we just haven't exhausted the fruit here.我们只是还没把水果用完
495CAOKEY NAPIER-MURPHY-MARTINEZ It's quite a controversial stance you've taken,考基·纳尼尔-穆里普西-马蒂内斯 你的立场相当有争议
496that the right way to do this is to put it out there in public这样做的正确方式是公开发表
497and then harness all this.然后把这一切。
498Instead of just your team giving feedback,不只是你的团队提供反馈,
499the world is now giving feedback.世界现在正在提供反馈。
500But if bad things are going to emerge,但是,如果坏事要发生,
501it is out there.它就在这里。
502So the original story that I heard on OpenAI我在OpenAI上听到的原创故事
503when you were founded as a nonprofit,当时,你们原是非营利的。
504where you were there as the great sort of check在那里你在那里 作为伟大的检查类型
505on the big companies doing their unknown, possibly evil thing with AI,在大公司 做他们未知的, 可能是邪恶的事情与AI,
506and you were going to build models that sort of somehow held them accountable而你打算建立一些模式 以某种方式追究他们的责任
507and was capable of slowing the field down, if need be.并且能够 放慢球场 如有必要。
508Or at least that's kind of what I had.或者至少这就是我拥有的
509And yet what's happened, arguably, is the opposite,然而,发生的事情,可以说是相反的,
510that your release of GPT, especially chat GPT,你释放GPT, 特别是聊天GPT,
511puts such shockwaves through the tech world科技界的冲击波
512现在Google和Meta等等都在赶时间
513And some of their criticisms have been,他们的一些批评是,
514you are forcing us to put this out here without proper guardrails,你强迫我们把这个放在这里 没有合适的护栏
515or we die.否则我们就死定了
516How do you make the case你怎么解释?
517that what you have done is responsible here and not reckless?你所做的一切 责任在这里,而不是鲁莽?
518CAOKEY NAPIER-MURPHY-MARTINEZ Yeah, we think about these questions all the time,考基·纳尼尔-穆里普西-马蒂内斯 我们一直在想这些问题
519like seriously all the time,像一直以来认真的,
520and I don't think we're always going to get it right.我不认为我们总是会得到它的权利。
521But one thing I think has been incredibly important,但我认为有件事很重要
522from the very beginning,從一開始,
523when we were thinking about how to build artificial general intelligence,当我们想着如何建立人工一般情报时
524actually have it benefit all of humanity.实际上,它有益于全人类。
525How are you supposed to do that?你怎么能这么做?
526And that the default plan of being like,而默认的计划是,
527well, you build in secret, you get the super-powerful thing,好吧,你秘密建造, 你得到超强的东西,
528and then you figure out the safety of it,然后你找出它的安全,
529and then you push go and you hope you got it right.然后再推下去,希望你是对的
530I don't know how to execute that plan.我不知道如何执行那个计划。
531Maybe someone else does, but for me, that was always terrifying.也许别人知道 但对我来说 那总是很可怕
532It didn't feel right.感觉不对劲
533And so I think that this alternative approach所以我认为这个替代方法
534is the only sort of other path that I see,是唯一一种其他路径 我看见,
535which is that you do let reality hit you in the face.那就是你让现实袭击你的脸。
536And I think you do give people time to give input.我认为你确实给人们一些时间提供投入。
537You do have what, before these machines are perfect,在这些机器完好之前 你确实有办法
538before they are super-powerful,在他们超能力之前,
539that you actually have the ability to see them in action.你确实有能力看到他们行动起来
540And we've seen it from GPT-3.我们从GPT -3看到它。
541GPT-3, we really were afraid that the number one thingGPT-3,我们真怕头号
542people were going to do with it was generate misinformation人们会为此制造错误信息
543and try to tip elections.并尝试给选举小费。
544Instead, the number one thing was generating Viagra spam.倒不如说,第一是产生威哥垃圾邮件。
545So, Viagra spam is bad, but there are things that are much worse.所以,伟哥的垃圾邮件是坏的, 但有些事情更糟糕。
546Here's a thought experiment for you.给你做个思想实验
547Suppose you're sitting in a room, there's a box on the table.假设你坐在房间里,桌子上有盒子
548You believe that in that box is something that there's a very strong chance你相信那个盒子里有很强的机会
549it's something absolutely glorious这是绝对光荣的东西
550that's going to give beautiful gifts to your family and to everyone.这会给你的家人和所有人 带来美好的礼物
551But there's actually also a one percent thing in the small print there但实际上还有百分之一的东西 在小的指纹那里
552that says,说,
553(原始内容存档于2018-09-27). Pandora.
554And there's a chance that you're going to see a picture of Pandora你会看到潘多拉的照片
555and there's a chance that this actually could unleash并有一个机会 这实际上可能释放
556unimaginable evils on the world.无法想象的邪恶在世界上。
557Do you open that box?你打开那个盒子吗?
558Well, so absolutely not.那么,绝对不是。
559I think you don't do it that way.我想你不会这么做的
560And honestly, I'll tell you a story that I haven't actually told before,老实说,我会告诉你一个我以前没有讲过的故事,
561我们开始OpenAI后
562I remember I was in Puerto Rico for an AI conference.我记得我在波多黎各参加AI会议
563I was sitting in the hotel room,我坐在酒店房间
564just looking out over this wonderful water,只是看着这美妙的水,
565all these people having a good time,这些人玩得很开心
566and you think about it for a moment.你考虑一下
567And if you could choose for a potentially, basically that Pandora's box,如果你可以选择一个潜在的, 基本上是潘多拉的盒子,
568to be five years away or 500 years away,5年或500年之后,
569which would you pick?你会选哪一个?
570And on the one hand, you're like,一方面,你喜欢,
571well, maybe for you personally, it's better to have it be five years away.好吧,也许对于你个人来说, 它最好有五年远。
572But if it gets to be 500 years away and people get more time to get it right,但是,如果它得到500年的远 和人们得到更多的时间 得到它的权利,
573which do you pick?你选哪一个?
574And I just really felt it in that moment.我当时真的感觉到了
575I was like, of course you do the 500 years.我当然喜欢你做500年。
576For real, my brother was in the military at the time,说真的,我哥哥当时是军人
577and he puts his life on the line in a much more real way他以更真实的方式 将他的生命置于现实之中
578than any of us typing things in computers比我们任何人都在电脑上打字
579and developing this technology at the time.和开发这一技术 在当时。
580And so, yeah, I'm really sold on the, you've got to approach this right.所以,是的,我真的很卖 这一点,你必须正视。
581But I don't think that's quite playing the field as it truly lies.但我不认为这是在玩 因为它真的谎言。
582If you look at the whole history of computing,如果你看看整个计算历史,
583I really mean it when I say that this is a industry-wide,我是认真的 当我说这是一个全行业,
584or even sort of just almost like a human development of technology-wide shift.甚至有点像整个技术转变的人类发展。
585And the more that you sort of don't put together the pieces that are there,你越是不把那些碎片拼凑在一起
586we're still making the faster computers,我们还在制造更快的电脑
587we're still improving the algorithms,我们还在改进算法
588like all these things, they are happening.像所有这些事情,它们正在发生。
589And if you don't put them together, you get an overhang.如果你不把它们拼凑起来 就会有麻烦了
590Which means that if someone does,也就是说,如果有人这样做,
591or the moment that someone does manage to connect to the circuit,或者当有人设法连接到电路时,
592then you suddenly have this very powerful thing,然后你突然有了这个非常强大的东西,
593no one's had any time to adjust,没有人有时间调整,
594like who knows what kind of safety precautions you get.就像谁知道你得到了什么样的安全防范措施。
595And so I think that one thing I take away is,所以我认为我拿走了一件事
596even if you think about development of other sort of technologies,即使你想着开发其他类型的技术
597think about nuclear weapons,想想核武器,
598people talk about being like a zero-to-one sort of change人们在谈论一个零对一的变化
599in what humans could do,以人类所能做的事发誓,
600but I actually think that if you look at capability,但我真的认为 如果你看看能力,
601it's been quite smooth over time.好久不见
602And so the history, I think, of every technology we've developed所以,我想,我们开发的每一个技术的历史
603has been you've got to do it incrementally,一直以来,你必须逐步地做到这一点,
604and you've got to figure out how to manage it你必须想出办法
605for each moment that you're sort of increasing it.每一瞬间你都在增加
606CAOKEY MCCLELLAN So what I'm hearing is that you,考基·梅克莱伦 所以,我听说的是,你,
607the model you almost have is that we have birthed你几乎拥有的模型是 我们出生了
608this extraordinary child that may have superpowers这个非凡的孩子 可能拥有超能力
609that take humanity to a whole new place.将人类带到一个全新的地方
610It is our collective responsibility这是我们的集体责任
611to provide the guardrails for this child,为这个孩子提供护栏
612to collectively teach it to be wise and not to tear us all down.集体地教它要聪明,不要把我们都撕碎.
613Is that basically the model?那基本上是模型吗?
614CAOKEY MCCLELLAN I think it's true,考基·梅克莱伦 我想这是真的
615and I think it's also important to say this may shift, right?我想也有必要说这可能会改变 对吧?
616We've got to take each step as we encounter it,我们必须在遇到它时采取每一步,
617and I think it's incredibly important today我觉得今天很重要
618that we all do get literate in this technology,我们都在这种技术上识字,
619figure out how to provide the feedback, decide what we want from it.找出如何提供反馈, 决定我们想要从它。
620And I think that my hope is that that will continue to be the best path,我认为,我希望,那将继续是最佳的道路,
621but it's so good we're honestly having this debate,但很高兴我们真的有辩论
622because we wouldn't otherwise if it weren't out there.因为我们不会 否则,如果它不在那里。
623CAOKEY MCELLLLAN GREG Brockman, 非常感谢你来到TED。
624and blowing our minds.吹动我们的思想。
625Thank you.谢谢
626CAOKEY MCCLELLAN Appreciate it.考基·梅克莱伦 谢谢
627CAOKEY MCCLELLAN Thank you.考基·梅克莱伦 谢谢
该视频共有字幕 627 条。解锁更多字幕为会员功能,请移动到 价格