普林斯顿-耶鲁“相约周末”品读汇第七场：荣登《时代》杂志AI
字幕摘录
| 时间 | 英文 | 中文 |
|---|---|---|
| 0:03 | Good day, everyone, and welcome to this event at Yale Center | 大家好,欢迎来到耶鲁中心 |
| 0:08 | Beijing, Yale University's home here in China. | 北京,耶鲁大学在中国的故乡. |
| 0:12 | And a special welcome to our speakers, | 特别欢迎我们的演讲者, |
| 0:16 | Professor Aravind Narayanan and Dr. Sayesh Kapoor, | 阿拉文德·纳拉亚南教授和萨耶什·卡普尔博士 |
| 0:22 | as well as our commenter, Thomas Luo. | 以及我们的评论家 托马斯·罗 |
| 0:27 | Welcome to the center. | 欢迎来到中心 |
| 0:29 | I'm Carol Rafferty, a Yale College alumna | 我是卡罗拉费蒂,耶鲁大学的校友 |
| 0:32 | and also executive director of the Yale Center Beijing. | 并兼任耶鲁北京中心执行董事. |
| 0:37 | And today we are so delighted to be hosting this book talk | 今天我们很高兴能主持这个书评会 |
| 0:42 | on AI snake oil, what artificial intelligence can do, | 在AI蛇油, 人工智能可以做什么, |
展开字幕全文（1442 条）
| 序号 | 英文 | 中文 |
|---|---|---|
| 1 | Good day, everyone, and welcome to this event at Yale Center | 大家好,欢迎来到耶鲁中心 |
| 2 | Beijing, Yale University's home here in China. | 北京,耶鲁大学在中国的故乡. |
| 3 | And a special welcome to our speakers, | 特别欢迎我们的演讲者, |
| 4 | Professor Aravind Narayanan and Dr. Sayesh Kapoor, | 阿拉文德·纳拉亚南教授和萨耶什·卡普尔博士 |
| 5 | as well as our commenter, Thomas Luo. | 以及我们的评论家 托马斯·罗 |
| 6 | Welcome to the center. | 欢迎来到中心 |
| 7 | I'm Carol Rafferty, a Yale College alumna | 我是卡罗拉费蒂,耶鲁大学的校友 |
| 8 | and also executive director of the Yale Center Beijing. | 并兼任耶鲁北京中心执行董事. |
| 9 | And today we are so delighted to be hosting this book talk | 今天我们很高兴能主持这个书评会 |
| 10 | on AI snake oil, what artificial intelligence can do, | 在AI蛇油, 人工智能可以做什么, |
| 11 | what it can't, and how to tell the difference. | 和如何分辨区别。 |
| 12 | This is a collaboration with the Princeton University Press | 这是与普林斯顿大学出版社的合作 |
| 13 | Beijing office and is part of the Princeton and Yale | 北京办事处是普林斯顿和耶鲁大学的一部分 |
| 14 | ideas series, where we connect Chinese readers with Princeton | 我们把中国读者和普林斯顿联系起来 |
| 15 | University Press's authors on the most cutting edge | 大学出版社最尖端的作者 |
| 16 | ideas in our world. | 思想在我们的世界。 |
| 17 | So what's today's talk about? | 那么今天谈什么? |
| 18 | I think the title of the book really perfectly | 我觉得书的题目很完美 |
| 19 | encapsulates what we're going to talk about, | 囊括了我们要谈的内容, |
| 20 | which is what AI can do, what it can't, | 这是AI能做的,它不能做的, |
| 21 | and how to tell the difference. | 和如何分辨区别。 |
| 22 | I don't know if our speakers online are able to tell, | 我不知道我们网上的演讲者能否说出 |
| 23 | but at 8 AM on a Friday night, you | 但星期五晚上8点,你 |
| 24 | have a full house of people who I see are probably | 有一整栋房子的人 我看见他们可能 |
| 25 | high school or middle school students to working | 高中或中学生参加工作 |
| 26 | professionals or diplomats or journalists or whatnot. | 专业或外交官或记者,等等。 |
| 27 | So I hope this would draw a wide audience for your book | 所以我希望这本书能吸引广大读者 |
| 28 | here in China. | 在中国 |
| 29 | We will also have live streaming audience joining online. | 我们也会有现场观众加入网络。 |
| 30 | So we'll be able to tell you the figures, | 所以我们可以告诉你数字 |
| 31 | although not as fast as you could | 虽然你没有最快的速度 |
| 32 | with the work you do in artificial intelligence. | 和你在人工智能中的工作有关 |
| 33 | So we're all very eager to hear what you have to say. | 所以我们都渴望听到你说的话 |
| 34 | A bit of introduction about our speakers and our commenter | 介绍一下我们的发言者和评论员 |
| 35 | today. | 今时. |
| 36 | 书作者Arvind Narayanan教授 | |
| 37 | and Dr. Sayesh Kapoor from Princeton University, | 和普林斯顿大学的萨耶什·卡普尔博士, |
| 38 | they have been recognized by Times Magazine | 被《时代杂志》认出来 |
| 39 | as two of the 100 most influential experts in AI. | 作为AI百大最具影响力的专家中的两位. |
| 40 | Arvind Narayanan是计算机科学的教授 | |
| 41 | at Princeton University and the director | 普林斯顿大学和院长 |
| 42 | of the Center for Information Technology Policy. | 信息技术政策中心 |
| 43 | He studies the societal impact of digital technologies, | 他研究了数字技术的社会影响, |
| 44 | especially AI, and has co-authored a number of books, | 特别是大赦国际,并共同撰写了一些书籍, |
| 45 | actually, including Fairness and Machine Learning, | 事实上,包括公平与机器学习, |
| 46 | as well as, I found out in the book, | 还有,我在书里发现, |
| 47 | Bitcoin and Cryptocurrency Technologies. | 比特币和密码货币技术。 |
| 48 | We need to invite you to speak for us more. | 我们需要邀请你代表我们发言。 |
| 49 | These are all very timely talks. | 这些都是非常及时的会谈。 |
| 50 | Sayesh Kapoor博士是博士候选人 | |
| 51 | at the Center for Information Technology | 信息技术中心 |
| 52 | at Princeton University. | 在普林斯顿大学 |
| 53 | Princeton University, he previously worked for Facebook, | 普林斯顿大学 他以前在Facebook工作过 |
| 54 | doing content moderation, and conducted AI research | 进行内容节制,并进行人工智能研究 |
| 55 | at Columbia University and the Swiss Federal Institute | 哥伦比亚大学和瑞士联邦研究所 |
| 56 | of Technology in Lausanne. | 在洛桑的技术。 |
| 57 | Our commenter today, Thomas Luo, | 我们今天的评论员 托马斯·罗 |
| 58 | is the founding partner of GenAI Assembling | 是GenAI的创始合伙人 |
| 59 | and founder and CEO of Sea Planet and PingWest. | 也是海星球和平西的创始人兼首席执行官. |
| 60 | Thomas is a prominent alumnus of Tsinghua University | 托马斯是清华大学的著名校友 |
| 61 | and has built a reputation as a keen observer | 并树立了观察者的声誉, |
| 62 | of technological innovation in both China | 中国和中国的技术创新 |
| 63 | and the Silicon Valley. | 还有硅谷 |
| 64 | And before we dive into the discussion, | 在讨论之前 |
| 65 | I'd like to especially thank | 我想特别感谢 |
| 66 | the Princeton University Press Beijing Office | 普林斯顿大学出版社北京办事处 |
| 67 | and our Yale Center Beijing team for hosting this event, | 我们的耶鲁中心北京小组 主办这次活动, |
| 68 | as well as this very fun series. | 还有这个很有趣的系列 |
| 69 | The Princeton University Press Beijing Office | 普林斯顿大学出版社北京办事处 |
| 70 | will also be gifting two copies of this book | 还会赠送两本书 |
| 71 | to the first two people in the audience who ask questions. | 给观众头两个问问题的人 |
| 72 | So this is very exciting, but the book is also available | 所以,这是非常令人兴奋的, 但书也有 |
| 73 | for sale after the event outside. | 外出活动后出售。 |
| 74 | So ka-ching, ka-ching, people will be buying your books. | 所以,卡青,卡青, 人们会买你的书。 |
| 75 | And without further ado, let's welcome our speakers | 不用多说了,欢迎我们的演讲者 |
| 76 | and commenter, thank you very much. | 和评论员,非常感谢。 |
| 77 | Thank you so much for that kind introduction. | 非常感谢你的介绍 |
| 78 | Hi everyone, my name is Arvind Narayanan. | 大家好 我叫阿文德·纳拉亚南 |
| 79 | I hope you can hear me okay. | 希望你听得见我说话 |
| 80 | I'm going to start us off with a few minutes | 我先从几分钟开始 |
| 81 | telling kind of the story of this book | 讲述这本书的故事 |
| 82 | and then Saish will do a presentation | 然后赛什会做一个演示 |
| 83 | of the contents of this book. | 本书的内容。 |
| 84 | And I will chime in from time to time | 我会不时地敲门 |
| 85 | during the presentation. | 介绍期间。 |
| 86 | And then I would love to hear your questions, | 然后我想听听你的问题 |
| 87 | have a discussion and so forth. | 有讨论等等。 |
| 88 | Thank you for being here on a Friday evening. | 谢谢你周五晚上能来 |
| 89 | Okay, so for me, the story of this book started | 对我来说,这本书的故事开始了 |
| 90 | many years ago when I was doing research | 多年前我做研究的时候 |
| 91 | on what could go wrong with artificial intelligence, | 人工智能有什么不对劲, |
| 92 | particularly I was looking at potential biases | 特别是我看 潜在的偏见 |
| 93 | that could result, AI as you might know, | 你可能知道,艾丽克丝 |
| 94 | uses machine learning technology and the data that's used | 使用机器学习技术和所使用的数据 |
| 95 | for training machine learning comes from data created | 用于培训机器学习的数据 |
| 96 | by people, data about people. | 由人,关于人的数据。 |
| 97 | And so it's going to reflect cultural stereotypes | 因此,它将反映 文化定型观念 |
| 98 | in various societies. | 在各种社会中。 |
| 99 | Those could be gender biases, | 这些可能是性别偏见, |
| 100 | those could be various other kinds of biases. | 这些可能是各种其他类型的偏见。 |
| 101 | And as you might know, there is an active community | 如你所知,有一个活跃的社区 |
| 102 | of researchers who are looking | 正在寻找的研究人员 |
| 103 | into these kinds of questions. | 进入这样的问题。 |
| 104 | However, around 2019, I started thinking about | 不过,2019年左右,我开始思考 |
| 105 | are there cases where there are problems deeper than bias? | 是否有比偏见更深层次的问题? |
| 106 | Is AI being used in situations where it doesn't work at all | 人工智能是否被用于完全无效的情况 |
| 107 | and we shouldn't expect it to work at all? | 我们不该指望它能成功吗? |
| 108 | And in particular, I saw over and over again | 尤其是,我一次又一次地看到 |
| 109 | that there were these hiring automation technologies | 有这些雇佣自动化技术 |
| 110 | where these AI vendors would go to companies and say, | 在那里,这些AI供应商 会去公司说, |
| 111 | look, whenever you advertise a job opening, | 你看,每当你广告 一个职位空缺, |
| 112 | you have hundreds of people, | 你有上百个人 |
| 113 | maybe a thousand people applying to this position. | 也许有一千人申请这个职位 |
| 114 | And if some of you are job seekers, you might know, | 如果你们有些人是求职者,你们可能知道 |
| 115 | I think this is true all around the world | 我觉得全世界都是这样 |
| 116 | that there are so many people competing | 有这么多人竞争 |
| 117 | for each individual job. | 每项工作。 |
| 118 | And so the people who are reviewing our job applications, | 所以那些审查我们工作申请的人 |
| 119 | they don't know what to do. | 他们不知道该怎么办。 |
| 120 | There are too many applications | 申请太多 |
| 121 | to review too many people to interview. | 以审查太多的人 采访。 |
| 122 | And so the AI companies were saying, | 所以AI公司说, |
| 123 | look, use our AI product, | 用我们的人工智能产品 |
| 124 | ask your job candidates to upload a video, | 请候选人上传视频 |
| 125 | a short 30 second video where they're talking about | 短短的30秒视频 他们谈论 |
| 126 | not even so much their job qualifications, | 甚至没有这么多的工作资格, |
| 127 | but about their hobbies or whatever. | 但他们的爱好或什么的。 |
| 128 | And our product will use AI to figure out their personality | 我们的产品会用人工智能 找出他们的个性 |
| 129 | and how good they're going to be at a particular job. | 以及他们在某项工作上有多好 |
| 130 | And it occurred to me that there was no possible way | 我突然想到,不可能有办法 |
| 131 | by which this could work. | 它可以通过它发挥作用。 |
| 132 | I did not know any evidence that AI is capable | 我不知道有证据表明AI有能力 |
| 133 | of this sort of thing. | 这种事 |
| 134 | And coincidentally at that time in 2019, five years ago, | 巧合的是,五年前的2019年, |
| 135 | almost to the day actually, | 几乎直到今天,实际上, |
| 136 | I was invited to give a talk at MIT | 我被邀请去麻省理工学院演讲 |
| 137 | and I gave a talk called AI Snake Oil. | 我做了一个叫AI蛇油。 |
| 138 | And I pointed to this and other kinds of AI technologies | 我指出这种和其它类型的AI技术 |
| 139 | that I thought couldn't work. | 我认为不能工作。 |
| 140 | I called it an elaborate random number generator. | 我把它称为一个精心的随机数生成器。 |
| 141 | And I pointed to evidence from research | 我指出研究的证据 |
| 142 | led by my Princeton colleague, | 在我的普林斯顿同事的带领下 |
| 143 | sociologist professor, Matthew Salkonic. | 马修·萨尔科尼奇社会学家教授. |
| 144 | When you look at how well machine learning is able | 当你看机器学习有多好 |
| 145 | to predict what's going to happen in a person's life, | 预测一个人的生活中会发生什么 |
| 146 | the ability to predict that | 能够预测 |
| 147 | is only slightly better than random. | 仅略优于随机。 |
| 148 | And that should not be surprising, right? | 这不奇怪吧? |
| 149 | The future is not determined yet. | 未来尚未确定。 |
| 150 | No one can know what's going to happen in the future. | 未来无量众生能知. |
| 151 | Yes, if you have a lot of data and algorithms, | 是的 如果你有很多数据和算法 |
| 152 | you can do slightly better than random, | 你可以做比随机稍好一点, |
| 153 | but it's really just picking up crude statistical patterns | 但它真的只是 收集粗糙的统计模式 |
| 154 | in the data. | 在数据中。 |
| 155 | It's not some magic technology that can see the future. | 并不是某些魔法技术能看见未来. |
| 156 | So I gave that talk. | 所以我就说了 |
| 157 | The next day I put up my slides online. | 第二天我在网上放幻灯片 |
| 158 | I thought maybe 20 of my colleagues would look at it, | 我以为我的20个同事会看看 |
| 159 | but instead the slides went viral online. | 但幻灯片却在网络上传播 |
| 160 | And I didn't know that this is something | 我不知道这是什么 |
| 161 | that could happen with academic research. | 在学术研究中可能发生这种情况。 |
| 162 | And I was very surprised. | 我很惊讶。 |
| 163 | Within two days, I had 20 or actually more than that, | 在两天内,我还有20个或更多 |
| 164 | 30 or 40 invitations in my inbox, | 我收件箱里有三四十份邀请函 |
| 165 | asking me to turn that talk into a book or an article | 叫我把这段话变成书或文章 |
| 166 | or something like that. | 或者类似的东西。 |
| 167 | And I was very surprised. | 我很惊讶。 |
| 168 | Why are people so interested in this topic? | 为什么人们对这个话题这么感兴趣? |
| 169 | And I realized it was not because I had said | 我意识到这不是因为我说过 |
| 170 | something profound, but in fact, | 一些深刻的,但事实上, |
| 171 | precisely because I had something | 正是因为我有东西 |
| 172 | that many people have noticed independently, | 许多人独立地注意到, |
| 173 | that a lot of what is being sold as AI doesn't | 作为人工智能出售的很多东西 |
| 174 | and probably can't work. | 而且可能无法工作。 |
| 175 | Although, of course, there is a lot of genuine progress in AI. | 当然,大赦国际取得了许多真正的进展。 |
| 176 | We talk about that in the book as well. | 我们在书里也谈到了这一点。 |
| 177 | But while there are many people pushing back against faulty AI, | 但是,虽然有很多人反对错误的AI, |
| 178 | there are relatively few computer scientists saying, | 计算机科学家说, |
| 179 | look, I understand how AI works. | 听着,我知道AI是怎么工作的 |
| 180 | I built AI technology. | 我建立了AI技术。 |
| 181 | And I'm telling you that there is no way | 我告诉你,没有办法 |
| 182 | that this particular application of machine learning | 机器学习的特殊应用 |
| 183 | can possibly work. | 可能会工作。 |
| 184 | So I felt like that's a story I wanted to tell. | 所以我觉得那是我想讲的故事 |
| 185 | But at that time, I didn't feel ready to tell that story. | 但当时,我还没有准备好讲述这个故事. |
| 186 | I felt like a lot more research needed to be done. | 我觉得需要做更多的研究。 |
| 187 | 我很高兴Sayaj Kapoor加入我的队伍 | |
| 188 | And as you heard in the introduction, | 正如你在介绍中听到的, |
| 189 | he has built machine learning technologies | 他建立了机器学习技术 |
| 190 | 在Facebook这样的地方。 | |
| 191 | So he has seen what they can be useful for, | 故他看见他们对于什么是有用的, |
| 192 | but also what their limitations have been firsthand. | 但他们的局限性是第一手的 |
| 193 | And we've been doing research for many years. | 我们已经做了多年的研究。 |
| 194 | And we've found many surprising things | 我们发现很多奇怪的事情 |
| 195 | in the course of that research. | 在研究过程中 |
| 196 | For instance, we've been looking at AI for science. | 例如,我们一直在寻找AI的科学。 |
| 197 | And of course, it can be very powerful. | 当然,它可能非常强大。 |
| 198 | We've had a couple of Nobel prizes | 我们有过几次诺贝尔奖 |
| 199 | for the application of AI to scientific discovery. | 应用人工智能进行科学发现。 |
| 200 | But we've also seen how it can really | 但我们也看到了 如何真正可以 |
| 201 | go wrong when scientists carelessly | 科学家粗心大意,就错了 |
| 202 | apply machine learning. | 应用机器学习。 |
| 203 | There can be errors in the claims | 索赔中可能有错误 |
| 204 | that you make that go undetected that are sitting | 你使那未被发现的坐着, |
| 205 | in the scientific record until years later, | 在科学记录中,直到数年后, |
| 206 | someone realizes that this discovery is actually bogus. | 有人意识到这个发现其实是虚假的. |
| 207 | And we think that's a little bit of a crisis going on right | 我们认为,这是 一点点的危机正在发生 |
| 208 | now in the scientific world. | 现在在科学世界。 |
| 209 | And we've written papers exposing | 我们写了论文 揭露 |
| 210 | the scale at which these mistaken claims are being made. | 这些错误说法的提出规模。 |
| 211 | So it's been a learning process for us. | 所以这对我们来说是一个学习的过程。 |
| 212 | And we're so excited now to share some of those learnings | 我们非常兴奋地分享这些学习 |
| 213 | with you. | 和你们一起 |
| 214 | So with that, I will turn it over to Sayaj. | 这样我就把它交给Sayaj |
| 215 | Thank you again for being here. | 再次感谢你来到这里。 |
| 216 | Absolutely, yeah. | 当然,是的。 |
| 217 | Thank you so much for being here, everyone. | 非常感谢你们能来,各位 |
| 218 | And thanks, Arvind, for that wonderful intro. | 谢谢你 阿文德 精彩的介绍 |
| 219 | So I wanted to start us off with the three words | 所以我想从三个字开始 |
| 220 | in the title of the book, just AI and snake oil, | 在书名,只是AI和蛇油, |
| 221 | and basically talk about what we even mean | 并基本上谈论 我们甚至意味着什么 |
| 222 | when we use these words, AI. | 当我们用这些词,AI。 |
| 223 | So I think it's very surprising that despite 80 years of work | 所以,我觉得这非常令人惊讶的是, 尽管80年的工作 |
| 224 | in this area, it's still really hard | 在这个区域,还是很难 |
| 225 | to give an overarching definition of what constitutes AI. | 对什么构成大赦国际给出总体定义。 |
| 226 | But in the book, we use a three-factor test, | 但是在书中,我们用一个三要素测试, |
| 227 | three loose criteria to determine if something | 三条松散的标准 来确定什么 |
| 228 | is or isn't AI. | 是还是不是AI。 |
| 229 | So I found this definition to be useful. | 所以我觉得这个定义很有用。 |
| 230 | I hope it is for you as well. | 我也希望你也会这样 |
| 231 | The first factor is whether the task | 第一个因素是任务是否 |
| 232 | that we are using AI or using a tool to solve | 我们使用AI或工具来解决 |
| 233 | requires some creative effort or training for humans. | 需要一些创造性的努力或对人类的培训。 |
| 234 | So one example is text-to-image tools | 所以一个例子是文本到图像工具 |
| 235 | that generate an image based on a prompt in the text. | 生成基于文本中提示的图像。 |
| 236 | This would be an example of AI by this definition, | 根据这一定义,这将是大赦国际的一个例子, |
| 237 | because creating images or artwork | 因为创建图像或艺术品 |
| 238 | requires creative effort or training for humans. | 需要人类的创造性努力或培训。 |
| 239 | The second criteria is that the behavior of the tool | 第二个标准是工具的行为 |
| 240 | should not directly be specified in the code by the developer. | 开发者不应在代码中直接指定。 |
| 241 | So one example of such a tool might be a rule-based tool, | 因此,这种工具的一个例子是基于规则的工具, |
| 242 | which sort of looks at your conditions | 什么样的情况 |
| 243 | and takes a specific decision based on that. | 并据此作出具体决定。 |
| 244 | In fact, even a thermostat, in some sense, | 事实上,即使是自动调温器, 在某种程度上, |
| 245 | has its code or has its behavior directly specified. | 已直接指定其代码或行为。 |
| 246 | So if the temperature goes above a certain degree, | 如果温度超过一定程度 |
| 247 | you turn on the cooling. | 你打开冷却器。 |
| 248 | If the temperature goes below a certain degree, | 如果温度低于一定程度 |
| 249 | you turn on the heating. | 你打开暖气。 |
| 250 | That is not an example of AI. | 这不是大赦国际的例子。 |
| 251 | We do want the behavior to be learned | 我们确实希望人们能够了解我们的行为 |
| 252 | from patterns in the data rather than directly being | 从数据中的图案而不是直接 |
| 253 | hard-coded by the developers. | 由开发者硬编码. |
| 254 | The third and final criteria that we are applying | 我们适用的第三项也是最后标准 |
| 255 | is that there should be some flexibility in the inputs | 即投入应有一定的灵活性。 |
| 256 | that we use to give to the tool. | 我们用来给工具。 |
| 257 | So for example, if we are using an AI tool | 比如说,如果我们在使用人工智能工具 |
| 258 | to distinguish between images of cats and dogs, | 区分猫和狗的画面, |
| 259 | the tool should work about as well. | 工具也应该发挥作用。 |
| 260 | The tool should be able to take in inputs | 该工具应能够吸收投入 |
| 261 | that it hasn't seen before. | 从来没有见过。 |
| 262 | So it shouldn't just work on images of dogs and cats | 所以,它不应该只是工作 在狗和猫的形象 |
| 263 | that the tool has seen before, | 工具已经看到, |
| 264 | but rather it should generalize to some extent | 但它应该在某种程度上概括 |
| 265 | to images that have not been seen before, | 给以前从未见过的图像, |
| 266 | even if there is some accuracy loss. | 即使有 某些准确性损失。 |
| 267 | So this is like an overarching definition of what we mean | 所以,这就像一个总体的定义 我们的意思 |
| 268 | when we use the term AI. | 当我们使用AI这个词。 |
| 269 | Now, of course, as you can tell, | 现在,当然,如你所知, |
| 270 | this is a really broad definition | 这是一个非常宽泛的定义 |
| 271 | and we'll come back to that in a minute. | 一会儿再谈这个 |
| 272 | But I want you to dive a little bit further | 但我希望你更远一点 |
| 273 | for just a minute, | 就一会儿, |
| 274 | which is what is the dominant paradigm for using AI today? | 今天使用AI的主要模式是什么? |
| 275 | And that is machine learning. | 这就是机器学习。 |
| 276 | So machine learning refers to learning from examples | 所以机器学习是指从实例中学习 |
| 277 | and learning patterns in the data | 数据中的学习模式 |
| 278 | using these past examples. | 使用这些过去的例子。 |
| 279 | So for example, given enough photos of cats | 比如说,给猫的照片足够多 |
| 280 | on the one hand and dogs on the other, | 一方面是狗 一方面是狗 |
| 281 | a machine learning system | 机器学习系统 |
| 282 | can learn to distinguish between the two. | 可以学会区分两者。 |
| 283 | I find this a hilarious example | 我觉得这个例子很有趣 |
| 284 | of the difficulty of trying to tell apart a labradoodle, | 很难分辨拉布拉多 |
| 285 | which is a type of dog from fried chicken in these images. | 这是一种狗 从炸鸡在这些图像。 |
| 286 | But given enough images of labradoodles on the one hand | 但只要一面足够多的拉布拉多面 |
| 287 | and fried chicken on the other, | 和煎鸡在另一边, |
| 288 | AI systems can in fact today learn to distinguish the two. | AI系统实际上今天可以学会区分两者. |
| 289 | And to create tools like chat GPT, | 为了创造像聊天GPT这样的工具, |
| 290 | a very similar process is followed. | 遵循的程序非常相似。 |
| 291 | So instead of these labels being provided by a human, | 而不是人类提供的标签 |
| 292 | the labels are generated automatically from existing data, | 标签由现有数据自动生成, |
| 293 | say on the internet. | 在互联网上说。 |
| 294 | So for example, if the sentence, | 比如说,如果句子 |
| 295 | what is your name appears in the training set | 你叫什么名字出现在训练场 |
| 296 | or appears on some internet data, | 或出现在一些互联网数据, |
| 297 | it might turn into four different examples | 可能会变成四个不同的例子 |
| 298 | used in the training data for tools like chat GPT. | 在训练数据中用于聊天GPT等工具. |
| 299 | So the first example is what is your blank | 所以第一个例子就是你的空白 |
| 300 | and the correct answer would be name. | 正确的答案将是名字。 |
| 301 | The second example is what is blank name | 第二个例子是空白名称 |
| 302 | and the correct answer is your and so on. | 正确的答案是你,等等。 |
| 303 | And so machine learning really is what powers | 所以机器学习才是真正的力量 |
| 304 | a lot of the advances that we see in AI, | 在AI中看到的很多进步, |
| 305 | especially in the last few years. | 特别是过去几年 |
| 306 | So this brings me to the next part of the title, | 因此,我来到标题的下一个部分, |
| 307 | which is what is the snake oil in the title AI snake oil? | AI蛇油标题中的蛇油是什么? |
| 308 | So snake oil refers to this long standing | 蛇油是指这个长长的姿势 |
| 309 | or rather old tradition, | 或相当古老的传统, |
| 310 | especially within the United States, | 特别是在美国, |
| 311 | but all over the world of salesmen | 全世界都有推销员 |
| 312 | selling snake oil ointments | 销售蛇油膏 |
| 313 | as a cure for every single possible disease. | 作为治疗每一种可能疾病的方法。 |
| 314 | So for example, this poster here | 比如说,这张海报 |
| 315 | shows an example of Clark Stanley snake oil, | 以克拉克・斯坦利蛇油为例 |
| 316 | which was claimed to cure everything from rheumatism | 用来治愈风湿病 |
| 317 | to kidney stones to back pains and so on. | 肾结石的背痛等等。 |
| 318 | Now, of course, these medicines or oils | 现在,当然,这些药品或油 |
| 319 | did not really work very well. | 效果不怎么样 |
| 320 | And in the United States, | 在美国, |
| 321 | the Food and Drug Administration was established | 成立了食品和药品管理局 |
| 322 | in part because of these false advertising claims. | 部分原因是这些虚假的广告主张。 |
| 323 | And as I've been mentioned, | 正如我所提到的, |
| 324 | we think there's something similar going on with AI today. | 我们认为今天AI也有类似的事情发生. |
| 325 | So this is a screenshot of the example | 这是一个例子的截图 |
| 326 | he mentioned in the beginning. | 他一开始提到过 |
| 327 | This is a tool that claims to predict | 这是一个工具 声称预测 |
| 328 | how well a candidate might do at a job | 候选人在工作上做得怎么样? |
| 329 | and is being used by or is used by companies | 公司正在使用或正在使用 |
| 330 | that are seeking to hire people | 试图雇用人员 |
| 331 | to decide whose application to move on to the next stage | 决定谁申请进入下一阶段 |
| 332 | and whom to reject. | 且加以否认者, |
| 333 | And this is the sort of information this tool gives out. | 这就是这个工具给出的信息。 |
| 334 | So note that it has a lot of detailed information | 所以请注意,它有很多详细的信息 |
| 335 | about every single person. | 关于每个人。 |
| 336 | In this case, on the top right, | 在这种情况下,在右上方, |
| 337 | you'll see that this person is flagged as being assertive | 你会看到,这个人被标记为 坚定的 |
| 338 | and it even gives out a score. | 它甚至给出一个分数。 |
| 339 | So in this case, the score is 8.98. | 因此,在这种情况下,得分是8.98. |
| 340 | Notice the two decimal points of precision. | 注意精确度的小数点。 |
| 341 | So it's really talking about how accurately | 所以,它真的是在谈论 如何准确 |
| 342 | this tool can determine how well this candidate would do. | 这一工具可以确定该候选人的工作情况。 |
| 343 | And as I've been said, | 如我所说, |
| 344 | we think that many of these types of tools | 我们认为,许多这类工具 |
| 345 | are essentially elaborate random number generators. | 基本上是精心设计的随机数生成器。 |
| 346 | There is no evidence, very little peer reviewed work | 没有证据,同行评审的工作很少 |
| 347 | showing that these tools have any efficacy at all, | 表明这些工具有任何效力, |
| 348 | let alone at predicting how well someone will do at their jobs. | 更别提预测别人的工作能做得多好了 |
| 349 | And I think this is the core of a message. | 我认为这是信息的核心。 |
| 350 | So our key message in the book | 所以书中的关键信息 |
| 351 | is that the most confusing thing about AI | 这是关于AI最令人困惑的事情 |
| 352 | is that AI is an umbrella term. | AI是一个总括术语。 |
| 353 | It is the same term is used to refer | 这个词也用来指 |
| 354 | to many different types of technologies, | 许多不同类型的技术, |
| 355 | which have very little to do with each other under the hood. | 和头罩下对方没什么关系 |
| 356 | So some of these technologies have made massive progress | 这些技术已经取得了巨大进步 |
| 357 | in the last 10 years. | 过去10年 |
| 358 | For example, generative AI. | 例如,遗传性AI. |
| 359 | So the image here is an image generated using OpenAI's | 因此这里的图像是使用 OpenAI 生成的图像 |
| 360 | text to image tool called Dali. | 文本到图像工具,名为 Dali。 |
| 361 | It was an image generated using just a very simple prompt, | 这是用一个非常简单的提示生成的图像, |
| 362 | an astronaut riding a horse. | 太空人骑着马 |
| 363 | And this is the image it came up with. | 这就是它产生的形象。 |
| 364 | And clearly we've been making improvements | 显然我们一直在改进 |
| 365 | to text generation models like chat GPT, | 到文本生成模型,如聊天GPT, |
| 366 | to image generation models like Dali, | 像达利这样的图像生成模型 |
| 367 | and also to generative AI in other domains | 以及其它领域的遗传性AI |
| 368 | like protein folding. | 像蛋白质折叠。 |
| 369 | All of these have indeed been transformative. | 所有这些确实是变革性的。 |
| 370 | On the other hand, we have tools which we call predictive AI. | 另一方面,我们有我们称之为预测性AI的工具。 |
| 371 | These are tools that are used to make predictions | 这些都是用来预测的工具 |
| 372 | about individual people's futures. | 关于个人的未来。 |
| 373 | And on that basis, make decisions about whether or not | 在此基础上,决定是否 |
| 374 | to hire them for a job, | 雇他们做工作, |
| 375 | or whether they should be released on bail | 或应否保释他们 |
| 376 | when they're being criminally charged, | 当他们受到刑事指控时 |
| 377 | or what their insurance rates should be, | 或他们的保险费率是多少, |
| 378 | or whether they should receive insurance and so on. | 或者他们是否应该得到保险等等。 |
| 379 | And I think this is where we find a lot of snake oil. | 我想这就是我们发现很多蛇油的地方。 |
| 380 | And we'll come back to predictive AI in a minute. | 我们马上就回来预测AI。 |
| 381 | But in the book, we also cover two other types | 但是在书中,我们也涵盖另外两种类型 |
| 382 | of applications of AI, including social media algorithms. | 包括社交媒体算法。 |
| 383 | Both the algorithms used to optimize for engagement | 两种算法都用来优化交战 |
| 384 | on social media, but also more importantly, | 但更重要的是, |
| 385 | the algorithms that are used to moderate content | 用于调和内容的算法 |
| 386 | or to take down content if it does not agree | 或删除内容,如果它不同意 |
| 387 | with the policies that a social media platform has. | 以及社交媒体平台的政策。 |
| 388 | And finally, we briefly talk about robotics, | 最后,我们简短地谈到机器人 |
| 389 | including in self-driving cars in the book as well. | 包括书中的自驾车 |
| 390 | Okay, so let's get back to predictive AI, | 好吧,让我们回到预测AI, |
| 391 | which is where I mentioned there's a lot of snake oil. | 我提到那里有很多蛇油 |
| 392 | So what we think is happening with predictive AI | 所以,我们认为发生 预测AI |
| 393 | is that vendors who are selling predictive AI | 是卖预言AI的卖家 |
| 394 | exploit the public's confusion | 利用公众的困惑 |
| 395 | over the different types of AI. | 超越不同类型的AI。 |
| 396 | They sell their tools as if they're state of the art, | 他们卖自己的工具 仿佛他们最先进的, |
| 397 | but under the hood, | 但是在引擎盖下 |
| 398 | these tools are doing something fairly unsophisticated. | 这些工具正在做一些相当不成熟的事情。 |
| 399 | For example, I really like this sort of case study | 比如说,我真的很喜欢这种案例研究 |
| 400 | of Ritorio. | 里托里奥 |
| 401 | Ritorio是一家雇佣公司。 | |
| 402 | On its website, it claims to identify | 在其网站上,它声称识别 |
| 403 | and drive winning behaviors for different types of jobs, | 推动不同类型工作的胜利行为, |
| 404 | including customer service. | 包括客户服务。 |
| 405 | And they use AI-powered video analysis | 他们使用人工智能的视频分析 |
| 406 | to make these determinations. | 作出这些决定。 |
| 407 | Now, in most cases, when companies make these claims, | 现在,在大多数情况下, 当公司提出这些要求, |
| 408 | we don't really have the tools to assess | 我们真的没有工具来评估 |
| 409 | how well their tools work. | 他们的工具如何有效。 |
| 410 | But in this case, an investigative journalism group | 但在这个案子里 一个调查新闻组 |
| 411 | was able to get access to the tool | 能够访问工具 |
| 412 | and test out how well it works. | 并测试它的工作原理。 |
| 413 | What this found was astonishing. | 这个发现令人吃惊。 |
| 414 | So the same video of a person | 所以,同一个人视频 |
| 415 | with just the background in the back changed | 后面的背景就变了 |
| 416 | led to dramatically different scores. | 导致得分大不相同。 |
| 417 | So for example, in this image here, | 例如,在这个图像中, |
| 418 | if someone was speaking in front of a plain background, | 如果有人在普通背景面前说话 |
| 419 | they would get a much lower score | 他们得到的分数会低得多 |
| 420 | compared to if they have an image of a bookshelf | 以有书架的形象发誓, |
| 421 | in the background. | 在背景中显示。 |
| 422 | And they found many other such cases, | 他们发现了许多其他这样的案件, |
| 423 | which highlighted that these tools | 强调了这些工具 |
| 424 | are essentially relying on correlations in the data. | 基本上依靠数据的关联性。 |
| 425 | They don't really look at what the job candidate is saying, | 他们并不真正看 候选人在说什么, |
| 426 | but rather on superficial appearances, | 但其实是表面的, |
| 427 | such as whether there's a bookshelf in the camera. | 比如摄像机里有没有书架 |
| 428 | Now, this might seem like a funny example, | 现在,这似乎是一个有趣的例子, |
| 429 | but indeed it can be consequential if you're a job seeker. | 但是如果你是求职者,那就可能因此发生。 |
| 430 | But predictive AI has actually been used | 但预言AI实际上被使用 |
| 431 | for far more consequential applications. | 更具有后果的应用。 |
| 432 | For example, in 2013, the Netherlands deployed an algorithm | 例如,2013年,荷兰采用了算法 |
| 433 | to detect welfare fraud. | 发现福利欺诈。 |
| 434 | That is if families with children were taking money | 如果有孩子的家庭拿钱 |
| 435 | from the state in a fraudulent manner. | 从国家以欺诈的方式。 |
| 436 | The algorithm wrongly accused thousands of families | 算法错误地指责了成千上万的家庭 |
| 437 | and sent many into debt. | 并且使许多人背负债务。 |
| 438 | In some cases, families were asked to pay back | 在有些情况下,家庭被要求偿还 |
| 439 | over 100,000 euros. | 超过10万欧元。 |
| 440 | Six years after the algorithm was introduced, | 算法推出六年后 |
| 441 | it was finally found out that the algorithm had these flaws | 最后发现算法有这些缺陷 |
| 442 | and it was discontinued. | 它被中止了。 |
| 443 | And over the fallout over the algorithm was used, | 在算法的反射上, |
| 444 | the Dutch prime minister | 荷兰总理 |
| 445 | and his entire cabinet had to resign. | 他的整个内阁不得不辞职。 |
| 446 | There are also examples within the US. | 在美国也有这方面的例子。 |
| 447 | So in 2017, US healthcare technology company Epic | 2017年,美国医疗科技公司Epic |
| 448 | launched a sepsis prevention algorithm | 启动预防败血症算法 |
| 449 | and claimed to have an extremely high accuracy. | 并声称其准确度极高。 |
| 450 | Four years later, when an independent investigation | 四年后,当独立调查 |
| 451 | was conducted on the efficacy of this algorithm, | 以这种算法的功效进行, |
| 452 | it found that the accuracy was actually much lower. | 它发现准确性实际上要低得多。 |
| 453 | So while Epic had claimed a relative accuracy of about 80%, | 虽然Epic声称相对准确度约为80%, |
| 454 | the actual accuracy was about 60% or so, | 实际准确度约为60%左右, |
| 455 | much closer to a coin toss. | 离掷硬币更近 |
| 456 | And one year after this investigation, | 调查一年后 |
| 457 | Epic discontinued this one size fits all | Epic 停止使用此尺寸 |
| 458 | sepsis prediction model. | 败血症预测模型. |
| 459 | Now, if companies can indeed predict sepsis early, | 现在,如果公司真的可以 早预测败血症, |
| 460 | that would be a huge advance | 这将是一个巨大的进步 |
| 461 | because sepsis is one of the deadliest diseases, | 因为败血症是最致命的疾病之一 |
| 462 | especially in hospitals. | 特别是在医院里 |
| 463 | But here what we're seeing is there is inadequate validation | 但这里我们所看到的是 验证不充分 |
| 464 | of these tools in the real world. | 这些工具在现实世界。 |
| 465 | And finally, another example from the US | 最后,美国的另一个例子 |
| 466 | is the US state of Oregon in 2018 | 2018年是美国俄勒冈州 |
| 467 | implemented an algorithm to predict | 执行一个预测算法 |
| 468 | if children are at risk of maltreatment. | 如果儿童有遭受虐待的危险。 |
| 469 | But after widespread reports of bias | 但是在广泛报道偏见之后 |
| 470 | against black families in particular, | 特别是针对黑人家庭, |
| 471 | the tool was discontinued in 2022. | 该工具于2022年中止。 |
| 472 | So you'll note that all three tools here | 所以,你会注意到,所有三个工具在这里 |
| 473 | are examples of predictive AI. | 它们是预测性AI的例子。 |
| 474 | In many cases, these tools just don't work | 在许多情况下,这些工具是没用的 |
| 475 | as well as they are promised to. | 他们确是被警告的。 |
| 476 | And that eventually leads to fallout over these tools use. | 这最终导致了这些工具的使用。 |
| 477 | But in the meantime, they're actively causing harm | 但与此同时,他们正在积极造成伤害 |
| 478 | to job seekers, to people who are taking child benefits | 向求职者、领取子女津贴的人发放 |
| 479 | and so on. | 诸如来. |
| 480 | And in fact, there have been hundreds of such incidents | 事实上,发生了数百起这样的事件 |
| 481 | in the last few years in fields like insurance, | 在过去几年中, 在保险等领域, |
| 482 | to mortgage, to the DHS, | 向人口与健康调查提供抵押贷款, |
| 483 | to the Department of Homeland Security and so on. | 给国土安全部等等 |
| 484 | And I think this is what leads to our pessimism | 我觉得这就是导致我们悲观的原因 |
| 485 | about predictive AI. | 关于预测AI。 |
| 486 | And in fact, we have an entire chapter in the book | 事实上,书中有一章 |
| 487 | about how difficult it is to predict these life outcomes. | 预测这些生命结果有多难 |
| 488 | So it's not just about AI going wrong in these cases, | 所以这不仅仅是关于AI在这些情况中出错, |
| 489 | but rather that many of the types of things | 但其实很多种类的东西 |
| 490 | we are trying to predict with AI | 我们试图与AI预测 |
| 491 | might have fundamental limits to their predictability. | 它们的可预测性可能会受到根本性的限制。 |
| 492 | So maybe I'll stop here for a minute | 所以,也许我会在这里停留一分钟 |
| 493 | and let Arvind chime in. | 让阿文德进来 |
| 494 | Sure, thank you. | 好的 谢谢 |
| 495 | Yes, I don't have too much to add. | 是的,我没有太多补充。 |
| 496 | I'll talk for a couple of minutes | 我谈几分钟 |
| 497 | and then we'll, at the end of the presentation, | 然后,在演讲结束时, |
| 498 | I'll say a little bit more about what's in the book. | 略说经中之事. |
| 499 | Unless you actually wanna go to that slide, Sayash, | 除非你真的想去看那张幻灯片 Sayash |
| 500 | then I can go through that slide if that's okay. | 如果可以的话 我可以通过幻灯片 |
| 501 | Great. | 不错 |
| 502 | So this is our way of thinking about | 这就是我们思考的方式 |
| 503 | all of the things that can go right or wrong with AI. | 所有那些可以对错的AI。 |
| 504 | So we have placed, | 所以,我们安置, |
| 505 | and this is from the introductory chapter of the book, | 这是书中介绍的一章, |
| 506 | we have placed various applications of AI | 我们提出了各种AI的申请。 |
| 507 | into this two-dimensional chart. | 输入此二维图表。 |
| 508 | On the X axis, what you see is a spectrum. | 在X轴上,你看到的是光谱. |
| 509 | Some applications of AI are completely fine. | AI的一些应用完全没问题. |
| 510 | For instance, autocomplete in our phones, | 例如,自动完成我们的手机, |
| 511 | that's an example of AI, | 这是AI的例子, |
| 512 | or at least it would have been called AI at one point, | 或者至少应该叫AI, |
| 513 | but it's no longer necessarily called AI today. | 但它今天不一定叫AI了 |
| 514 | I'll come back to that point. | 我会回到这一点。 |
| 515 | We have a slide on that. | 我们有一个幻灯片。 |
| 516 | And on the right-hand side, | 在右边 |
| 517 | we have many applications of AI that we think are harmful. | 我们有很多AI的应用 我们认为是有害的。 |
| 518 | So if you look at the top right, for instance, | 所以,如果你看看右上方,例如, |
| 519 | we have criminal risk prediction that Sayash talked about. | 萨亚什所说的犯罪风险预测 |
| 520 | We don't think AI should be used to predict | 我们认为不应该用AI来预测 |
| 521 | who might commit a crime in the future. | 将来可能会犯罪 |
| 522 | That's just not something we should use AI for. | 我们不该用人工智能的 |
| 523 | And similarly, video interviews, | 同样,视频采访, |
| 524 | we don't think AI should be used | 我们认为不应该使用AI |
| 525 | for analyzing people's videos | 分析人们的视频 |
| 526 | to see how well they will perform at a particular job. | 看看他们在某项工作上表现如何。 |
| 527 | There's just no basis to expect that to work. | 没有理由期望它能成功 |
| 528 | So that's the X axis. | 这就是X轴 |
| 529 | But on the Y axis, what you see here | 但是在Y轴上,你在这里看到的 |
| 530 | is that regardless of whether something works well | 是,不管 东西是否工作良好 |
| 531 | or doesn't work at all, that's the Y axis, | 或者根本没用 这就是Y轴 |
| 532 | it can be either benign or harmful. | 可以是良性的,也可以是有害的。 |
| 533 | So for instance, autocomplete works really well, | 比如说,自动完成非常有效 |
| 534 | but if you look at the bottom middle, | 但是如果你看看中间的底部 |
| 535 | image generation for stock photography, | 用于 stock 摄影的图像生成, |
| 536 | it's true that image generators like Dolly or Majourney | 是真的,像多莉或Majourney这样的图像生成器 |
| 537 | or even video generators can do a really good job. | 甚至是视频生成器都能做得很好 |
| 538 | That doesn't mean that that application of AI | 这并不意味着AI的应用 |
| 539 | is completely okay. | 完全没事。 |
| 540 | There are some problems with that. | 有一些问题。 |
| 541 | The big problem with that application of AI | 适用大赦国际的重大问题 |
| 542 | is that it is trained on data created by artists, | 由艺术家创造的数据 |
| 543 | photographers, and other creative people, | 摄影师和其他有创造力的人 |
| 544 | and they are not being compensated in any way. | 而且他们没有得到任何补偿。 |
| 545 | So this is something we talk about in the book. | 所以我们在书里谈这个 |
| 546 | As a society, how can we remedy this injustice? | 作为一个社会,我们如何纠正这种不公正? |
| 547 | I think that's an important question we should ask. | 我认为这是一个我们应该问的重要问题。 |
| 548 | It's not that we should stop using image generators. | 我们不应该停止使用图像生成器。 |
| 549 | That's not what we're recommending. | 这不是我们的建议 |
| 550 | We use image generators ourselves. | 我们自己使用图像生成器。 |
| 551 | This is less of an issue for individuals to be aware of | 这对人们来说不是一个了解的问题。 |
| 552 | and more of an issue for society, for regulators, | 对于社会,对于监管者来说, |
| 553 | for the media to bring attention to and so forth. | 让媒体引起注意等等。 |
| 554 | So on the bottom left here are things | 左边的底部有东西 |
| 555 | that both work well and are not really problematic. | 两者都运作良好,而且没有真正的问题。 |
| 556 | So another example is code generation. | 所以另一个例子是代码生成。 |
| 557 | As you might know, you can use AI | 你可能知道,你可以使用人工智能 |
| 558 | to automatically generate code. | 以自动生成代码。 |
| 559 | It can even create entire apps, relatively simple apps. | 它甚至可以创建整个应用程序,相对简单的应用程序. |
| 560 | Again, these are things we make use of ourselves | 再说一遍,这些是我们利用的 |
| 561 | pretty heavily. | 相当严重。 |
| 562 | It's not a really problematic application of AI. | 这不是一个真正的问题 应用AI。 |
| 563 | We do have to think about, are there security bugs | 我们一定要考虑一下 是否有安全漏洞 |
| 564 | in the generated code, for instance? | 例如,在生成的代码中? |
| 565 | There is research showing that AI-generated code | 有研究表明AI生成的代码 |
| 566 | can have more security vulnerabilities, | 安全漏洞可能更多, |
| 567 | making it more vulnerable to hackers. | 让它更容易被黑客攻击 |
| 568 | That's something we all need to be aware of. | 这是我们都需要意识到的。 |
| 569 | But overall, that's an application of AI | 但总的来说,这是AI的应用 |
| 570 | that we very much approve of. | 我们非常赞成 |
| 571 | And going back to the top right, sorry, | 回到右上方 对不起 |
| 572 | if you could stay on that slide, Sayash, | 如果你可以留在幻灯片上,萨亚什, |
| 573 | going back to the top right, | 回到右上方, |
| 574 | we talked about criminal risk prediction. | 我们谈了犯罪风险预测 |
| 575 | We talked about video interviews. | 我们谈论了视频采访。 |
| 576 | Another one here is cheating detection. | 这里还有一个是作弊检测 |
| 577 | So what this is talking about is, as you might know, | 所以,这是在谈论, 正如你可能知道, |
| 578 | many educators are concerned that students are using | 许多教育工作者担心学生会使用 |
| 579 | ChatJPT和其他AI工具做功课. | |
| 580 | And so there are a lot of AI products | 所以有很多AI产品 |
| 581 | claiming to detect which homeworks are AI-generated. | 声称检测哪些作业是AI生成的. |
| 582 | And perhaps some of you have experience with this yourselves | 也许你们中有些人有经验 |
| 583 | that these tools don't work well, | 这些工具不起作用, |
| 584 | and all over the world, | 和全世界, |
| 585 | students are getting falsely accused of cheating using AI. | 学生被诬告利用AI作弊. |
| 586 | And we think that is a huge, huge problem. | 我们认为这是一个巨大的问题。 |
| 587 | Instructors shouldn't be using these. | 教官不应该用这些 |
| 588 | They should be modifying the education system | 他们应该改变教育制度 |
| 589 | so that they don't have to worry about | 让他们不必担心 |
| 590 | 学生是否使用 ChatJPT 。 | |
| 591 | We have been doing a lot of that work ourselves | 我们自己做了很多工作 |
| 592 | in our own teaching here at Princeton. | 我们自己在普林斯顿教书 |
| 593 | So that's the bottom left, and that's the top right. | 这就是左下方,右上方。 |
| 594 | On the other hand, | 另一方面, |
| 595 | there is something like predicting civil wars, | 有类似预言的内战, |
| 596 | which you see in the closer to the top left here. | 在离这里最近的地方。 |
| 597 | So that is a case where it's not a commercial application | 因此,这是一个案例,它不是一个商业应用 |
| 598 | of AI, but instead it is a scientific application of AI. | 但是,这是大赦国际的科学应用。 |
| 599 | A lot of papers in political science | 很多政治学论文 |
| 600 | that said we can accurately predict | 说我们可以准确预测 |
| 601 | where civil wars are going to happen using AI. | 使用AI进行内战。 |
| 602 | And our investigation showed, we have a paper on this, | 我们的调查显示,我们有一篇论文 |
| 603 | that all of those papers had errors. | 所有的文件都有错误 |
| 604 | That doesn't work at all. | 这根本行不通 |
| 605 | And just like we said, | 正如我们所说, |
| 606 | we shouldn't expect to be able to predict | 我们不应该指望能够预测 |
| 607 | people's future behavior, | 人们的未来行为, |
| 608 | who might commit a crime or who will be good at a job. | 谁可能犯罪,谁将善于工作。 |
| 609 | We also shouldn't expect to be able to predict | 我们也不应该指望能够预测 |
| 610 | where wars will happen. | 在那里,战争将发生。 |
| 611 | There are theoretical reasons to expect | 理论上是有原因的 |
| 612 | that wars are fundamentally unpredictable. | 战争根本无法预测 |
| 613 | So this is a theme that we keep coming back to | 因此,这是一个主题 我们继续回来 |
| 614 | over and over in the book, | 一遍又一遍地在书中, |
| 615 | that the future is fundamentally unpredictable. | 未来根本无法预测。 |
| 616 | So one last thing I will end with, | 所以我最后要说一件事 |
| 617 | Sayesh, if you could go to that slide. | 萨耶什,如果你可以去那个幻灯片。 |
| 618 | There is a funny definition of AI that says, | AI有个有趣的定义说, |
| 619 | AI is whatever hasn't been done yet. | AI是所有还没有做过的事情。 |
| 620 | So what does that mean? | 这是什么意思? |
| 621 | AI is whatever hasn't been done yet. | AI是所有还没有做过的事情。 |
| 622 | So when an application of AI starts working really well, | 所以当人工智能的应用开始有效时 |
| 623 | like we have a bunch of examples of these in the slide. | 就像我们在幻灯片里有很多例子 |
| 624 | So things like the Roomba or other robot vacuum cleaners, | 所以像Roomba或其他机器人吸尘器 |
| 625 | or even a web search and so forth. | 甚至网络搜索等等 |
| 626 | So all of these examples on the slide, | 因此,所有这些例子在幻灯片上, |
| 627 | we probably use on a daily basis, right? | 我们可能每天使用,对不对? |
| 628 | Speech recognition, we all dictate texts to our phones. | 语音识别 我们都会给手机发短信 |
| 629 | They work really well. | 他们的工作真的很好。 |
| 630 | And when that happens, | 而当它发生的时候, |
| 631 | it kind of fades into the background. | 它会消失在背景中 |
| 632 | We take it for granted. | 我们认为这是理所当然的。 |
| 633 | We stop calling it AI. | 我们不再叫它AI |
| 634 | It's when an application of AI is new, | 这是当一个应用AI是新的, |
| 635 | like much of generative AI today, | 就像今天的许多基因AI, |
| 636 | when it can work well, | 当它可以运行良好, |
| 637 | but also can really go wrong | 也会出错的 |
| 638 | when it has societally double edged implications. | 它具有双重社会边缘影响。 |
| 639 | That's when we are more likely to call it AI. | 那时我们更可能把它称为AI. |
| 640 | And so that's what it means. | 故其义也. |
| 641 | AI is whatever hasn't been done yet. | AI是所有还没有做过的事情。 |
| 642 | The definition of AI constantly keeps getting redefined | AI的定义不断被重新定义 |
| 643 | so that it's always at the frontier of what is possible. | 让它永远处于可能的前沿 |
| 644 | And then that's part of the reason | 然后这就是原因之一 |
| 645 | why AI doesn't have a technical definition. | 为什么AI没有技术定义. |
| 646 | It's more of a sociological definition | 这更是一个社会学的定义 |
| 647 | of what we collectively choose to call AI | 我们集体选择的AI |
| 648 | because we want to indicate | 因为我们想表明 |
| 649 | that it's cutting edge technology. | 这是尖端技术。 |
| 650 | And so what we predict | 所以,我们预测 |
| 651 | is that much of what we call AI today, | 这就是我们今天所说的AI, |
| 652 | like self-driving cars, | 像自驾车, |
| 653 | those are called AI today. | 他们今天被称为AI。 |
| 654 | But one day, even though self-driving cars today | 但有一天,尽管今天自驾车 |
| 655 | occasionally get into crashes, | 偶尔会撞车 |
| 656 | there are even deaths due to self-driving cars, | 甚至因为自驾车而死亡 |
| 657 | one day that'll be a solved engineering problem. | 有一天,那将是一个解决的工程问题。 |
| 658 | We think the number of accidents from self-driving cars | 我们认为自驾车事故的数量 |
| 659 | will drop to nearly zero. | 将降至近零。 |
| 660 | There are historical examples of this. | 有历史的例子可以说明这一点。 |
| 661 | When elevators first were a thing, | 当电梯第一次是一件事情, |
| 662 | there was a huge amount of concern about their safety. | 他们的安全受到极大关注。 |
| 663 | But now, of course, | 但现在,当然 |
| 664 | elevator accidents are extremely, extremely rare. | 电梯事故极为罕见。 |
| 665 | And so we think this will happen with self-driving cars. | 所以,我们认为这会发生 在自驾车。 |
| 666 | We won't call it AI anymore. | 我们不会再叫它人工智能了 |
| 667 | We'll take it for granted. | 以是义故. |
| 668 | In fact, we will even drop the term self-driving. | 事实上,我们甚至会放弃自我驾驶这一术语。 |
| 669 | We will just call them cars most likely, | 我们只管叫他们车子 |
| 670 | and we'll need a new term like manual car | 我们需要像手动汽车这样的新术语 |
| 671 | for what we call cars today. | 今天我们称之为汽车。 |
| 672 | And so that's our optimistic vision for the future of AI. | 这就是我们对AI未来的乐观愿景. |
| 673 | This is the kind of AI we want more of, | 这是那种我们想要更多的AI, |
| 674 | tools that do a specific task that work well | 执行具体任务的工具 |
| 675 | and don't have these kind of dubious societal implications | 并且没有这种可疑的社会影响 |
| 676 | like many other technologies like criminal risk scoring. | 就像许多其他技术 比如犯罪风险评分。 |
| 677 | Self-driving cars are going to save | 自驾车会省钱的 |
| 678 | perhaps on the order of a million lives per year | 也许每年有100万条生命 |
| 679 | that are lost today due to car accidents. | 今天因为车祸而失去的 |
| 680 | On the other hand, criminal risk scoring | 另一方面,犯罪风险得分 |
| 681 | will never become this technology | 永远不会成为这种技术 |
| 682 | that we take for granted, I think, | 我认为我们是理所当然的, |
| 683 | because the problems with it are not technical. | 因为它的问题不是技术性的。 |
| 684 | It's not that we need more data to better predict the future. | 我们不需要更多的数据来更好地预测未来。 |
| 685 | It's rather that the future is fundamentally unpredictable, | 更确切地说,未来根本上是不可预测的, |
| 686 | and it's not morally justified to predict | 而在道德上无法预测 |
| 687 | what someone is going to do | 别人会怎么做? |
| 688 | and punish them on the basis of that. | 并据此惩罚他们。 |
| 689 | It's only morally justified to punish someone | 惩罚别人才有道德上的理由 |
| 690 | on the basis of what they have already done. | 以他们已经做的为基础。 |
| 691 | And so because of that, criminal risk scoring | 因此,犯罪风险得分 |
| 692 | will not become this acceptable technology | 将不会成为这种可接受的技术 |
| 693 | that we take for granted, but self-driving cars will. | 我们当然可以 但自驾车会 |
| 694 | That's our prediction. | 这是我们的预测。 |
| 695 | And so hopefully that gives you an idea of the distinction | 希望这能给你一个区别的想法 |
| 696 | between the kinds of AI that we want more of | 我们想要更多的人工智能 |
| 697 | and the kinds of AI that we want to push back on. | 以及我们想继续的人工智能 |
| 698 | Okay, back to you, Sayas. | 好吧,回到你,萨亚斯。 |
| 699 | Fantastic, thank you, Arvind. | 太棒了 谢谢你 阿文德 |
| 700 | So maybe in the next like 10 minutes or so, | 也许在接下来的10分钟左右, |
| 701 | we'll quickly jump into one other chapter in the book | 我们很快会跳进书的另一章 |
| 702 | which talks about how AI hype persists. | 它讲述了AI的喧闹是如何持续的. |
| 703 | So I think one reason for AI hype persisting is clear, | 所以我认为一个原因 AIhype坚持是清楚的, |
| 704 | which is that companies conflate the different types of AI | 即公司将不同类型的AI混为一谈 |
| 705 | which have very little to do with each other, | 他们彼此之间几乎没有任何关系, |
| 706 | things like generative AI and predictive AI. | 比如基因AI和预测AI |
| 707 | And while generative AI has made a lot of advances | 虽然基因AI取得了很大进步 |
| 708 | over the last decade, | 过去十年, |
| 709 | predictive AI essentially relies on decades-old technology. | 预测性AI主要依靠几十年的技术. |
| 710 | But when companies conflate these two, | 但当公司把这两个人混为一谈时, |
| 711 | that results in hype about these products. | 导致这些产品的喧嚣。 |
| 712 | But I think there are also companies | 但我认为还有公司 |
| 713 | that sell generative AI tools or like, | 出售基因AI工具或类似, |
| 714 | there's essentially a lot of hype | 基本上有很多杂音 |
| 715 | around generative AI tools themselves. | 围绕基因AI工具本身。 |
| 716 | Here's what I mean by that. | 这就是我的意思。 |
| 717 | So this is a screenshot of the abstract | 这就是抽象的截图 |
| 718 | OpenAI的技术报告 | |
| 719 | when it released the GPT-4 series of models. | 当它发布GPT-4系列模型时. |
| 720 | And in that abstract was a very interesting sentence. | 在这个抽象中,有一个非常有趣的句子。 |
| 721 | So OpenAI said that GPT-4 exhibits human level performance | 所以OpenAI说,GPT-4展示了人类层面的表现. |
| 722 | on various professional and academic benchmarks, | 关于各种专业和学术基准, |
| 723 | including passing a simulated bar exam | 包括通过模拟酒吧考试 |
| 724 | with a score in the top 10% of test takers. | 在考试录取者中得分最高的10%。 |
| 725 | What this sentence was interpreted to mean very widely, | 这句话被解释为非常广泛的含义, |
| 726 | including in the press, | 包括媒体, |
| 727 | was that GPT-4 is about to replace lawyers. | 是GPT -4即将取代律师。 |
| 728 | But if you think that's what's likely to happen, | 但如果你认为这有可能发生 |
| 729 | I think that is an example of AI hype as well, | 我认为这也是AI hype的例子, |
| 730 | which was peddled quite heavily by these news organizations | 被这些新闻组织大肆推崇 |
| 731 | once GPT-4 was launched. | 在GPT-4发射后 |
| 732 | And that's because it's not a lawyer's job | 因为这不是律师的工作 |
| 733 | to answer bar exam questions all day. | 整天回答律师考试的问题 |
| 734 | And so when we rely on these benchmarks | 所以,当我们依靠这些基准 |
| 735 | to see how well these tools can be used | 以了解这些工具的使用情况 |
| 736 | to solve real world tasks, | 解决现实世界的任务, |
| 737 | or perhaps even do the job of a real person, | 或者甚至做一个真正的人的工作, |
| 738 | I think that leads to exaggerated claims. | 我认为这导致了夸张的说法。 |
| 739 | Now, these exaggerated claims are not new at all. | 现在,这些夸张的说法根本不新鲜。 |
| 740 | So this is Jeffrey Hinton, | 这是杰弗里·欣顿 |
| 741 | the recent Nobel Prize winner in 2016. | 2016年的诺贝尔奖获得者 |
| 742 | In 2016, he said, and I'm quoting him, | 2016年,他说, 我在引用他的话, |
| 743 | is, if you work as a radiologist, | 如果你是放射学家 |
| 744 | you like the coyote that's already over the edge | 你喜欢野狼已经越过边缘 |
| 745 | of the cliff, but hasn't yet looked down. | 但还没有往下看 |
| 746 | So he doesn't realize there's no ground underneath him. | 所以他不知道下面没有地盘 |
| 747 | He goes on to say, | 他接着说, |
| 748 | people should stop training radiologists now. | 人们应该停止训练放射科医生 |
| 749 | It's just completely obvious that within five years, | 很明显,五年内, |
| 750 | deep learning is going to do better than radiologists. | 深入学习会比放射科医生做得更好。 |
| 751 | Mind you, the statement was in 2016. | 请注意,声明是在2016年。 |
| 752 | In 2024, there was a worldwide shortage of radiologists. | 2024年,全世界缺乏放射学家. |
| 753 | So why is this disconnect happening? | 那么,为什么这种断开? |
| 754 | I think one of the reasons is that methods experts | 我认为原因之一是 方法专家 |
| 755 | like Jeff Hinton certainly is, | 就像杰夫・欣顿当然是, |
| 756 | so experts who sort of think about AI very deeply, | 因此,专家 排序思考AI非常深刻, |
| 757 | perhaps don't really understand | 也许不太明白 |
| 758 | what it is for domain experts, | 对于域专家来说, |
| 759 | people who work as radiologists or lawyers or what have you, | 那些做放射科医生或律师的人 或者你有什么 |
| 760 | to actually, what it takes | 实际上,它需要什么 |
| 761 | to actually do their jobs very well. | 真正做好他们的工作 |
| 762 | So they're not in the best position | 所以他们不是最好的人选 |
| 763 | to talk about the impact of AI on those jobs. | 讨论AI对这些工作的影响。 |
| 764 | So that's certainly one factor. | 这当然是一个因素。 |
| 765 | Another factor is that the AI we have today | 另一个因素是今天的人工智能 |
| 766 | perhaps might not be up to the mark itself | 也许不能达到标记本身 |
| 767 | when it comes to replacing radiologists. | 在替换放射科医生的时候 |
| 768 | So here's what happened, and this is like an example | 事情是这样的 这就像一个例子 |
| 769 | of how an AI tool that was trying to diagnose pneumonia | 试图诊断肺炎的AI工具 |
| 770 | from chest x-rays went wrong. | 胸部X光检查出了问题 |
| 771 | So this is an example of the chest x-ray, | 这是胸部X光的例子 |
| 772 | and the tool was meant to sort of diagnose pneumonia | 工具是用来诊断肺炎的 |
| 773 | on this basis. | 在这方面。 |
| 774 | But what instead happened was, | 但结果却是 |
| 775 | as you can see on the slide, on the top right, | 正如你可以看到在幻灯片, 在右上方, |
| 776 | almost the entire sort of reason the AI model | 几乎都是人工智能模型的原因 |
| 777 | was making its decisions about pneumonia | 正在决定肺炎 |
| 778 | was based on whether or not there was a hospital token | 是因为有没有医院的标志 |
| 779 | on the top right side of the image. | 在图像的右上方。 |
| 780 | So instead of looking at the lungs for evidence of pneumonia, | 所以,与其看肺部 肺炎的证据, |
| 781 | all the AI tool was doing was recognizing | AI工具所做的一切 正在承认 |
| 782 | what token the hospital had put on the top right side, | 医院在右上方放了什么标志 |
| 783 | and on this basis, making its decisions. | 并在此基础上作出决定。 |
| 784 | So this is an example of a failure in AI tools | 所以这是AI工具失败的例子 |
| 785 | that we'll come back to. | 我们会回来的 |
| 786 | But for the moment, let's come back to this sort of map | 但暂时,让我们回到这种地图 |
| 787 | of how AI hype persists. | 爱尔莎·霍普如何坚持下去。 |
| 788 | So the first reason for AI hype | 所以,第一个原因 AIhype |
| 789 | is that companies make tall claims, | 是公司提出高要求, |
| 790 | whether it's predictive AI companies | 是否是预测性的AI公司 |
| 791 | sort of claiming to sell AI | 那种声称出售AI |
| 792 | and conflating between different types of AI, | 以及不同类别AI之间的混杂, |
| 793 | or generative AI companies that overcame | 或基因化AI公司 克服 |
| 794 | how well their tools work and how general they are. | 他们的工具如何运作,他们如何是普通的。 |
| 795 | Now, let's head back to the sphere of research | 现在,让我们回到研究领域去 |
| 796 | for a minute. | 一分钟。 |
| 797 | So how do we tell in AI research | 所以,我们如何在AI的研究中说 |
| 798 | how well a system performs? | 系统表现如何? |
| 799 | The best way of doing this | 最好的办法 |
| 800 | and the most widespread way of doing this today | 以及今天最普遍的方法 |
| 801 | is by using something called benchmark datasets. | 是通过使用一个叫做基准数据集。 |
| 802 | So for OpenAI's example, the dataset that they used | 以OpenAI为例,他们使用的数据集 |
| 803 | was that of a simulated bar exam, | 是模拟律师考试, |
| 804 | where they took a number of questions, | 他们问了一些问题, |
| 805 | mostly multiple choice questions, | 多数选择问题, |
| 806 | and asked the model to respond with answers. | 并要求模特儿回答 |
| 807 | And what you then do is you train the AI systems | 然后你做的是训练人工智能系统 |
| 808 | on one part of it. | 在其中的一部分。 |
| 809 | But crucially, you set aside, | 但关键是,你让开 |
| 810 | let's say around 30% or half of the test of the dataset | 假设数据集测试的30%或一半左右 |
| 811 | to evaluate the AI models on that test set. | 评估测试组上的AI模型。 |
| 812 | And this part is essential to machine learning more broadly | 而这部分对于机器学习更广泛至关重要 |
| 813 | because if you don't separate out this dataset, | 因为如果你不分开这个数据集, |
| 814 | if you evaluate your AI on the same dataset | 如果你在同一个数据集上评价你的AI |
| 815 | it was trained on, | 它被训练在, |
| 816 | then you're essentially teaching to the test. | 那你基本上就是在教测试 |
| 817 | So the AI tool can essentially memorize the examples | 因此,AI工具基本上可以记住实例 |
| 818 | in the dataset that it was trained on. | 在它所训练的数据集中。 |
| 819 | And this is an example of what we call leakage. | 这就是我们所谓的渗漏的例子。 |
| 820 | So there is information that is leaked about the dataset | 因此,有信息泄露 关于数据集 |
| 821 | to the AI tool by virtue of being trained | 通过培训,加入AI工具 |
| 822 | on the entire dataset. | 整个数据集。 |
| 823 | And there's no sort of hidden test set | 没有隐藏的测试集 |
| 824 | that it can be evaluated on. | 它可以被评价。 |
| 825 | So this is one example of leakage. | 所以这是渗漏的一个例子。 |
| 826 | There are other examples as well. | 还有其他例子。 |
| 827 | So the chest x-ray image that we saw a few slides earlier | 因此,胸部X射线图像,我们看到一些幻灯片之前 |
| 828 | is an example of what happens when we rely on these AI tools | 就是我们依靠这些人工智能工具 会发生什么的例子 |
| 829 | {\fn黑体\fs22\bord1\shad0\3aHBE\4aH00\fscx67\fscy66\2cHFFFFFF\3cH808080}只有像那些已经隐藏的答案 | |
| 830 | in the chest x-ray. | 胸前X光检查 |
| 831 | So if there's like a token that indicates | 所以,如果有像一个标志 显示 |
| 832 | that a chest x-ray scan is from a particular hospital | 胸部X光扫描来自某家医院 |
| 833 | that makes it more likely for the AI tool to work well | 这使得人工智能工具更容易运作 |
| 834 | on images from that hospital, but not from other ones. | 在医院的照片上,但不是其他的。 |
| 835 | So as I mentioned a few years ago, | 所以,正如我几年前提到的, |
| 836 | we started looking at this problem | 我们开始研究这个问题 |
| 837 | in the field of civil war prediction. | 在内战预测领域。 |
| 838 | And we found that every single paper | 我们发现每张纸 |
| 839 | which claimed that AI tools did better | 声称AI工具做得更好 |
| 840 | than decades old regression methods, | 数十年来的回归方法, |
| 841 | every single one of those papers | 每一个文件 |
| 842 | suffered from some of the other form of leakage. | 受到某种其他形式的渗漏的影响。 |
| 843 | And in fact, this sort of led us to investigate | 事实上,这导致我们调查 |
| 844 | whether leakage is also an issue in other science fields | 渗漏是否也是其他科学领域的一个问题 |
| 845 | that are adopting AI and machine learning. | 正在通过人工智能和机器学习。 |
| 846 | And when we looked at ML-based science, | 当我们看着基于ML的科学, |
| 847 | that is scientific research that uses machine learning, | 这是利用机器学习的科学研究, |
| 848 | we found that leakage is widespread in these fields. | 我们发现在这些领域渗漏很普遍。 |
| 849 | So this is a screenshot from a paper from 2023, | 这是2023年一篇论文的截图 |
| 850 | where we found that over 294 papers | 我们发现294份论文 |
| 851 | across 17 different scientific fields suffered from leakage. | 共有17个不同的科学领域受到渗漏的影响。 |
| 852 | And perhaps more surprisingly, | 也许更令人惊讶的是, |
| 853 | we found that each of these fields | 我们发现 每一个领域 |
| 854 | was independently rediscovering what it means | 是独立地重新发现它的意思 |
| 855 | for AI results to be impacted by leakage. | AI结果受到渗漏的影响。 |
| 856 | We have since updated this survey. | 我们自此更新了这项调查。 |
| 857 | So in our latest run, we found over 600 papers | 所以最近我们发现600多份论文 |
| 858 | across 30 different fields suffering from leakage. | 覆盖30个不同田地 受到渗漏影响 |
| 859 | And we think it very much is contributing | 我们认为它很有贡献 |
| 860 | to a crisis in science. | 在科学危机。 |
| 861 | In fact, this crisis has been outlined | 事实上,这场危机已经概述 |
| 862 | by several sort of journalistic pieces, | 以数种新闻片, |
| 863 | and others have also caught onto it. | 其他人也发现了 |
| 864 | We've heard about data leakage causing issues | 我们听说数据泄露引起问题 |
| 865 | in healthcare, in machine learning, | 在保健,机器学习, |
| 866 | research in other scientific disciplines, and so on. | 其他科学学科的研究,等等。 |
| 867 | And while there are genuine advances absolutely | 虽然有真正的进步 绝对 |
| 868 | that AI has led to things like alpha fold | AI导致了类似α折叠的东西 |
| 869 | and protein folding models | 和蛋白质折叠模型 |
| 870 | that have rightly been recognized, | 他们确已被确认, |
| 871 | there is also a vast amount of research | 还有大量的研究 |
| 872 | that suffers from this reproducibility crisis | 受到这种可复制危机的伤害 |
| 873 | primarily because of leakage. | 主要因为渗漏。 |
| 874 | Okay, so let's come back to our map of how AI hype persists. | 好吧,让我们回到我们的地图 如何AI的hype坚持。 |
| 875 | So our first part was companies making tall claims | 所以,我们的第一部分是公司 提出高要求 |
| 876 | without transparency. | 没有透明度。 |
| 877 | In addition to that, AI research is suffering | 除此之外,AI的研究也在受苦 |
| 878 | from a reproducibility crisis. | 从可复制的危机。 |
| 879 | And what these two things together mean | 这两件事的意义 |
| 880 | is that most of the prominent AI results that we see | 大部分显著的AI结果,我们看到 |
| 881 | are likely to be exaggerated. | 可能会被夸大 |
| 882 | On top of that, we've also seen a lot of input | 此外,我们还看到很多投入 |
| 883 | by public figures. | 根据公众人物。 |
| 884 | So for example, when GPT-4 came out, | 例如GPT-4出来后 |
| 885 | many influential people signed this open letter | 许多有影响力的人都签署了这份公开信 |
| 886 | that called for a pause on giant AI experiments. | 这要求暂停巨型AI实验 |
| 887 | This letter also mentioned that AI that is more powerful | 这封信还提到,大赦国际的权力更大。 |
| 888 | than GPT-4 could pose civilization level risks, | 超过GPT-4可能造成文明程度的风险, |
| 889 | perhaps even the risk of extinction. | 甚至还有灭绝的危险 |
| 890 | And on that basis, all of these signatories | 在此基础上,所有签署者 |
| 891 | were asking to pause these experiments. | 他们要求暂停这些实验 |
| 892 | We call this an example of critique hype, | 我们称这为批评胡言乱语的例子, |
| 893 | that is criticism at face value, | 也就是正面的批评, |
| 894 | which also leads to hype about AI. | 这也导致了关于AI的喧嚣. |
| 895 | In this case, for instance, | 例如,在这种情况下, |
| 896 | by saying that AI will soon be powerful enough | 说AI很快会足够强大 |
| 897 | to pose civilization wide risks. | 给文明带来巨大的风险。 |
| 898 | And I think this is an example of something | 我想这是个例子 |
| 899 | that's quite pervasive today, | 今天很普遍, |
| 900 | and especially by public figures | 特别是公众人物 |
| 901 | who talk about the harms of AI. | 他们谈论AI的伤害。 |
| 902 | So coming back to the map, | 所以回到地图上 |
| 903 | in addition to prominent AI results being exaggerated, | 除了显著的AI结果被夸大外, |
| 904 | public figures also distract from the real issues | 公众人物也分散对实际问题的注意力 |
| 905 | of misinformation that is likely to arise | 可能发生的错误信息 |
| 906 | when people over rely on these tools | 当人们依赖这些工具时 |
| 907 | or perhaps the labor implications of tools | 或也许工具对劳动力的影响 |
| 908 | by pointing to sci-fi threats like civilizational collapse. | 指向科幻威胁 比如文明崩溃 |
| 909 | And both of these together mean | 两者都意味着 |
| 910 | that there is rampant AI hype | 有猖獗的AI 谣言 |
| 911 | in many of the sources of information | 在许多信息来源中 |
| 912 | that are available to people. | 给人们提供 |
| 913 | Finally, the way this information is conveyed to the public | 最后,如何向公众传播这一信息 |
| 914 | closes the feedback loop of hype. | 关闭hype的反馈循环。 |
| 915 | So here's a screenshot | 这是一张截图 |
| 916 | of what popular news headlines look like | 大众新闻头条是什么样子 |
| 917 | right after Microsoft launched their Bing's AI chat, | 就在微软推出他们的AI聊天后 |
| 918 | which relied on GPT-4 under the hood. | 它依赖于GPT-4 在引擎盖下。 |
| 919 | The New York Times said, | 纽约时报说, |
| 920 | Bing's AI chat, I want to be alive. | Bing的AI聊天 我想活着 |
| 921 | There were other outlets, for instance, | 还有其他渠道,例如, |
| 922 | saying things like, are AI chat bots turning sentient? | 说着什么 AI聊天会变得有灵性吗? |
| 923 | And we think that when the public is faced | 我们认为当公众面对 |
| 924 | with headline after headline, | 标题后为标题, |
| 925 | claiming that AI could perhaps be sentient, | 声称大赦国际可能具有灵敏性, |
| 926 | that has a really negative effect | 这真的有负面效应 |
| 927 | on how the public could treat AI systems, | 关于公众如何对待AI系统, |
| 928 | perhaps as being sentient. | 也许是有意识的 |
| 929 | The really sad thing about all of this | 这一切真可悲 |
| 930 | is that we have known for at least six decades now | 我们至少已经知道60年了 |
| 931 | that when humans interact with AI-based chat bots, | 当人类与基于AI的聊天室互动时 |
| 932 | they are likely to anthropomorphize them, | 他们有可能使他们变成人类, |
| 933 | to treat them as if they were human-like. | 把他们当人一样对待 |
| 934 | So for example, this is an image of Eliza, | 比如说 这是伊丽莎的影像 |
| 935 | a chat bot that was developed by Joseph Weissenbaum in 1966. | 1966年由约瑟夫·魏森鲍姆开发的聊天机器人. |
| 936 | Eliza is an example of a really simple chat bot. | 伊丽莎是一个非常简单的聊天机器人的例子。 |
| 937 | All it is doing is sort of parroting back your responses | 它所做的只是 某种程度地解释你的答复 |
| 938 | in the form of a question, | 以问题的形式, |
| 939 | and even technically, it is extremely rudimentary. | 即使是在技术上,它也是极其初级的。 |
| 940 | It just has a list of rules on the basis of which | 它有一份规则清单 在此基础上 |
| 941 | it asks the next question or sends the next message. | 它问下一个问题 或发送下一个信息。 |
| 942 | So for example, Eliza could start with something like, | 比如说,伊丽莎可以从... |
| 943 | is something troubling you? | 有什么问题吗? |
| 944 | And the person might respond, men are all alike. | 这个人可能会回应 男人都一样 |
| 945 | Eliza would just respond | 伊丽莎会回应的 |
| 946 | with a fairly generic question in this case. | 还有一个很普通的问题 |
| 947 | What is the connection, do you suppose? | 你觉得有什么联系? |
| 948 | And so on and so forth. | 及诸众生. |
| 949 | The really interesting thing about Eliza | 关于伊丽莎的有趣的事 |
| 950 | is not how it was designed, | 这不是设计它的方式, |
| 951 | but rather the effect it had | 但它的影响 |
| 952 | on people who interacted with it. | 与它互动的人。 |
| 953 | Many of the people who talked to this chat bot | 很多和这个聊天机器人交谈的人 |
| 954 | would, after the conversation, refuse to believe | 在谈话后 拒绝相信 |
| 955 | that they just talked to an AI chat bot, | 他们刚跟一个AI聊天机器人谈过了 |
| 956 | that they hadn't really talked to another human. | 他们没有真的跟另一个人说话。 |
| 957 | This was termed the Eliza effect | 这被称为伊丽莎效应 |
| 958 | as an example of the implications of treating AI systems | 作为处理AI系统所涉问题的一个例子 |
| 959 | as if they are human-like. | 就像他们像人类一样 |
| 960 | And so given that we have this vast amount of understanding | 所以考虑到我们有 如此巨大的理解 |
| 961 | and like many experiments, | 像很多实验一样 |
| 962 | showing that humans tend to treat AI systems as human-like, | 显示人类倾向于将人工智能系统视为人一样, |
| 963 | I think it is particularly sad in some sense | 我觉得从某种意义上说,这特别可悲 |
| 964 | that these news outlets claimed that Bing's AI chat | 这些新闻声称Bing的AI聊天 |
| 965 | in this case was turning sentient. | 在这个案子中 变得很敏感 |
| 966 | So rather than informing their readers | 而不是通知读者 |
| 967 | about how to watch out for symptoms like the Eliza effect, | 如何注意象伊丽莎那样的症状 |
| 968 | they were actually feeding in to the hype. | 他们其实是喂进 杂音。 |
| 969 | And this, I think, completes the feedback loop of AI hype. | 我认为这完成了AI的回馈循环。 |
| 970 | So when journalists uncritically report on AI | 所以当记者不严谨地报道AI时 |
| 971 | and exploit our cognitive biases like anthropomorphism, | 利用我们的认知偏见 像人类形态学, |
| 972 | this essentially completes the feedback loop of AI hype, | 这基本上完成了AIhype的反馈循环, |
| 973 | especially in the public sphere. | 特别是在公共领域。 |
| 974 | Okay, and I'll quickly hand it back to Arvind | 好,我马上交给阿文德 |
| 975 | for a brief overview of what else is in the book. | 简略地概述书中的其他内容。 |
| 976 | Sure, maybe I will just take one minute | 当然,也许我只需要一分钟 |
| 977 | just to share some final thoughts | 只是想谈谈最后的想法 |
| 978 | and then really look forward to what Thomas has to say. | 然后真的期待 托马斯要说的话。 |
| 979 | So many people started paying attention to AI | 很多人开始关注AI |
| 980 | after ChatGPT came out, but in fact, it goes back 80 years. | 在ChatGPT出来后,但事实上,它可以追溯到80年前. |
| 981 | So this is a screenshot from the 1950s | 这是1950年代的截图 |
| 982 | and the history of neural networks | 神经网络的历史 |
| 983 | actually goes back to the 1940s. | 实际上可以追溯到1940年代. |
| 984 | Let's keep going on those slides, Sayesh. | 让我们继续这些幻灯片,萨耶什。 |
| 985 | So we have a whole chapter discussing why it is that | 所以我们有一整章讨论为什么 |
| 986 | we've talked a little bit about what goes wrong | 我们聊了一下出了什么事 |
| 987 | when we try to use AI to predict the future. | 当我们试图使用AI来预测未来时. |
| 988 | And I briefly mentioned research | 我简短地提到研究 |
| 989 | by our Princeton colleague, Matt Salgonik. | 我们普林斯顿的同事 马特·萨尔戈尼克 |
| 990 | We summarized that research. | 我们总结了这一研究。 |
| 991 | We have a lot of other research. | 我们有很多其他的研究。 |
| 992 | And there's a whole chapter in the book | 书里有一整章 |
| 993 | that doesn't even talk about AI, | 连人工智能都没说 |
| 994 | but it's more about a sociological understanding | 但更多的是社会学的理解 |
| 995 | of why the future is hard to predict. | 为什么未来很难预测。 |
| 996 | Then there is existential risk | 那还有生存风险 |
| 997 | that Sayesh already mentioned. | 萨耶什已经提到。 |
| 998 | We talk about the role of institutions. | 我们谈到了机构的作用。 |
| 999 | I'll save this for the Q&A. | 我会留着给QQA |
| 1000 | What we mean by broken AI | 我们的破解AI的意思 |
| 1001 | is appealing to broken institutions. | 正在吸引破碎的机构。 |
| 1002 | We talk about regulations a little bit. | 我们谈一些条例 |
| 1003 | We talk about AI and children, | 我们谈论AI和孩子们, |
| 1004 | the role that AI is going to play | 大赦国际将发挥的作用 |
| 1005 | in the life of a child born today, for instance, | 例如在今天出生的孩子的生活中, |
| 1006 | we think is going to be much more significant | 我们认为会更重要 |
| 1007 | than it plays in our lives today. | 而不是它在我们今天的生活里 |
| 1008 | But anyway, we'll save all that for later. | 但是,无论如何,我们会保存这一切 以后。 |
| 1009 | For now, we just want to mention | 现在,我们只想提一下 |
| 1010 | that the book is available for purchase. | 书可以买到 |
| 1011 | And with that, we look forward to hearing | 因此,我们期待听到 |
| 1012 | what you have to say. | 你该说什么? |
| 1013 | Thank you. | 谢谢 |
| 1014 | Thanks, Mr. Narayana and Mr. Kapoor | 谢谢,纳拉亚纳先生和卡普尔先生 |
| 1015 | to give us this presentation | 给我们介绍一下 |
| 1016 | and to give us a very fundamental definition | 给我们一个非常基本的定义 |
| 1017 | 何谓AI,何谓AI, | |
| 1018 | and what AI can do, | 以及大赦国际能做什么, |
| 1019 | and what those kind of the problems | 什么样的问题 |
| 1020 | that AI cannot solve, | AI无法解决, |
| 1021 | especially I really admire this part, | 特别是我真的很欣赏这个部分, |
| 1022 | like just Mr. Narayana said, | 就像纳赖亚纳先生说的 |
| 1023 | okay, AI cannot predict the future. | 好吧,AI无法预测未来。 |
| 1024 | And he just gave us some kind of the examples | 他只是给我们一些例子 |
| 1025 | and the reasons why AI cannot predict future, | 以及大赦国际无法预测未来的原因, |
| 1026 | which is really, I think this is a really crucial part | 这是真的,我认为 这是一个真正的关键部分 |
| 1027 | for us to understand what AI can do | 为了让我们明白AI能做什么 |
| 1028 | and what AI cannot do, right? | 和什么AI不能做的,对不对? |
| 1029 | So, but my question, | 所以,但我的问题, |
| 1030 | yeah, I will raise some kind of the question. | 是的,我会提出 某种问题。 |
| 1031 | So my question may pretty focus | 所以,我的问题可能相当集中 |
| 1032 | on the part of the AI hype, | 由AI的歇斯底里, |
| 1033 | because we really every day, | 因为我们真的每天都, |
| 1034 | you know, I'm a journalist entrepreneur | 我是记者企业家 |
| 1035 | and we pretty focused on AI | 我们相当专注于AI |
| 1036 | and we do a bunch of this kind of the AI coverages. | 我们做一堆这样的AI覆盖。 |
| 1037 | We talk with the AI figures and AI companies | 我们跟AI的人物和AI公司谈过 |
| 1038 | and to get a deeply experienced | 并获得一个深刻的经验 |
| 1039 | down this kind of the AI hardware, | 在这种AI硬件, |
| 1040 | AI devices, AI language models, | AI设备,AI语言模型, |
| 1041 | or this kind of things. | 或这种东西。 |
| 1042 | And we do a lot of this kind of things | 我们做很多这样的事情 |
| 1043 | and sometimes we sort of like, | 有时候我们喜欢, |
| 1044 | and we are in the way to blindfold it, | 并且阻碍我们遮住它。 |
| 1045 | run towards this kind of the AI papers, | 跑到这种AI文件, |
| 1046 | or the AI companies, or the AI prototypes, | 或AI公司,或AI原型, |
| 1047 | or AI products. | 或AI产品。 |
| 1048 | We strongly feel that there are a lot of this kind | 我们强烈地感到 有很多这样的 |
| 1049 | of the AI hypes. | 爱尔莎·霍普斯 |
| 1050 | Yeah, I can just raise with an example | 对,我可以举个例子 |
| 1051 | that I like a week ago, | 我喜欢一个星期前, |
| 1052 | my company hosted our own AI event | 我公司承办了我们的AI活动 |
| 1053 | in Beijing right over here in Beijing. | 在北京就在这里 |
| 1054 | Before the event, I set up a rule, | 事发前 我订了一条规矩 |
| 1055 | I set up a rule to prohibit our speakers | 我订了一条禁止发言的规则 |
| 1056 | to use the buzzwords like AGI or scatting law | 使用诸如 AGI 或 分解定律 的 蜂鸣词 |
| 1057 | or this kind of things. | 或这种东西。 |
| 1058 | The rule I set up, yeah, | 我规定的规则,是的, |
| 1059 | this rule is established because each speaker | 本规则的确立是因为每个发言者 |
| 1060 | or panelist can put his or her own stamp | 或主讲人可以自己盖章 |
| 1061 | on the envelope of AGI or generative AI | 关于AGI或基因AI的信封 |
| 1062 | or scatting law. | 或散诸法. |
| 1063 | Each one can give their own understanding on that | 每个人可以给出自己的理解 |
| 1064 | which causes more misconceptions or confusions, I think. | 我认为这会引起更多的误解或混淆。 |
| 1065 | So during the event, I think I tried my best | 所以在活动中,我想我尽力了 |
| 1066 | to remind each one, okay, we won't talk about AGI today. | 提醒每个人,好吗,我们今天不谈AGI。 |
| 1067 | Don't mention that word. | 别提这个词了 |
| 1068 | Don't mention scatting law. | 别提散乱法. |
| 1069 | What you are doing is not based on the scatting law, okay? | 你这样做不是根据 散乱法,好吗? |
| 1070 | We cannot say that. | 我们不能这么说。 |
| 1071 | I tried my best to do that, | 我尽力了 |
| 1072 | but still, in some of the sessions | 但是,在一些会议上 |
| 1073 | and we unconsciously talk about AGI | 我们无意识地谈论AGI |
| 1074 | or this kind of things, yeah. | 或这种事情,是的。 |
| 1075 | I think this reflects how we get used | 我觉得这反映了我们是如何被利用的 |
| 1076 | or how the whole industry, I mean the AI industry, | 或者整个行业,我的意思是AI行业, |
| 1077 | AI industry got used to enjoying this kind of the best words. | AI行业习惯了享受这种最好的词. |
| 1078 | Even most of the time we are probably | 即使大多数时候我们可能 |
| 1079 | unaware of this kind of a situation. | 不知道这种情况 |
| 1080 | Yeah, so let's talk deeper about the AI hype. | 是啊,让我们更深入地谈谈AI的喧闹。 |
| 1081 | Yeah, I think the last chart is talking about | 是啊,我想最后的图表 是谈论 |
| 1082 | 谁做了这种AI的hype | |
| 1083 | or this kind of the exaggerated AI claims. | 或这种夸张的AI声称。 |
| 1084 | Who made that kind of things? | 谁做的那种东西? |
| 1085 | I think both of you conclude | 我想你们两个都说完了 |
| 1086 | this kind of the people into three groups. | 这种人分成三组 |
| 1087 | Three groups, one is that the companies, | 三组,一个是公司, |
| 1088 | the AI companies who are selling their products, | 销售其产品的AI公司, |
| 1089 | their AI products or their AI solutions. | 他们的AI产品或AI解决方案。 |
| 1090 | The second group is the AI researchers, | 第二组是AI研究者 |
| 1091 | like you guys, the AI researchers | 像你们一样,人工智能研究人员 |
| 1092 | who spread out a lot of these papers | 他们散发了很多这些文件 |
| 1093 | or this kind of things. | 或这种东西。 |
| 1094 | And third group is the people like us, | 第三组是像我们这样的人 |
| 1095 | the journalists or the reporters or editors | 记者或记者或编辑 |
| 1096 | who are covering AI. | 他们正在掩护大赦国际。 |
| 1097 | So my first question may start from the group like us. | 我的第一个问题可能来自我们这样的群体。 |
| 1098 | Yeah, so no, I just noticed different kind of the medias | 是啊,所以没有,我只是注意到 不同的媒体类型 |
| 1099 | have different attitudes or standing points | 有不同的态度或立场 |
| 1100 | on AI companies, I think, yeah. | 有关AI公司,我想,是的。 |
| 1101 | For professional or very vertical tech | 专业或非常垂直的技术 |
| 1102 | or AI media or AI outlets, I think we always | 或AI媒体或AI机构, 我认为我们总是 |
| 1103 | just portray over optimistic light | 只是描绘出乐观的光芒 |
| 1104 | to AI products or AI companies. | - AI产品或AI公司。 |
| 1105 | While the mass media always remind people | 虽然大众传媒总是提醒人们 |
| 1106 | the harm or this kind of the risk of AI, | AI的伤害或这种风险, |
| 1107 | I think this is very interesting. | 我觉得这很有趣 |
| 1108 | But each of this kind of the media get traffic, | 但每类媒体都有流量 |
| 1109 | get traffic, get this kind of the people's attraction, | 堵车 吸引人们 |
| 1110 | I think. | 我觉得 |
| 1111 | In my team, I'm always struggling on this kind of things. | 在我的团队里,我总是在为这种事挣扎 |
| 1112 | I always, I sometimes guard my team | 我总是,我有时守护我的团队 |
| 1113 | to use those kind of the very moderate words | 用那种温和的词 |
| 1114 | or very balanced tones to tell AI stories, | 或非常平衡的音调来讲述AI的故事, |
| 1115 | to describe AI product or to comment | 描述AI产品或评论 |
| 1116 | how AI company make senses to the society, to its users. | AI公司如何给社会,给用户带来意义. |
| 1117 | But I sometimes struggle to get traffics | 但我有时会拼命去堵车 |
| 1118 | competing with those kind of the other medias | 与其他媒体竞争 |
| 1119 | who have very strong opinion, | 他们有很强的意见, |
| 1120 | who provides very fruitful emotional value | 提供非常有成果的情感价值 |
| 1121 | instead of insight provider. | 而不是提供洞察力 |
| 1122 | So I think, I'm not sure, how can you guys tell me, okay, | 所以我想,我不知道, 你们怎么能告诉我,好吗, |
| 1123 | whether is this a unsolvable dilemma | 这是否是一个无法解决的难题 |
| 1124 | or how the mass media or no matter the mass media | 或大众传媒如何或不论大众传媒如何 |
| 1125 | or even the tech media can do this kind of things better | 或甚至科技媒体 可以做这样的事情更好 |
| 1126 | to let more people, not only for the industry, | 让更多的人, 不仅仅是为行业, |
| 1127 | but the masses of all the people | 但人民大众 |
| 1128 | to have a more clearer and neutral understanding | 更清楚和中立的理解 |
| 1129 | on what AI is, how it works, | 关于AI是什么,它是如何工作的, |
| 1130 | and what's the background technology behind that. | 背后的背景技术是什么? |
| 1131 | Thank you. | 谢谢 |
| 1132 | Yeah, that's a very hard problem | 是啊,这是一个非常困难的问题 |
| 1133 | from the perspective of journalists | 从记者的角度来看 |
| 1134 | to know how to cover AI in a more responsible way | 了解如何以更负责的方式覆盖大赦国际 |
| 1135 | while also competing with other journalists | 同时也与其他记者竞争 |
| 1136 | who might not have the same standards. | 他们可能没有同样的标准。 |
| 1137 | I think in general, this is a challenge | 我觉得总的来说,这是个挑战 |
| 1138 | all throughout journalism, not just in AI. | 整个新闻,不只是在AI。 |
| 1139 | And so I think the solutions are going to have to be broader | 所以我认为解决方案必须更加宽泛 |
| 1140 | and maybe not specifically about AI. | 也许不是关于AI。 |
| 1141 | I can share a couple of thoughts, but you're the expert. | 我可以分享一些想法 但你是专家 |
| 1142 | I'm curious to hear what has worked for you | 我很想知道什么对你有用 |
| 1143 | and what has worked less well for you as well. | 以及那些对你不太有用的东西 |
| 1144 | So one thought is that as you pointed out, | 所以,有一个想法是, 正如你指出, |
| 1145 | this is very tied to the business models. | 这与商业模式密切相关。 |
| 1146 | If the business model is based on clicks, of course, | 如果商业模式是基于点击,当然, |
| 1147 | it's going to be hard to have more in-depth coverage. | 要进行更深入的报导是很难的. |
| 1148 | One thing I've noticed | 我注意到一件事 |
| 1149 | in the United States media ecosystem | 美国媒体生态系统 |
| 1150 | is that the journalistic organizations | 是记者组织 |
| 1151 | that want to provide in-depth factual coverage | 希望提供深入的事实报道 |
| 1152 | tend to move towards a different business model. | 倾向于走向不同的商业模式。 |
| 1153 | It's subscription-based, | 它的订阅, |
| 1154 | where readers want to come back month after month | 读者希望月复一月 |
| 1155 | because they know that this outlet | 因为他们知道这个出口 |
| 1156 | is going to give them good information. | 将会给他们好的信息。 |
| 1157 | Some of them are becoming nonprofits | 他们中有些人正在成为非营利组织 |
| 1158 | where they are funded by donations | 由捐款供资 |
| 1159 | in order to do in-depth reporting. | 以便进行深入的报告。 |
| 1160 | And that's not necessarily just for AI. | 这不一定只是AI的。 |
| 1161 | That has been true historically | 历史上就是这样 |
| 1162 | where a lot of the most in-depth reporting | 其中很多最深入的报告 |
| 1163 | comes from more nonprofit journalistic organizations | 更多来自非营利新闻组织 |
| 1164 | than the ones who are competing for clicks. | 胜过那些竞拍者。 |
| 1165 | Another is just to have self-regulation, | 另一种就是自我约束 |
| 1166 | I think, among journalists. | 我想,在记者中间。 |
| 1167 | And we have been to a few journalism conferences, | 我们参加了几次新闻会议, |
| 1168 | for instance, again, here in the US, | 例如,在美国这里, |
| 1169 | where journalists are discussing how we can all do better | 记者们正在讨论我们如何做得更好 |
| 1170 | and hold each other accountable, right? | 互相责备,对吧? |
| 1171 | Where they work together | 他们一起工作的地方 |
| 1172 | instead of necessarily always competing with each other. | 而不是总是相互竞争 |
| 1173 | And one last third suggestion | 最后三分之一的建议 |
| 1174 | is perhaps a different kind of content | 也许是另一种内容 |
| 1175 | when you're covering something that's in the news, | 当你在报道新闻时 |
| 1176 | when you want to cover a product | 当你想要覆盖一个产品时 |
| 1177 | that a company put out, for instance. | 比方说,一家公司已经解散了 |
| 1178 | It's hard to do that without hype | 胡说八道很难做到 |
| 1179 | because you have to make it seem | 因为你必须让它看起来 |
| 1180 | like this is something remarkable, | 像这样的东西是了不起的, |
| 1181 | something that is worth the reader or viewer or listener | 值得读者、观众或听众欣赏的东西 |
| 1182 | paying attention to. | 关注。 |
| 1183 | And when you're covering something dangerous about AI, | 当你掩盖AI危险的事情时 |
| 1184 | again, you have to hype it up | 再说一次,你必须把它弄乱 |
| 1185 | because you have to make it seem like this is very special. | 因为你得让它看起来很特别 |
| 1186 | But there is another kind of content, right? | 但还有另一种内容,对不对? |
| 1187 | Another kind of content might be an interview | 另一种内容可能是采访 |
| 1188 | with someone who is working in AI, for instance. | 例如在AI工作的人 |
| 1189 | And it seems like people really love | 好像人们真的喜欢 |
| 1190 | these in-depth interviews, podcasts, you know? | 这些深入的采访,播客,你知道吗? |
| 1191 | So for instance, in the US, | 例如,在美国, |
| 1192 | we have podcasts like Joe Rogan, | 我们有播客 像乔罗根, |
| 1193 | which have become enormously successful, | 他们已成功, |
| 1194 | have become media empires on their own | 自己成为媒体帝国 |
| 1195 | and really competing with traditional media outlets. | 与传统媒体竞争 |
| 1196 | So that is surprising to me. | 这让我很惊讶 |
| 1197 | And it's a sign that people really want | 这是人们真正想要的标志 |
| 1198 | this kind of in-depth content | 这种深入的内容 |
| 1199 | that comes from our journalists talking to an expert. | 这来自于我们的记者 与专家交谈。 |
| 1200 | So that could be another kind of content | 所以,这可能是另一种内容 |
| 1201 | where you can have more depth. | 在那里你可以有更深的深度。 |
| 1202 | That'd be great. | 这将是伟大的。 |
| 1203 | You know, just for my own experience, | 你知道,只是为我自己的经验, |
| 1204 | recently, I mean, just during the past probably a year, | 最近,我的意思是, 只是在过去可能一年, |
| 1205 | I mean, the major channels for me to get deeper | 我的意思是,主要渠道 我更深入 |
| 1206 | and understand on the most cutting edge AI technology | 并了解最前沿的AI技术 |
| 1207 | or this kind of the models or this kind of the researchers | 或这种模型或这种研究者 |
| 1208 | or the business ideas are coming from | 或商业想法来自 |
| 1209 | those kind of the podcasts hosted by VC firms | 此类播客由《维也纳公约》公司主办 |
| 1210 | like Andreus and Horace or Y Combinator | 像安德烈乌斯和贺拉斯或Y组合 |
| 1211 | or NoPriori or this kind of the VC guys | 或无主或这种越共的家伙 |
| 1212 | who are hosting this kind of the postcards | 他们主持这种明信片 |
| 1213 | 而不是 TechCrunch、 Verge 或 商业内幕 | |
| 1214 | or this kind of the mass media. | 或这种大众传媒。 |
| 1215 | I think this is because this kind of the, | 我觉得这是因为这种, |
| 1216 | I mean, the insiders, | 我的意思是,内线, |
| 1217 | no matter the VC guys or entrepreneurs, | 不管越共的人和企业家 |
| 1218 | they can get better understanding on what they're doing. | 他们可以更好地了解自己在做什么。 |
| 1219 | But one following question, | 但有一个问题 |
| 1220 | highly relevant to that is that, | 与此密切相关的是, |
| 1221 | do you think that the, I mean, the reporters | 你觉得,我的意思是,记者 |
| 1222 | or the journalists who are covering AI | 或报道大赦国际的记者 |
| 1223 | need to read AI research papers frequently, | 需要经常阅读AI研究论文, |
| 1224 | probably in a weekly basis | 也许每周一次 |
| 1225 | or research papers frequently to know more about that | 或研究论文 经常了解更多 |
| 1226 | because most of the journalists include myself. | 因为大多数记者包括我自己 |
| 1227 | I read papers frequently, but for myself, I think, | 我经常看报纸,但我觉得 |
| 1228 | I'm still with not an AI background | 我还是没有AI的背景 |
| 1229 | or a computer science background. | 或计算机科学背景。 |
| 1230 | I'm not with that background. | 我没有那种背景 |
| 1231 | It's sometimes just give me some kind of the challenges | 有时候只是给我一些挑战 |
| 1232 | to read some details, | 阅读一些细节, |
| 1233 | but thanks for AI tools like Chaijibiti or Cloud, right? | 但是谢谢你的人工智能工具 比如Chaijibiti或Cloud,对不对? |
| 1234 | Do you think it's really needed for, | 你觉得它真的需要, |
| 1235 | Do you think it's really needed for AI tools | 你觉得真的需要人工智能工具吗? |
| 1236 | for non-AI background journalists to read AI papers? | 非AI背景记者阅读AI论文?. |
| 1237 | Yeah, I think that's a great question. | 是啊,我觉得这是一个很好的问题。 |
| 1238 | So, I mean, in some sense, it depends on the sort of | 所以,我的意思是,从某种意义上说, 这取决于什么类型的 |
| 1239 | journalism that you are doing. | 你做的新闻工作 |
| 1240 | So, for example, for science journalists who are reporting | 所以,比如说,对于正在报道的科学记者来说 |
| 1241 | specifically on the contents of a paper and comparing it to the previous state of the art, | 特别是一篇论文的内容,并将其与以前的艺术水平进行比较, |
| 1242 | I think that is a very important skill to have. | 我认为这是一种非常重要的技能。 |
| 1243 | But for other types of journalists, for example, | 但对其他类型的记者来说, |
| 1244 | people looking at the impact of AI on society, I'm not sure if that's the most high-value way. | 人们看AI对社会的影响,我不确定这是否是最有价值的方式. |
| 1245 | So, I'll give you an example. | 所以,我给你一个例子。 |
| 1246 | One of the journalists whose work I really like in the US | 我很喜欢在美国工作的记者之一 |
| 1247 | is Timothy B. Lee, and a lot of his work is related to self-driving cars. | 是Timothy B. Lee 他的很多作品都和自驾车有关 |
| 1248 | So, he reports | 所以,他汇报 |
| 1249 | on the state of progress on self-driving cars and so on. | 关于自驾车等进展状况. |
| 1250 | But a lot of the insights that he has | 但他有很多见解 |
| 1251 | are not by reading papers on the state of the technology, but rather by looking at the | 并不是通过阅读有关技术状况的论文,而是通过观察 |
| 1252 | statistics from companies like Waymo and Cruise, by looking at the statistics released by government | 从Waymo和Cruise等公司获得的统计数据, |
| 1253 | organizations, by comparing the regulatory stances on self-driving cars across the states in the US. | 通过比较美国各州自驾汽车的监管立场 |
| 1254 | And that does not necessarily require deep academic expertise, but it does require being able to find | 这并不一定需要深层次的学术专业知识,但确实需要找到 |
| 1255 | the right sources and being able to read government documents or being able to track down | 正确来源和能够阅读政府文件或能够追查 |
| 1256 | companies' reports when they are releasing them. | 公司的报告,当他们被释放。 |
| 1257 | And I think we're at the point, or at least slowly | 我想我们到了点,或者至少慢一点 |
| 1258 | beginning to get to the point, where AI companies are also starting to figure out what they communicate | AI公司也开始找出他们沟通的内容 |
| 1259 | to the public. | 告大众曰. |
| 1260 | And many of these sort of nuggets of information are hidden within this large | 许多这类信息 隐藏在这个巨大的 |
| 1261 | reports that companies release. | 报告公司释放。 |
| 1262 | So, I was part of this initiative called the Foundation Model | 所以,我是这个倡议的一部分 叫做基础模型 |
| 1263 | Transparency Index. | 透明度指数。 |
| 1264 | And companies were required to or asked to report a hundred different things | 公司被要求或被要求 报告百种不同的东西 |
| 1265 | about how they train their AI models. | 他们是如何训练他们的AI模型。 |
| 1266 | And I think there were some very interesting insights | 我觉得有一些非常有趣的见解 |
| 1267 | in a lot of what the companies reported, which were not related to academic insights really, | 在很多公司的报道中 与学术见解无关 |
| 1268 | but which were more about where do they get the training data from, or how much do they pay their | 但他们从哪里得到培训数据 或支付多少钱 |
| 1269 | workers when they're annotating data, and so on. | 工人们在分析数据时, |
| 1270 | And so, if people are looking at sort of the impact | 所以,如果人们在看 那种影响 |
| 1271 | of AI on society, then I think this type of secondary information can be quite useful and | 我认为这种次要信息可以很有用 |
| 1272 | important. | 这很重要 |
| 1273 | That'd be great to get more access to the data and the facts to get better than standing | 最好能让更多的人获得数据 和事实 得到比站立更好的 |
| 1274 | on the AI. | 在人工智能上。 |
| 1275 | Okay, let's talk further about the researchers like you guys. | 好吧,让我们进一步谈论 研究人员喜欢你们。 |
| 1276 | Researchers sometimes | 研究人员有时 |
| 1277 | may still spread some kind of the AI hype to the public. | 可能还会向公众传播某种AI的谣言. |
| 1278 | So, and now in some kind of the, | 所以,现在在某种情况下, |
| 1279 | could you please to, yeah, I will list you some kind of the AI researcher figures like | 麻烦你,是的,我会列出一些 AI研究者的数字像 |
| 1280 | Jeffrey Hinton, and you just mentioned, right, or Yellow Koon, a guy I pretty, yeah, I like him a | Jeffrey Hinton, 你刚才提到,是的,还是黄坤,一个我漂亮,是的,我喜欢他 |
| 1281 | lot, I think. | 我觉得是很多 |
| 1282 | And probably Fei-Fei Li, I think. | 还有李飞飞吧 |
| 1283 | And I think, yeah, would you please to give some | 我想,是的,请你给一些 |
| 1284 | kind of the comments on their role to be a, or would you please to give the comments on | 或请各位发表评论。 |
| 1285 | these figures and their roles to the public to tell about what, are they acting in the right way | 这些人物和他们的角色 告诉公众什么, 他们的行为是正确的方式 |
| 1286 | to tell the people the right thing about AI? | 告诉人们AI的正确之处? |
| 1287 | How do you think about it? | 你觉得怎么样? |
| 1288 | Would you like to give some | 你想给一些 |
| 1289 | kind of the comments about these kind of the peers in the AI researcher field? | 对AI研究者领域这类同行的评论是什么? |
| 1290 | Thank you. | 谢谢 |
| 1291 | I'm happy to share some comments on that. | 我很乐意就此发表一些看法。 |
| 1292 | I think there are a couple of reasons why | 我想有几个原因 |
| 1293 | AI researchers are often hyping AI too much, and I'll share some thoughts on how they can do better. | AI研究者们经常对AI叹气过多,我将分享一些关于他们如何能做得更好的想法. |
| 1294 | One is, if you look at the reason why researchers are getting into AI. | 第一,如果你看看 研究人员进入AI的原因。 |
| 1295 | I mean, let me tell you my | 我的意思是,让我告诉你我的 |
| 1296 | own story. | 自己编的故事。 |
| 1297 | 25 years ago, when I decided that my undergrad major would be in computer science, | 25年前,当我决定 我的研究生将进入计算机科学, |
| 1298 | it's because one day I wanted to build AGI. | 因为有一天我想建造AGI |
| 1299 | I wanted to help build it. | 我想帮助建造它。 |
| 1300 | I really liked your rule, | 我真的很喜欢你的规矩 |
| 1301 | by the way, and your conference of not talking about AGI. | 顺便说一句,还有你的会议 不谈论AGI。 |
| 1302 | I think we need more of that at more | 我认为我们需要更多的 更多 |
| 1303 | events. | 事件。 |
| 1304 | But nonetheless, the fact remains that many AI researchers are thinking about this. | 但是,事实上,许多AI研究者仍在思考这个问题。 |
| 1305 | It's | 这是 |
| 1306 | this kind of North Star, and obviously many researchers believe that it can completely | 很明显,许多研究者认为它可以完全 |
| 1307 | transform the world if you had AI that could do any job that any person could do. | 如果你有人工智能 做任何人都能做的任何工作 就能改变世界 |
| 1308 | And so a lot of | 这么多 |
| 1309 | people are coming into the AI field because they have a kind of religious belief that this world | 人们进入AI领域 因为他们有某种宗教信仰 这个世界 |
| 1310 | changing thing is achievable. | 改变事物是可以实现的。 |
| 1311 | And so you're starting from a baseline of people who believe | 所以,你开始 从一个基线的人相信 |
| 1312 | in something radical. | 在激进的东西。 |
| 1313 | I'm not talking about whether that belief is true or false, | 我不是说这种信念是真还是假 |
| 1314 | but you can imagine if someone is really committed to this mission, if you will, | 但你可以想象,如果有人 真的致力于这个任务, 如果你愿意, |
| 1315 | then they are going to believe some things that might sound really radical or crazy. | 然后他们会相信一些事情 听起来非常激进或疯狂。 |
| 1316 | So I think that's part of the reason. | 所以我认为这就是原因之一。 |
| 1317 | And a second more mundane reason is that everybody needs to get | 还有一个更普通的理由是 每个人都需要得到 |
| 1318 | funding. | 供资。 |
| 1319 | Yann LeCun和Jeff Hinton, 对他们来说, 他们的布局相当不错. | |
| 1320 | They don't need to | 他们不需要 |
| 1321 | hype AI to get funding. | 高调AI来获得资金。 |
| 1322 | But for a lot of other researchers, unfortunately, if you | 但对于很多其他研究者来说 不幸的是 如果你 |
| 1323 | put out a press release touting how amazing your new invention is, you get more attention, | 发布一份新闻稿 说出你的新发明多么惊人 你得到更多的关注, |
| 1324 | potentially more funding and so forth. | 可能有更多的资金等等。 |
| 1325 | And even in our research, we are often thinking carefully | 即使在我们的研究中 我们经常仔细思考 |
| 1326 | about our paper titles. | 关于我们的论文标题。 |
| 1327 | Are we exaggerating those too much? | 我们是不是夸大了这些? |
| 1328 | So I think that's a struggle | 所以我觉得那是一场斗争 |
| 1329 | for every single AI researcher. | 每一个AI研究员。 |
| 1330 | And a third thing, I think something new... | 第三件事,我觉得有些新... |
| 1331 | Sorry, go ahead. | 对不起,请便。 |
| 1332 | Can I | 可以吗? |
| 1333 | share one last thought on that? | 和大家分享最后的想法吗? |
| 1334 | In terms of how people can do better. | 论人如何能更善. |
| 1335 | So one thing that's changed, | 所以有一件事已经改变了, |
| 1336 | AI used to be a really niche topic, and then it was okay for AI researchers to say some crazy | AI曾经是一个非常合适的话题, 然后AI研究人员可以说一些疯狂 |
| 1337 | things because the people who were paying attention were mostly other AI researchers. | 因为关注的人大多是其他AI研究者 |
| 1338 | But the extent to which the public and the media are now paying attention to what AI researchers | 但公众和媒体关注AI研究人员的程度 |
| 1339 | are saying is, of course, on a completely different scale. | 意思是,当然, 是一个完全不同的规模。 |
| 1340 | Now AI researchers have become | 现在AI研究者变成了 |
| 1341 | public figures. | 公众人物. |
| 1342 | And many researchers have not yet understood this fact that for public researchers, | 而许多研究者尚未意识到这一事实,对于公共研究者来说, |
| 1343 | there is a higher ethical standard for any kind of public figure that we should hold them to. | 任何类型的公众人物都应该遵守更高的道德标准。 |
| 1344 | And | 还有 |
| 1345 | what they say matters, and they can't just say whatever is on their mind all the time. | 他们说的话很重要 他们不能老是说他们心里想的 |
| 1346 | And I | 还有我 |
| 1347 | think there needs to be a culture change. | 认为需要改变文化 |
| 1348 | Oh, that'd be great. | 哦,这将是伟大的。 |
| 1349 | So is that to say that researchers, | 这么说来 研究人员 |
| 1350 | I mean, the entrepreneur, researchers being entrepreneur will be harmful for the original AI | 我的意思是,创业者, 研究人员是创业者 将对原AI有害 |
| 1351 | research, because then this will give them the added motivation to make the AI hype. | 研究,因为这样他们就会得到更多的动力 来制造AI的热闹。 |
| 1352 | Is that right? | 是吗? |
| 1353 | I mean, AI researchers, yeah, be entrepreneurs. | 我的意思是,AI研究人员,是的,做企业家。 |
| 1354 | Sanaj, do you want to? | 萨纳伊,你想吗? |
| 1355 | Yeah, sure. | 当然 |
| 1356 | I mean, I don't | 我的意思是,我没有 |
| 1357 | think there's an issue with researchers becoming entrepreneurs per se. | 认为研究者成为企业家本身有问题. |
| 1358 | I do think like a big | 我确实像一个大 |
| 1359 | challenge with the AI tools that we have today is that we don't really have enough real world users | 对AI工具的质疑,我们今天有的是 我们真的没有足够的真实世界用户。 |
| 1360 | that are being sort of productionized. | 正在生产 |
| 1361 | And so maybe that is something we need more of. | 也许我们需要更多 |
| 1362 | But I think the specific challenge with AI hype in entrepreneurs is, I think researchers are seen | 但我认为在创业者中 AI hype的具体挑战 是,我认为研究者被看到 |
| 1363 | as this trusted public figure who talks sort of dispassionately about scientific research and | 作为这个值得信赖的公众人物, 他有点冷静地谈论科学研究和 |
| 1364 | communicates that to the public. | 向大众传达。 |
| 1365 | And when they turn into entrepreneurs, then they become people | 当他们变成企业家, 然后他们成为人 |
| 1366 | who have something to sell. | 他们有东西卖。 |
| 1367 | And I think this sort of transition, even though there is no clean line, | 我认为这种转变 即使没有干净的线条 |
| 1368 | this sort of transition needs to be appreciated by like journalists writing articles about people | 这种转变需要像记者那样 写关于人的文章 |
| 1369 | who now have like a business incentive to sell things and not just treat them as neutral, | 他们现在有 商业激励 出售的东西, 不只是把它们视为中立, |
| 1370 | independent parties. | 独立党派。 |
| 1371 | And I think that is what needs to be emphasized a lot more. | 我认为,必须更加强调这一点。 |
| 1372 | So when an | 所以当一个 |
| 1373 | entrepreneur or like someone who is trying to sort of set a company of the ground says something | 企业家或喜欢有人 谁试图设置一个公司 地面说 |
| 1374 | about that company, that should be seen as a statement by someone who has a financial interest | 关于那家公司,那应该被看作是有经济利益的人的声明 |
| 1375 | in that company, rather than a statement by a third party expert who has a dispassionate view | 而不是由持有冷静观点的第三方专家的陈述 |
| 1376 | to some extent on the workings of that company. | 在某种程度上,关于公司的运作。 |
| 1377 | And I think that is what we are sort of trying to | 我想这就是我们试图 |
| 1378 | wrap our head around right now, simply because we have seen this huge boom in AI in the last few | 把我们的头围起来,仅仅因为我们看到 在AI的这个巨大的繁荣 在最后几个 |
| 1379 | years. | 岁月 |
| 1380 | This wasn't a problem maybe 20 years ago when AI was useful for a few things other than, | 也许20年前 AI对一些事情有用 |
| 1381 | I don't know, like this type of AI was useful for maybe handwriting recognition or whatever. | 我不知道,像这种AI 有用 也许笔迹识别什么的。 |
| 1382 | But all of a sudden it has become a consumer technology. | 但突然间它变成了一种消费技术. |
| 1383 | And so we're seeing a huge boom | 所以我们看到一个巨大的繁荣 |
| 1384 | of researchers turning into entrepreneurs. | 研究者变成企业家。 |
| 1385 | And I think we need to treat their claims with a | 我认为我们需要用 |
| 1386 | little bit more skepticism, especially when it's about their products or their industry. | 更有一点怀疑,特别是当它涉及到他们的产品或工业时. |
| 1387 | Hey, that'd be great. | 嘿,这将是伟大的。 |
| 1388 | Yeah, sure. | 当然 |
| 1389 | Still talk a little bit more about the AI companies | 还要多谈谈AI公司 |
| 1390 | who have the strongest motivation to do AI hype. | 谁有最强的动机做AI的hype。 |
| 1391 | In a way, you may rather say | 在某种程度上,你可能宁愿说 |
| 1392 | this kind of the advertisements displayed on YouTube about AI companies or AI products, | 在YouTube上刊登的关于AI公司或AI产品的广告, |
| 1393 | or even their websites. | 甚至是他们的网站 |
| 1394 | I mean, the AI companies' websites are all AI hype, I think. | 我的意思是,AI公司的网站 都是AI的hype,我认为。 |
| 1395 | In a way, | 在某种程度上, |
| 1396 | we can say that. | 我们可以这么说 |
| 1397 | So how, I mean, just how the public can get a better understanding | 所以,我的意思是, 如何让公众得到更好的理解 |
| 1398 | from those kind of the advertisement materials or those kind of the advertisement videos. | 从那些广告材料 或那种广告视频。 |
| 1399 | And we always see this kind of the Grammar, the Grammar Lake or the Chachibee TV, even now Cloud, | 我们总是看到这种语法, 语法湖或Chachibee电视, 即使现在云, |
| 1400 | 我的意思是,云 动力由Anthropic。 | |
| 1401 | And they just display their advertisement in San | 他们只是展示他们的广告 在桑 |
| 1402 | Francisco airport on the YouTube and everywhere. | 弗朗西斯科机场在YouTube和各地。 |
| 1403 | And how people just get the right understanding | 和人们如何得到正确的理解 |
| 1404 | from those kind of the advertisement materials and to have a more balanced and objective | 从这类广告材料 并有一个更平衡和客观 |
| 1405 | understanding about AI by reading or by hearing or by watching these kind of things. | 通过阅读、听觉或观看这些东西来理解AI。 |
| 1406 | Or anyone or anybody or any organization can play a more moderate role to change this kind | 或任何人 或任何组织可以发挥更温和的作用 改变这种 |
| 1407 | of the situation, I think. | 我认为,情况。 |
| 1408 | Advertisement about AI now is everywhere, it's everywhere, especially | 关于AI的广告现在到处都是,它到处都是,特别是 |
| 1409 | Silicon Valley and Bay Area. | 硅谷与湾区. |
| 1410 | I'll say that there's one really big good thing about generative AI | 我会说,有一个真正的大好事 关于基因AI |
| 1411 | compared to, let's say, predictive AI. | 比起预测性AI |
| 1412 | If a company comes out and claims, oh, our AI can predict who | 如果一个公司出来并声称, 哦,我们的AI可以预测谁 |
| 1413 | is going to commit a crime, and that's their advertisement, there's nothing you can do to | 这是他们的广告,你无能为力 |
| 1414 | check that claim for yourself. | 你自己去查查那份索赔 |
| 1415 | Generative AI is very different. | 遗传性AI非常不同. |
| 1416 | 如果Chachibee或云声称 | |
| 1417 | I don't know, whatever it is, legal work, 如果你是律师或一些法律工作 | |
| 1418 | expertise, that's a claim you can check for yourself. | 专家,这是一个权利主张 你可以检查自己。 |
| 1419 | And that's a much more efficient way | 这是更有效率的方法 |
| 1420 | of doing it than trying to figure out if their advertisement is hyped or not or reading journalistic | 而不是试图找出他们的广告 被赞美与否 或阅读新闻 |
| 1421 | articles or even listening to this talk or reading our book. | 写文章,甚至听这些话 读我们的书 |
| 1422 | There's something much more | 还有更多 |
| 1423 | high value you can do with your time, which is just to play with generative AI products yourself, | 高价值你可以做你的时间, 这只是玩 基因AI产品自己, |
| 1424 | just within a few hours of using it, you're going to get a pretty good understanding of its | 在使用后几小时内 你就会非常了解它 |
| 1425 | potential as well as its limitations for the specific use cases that you care about. | 以及它对于你所关心的具体用途的局限性。 |
| 1426 | And I'm sure many of you are already doing that. | 我敢肯定,你们很多人已经这样做了。 |
| 1427 | I very much encourage you to continue doing that. | 我非常鼓励你继续这样做。 |
| 1428 | 记住,AI不仅仅是Chachibee或云。 | |
| 1429 | It doesn't have to be specifically a | 不一定是特别的 |
| 1430 | bot that you go and type into. | 机器人,你进入。 |
| 1431 | AI is integrated into a number of other software products that we | AI被整合到许多其他软件产品中,我们 |
| 1432 | use on an everyday basis. | 每天使用。 |
| 1433 | So when you're interacting with AI, be mindful of that. | 所以当你与AI互动时,要注意这一点. |
| 1434 | Use that as a way to | 用它来作为方法 |
| 1435 | develop your intuition on what it's doing well, what it's not doing well. | 发展你的直觉 关于它做什么好, 做什么不好。 |
| 1436 | When you're on social | 当你在社交上 |
| 1437 | media, remember that so much of content out there is AI generated. | 媒体,请记住, 这么多的内容 有AI生成。 |
| 1438 | So use that as a way to update | 所以用它来更新 |
| 1439 | your understanding of the kind of realistic images that are possible to create with social media. | 你对社交媒体所能创造的现实形象的理解。 |
| 1440 | So yeah, basically, my advice is through the course of our everyday interaction, both with AI | 所以,基本上,我的建议是通过 我们日常的互动, 无论是与AI |
| 1441 | products and other kinds of software products, we should be constantly reflecting on that and | 产品和其他类型的软件产品,我们应当不断对此进行反思。 |
| 1442 | updating our intuition for what AI can and cannot do, generative AI specifically, at any given point | 更新我们的直觉,说明AI能做什么和不能做什么, 具体地说,在任何特定的时间点, |
该视频共有字幕 1443 条。解锁更多字幕为会员功能，请移动到 价格