One Year In, ChatGPT’s Legacy Is Clear

The technology is less important than the ideas it represents.
技术不如它所代表的理念重要。

ChatGPT is one year old today, and it’s accomplished a lot in its first trip around the sun. The chatbot has upended or outright killed high-school and college essay writing and thoroughly scrambled the brains of academics, creating an on-campus arms race that professors have already lost. It has been used to write books, article summaries, and political content, and it has flooded online marketplaces with computer-generated slop.
ChatGPT 今天一岁了,在它的第一次环绕太阳之旅中取得了很多成就。这个聊天机器人彻底颠覆了高中和大学的论文写作,甚至直接导致其消亡,并且彻底搅乱了学术界人士的思维,造成了校园内的军备竞赛,教授们已经输掉了这场比赛。它被用来写书籍、文章摘要和政治内容,并且已经用计算机生成的废话充斥了在线市场。

As we’ve gotten to know ChatGPT, we’ve noticed how malleable it is. The li’l bot loves clichés. Its underlying technology has been integrated into internet search. ChatGPT is a time waster—a toy—but also, potentially, a labor-force destroyer and a way for machines to leech the remaining humanity out of our jobs. It may even be the harbinger of an unrecognizable world and a “textpocalypse” to come.
随着我们逐渐了解ChatGPT,我们注意到它是多么的可塑。这个小机器人喜欢陈词滥调。它的底层技术已经被整合到互联网搜索中。ChatGPT是一个消磨时间的玩具——但同时,也可能是一个劳动力破坏者,以及一种让机器从我们的工作中吸走剩余人性的方式。它甚至可能是一个我们无法识别的世界和即将到来的“文本末日”的先兆。

Even for a large language model with billions of parameters, trained off of perhaps terabytes of potentially questionable and opaquely scraped data ... that’s quite a year.
即使是一个拥有数十亿参数的大型语言模型,它也可能是通过数以太字节计的、潜在质量存疑且来源不透明的数据训练而成......这确实是非同寻常的一年。

As evidenced above, my colleagues and I have spent a great deal of time and many words trying to understand exactly what ChatGPT means to the world. It’s been a long process, in part because its launch took everyone by surprise. As Karen Hao and I reported earlier this month, OpenAI didn’t expect ChatGPT to amount to much more than a passing curiosity among AI obsessives on Twitter: In a company pool wagering how many people might use ChatGPT during its first week, the highest guess was 100,000 users. (The tool hit 1 million within the first five days.) The product was intended to be the software equivalent of a concept car. Instead, it became one of the most popular applications in the history of the internet.
正如上文所示,我和我的同事们花费了大量时间和文字,试图准确理解ChatGPT对世界意味着什么。这是一个漫长的过程,部分原因是它的推出让所有人都感到意外。正如我和Karen Hao在本月早些时候报道的,OpenAI并没有期望ChatGPT会成为Twitter上AI迷们的一时新奇:在一次公司内部的赌注中,关于第一周可能会有多少人使用ChatGPT,最高的猜测是10万用户。(这个工具在前五天内就达到了100万用户。)这个产品本意是作为软件版的概念车。然而,它却成为了互联网历史上最受欢迎的应用之一。

The reason for its success is obvious. ChatGPT is not sentient or intelligent in any way—I still love the early description that it’s just “spicy autocomplete”—but it frequently offers a decent simulation of talking with a vaguely boring person or, say, a polite customer-service representative. Sure, ChatGPT can “hallucinate,” delivering misinformation as though it’s fact, but it can also write ballads, pass an M.B.A. exam, and debug code. Enough of these little interactions felt to a lot of people like magic, or at very least, the beginning of a bona fide technological breakthrough.
它成功的原因显而易见。ChatGPT 既不具有感知能力也不具有任何形式的智能——我仍然喜欢它早期的描述,就是“辣味自动补全”——但它经常能够提供一种与一个略显无聊的人交谈,或者说,一个礼貌的客服代表交谈的相当不错的模拟。当然,ChatGPT 有时会“幻想”,把错误信息当作事实来传递,但它也能写歌谣,通过M.B.A.考试,以及调试代码。对很多人来说,这些小互动中有足够多的感觉像是魔法,或者至少是一个真正的技术突破的开始。

Evangelists have told me that they employ ChatGPT like they would an enthusiastic intern, or a copy editor, or a debate partner. Others have said they use it to automate the drudgery of small tasks such as writing emails, and those who program or work with databases seem to find ChatGPT and its extensions akin to having an extra arm to work with, maybe even an extra brain. A ChatGPT screen is open on their computer at all times.
传道者告诉我,他们像使用一个热情的实习生、一个文案编辑或一个辩论伙伴那样使用ChatGPT。其他人说他们用它来自动化写电子邮件等小任务的苦差事,那些编程或与数据库打交道的人似乎发现ChatGPT及其扩展就像拥有一只额外的手臂来工作,甚至可能是一个额外的大脑。他们的电脑上始终开着一个ChatGPT的屏幕。

And then there are those who make a bolder case: that ChatGPT is “a thinking companion,” a way to summarize large texts into digestible nuggets, to brainstorm and generate ideas, to help build and execute business plans, and, most important, to get a machine with computational power to do tasks quickly that might take a human countless hours. It’s not just that ChatGPT can solve a problem. As the programmer James Somers wrote in The New Yorker recently, it’s that “from a deep well of knowledge, it could suggest ways of approaching a problem” altogether. It offers a way to unlock a new mode of thinking.
然后还有一些人提出了更大胆的观点:ChatGPT是“一个思考伴侣”,一种将大量文本总结成易于消化的小块的方式,用来头脑风暴和生成想法,帮助构建和执行商业计划,最重要的是,利用机器的计算能力快速完成可能需要人类无数小时才能完成的任务。问题不仅仅在于ChatGPT能解决问题。正如程序员詹姆斯·索默斯最近在《纽约客》上写的那样,关键是“从深厚的知识井中,它能够提出全新的解决问题的方法”。它提供了一种解锁新思维模式的方式。

When I hear these descriptions, I feel a sense of panic, as if a technical revolution has passed me by. I find ChatGPT too untrustworthy for research tasks. (I don’t particularly need a research assistant who may, out of nowhere, imagine misinformation and present it as fact.) I’ve spent time refining prompts and even building my own bot to try and edit my own writing, and I’ve found the output lacking in almost every way when it comes to replicating or even streamlining the job I get paid to do. My brain, I’ve come to realize, is bad at constructing prompts, a skill that seems to have more in common with programming than it does prose writing. The experience can feel akin to being present for the invention of the iPod but hating music.
当我听到这些描述时,我感到一种恐慌,好像一场技术革命已经悄然从我身边溜走。我发现ChatGPT对于研究任务来说太不可靠了。(我并不特别需要一个可能凭空想象错误信息并将其当作事实呈现的研究助理。)我花时间精炼提示词,甚至构建了我自己的机器人来尝试编辑我的写作,但我发现在复制甚至简化我所从事的工作方面,其输出在几乎所有方面都不尽人意。我逐渐意识到,我的大脑不擅长构建提示词,这项技能似乎与编程有更多的共同点,而不是与写作散文。这种经历感觉就像亲临iPod的发明之际,却讨厌音乐。

A good ChatGPT whisperer understands how to sequence commands in order to get a machine to do its bidding. That’s a genuine skill, but one that eludes me as well as some other humanities types I know. The best ChatGPT prompters I know tend to be good systems thinkers or at least well-organized people—the kind who might create a series of automated protocols and smart-home integrations to turn their lights on and off. I’m the guy who sees romance in wandering around in the dark, bumping into a coffee table, to find the switch.
一个优秀的ChatGPT驯服师懂得如何排列命令,以便让机器听从他们的指令。这是一项真正的技能,但它却让我以及我所认识的一些其他人文学科的人望尘莫及。我所知道的最佳ChatGPT提示者往往是优秀的系统思考者,或者至少是组织有序的人——那种可能会创建一系列自动化协议和智能家居集成来开关灯光的人。而我是那种在黑暗中徘徊,撞到咖啡桌,去寻找开关的浪漫主义者。

This is all to say that I both recognize and respect ChatGPT’s impact over the past year while also feeling a bit gaslit by it. Academically, I understand what’s happening, how the tool may unlock productivity and creativity that leaves me behind. It feels perfectly reasonable to me that chatbots will automate tasks and jobs and change the way a lot of companies handle workflows behind the scenes. I believe that it will continue to flood the internet with text of varying quality and, in many cases, be used by greedy, amoral people to generate vapid, large language lorem ipsum to turn a quick profit at the expense of whatever humanity exists on the web.
这就是说,我既认识到也尊重ChatGPT在过去一年的影响,同时也感到有些被它搞得心神不宁。从学术上来说,我理解正在发生的事情,理解这个工具可能会释放生产力和创造力,让我落后。对我来说,聊天机器人将自动化任务和工作,改变许多公司在幕后处理工作流程的方式,这似乎完全合理。我相信它将继续向互联网灌输不同质量的文本,并且在许多情况下,会被贪婪、没有道德的人用来生成空洞的、大型语言的“乱数假文”,以牺牲网络上存在的任何人性为代价,快速牟利。

The legacy of ChatGPT may not have much to do with its utility at all. ChatGPT, the tool, is likely less important than ChatGPT, the cultural object. ChatGPT was actually what OpenAI intended to create all along: proof of concept for the bigger idea of a breakthrough in generative artificial intelligence. Even if you can’t get the bot to spit out Faulkner, ChatGPT still feels like a paradigm shift—a glimpse at a technology that had been teased in movies and popular culture for decades but never really seemed to arrive in a way that was functional for the general public. Now it’s here: proof that the generative-AI era has arrived, even if the conversations we’ve all had about the technology seem more immediately consequential than the product itself.
ChatGPT的遗产可能与其实用性根本无关。ChatGPT这个工具可能不如ChatGPT这个文化对象重要。ChatGPT实际上是OpenAI一直以来想要创造的东西:对于生成性人工智能重大突破这一更大理念的概念验证。即使你让机器人吐不出福克纳的作品,ChatGPT仍然感觉像是一次范式转移——对于一个在电影和流行文化中被暗示了几十年但从未真正以对大众有用的方式到来的技术的一瞥。现在它来了:证明生成性AI时代已经到来,即使我们所有人关于这项技术的讨论看起来比产品本身更具有直接的后果。

Whatever its limitations, ChatGPT is still most valuable as a symbol and a placeholder—a stepping stone to an age when technologists might someday replicate human intelligence. ChatGPT gave true believers and hype cyclists a prototype to gesture toward. This year, we got excited about the concept car, even if, deep down, we know that most models never see the open road.
尽管有其局限性,ChatGPT 仍然作为一个象征和占位符最有价值——它是迈向技术专家有朝一日可能复制人类智能时代的一块垫脚石。ChatGPT 为真正的信徒和炒作者提供了一个可以指向的原型。今年,我们对概念车感到兴奋,即使在内心深处,我们知道大多数车型永远不会上公路。