GPT4横空出世，互联网霸主Google跌落神坛 | 纽约客

1

Bing A.I. and the Dawn of the Post-Search Internet

必应AI与后搜索时代之前奏

英文部分选自纽约客20230321

Bing A.I. and the Dawn of the Post-Search Internet

必应AI与后搜索时代之前奏

So much of the current Web was designed around aggregation. What value will legacy sites have when bots can do the aggregation for us?

当今的互联网是围绕聚合思想进行设计的。若机器人能为我们聚合信息，传统网站还有什么价值？

Around a year ago, I wrote a column about users’ growing frustration with Google Search, as automated summaries, sponsored content, and S.E.O.-tailored spam increasingly crowded out the kinds of useful Web site results that Googling was supposed to produce. Google’s search algorithm wasn’t directing us to the information that we wanted to find (for instance, in my case at the time, the elusive perfect toaster) so much as bombarding us with the half-baked recommendations of content mills. Yet Google Search has maintained its dominance partly out of habit and partly because no competing service has offered a viable alternative—until now. On February 7th, Microsoft began a beta launch of a version of its search engine, Bing, in the form of an A.I. chatbot, powered by GPT-4, the latest iteration of OpenAI’s large language model ChatGPT. Instead of directing users to external sites, the new Bing can simply generate its own answers to any query. Google considers the tool to be an existential threat to its core business, for good reason. At the end of last year, the Times reported that the company had declared a “code red.” Microsoft’s vice-president of design Liz Danzico, who helped to develop Bing A.I.’s interface, told me recently, “We’re in a post-search experience.”

大约一年前，我写了一篇专栏文章，主题是用户对谷歌搜索越来越失望。究其原因是自动汇总、广告内容和针对搜索引擎优化的垃圾信息越来越多，挤占了谷歌搜索本应呈现的实用内容。谷歌的搜索算法并没有引导我们找到所需信息（比如怎么也找不到的理想烤面包机），而是给我们推送了海量的由内容生产者推荐的物品或信息。然而谷歌搜索仍占据统治地位，一方面是因为用户习惯问题，另一方面是因为缺乏能与之匹敌的竞品——然而，大人，时代变了：2月7日，微软发布了测试版新版必应（Bing）搜索引擎，搭载了GPT-4内核的人工智能聊天机器人（GPT-4是OpenAI的大型语言模型GPT的最新迭代）。新的必应搜索引擎无需将用户引导至外部站点，而是为任何搜索请求直接生成自己的答案。谷歌认为这一工具对其核心业务构成了生存威胁，这不无道理。去年年底，《泰晤士报》（Times）报道称谷歌拉响了红色警报。近日，参与开发必应AI界面的微软设计副总裁利兹·丹姿克（Liz Danzico）告诉我：“我们正迈入后搜索时代。”

注释：

code red: The highest possible state of alert to address or prepare for some imminent danger, threat, trouble, or difficulty. e.g. The Pentagon issued a nationwide code red following warnings of a foreign invasion.

来源：https://idioms.thefreedictionary.com/Code+red!

Bing A.I., which I’ve had the chance to test-drive in recent weeks, is essentially ChatGPT hooked up to Microsoft’s search directory. Using it is like talking to a very powerful librarian whose purview is the breadth of the Internet. The old search experience is almost second nature to today’s Internet users: enter relevant keywords into Google, press Enter, and sift through the list of links that show up on the results page. Then click through to find the info you’re looking for. If you don’t find it, perhaps return to the Google Search page and tinker with your keywords to try again. With Bing A.I., Web sites are the source material, not the destination, and results are produced through what Danzico called a “co-creation process” between user and bot. Bing A.I. recaps the recappers and aggregates the aggregators, sifting through the barrage of online information for you. I asked it which toaster Wirecutter recommended, and it told me the Cuisinart CPT-122 2-Slice Compact Plastic Toaster; then I asked it to draw up a list of other recommendations, and it gathered ones from outlets including The Kitchn, Forbes, and The Spruce Eats. Within a few seconds, and without ever leaving the Bing A.I. page, I had a digest of reputable devices. The chatbot wouldn’t tell me precisely which one to buy, however. “I can’t make decisions for you because I’m not a human,” it said.

过去几周，我有幸试用了必应AI。本质上，它是将ChatGPT连接到微软的搜索目录中：使用必应AI时，好似在与一位博览全网资源的图书馆管理员交谈。如今的互联网用户早已对传统搜索方式烂熟于心，在谷歌中输入相关的关键词，按回车键，细看结果页面上列出的各种链接。接着，点击链接查找所需的信息，若一无所获，则返回到谷歌搜索页面，调整关键词后再次尝试。然而，必应AI是将网站作为信息源，而非目的地，并通过丹姿克所称的用户与机器人的“协同创作”来返回搜索结果。必应AI是在汇总的基础上进一步汇总、归纳的基础上进一步归纳，替用户筛选海量网上信息。当我问必应AI，评论网站Wirecutter最推荐哪一款烤面包机时，它会告诉我是美膳雅（Cuisinart）CPT-122两片式小型塑料烤面包机。我问必应AI有无其他烤面包机推荐时，它又参考了The Kitchn网站、《福布斯》（Forbes）杂志、The Spruce Eats网站的推荐款。几秒钟内，我就得到了一份烤面包机的推荐列表，这个过程中我还不用离开必应AI界面。然而，AI机器人不会告诉我具体买哪一款，它回复称：“因为我不是人类，所以无法替您做决定”。

In some ways, a Bing A.I. user has more agency than a user of Google Search. Communicating with the chatbot involves what Sarah Mody, a senior product marketing manager at Microsoft, described as “learning not to speak the search of twenty years ago.” You communicate not in isolated keywords but in straightforward sentences, and you can narrow or refine your query results by asking the machine follow-ups. If you request an itinerary for a trip to Iceland, for instance, and then ask, “What time does the sun set there?” the bot will understand which “there” you’re referring to. But, in other ways, the Bing user is more limited and passive, encouraged to let the machine decide which information is worthwhile rather than doing any searching on their own. “It’s your travel guide, your bank, your confidante, your guide,” Danzico said. The “conversation mode” interface—a single chat box on top of a gentle color gradient—reflects that one-stop-shop objective. Where using Google Search sometimes feels like engineering the right equation to solve a problem, using Bing A.I. is a bit like a series of text-message conversations. It even punctuates answers with a smiling, blushing emoji: “I’m always happy to chat with you. ���,” it told me. To the left of the chat box, there’s a button that reads “new topic” and shows a broom sweeping away dust. Clicking it erases the current chat and starts over. The module, Danzico told me, was developed with the help of the A.I. itself.

在某种程度上，比起谷歌搜索，必应AI的用户拥有更大的自主权。正如微软资深产品营销经理萨拉·穆迪（Sarah Mody）所言，（使用新版必应时）“不要还抱着二十年前那一套不放”。和必应机器人聊天时不要只蹦出单个的关键词，而要直接说出完整的句子。同时，你还可以通过追问聊天机器人来获得更精准、更完善的回答。举个例子，如果你让必应AI帮你规划冰岛游行程，紧接着问“那个地方几点日落？”这种情况下，必应AI就会理解你所说的“那个地方”就是冰岛。然而，反过来说，必应AI的用户也会更加受限、更加被动，因为他们会更倾向于让必应AI决定哪些信息是有价值的，而不是自己亲自去搜一搜这些信息。丹姿克称：“必应AI将成为你的导游、你的银行、你的知己，甚至是你的人生导师。”选择位于一片颜色渐变块上方的“对话模式”界面，你就能享受这种一站式服务。使用谷歌搜索有时就好比设计一个正确的方程式来解决一个问题，但使用必应AI就有点像一系列短信对话：它甚至还会在回答里添加一个笑脸表情。比如它会和我说：“和你聊天很开心☺”。对话框左边有一个“新话题”按钮，图标是一个清除灰尘的扫帚。点击这个按钮，当前对话就会清空，并开启新对话。丹姿克告诉我这一模块是在必应AI自己的帮助下开发的。

Though tools like Bing A.I. promise extreme, almost unimaginable convenience for users, they are likely to be even worse for content creators than the search and social-media companies that have siphoned up the majority of digital-advertising dollars over the past decade. Bing A.I. does offer referrals to Web sites in the form of footnotes linking to URLs. But the URLs are intentionally unobtrusive, to minimize for users what one Microsoft staffer described to me as the “cognitive load” of having to click on and scroll through links. The other day, Mody demonstrated over video chat how she could ask Bing A.I. to find a good vegetarian recipe for dinner. The bot pulled up a Bon Appétit recipe for vegetarian lasagna (like The New Yorker, Bon Appétit is owned by Condé Nast) and reprinted it in full within the chat. Then Mody asked it to list all of the ingredients and arrange them by grocery-store aisle—a request that no cooking Web site could hope to fulfill.

必应AI这样的工具给普通用户带来的是几乎难以想象的极大便利，但其给内容创作者造成的影响之大，远甚于之前的搜索和社交媒体公司（这些公司在过去十年里吸走了数字广告收入的大头）。的确，必应AI通过脚注链接的形式注明信息来源网站，但这些链接都有意做得不起眼，以期为用户省去点击链接并阅读网页的功夫，也就是微软员工所谓的减轻用户“认知负荷”。几天前，穆迪（Mody）通过视频展示了她如何通过询问必应AI找到一份不错的素食晚餐食谱。必应AI从《好胃口》（Bon Appétit）杂志（和《纽约客》（The New Yorker）杂志同属康泰纳仕集团（Condé Nast））调出了一份素食千层面食谱，并在聊天页面里全部转发了过来。随后，穆迪让必应AI列出食谱的所有原材料，并标明其摆在食品杂货店的哪个货架上——这是任何烹饪网站都无法实现的功能。

Later, on my own, I asked Bing A.I. to give me the latest news about First Republic Bank and the larger unfolding banking crisis. It generated a summary of breaking news, footnoting articles from NBC, CNN, and the Wall Street Journal, which is paywalled. (The Journal has said that any A.I. referencing its material should pay for a proper license, though it may have a hard time enforcing that for publicly accessible articles, because A.I. search crawls the entire Web just as Google does.) Then I asked Bing to present the news in a bulleted list in the style of the newsletter publication Axios. The result was a somewhat dry but otherwise convincing imitation. Another time, when I asked Bing for wallpaper options suitable for bathrooms with showers, it delivered a bulleted list of manufacturers. Instead of Googling for a listicle, I’d “co-created” one with the bot.

之后，我亲自体验了一下。我让必应AI为我提供关于第一共和银行（First Republic Bank）的危机以及更大规模的银行业危机的最新消息，然后它给我生成了一份突发新闻小结，其中包含了来自NBC、CNN和《华尔街日报》（Wall Street Journal）的文章。《华尔街日报》网站上的文章是付费内容，其也曾表示，任何引用其内容的人工智能工具都应支付适当的许可费用，不过对于公开的文章，可能很难强制执行这一规定，因为人工智能工具会像谷歌一样进行全网检索。然后我让必应AI以Axios简报列表的形式将这些新闻呈现给我。必应AI模仿得很生硬，但效果还是让我惊叹。还有一次，我让必应AI帮我寻找适合贴在浴室内的壁纸，它给我生成了一个壁纸商列表。这一列表并非是我在谷歌上搜索得来，而是由我和聊天机器人“协同创建”。

So much of the current Web was designed around aggregation—lists of product recommendations on The Strategist, summaries of film reviews on Rotten Tomatoes, restaurant reviews on Yelp. What value will those sites have when A.I. can do the aggregation for us? If Google Search is an imperfect book index, telling us where to find the material we need, Bing A.I. is SparkNotes, allowing us to bypass the source material altogether. Users might simply “read” publications in the form of A.I. chat summaries, as if listening to a mechanized butler reciting newspaper headlines aloud. The paradox of A.I., though, is that it relies on the source material—the vast sea of information that other sites create—to generate its answers. For that reason, it’s easy to imagine a kind of vicious cycle caused by the widespread adoption of tools like Bing A.I. If users don’t have to visit sites directly anymore, then those sites’ business models, based on advertising and subscriptions, will collapse. But if those sites can no longer produce content then A.I. tools won’t have fresh, reliable material to digest and regurgitate.

如今很多网站都是聚合类网站，如The Strategist杂志的产品推荐列表、烂番茄的电影评论汇总、Yelp的餐馆评论。当人工智能可以为我们聚合信息时，这些网站还有什么价值？如果说谷歌搜索是一个不完美的图书索引，告诉我们在哪里可以找到需要的材料，那么必应AI就类似文学指南网站Sparknotes，让我们可以完全绕过原始材料。用户通过浏览人工智能生成的汇总信息来“阅读”出版物，就像听一个机械管家大声朗读报纸头条。但人工智能的矛盾之处就在于，它所生成的回答，正是依赖于其他网站创造的海量信息。因此，不难想象，大范围使用诸如必应AI的工具会带来恶性循环。如果用户再也无需直接访问网站，那么这些基于广告和订阅模式盈利的网站将轰然倒塌。而如果这些网站不再生产新的内容，那么AI工具也就没有新鲜可靠的材料来吸收和整合。

注释：

SparkNotes：Sparknotes是美国著名文学指南网站，它提供了海量优质的文学指南学习资料，很多名校英美文学考研试题都可以在该网站找到权威答案，国内的大部分英美文学考研指导书都曾大段引用过Sparknotes里的内容。

http://www.sparknotes.com/

Speaking of fresh and reliable material, what happens when more and more of the material found online is generated by artificial intelligence? Last week, Google and Microsoft both announced a suite of A.I. tools for office workers, with applications that can compose new e-mails, reports, and slide decks or summarize existing ones. Similar tools will almost certainly soon be invading other aspects of our digital lives. In New York, John Herrman described the likely effects as “textual hyperinflation.” When the bot is generating material that workers used to produce, it will become difficult to determine which pieces of content in your in-box are meaningful; there will be no guarantee that any human labor or thought went into a given e-mail or report. We are all already bombarded by spam, mostly of a man-made variety. This will be a new kind of A.I.-generated spam on an unprecedented scale, much of it difficult to distinguish from content made by humans. Sooner than we think, content mills may be using A.I. to generate full articles, publicists may be using it to write press releases, and cooking sites might use it to devise recipes. Humans will need help navigating the glut, but media companies will have fewer resources to devote to serving that need. In such a scenario, A.I. could solve the very problem that it creates: eventually, if tools like Bing A.I. cause the well of original material online to run dry, all that will be left is self-referential bots, dredging up a generic set of answers that machines created in the first place.

说到新鲜可靠的材料，当网上越来越多的材料都是由人工智能生产的，结果会怎样？上周，谷歌和微软都发布了一套面向办公群体的人工智能工具，其中的应用程序可以编写新的电子邮件、报告和幻灯片，还可总结现有的材料。毋庸置疑，类似的工具将会渗透到我们电子生活的其他方面。《纽约》杂志记者约翰·赫尔曼（John Herrman）将其潜在影响描述为“文本恶性通胀”。若由机器人来生产曾经由人类生产的材料，我们就难以确定收件箱中的哪些内容是有意义的，我们无法确保某份邮件或报告中包含了人类的劳动或思想。我们已经苦于垃圾邮件的狂轰滥炸；至少现在这些垃圾邮件大部分还是人造的，然而未来我们将要面对的是规模空前的AI生成的垃圾邮件，且其中很多内容难以与人类制造的内容区分开来。也许很快，内容生产者会使用AI生成完整的文章，宣传人员用AI来撰写新闻稿，烹饪网站用AI来设计食谱。人们将需要帮助，来消化过剩的信息，但媒体公司能够用于满足这一需求的资源将会更少。在此情况下，人工智能可以解决自身创造出的问题：最终，如果像必应AI这类工具造成网络原始资料的枯竭，那么剩下的就会是不断自我引用的机器人，它们只能重新翻找机器最开始创造的答案。

In the meantime, non-automated text online will become an artisanal good, a product that we seek out for its unadulterated quality: instead of natural wine, “natural language.” On Tuesday, Google announced the release of its own A.I. chatbot. Called Bard—a more evocative and lofty title than Microsoft’s dry “A.I.-powered copilot”—it amounts to a broadside in the A.I. arms race among tech giants. But Google has conspicuously kept the chatbot separate from its signature product. “We think of Bard as complementary to Google Search,” one company executive told the Times. It’s a tacit acknowledgment of A.I.’s threat to the company’s current model. In its race to catch up with Microsoft, Google must try to avoid cannibalizing itself. Bing, however, is cheerily ushering us into the post-search future. 

与此同时，互联网上非AI生成的文本将会成为不添加“杂质”的“手工艺品”，只不过我们此处谈论的主题不是酿造葡萄酒，而是产出自然语言。周二，谷歌发布其人工智能聊天机器人Bard。相比于微软发布的“人工智能助力的copilot”，Bard这个名字更引人共鸣，更意境高远。这意味着科技巨头之间一场激烈的人工智能军备竞赛已然开始。但很明显，谷歌将其聊天机器人与其招牌产品谷歌搜索区分开来。谷歌一位高管告诉《泰晤士报》：“我们认为Bard是对谷歌搜索的补充。”这一言论默认了人工智能对谷歌现有模式的威胁。在追赶微软的竞赛中，谷歌必须努力避免自身产品的互相残杀。然而必应正兴高采烈地引领我们走进后搜索时代。

Yo，女，种下过流星，立志不做大鸵鸟
Humi，在躺平、侧卧和睡梦中寻找诗和远方

Rex，男，集书狂魔，经学钢粉，总觉得24小时不够用

Des，谷歌前员工
Iris，博物馆小可爱
Ithil，胡辣汤爱好者

Colin，男，崇拜科比的一枚小翻译

4