ChatGPT加入搜索后，会颠覆Google吗？
这是过去几周里，在科技产业圈里热议的话题。数字化转型网www.szhzxw.cn
1月3日，科技媒体The Information爆料了人工智能研究实验室OpenAI与微软Bing最新的合作计划。两位直接参与该计划的人士称，微软正与OpenAI一起研究一款具备AI对话能力的新版Bing搜索引擎，预计将于3月底前公布。该版本将融合了OpenAI最新推出的对话大模型ChatGPT，与以往单纯罗列网页链接的搜索模式不同，ChatGPT可以直接回答搜索查询。
DA Davidson分析师甚至因此给予了微软“买入”评级，并称，短期内OpenAI的ChatGPT，将会转化为Azure的增量。从长远来看，将ChatGPT功能整合到必应（Bing）可能会为微软提供十年一次的推翻谷歌搜索主导地位的机会。
在此合作之前，微软一直出钱出力，支持OpenAI业务。2020年，微软以10亿美元投资OpenAI。同年，微软Azuer云也成为了OpenAI的独家算力伙伴。在此之前OpenAI曾一度成为Google云的五大企业客户之一，年单1.2亿美元。不过，微软收购OpenAI之后，似乎从微软拿到了极低的云计算折扣。有数据显示，在2021财年中，OpenAI在微软产品上的开支仅为92000美元，而以OpenAI运维、研发GPT-3和GPT-3.5等大模型的进度来看，云计算用量并未有明显下降。数字化转型网www.szhzxw.cn
与Bing的合作被曝出之后，OpenAI 在1月6日早些时候又传出出售股票的消息。据《华尔街日报》报道，OpenAI正在与Thrive Capital、Founders Fund两家基金谈判，以 290 亿美元的估值出售至少价值 3 亿美元的股票。2021年，OpenAI员工向红杉资本、老虎环球出售股票时的估值约为200亿美元。
一、AI对于搜索行业太超前？
既然通用大模型如此厉害，为何在搜索上积累最深的Google不做，毕竟要论搜索，Google在业内有绝对优势。
据《纽约时报》报道，ChatGPT发布之后不久，Google首席执行官Sundar Pichai就在内部动员研究人员团队想办法应对ChatGPT，并宣布这种情况是“红色代码”的威胁。
其实Google也并非没有尝试，在2021年5月的I/O大会上Google就展示了语言类通用大模型LaMDA，并持续对其进行迭代研发。然而直至目前，Google都没有向公众发布LaMDA，Google方面称，部分原因在于LaMDA目前尚不完美，其中存在较多错误。数字化转型网www.szhzxw.cn
事实上，在大模型尚未成熟的时代，搜索引擎就已开始尝试AI技术了。在查询过程中，搜索引擎利用自然语言处理（Natural Language Processing，NLP）分析和理解用户输入的查询词语。在搜索过程中，利用深度神经网络（Deep Neural Networks，DNN），解决“隐式搜索”问题。
“隐式搜索是百度内部的说法，也可以叫做‘真正的语义搜索’。”一位百度搜索技术专家告诉虎嗅，隐式搜索不再是单纯的字面搜索，而是利用AI分析用户关键词语的深层意思。例如，搜索“现实版钢铁侠”，得到的搜索结果就都是马斯克相关的内容。
此外，Google、百度、Bing等还利用AI实验了类似ChatGPT的AI功能，使搜索呈现出的结果不只是罗列网页，更能从搜索结果中筛选内容，直接回答用户的问题。
例如搜索“ChatGPT是什么”。
Google会在侧边栏直接引用维基百科：“ChatGPT是由OpenAl开发的一个人工智能聊天机器人程序，于2022年11月推出……”
百度则会筛选搜索内容，并摘取答案直接回答在第一条的显著位置：“聊天机器人，chatgpt是OpenAl开发的一个大型预训练语言模型……”数字化转型网www.szhzxw.cn
不过这种技术在搜索行业并没有得到广泛认可，甚至没有统一的名字。“你可以管这种技术叫精准搜索，也可以叫唯一答案或阿拉丁。”
即便是ChatGPT搜索，也仍存在一些问题。
首先，AI挑选出的内容很难确保准确性，连OpenAI的CEO Sam Altman都认为，“现在依赖它（ChatGPT）来做任何重要的事情都是错误的。”如果把ChatGPT提供的搜索结果作为“唯一答案”呈现给用户，却无法保证其准确性的话，那么在错误的答案面前，再好的交互体验，都会成为笑话。
此外，有的时候用户使用搜索引擎的目的是直接得到答案，比如你搜“ChatGPT是什么”时，你想看到的就是一个定义。但当你搜“关于AI通用大模型的技术文章”时，你可能需要的是搜集广泛、全面的资料，如果AI筛选了部分内容，则可能影响你在搜索结果中博采众长。
二、更重要的是经济账
不过，虽然搜索引擎很早就融入了AI技术，但却很难有更深入的应用，其中的原因就是算不好经济账。
通用AI大模型的研发、运维投入都不是小数目，更别说加入了搜索。
“搜索”是一项开销巨大的业务，要利用爬虫从互联网海量数据中抓取搜索结果，背后需要大量的算力支持。据业内人士透露，百度搜索处理的网页数量在七八年前就已达到千亿级别。服务器用量超过一万台，每年纯设备成本超50亿。如果算上研发、运营和维护的人工成本，总开销可能会接近100亿。虽然近两年服务器成本或有下滑，但人力成本会一直随着通胀逐年上升。
而这只是不包含ChatGPT情况下的纯搜索成本。
数据显示，GPT-3的训练成本约1200万美元，而ChatGPT的运行成本大概是单轮回答（Single Turn）几美分，听起来不多，但叠加到搜索引擎每天处理的数亿、数十亿次搜索中，产生的开销就高得过分了。如果按照每天10亿次搜索量来计算，把ChatGPT的回答成本压到最低的1美分，那么按照现在的汇率看，这样一个搜索引擎2天的“答题成本”就要一个半“小目标”。数字化转型网www.szhzxw.cn
由于ChatGPT的训练并不是基于非实时更新的数据库（数据皆采集于2021年以前），它的搜索功能无法解答2021年以后的新问题。目前要对ChatGPT进行实时训练，就要面临新的技术挑战和成本问题。
不过也有AI行业专家认为，实时更新的AI搜索引擎实用价值并没有想象中那么大。虽然互联网中每天新增的数据量很大，但真正有价值，值得沉淀下来的知识非常有限。用户对于这些知识的需求，也是随时间推移阶段性增长的。这样就可以把重新训练大模型或更新数据库的时间拉长，从而摊薄成本。
那么花大价钱融入搜索引擎的ChatGPT能带来多少回报呢？
与上边提到过的“精准搜索、唯一答案、阿拉丁”技术相似，ChatGPT也无法直接提供经济效益。曾经的谷歌广告团队负责人Sridhar Ramaswamy就表示，ChatGPT的使用，从一定程度上阻止了用户点击带有广告的谷歌链接。Google在2021年的广告业务收入为2080亿美元，占Alphabet总收入的81%。
为何Bing可以去做与ChatGPT结合的尝试？
“单看搜索业务，国内头部搜索引擎的利润率大概在65%左右。”一位搜索行业专家向虎嗅透露，虽然成本会随着服务器和人员的多少增减，但要保证搜索引擎可用，单个服务器的成本差距不大，广告规模却会差很多。因此营收规模越大，利润率也越高。放在任何一个成熟的海外市场，头部搜索引擎的利润率也都差不多是这个数。
据微软财报显示，在截至2022年6月的财年数据中，搜索、MSN.com和其他新闻产品的广告收入共116亿美元，其中搜索业务占比较大。相比之下，Google搜索业务的同期收入至少是Bing的10倍。
据海外多家调研机构的统计显示，Bing在美国本土的市场占有率远低于Google，目前全球搜索市场中Bing的市占率不超过10%，而Google的市占率保守估计也不低于80%。从市占率看来，Bing和Goggle二者的竞争关系似乎并不明显。数字化转型网www.szhzxw.cn
因为市场占比小，搜索利润率相对较低，Bing很自然地不承担过多的营收压力，也就可以放心大胆地尝试ChatGPT这种未来的颠覆式创新，而不需要考虑太多成本。
技术创新的故事就是如此充满戏剧性和残酷性。有时候赌对了一个技术路线，就能赢下未来十年甚至几十年的商业战争。而原有技术路线的优势者，往往由于自身的路径依赖，反过来被自己的优势地位掣肘，难以在新技术上勇往直前。正如曾经的光刻机霸主尼康赌错了技术路线，在浸润式光刻机上被荷兰ASML公司捷足先登，丧失了此后几十年的市场领先地位；诺基亚在智能手机时代被后继者超越，跌下神坛。ChatGPT大规模商用后会怎么样，谁也不知道，而一旦爆发，将是颠覆式创新，谁赢下了这场赌注，谁就能赢下下一个人工智能时代。
翻译：
Will ChatGPT upend Google when it joins Search?
That’s been the talk of the tech industry for the past few weeks.数字化转型网www.szhzxw.cn
On January 3, The Information broke the news about the latest partnership between AI research lab OpenAI and Microsoft’s Bing. Microsoft is working with OpenAI on a new version of its Bing search engine with AI-enabled conversations, which is expected to be announced by the end of March, said two people directly involved in the plans. This version will incorporate OpenAI’s new ChatGPT conversation model. Instead of simply listing web links, ChatGPT can answer search queries directly.
Analysts at DA Davidson even have a “buy” rating on Microsoft, saying OpenAI’s ChatGPT will translate into incremental Azure growth in the near term. In the long run, integrating ChatGPT features into Bing could provide Microsoft with a once-in-a-decade opportunity to topple Google’s search dominance.
Prior to this partnership, Microsoft had been supporting the OpenAI business.
In 2020, Microsoft invested $1 billion in OpenAI. That same year, Microsoft’s Azuer cloud also became OpenAI’s exclusive computing partner. Previously, OpenAI had been one of Google Cloud’s top five enterprise customers, with an annual revenue of $120 million. However, Microsoft seems to have gotten a very low cloud discount from Microsoft by acquiring OpenAI. In FY2021, OpenAI only spent $92,000 on Microsoft products, and the pace of OpenAI operations and development of larger models such as GPT-3 and GPT-3.5 did not show a significant decline in cloud computing usage.
After the Bing partnership was announced, OpenAI announced a stock sale earlier this week. According to the Wall Street Journal, OpenAI is in talks with two funds, Thrive Capital and Founders Fund, to sell at least $300 million worth of shares at a valuation of $29 billion. In 2021, OpenAI employees sold shares to Sequoia Capital and Tiger Global at a valuation of about $20 billion.
Is AI too ahead of its time for search?
If the general model is so powerful, why not Google, which has accumulated the most in search, after all, Google has an absolute advantage in the industry.数字化转型网www.szhzxw.cn
Shortly after ChatGPT was released, Google CEO Sundar Pichai internally mobilized a team of researchers to figure out how to deal with ChatGPT and declared the situation a “Code red” threat, according to The New York Times.
In fact, Google is not without attempts. At the I/O conference in May 2021, Google demonstrated LaMDA, a universal language model, and continued to carry out iterative research and development on it. Until now, however, Google hasn’t released LaMDA to the public, in part because, Google says, LaMDA isn’t perfect yet and has a lot of bugs.
In fact, search engines have been experimenting with AI since before the big models matured. In the query process, the search engine uses Natural Language Processing (NLP) to analyze and understand the query words entered by users. In the search process, Deep Neural Networks (DNN) are used to solve the “implicit search” problem.
“Implicit search is Baidu’s internal term, which can also be called ‘true semantic search.'” A Baidu search technology expert told Husniff that implicit search is no longer a simple literal search, but uses AI to analyze the deep meaning of users’ key words. A search for “real life Iron Man,” for example, brings up all Musk-related results.数字化转型网www.szhzxw.cn
In addition, Google, Baidu, and Bing are also experimenting with Chatgpt-like AI features that allow search results to not only list web pages, but also filter content from search results to directly answer users’ questions.
For example, search “What is ChatGPT?”.
Google will quote Wikipedia directly in the sidebar: “ChatGPT is an AI chatbot program developed by OpenAl and launched in November 2022…”
Baidu, in turn, screens search content and extracts the answers directly to the first prominent position: “Chatbot, chatgpt is a large pre-trained language model developed by OpenAl…”
But the technology isn’t widely recognized in the search industry, or even has a name. “You can call it precision search. You can call it single answer or Aladdin.”
Even ChatGPT search still has some problems.数字化转型网www.szhzxw.cn
For one thing, it’s so difficult to ensure accuracy in what the AI picks out that even OpenAI CEO Sam Altman believes that “it’s a mistake to rely on it [ChatGPT] to do anything important right now.” If ChatGPT’s search results are presented to the user as the “only answer,” with no guarantee of accuracy, then a good interaction becomes a joke in the face of the wrong answer.
In addition, sometimes the purpose of a search engine is to get a direct answer. For example, when you search “What is ChatGPT?”, you want to see a definition. But when you search for “technical articles on general AI large models,” you may want to collect broad, comprehensive information, and if the AI filters some of the content, it may affect your ability to learn from others in the search results.
More important is the economic ledger
However, although search engines have long incorporated AI technology, it has been difficult to make further use of it, in part because of poor economic calculations.
The research and development, operation and maintenance of general AI models are not small amounts, let alone the addition of search.
“Search” is a costly business. To use crawler to grab search results from massive data on the Internet, a lot of computing power is needed behind it. According to industry insiders, the number of web pages handled by Baidu search has reached 100 billion level seven or eight years ago. The number of servers exceeds 10,000, and the annual pure equipment cost exceeds 5 billion. If labor costs for research and development, operations and maintenance are included, the total could be closer to $10 billion. Although server costs may have fallen in the last two years, labor costs will continue to rise with inflation.
And that’s just the pure search cost without ChatGPT.
According to the data, GPT-3 costs about $12 million to train, while ChatGPT costs about a few cents to run in a Single Turn, which doesn’t sound like much, but when you add it to the hundreds of millions and billions of searches that search engines handle every day, the overhead is excessive. If you take ChatGPT’s answer costs down to a penny, based on a billion searches per day, then at today’s exchange rates, the “answer costs” for such a search engine in two days would be a “small target” of one and a half.
Because ChatGPT’s training isn’t based on a database that isn’t updated in real time (the data was collected before 2021). Its search function won’t be able to answer new questions beyond 2021. To train ChatGPT in real time, there are new technical challenges and cost problems.
However, some AI industry experts believe that the real-time update AI search engine practical value is not as great as imagined. Although the amount of data added to the Internet every day is huge. The amount of knowledge that is truly valuable and worth depositing is very limited. Users’ demand for such knowledge also increases gradually over time. This can spread the cost by extending the time it takes to retrain a large model or update a database.数字化转型网www.szhzxw.cn
So what’s the return on ChatGPT, which pays a lot of money to be integrated into a search engine?
Like the “precise search, single answer, Aladdin” technology mentioned above, ChatGPT does not provide direct economic benefits. Sridhar Ramaswamy, a former head of Google’s advertising team. Said the use of ChatGPT partly prevented users from clicking on Google links with ads. Google will generate $208 billion in AD revenue in 2021, accounting for 81% of Alphabet’s total.
Why did Bing try to combine ChatGPT?
“Looking at the search business alone, the profit margin of the top search engine in China is probably about 65 percent.” A search industry expert told Tiger that although the cost will increase and decrease with the number of servers and personnel, but to ensure that the search engine is available. The cost difference of a single server is not big, but the advertising scale will be much different. So the bigger the revenue, the higher the profit margin. In any mature overseas market, the head search engine’s profit margin is about the same.
Search accounted for the lion’s share of the $11.6 billion in advertising revenue from search, MSN.com and other news products in the fiscal year ending June 2022, according to Microsoft’s financial results. By comparison, Google’s search revenue over the same period was at least 10 times that of Bing.
According to the statistics of several overseas research institutions. Bing’s market share in the United States is far lower than Google’s.
At present, Bing’s market share in the global search market is no more than 10%, while Google’s market share is conservatively estimated to be no less than 80%. The competition between Bing and Goggle doesn’t seem obvious in terms of market share.
Because of its small market share and relatively low search margins, Bing naturally doesn’t have a lot of revenue pressure on it. So it can take a bold shot at future disruptive innovations like ChatGPT without much cost.数字化转型网www.szhzxw.cn
That’s how dramatic and cruel the story of technological innovation is. Sometimes the right bet on a technology route can win the business wars of the next decade or even decades. However, those who have advantages in the original technological route are often hampered by their own advantageous position due to their own path dependence, and it is difficult to go forward in the new technology. Just as Nikon, the former lithographic machine leader, gambled on the wrong technical route and was beaten by ASML of the Netherlands in infiltrating lithography machine, losing its market leading position for decades. Nokia was overtaken by its successor in the smartphone era and fell from its pedestal. No one knows what ChatGPT will do when it is commercialized on a large scale. But it will be a disruptive innovation, and whoever wins the bet will win the next era of AI.
本文由数字化转型网（www.szhzxw.cn）转载而成，来源：虎嗅APP；作者：齐健；编辑/翻译：数字化转型网宁檬树。
免责声明: 本网站(http://www.szhzxw.cn/)内容主要来自原创、合作媒体供稿和第三方投稿，凡在本网站出现的信息，均仅供参考。本网站将尽力确保所提供信息的准确性及可靠性，但不保证有关资料的准确性及可靠性，读者在使用前请进一步核实，并对任何自主决定的行为负责。本网站对有关资料所引致的错误、不确或遗漏，概不负任何法律责任。
本网站刊载的所有内容(包括但不仅限文字、图片、LOGO、音频、视频、软件、程序等) 版权归原作者所有。任何单位或个人认为本网站中的内容可能涉嫌侵犯其知识产权或存在不实内容时，请及时通知本站，予以删除。
