导读：微软重新用ChatGPT挑战谷歌的搜索霸主地位，谷歌又将如何应对，当然本文还有更详细的内容。
在未来几周内，微软将 OpenAI 即将推出的大型语言模型 GPT-4 整合到 Bing 搜索引擎中。
当今搜索引擎霸主谷歌起初对ChatGPT并不在意，本来自己是AI的领养羊，而今却如坐针毡，它开始不断测试竞争对手的人工智能产品，担心自己可能会失去在线搜索领域的主导地位，并准备备战应对微软。
谷歌员工们在应对微软挑战的同时，感慨的让 ChatGPT 写了一首有关谷歌裁员的诗：
内容也算有些韵味。但是微软的GPT搜索引擎已经箭在弦上，必然要迎接挑战。
何为GTP？GTP 文本生成模型已经从能够自动完成文本，发展到能够执行更通用的自然语言任务。最新的GPT-4 技术预计将比其前身 ChatGPT 与 GPT-3 更加强大，能够生成更高质量文本和响应更广泛输入查询等新功能。
作为 OpenAI 的主要投资者，微软获得了 OpenAI 产品的独家部署权。据SemaFor称，微软将使用 GPT-4 为Bing 提供支持，在未来几周内将新系统词整合并部署到全球 Bing 搜索产品体系中。
目前尚不清楚这些类型的模型如何应用到搜索引擎产品中。有一些数据表明大型语言模型可以从其训练数据中回忆与提取信息，但它们常常会弄错事实。因为互联网在不断变化；新的网站和页面会随着新知识的出现而不断更新与添加。然而像 ChatGPT 这样的工具只能在其训练数据的范围内运行。比如GPT-3 的最初版本是根据 2019 年之前收集的互联网文本进行训练的，它仍然会说特朗普还是现任美国总统。
微软笃定是要用GTP改造 Bing，而谷歌肯定不肯居于人后。
谷歌美国本部正在加紧测试与 ChatGPT 相媲美的大型语言模型，构建新的人工智能搜索产品，以保持最高竞争力。谷歌高管们启动了一个代号为“Atlas”的项目，并召集了工程师，并指派他们开发一个类似的系统。据CNBC报道，谷歌开发人员还在试验一种名为“Apprentice Bard”的聊天机器人，用于构建基于其 LaMDA 语言模型的搜索桌面应用程序。
除了收购以外，最近也有消息显示谷歌正在创始人拉里・佩奇、谢尔盖・布林的带领下自研 ChatGPT 竞品的消息。自 OpenAI 的 ChatGPT 上线以来，谷歌就一直承受压力，这让谷歌高管对公司的发展方向进行了紧急修正，在大裁员的同时集中精力推动“最高优先事项”。
严峻的形势，让一向引领 AI 方向的谷歌不得不跟进。谷歌即将在 2 月 8 日举行的 YouTube 直播活动，将介绍公司在“如何利用人工智能的力量重新构想搜索信息、探索问题和与信息互动的方式，让人们比以往更自然、更直观地找到所需信息”。
其内部正在测试一款名为“Apprentice Bard“的 AI 模型，它基于一种”突破性的对话技术”——LaMDA ，可提供问答搜索能力。
这还没完，甚至连布林本人都开始下场改代码了：被曝亲自上手改LaMDA代码，具体针对的是训练LaMDA的数据。
谷歌 CEO Sundar Pichai 在财报电话会议中表示，谷歌可能正准备通过让人们”与其作为搜索伴侣的最新、最强大的语言模型直接交互“，以此和 OpenAI 的 ChatGPT 竞争。
一、ChatGPT原班人马新立门户，谷歌急投4亿美金
在微软与 ChatGPT 创建者 OpenAI 绑定在一起之后，谷歌除了自研AI产品应对之外，还选择支持一家由前 OpenAI 员工创立的新AI公司，名字称为：
在2022 年底，谷歌向这家初创公司投资了约 3 亿美元（彭博社称近 4 亿美元），不过并未第一时间披露这一消息。英国《金融时报》称，作为这笔钱的回报，谷歌获得了该公司 10% 的股份，Anthropic 被要求从这家搜索巨头那里购买谷歌云计算资源。新融资将使这家总部位于旧金山的初创公司估值增至近 50 亿美元。
可以说，谷歌完整复制了微软和 OpenAI 之间的合作关系链。
Anthropic 是由几位离开 OpenAI 的研发人员在 2021 年创立。包括 OpenAI 前研究副总裁 Dario Amodei、其妹妹Daniela、GPT-3 论文主创 Tom Brown 等人，并于去年 4 月获得更多关注，成立不到一年，它宣布获得高达 5.8 亿美元的资金。
Anthropic 的两个创始人均提出让AI更易用的目标。
Anthropic近期推出的生成式 AI 模型“Claude”，被业界认为是 ChatGPT 的有力竞争者。Claude 似乎在关键方面对原版系统做了优化改进，可通过 Slack 集成访问，但处在封闭测试阶段并未公开。
去年 12 月，Anthropic 发布了一篇题为《Constitutional AI: Harmlessness from AI Feedback》的论文。这篇论文描述了一个 520 亿参数的模型 ——AnthropicLM v4-s3。该模型是在一个大型文本语料库上用无监督方式训练的，非常像 GPT-3。
Anthropic 表示，Claude 是一个新的、更大的模型，其架构选择与已发表的研究结果相似，Claude 便以此为基础来构建。
Claude 也通过从网络上获得的大量文本示例进行训练，根据语义上下文等模式了解单词出现的可能性。它可以就广泛的主题进行开放式对话，讲笑话和讲哲学都可以。
同时，它也有着与 ChatGPT 一样的缺陷：有的用户说 Claude 的数学比 ChatGPT 差，犯了明显的错误并且未能给出正确的后续响应；它的编程水平也有所欠缺，可以更好地解释自己写的代码，但在 Python 以外的语言上目前还不太在行。
目前，尚不清楚谷歌是否计划像微软对 ChatGPT 所做的那样将 Claude 集成到Google服务中。谷歌已拥有大量开发人工智能语言系统的内部专业知识，《金融时报》表示，这笔交易的一个动机就是建设谷歌的云计算业务。
根据《金融时报》所指出，Dario Amodei 是“在对公司的发展方向存在分歧之后”从 OpenAI 中分离出来的 —— 即 OpenAI 公司在 2019 年与微软的首次交易后越来越关注商业化。
一些 AI 研发人员认为 OpenAI 的行为商业化严重，对模型质量、性能的考虑欠缺。即便 ChatGPT 能够迅速生成问题的详细回答，但 OpenAI 没有推出适当的检验方法来判断 ChatGPT 输出的正确与否。
Anthropic 强调其目标在构建“可靠、可解释和可操纵的人工智能系统”。
二、各家科技巨头跟进
OpenAI 首席执行官 Sam Altman 表示，”ChatGPT 比我想象得要有用得多“。他认为 ChatGPT 除了写长文的简短总结之外，对他自己实际工作有用的能力还包括写代码，”它能够提出深奥的编程问题或帮助调试代码，感觉就像有一个可以与之交谈的超级聪明的程序员。“
无论如何，Altman 都打赌已经有足够多的人发现 ChatGPT 足够有用，能值回每月 42 美元的ChatGPT Plus会员价了。在动作最快的微软那里，ChatGPT 本周已经加入了 Teams 产品中，还可以在 Outlook 上帮人写电子邮件。
比尔盖茨表示，像 ChatGPT 这样的人工智能的兴起将与互联网的诞生或个人PC的发展一样重要。
近来，一些科技公司在尝试将聊天机器人与网络搜索的功能结合起来，微软、谷歌、百度等大型科技公司都已入局。
而以元宇宙为重心的科技巨头 Meta 感受到的情绪似乎有所不同。作为 Meta AI 首席科学家，图灵奖得主 Yann LeCun 最近讨论的主要是”ChatGPT 乃至 OpenAI 这家公司都没有实质创新“的问题：
Yann LeCun：一些机构发布知识和开源代码供全世界使用，而其他机构只是消费它。
Meta的表现，有点吃葡萄的感觉，也有技术人员相轻的色彩，也有可能有自己的技术原则与哲学。
不过微软可是讯雷不及掩耳之势，在今天凌晨上线了加入ChatGTP的Bing搜索引擎。
有一些 Bing 用户表示，“新 Bing”界面今天早上刚刚出现——然后又消失了。
在Twitter有多个用户均发表了报告，更新的Bing搜索引擎的变化，也可能对不同的用户有增量的的访问权限。
这个增加了 AI 增强型的 Bing 屏幕截图显示，菜单栏中“搜索”旁边出现了一个新的“聊天”选项。当选择它后，用户将被带到一个聊天界面，上面写着“欢迎使用新的必应：您 AI 驱动的答案引擎。” 接着是三个建议：用户“提出复杂的问题”、“获得更好的答案”和“获得创意灵感”。还有一个免责声明：“Bing 由 AI 提供支持，因此可能会出现意外与错误。确保检查事实并分享反馈，以便我们学习和改进。”
这些看起来都与 ChatGPT 本身 UI 相似，它还提供了有关如何最好地使用 AI 聊天机器人的建议，并警告其答案可能不可靠。
当 AI 语言模型加入搜索引擎最有用时，最后一点灵感是关键。GTP系统会以一定的频率让用户“产生幻觉”（这是 AI 所说的“编造东西”）。有专家警告说，将它们集成到搜索引擎中几乎肯定会提供虚假信息，问题是这些信息也可能是危险的。如果这些系统在没有适当保护的情况下上线，它们可能会提供有害的信息，如医疗建议、种族主义或存在偏见的历史等。
尽管存在这样那样的潜在问题，但 ChatGPT 的爆炸性流行似乎正在鼓励微软和谷歌加快其人工智能辅助未来的计划。
有网友称，微软一直在等待着一个胜利，鲍尔默时代低估了 Apple，从而搞砸了微软公司。现在他们找到了，如果谷歌没有现成的 AI 解决方案，后者很可能就完蛋了。Apple 不得不终止放弃 Google 每年 15B亿美元的使用费，转而从微软那里获得 ChatGPT 许可，除非Apple再为 Siri 构建一个AI平台。
今天的Bing或许是个彩蛋，微软必将在未来几周内发布正式版本的 Bing 。可以预见的是，两雄相争，用户得利。
今天是立春，一年之计在于春，一切刚刚开始。
翻译：
ChatGPT relaunches Microsoft’s challenge to Google’s dominance in search, and how Google will respond.
In the coming weeks, Microsoft will integrate OpenAI’s upcoming large-scale language model GPT-4 into its Bing search engine.
Google, the current dominant search engine, initially ignored ChatGPT. It had been adopted by AI, but now it was on tenterspits, testing rival AI products, fearing it might lose its dominance in online search, and preparing to take on Microsoft.
While Google employees are grappling with Microsoft’s challenges, they let ChatGPT write a poem about Google’s layoffs:
The content is also somewhat interesting. But Microsoft’s GPT search engine is ready to meet the challenge.
What is GTP? The GTP text generation model has evolved from being able to automate text to being able to perform more general natural language tasks. The latest GPT-4 technology is expected to be more powerful than its predecessors ChatGPT and GPT-3, with new capabilities such as generating higher-quality text and responding to a wider range of input queries.
As a major investor in OpenAI, Microsoft has the exclusive right to deploy OpenAI products. According to SemaFor, Microsoft will use GPT-4 to power Bing, integrating and deploying the new system word into the global Bing search product system in the coming weeks.
It is not clear how these types of models apply to search engine products. There is some evidence that large language models can recall and extract information from their training data, but they often get the facts wrong. Because the Internet is constantly changing; New sites and pages are constantly updated and added as new knowledge becomes available. Tools like ChatGPT, however, can only work within the scope of their training data. The original version of GPT-3, for example, trained on Internet texts collected before 2019, would still say that Mr. Trump was the current president of the United States.
Microsoft is definitely going to revamp Bing with GTP, and Google is definitely not going to be second best.
The U.S. headquarters of Google is testing a large language model similar to ChatGPT and building a new AI search product to stay competitive. Google executives started a project code-named Atlas, gathered engineers and assigned them to develop a similar system. Google developers are also experimenting with a chatbot called “Apprentice Bard” to build a search desktop application based on its LaMDA language model, according to CNBC.
In addition to the acquisition, it was recently revealed that Google is developing a rival to ChatGPT under the leadership of founders Larry Page and Sergey Brin. Google has been under pressure ever since OpenAI’s ChatGPT went live, leading Google executives to make an emergency correction in the company’s direction, focusing on pushing “highest priorities” amid sweeping layoffs.
The grim situation, so that has always led the direction of AI Google had to follow. Google’s upcoming YouTube live event on Feb. 8 will describe how the company is “harnessing the power of artificial intelligence to reimagine the way we search for information, explore questions, and interact with information to make finding the information we need more natural and intuitive than ever before.”
Internally, it is testing an AI model called the Apprentice Bard, based on LaMDA, a “breakthrough conversation technology” that provides question-and-answer search capabilities.
That’s not the end of it. Even Brin himself has been making changes to the LaMDA code himself: it’s been revealed that he changed the LaMDA code himself, specifically on the data that trained Lamda.
Google CEO Sundar Pichai said on the earnings call that Google may be preparing to compete with OpenAI’s ChatGPT by letting people “interact directly with the latest and most powerful language model as a search companion.”
Google invested $400 million in ChatGPT’s new portal
After Microsoft tied up with ChatGPT creator OpenAI, Google, in addition to developing its own AI product response, chose to support a new AI company founded by former OpenAI employees called:
At the end of 2022, Google invested about $300 million (Bloomberg says nearly $400 million) in the startup, though it didn’t immediately disclose the news. In return for the money, Google received a 10 percent stake in the company, and Anthropic was ordered to buy Google cloud computing resources from the search giant, the Financial Times said. The new funding will value the SAN Francisco-based startup at nearly $5 billion.
In a way, Google has completely copied the chain of relationships between Microsoft and OpenAI.
Anthropic was founded in 2021 by several developers who left OpenAI. Including former OpenAI research Vice President Dario Amodei, his sister Daniela, GPT-3 paper creator Tom Brown and others, and gained more attention last April, less than a year after it was founded, it announced up to $580 million in funding.
Dario Amodei, CEO of Anthropic
Daniela Amodei, president of Anthropic, is a former vice president of Security and policy at OpenAI
The two founders of Anthropic share the goal of making AI easier to use.
Anthropic’s recently launched generative AI model Claude is considered a strong competitor to ChatGPT in the industry. Claude seems to have improved on the original in key ways and is accessible through Slack integration, but is in closed beta and not publicly available.
In December, Anthropic released a paper titled “Constitutional AI: Harmlessness from AI Feedback.” The paper describes a 52-billion-parameter model, AnthropicLM v4-s3. The model was trained on a large textual corpus in an unsupervised manner, much like GPT-3.
Anthropic states that Claude is a new, larger model based on architectural choices similar to published research findings.
Claude is also trained on the possibility of words based on patterns such as semantic context, using numerous text examples from the Internet. It allows for open-ended conversations on a wide range of topics, jokes and philosophy.
At the same time, it suffers from the same flaws as ChatGPT: some users say Claude’s math is worse than ChatGPT’s, that he makes obvious mistakes and fails to give correct follow-up responses; It also lacks programming skills that allow it to interpret its own code better, but is currently not very good at languages other than Python.
At present, it is not clear whether Google plans to integrate Claude into Google services, as Microsoft did with ChatGPT.
Google already has considerable in-house expertise in developing AI language systems, and the Financial Times said one motivation for the deal was to build up Google’s cloud computing business.
According to the Financial Times, Dario Amodei was split from OpenAI “after disagreements over the direction of the company” — with OpenAI increasingly focused on commercialization following its initial deal with Microsoft in 2019.
Some AI developers believe that OpenAI’s behavior is heavily commercialized and lacks consideration for model quality and performance. Even though ChatGPT can quickly generate detailed answers to questions, OpenAI does not introduce proper checks to determine whether ChatGPT’s output is correct or not.
Anthropic emphasizes its goal of building “reliable, interpretable, and maneuverable artificial intelligence systems.”
The tech giants follow suit
“ChatGPT is much more useful than I thought it would be,” says Sam Altman, CEO of OpenAI. He credits ChatGPT’s ability to write code, in addition to short summaries of long articles, for his own practical work, “the ability to ask esoteric programming questions or help debug code, which feels like having a super-smart programmer to talk to. “
In any case, Altman is betting that enough people will find ChatGPT useful enough to justify the $42 a month ChatGPT Plus membership. At Microsoft, the fastest, ChatGPT was added to Teams this week and can also be used to write emails for people in Outlook.
Bill Gates says the rise of artificial intelligence like ChatGPT will be as important as the birth of the Internet or the development of the personal PC.
Recently, some tech companies have been trying to combine chatbots with web search capabilities. And major tech companies like Microsoft, Google and Baidu have all joined in.
Meta, a tech giant with a focus on the meta-universe, seems to feel a different mood. Turing Award winner Yann LeCun, chief scientist at Meta AI, recently discussed the “lack of real innovation in ChatGPT or even OpenAI as a company” :
Yann LeCun: Some organizations release knowledge and open source code for the world to use, while others just consume it.
The performance of Meta is a bit like eating grapes. But also has the light color of technicians, and may have its own technical principles and philosophy.
Microsoft, however, was surprised and launched its Bing search engine with ChatGTP in the early hours of this morning.
Some Bing users said the “new Bing” interface just appeared this morning — and then disappeared.
Multiple users on Twitter have reported changes to the updated Bing search engine, which may also have incremental access to different users.
This AI enhanced Bing screen shot shows a new “chat” option next to “Search” in the menu bar. When selecting it, users will be taken to a chat screen that says “Welcome to the new Bing: Your AI-powered Answer engine.” This is followed by three tips: users “ask complex questions,” “get better answers,” and “get creative.” Here’s a disclaimer: “Bing is powered by AI, so it’s subject to accidents and errors. Make sure to check the facts and share feedback so we can learn and improve.”
These all look similar to the ChatGPT UI itself, which also offers advice on how best to use the AI chatbot, with a warning that its answers can be unreliable.
New Bing can’t answer programming questions yet. The UI may be an early development project
This last bit of inspiration is key when AI language models are most useful in search engines. The GTP system causes users to “hallucinate” (which is what the AI calls “making things up”) with a certain frequency. Integrating them into search engines will almost certainly provide false information, some experts warn. The problem is that the information can also be dangerous. If these systems go online without proper protections, they can provide damaging information. Such as medical advice, a history of racism or bias.
Despite potential problems of one kind or another, ChatGPT’s explosive popularity seems to be encouraging Microsoft and Google to accelerate their AI-assisted future plans.
Some said Microsoft had been waiting for a victory and that the Ballmer era had ruined the company by underestimating Apple. Now they have, and if Google doesn’t have an AI solution in place, the latter is probably doomed. Apple had to stop giving up Google’s $15.B a year royalties and license ChatGPT from Microsoft unless it built an AI platform for Siri.
Today’s Bing may be an Easter egg, but Microsoft will release the official version of Bing in the next few weeks. Predictably, the two males fight, and the user benefits.
Today is the beginning of spring, the year’s plan in spring, everything just started.
本文由数字化转型网（www.szhzxw.cn）转载而成，来源：图灵人工智能；编辑/翻译：数字化转型网宁檬树。
免责声明: 本网站(https://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。
