What Is ChatGPT? What to Know About the AI Chatbot That Will Power Microsoft Bing
关于爆火的AI聊天机器人ChatGPT，你需要了解什么？
OpenAI’s chatbot has triggered a new artificial-intelligence race and may reshape the future of work
OpenAI的聊天机器人引发了一场新的人工智能竞赛，并可能重塑未来职场。
Within weeks of its launch, OpenA’s ChatGPT triggered a new global race in artificial intelligence. The chatbot is part of a fresh wave of so-called generative AI – sophisticated systems that produce content from text to images – that is set to be one of the most disruptive forces in a decade to Big Tech, industries and the future of work.
在OpenAI推出ChatGPT几周后，该产品就引发了一场新的全球人工智能（AI）竞赛。这款聊天机器人是新一波所谓的生成式AI浪潮的一部分。该技术指的是生成文本、图像等内容的复杂系统，或许是10年来对大型科技公司、众多行业和未来职场最具颠覆性的力量之一。
Microsoft Corp. has added the technology to its products, including search engine Bing, while competitors Google and Baidu are pushing to launch similar tools. Despite its sudden burst in popularity, the technology currently has serious limitations and potential risks that include spewing misinformation and infringing on intellectual property.
微软（Microsoft Corp., MSFT）已将该技术添加到搜索引擎必应（Bing）等产品中，而竞争对手谷歌（Google）和百度（Baidu, BIDU, 9888.HK）正努力推出类似工具。尽管突然大受欢迎，但该技术目前面临一些严重的局限性和潜在风险，包括传播错误信息和侵犯知识产权等。
Here’s what to know.
以下是需要了解的内容。
What is ChatGPT?
什么是ChatGPT？
ChatGPT is an artificial-intelligence chatbot developed by San Francisco-based AI research company OpenAI. Released in November 2022, it can have conversations on topics from history to philosophy, generate lyrics in the style of Taylor Swift or Billy Joel, and suggest edits to computer programming code.
ChatGPT是总部位于旧金山的AI研究公司OpenAI开发的一款AI聊天机器人。该产品2022年11月发布，可以进行从历史到哲学等话题的对话，生成泰勒·斯威夫特（Taylor Swift）或比利·乔尔（Billy Joel）风格的歌词，并对计算机编程代码提供编辑建议。
ChatGPT is trained on a vast compilation of articles, websites and social-media posts scraped from the internet as well as real-time conversations – primarily in English – with human contractors hired by OpenAI. It learns to mimic the grammar and structure of the writing and reflects frequently-used phrases.
ChatGPT是根据从互联网上搜罗的大量文章、网站和社交媒体帖子以及用户与OpenAI雇用的人工雇员的实时对话（主要是英语）进行训练的。ChatGPT学着模仿写作的语法和结构，并使用常用语。
The chatbot isn’t always accurate: its sources aren’t fact-checked, and it relies on human feedback to improve its accuracy.
ChatGPT并不总是准确的：它的消息来源没有经过事实核查，它依靠人类的反馈来提高准确性。
OpenAI developed ChatGPT as part of a strategy to build AI software that will help the company turn a profit. In January, Microsoft, its strategic partner, unveiled a fresh multibillion-dollar investment in OpenAI and said it plans to infuse ChatGPT into its Bing search app and other products.
作为建立AI软件策略的一部分，OpenAI开发了ChatGPT，该策略将有助于公司实现盈利。今年1月，其战略合作伙伴微软披露将对OpenAI进行价值数以十亿计美元的投资，并表示计划将ChatGPT嵌入其必应（Bing）搜索应用程序和其他产品。
How do ChatGPT and other AI chatbots work?
ChatGPT和其他AI聊天机器人是如何工作的？
The technology that underlies ChatGPT is referenced in the second half of its name, GPT, which stands for Generative Pre-trained Transformer. Transformers are specialized algorithms for finding long-range patterns in sequences of data. A transformer learns to predict not just the next word in a sentence but also the next sentence in a paragraph and the next paragraph in an essay. This is what allows it to stay on topic for long stretches of text.
支撑ChatGPT的技术在该软件名字的后半部分被提及，即GPT，它是Generative Pre-trained Transformer（生成预训练转换器）的缩写。转换器是在数据序列中寻找长程模式的专门算法。转换器不仅能学会预测一个句子中的下一个词，还能学会预测一个段落中的下一个句子以及一篇文章中的下一个段落。这就是为什么它能够在长文本中紧扣主题。
Because a transformer requires a massive amount of data, it is trained in two stages: first, it is pretrained on generic data, which is easier to gather in large volumes, and then it is fine-tuned on tailored data for the specific task it is meant to perform. ChatGPT was pretrained on a vast repository of online text to learn the rules and structure of language; it was fine-tuned on dialogue transcripts to learn the characteristics of a conversation.
由于转换器需要大量的数据，它的训练分为两个阶段：首先，它在通用数据上进行预训练，这种数据更容易大量收集；然后，根据它要执行的具体任务，利用定制的数据进行微调。ChatGPT利用一个庞大的在线文本库进行预训练，以学习语言的规则和结构：它通过对话记录进行微调，以学习对话的特征。
Developed by researchers at Alphabet Inc.’s Google in 2017, transformers have since become pervasive across dozens of technologies. Google uses them in its new experimental service Bard, which gives users conversational answers to their search queries. Chinese tech giant Baidu is using them to develop its own ChatGPT to integrate into its search engine by March.
转换器（Transformer）于2017年由Alphabet Inc.（GOOG）旗下谷歌的研究人员提出，自那以来普遍存在于众多技术。谷歌在其新的实验性服务Bard中用到了转换器；Bard针对用户的搜索查询提供对话式答案。中国科技巨头百度正使用转换器来开发自己的类ChatGPT聊天机器人，计划到3月将其整合到自己的搜索引擎中。
Transformers, which can be trained on images or images and captions simultaneously, are also the basis of image-generation software systems such as OpenAI’s Dall-E 2 and Stability.ai‘s Stable Diffusion.
转换器也是OpenAI旗下Dall-E 2、Stability.ai旗下Stable Diffusion等图像生成软件系统的基础。对转换器可以进行图像方面的训练，或同时进行图像和图说方面的训练。
I tried using ChatGPT, but it says it’s ‘at capacity.’ What does that mean?
我试着使用ChatGPT，但它说自己在“满负荷运行”。这是什么意思？
The message “at capacity” means OpenAI’s servers are overwhelmed with requests and can’t handle new ones right now. Try again later. An OpenAI spokesperson said the company has been working to expand server capacity to meet the unexpected demand.
“满负荷运行”这一讯息意味着OpenAI对查询请求应接不暇，现在无法处理新请求。请稍后再试。OpenAI的一位发言人说，该公司已在努力扩大服务器容量，以满足这种超预期的需求。
How much does ChatGPT cost?
ChatGPT收费多少？
ChatGPT is free. OpenAI released the chatbot as a research preview and users can try it through a dedicated website. On Feb. 1, OpenAI also launched a premium version for $20 a month, starting in the U.S., that will give subscribers priority access.
ChatGPT是免费的。OpenAI把该聊天机器人作为一个研究预览版本发布了出来，用户可以通过一个专门的网站来试用。此外，2月1日，OpenAI先在美国推出了一个高级版本，月费为20美元，该版本将为订户提供优先访问权。
Both Microsoft and OpenAI plan to release an API, or application programming interface, allowing companies to integrate the technology into their products or back-end solutions. Microsoft’s API will be available through its Azure cloud-computing platform. Both companies already offer OpenAI’s earlier AI technologies.
微软和OpenAI都计划发布一个API（应用程序编程接口），允许各公司将该技术整合到自家产品或后端解决方案中。微软的API将通过旗下Azure云计算平台提供。这两家公司都已经提供OpenAI早期的人工智能技术。
How are people and businesses using ChatGPT?
个人和企业如何使用ChatGPT？
Media companies including BuzzFeed and the publisher of Sports Illustrated have announced plans to generate content such as quizzes and articles with ChatGPT. Some schools have blocked access to the service on their networks to stave off cheating, while others are actively encouraging students to use the tools ethically.
包括BuzzFeed和《运动画刊》（Sports Illustrated）出版商在内的媒体公司已经宣布，计划用ChatGPT生成趣味问答和文章等内容。一些学校已经在校内网上屏蔽了ChatGPT，以防有人作弊，而另一些学校则积极鼓励学生以合乎道德的方式使用这些工具。
Keep in mind that OpenAI has access to your inputs and outputs for ChatGPT and its employees and contractors may read them as part of improving the service. Avoid providing private data or sensitive company information.
应注意的是，为优化ChatGPT，OpenAI可以访问您因为使用该服务而产生的输入和输出信息，OpenAI的员工和承包商也可能会阅读这些信息。请不要输入隐私数据或敏感的公司信息。
Other generative AI technologies such as Dall-E 2 and avatar-generator Lensa have become popular with internet users for producing fantastical images and illustrations, and are finding use among independent writers to create artwork for their articles.
包括Dall-E 2和自拍生成器Lensa在内的其他生成式AI技术已在互联网用户中流行起来，可用于制作奇幻的图像和插图，独立作家也在运用这些技术为他们的文章创作插图。
What are the pitfalls of AI chatbots?
AI聊天机器人有何缺点？
AI chatbots and other generative AI programs are mirrors to the data they consume. They regurgitate and remix what they are fed to both great effect and great failure. Transformer-based AI program failures are particularly difficult to predict and control because the programs rely on such vast quantities of data that it is almost impossible for the developers to grasp what that data contains.
AI聊天机器人和其他生成式AI程序能够像镜子一样反映“投喂给”它们的数据。它们对这些数据进行“反刍”并将其重新混合，有时结果令人惊艳，有时则一塌糊涂。基于转换器的AI程序的故障尤其难以预测和控制，因为这些程序依赖于海量数据，开发人员几乎不可能掌握这些数据包含的信息。
ChatGPT, for example, will sometimes answer prompts correctly on topics where it ingested high-quality sources and frequently conversed with its human trainers. It will spew nonsense on topics that contain a lot of misinformation on the internet, such as conspiracy theories, and in non-English languages, such as Chinese.
例如，ChatGPT有时能正确回答一些问题，因为对于这类话题“摄入”了高质量的信息来源，并经常与人类训练员对话。而对于互联网上含有大量错误信息的话题，比如阴谋论，以及非英语语言，比如中文，它就会胡言乱语。
Some artists have also said that AI image generators plagiarize their artwork and threaten their livelihoods, while software engineers have said that code generators rip large chunks of their code.
一些艺术家也表示，AI图像生成器会剽窃他们的艺术作品，并威胁到他们的生计，而软件工程师则表示，代码生成器会抄袭他们的大段代码。
For the same reasons, ChatGPT and other text generators can spit out racist and sexist outputs. OpenAI says it uses humans to continually refine the chatbot’s outputs to limit these mishaps. It also uses content-moderation filters to restrict ChatGPT’s responses and avoid politically controversial or unsavory topics.
出于同样的原因，ChatGPT和其他文本生成器可能输出带有种族主义和性别歧视的内容。OpenAI称其使用人工来不断完善ChatGPT的输出，以减少这类错误的发生。OpenAI还使用内容审核过滤器来限制ChatGPT的回答，并避开政治上有争议或令人不快的话题。
Ridding the underlying technology of bias – which has for years been a recurring problem, including for an infamous Microsoft chatbot in 2016 known as Tay – remains an unsolved problem and a hot area of research.
消除底层技术偏见仍然是一个未解决的问题，也是一个热门的研究领域。多年来，底层技术偏见是一个反复出现的问题，包括2016年被称为Tay的臭名昭著的微软聊天机器人。
“ChatGPT is incredibly limited, but good enough at some things to create a misleading impression of greatness,” tweeted OpenAI Chief Executive Sam Altman shortly after the chatbot’s release, adding that it is “a mistake to be relying on it for anything important right now.”
“ChatGPT存在令人难以置信的局限性，但在某些方面的能力足以让人误以为它是伟大的，”OpenAI首席执行官Sam Altman在该聊天机器人发布后不久发推文称，“现在依靠它来做任何重要的事情都是错误的。”
What is Microsoft’s relationship to OpenAI?
微软与OpenAI是什么关系？
Microsoft is OpenAI’s largest investor and exclusively licenses its technologies. The tech giant invested $1 billion into the AI startup in 2019, an undisclosed amount in 2021 and an additional amount of up to $10 billion in January, according to people familiar with the latest deal. Under the agreement, Microsoft can use OpenAI’s research advancements, including GPT-3 and ChatGPT, to create new or enhance existing products. It is the only company outside of OpenAI that can provide an API for these technologies.
这家科技巨头是OpenAI最大的投资者，并独家授权其技术。微软2019年向这家AI初创公司投资了10亿美元，2021年再度投资，金额未披露，据知情人士透露，今年1月又追加了不超过100亿美元投资。根据协议，微软可以使用OpenAI的研究进展，包括GPT-3和ChatGPT，来创建新的产品或增强现有产品。微软是OpenAI之外唯一一家可以为这些技术提供API的公司。
Is AI going to replace jobs?
AI会抢人类饭碗吗？
As with every wave of automation technologies, the latest will likely have a significant impact on jobs and the future of work. Whereas blue-collar workers bore the brunt of earlier waves, generative AI will likely have a greater effect on white-collar professions. A 2019 study from the Brookings Institution found that AI would most affect jobs such as marketing specialists, financial advisers and computer programmers.
与每一波自动化技术一样，最新的技术可能会对当下的岗位和未来的就业产生重大影响。在此前的几波浪潮中，蓝领工人往往首当其冲，而生成性AI将可能对白领职业产生较大影响。布鲁金斯学会（Brookings Institution） 2019年的一项研究发现，AI对营销专家、财务顾问和计算机程序员等工作影响最大。
Those effects will be mixed. Economists who study automation have found that three things tend to happen: some workers improve their productivity, some jobs are automated or consolidated, and new jobs that didn’t previously exist are also created.
这些影响将有利有弊。研究自动化的经济学家已经发现，相关影响往往包含三个方面：一些劳动者提高了生产力；部分岗位实现自动化或被合并；产生出以前不存在的新岗位。
The final scorecard is difficult to predict. In company-level studies of automation, researchers have found that some companies that adopt automation may increase their productivity and ultimately hire more workers over time. But those workers can experience wage deflation and fewer career growth opportunities.
最终结果很难预测。在公司层面的自动化研究中，研究人员发现，一些采用自动化的公司可能会提高生产力，并最终在一段时间内雇用更多的工人。但这些工人可能会经历工资缩水，职业发展机会减少。
Newly created jobs often go one of two ways: they either require more skill, or a lot less, than the work that was automated. Self-driving cars, for example, create new demand for highly skilled engineers but also for low-skilled safety drivers, who sit in the driver’s seat to babysit the vehicle.
新创造的岗位通常有两种情况：与被自动化的岗位相比，要么需要更多技能，要么所需的技能少得多。例如，自动驾驶汽车创造了对高技能工程师的新需求，但也创造了对不怎么需要技能的安全司机的新需求，这些司机只需坐在驾驶座上照看车辆。
