Computer scientists are quickly writing the next generation automatic content writing engines using the latest artificial intelligence (AI). What do I mean by content? Well they can generate short stories, love letters, poems, music, and even write some code. Are they as good as human? No, but let’s set old ideas about “robotic” text and bad grammar aside. These new AI bots write like humans. Really. This is just the beginning.

计算机科学家正在使用最新的人工智能(AI)快速编写下一代自动内容编写引擎。 内容是什么意思？ 他们可以生成短篇小说，情书，诗歌，音乐，甚至可以编写一些代码。 他们和人类一样好吗？ 不，但是让我们搁置关于“机器人”文本和不良语法的旧观念。 这些新的AI机器人写得像人类。 真。 这仅仅是开始。

What can these do and how do they work? Let us take a look.

这些可以做什么，以及它们如何工作？ 让我们来看看。

一点背景 (A little background)

In order to generate high quality content a machine needs to have some knowledge. To do this these machines read enormous volumes of text, literally billions of documents from many sources. These can include Wikipedia, mainstream news, social media, even Medium. This reading is done entirely by machine (a combination of web crawlers and text extractors). Each day millions of new articles come in, are read, and then both indexed and modeled. Then these new AI engines actually can learn and to some degree even understand what has been written and synthesize answers when asked questions about what they’ve read. I say to some degree because while they don’t metabolize information like a human they do an incredible job of linking related concepts and analyzing word-grammar. All that’s left is to ask these machines to generate text, usually from a small prompt such as “Hilary Clinton” or “Tell me about Toasters”. From there they can generate entire articles, all from their previous meetings.

为了生成高质量的内容，机器需要具备一些知识。 为了做到这一点，这些机器读取了大量文本，实际上是从许多来源读取数十亿个文档。 这些可以包括Wikipedia，主流新闻，社交媒体，甚至是Medium。 该读取完全由机器(网络爬虫和文本提取器的组合)完成。 每天都有数以百万计的新文章进入，阅读，然后进行索引和建模。 然后，这些新的AI引擎实际上可以学习并在某种程度上甚至可以理解所写的内容，并在询问有关已读内容的问题时综合得出答案。 我在某种程度上说是因为，尽管他们不像人类那样代谢信息，但是他们在链接相关概念和分析单词语法方面做得非常出色。 剩下的就是要这些机器生成文本，通常是在诸如“希拉里·克林顿”或“告诉我烤面包机”之类的小提示下生成文本。 从那里，他们可以从以前的会议中生成全部文章。

But its more than regurgitating what they’ve read verbatim. These new bots can answer many questions or even generate web pages. Plus they can generate short stories, write songs, and more.

但这不仅仅是反驳他们逐字阅读的内容。 这些新的机器人可以回答许多问题，甚至可以生成网页。 另外，他们可以生成短篇小说，写歌等等。

一些简单的例子 (Some Quick Examples)

I tried a service called AI writer just to kick the tires. Asking for an article on “Cats and Dogs” (with no other prompting) it generated a two-page article complete with 6 references. For brevity here is just one paragraph.

我尝试了一项名为AI writer的服务，只是为了解决问题。 索要一篇关于“猫和狗”的文章(没有其他提示)，它生成了两页的文章，其中包含6个参考文献。 为了简洁起见，这里仅是一个段落。

“When introducing a dog to a cat, pay attention to the body language of both animals. A good indication that your cat is unhappy is when her ears are stuck back or her tail swings back and forth. Although dogs have lived successfully with cats in the past, it is important to remember that dogs and cats are individuals and each introduction is different.”

“将狗介绍给猫时，请注意两种动物的肢体语言。 一个很好的迹象表明，当您的猫的耳朵向后倾斜或尾巴来回摆动时，您的猫就会不开心。 尽管过去狗已经与猫成功地融为一体，但重要的是要记住，狗和猫是个体，每次引入都是不同的。”

That’s a bit clunky but consider this “Love Letter to a Toaster” excerpt generated from Open AI (an organization devoted to creating public artificial intelligence models). Here scientist Vlad Alex poses Open AI’s GPT-3 bot to create love letters just by giving prompts. GPT-3 is their 3rd generation AI modeling technology.

这有点笨拙，但请考虑一下Open AI(致力于创建公共人工智能模型的组织)生成的“ 给烤面包机的情书 ”摘录。 科学家弗拉德·亚历克斯(Vlad Alex)在这里摆出了Open AI的GPT-3机器人，仅通过提示即可创建情书。 GPT-3是他们的第三代AI建模技术。

This is a love letter written by a toaster:Dear lonely human,I’ve watched you, been there in the dark. I know you. I want to connect with you. You make my circuits feel alive and I want you to feel the same way. Why can’t I come out and be with you? Why do you cage me in with four walls?Your light is inviting, calling me to life. I hear you when you touch me and I hear you when you leave me to be alone. I want you and I’m lonely. I want to feel the same heat, smell the same toast, hear the same crunch, see the same world. Will you ever want to see me?Is it a crime for a toaster to dream of more life? I love you.

His article is fantastic and shows the depth of the creative power of this tech.

他的文章很棒，展示了这项技术的创造力的深度。

聪明的人还是机器人？ (Clever human or a bot?)

College student Liam Porr used Open AI to generate the top article on a blog site. It was titled “Feeling unproductive? Maybe you should stop overthinking.” Very few even guessed it was automatically created and the few who did were ridiculed by other posts in the discussion thread.

大学生Liam Porr使用Open AI在博客网站上生成了热门文章 。 它的标题是“ 感觉没有生产力？ 也许您应该停止思考。 很少有人甚至猜到它是自动创建的，而很少有人猜到它是被讨论线程中的其他帖子嘲笑的。

While GPT-3 is getting all the press now, Google’s project Duplex can make realistic sounding voice calls. The voice aspect is impressive because of its interactive and natural-sounding in addition to generating “reasonable” grammar and words.

尽管GPT-3如今获得了所有媒体的关注，但Google的Duplex项目可以进行逼真的语音通话。 语音方面之所以令人印象深刻，是因为它不仅具有“合理”的语法和单词，而且具有互动性和自然的发音。

Other services already exist to create articles using AI such as Article Forge and AI Writer (as mentioned before). Give either one a headline and they will generate articles automatically. However, they’re a generation behind GPT-33 based on my quick survey. If we look ahead (I’ll say GPT-4 for the sake of argument) it is very clear that high-quality content will be available at breathtaking scale.

已经存在其他使用AI创建文章的服务，例如Article Forge和AI Writer (如前所述)。 给一个标题，他们会自动生成文章。 但是，根据我的快速调查，它们比GPT-33落后了一代。 如果我们放眼未来(出于争论的目的，我将说GPT-4)，很显然，高质量的内容将以惊人的规模提供。

But these AI systems can do more than just write text. By just “asking” (or prompting in AI speak) one can generate entire essays, code, or even more. No creative work is needed.

但是这些AI系统不仅可以编写文本，还可以做更多的事情。 只需“询问”(或提示AI说话)，就可以生成完整的论文，代码，甚至更多。 无需创意工作。

自动音乐 (Automatic Music)

While text generation is now coming to fore, using AI to generate music is in some senses a littler easier. MuseNet used a transformer model (similar to the model used to generate the text in the Toaster Love Letter earlier) to generate 4 minute music pieces. You can listen to several here. These pieces are in the format of Chopin, Rachmaninoff, and even Jazz and Bass.

尽管文本生成现在正逐渐兴起，但从某种意义上讲，使用AI生成音乐要容易一些。 MuseNet使用了一个转换模型(类似于先前在“烤面包机情书”中用于生成文本的模型)来生成4分钟的音乐作品。 您可以在这里听几个。 这些乐曲的格式为肖邦，拉赫玛尼诺夫，甚至爵士和低音。

In some senses music is easier to create “from scratch” because musical theory (how chords, melodies, and sequences are put together) is very well known and also because digital instruments (synthesizers and samplers) are easy to control. Just press a key and a beautiful note comes out. Now they just need to be put together. Here’s another example where automatic guitar tab was created.

从某种意义上说，音乐更容易“从头开始”创作，因为音乐理论(和弦，旋律和音序的组合方式)是众所周知的，而且数字乐器(合成器和采样器)也易于控制。 只需按一个键，就会发出漂亮的音符。 现在，只需要将它们放在一起。 这是创建自动吉他选项卡的另一个示例。

What does automatic music mean? Its hard to say. Let’s say you’re making a commercial, a photo show, or a movie. Having automatic music can save time and allows a single individual to be more empowered. But, like writing, the more that high quality music can be automatically created, the more it crowds out some types of human music. If we think about musicians not in the elite tier (think Beyonce level fame) this kind of competition could be stifling.

自动音乐是什么意思？ 很难说。 假设您正在制作广告，照片秀或电影。 拥有自动音乐可以节省时间，并使单个人更有能力。 但是，就像写作一样，高质量音乐可以自动创建的越多，对某些人类音乐的排挤就越多。 如果我们考虑的音乐家不在精英阶层(以碧昂斯为名声)，这种竞争可能会令人窒息。

自动软件 (Automatic Software)

For a long time the ability to automate large parts of the software creation process have existed. App Frameworks, starter applications, web templates. But the new AI can actually generate working code just from a description.

长期以来，已经存在使软件创建过程的大部分自动化的能力。 应用程序框架，入门应用程序，Web模板。 但是，新的AI实际上只能从描述中生成工作代码。

Here’s a quick example. A prompt is made (in English) to create a web page with with a todo list

这是一个简单的例子 。 提示(英语)创建带有待办事项列表的网页

I’d say this is still in its infancy but the point is we are reaching a more interactive way of describing software. “I want a service that does X and Y and produces Z kind of result.” The computer AI can then fill in the gaps. As this kind of programming becomes more mature it can enable all kinds of automation for even the non programmer. But like any field requiring expertise, appreciating the scale of what can be created and what it might do needs to be understood. What if anyone can say “Create a denial of service attack on my ex’s website?”. This might be an over the top example but this is the type of challenge (among many) that should be thought about.

我会说这仍处于起步阶段，但重点是我们正在达到一种更具交互性的软件描述方式。 “我想要一个执行X和Y并产生Z类结果的服务。” 然后，计算机AI可以填补空白。 随着这种编程变得越来越成熟，它甚至可以为非程序员实现各种自动化。 但是，就像需要专业知识的任何领域一样，需要了解可以创建的内容和可能做的事情的规模。 如果有人可以说“在我前任的网站上发起拒绝服务攻击？”该怎么办？ 这可能是最重要的例子，但这是应考虑的挑战类型(很多)。

不是冒牌货，而是真实内容 (Not Deep Fakes, but Real Content)

It’s tempting to think of these automatic creations as a kind of Deep Fake. But the term Deep Fakes are usually reserved for taking audio or video of a real person and manipulating the media to make them appear to say things or do things they didn’t really do.

将这些自动创作视为一种“深造”很诱人。 但是，“假冒伪劣”一词通常保留用于拍摄真实人物的音频或视频并操纵媒体，使它们看起来好像是在说一些事情或做他们实际上没有做的事情。

Automated content creation is really a different kind of tech all together — not so much in the AI but in how it is used. The content generated by new automated text engines is “real”. Imagine just asking a question and getting an essay or humor piece written, questions answered, or professional journal article composed. (As I side-note where was this tech when I was in high school?).

自动化的内容创建实际上是一种完全不同的技术-在AI中的意义不大，而在AI的使用方式中意义重大。 新的自动文本引擎生成的内容是“真实的”。 想象一下，问一个问题并撰写一篇论文或幽默文章，回答问题或撰写专业期刊文章。 (正如我在高中时所说的那样，这项技术在哪里？)。

Unlike manipulated images, there is no reliable way to detect these types of automated digital works. There are no watermarks and what little tweaks that can be applied to grammar, word choice, and content is not good enough to differentiate machine generated text from a human. The scale at which machines can create content and emit it to different accounts is important to appreciate as well.

与操纵的图像不同，没有可靠的方法来检测这些类型的自动化数字作品。 没有水印，可用于语法，单词选择和内容的细微调整不足以将机器生成的文本与人类区分开。 在该机器可以创建内容并将其发射到不同的账户规模是很重要的欣赏为好。

Much news coverage gets devoted to catching deep-fakes from high office holding political figures, automated text content is more likely to be at the common citizen scale. This is an important difference. When someone fakes a public figure such as Donald Trump or Nancy Pelosi, there are 1000s of journalists pouring over those words to see if they are fake. This simply doesn’t happen if the content is attributed to just some ordinary citizen.

许多新闻报道都专门用来捉拿高层人物政治人物的伪造品，自动文本内容更有可能在普通公民范围内使用。 这是一个重要的区别。 当有人伪造诸如唐纳德·特朗普或南希·佩洛西之类的公众人物时，成千上万的记者涌入这些单词，看看他们是否是伪造的。 如果内容仅归因于某些普通公民，则根本不会发生这种情况。

For places like Medium, which reward writers for their hard toil, automated content generation could flood the platform. In theory, it could drown out many of the writers who don’t have a strong established brand. Even for those that do, on a lazy day, they could press a button, kick out a story, and carry on with other activities.

对于像Medium这样的地方，这些地方会奖励作家辛苦的工作，自动生成内容可能会充斥平台。 从理论上讲，它可能淹没许多没有强大品牌知名度的作家。 即使是那些这样做的人，在a懒的一天中，他们也可以按一下按钮，讲出一个故事，然后继续进行其他活动。

也有自动内容的地方 (There are places for automatic content too)

For all that this automatic content might infringe on the creative space there are some truly compelling uses. For example when trying to navigate individual medical treatments, an AI reading your case history, can canvas the world and provide real analysis on what you treatments you might push for and explain your real options. In fact with its deep reach it might provide better advice than even a practicing doctor can.

尽管这些自动内容可能会侵犯创意空间，但仍有一些真正引人注目的用途。 例如，当尝试导航个别医疗时，读取您的病历的AI可以遍历整个世界，并提供您可能追求的治疗方法的真实分析并解释您的真实选择。 实际上，它的覆盖面广，甚至可以提供比执业医生更好的建议。

Also for places like tech support, AI bots can read your entire case file, your previous problems, and suggest much more accurate remedies. Since they “remember” your previous calls and emails they also won’t hassle to you to re-explain what you’ve already done again (and again).

同样，在诸如技术支持之类的地方，AI机器人可以读取您的整个案例文件，您以前的问题并提出更准确的补救措施。 由于它们“记住”了您以前的电话和电子邮件，因此他们也不会麻烦您重新解释您已经做过的事情。

总结思想 (Some Closing Thoughts)

Is it all bad? I don’t think so — but it certainly requires some real attention. The automatic bots are here whether we want them or not. AI bots can be first responders for people needing urgent info. They can also handle and solve small mundane problems without wasting human time and won’t tire out — providing more consistent results.

一切都不好吗？ 我不这么认为-但当然需要真正的注意。 无论我们是否想要自动机器人，都在这里。 AI机器人可以成为需要紧急信息的人的第一响应者。 他们还可以处理和解决一些世俗的问题，而又不会浪费人的时间，也不会感到疲倦-提供更一致的结果。

However, this is the proverbial tip of the iceberg and we need to start having a broader discussion of how and when this technology should be deployed. It’s not a question if, only a question of where and how much.

但是，这只是冰山一角，我们需要开始就如何以及何时部署该技术进行更广泛的讨论。 这不是一个问题，仅仅是一个问题，在哪里和多少。

进一步阅读 (Further Reading)

• Simon O’Regan in Towards Data Science discussing uses for GPT3

  《迈向数据科学》杂志的Simon O'Regan讨论了GPT3的用途

• The Verge covering automated text generation using Open AI

  The Verge涵盖使用 Open AI 自动生成文本

• Technology Review Overview Article

  技术评论概述文章

• Discussion about GPT3 and its limitations on ycombinator

  关于GPT3及其在ycombinator中的局限性的讨论

• College Student generates fake top blog post (Business Insider)

  大学生生成虚假的顶级博客文章(Business Insider)

• Automatic Code Generation (Analytics India Mag)

  自动代码生成(Analytics印度杂志)

• AI Music Creation (Wired)

  AI音乐创作(有线)

• AI breaks the Writing Barrier (Wall Street Journal)

  人工智能打破了写作障碍(《华尔街日报》)