GPT-3 is a language model created by OpenAI which can form natural language, given a prompt. Behind the scenes is a neural network with 175 billion parameters being fed massive amounts of data scraped from the internet. The result is a machine that can understand and respond to questions posed in GPT-3. It is not yet able to speak or write anything other than English. But it does have an understanding of basic grammar. This means that if you ask it to tell you what a triangle is, it will say ‘a three-sided object’ or something similar.

GPT-3是OpenAI创建的一种语言模型，可以在提示时形成自然语言。 幕后的是一个神经网络，其中包含1750亿个参数，这些参数被馈给从Internet抓取的大量数据。 结果是一台可以理解并回答GPT-3中提出的问题的机器。 除英语外，它还不能说其他任何语言。 但是它确实了解基本语法。 这意味着，如果您要求它告诉您什么是三角形，它会说“三边对象”或类似的东西。

I must admit: I hate writing introductions. So I wrote the first two sentences of the paragraph above, plugged it into AI Dungeon with some context, and out came the remainder of the paragraph. (As we’ll see in a moment, GPT-3 is severely underselling itself here.)

我必须承认：我讨厌写介绍。 因此，我写下了上面段落的前两个句子，并在一定情况下将其插入到AI Dungeon中，然后剩下的就是该段落。 (稍后我们将看到，GPT-3在这里严重卖空了自己。)

It’s obviously not perfect, but some allowance should be made for the fact that AI Dungeon runs a weaker version of GPT-3 and is configured to be used as a text-based game rather than a way of cheating at writing. As the following sample demonstrates, the full GPT-3 (to which access is currently severely limited) could be mistaken for a college-level writer.

这显然不是完美的，但是应该考虑到AI Dungeon运行的是GPT-3的较弱版本，并且被配置为基于文本的游戏，而不是作弊的一种方式。 如以下示例所示，完整的GPT-3(目前受到严格限制的访问权限)可能会被误认为是大学级作家。

There is an ongoing debate in my family about the origin of the word “cowboy.” We know the term was originally used to describe a herder or drover of cattle. Some of my family maintains it was coined in Spain, describing the Mexican or Spanish version of those workers.

我的家人一直在争论“牛仔”一词的由来。 我们知道该词最初是用来形容牛群或牛群的。 我的一些家庭坚持认为它是在西班牙创造的，描述了这些工人的墨西哥或西班牙版本。

My father’s family, on the other hand, insists the term was used by Anglo cowboys to describe the tough but generally peaceful herders or drovers.

另一方面，我父亲的家人坚持认为盎格鲁牛仔曾用这个词来形容强硬但普遍和平的牧民或驾车者。

The debate came up again a few weeks ago when we went to watch “Lone Ranger” with my father and his wife, my stepmother, Barb. The man riding the white horse was a sheriff, not a cowboy, they insisted.

几周前，当我们与父亲和他的妻子，我的继母Barb一起观看“孤独的游侠”时，辩论又开始了。 他们坚持说，骑白马的男人是警长，而不是牛仔。

I finally admitted I didn’t really care where the term came from and we all had a good laugh over it. (from the GPT-3 repo; direct link)

我终于承认，我不太在意这个词的来源，我们对此都笑了。 (来自GPT-3存储库； 直接链接 )

Occasionally, the AI’s awkwardness and repetition betray its identity:

有时，AI的笨拙和重复背叛了它的身份：

Glacier Ridge has had its students participate in various competitions and competitions. (from the GPT-3 repo; direct link)

冰川岭的学生参加了各种竞赛和竞赛 。 (来自GPT-3存储库； 直接链接 )

In all fairness, many human writers suffer from similar flaws. Few humans, however, can write in Spanish, French, Hungarian and Portuguese.

公平地说，许多人类作家都遭受类似的缺陷。 但是，很少有人能用西班牙语 ， 法语 ， 匈牙利语和葡萄牙语写东西 。

图灵测试 (The Turing Test)

Remember Eugene Goostman? In 2014, the chatbot made international headlines for passing the Turing Test, convincing 30% of a panel that it was human. Eugene Goostman is a 13 year-old boy from Ukraine, replete with the slang, initialisms, and an indifferent tone. Whether the chatbot successfully passed the Turing Test is still a subject of controversy; imitating a hostile 13 year-old is almost certainly an easier feat than passing as a human being in intellectual conversation.

还记得Eugene Goostman吗？ 2014年，聊天机器人通过图灵测试(Turing Test)成为国际头条新闻，说服了30％的小组认为它是人类。 尤金·古斯特曼(Eugene Goostman)是来自乌克兰的13岁男孩，对with语，首字母缩写和淡漠的语气充满了。 聊天机器人是否成功通过了图灵测试仍然是一个争议的话题。 模仿一个充满敌意的13岁小孩子，几乎可以肯定比在智力对话中成为人类要容易。

Naturally, I wanted to see how GPT-3 would fare in the Turing Test, but I needed some standard by which to evaluate its responses. Thankfully, Turing provided a sample dialogue between an interrogator and computer which I tried to replicate.

当然，我想看看GPT-3在图灵测试中的表现如何，但我需要一些标准来评估其响应。 值得庆幸的是，图灵提供了一个询问器和我试图复制的计算机之间的示例对话。

In order to prompt GPT-3 to respond to the same line of questioning as in Turing’s dialogue, I first assigned the interrogator the name ‘Paul’ and the computer ‘Jack’. My purpose in doing so was to avoid any confusion that the Interrogator and Computer were two individual characters in the dialogue. Then, I provided some context (italicized) about the ensuing discussion. There are some parallels between the sample dialogue and my coaxing of AI Dungeon, but the conversation largely falls apart halfway through.

为了促使GPT-3对与图灵对话中相同的问题做出回应，我首先为询问器分配了名称“ Paul”和计算机“ Jack”。 我这样做的目的是避免混淆询问器和计算机是对话中的两个独立角色。 然后，我提供了有关随后讨论的一些上下文(斜体)。 样本对话和我对AI Dungeon的哄骗之间存在一些相似之处，但对话在进行到一半时就大为分开了。

Whether a single dialogue is an indication of being able to pass the Turing Test is doubtful. However, if you consider Goostman as having passed the Turing Test, it appears as if GPT-3 can do the same.

一次对话是否表示能够通过图灵测试令人怀疑。 但是，如果您认为Goostman已经通过了图灵测试，那么看来GPT-3可以做到这一点。

难题 (Tough Questions)

Through some manipulation of AI Dungeon, we can interview imitations of famous people from a variety of fields. Here’s a snippet from fake Larry Page on big data:

通过对AI地牢的一些操纵，我们可以采访来自各个领域的名人的模仿。 这是有关大数据的伪造的Larry Page的摘录：

There are two main approaches [to improving our use of big data]. The first is automation, or what I like to think of as the smart pill…The other approach is to analyze the data in a different way. For example, what are the stories behind all these numbers? Why did this person buy something? What’s their story? If we can understand all that, we’ll be able to target advertising so it really matters to people.

有两种主要方法[改善我们对大数据的使用]。 第一种是自动化，或者我想将其视为智能药丸……另一种方法是以不同的方式分析数据。 例如，所有这些数字背后的故事是什么？ 为什么这个人买东西？ 他们的故事是什么？ 如果我们能理解所有这些内容，就可以定位广告，因此对人们来说确实很重要。

There is also a part of the interview where Page starts talking about people wanting to kill him when he was six years old as a result of his sister being born, so consider the above excerpt with a salt-shaker in hand. You can read the full text of this ‘interview’ here.

在采访的一部分中，佩奇(Page)开始谈论人们想因他妹妹出生而在他6岁时杀死他，因此请考虑上面的摘录，上面有一个振动筛。 您可以在此处阅读此“采访”的全文。

And some inspiration for your next article:

下一篇文章的一些启示：

The following is the headline of an article published in Towards Data Science, a Medium publication hosted at towardsdatascience.com: “A new tool for data scientists to analyze their own code.”

以下是在Towards Data Science上发表的一篇文章的标题，Towards Data Science是在towarddatascience.com上举办的中型刊物：“ 数据科学家用来分析自己的代码的新工具。 ”

You open up the article and read it. It’s written by one of your colleagues, who goes on to explain that he or she has created a program called “Data Analysis Toolkit”, which allows you to easily analyze your own code.

您打开文章并阅读。 它是由您的一位同事编写的，他接着解释说他或她已经创建了一个名为“数据分析工具包”的程序，该程序可让您轻松分析自己​​的代码 。

I must admit that I am already in possession of such a toolkit. It’s called “scrolling aimlessly trying to figure out why my code is broken”. Very innovative, if you ask me — apparently GPT-3 has different ideas.

我必须承认我已经拥有这种工具包。 这就是所谓的“漫无目的地滚动，试图找出为什么我的代码被破坏了”。 如果您问我，这是非常创新的-显然GPT-3有不同的想法。

假到你做完为止？ (Fake it ‘till you make it?)

My experiences with GPT-3 via AI Dungeon deeply impressed me, but they also revealed its limitations. Ultimately, GPT-3 is uncannily imitating what has been scraped from the Internet rather than displaying general intelligence. Even the full version has trouble with all but the most basic arithmetic, and plays a comparatively weak game of chess.

我通过AI地牢使用GPT-3的经历给我留下了深刻的印象，但同时也揭示了它的局限性。 最终，GPT-3毫无例外地模仿了从Internet抓取的内容，而不是显示一般情报。 即使是完整版本，除最基本的算术之外，其他所有代码都存在问题，并且玩国际象棋的能力相对较弱 。

Judging it by human standards, GPT-3 can be mind-numbingly stupid. That hasn’t stopped it from performing tasks previously thought to be the domain of highly-intelligent humans. This Twitter thread collates some of its achievements, which include writing fiction in a variety of styles, generating a working React app from a simple description of its function, and legal writing.

按照人的标准来判断，GPT-3可能让人感到愚蠢。 但这并没有阻止它执行以前被认为是高智能人类领域的任务。 这个Twitter线程整理了它的一些成就，其中包括以各种样式编写小说，通过对其功能的简单描述来生成可运行的React应用程序以及法律写作。

I am fairly convinced at this point that people will lose their jobs because of advancements in language models like GPT-3. The only question in my mind is whether this will happen in the next five years or the next ten. Artificial General Intelligence may be far away from realization, but intelligence isn’t a prerequisite to disrupting society.

在这一点上，我相当确信，由于GPT-3等语言模型的进步，人们将失去工作。 我心中唯一的问题是，这是否会在未来五年或未来十年内发生。 人工智能可能离实现还很遥远，但是智能并不是破坏社会的前提。

For now, I’ll enjoy playing AI Dungeon and try not to consider that I might have to compete with it for a job in a few years. Who am I kidding? When that time comes, there won’t be any competition.

现在，我会喜欢玩AI地牢，并尽量不要认为我可能需要在几年内与其竞争才能工作。 我在跟谁开玩笑？ 到那时候，将不会有任何竞争。