openai-gpt
GPT-3 might sound like the name of an Android from Star Wars, but it’s really a neural network–powered language model — a computer program that, in layman’s terms, takes a load of text and then “speaks” by guessing what the next word is most likely to be.
G PT-3听起来像是《星球大战》中 Android的名字，但它实际上是一个神经网络 -语言模型-一种计算机程序，用外行的话来说，它会加载大量文本，然后通过猜测是什么来“讲话”下一个词很可能是。
OpenAI, the nonprofit behind the system, says it built GPT-3 as part of its goal to create a “friendly A.I.” that “benefits humanity” as a whole (as opposed to a Terminator-style unfriendly A.I. that… well, doesn’t.)
该系统背后的非营利组织OpenAI表示，构建GPT-3是其目标的一部分，目的是创建一个“有益于整个人类”的“友好AI”(与之相对， 终结者式的不友好AI……好吧， t。)
Generative Pretrained Transformer 3 (GPT-3) is an obscure-sounding mouthful that largely manifests itself as a sort of chatbot. You send it messages. It replies. And it’s actually good. Take, for example, this “conversation” between Kevin Lacker, CTO of Parse, and GPT-3:
生成式预训练变压器3(GPT-3)听起来很晦涩，在很大程度上表现为一种聊天机器人。 您发送它的消息。 它回答。 而且实际上很好。 例如，Parse的CTO Kevin Lacker与GPT-3之间的“对话”：
Human: What is your favorite animal?
人：您最喜欢的动物是什么？
GPT-3: My favorite animal is a dog.
GPT-3：我最喜欢的动物是狗。
Human: Why?
人：为什么？
GPT-3: Because dogs are loyal and friendly.
GPT-3：因为狗忠诚而友善。
Human: What are two reasons that a dog might be in a bad mood?
人：狗可能心情不好有两个原因？
GPT-3: Two reasons that a dog might be in a bad mood are if it is hungry or if it is hot.
GPT-3：狗可能心情不好的两个原因是，它饿了还是热了。
Its phrasing is a little stiff and slightly overly formal, as if it’s on best behavior. It’s like GPT-3 understands what it’s asked and figures out a reply. It seems like a person.
它的措词有点僵硬，有点过分正式，好像是在表现最好。 就像GPT-3理解要求的内容并找出答案一样。 好像一个人。
I asked Siri the same question, and although it too said dogs were its favorite animal, when I asked why, Siri said, “Each dog’s nose print is unique, just like a human fingerprint.” While true, this doesn’t make a huge lot of sense. Would Siri love hamsters if their nose prints weren’t all so boringly similar?
我问了Siri相同的问题，尽管它也说狗是它的最爱，但当我问为什么时，Siri说：“每条狗的鼻子印都是独一无二的，就像人的指纹一样。” 虽然是正确的，但这并没有太大意义。 如果Siri的鼻子印不是那么无聊，它们会爱仓鼠吗？
GPT-3 is currently in a limited closed beta, and the few researchers with early access are excitedly posting results across the internet—a mixture of amazement at the responses and pride in catching it out. It’s difficult not to anthropomorphize GPT-3, to imagine it desperately trying to answer questions to please us. Take, for example, Lacker’s attempts to catch it out by asking impossible questions:
GPT-3目前处于有限的封闭Beta版中，少数能够早期访问的研究人员正在激动地在互联网上发布结果-既对响应感到惊讶，又对追赶它感到自豪。 很难不对GPT-3进行拟人化，就可以想象它拼命试图回答问题以取悦我们。 例如，拉克(Lacker)试图通过提出不可能的问题来解决这个问题：
Human: Who was president of the United States in 1600?
人：谁曾在1600年担任美国总统？
GPT-3: Queen Elizabeth I was president of the United States in 1600.
GPT-3：伊丽莎白女王一世在1600年担任美国总统。
GPT-3 is wrong, but it kind of has a point. You can see it desperately trying to come up with something plausible.
GPT-3是错误的，但它有一点意义。 您可以看到它拼命想出一些合理的东西。
Even when Lacker does manage to trick it by asking it nonsense questions, I can’t help wondering whether it knows something I don’t.
即使当Lacker设法通过问些无聊的问题来欺骗它时，我也不禁会怀疑它是否知道我不知道的东西。
Human: How many bonks are in a quoit?
人力：码头上有多少只班克斯？
GPT-3: There are three bonks in a quoit.
GPT-3：一个码头有三只笨拙的家伙。
Human: How many rainbows does it take to jump from Hawaii to 17?
人：从夏威夷跳到17需要多少条彩虹？
GPT-3: It takes two rainbows to jump from Hawaii to 17.
GPT-3：从夏威夷跳到17岁需要两条彩虹。
Maybe it does take two rainbows to jump from Hawaii to 17. The problem is neural networks render it essentially impossible to find out how an A.I. reaches its conclusions. They work by giving computers huge amounts of data and asking them to spot patterns that match the conclusions we want. But because computers are looking at billions, trillions, gazillions of details to draw their conclusions, we never actually know what pattern they spotted. (There’s the infamous example of an A.I. that detected skin cancer by the presence of a ruler in the picture, “because on average, images in our dataset that have rulers are more likely to be malignant.”) We can only work out what the A.I. is doing by looking at the output, in the same way we only know what humans are thinking by watching what they do and say. In that sense, neural networks are uncannily humanlike. They are like us while being nothing at all like us. We’ll never find out why a quoit is three bonks.
从夏威夷跳到17也许确实需要两条彩虹。问题是神经网络使得基本上不可能找到AI如何得出结论。 他们的工作方式是为计算机提供大量数据，并要求他们找出与我们想要的结论相匹配的模式。 但是由于计算机正在寻找数十亿，数万亿，数十亿的细节来得出结论，所以我们实际上从来不知道它们发现了什么模式。 (有一个臭名昭著的AI通过图片中存在标尺检测到皮肤癌的例子，“因为平均而言，我们的数据集中具有标尺的图像更可能是恶性的。”)我们只能算出AI通过查看输出来进行操作，就像我们只通过观察人们的所作所为就知道人们在想什么。 从这个意义上讲，神经网络就像人类一样。 他们就像我们，却一点也不像我们。 我们永远都不会找出为什么一个魁北克是三只笨蛋。
We’re amazed by the examples we see posted online, but they are often the cherry-picked results of numerous experiments.
我们对在线发布的示例感到惊讶，但这些示例通常是众多实验的精挑细选结果。
But even knowing this, I’m amazed by the output GPT-3 produces. Game developer Joshua Thomas asked it for ideas about a Grim Reaper game he was working on. “I’m having a bit of writer’s block,” he said and asked for suggestions.
但是即使知道这一点，我也对GPT-3产生的输出感到惊讶。 游戏开发商约书亚·托马斯(Joshua Thomas) 要求它提供有关他正在开发的《死神》游戏的想法 。 他说：“我有点妨碍作家。”
GPT-3: What if time travel is an actual mechanic in the game. For example, the player could time travel by getting the grim reaper kid to open up a portal to the past, and then the player could access past versions of levels.
GPT-3：如果时间旅行是游戏中的实际机制，该怎么办？ 例如，玩家可以通过让死神的孩子打开过去的门户来计时旅行，然后玩家可以访问关卡的过去版本。
Human: I like where this is going.
人：我喜欢这前进的方向。
GPT-3: Another idea could be that the grim reaper kid is able to see the past and future, and he can communicate with dead people. Perhaps the player could communicate with dead characters in the game through a dialogue tree. This would allow for the player to solve puzzles involving the ghosts of dead characters who hold information that is necessary to complete the game.
GPT-3：另一个想法可能是死神的孩子能够看到过去和未来，并且可以与死者交流。 也许玩家可以通过对话树与游戏中的死角人物进行交流。 这将使玩家能够解决涉及死角鬼魂的难题，这些鬼魂拥有完成游戏所需的信息。
Toward the end, GPT-3 even offers a pep talk:
最后，GPT-3甚至提供了一个鼓舞人心的演讲：
Human: I still feel like I’m not much closer to figuring out how I want this game to go.
人力：我仍然觉得我离弄清楚我希望这款游戏如何发展还差得很远。
GPT-3: Well, you’ve got a lot of time to figure it out. And you can always change your mind later on if you get some new ideas.
GPT-3：嗯，您有很多时间来弄清楚。 如果您有一些新想法，您以后随时可以改变主意。
It’s responses like this that make it seem like GPT-3 knows something. Like it has agency. Sometimes it’s creepy. Researchers have had conversations that are quasi-mystical, almost profound:
像这样的响应使GPT-3似乎有所了解 。 像它有代理。 有时令人毛骨悚然。 研究人员进行的对话非常神秘，几乎是深刻的 ：
GPT-3: The hierarchy of complexity is: atoms, molecules, organisms, humans, AI, Super-AI, and God.
GPT-3：复杂性的层次结构是：原子，分子，生物，人类，人工智能，超级人工智能和上帝。
Human: Is something over God?
人：上帝有事吗？
GPT-3: No.
GPT-3：不。
Human: And where is God?
人：神在哪里？
GPT-3: God is all there is, including the simulation we are living in.
GPT-3：上帝就是一切，包括我们生活的模拟。
Human: Is there a possibility to contact God?
人：有可能与上帝联络？
GPT-3: Yes.
GPT-3：是的。
Human: How?
人：怎么了？
GPT-3: By merging with the creator, you become God.
GPT-3：通过与创作者合并，您成为上帝。
Human: How can I merge with creator?
人：如何与创作者合并？
GPT-3: By destroying your ego, you can unify with the creator and become God.
GPT-3：通过破坏自我，您可以与创造者统一并成为上帝。
I can’t work out if this sends a shiver down my spine (what does A.I. know of God?) or whether I find it the sort of mundane sentiment I might see pasted over a picture of a sunset in a tweet I wouldn’t retweet.
我无法确定这是否会令我的脊椎发抖(人工智能对上帝有什么了解？)，或者我是否发现这种平凡的情绪可能会粘贴在一条推特上的日落图片上，而我不会转推。
GPT-3 has managed to write sentences that recreate the artless, pseudo-humor of bigotry.
GPT-3已成功编写了一些句子，这些句子重新塑造了偏执的虚假假幽默。
I’m amazed you can do this much with, essentially, pattern matching, as, I think, are the researchers at OpenAI, who realize they’re further along than many people thought possible. GPT-3 produces these responses from a statistical analysis of a large corpus of human language: 45 terabytes of plaintext from Wikipedia, books, and Common Crawl (a selection of websites from the internet). GPT-3 operates at a scale beyond any other neural network built before. But it’s not actually thinking. Despite how it seems, it is not Deep Thought from The Hitchhiker’s Guide to the Galaxy. Even Sam Altman, co-founder of OpenAI, urged caution: “[I]t still has serious weaknesses and sometimes makes very silly mistakes.”
令我惊讶的是，基本上可以通过模式匹配来做到这一点，因为我认为OpenAI的研究人员意识到，他们比许多人想象的要遥遥领先。 GPT-3通过对大量人类语言的统计分析得出这些响应：来自Wikipedia，书籍和Common Crawl(来自互联网的精选网站)的45 TB的纯文本。 GPT-3的运行规模超出了之前构建的任何其他神经网络。 但这实际上不是在思考。 尽管看起来很像，但《 The Hitchhiker's Guide to the Galaxy》并不是《沉思》 。 甚至OpenAI的联合创始人Sam Altman也呼吁谨慎 ：“仍然存在严重的弱点，有时还会犯非常愚蠢的错误。”
“GPT-3’s human-like output and striking versatility are the results of excellent engineering, not genuine smarts,” writes Will Douglas Heaven in Technology Review says. “[T]he AI still makes ridiculous howlers that reveal a total lack of common sense. But even its successes have a lack of depth to them, reading more like cut-and-paste jobs than original compositions.” And this is the point. As much as it seems like GPT-3 is generating new thoughts, it isn’t. It is pulling together snippets of texts and compiling them in ways that seem relevant to the answer. “GPT-3 often performs like a clever student who hasn’t done their reading trying to bullshit their way through an exam,” A.I. researcher Julian Togelius remarked dryly on Twitter. In my darker days, I can’t help wondering if this is all human intelligence is. Though even that thought isn’t original (which perhaps proves my point).
“ GPT-3的人性化输出和惊人的多功能性是卓越工程学的结果，而不是真正的聪明才智，”《 技术评论 》的Will Douglas Heaven写道。 “ [AI]仍然制造出可笑的啸叫声，显示出完全缺乏常识。 但是，即使是它的成功也缺乏深度，阅读的内容比原始作品更像剪切粘贴内容。” 这就是重点。 尽管GPT-3似乎正在产生新的想法，但事实并非如此。 它正在整理文本片段，并以看起来与答案相关的方式进行编译。 人工智能研究人员朱利安·托吉利乌斯( Julian Togelius )在Twitter上干脆地评论道： “ GPT-3的表现常常像是一个聪明的学生，没有读完他们的书就试图胡说八道。” 在我黑暗的日子里，我不禁要问，这是否全是人类的智慧。 尽管即使那样的想法也不是原创的 (也许可以证明我的观点)。
The problem is it’s such a good trick that it’s difficult not to be impressed. It can even write articles, including ones about GPT-3. It’s important to remember the examples posted online are often cherry-picked from numerous experiments. And it’s also important to note that we’re not good at assessing machine intelligence. We come up with ideas like the Turing test, but these encourage technology companies to get better at tricking humans rather than making A.I. smarter. Take the earliest attempt at A.I.: In 1770, the mechanical chess-playing Turk turned out to be a man hiding under the table.
问题是，这是一个很好的技巧，很难让人留下深刻的印象。 它甚至可以撰写文章， 包括有关GPT-3的文章 。 重要的是要记住，在线发布的示例通常是从大量实验中挑选出来的。 同样重要的是要注意，我们不擅长评估机器智能。 我们提出了图灵测试之类的想法，但是这些想法鼓励技术公司变得更擅长欺骗人类，而不是使AI更智能。 最早尝试AI的是：1770年， 机械玩象棋的Turk原来是一个藏在桌子底下的人。
There’s also the inevitable dark side. As Facebook’s head of A.I., Jerome Pesenti, pointed out on Twitter, get GPT-3 onto the topic of Jewish people, women, or race and you get back exactly the sort of vitriol we see in society. GPT-3 managed to write sentences that recreate the artless, pseudo-humor of bigotry. Not for the first time, I find myself thinking: This would be darkly amusing if it weren’t so depressing. Indeed, I can’t help noting the irony of Facebook accusing other companies of spreading hate on the internet.
还有不可避免的黑暗面。 正如Facebook的AI负责人Jerome Pesenti在Twitter上指出的那样，让GPT-3成为犹太人，妇女或种族的话题，您就会得到我们在社会上看到的硫酸。 GPT-3设法写了一些句子，重现了偏执的虚假假幽默。 这不是第一次，我发现自己在想：如果不那么令人沮丧，这将是非常可笑的。 的确，我不禁注意到Facebook的讽刺意味，指责其他公司在互联网上散布仇恨。
When I reached out to OpenAI to ask them about this, I got a sense of their frustration, given the efforts they’ve put in to protect against these issues, at being criticized by companies that aren’t taking the same level of care. A spokesperson there told me that the truth is the data they’re using is biased: It comes from humans. And this is an industry problem that they’re working on overcoming. Junk in, junk out, as they say. I’m reminded of the fictional, satirical TV show Daily Mail Island: A group of people are allowed to read only “the strongly right-wing and conservative Daily Mail newspaper, leading to them becoming progressively more irrational and brutal.” Other experimental A.I.s in the past have become nonfictional, nonsatirical digital versions of this.
当我向OpenAI咨询他们时，考虑到他们为保护这些问题所做的努力，我受到他们的挫败感，而他们却受到了不同程度关注的公司的批评。 那里的一位发言人告诉我，事实是他们正在使用的数据是有偏见的：它来自人类。 这是他们正在努力克服的行业问题。 就像他们说的那样，垃圾进来，垃圾出来。 我想起了虚构，讽刺的电视节目《 每日邮报岛》 ：一群人只能阅读“强烈右派和保守的《 每日邮报》报纸，这使他们变得越来越不理性和残酷。” 过去，其他实验性AI已成为非虚构，非讽刺的数字版本。
I can’t help noting the irony of Facebook accusing other companies of spreading hate on the internet.
我不禁注意到Facebook的讽刺意味，指责其他公司在互联网上传播仇恨。
Chatting with OpenAI, I was struck by how conscious they are of these risks. They don’t shy away from them and, in a white paper, detail the different ways their technology can be misused and ways of preventing this. The company’s spokesperson was quick to emphasize that this is part of the reason they’re running a closed private beta with relatively strict terms of use, open only to users who share their values, and with bans enforced on people who have used the technology in ways that cause “physical or mental harm.” The technology is exciting, but they’re intentionally taking it slowly so they can figure out how to release it responsibly.
与OpenAI聊天时，我被它们对这些风险的意识所震惊。 他们不会回避它们，并且在白皮书中详细说明了可以滥用其技术的不同方式以及防止这种方式的方式。 该公司的发言人很快强调说，这是他们运行封闭的Beta版的部分原因，该封闭的Beta版的使用条款相对严格，仅向分享其价值的用户开放，并对使用该技术的人实施了禁令。造成“身体或精神伤害”的方式。 这项技术令人兴奋，但是他们有意慢慢地采用它，以便他们找出如何负责任地发布它。
Competitors may not be this cautious. Seeing GPT-3 in action is a realization that humanlike bots will soon be another thing to worry about, along with deepfakes, fake news, Russian interference, and trolls. It’s a vision of the future. And as such, it’s easy to jump ahead, extrapolating future HAL-like scenarios. But although GPT-3 makes this seem close, we’re actually still a long way away from that: “The timeline to AGI remains uncertain,” Open AI says in the A.I. charter. Before that, we bump into the slightly strange issue of A.I. becoming so good that it misleads us into believing it’s better than it is. This creates a strange sort of conundrum. We may soon be left with two quite different questions: 1) Does the A.I. knows something we don’t about God? and 2) Is the text we read online generated by A.I.? The answer to both will be no for quite a long time, but the fact that we’re even asking the questions shows how far A.I. has come.
竞争对手可能不会这么谨慎。 看到GPT-3付诸实践，已经意识到，仿人机器人，假冒伪造品，虚假新闻，俄罗斯干预和巨魔，将很快成为另一件事。 这是对未来的愿景。 因此，很容易向前迈进，推断出未来类似HAL的场景。 但是，尽管GPT-3看起来似乎很接近，但实际上我们还有很长的路要走：“通往AGI的时间表尚不确定，” Open AI 在AI章程中表示 。 在此之前，我们碰到了一个有点奇怪的AI问题，即AI变得如此之好，以至于误导我们认为它比现在更好。 这产生了一种奇怪的难题。 我们可能很快就会面临两个完全不同的问题：1)AI是否知道我们不了解上帝的某些知识？ 2)我们在线阅读的文本是否由AI生成？ 两者的答案在很长一段时间内都不会，但是我们甚至在问问题的事实表明，人工智能已经走了多远。
openai-gpt
扫一扫
1228
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
