open ai gpt
Immersive storytelling, customer-centricity, and smarter programming are just the beginning
身临其境的讲故事，以客户为中心和更智能的编程仅仅是开始
By Nicolás Rodríguez
尼古拉斯·罗德里格斯(NicolásRodríguez)
OpenAI, the San Francisco-based AI lab, just released the third iteration of its GPT (Generative Pre-trained Transformer) model, or GPT-3 for short. After investing around $4.6 million, the program has shaken up every corner of the Internet, generating a mix of excitement and trepidation. But what is GPT-3, exactly?
总部位于旧金山的AI实验室OpenAI刚刚发布了其GPT(通用预训练变压器)模型(简称GPT-3)的第三次迭代。 在投资了460万美元之后，该计划动摇了互联网的各个角落，令人兴奋不已。 但是到底什么是GPT-3？
Generally speaking, it’s a language model based on automated learning, predicting word(s) based on a certain entry or text, building on what it has learned from previous data. This learning process, called training, is focused on “feeding” the algorithm with as many examples as possible, to allow it to identify patterns in data. It can then estimate the most likely answer based on the rules it has inferred from the information and context it has ingested
一般而言，它是一种基于自动学习的语言模型，可以基于从以前的数据中学到的内容，根据特定的条目或文本来预测单词。 这种称为训练的学习过程专注于通过尽可能多的示例“馈送”算法，以使其能够识别数据中的模式。 然后，它可以根据从信息和上下文中推断出的规则来估计最可能的答案
GPT-3 has been trained with data from five sources: 60 percent comes from Common Crawl, a global search engine crawler; 22 percent from ebText2, a text collection extracted from different web pages and curated by people; and 18 percent from books and Wikipedia, with more than 6 million articles. Altogether, they gather around 500 billion tokens or word sequences, usually separated by a comma.
GPT-3接受了来自五个来源的数据的培训：60％来自全球搜索引擎爬虫Common Crawl； 来自ebText2的22％，ebText2是从不同网页提取并由人们策划的文本集合； 18％的书籍和维基百科，超过600万篇文章。 它们总共收集了约5000亿个令牌或单词序列，通常用逗号分隔。
With content from diverse sources including novels, blogs, and code lines in different programming languages, GPT-3 offers depth and complexity that is still being explored and developed.
GPT-3的内容来自小说，博客和不同编程语言的代码行等各种来源，其深度和复杂性仍在探索和开发中。
Understanding its origins
了解其起源
When thinking about AI’s transformative powers, you must first define intelligence. Most definitions agree it involves the ability to learn, analyze, and comprehend information in a way that can be understood as knowledge and be applied in adaptive behavior. In this sense, we humans have developed specialized“life skills”, primarily for decision-making problem-solving, creative and critical thinking, communications, and interpersonal skills. Can we say the same about Artificial Intelligence?
在考虑人工智能的变革能力时，您必须首先定义智能。 大多数定义都同意它涉及以一种可以理解为知识并应用于适应性行为的方式来学习，分析和理解信息的能力。 从这个意义上讲，我们人类已经开发出专门的“生活技能”，主要用于决策问题解决，创造性和批判性思维，沟通以及人际交往能力。 我们可以对人工智能说同样的话吗？
For a start, AI represents the ability of a machine to imitate intelligent human behavior. This means that behavior can look intelligent even if it doesn’t include comprehension, let alone knowledge. The key question: Is GPT-3 really intelligent, or just pretending to be intelligent?
首先，AI代表机器模仿人类智能行为的能力。 这意味着即使不包含理解力，行为也可以看起来很聪明，更不用说知识了。 关键问题：GPT-3真的是智能的，还是只是假装是智能的？
Historically, “robots” were based on traditional programming, and told what to do, when, and how. In due course, machines could walk, keep up a conversation, and classify products, with rudimentary behaviors limited by the rules or code created by humans. Increased sophistication and complexity were directly proportional to the massive human effort to codify such behavior.
从历史上看，“机器人”是基于传统编程的，并告诉他们做什么，何时以及如何做。 在适当的时候，机器可以行走，保持对话并对产品进行分类，其基本行为受到人类创建的规则或代码的限制。 复杂程度的提高与人类为编纂这种行为付出的巨大努力成正比。
Everything changed in the ’60s when American psychologist Frank Rosenblatt created the Perceptron, the first artificial neuron. This was a turning point for automated learning as a discipline whose goal is to develop learning among machines, making computers able to identify underlying rules or patterns in the information and act accordingly. Rosenblatt’s investigation was overlooked for almost 20 years, mainly because of unfounded criticism from MIT professors Marvin Minsky (an AI pioneer who won the Turing Prize in 1969) and Seymour Papert (a fellow computer scientist). They saw the Perceptron model as limited and practically useless, which impeded Rosenblatt’s progress.
美国心理学家弗兰克·罗森布拉特(Frank Rosenblatt)创造了第一个人工神经元感知器，一切都在60年代发生了变化。 这是自动学习作为一门学科的转折点，其目标是在机器之间发展学习，使计算机能够识别信息中的基本规则或模式并据此采取行动。 Rosenblatt的研究被忽视了将近20年，主要是因为MIT教授Marvin Minsky(1969年获得图灵奖的AI先驱)和Seymour Papert(同为计算机科学家)的批评毫无根据。 他们认为Perceptron模型有限且几乎无用，这阻碍了Rosenblatt的发展。
Fast-forward to the ’80s, when the rise of neuronal networks research and the principles of Perceptron gained relevance while being applied to more complex structures. This gave birth to a still-relevant trend of interconnected multilayer artificial neurons. As a result of these decades of research, most of our modern world is ruled by increasingly complex systems. Without them, we would live in a world with self-driving cars, Facebook recognizing your friends, or voice assistants making recommendations about where to eat or what to buy.
快进到80年代，神经网络研究的兴起和Perceptron的原理在应用于更复杂的结构时获得了相关性。 这催生了仍然相关的多层人工神经元互连的趋势。 经过数十年的研究，当今的大多数世界都由日益复杂的系统所统治。 没有他们，我们将生活在无人驾驶汽车，Facebook识别您的朋友或语音助手为您在哪里吃饭或买什么的建议的世界中。
Competing with human intelligence
与人类智力竞争
Considering this technology into the equation, it is possible to think of an analogy between how we humans think and what machines do, thanks to automated or machine learning.
考虑到这项技术，可以通过自动化或机器学习在人类思维方式和机器行为之间进行类比。
One could say that our intellectual capacity comes from the complexity emerging from our brain, with synapses (the link or communications between neurons) as the biological enabler of perception and thinking. Thanks to developments in the deep learning field, machines now have similar elements to neurons and synapses called artificial neural networks (layers with interconnected artificial neurons) and parameters, respectively.
可以说我们的智力来自大脑的复杂性，突触(神经元之间的联系或交流)是感知和思考的生物促成因素。 由于深度学习领域的发展，机器现在具有与神经元和突触类似的元素，分别称为人工神经网络(具有相互连接的人工神经元的层)和参数。
Interaction between these components allows a model to learn how to recognize human faces, generate text, or identify the sentiment expressed in a tweet. That justifies the interest for models to be increasingly deep, with more neuron layers and parameters.
这些组件之间的交互作用使模型可以学习如何识别人脸，生成文本或识别推文中表达的情感。 这证明了对具有更多神经元层和参数的模型越来越深入的兴趣。
GPT-3 excels in this regard: it has 175.000 million parameters, a hundred times more parameters than GPT-2, and nine times more than Turing-NLG, the second biggest model in the world. That number of components is key to its performance, only that number of parameters allows it to adapt, generalize and consume the great amount of data collection for its training giving, as a result, a highly flexible, comprehensive, and surprising AI. However, is all this enough to speak about an approach to general artificial intelligence, capable of performing virtually any human task?
GPT-3在这方面表现出色：它拥有1750亿个参数，是GPT-2的一百倍，是世界第二大模型Turing-NLG的九倍。 如此多的组件是其性能的关键，只有如此多的参数才能使其适应，归纳和使用大量数据收集进行训练，从而提供高度灵活，全面且令人惊讶的AI。 但是，这一切足以说明一种能够执行几乎任何人类任务的通用人工智能方法吗？
GPT-3 can solve tasks related to languages, such as writing a text or answering a question, and also solve math problems (such as adding and subtracting), as well as identifying a sequence or rudimentary code. At first sight, it seems that its path for general intelligence is reduced to achieving deeper models with more parameters. After all, GPT-3 is nothing more than GPT-2 on steroids.
GPT-3可以解决与语言相关的任务，例如编写文本或回答问题，还可以解决数学问题(例如加法和减法)，以及识别序列或基本代码。 乍一看，它的通用情报之路似乎简化为使用更多参数实现更深层次的模型。 毕竟，GPT-3只不过是类固醇的GPT-2。
However, if we look at human evolution, we find a completely different answer. Human beings are not animals with the highest number of neurons, and we don’t excel for our synapses. What makes our brains truly unique is our frontal lobe: the center of command where most of what we consider intelligence originates from and that is more developed than most complex animals. All our executable functions (the ability to select the most effective option according to context) and flexibility (ability to reach correct answers according to a situation), or going through updates (ability to acquire and manipulate new useful information for solving innovative problems), are only possible thanks to our frontal lobes, which have advanced and been perfected for the past million years.
但是，如果我们看人类的进化，我们会找到完全不同的答案。 人不是神经元数量最多的动物，我们的突触并不出色。 使我们的大脑真正独特的是我们的额叶：大脑的指挥中心，我们认为智力的大部分起源于此，并且比大多数复杂的动物发达。 我们所有的可执行功能(根据上下文选择最有效的选项的能力)和灵活性(根据情况获得正确答案的能力)，或进行更新(获取和操纵新的有用信息以解决创新问题的能力)，由于我们的额叶在过去一百万年中不断发展和完善，这才有可能实现。
Maybe AI is also about quality rather than quantity, meaning that more efficient, complex, and deeper algorithms are still waiting to be developed in the pursuit of human-like intelligence.
也许AI也是关于质量而不是数量，这意味着在追求类似人类的智能时，仍然需要开发更高效，更复杂和更深入的算法。
The effort to get out of the lab and into the streets
离开实验室进入街头的努力
So what does all that history mean for AI? When thinking about its uses and applications, we can make the mistake of undervaluing the effort of integrating models into our chaotic and complex world. In fact, one of the major criticisms of GPT-3 is the sexist and racist content it can generate, which indicates that it is not ready for mass adoption. This is less of criticism on its developers and more of a recognition of the tension between research and application. One has the objective of expanding limits of human knowledge in the widest way, while the other tries to solve specific and practical problems.
那么，所有的历史对人工智能意味着什么？ 在考虑其用途和应用时，我们可能会犯一个错误，那就是低估了将模型集成到我们混乱而复杂的世界中的努力。 实际上，对GPT-3的主要批评之一是它可能产生的性别歧视和种族主义内容，这表明它尚未为大规模采用做好准备。 这减少了对开发人员的批评，而更多地是对研究与应用之间的紧张关系的认识。 一个目标是以最广泛的方式扩大人类知识的范围，而另一个则试图解决具体的实际问题。
If we think of another example like the automotive industry, this distinction becomes clearer. Is the engine the essence of a sports car? Yes, of course. Is it enough to know about engines to manufacturing a sports car? No way; even if we are able to design and manufacture an engine, we need the entire car body to support its features. Can we use the same engine in other vehicles? Absolutely, as long as we respect certain guidelines. Can we manufacture cars without manufacturing engines? Yes, indeed; consider McLaren, a prestigious super sports brand that has used Mercedes-Benz engines. In this sense, GPT-3 is nothing else but a V12 with 3,000 horsepower, a monster with a tremendous brute performance able to go from zero to 100 in three seconds. The engine is provided by OpenAI, and it’s in our hands to adapt it and provide it with value as part of a comprehensive solution.
如果我们想到另一个示例，例如汽车行业，这种区别将变得更加清晰。 发动机是跑车的本质吗？ 当然是。 是否足够了解制造跑车的引擎？ 没门; 即使我们能够设计和制造发动机，我们也需要整个车身来支撑其功能。 我们可以在其他车辆中使用相同的发动机吗？ 绝对可以，只要我们遵守某些准则即可。 我们可以在不制造引擎的情况下制造汽车吗？ 确实是的; 想想迈凯轮(McLaren)，这是一个享有盛誉的超级运动品牌，已经使用了梅赛德斯-奔驰发动机。 从这个意义上说，GPT-3就是拥有3,000马力的V12，这是一款具有出色残酷表现的怪物，能够在三秒钟内从零变到100。 该引擎由OpenAI提供，它在我们手中，可以对其进行调整并为它提供价值，作为全面解决方案的一部分。
At R/GA, we understand these subtle but key differences. We are a creative company focused on driving Business, Experience, and Marketing transformation for our clients and partners, converging multiple skill sets and talent: from creatives, designers, programmers, and marketers to strategists, technologists, and data scientists. No single team has a monopoly on the development or practical data application of emerging technology such as GPT-3.
在R / GA，我们了解这些细微但关键的差异。 我们是一家富有创造力的公司，致力于为我们的客户和合作伙伴推动业务，体验和市场营销转型，融合了多种技能和才华：从创意，设计师，程序员和营销人员到战略家，技术人员和数据科学家。 没有哪个团队可以垄断GPT-3等新兴技术的开发或实际数据应用。
We are specialized in building custom “automobiles”, where models based on AI are a means to an end: providing solutions to complex problems with added value for our clients. Of course, we also manufacture “engines” when the project demands it, and the challenge invites us to go further. It is in our DNA to explore, test, and apply emerging technologies, but most of all we are committed to our delivery, aligning our values and capabilities in an industry increasingly filled by empty promises. That is why we have been hiring the best talent, searching expansively and inclusively to recruit and engage people who understand the world through data and see the possibilities through a human-first lens
我们专注于构建定制的“汽车”，其中基于AI的模型是达到目的的一种手段：为客户提供具有附加值的复杂问题的解决方案。 当然，我们还可以在项目需要时制造“发动机”，而挑战也促使我们走得更远。 探索，测试和应用新兴技术是我们的DNA，但是最重要的是，我们致力于交付，使我们的价值和能力与日益充斥着空洞承诺的行业保持一致。 这就是为什么我们一直聘请最优秀的人才，进行广泛而广泛的搜索以招募和吸引那些通过数据了解世界并以人为本的视角看待可能性的人
From telecom to retail, healthcare to financial services, and beyond, all industries today are being challenged and renewed thanks to increasingly intelligent systems — and at R/GA, we want every day to go a step further. Artificial Intelligence is inexhaustible in the right hands: with creativity, humanity, and innovation at the center, new insights and applications emerge, leading to smarter businesses, agile systems, and truly game-changing products and services.
从电信到零售，医疗保健到金融服务，等等，如今，由于智能系统的日益成熟，所有行业都面临着挑战和更新。在R / GA，我们希望每一天都走得更远。 人工智能在右手无穷无尽：以创造力，人性和创新为中心，新的见解和应用不断涌现，从而带来了更智能的业务，敏捷系统以及真正改变游戏规则的产品和服务。
While AI has a long way to go, every great adventure starts with an open mindset of discovery, exploration, and purpose. So will GPT-3 live up to all the hype? We don’t know what the future holds, but we will surely keep contributing as we’ve been doing for more than 40 years at R/GA, pushing the limits of what’s known to set new benchmarks for innovation and exploration that will improve lives, organizations, and even societies. It’s what makes us leap out of bed each morning and roll up our sleeves. And if that’s an exciting prospect to your career or business too, by all means — please reach out and continue the conversation.
尽管AI还有很长的路要走，但每一次伟大的冒险都始于开放的发现，探索和目标思维。 那么，GPT-3能否如愿以偿？ 我们不知道未来会怎样，但是随着我们在R / GA上所做的40多年的努力，我们一定会继续做出贡献，突破已知的极限，为创新和探索设定新的基准，以改善生活，组织甚至社会。 这就是使我们每天早晨跳下床然后卷起袖子的原因。 而且，如果这对您的职业或业务而言也是令人兴奋的前景，请与他人联系并继续进行对话。
— Nicolás Rodríguez is Lead Data Scientist at R/GA Buenos Aires. He can be reached at [email protected].
—NicolásRodríguez是布宜诺斯艾利斯R / GA的首席数据科学家。 可以通过[email protected]与他联系。
翻译自: https://medium.com/@RGA/gpt-3-and-the-artificial-intelligence-that-surrounds-us-98572617fd05
open ai gpt