少女野外调教 外刊精读|《经济学东谈主》东谈主工智能简史

A short history of AI少女野外调教

For decades， artificial intelligence failed to achieve its lofty goals.几十年来，东谈主工智能没能达成其宏伟打算In the first of six weekly briefs we ask what changed.在一语气六期每周提要著作的第一篇中，咱们探寻了什么如故发生了窜改。

图片

Over the summer of 1956 a small but illustrious group gathered at Dartmouth College in New Hampshire; it included Claude Shannon， the begetter of information theory， and Herb Simon， the only person ever to win both the Nobel Memorial Prize in Economic Sciences awarded by the Royal Swedish Academy of Sciences and the Turing Award awarded by the Association for Computing Machinery.They had been called together by a young researcher， John McCarthy who wanted to discuss “how to make machines use language， form abstractions and concepts” and “solve kinds of problems now reserved for humans”. It was the first academic gathering devoted to what McCarthy dubbed 'artificial intelligence' And it set a template for the field's next 60-odd years in coming up with no advances on a par with its ambitions.

1956年的夏天，一小群顶尖的学者在新罕布什尔州的达特茅斯学院约聚。这内部有信息论的创举者克劳德·香农，和惟一同期得到瑞典皇家学院授予的诺贝尔经济学奖和好意思国计较机学会授予的图灵奖的赫伯特·西蒙。他们被年青的学者约翰·麦卡锡召唤到一谈，他想要辩论“怎样让机器使用话语、酿成概括意见”和“处理多样现时由东谈主类处理的各类问题”。这是致力于麦卡锡称之为“东谈主工智能”的第一次学术约聚，为这个领域在接下来的60多年的发展设定了模版，尽管在此时间得到的进展还无法配得上其志在千里。

The Dartmouth meeting did not mark the beginning of scientific inquiry into machines which could think like people. Alan Turing， for whom the Turing prize is named， wondered about it; so did John von Neumann， an inspiration to McCarthy. By 1956 there were already a number of approaches to the issue; historians think one of the reasons McCarthy coined the term artificial intelligence， later AI， for his project was that it was broad enough to encompass them all， keeping open the question of which might be best. Some researchers favoured systems based on combining facts about the world with axioms like those of and symbolic logic so as to infer appropriate responses; others preferred building systems in which the probability of one thing depended on the constantly updated probabilities of many others.

科学界对不详像东谈主类一样想考的机器的探寻，达特茅斯会议并不是开头。阿兰·图灵想考过这个问题，图灵奖等于以他定名的；麦卡锡的偶像冯·诺依曼亦然。这个问题在1956年就如故有了一些解法；历史学家以为，麦卡锡创造东谈主工智能（AI）这个意见的其中一个原因，是为了让他的名目不错包罗万象，让这个问题保捏通达性可能是最佳的遴选。有些商酌东谈主员倾向于基于寰宇表象和访佛几何公理、标识逻辑磋磨的系统，不错给出合理反馈；其他商酌员更倾向于开发在事物之间无间更新的概率系统。

The following decades saw much intellectual ferment and argument on the topic， but by the 1980s there was wide agreement on the way forward: 'expert systems' which used symbolic logic to capture and apply the best of human know-how. The Japanese government， in particular， threw its weight behind the idea of such systems and the hardware they might need. But for the most part such systems proved too inflexible to cope with the messiness of the real world. By the late 1980s AI had fallen into disrepute， a byword for overpromising and underdelivering. Those researchers still in the field started to shun the term.

在随后的几十年间，针对这个主题有好多学问分子的想想摇荡和争论。但到了1980年代，关于使用标识逻辑图捕捉和支配东谈主类最佳学问训戒的“大众系统”如故得到了泛泛认同。尤其日本政府在基于这个意见的访佛系统和所需的硬件上参预宏大。但是，绝大多数这类系统被说明，在处理确切寰宇的复杂性时零落活泼性。1980年代末，东谈主工智能如故端淑扫地，成为了过度愉快却无法闭幕的代名词。仍然在这个领域的商酌东谈主员运行隐敝这个术语。

It was from one of those pockets of perseverance that today's boom was born. As the rudiments of the way in which brain cells - a type of neuron - work were pieced together in the '90s， computer scientists began to wonder if machines could be wired up the same way. In a biological brain there are connections between neurons which allow activity in one to trigger or suppress activity in another; what one neuron does depends on what the other neurons connected to it are doing. A first attempt to model this in the lab (by Marvin Minsky， a Dartford attendee) used hardware to model networks of neurons. Since then， layers of interconnected neurons have been simulated in software.

今天东谈主工智能领域的大爆发，恰恰源于繁多依然坚捏的群体中的一个。这种关节的基本理念，源于脑细胞（一种神经元）的职责旨趣在90年代被从容领略，计较机科学家运行想考，是否机器不错以同样的式样搭建起来。在生物的大脑里，神经元的勾搭使得某个神经元的举止被激活或者扼制另一个神经元的举止。一个神经元作念什么，依赖于与其勾搭的其他神经元在作念什么。参加那次达特茅斯会议的马尔文·明斯基，第一个在试验室中尝试搭建这种模子，他使用硬件去模拟神经网罗。从此以后，多层勾搭的神经网罗也在软件中被模拟出来。

These artificial neural networks are not programmed using explicit rules; instead， they 'learn' by being exposed to lots of examples. During this training the strength of the connections between the neurons (known as 'weights') are repeatedly adjusted so that， eventually， a given input produces an appropriate output. Minsky himself abandoned the idea， but others took it forward. By the early 1990s neural networks had been trained to do things like help sort the post by recognising handwritten numbers. Researchers thought adding more layers of neurons might allow more sophisticated achievements. But it also made the systems run much more slowly.

这些东谈主工神经网罗莫得使器具体的端正进行编程，状貌全非是，它们通过被败露在无数样本中来“学习”。在考试中，神经元之间的相接（被称为“权重”）被反复诊疗，是以最终一个给定的输入会产生一个合适的输出。明斯基我方覆没了这个方针，但其他东谈主在这个方朝上赓续前行。到了90年代早期，神经网罗如故不错被考试得不错作念通过识别手写数字来完成邮件分拣这么的事情。商酌东谈主员以为增多更多层的神经网罗也不错得到愈加复杂的后果。但这么也会使得系统运行得更慢。

A new sort of computer hardware provided a way around the problem. Its potential was dramatically demonstrated in 2009， when researchers at Stanford University increased the speed at which a neural net could run 70-fold， using a gaming PC in their dorm room. This was possible because， as well as the 'central processing unit' (CPU) found in all PCs， this one also had a graphics processing unit (GPU) to create game worlds on screen. And the GPU was designed in a way suited to running the neural-network code.

一种新的计较机硬件为处理这个问题提供了关节。这种关节的后劲在2009年被戏剧性地展现出来，斯坦福大学的商酌员在寝室的游戏电脑上将一个神经网罗的计较速率耕种了70倍。这可能是因为除了通盘电脑皆有的中央处理器CPU之外，这台电脑还有在屏幕上创建游戏寰宇的图形处理器GPU。GPU的想象式样适值合适来运行神经网罗的代码。

Coupling that hardware speed-up with more efficient training algorithms meant that networks with millions of connections could be trained in a reasonable time; neural networks could handle bigger inputs and少女野外调教， crucially， be given more layers. These “deeper networks” turned out to be far more capable.

磋磨硬件的提速和愈加高效的考试算法，领有百万勾搭的神经网罗不错在可禁受的时候里被考试；神经网罗不错处理更大的数据量输入，最弱点的是，不错领有更多层的网罗。这些“深层网罗”说明不错领有更强的智力。

The power of this new approach， which had come to be known as 'deep learning'， became apparent in the ImageNet Challenge of 2012. Image-recognition systems competing in the challenge were provided with a database of more than a million labelled image files. For any given word， such as 'dog' or 'cat'， the database contained several hundred photos. Image-recognition systems would be trained， using these examples， to 'map' input， in the form of images， onto output in the form of one-word descriptions. The systems were then challenged to produce such descriptions when fed previously unseen test images. In 2012 a team led by Geoff Hinton， then at the University of Toronto， used deep learning to achieve an accuracy of 85%. It was instantly recognised as a breakthrough.

这种被称作“深度学习”新关节的实力，在2012年的图像网罗挑战赛中展露无遗。在挑战赛中，提供给图像识别系统存储了一百万张打标图像文献的数据库。关于任何输入的单词，比如“狗”或者“猫”，数据库中皆包含几百张像片。图像识别系统将使用这些样例进行考试，将图像风光的输入映射至以一个单词态状的笔墨输出。在挑战中，系统会被提供一些从未见过的测试图片，条目对其输出笔墨态状。2012年在多伦多大学，由杰夫·亨敦教学的小组使用深度学习取得了85%的准确率。这坐窝被以为是一个打破。

By 2015 almost everyone in the image-recognition field was using deep learning， and the winning accuracy at the ImageNet Challenge had reached 96% - better than the average human score. Deep learning was also being applied to a host of other 'problems...reserved for humans' which could be reduced to the mapping of one type of thing onto another: speech recognition (mapping sound to text)， face-recognition (mapping faces to names)， and translation.

到了2015年，险些通盘在图像识别领域的东谈主皆在使用深度学习，图像网罗挑战赛的最高识别准确率如故达到了96%，这也高出了东谈主类的平均水平。深度学习也被支配到一系列其他不错被简化为模式匹配的“东谈主类才能处理的问题“，举例语音识别（将声息动荡为笔墨）、东谈主脸识别（将东谈主脸动荡为名字）、还有翻译。

In all these applications the huge amounts of data that could be accessed through the internet were vital to success; what was more， the number of people using the internet spoke to the possibility of large markets. And the bigger (ie， deeper) the networks were made， and the more training data they were given， the more their performance improved.

在通盘的支配中，不错通过互联网战役到海量的数据是得手的弱点。使用互联网的东谈主越多意味着领有更大阛阓的可能性。搭建更大（举例，更深层）网罗和提供更多的考试数据，神经网罗的理解越好。

Deep learning was soon being deployed in all kinds of new products and services. Voice-driven devices such as Amazon's Alexa appeared. Online transcription services became useful. Web browsers offered automatic translations. Saying such things were enabled by AI started to sound cool， rather than embarrassing， though it was also a bit redundant; nearly every technology referred to as AI then and now actually relies on deep learning under the bonnet.

深度学习很快就被部署在通盘种类的新址品和工作中。语音驱动的开采，举例亚马逊的Alexa出现了。在线转录工作变得有效。网罗浏览器提供了自动翻译功能。淌若说这些东西皆因为AI成为可能，终于运行听起来很酷而不是无语了，尽管有些富余。险些每个被称之为AI的科技当今的确简直在底层相配依赖深度学习。

In 2017 a qualitative change was added to the quantitative benefits being provided by more computing power and more data: a new way of arranging connections between neurons called the transformer. Transformers enable neural networks to keep track of patterns in their input， even if the elements of the pattern are far apart， in a way that allows them to bestow “attention” on particular features in the data.

2017年，由更多算力和更多数据带来的量化收益又迎来了质变：一种摆设神经元勾搭的新关节-变形器。变形器不错让神经网罗对输入数据中的模式保捏追踪，即使这些模式的元素距离很远，也不错聚拢留心力在数据的某些特定特征上。

Transformers gave networks a better grasp of context， which suited them to a technique called self-supervised learning. In essence， some words are randomly blanked out during training， and the model teaches itself to fill in the most likely candidate. Because the training data do not have to be labelled in advance， such models can be trained using billions of words of raw text taken from the internet.

变换器使得神经网罗对高下文有了更好的领略，这种时候被称之为自监督学习。本色上，某些词会在考试中被立时去掉，模子教它我方填上访佛的替代内容。因为考试数据无需被提前打标，这种模子不错使用网上获取的数以亿计的原始语料进行考试。

图片

Mind your language model留心你的话语模子

在线看三级片

Transformer-based large language models (LLMs) began attracting wider attention in 2019， when a model called GPT-2 was released by OpenAI， a startup (GPT stands for generative pre-trained transformer). Such LLMs turned out to be capable of 'emergent' behaviour for which they had not been explicitly trained. Soaking up huge amounts of language did not just make them surprisingly adept at linguistic tasks like summarisation or translation， but also at things like simple arithmetic and the writing of software - which were implicit in the training data. Less happily， it also meant they reproduced biases in the data fed to them， which meant many of the prevailing prejudices of human society emerged in their output.

当一种名叫GPT-2的模子被创业公司OpenAI在2019年发布，基于变换器的谎言语模子LLMs运行引起泛泛热心（GPT代表着预考试的生成式变换器）。这么的大模子不错应答莫得被具体考试过的”突发“行为。罗致无数的话语不仅使他们在举例追念或翻译这么的话语任务时理解喜东谈主，还不错完成肤浅算数和编程职责，这些智力隐含在考试数据中。稍有起火的是，这也意味着他们能从喂给他们的数据中复制了偏见，这意味着好多东谈主类社会中广博存在的偏见会出当今他们的输出中。

In November 2022 a larger OpenAI model， GPT-3.5， was presented to the public in the form of a chatbot. Anyone with a web browser could enter a prompt and get a response. No consumer product has ever taken off quicker. Within weeks ChatGPT was generating everything from college essays to computer code. Al had made another great leap forward.

2022年11月，一个更大的OpenAI模子GPT-3.5以聊天机器东谈主的风光出当今公众眼前。任何东谈主使用网罗浏览器就不错输入一条请示得到一个反应。从来莫得花消居品有过这么快速率的爆火。几个星期内，ChatGPT产生了从大学作文到计较机代码的一切。东谈主工智能如故又上前起原了一大步。

Where the first cohort of AI-powered products was based on recognition， this second one is based on generation. Deep-learning models such as Stable Diffusion and DALL-E， which also made their debuts around that time， used a technique called diffusion to turn text prompts into images. Other models can produce surprisingly realistic video， speech， or music.

东谈主工智能加捏的居品在第一个阶段的功能基于识别，而在第二个阶段基于生成。像Stable Diffusion和DALL-E这么的深度学习模子，在初度公开亮相的时候使用了被称作扩散的时候将笔墨转念为图像。其他模子不错生成相配传神的视频、语音或音乐。

The leap is not just technological. Making things makes a difference. ChatGPT and rivals such as Gemini (from Google) and Claude (from Anthropic， founded by researchers previously at OpenAI) produce outputs from calculations just as other deep-learning systems do. But the fact that they respond to requests with novelties makes them feel very unlike software which recognises faces， takes dictation or translates menus. They really do seem to 'use language' and 'form abstractions'， just as McCarthy had hoped.

这个飞跃不单是是时候上的。生成事物带来了很大的不同。ChatGPT和Gemini（来自谷歌）和Claude（来自Anthropic，由前OpenAI商酌员创立的公司）这么的竞争敌手，就像其他深度学习系转圜样通过计较输出扬弃。但事实是，他们是以新奇的扬弃反应肯求，这使得他们与东谈主脸识别、听写或翻译菜单嗅觉很不一样。他们简直好想在”使用话语“和”酿成概括意见“，就像麦卡锡但愿的那样。

This series of briefs will look at how these models work， how much further their powers can grow， what new uses they will be put to， as well as what they will not， or should not， be used for.

这一系列的提要著作，将会热心这些模子是怎样职责的，它们的智力将会变得何等刚毅，它们将会带来什么新支配，以及它们不成作念什么或者不应该被用来作念什么。

*本文翻译自《经济学东谈主》2024年7月20日科技著作《A short history of AI》，仅供英文相通学习使用，原图文版权归经济学东谈主杂志通盘

关联主题著作：

外刊精读|《经济学东谈主》到底AI是怎样想考的？

外刊精读|《经济学东谈主》AI东谈主才战日趋尖锐化少女野外调教

本站仅提供存储工作，通盘内容均由用户发布，如发现存害或侵权内容，请点击举报。