最近双语君搞了个AI内测账号,每天沉迷于请AI作画。下边是几张双语君请AI画的画,大家猜猜都输入了什么关键词?

  答案

  五彩斑斓的黑

  答案

  鱼香肉丝

  答案

  过桥米线

就在过去这几个月里,以ChatGPT为典型的一类“善解人意”的AI忽然火了起来。这类AI的实用性特别高,能画画、能翻译句子、能写报告,最重要的是人指挥TA做这些事情都不用费心费力输入命令,只需要像聊天一样打字即可,不费劲。

AI理解人类语言,有点像坐标系

中国人民大学高瓴人工智能学院助理教授周骁表示,ChatGPT这类AI可以叫做自然语言型AI,其主要学习方向就是理解分析人类语言。在这之前,人机博弈型AI早在2016年左右就实现了突破,阿尔法狗战胜世界冠军柯洁就是最好的例子;图像识别型AI则是在过去几年达到满意的精度,能广泛应用于各种人脸识别、车牌识别的场景。

Zhou Xiao, an assistant professor at Gaoling School of Artificial Intelligence, Renmin University of China, listed a few examples of AIs that people can use in daily life. There is the gaming AIs that can compete with humans playing games, such as AlphaGo developed by DeepMind around 2016, which features a search algorithm, deep neural networks and reinforcement learning.

There are also image recognition AIs that can recognize car plate numbers for traffic management and human faces in the hunt for suspects, which made breakthroughs around 2020 with their precision rate reaching 99 percent.

至于最近大火的ChatGPT则是语言模型,通过多种人工智能技术和算法实现了对自然语言的理解和生成,从而能够与人类进行自然的对话交互,它的出现让人工智能强势地摘掉了“人工智障”的帽子。GPT的全称叫生成预训练转换器(Generative Pretrained Transformer),复旦大学附属中山医院院长、中国科学院院士葛均波曾经很形象地解释过这三个词:

Generative是生成的意思,表示它区别于既往搜索引擎,能自己生成创作出新的内容;Pretrained是预先训练意思,表示它之前已经在一些数据库里经过训练过,已经具有一些自身逻辑及判断;Transformer是一种全新的算法架构,能够在处理长文本时更高效,训练时间更短。

百姓网创始人、CEO王建硕试图通过坐标系的概念解释GPT识别人类语言的原理。“好比在三维空间中,知道三个坐标值就能确定一个点的位置,这三个坐标可以叫做参数,”他边比划边说:“但当我们提起某个物体时,可能需要上千个参数才能确定这一物体究竟是什么。比如苹果的参数有能吃、水果、颜色红或绿、在树上生长、一般直径不超过10厘米等等。类似参数越多,我们就越能精准确定某物体。”

“That algorithm works in a way like finding a point in a space,” Wang Jianshuo, founder and CEO of Baixing AI, a company building the basic infrastructure for a world that bots talk with bots, tried to explain how it works in plain language: “With three parameters one can locate a point in a 3-dimensional coordinate system. In human language one might need thousands of parameters to describe an object; For example an apple needs parameters such as edible, fruit, green or red in color, grown from a tree, generally smaller than 10 cms in diameter to be defined. The more parameters that can be defined, the more accurately AI can find the right point.”

当苹果和香蕉同时出现时,能吃、甜、水果等参数就不够了,光凭这三个参数谁也没有办法确定讲述的是苹果还是香蕉。这时候就得增加参数,比如形状是长条还是类似圆形?颜色是红绿还是黄?“这就是GPT的基本算法,本质上是建了一个上千甚至更高维度的模型。其实人脑中也有无数个这样的模型并且在随时调用,只是我们不会感受到罢了。”

Wang cited apples and bananas as two examples. With parameters such “edible”, “sweet” and “fruit”, neither humans nor an AI could distinguish them from each other. But with the parameter concerning the shape included, namely “long” or “round”, one could make a guess. With the parameter of the color being red, green or yellow, one could be surer about his/her judgment. “That’s also how GPT works — Most of the times it takes thousands of parameters for AI to define an object like we humans do,” he said: “we just do not realize that we are doing that.”

所以有时候AI对句子的理解异于我们常人。比如下图就是AI画的驴肉火烧,肉、火、烧的元素都齐备了,驴则以两只长耳朵的形式出现。

AI画的驴肉火烧

要从如此多的维度描述一个物体,其计算量自然可想而知。据复旦大学计算机科学技术学院教授张军平表示,ChatGPT大概用了285000个CPU和10000多颗GPU、其训练文本数据达到45Tb (1Tb=1024Gb),训练成本高达1200万美元。

在一篇题为“ChatGPT:潜力、期待与限制”的论文中,张军平进一步解释道,ChatGPT主要受益于大型语言模型(Large Language Model),使用语言模型(LM)用大规模数据训练庞大的神经网络模型。

“ChatGPT会生成与用户意图相匹配的多个回合的响应。ChatGPT捕获先前的会话上下文来回答某些假设问题,这大大增强了会话交互模式下的用户体验。”

“It was based on the training of 45 Terabytes of data that ChatGPT made its breakthrough early this year, for which purpose it used about 285,000 CPUs and over 10,000 model A100 GPUs,” said Zhang Junping, a professor on computer technology at School of Computer Science, Fudan University, who stressed that the amount of data the language model involves is of key importance to AIs like GPT in an academic essay ChatGPT: Potentials, prospects, and limitations: “ChatGPT benefits mainly from Large Language Models(LLMs) that train huge neural network models with large-scale data using Language Models(LMs).”

“ChatGPT generates responses that match the user’s intent with multiple turns. ChatGPT captures previous conversational contexts to answer certain hypothetical questions, which greatly enhances the user experiences in conversational interaction.”

AI画的日之夕矣羊牛下来

为测试这些AI对人类语言的理解能力,双语君分别在三个AI对话框中输入“anlyze how Artifical Intelligenz works”,拼写错误都被轻松忽略。王建硕则教了个更直观的测试方法:问AI《家庭问答》的主播年龄的平方根是多少?要回答这一问题,AI需要搞清楚《家庭问答》是什么、其主播是谁、年龄多大了,最后再算一次平方根。

简直完美。

For all three AIs of the type, a test sentence is input to “anlyze how Artifical Intelligenz works” with the spelling errors intentionally left uncorrected. All three neglected the errors as if they didn’t exist and just talked on.

Wang offered as a test of the ability of ChatGPT the following question: “What’s the root of The Family Feudhost’s age?” To solve that problem, ChatGPT has to find out what The Family Feudis, who hosts it, how old he is and then calculate the root.

AI成熟后,连上班关门都会变得简单

前边讲的几种AI都深刻改变了我们的生活。比如人机博弈AI成熟后,游戏行业发展得到加速,人机对战不再是“虐菜”的代名词;人脸识别AI成熟后,“刷脸”也成为了人们的主流支付方式。

AI has already launched a revolution not only in the computer industry, but also in daily lives. In gaming, AIs have made major progress over the past half a decade, people now can play games such as chess or GO with smart AIs as opponents so as to sharpen their own skills. Image recognition AIs have also become mature, people could easily open a lock or pay for a deal by holding theirsmartphone in front of their face.Now with GPT making fast developments, the way people interact with computers might changeagain.

在风云学会陈经看来,GPT的发展将给人机交流方式带来革命性的改变。看看电脑的发展历史,电脑“输出”一直在进步,从文字到图像再到今天绚丽的建模与视频,但“输入”始终还是电脑鼠标,直到几年前声控才走入大家的生活,Siri还经常听不懂人话。随着GPT的发展与声音识别的结合,声控的准确性将大大提高。

According to Chen Jing, a researcher at Fengyun Institute of Science, Technology and Strategy, computers still receive instructions and requests from human users in a way that is quite inefficient, via akeyboard and mouse, as the user must click on some icon or type in something to make the computer act as required. With speech recognition and face recognition AI technologiesimproving significantly in the 2010s, voice recognition became a newmeansof input.

3月15日,Open AI发布GPT-4,在GPT-3的基础上增加了图像识别功能。这就为人机进一步高效互动提供了可能性,比如目前还比较初级的手势控制将更快成熟。

The GPT-4 technology, released by Open AI on March 15, propels the process another step forward. Being able to recognize images in a more accurate, reliable way, the GPT-4 technology can understand humans in a more efficient way.

In Chen’s view, the GPT-4 technologymay enable humans to make commands to machines via gestures. For example, currently there are already smart appliances that people can talk to and tell them what to do. In the future they will be able to wave and be caught by a camera on the appliance, a technology that exists now but is not yet ripe.

陈经甚至设想了一个场景:早上离家上班时,不用再锁门关灯,而只需要对着大门口的摄像头挥手道别。AI立即会领会主人的意思,关灯反锁以待主人归来;等主人回来了,再对大门招招手,门就自动开了,走到哪个屋子里灯光也自动打开,躺到沙发上电视也会自己开开。

“Imagine that you are leaving home for office,” Chen said: “currently you have to turn off the lights and lock the door. With GPT-4, all you need is to wave goodbye to the camera at your gate, then the AI will understand you and turn off all unnecessary appliances, close the door and lock it for you. When you come home, just smile at the camera and it will wake everythingup. That’s not only because the AI can recognize your face based on the technology that became ripe around 2020, but also because it can understand your gestures and facial expressionsbased on GPT-4.”

AI会取代人类的工作吗?

大家并不是都认同人工智能的出色表现的。据双语君观察,这种认同与否往往与一个人从事的行业有关系。

比如,一位画家就对AI作画这件事比较抵触。“AI画的画是没有灵魂的”,他不无愤慨地表示:“其实那都不应该叫作画啦。AI不过是从网上扒拉点素材,再拼起来而已啦。”

Not everybody is happy with AI’s growing abilities. “The images drawn by AI are soulless”, said Xi Li (pseudo name), a 41-year-old painter, with a little anger: “They can help humans, but never replace human hands.”

Then he went further: “We don’t call the deed drawing — AI is only collecting materials online and sticking them together.”

AI获奖画作《太空歌剧院》引发争议

再比如,一位强烈要求匿名的科幻作者对AI写稿表示鄙视,认为AI不可能像人类一样把握各种词汇的运用。那怎么看“文无第一、武无第二”呢?这位作家表示同意这句话,但再三强调“AI写的小说抓不住读者的心”。

Similarly, a science fiction writer who hopes to stay anonymous said that AI can never write as good as professional writers do. “There is no yardstick to measure how good an essay is, but the ones written by AI just do not strike people in the heart.”

社科院人口与劳动经济所研究员屈小博教授一直研究AI对就业的影响,他也很理解这些人的感受。“大家都担心自己的工作被取代嘛,”他笑笑说:“我们这些研究员也担心啊。”

For Qu Xiaobo, a professor at the Institute of Population and Labor Economics and deputy director of the Human Resource Center, Chinese Academy of Social Sciences(CASS), that reflects people’s worries about their jobs being taken away by an AI. “Such concerns are understandable,” he said. “Even we professors worry about being replaced by AI professors.”

不过屈教授总体还是持乐观态度的,因为他觉得AI消灭旧岗位的同时也会带来新岗位。他表示劳动经济学里有个“全要素生产率”的概念,AI能在很大程度上提高全社会的生产率,从而创造更多的社会财富以供分配,也需要更多的岗位。

But Qu holds a positive attitude toward employment in the age of AI. “While taking jobs, AI creates new jobs, too,” he said. “Its general effect upon the market depends on how many jobs will disappear and how many will emerge. AI willtheoretically increase the Total Factor Productivity(TFP) of society, which means it should create more jobs than it replaces.”

如果在招聘网站上搜索带AI的招聘岗位,“AI顾问”、“AI工程师”、“AI指导”基本都能占据好几页,而且薪水看起来不算低。屈教授管这些岗位叫“人机交互岗位”,认为这种岗位会越来越多。

If one searches “AI” on several domestic job websites, new jobs involving AI released for weeks even days such as “AI consultant”, “AI engineer”, “AI tutor” will emerge and cover almost tens of pages, with promised wages ranging from 15,000 to 30,000 yuan ($4,368) in Beijing, the capital with an average monthly salary around 15,000 yuan for 2022. Qu said these are the newly emerging jobs he refers to, as quite many are dealing with AI, or what he called “jobs for human-AI interactions”.

AI相关招聘岗位

王建硕更是比较乐观,他还表示每次科技进步都会消灭旧岗位带来新岗位。“汽车发明了,马车夫这个工种慢慢淘汰了,但汽车司机是不是比那时的马车夫多多了?待遇也好吧?飞机发明出来,不就有了飞行员这个职业?而且有了飞机大家的生活水平也被带着提高了不少。AI进步越快,我们的总体生活就越好。”

His view is echoed by Wang, who said that new jobs will be created with the fast development of AI. “Actually with every major technological progress new jobs emerged,” he said. “With the invention of automobiles there was the need for drivers, while with the invention of planes we got pilots, and people’s general living standards have been increasing with these progresses because the general productivity of society was increased. The faster AI progresses, the easier life for people will be.”

不过屈教授强调,对个人来说压力还是有的,而且不小,因为自己的技能未必能适配新的岗位,要小心被劳动力市场的结构性矛盾所挤压。他的意见是,作为个人要终身学习、努力适应新情况,作为国家与社会则需要重视这一问题,给劳动者提供足够的培训与适应机制,尽量最小化甚至避免结构性矛盾带来的阵痛。

However, for individuals, how to keep one’s job remains a challenge because possibly not everyone has the skills needed for the new jobs. Qu said that phenomenon has an academic term “structural mismatch”. But he added that’s not unique to AI,but has been a common occurrence throughouthistory because technology progresses all the time. The current difficulties for college graduates to find jobs, he said, are also partly the result of a “structural mismatch”. To solve that, individuals need to keep learning new things so as to both have experiences and skills. Qu also stressed the State’s role in providing employees with better training so that they can constantly update their skills to meet the changes in the job market.

[Photo/pexels]

不会“追求幸福”是AI的死穴

作为一名毕业于自动化专业、参与过人工智能研发的记者,笔者认为,人类有一个特性AI目前赶不上,在肉眼可见的将来也赶不上,那就是对美好生活的向往。

AI is smart, but the inability to pursue happiness is an uncrossable line that distinguishes humans from AI, so far at least.

这个表述听起来似乎有点大,其实和人们的生活连接十分紧密,因为人类社会就是在不断地追求幸福中前进的。原始人厌倦了靠双脚丈量大地的日子,所以发明了车轮,后来又发明了火车汽车;人类不满足于被囚禁在陆地上,老想去探索海洋,由此才发明了独木舟、木船以及后来的轮船;人类羡慕鸟儿的翅膀,想在天空自由地翱翔,于是飞机火箭一个个从梦想成为了现实。

某种意义上,是人类对美好生活的向往倒逼了技术进步,从而推动了历史的车轮滚滚向前。

The word might sound too high-tone, but almost every major technological progress has been the outcome of people’s pursuit of a better life. People wanted to travel over the sea, so they invented the canoe, boats and ships. People wanted to save the trouble of walking on foot, so they invented the train and automobiles. People wanted to fly in the sky and to the moon, so they invented the airplane and the rocket. To a great degree, it is our ancestors’ pursuit of a better, more convenient life that has shaped how we live today.

到目前为止所有的AI都没有追求幸福的能力。它们就像最顶尖的仆人,训练有素反应迅捷,能在最短时间内解决人类提出的所有问题与需求,却唯独没有自己提出需求的能力。不提出需求也就没法进步,就只能止步于满足当下的循环。

如果以人类目前的技术水平创造出一个人工智能世界,那这个世界将能够顺畅运行却无法进步。原因很简单,没有进步的需求,人工智能在本质上只是人类编写的一段段程序,能够在已有的材料中精中取精建造大厦,却永远不会想着给这个世界增添片瓦。

But that’s exactly what ChatGPT, or any other kind of AI, lacks. AI in essence is sequence of codes written by humans to make daily life better. AI does not have its own pursuit. It does not dream of improving the conditions for itself, which prevents it from progressing on its own. It has the potential of designing an airplane, but it will never do so unless given instruction to do so by a human, which makes its progress dependent on humans.

有时候“求生欲”也是进步的一种动力,而这恰恰是生命的专属。当生命在38亿年前诞生于地球上时,就已经注定了生存繁衍的本能,其一切活动都是为了更好地生存,为此才能改变世界;人工智能就不一样了,不过是代码,是人类为解决问题而创造出来的,其本能是解决问题。这就决定了二者“主”和“仆”的区别,是被服务与服务的根本差别。

只有生命才想着给自己创造一个更好的环境,只有人类作为智慧生命有能力给自己创造这样的环境,人工智能还没法达到这样的境界。

In a philosophical sense, the pursuit of a better life has been carried in the genes of every creature for 3.8 billion years, since the primal life came into being on this planet. Since that moment, each species has continually evolved so as to win the competition to survive. To be more accurate, the instinct to survive, to continue existing as a life is the original driving force of the pursuit of happiness, but AI, whether GPT and other AI technologies, lack that driving force. They are merely long sequences of code that people write to make human lives better, for which purpose their own existence is not important. After all, AI is with a history of only decades, while that of life is hundreds of millions years.

王建硕说他们曾经做过测试,问AI生命的本质是什么,根据回答再一句句追问,结果AI答了三天三夜。他觉得有了这么强的思考能力,AI终究会获得这种求生欲,从而想着改变自己的境遇。至于完成这个突变所需要的时间,他认为是五年甚至更短,但机器真的实施求生欲可能要非常久远。

五年不过一瞬间,让我们拭目以待。

Talking about the question, Wang Jianshuosaid he once asked his AI the question about the meaning of life, which inspired the latter to think and think and think, answer and answer and answer, and for two whole days it had been explaining this. “Ultimately AI will gain the instinct to survive”, he said at last: “I trust these products we humans make and maybe all it needs is another five years or even shorter time span, but it might take longer time to implement it.”

On which side will you bet?

记者:张周项编辑:朱迪齐实习生:姚欣可