“所以我一旦将你的订阅源添加到我们网站
“ So once I added your RSS feed to our site,
每篇发表在
every single article that is published on
Vox.com网上的文章
Vox.com is getting sent
都会自动上传 然后我们的工作就像是为它自动创造视频”
through a feed and we ’ re just like automatically creating a
“天哪 太疯狂了
video for it.” “Oh my god that is so crazy.
我们提供的每篇文章”
Every single article through our feed. ” “
“每篇文章” 这是Wibbitz
Every single article. ” This is Wibbitz.
它是自动化制作新闻视频的公司之一
It’s one of the companies automating news video production.
你可以理解为机器人要抢我的饭碗了
You might call this the robot coming for my job.
“所以对于这篇文章
“ So this article,
我们的算法会智能地总结出
our algorithm will just intelligently summarize it into just a quick
一个三十秒到一分钟的短视频
30 second to 1 minute video.
然后根据文章的关键词
And then based on the keywords in the article,
就能将它和相关媒体相匹配”
it ’ s gonna match relevant media to it. ”
当想到它可能会遇到的各种困难的时候
It ’ s pretty impressive when you think
你就会对它现在的表现感到十分惊讶
about all the ways it could get confused.
“开始异常艰难
“In the beginning it was very rough.
遇到重名现象会让人工智能感到困惑
People with the same names would confuse it.
比如土耳其（Turkey）这个城市和火鸡（turkey）这个动物”
Turkey the country and turkey the animal wouldbe another example.”
他们的产品建立在机器学习算法之上
Their product was built with machine learning algorithms,
随时间变得越来越精确
and it became more accurate over time.
成果就是在同等情况下
The result is a video made
通过人工智能几秒钟制造出来的视频
in a few seconds that ’ s not drastically different from what
和一个人用几个小时制造出来的视频 没有很大差异
a human would make in several hours, giventhe same constraints.
Wibbitz是目前快速发展的所谓”人工智能”产品行业的一部分
Wibbitz is part of a rapidly growing industry of so-called “ AI-powered ” products.
在公司的财报会议上提到人工智能的公司数量
The number of companies mentioning artificial intelligence
在过去三年里
in their earnings calls has skyrocketed
突飞猛涨
in the past 3 years.
但事实上
But the truth is
“人工智能”这个术语定义并不明确
that the term “ artificial intelligence ” isn ’ t very well defined.
“实际情况是
“ What happens with AI is that initially lots
最初各种东西都被叫做‘人工智能’
of things are called artificial intelligence.
在过去它是专家系统
It used to be the expert systems;
这种控制飞机起飞降落的系统
the kind of systems that fly airplanes were called
被称为人工智能
artificial intelligence.
一旦它行之有效 成为惯例
Then once they were working and routine and everyone takes them
所有人便将它视为理所当然
for granted, then they
接着人们便不再称它为‘人工智能’”
are not called AI anymore.”
如今当人们讨论“人工智能”
Right now when people talk about AI,
主要讨论的是机器学习
they ’ re mostly talking about “ machine learning ”
这是至少可追溯到上个世纪五十年代的计算机科学的一个分支
– a subfield of computer science that dates back at least to the 1950s.
如今流行的方法
And the methods that are popular
与数十年前发明的算法并无本质区别
today aren ’ t fundamentally different from algorithms invented
那么为何如今吸引了众人的关注以及投资呢？
decades ago,So why all the interest and investment right now?
我采访了Manuela Veloso
I asked Manuela Veloso,
卡耐基梅隆大学机器学习学院的院长
the head of the machine learning department at Carnegie Melon.
“你必须明白
“ You have to understand that there is something very important
近几年有一件非常重要的事
about these past years.
那就是数据
It’s data.
我们人类成为数据收集专家 健身数据
We humans became collectors of data. Fitbits,
GPS 图片等等
GPSes, pictures,
环顾周围有多少信用卡消费
I mean look how much credit card purchases, how much data
周围就有多少数据
is around.”
只要电脑有能够处理大数据的能力
Certain machine learning algorithms really thrive
某些机器学习算法就能像如今这样
on big data, as long as computers have
在大数据背景下蓬勃发展
the processing power to handle it, which theydo now.
如果电脑是大炮 网络是火药
If computers are the cannon and the internet is gunpowder,
那么人工智能就是烟花
these are the fireworks and
且这仅仅是序幕
they have only just begun.
Pedro Domingos在其著作中
In his book,
提供了一种简单地去理解
Pedro Domingos offers a nice simple way of understanding
有监督的机器学习方法
supervised machine learning.
他说“每一种算法都有输入和输出
He says:“Every algorithm has an input and an output:
数据被输入电脑
the data goes into the computer,
经由算法处理
the algorithm does what it will with it, and out comes the
最后得出结果
result.
然而机器学习的路径则相反
Machine learning turns this around:
输入进去的是数据及想要的结果
in goes the data and the desired result and out comes
训练出来的是能将输入转换为想要的结果的算法”
the algorithm that turns one into the other.”
算法为找到数据间的统计关系而被训练
The algorithms aretrained to find statistical relationships
从而使其在处理新实例时
in the data that allow it to make
也能做出较为精准的预测
good guesses when presented with new examples.
这意味着我们对电脑
That means we no longer have an easy rule
能做什么以及不能做什么
for what kinds of tasks computers can and
不再有简单的规则
cannot do.
“十年以前 我可以自信地说
“ Ten years ago, I could have said with confidence,
为了明白计算机化如何工作
we know how this works to computerize something
你必须理解每一个步骤
you need to understand all the steps,
然后你编程每一个步骤 并命令愚蠢的电脑工作
then you script the steps and get a dumb machine
这仅仅是机械地按照你设定的步骤一步步完成
to do it and just follow mechanistically the process that you would have followed.
但如今 我们的机器 我不应该说我们
But now we have machines, I shouldn’t say we,
我没有创造它们
I don’t make them.
人们已经研究出能从数据中学习的机器
People have developed machines that learnfrom data.
很难说哪些种类的工作会被
That makes it harder to say what set
自动化取代 或者被自动化轻而易举地取代
of jobs are going to become substituted, readily substituted
而哪些工作会增加新的岗位”
by automation, and which will be complemented.”
就这一问题麦肯锡全球研究所展开了一项研究
A study by the McKinsey Global Institute gets
调研了八百种不同的工作
at this question by looking at the many tasks
所需要完成的任务
that make up 800 different occupations.
并将这些任务分成了七大类
And they grouped those tasks into 7 categories:
有三类会随着如今不断发展的科技
3 that are highly susceptible to automation
而极有可能会自动化 而另外四类则不会
with currently-demonstrated technologies,and 4 that are not.
“像人事工作这样
“ Things like managing people,
包含了创造
they include things like creativity, they include things
决策或判断
like decision-making or judgment.
以及像需要共情 人际交流
And caring work that requires empathy or human interaction,
和情感支撑的
with an emotional content to
护理工作
associate with it.
都较难实现自动化”
Those are much harder things to automate.”
该研究得出的结论是
The report concluded
虽然大部分工作的部分内容能够被自动化
that while most jobs include some tasks that can be automated,
但不足百分之五的工作会完全自动化
less than 5% of occupations can be fully automated.
“所以 职业变化这个问题
“ So this idea
可能比职业消失这一问题
of occupations and jobs changing may actually be a bigger effect than the question
影响更大
of jobs disappearing, although of course,
诚然 有部分工作会消失
there are some jobs that will disappear or
或者至少会减少”
at least decline.”
这主要是因为大多数工作
That ’ s because most jobs are made up of a bunch
都需要执行多任务
of different tasks and most of today ’ s
而如今人工智能却只能完成某项任务
AI can only do one task.
不要误解
Don’t get me wrong.
人工智能在这项任务上十分出色
They can be really good at that task.
深度神经网络在观看了
A deep neural network watched 5000 hours of BBC news
带有标题的BBC新闻五千小时后
with captions and now it can read
其读唇语能力比专家更突出
lips \u00a0better than human professionals.
学习了肿瘤影像的机器学习算法
And machine learning algorithms trained on images
相较于病理学家
of tumors can predict lung cancer survival
能更准确地预测肺癌存活率
better than human pathologists.
实际上错误是
The mistake is to assume
去认为这些专项功能的应用
that these focused applications can add up to
能逐渐积累成为更趋向于人类大脑的智能
a more general intelligence.
或者认为它能像人类一般学习 但事实却并非如此
Or that they learn like we do, which is simply not the case.
当人工智能得出正确的结论后
When they get the right answer it ’
人们很容易认为它们理解所观测到的内容
s tempting to assume they understand what they see.
只有当它犯错时
Only when they make a mistake do we get a glimpse
我们才会窥测到
at how different their process is
它的理解过程与我们的有天壤之别
from our own.
这种模式性识别被误认为是理解
It’s pattern recognition masquerading asunderstanding.
这便是为什么研究人员很容易使得
That ’ s why researchers can easily trick a learning algorithm
一个学习算法错误地标记图片
into mislabeling a picture.
就这一点 众多机器学习
“ A lot of machine learning, at this point,
是非常表面且不稳定的
is very superficial and very brittle.
它基于瞬间可观测的特征
It’s based on immediately observable features,
这些特征对与当下进程
which may or may not be essential to what’s
可能也有可能不重要
going on.”
去年 导演奥斯卡•夏普制作了一部微电影
Last year the director Oscar Sharp produced a
该剧本是由
short film that was written by a neural
科幻电影剧本训练出的神经网络撰写的
network trained on sci-fi movie scripts.
“撰写原则完全是同时完成的”
“The principle is completely constructedof the same time.”
“关于你的一切都是真实的”
“It was all about you to be true.”
“你甚至完全不会发现
“ You didn ’ t even see the movie
还有别的内容”
with the rest of the base. ”
“我不知道” “我不关心” 太棒了
“I don’t know.” “I don’t care.” It’s great.
完全不符合逻辑
It makes no sense.
因为神经网络没有一个五岁小孩所拥有的
Because it doesn ’ t have what a 5-year-old child has,
关于世界如何运行
which is an abstract model of how
事情因何发生 或者什么是故事
the world works, why things happen, or whata story is.
以及为何如此的抽象理解
And why should it?
我们经数百万年进化出这些东西
We evolved these things over millions of years.
“人工智能可以做到很多事
“ So there’s a lot it can do,
相较于以前要多得多 但是我想说的是
much more than before but I mean, we humans are amazing,
人类太神奇了 看 我们人类是很丰富的”
I think. We are very broad, see.”
人工智能运用日益精进
AI applications will keep getting better.
过去机器人的声音是这样的
Robot voices used to sounds like this.
如今已经是这样的了
Now they can sound like this.
这意味着Wibbitz将能给视频提供较为真实的语音
Which means Wibbitz will so be able to offernatural-sounding narration.
算法也开始分析视频的结构
Algorithms are also starting to analyze videoframes.
IBM设计了一个系统为电影预告片选取场景
IBM trained a system to select the scenes for a movie trailer.
所以除了简单地生成一般性视频 Wibbitz可能能生成特制视频
So instead of just pulling generic clips, Wibbitz might pull specific ones.
但还没有明确路径
But there ’ s no clear path
通向更接近人类的人工智能 这包含常识
toward a more human-like intelligence which includes common sense,
好奇心 抽象推理
curiosity, and abstract reasoning.
“我认为人工智能
“I think AI is as good as the content that
与它学习内容同等优秀
goes through it.
所以人们不能真的期待人工智能
So you can ’ t really expect AI
可以如少数人期待的那样施展魔法”
to do magic which some people expect it to do. ”
机器学习算法可以翻译三十七种语言
Machine learning algorithms can translate 37
但它并不知道
languages but they don ’ t know what a
椅子的作用
chair is for.
它们与人类不同
They ’ re nothing like us,
这便是它成为有力工具的重要原因
and that ’ s what makes them such a powerful tool.
Wibbitz不会做出这个视频
Wibbitz will never make this video,
但人工智能可以帮我做得更好
but AI could help me make a better one.