POI幕后访谈

Harold 这就是撒玛利亚人
Harold, meet Samaritan.
人工智能 你总说这是不可能的
A.I. Well, you always said it was impossible.
而你总是对的
And you were always right about everything.
大家好 欢迎来到《疑犯追踪》第三季结尾
Hello. Welcome to Person of lnterest, the end of season three.
今天和我们一起的是史蒂夫•奥莫杭德罗
We have with us today Steve Omohundro,
史蒂夫•奥莫杭德罗
科学家/作家
当然了 还有乔纳森诺兰 《疑犯追踪》的主创
and obviously Jonah Nolan, the creator of Person of lnterest.
我们今天所讨论的主题
And we thought it'd be fantastic to talk today
同时也是剧中正在讲述的主题 是有关人工智能的
about an ongoing theme on our show of artificial intelligence.
这一直是我们剧集中的元素之一
It's always been an element of the show
例如Harold Finch的机器
in terms of Harold Finch's machine.
我们能看到 在第三季末尾
Now what we've seen, of course, at the end of season three
出现了第二个AI
is the advent of a second A.I.
格雷格·普拉格曼
执行制片人
这就为我们的剧集增添了一些问题
So it is becoming an issue of proliferation on our show.
没错
Yeah.
这也是我们想和你探讨的问题
Something we'd love to talk to you about
听一听你的见解
and, and get your opinion on as well.
我喜欢这部剧 非常非常有意思
I'm loving the show. Very, very interesting.
我们迎来了一些关于AI的有趣问题
We're getting into some juicy A.I. issues,
AI之战
a battle of the A.I.s
没错 -是的
Yeah. - Yeah.
乔纳森•诺兰
执行制片人
史蒂夫 从学术上来说 不是无聊的那种
Steve, on, on a sort of pedantic, not to get too, too prosaic with it, but,
你会怎样定义人工智能
how would, how would you define artificial intelligence?
对于你来说 AI的定义是怎样的 -对我个人来说
For you, what would be a definition of A.I.? - For me personally,
我认为智能就是
I think of intelligence
运用有限资源来解决问题的能力
as just the ability to solve problems using limited resources.
因此 根据这个定义
And so, by that definition,
我认为病毒在某种程度上是有智力的
I think viruses have a kind of intelligence.
细菌当然也有
Bacteria certainly do.
猫狗 毋庸置疑
Dog and cats, oh, my god, yes.
人类 更不用说了
And humans, certainly.
所以根据这个定义
And so by that definition,
如今的机器可以说是某种程度的智能
today's machines are a certain level of intelligence.
但它们还没有达到
They haven't reached a kind of, you know,
能够自我提升的那种智能
recursively improving level of intelligence.
没错 -但是
Right. - But,
它们不能自己变聪明
They can't make themselves smarter.
它们可以 但只能按照程序员预先设定的那样
They can, but only in ways that are very predefined by the programmer.
所以我们有许多机器学习系统
So we have lots of machine learning systems.
那些为你的Nerflix推荐电影的算法
The algorithms that choose your Netflix movie
会通过你的反馈来自我调整
are changing themselves based on your responses.
这是学习的一种形式
So that's a form of learning.
但让它们学习的算法是没有变的
The learning algorithm they're using isn't changing.
程序员编了那些代码
The programmer wrote that.
它得到了这一个学习算法
It's got this one learning algorithm.
它会获取数据 做它该做的事
It's taking its data, and it's doing what it's doing.
下一代系统 也是我所研究的一类系统
The next generation, sort of one class of systems that I work on,
它们会自我模仿并进行自我改变
systems actually model themselves and change themselves.
这时不确定性开始增加
That's when things start getting a little more dicey,
因为本来一个系统要干什么
because today it's pretty predictable
如果你编好程序的话
what a system's gonna do,
是完全可预测的
you know, once you've programmed it.
但一旦到了它们可以自我衍变的程度
Once you get to the level where they can really change themselves,
它们就会开始自己做决定
now all of a sudden they may start making decisions that
连最初的程序员都不知道它们要干什么
the original programmer had no idea what's it gonna do.
这可能会赋予它难以置信的强大潜能
And that's where it's both incredibly potentially powerful
但也可能会使它变得非常危险
and where it can get really dangerous.
911之后 政♥府♥发动所有的力量
After 911, the government had everyone and their cat
想造一个监视一切的系统
trying to build a system that could watch everything.
他们也找上你了
They came to you as well.
是的 我告诉他们 要想如愿
Yes. I told them the only way they could achieve what they wanted
所造的系统要不仅能监视一切
was to build something that not only watched,
还要能理解
but understood what it was watching.
人工智能
An A.I.
我认为将AI区别开来的
For me, I guess the distinction for A.I. has always been
杰•沃思
特效总监
是理性独♥立♥的思考与程序化响应之间的对抗
rational independent thought versus programmed response.
那个阈值点会被超过
That threshold that gets crossed, versus,
是既定的程序
programming
与不断进化的理性思考之间的对抗 -自反性的 没错
versus evolving rational thought. - Reflexive. Yeah.
马尔科姆·麦克弗
机械工程学 生物医学工程学教授 西北大学(美)
是为了能够义无反顾地执行任务
And being able to set up a parameter
而为自己设定限制条件
that then it's going to execute that regardless of obstacles.
所以这一实体被编程设定了某一目标
So an entity that is programmed to have a goal.
如果这个目标很难达成 它就会适应性地改变
And if the, if the goal is frustrated, then it comes up adaptively with,
适应 没错
Adapt, yeah.
但我们现在投入使用的机器人也可以那么做
But we, we have in the field a robot that does that now.
我就见过它们进行自我调整 那有点诡异
And I've seen it work. And it, it is creepy.
如果你把一个机器人的腿卸下来 它会马上
If you pull a leg off the robot, and it figures out immediately
用另外的方式走路
how to start walking a different way,
仍坚持它的目标
and still get to its goal,
之后你就会把这种情况归结于某些原则
Then you're gonna attribute it to some agencies,
某种能力 来让它自我驱动
some sort of ability to sort of, drive itself.
大卫·斯莱克
联合执行制片人
你觉得是它自己想要那样吗 -你确实会有那种感觉
Do you feel like it wants that? - You do have that feeling.
你绝对会有 这很吓人
You absolutely do. It's eerie.
没错 -尤其在你尝试阻止它的时候
Yeah. - Especially if you try and stop it.
是的 它会绕着你转 -就是这样
Yeah. Well, it's, it goes around you. - Exactly.
你会觉得 "它一定是有着意志的"
You go, "Well, it must have a will."
阿曼达•西格尔
联合执行制片人
没错 -它一定有着活下去的意志
Yes. - It must have a will to survive.
它一定有着达成目标的意志
It must have a will to, to its goal.
我们是给机器编程 让它实现某一目标
Well, is there, is there any substantive difference
还是说机器没有任何本能
between being programmed to achieve a goal,
所做一切都是由于程序
and the machine isn't having instincts,
这二者真的有区别吗
which are really programming?
我怀疑没有区别
My strong suspicion is no.
撒玛利亚人是革命性的
Samaritan is revolutionary,
不只是因为它能做什么 而是它如何做
Not so much for what it does, but how it does it.
我将系统设计成自主控制
The system was designed to function autonomously,
通过过去的经验来自我优化与升级
to continually adapt and upgrade itself through past experiences.
你是说它能记忆与学习
You're saying that it can remember and learn.
当这些系统可以在我们没有做出明确指令的情况下
Where you get really powerful systems, you begin to get these A.I. systems,
自己做出决定 采取行动
is where they can make--take actions and make decisions
这时就可以说它们是强大的AI系统了
that I didn't tell it exactly how to do that.
怎样做到那样 首先 给它设定一个目标
How do you do that? Well, you give it a goal.
然后让系统考虑不同的执行方式 看它们
And you have the system consider different actions and see how they,
是否能够达成目标
whether they meet or don't meet their goal.
如果可以 就执行最佳方案
And if it meets it, you take the action that meets it the best.
所以正如我在TED里面举的例子
And so in the example I did in that TED Talk,
不需要构建自我保护机制
I didn't build, the idea was you don't have to build in self-preservation.
也不需要构建资源获取机制
You don't have to build in resource acquisition.
只需给它设定一个简单的目标 这些都会自动生成
Those emerge automatically if you give it a simple goal.
在那时我举的是下棋机器人的例子 -没错
So in that case I used the example of a chess playing robot. - Sure.
它唯一的目标就是下好棋
So its one and only goal in life is to play good chess.
想象一下在某个世界 你将那个机器人关掉了
And imagine a world in which you turn that robot off.
它不再下棋了
It's not playing any chess.
从给它设定的"下棋"的目标来看
So from the perspective of the chess goal,
这简直太恐怖了 这是一个最差的世界
that's a terrible world. That's the worst world in the world.
之后它就进化出了一个子目标
And so it develops a subgoal:
不要让自己被关掉
Keep yourself from being turned off,
即使之前你没有这个设定
even though you didn't build that into it.
所以有些可怕的就是
And so that's the piece that's a little bit scary
对于你能想象到的任何一个目标
is that for almost any goal you can imagine,
除非你非常清楚地想明白
unless you very explicitly think through,
不让它干什么
"oh, what do we not want this thing to do?"
否则它就会产生一种 我称之为"AI内驱力"的东西
it's gonna have a certain thing I call the A.I. drives.
这种力量会驱使它进行自我保护
So, it's gonna have a drive toward self-preservation,