Alex Wissner-Gross: A new equation for intelligence

翻译人员: Phillip Feng 校对人员: Zhiting Chen

Intelligence -- what is it?

智能--它是什么？

If we take a look back at the history

当我们回顾在历史上

of how intelligence has been viewed,

智能是如何被看待的，

one seminal example has been

一个开创性的例子是

Edsger Dijkstra's famous quote that

艾兹格•迪杰斯特拉的著名引述，

"the question of whether a machine can think

"关于一台机器能否思考的问题

is about as interesting

与关于

as the question of whether a submarine

一艘潜艇是否会游泳的问题

can swim."

几乎同样有趣"。

Now, Edsger Dijkstra, when he wrote this,

当艾兹格•迪杰斯特拉 写下这句话的时候，

intended it as a criticism

他的用意是去批判那些

of the early pioneers of computer science,

早年间开辟了计算机科学的先锋，

like Alan Turing.

比如阿兰 · 图灵。

However, if you take a look back

然而，如果你回顾过去

and think about what have been

并予以思考，有哪些

the most empowering innovations

最有利于发展的创新，

that enabled us to build

让我们有机会能够制造出

artificial machines that swim

会游泳的机器

and artificial machines that [fly],

和会[飞]的机器，

you find that it was only through understanding

你会发现，只有通过了解

the underlying physical mechanisms

游泳和飞行

of swimming and flight

背后的物理机制，

that we were able to build these machines.

我们才有能力去制造这些机器。

And so, several years ago,

所以说，在几年前，

I undertook a program to try to understand

我着手了一个项目， 试图去了解

the fundamental physical mechanisms

智能背后的

underlying intelligence.

基础物理机制。

Let's take a step back.

我们先退一步说。

Let's first begin with a thought experiment.

首先，让我们从一个思维实验开始。

Pretend that you're an alien race

假装你是一个外星人，

that doesn't know anything about Earth biology

你对地球上的生物学、

or Earth neuroscience or Earth intelligence,

神经科学和智能一无所知，

but you have amazing telescopes

但你有绝佳的望远镜，

and you're able to watch the Earth,

因此你能观望地球，

and you have amazingly long lives,

你的寿命也惊人地长，

so you're able to watch the Earth

所以你可以观察地球

over millions, even billions of years.

超过数百万年，甚至几十亿年。

And you observe a really strange effect.

然后你观察到一个很奇怪的现象。

You observe that, over the course of the millennia,

你观察到，几千年来，

Earth is continually bombarded with asteroids

地球不断地与小行星发生碰撞

up until a point,

直到某一刻，

and that at some point,

而在那一刻，

corresponding roughly to our year, 2000 AD,

大约对应的是公元2000年，

asteroids that are on

那些在地球撞击轨道

a collision course with the Earth

上的小行星，

that otherwise would have collided

本该相撞

mysteriously get deflected

但却被神秘地弹开了

or they detonate before they can hit the Earth.

或者在碰到地球之前就引爆了。

Now of course, as earthlings,

当然，作为地球人，

we know the reason would be

我们知道其中的原因是

that we're trying to save ourselves.

我们正试图自我拯救。

We're trying to prevent an impact.

我们要防止撞击发生。

But if you're an alien race

但如果你是一个外星人，

who doesn't know any of this,

对这些一无所知，

doesn't have any concept of Earth intelligence,

对地球上的智能也没有任何概念，

you'd be forced to put together

这就迫使你去总结

a physical theory that explains how,

一种物理理论， 去解释其原因，

up until a certain point in time,

直到在某一刻，

asteroids that would demolish the surface of a planet

本应摧毁一个星球表面的小行星，

mysteriously stop doing that.

神秘地停止了这种行为。

And so I claim that this is the same question

因此我声称这个问题

as understanding the physical nature of intelligence.

与理解智能的物理本质的问题 是相同的。

So in this program that I undertook several years ago,

因此，在我几年前着手的 这个项目中，

I looked at a variety of different threads

我研究了许多不同的线程，

across science, across a variety of disciplines,

跨越科学界，跨越多个学科，

that were pointing, I think,

在我看来，他们都指向

towards a single, underlying mechanism

一个统一的、潜在的

for intelligence.

智能机制。

In cosmology, for example,

例如在宇宙学中，

there have been a variety of different threads of evidence

就存在着各种各样的线索，

that our universe appears to be finely tuned

它们显示我们的宇宙就 为了智能的开发，

for the development of intelligence,

而被精准地调试过，

and, in particular, for the development

和特别是的对于发展

of universal states

世界各国

that maximize the diversity of possible futures.

去实现有最大多样化可能性的未来。

In game play, for example, in Go --

在棋牌界，举个例子，围棋--

everyone remembers in 1997

大家都记得在1997年的时候

when IBM's Deep Blue beat Garry Kasparov at chess --

IBM制作的机器人“深蓝“打败了 世界象棋冠军加里·卡斯帕罗夫--

fewer people are aware

很少有人意识到

that in the past 10 years or so,

在过去10年左右的时间里，

the game of Go,

围棋，

arguably a much more challenging game

可以说是一个更具挑战性的游戏，

because it has a much higher branching factor,

因为它具有更高的分支系数，

has also started to succumb

也已开始屈服于

to computer game players

电脑这个游戏对手，

for the same reason:

出于同样的原因：

the best techniques right now for computers playing Go

现在，电脑下围棋的 最佳技术方法

are techniques that try to maximize future options

是在下棋的过程中， 试图最大化

during game play.

未来的各种可能性。

Finally, in robotic motion planning,

最后，在机器人的运动规划中，

there have been a variety of recent techniques

有各种各样的新颖技术，

that have tried to take advantage

它们有试图利用

of abilities of robots to maximize

机器人的能力去将

future freedom of action

未来的行动自由最大化，

in order to accomplish complex tasks.

从而完成复杂的任务。

And so, taking all of these different threads

因此，考虑所有这些不同的线程

and putting them together,

并把它们放在一起，

I asked, starting several years ago,

从几年前开始我就在问，

is there an underlying mechanism for intelligence

有没有一种潜在的智能机制

that we can factor out

可以让我们分解出

of all of these different threads?

所有这些不同的线程？

Is there a single equation for intelligence?

是否存在一个 关于智能的公式？

And the answer, I believe, is yes. ["F = T ∇ Sτ"]

而我相信答案是有。 ["F = T ∇ SΤ"]

What you're seeing is probably

你看到的可能是

the closest equivalent to an E = mc²

我所见过最接近于 E = mc²

for intelligence that I've seen.

的智慧。

So what you're seeing here

所以你在这里看到的

is a statement of correspondence

是一张对应表，

that intelligence is a force, F,

其中智能是一种力量，F，

that acts so as to maximize future freedom of action.

它的作用是为了便于将未来的 行动自由最大化。

It acts to maximize future freedom of action,

它的作用是将未来的 行动自由最大化，

or keep options open,

或是保留灵活的选择权，

with some strength T,

与一种力量 T，

with the diversity of possible accessible futures, S,

和有多种可能性的、 可实现的未来，S，

up to some future time horizon, tau.

一直到某个未来的开始， tau（希腊字母）。

In short, intelligence doesn't like to get trapped.

简而言之，智能不喜欢被困住。

Intelligence tries to maximize future freedom of action

智能试图将未来的行动自由最大化，

and keep options open.

并保留选择权。

And so, given this one equation,

所以，鉴于这一公式，

it's natural to ask, so what can you do with this?

你自然会问， 那么这些可以让你做什么？

How predictive is it?

它是预测性有多高？

Does it predict human-level intelligence?

它能否预测人类的智能水平？

Does it predict artificial intelligence?

它能够预测人工智能吗？

So I'm going to show you now a video

因此，我将要展示给你们一段视频，

that will, I think, demonstrate

我认为，它会展示出

some of the amazing applications

单是这一个公式的

of just this single equation.

一些惊人的应用。

(Video) Narrator: Recent research in cosmology

（视频）讲述人： 宇宙学的最近研究

has suggested that universes that produce

反应了那些产生更多混乱、

more disorder, or "entropy," over their lifetimes

或者"熵"的宇宙， 在他们的生命中

should tend to have more favorable conditions

应该倾向于产生更多 有利的情况，

for the existence of intelligent beings such as ourselves.

让像我们这样的智慧生物 得以存在。

But what if that tentative cosmological connection

但假如那个在熵与智能之间

between entropy and intelligence

暂定的宇宙链接

hints at a deeper relationship?

暗示着更深层的关系呢？

What if intelligent behavior doesn't just correlate

如果智能的行为不仅只与

with the production of long-term entropy,

长期熵的生产相关，

but actually emerges directly from it?

而是直接由其产生的呢？

To find out, we developed a software engine

为了找到答案， 我们开发了一个软件引擎

called Entropica, designed to maximize

称为 Entropica， 设计的意图是将

the production of long-term entropy

长期熵的生产最大化，

of any system that it finds itself in.

无论它身在任何系统内。

Amazingly, Entropica was able to pass

惊人的是，Entropica 通过了

multiple animal intelligence tests, play human games,

多个动物的智能测验、 玩人类的游戏、

and even earn money trading stocks,

甚至在股票交易中赚钱，

all without being instructed to do so.

而且完全没有被给出那些指示。

Here are some examples of Entropica in action.

下面是一些 Entropica 的行动实例。

Just like a human standing upright without falling over,

就像人类站立不会跌到，

here we see Entropica

这里我们可以看到 Entropica

automatically balancing a pole using a cart.

自动地使用购物车去平衡棍子。

This behavior is remarkable in part

这种行为可以说是非常卓越的

because we never gave Entropica a goal.

因为我们从未给 Entropica 设定一个目标。

It simply decided on its own to balance the pole.

它自己就决定去平衡那根棍子。

This balancing ability will have appliactions

这种平衡能力将能应用于

for humanoid robotics

人形机器人

and human assistive technologies.

和人类的辅助科技。

Just as some animals can use objects

正如一些动物可以使用

in their environments as tools

环境中的物体作为工具

to reach into narrow spaces,

去伸入狭窄的空间，

here we see that Entropica,

这里我们可以看到 Entropica，

again on its own initiative,

同样是自主的，

was able to move a large disk representing an animal

能够移动一个表示动物的大圆盘

around so as to cause a small disk,

去把一个表示工具的小圆盘，

representing a tool, to reach into a confined space

去深入一个狭窄的空间，

holding a third disk

那里有第三个圆盘，

and release the third disk from its initially fixed position.

并把第三个圆盘从它初始 的静态解放出来.

This tool use ability will have applications

这种工具的使用能力将能运用于

for smart manufacturing and agriculture.

智能制造业和农业。

In addition, just as some other animals

此外，正如其他一些动物

are able to cooperate by pulling opposite ends of a rope

能够合作起来同时去拉 一根绳子的两端

at the same time to release food,

从而释放食物，

here we see that Entropica is able to accomplish

这里我们可以看到 Entropica 有能力完成

a model version of that task.

这项任务的模型版本。

This cooperative ability has interesting implications

这种合作能力能够带来有趣的影响，

for economic planning and a variety of other fields.

在经济规划和各种其他领域中。

Entropica is broadly applicable

Entropica 可以广泛适用于

to a variety of domains.

各种各样的领域。

For example, here we see it successfully

例如，在这里我们看到它成功的

playing a game of pong against itself,

与自己玩乒乓球游戏，

illustrating its potential for gaming.

说明其在游戏界的潜力。

Here we see Entropica orchestrating

在这里我们看到 Entropica 指挥着

new connections on a social network

社交网络上新的关系，

where friends are constantly falling out of touch

在这朋友们不断的失去联系

and successfully keeping the network well connected.

并成功地保持有效的网络连接。

This same network orchestration ability

这种相同的网络指挥能力

also has applications in health care,

在医疗保健、能源、和智能方面

energy, and intelligence.

都有相关的应用。

Here we see Entropica directing the paths

这里我们可以看到 Entropica

of a fleet of ships,

指挥一支舰队的路径，

successfully discovering and utilizing the Panama Canal

成功地发现并利用巴拿马运河，

to globally extend its reach from the Atlantic

然后将其范围从大西洋到太平洋

to the Pacific.

全球性地扩大。

By the same token, Entropica

同样的，Entropica

is broadly applicable to problems

可以广泛地适用于

in autonomous defense, logistics and transportation.

自主防御、 物流和运输地应用。

Finally, here we see Entropica

最后，在这里我们看到 Entropica

spontaneously discovering and executing

自主地发现和执行

a buy-low, sell-high strategy

一个低买高卖的策略，

on a simulated range traded stock,

这是在模拟的范围交易股票上，

successfully growing assets under management

它成功地将其管理的资产

exponentially.

成指数升涨。

This risk management ability

这种风险管理的能力

will have broad applications in finance

将在金融和保险领域

and insurance.

有广泛的应用。

Alex Wissner-Gross: So what you've just seen

阿历克斯•维斯纳-格罗斯： 你刚看到的

is that a variety of signature human intelligent

是各种具有代表性的人类智能

cognitive behaviors

的认知行为，

such as tool use and walking upright

例如工具的使用、直立行走

and social cooperation

和社会合作，

all follow from a single equation,

它们都遵循一个公式，

which drives a system

该公式所驱动的系统

to maximize its future freedom of action.

是要将其未来的行动自由最大化。

Now, there's a profound irony here.

现在，这里存在一个深刻的讽刺。

Going back to the beginning

回到最初，

of the usage of the term robot,

机器人这个术语的用法，

the play "RUR,"

"RUR，"这出戏，

there was always a concept

总存在一种概念就是

that if we developed machine intelligence,

如果我们开发了机器智能

there would be a cybernetic revolt.

就会产生一个人工智能的叛变。

The machines would rise up against us.

机器会奋起反抗我们。

One major consequence of this work

这项工作的主要成果之一

is that maybe all of these decades,

就是也许这几十年间，

we've had the whole concept of cybernetic revolt

我们对于人工智能的叛变 的整个概念

in reverse.

是颠倒的。

It's not that machines first become intelligent

机器不是先有了智慧

and then megalomaniacal

然后才变得狂妄

and try to take over the world.

并试图接管世界的。

It's quite the opposite,

其实几乎是相反的，

that the urge to take control

那种迫切的欲望，

of all possible futures

想要控制所有未来的所有可能

is a more fundamental principle

是比智能更基本的

than that of intelligence,

一个原则，

that general intelligence may in fact emerge

综合智能事实上可能是从

directly from this sort of control-grabbing,

这种控制欲中直接产生的，

rather than vice versa.

而不是反之。

Another important consequence is goal seeking.

另一个重要的成果是寻找目标。

I'm often asked, how does the ability to seek goals

我经常被问道， 寻找目标的能力

follow from this sort of framework?

怎么会遵循这种框架结构呢？

And the answer is, the ability to seek goals

答案是，寻找目标的能力

will follow directly from this

将直接遵循它，

in the following sense:

道理是这样的：

just like you would travel through a tunnel,

就像你要穿过一条隧道，

a bottleneck in your future path space,

你未来道路空间中的一个瓶颈，

in order to achieve many other

为了在以后实现许多

diverse objectives later on,

其他的各种目标，

or just like you would invest

或者就像你会投资

in a financial security,

于金融证券，

reducing your short-term liquidity

减少你的短期流动性

in order to increase your wealth over the long term,

从而长远的增加你的财富，

goal seeking emerges directly

目标的寻求直接涌现于

from a long-term drive

长期的驱动，

to increase future freedom of action.

为了增加未来的行动自由。

Finally, Richard Feynman, famous physicist,

最后，理查德 · 费曼， 这位著名的物理学家，

once wrote that if human civilization were destroyed

曾经写道， 如果人类文明被摧毁

and you could pass only a single concept

并且你只能将一个概念

on to our descendants

传承给我们的后代，

to help them rebuild civilization,

来帮助他们重建文明，

that concept should be

这个概念应该是

that all matter around us

我们身边的一切物质

is made out of tiny elements

都是由微小的元素组成的，

that attract each other when they're far apart

它们之间距离远的时候 会相互吸引，

but repel each other when they're close together.

但在靠的很近时 它们会互相排斥。

My equivalent of that statement

我与这句话等同的声明，

to pass on to descendants

来传递给后代，

to help them build artificial intelligences

帮助他们建立人工智能

or to help them understand human intelligence,

或是帮助他们理解 人类的智慧，

is the following:

是如下的话：

Intelligence should be viewed

智能应该被看作是

as a physical process

一个物理过程，

that tries to maximize future freedom of action

它试图将未来的行动自由最大化

and avoid constraints in its own future.

并且避免在自己的未来中的约束。

Thank you very much.

非常感谢。

(Applause)

（掌声）