Alex Wissner-Gross: A new equation for intelligence

譯者: Willy Feng 審譯者: Rowena Weng

Intelligence -- what is it?

智慧，是什麽？

If we take a look back at the history

如果我們回顧歷史

of how intelligence has been viewed,

對智慧的定義，

one seminal example has been

有一個基本的例子是，

Edsger Dijkstra's famous quote that

艾茲赫爾·戴克斯特拉說過的一句話： (註：著名電腦科學家)

"the question of whether a machine can think

“關於機械是否能思考的問題

is about as interesting

就有如在問

as the question of whether a submarine

潛水艇是否能游泳

can swim."

一樣有意思。”

Now, Edsger Dijkstra, when he wrote this,

當艾茲赫爾·戴克斯特拉寫下這句話，

intended it as a criticism

是在質疑

of the early pioneers of computer science,

早期的電腦科學先驅，

like Alan Turing.

譬如艾倫·圖靈。

However, if you take a look back

然而，如果你回顧

and think about what have been

並思考，

the most empowering innovations

是什麼重大的創新

that enabled us to build

使我們能夠製造出

artificial machines that swim

會游泳和會飛的

and artificial machines that [fly],

人造機器，

you find that it was only through understanding

你就會發現，

the underlying physical mechanisms

唯有透過了解

of swimming and flight

游泳和飛翔的基本物理機制，

that we were able to build these machines.

我們才能製造出這些機器。

And so, several years ago,

因此，幾年前，

I undertook a program to try to understand

我著手進行一個計劃，

the fundamental physical mechanisms

試圖去了解什麼是

underlying intelligence.

智慧的基本物理機制。

Let's take a step back.

先讓我們退一步，

Let's first begin with a thought experiment.

先從一個發想實驗開始。

Pretend that you're an alien race

假設你是一個外星人，

that doesn't know anything about Earth biology

對地球的生物完全不了解，

or Earth neuroscience or Earth intelligence,

也不了解地球的神經學和生物智慧，

but you have amazing telescopes

但你有很棒的望遠鏡，

and you're able to watch the Earth,

可以直接看到地球，

and you have amazingly long lives,

而且你有很長很長的壽命，

so you're able to watch the Earth

所以你有好幾百萬年甚至好幾十億年的時間

over millions, even billions of years.

來觀察地球。

And you observe a really strange effect.

你發現一個很怪異的事情。

You observe that, over the course of the millennia,

你發現，在千禧年這個過程中，

Earth is continually bombarded with asteroids

地球不斷地遭到小行星的撞擊，

up until a point,

直到某一天，

and that at some point,

在某一個時刻，

corresponding roughly to our year, 2000 AD,

大約就是我們現在的西元兩千年左右，

asteroids that are on

小行星原本運行在

a collision course with the Earth

會撞擊到地球的軌道上，

that otherwise would have collided

但是那個軌道

mysteriously get deflected

神奇地偏移了，

or they detonate before they can hit the Earth.

或者小行星在撞到地球前爆炸了。

Now of course, as earthlings,

當然，身為地球人，

we know the reason would be

我們知道那是因為

that we're trying to save ourselves.

我們試著拯救人類，

We're trying to prevent an impact.

試著避免撞擊發生。

But if you're an alien race

但如果你是外星人，

who doesn't know any of this,

不知道這些，

doesn't have any concept of Earth intelligence,

對地球上的智慧沒有任何概念，

you'd be forced to put together

那麼你只好勉強拼湊出一個

a physical theory that explains how,

物理理論來解釋，

up until a certain point in time,

直到某一個時刻，

asteroids that would demolish the surface of a planet

應該毀滅地表一切的小行星

mysteriously stop doing that.

神奇地不再發生。

And so I claim that this is the same question

而我認為這跟要了解

as understanding the physical nature of intelligence.

智慧的物理機制是一樣的問題。

So in this program that I undertook several years ago,

因此，在這項我幾年前開始進行的計劃中，

I looked at a variety of different threads

我研究各式各樣的想法，

across science, across a variety of disciplines,

橫跨科學以及不同領域，

that were pointing, I think,

我認為，

towards a single, underlying mechanism

這些都指向智慧的一個單一

for intelligence.

基本機制。

In cosmology, for example,

以宇宙論為例，

there have been a variety of different threads of evidence

有各種不同的證據顯示

that our universe appears to be finely tuned

我們所在的宇宙是被精心調整到

for the development of intelligence,

適合發展出智慧的，

and, in particular, for the development

尤其是發展出一個

of universal states

普遍性的狀態

that maximize the diversity of possible futures.

能使未來的可能性上做最大化。

In game play, for example, in Go --

以圍棋為例，

everyone remembers in 1997

大家都記得1997年

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

IBM 的深藍電腦打敗棋王卡斯帕羅夫，

fewer people are aware

但只有少數人知道

that in the past 10 years or so,

在過去的十年，

the game of Go,

圍棋，

arguably a much more challenging game

被視為是非常具挑戰性的遊戲，

because it has a much higher branching factor,

因為它有更多的分歧因素，

has also started to succumb

同時也開始讓

to computer game players

電腦玩家臣服，

for the same reason:

這些都是同樣的理由：

the best techniques right now for computers playing Go

現在讓電腦下棋最好的技巧

are techniques that try to maximize future options

就是將下棋過程可能發生的事件數

during game play.

最大化。

Finally, in robotic motion planning,

最後，在機器人的行動規劃中，

there have been a variety of recent techniques

最近的各種技術

that have tried to take advantage

都是試圖讓機器人

of abilities of robots to maximize

在未來能自由行動的可能性

future freedom of action

做最大化，

in order to accomplish complex tasks.

以完成某些複雜的任務。

And so, taking all of these different threads

所以，用這些不同的想法，

and putting them together,

把它們拼湊在一起，

I asked, starting several years ago,

在幾年前我開始問，

is there an underlying mechanism for intelligence

有沒有一個關於智慧的基本機制

that we can factor out

是我們可以從這些不同的想法中

of all of these different threads?

分解出來的？

Is there a single equation for intelligence?

有沒有一個屬於智慧的方程式？

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

我相信答案是，有的。 ["F = T ∇ Sτ"]

What you're seeing is probably

你現在看到的

the closest equivalent to an E = mc²

或許是我看過最接近 E = mc²

for intelligence that I've seen.

的屬於智慧的方程式。

So what you're seeing here

你所看到的

is a statement of correspondence

是相對應的詮釋，

that intelligence is a force, F,

智慧是一種力量，F

that acts so as to maximize future freedom of action.

它的作用是最大化行動的自由度。

It acts to maximize future freedom of action,

它的作用會最大化行動的自由度

or keep options open,

或是一直保有開放的選擇，

with some strength T,

配合某一強度 T，

with the diversity of possible accessible futures, S,

和可能發生的未來多樣性，S

up to some future time horizon, tau.

直到未來的某一個時間點，t。

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

簡單地說，智慧不喜歡被約束住。

Intelligence tries to maximize future freedom of action

智慧希望最大化未來行動的自由度，

and keep options open.

保持開放的選項。

And so, given this one equation,

所以，有了這一個方程式，

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

很自然地就會問，你能用它做甚麼？

How predictive is it?

它的預測能力如何？

Does it predict human-level intelligence?

它能否預測人類的智慧？

Does it predict artificial intelligence?

它能否預測人工智慧？

So I'm going to show you now a video

現在我要給各位看一段影片，

that will, I think, demonstrate

我認為可以說明

some of the amazing applications

一些令人驚訝的應用，

of just this single equation.

而且都只來自這一個方程式。

(Video) Narrator: Recent research in cosmology

(影片) 旁白：宇宙學最近的研究

has suggested that universes that produce

推論宇宙會產生愈來愈多的

more disorder, or "entropy," over their lifetimes

失序，或是熵 (entropy)，

should tend to have more favorable conditions

應該更容易擁有有利的環境，

for the existence of intelligent beings such as ourselves.

讓智慧存在。

But what if that tentative cosmological connection

但如果把這個宇宙學待驗證的

between entropy and intelligence

亂度和智慧的關係

hints at a deeper relationship?

再進一步加深會怎樣？

What if intelligent behavior doesn't just correlate

如果智慧和長期亂度的增加

with the production of long-term entropy,

不只是有正相關性，

but actually emerges directly from it?

而且是從中發展出來的呢？

To find out, we developed a software engine

為了解答這問題，我們開發了一個軟體

called Entropica, designed to maximize

叫做 "Entropica"，

the production of long-term entropy

可以把任何系統中

of any system that it finds itself in.

熵的長期成長最大化。

Amazingly, Entropica was able to pass

令人驚訝的是，Entropica 能夠通過

multiple animal intelligence tests, play human games,

多項動物智慧測試，玩人類的遊戲，

and even earn money trading stocks,

甚至從股票交易中賺到錢，

all without being instructed to do so.

而且事前完全不用去教導它。

Here are some examples of Entropica in action.

這裡有幾個 Entropica 的實例。

Just like a human standing upright without falling over,

像人可以直立站著不會跌倒，

here we see Entropica

我們可以看到，

automatically balancing a pole using a cart.

Entropica使用一台車來自動平衡桿子。

This behavior is remarkable in part

這個表現在某方面很了不起，

because we never gave Entropica a goal.

因為我們從來沒有為Entropica設定一個目標。

It simply decided on its own to balance the pole.

由它自己決定要去平衡這個桿子。

This balancing ability will have appliactions

這個平衡的能力可以應用在

for humanoid robotics

機器人上，

and human assistive technologies.

以及人類行動輔助技術。

Just as some animals can use objects

就像有些動物

in their environments as tools

會使用週遭的物品當作工具，

to reach into narrow spaces,

以便能伸及到窄小的地方，

here we see that Entropica,

我們可以再次看到 Entropica

again on its own initiative,

由它自己決定，

was able to move a large disk representing an animal

可以移動代表動物的大圓圈，

around so as to cause a small disk,

讓代表工具的小圓圈

representing a tool, to reach into a confined space

進入一個有第三個圓圈的

holding a third disk

狹小空間，

and release the third disk from its initially fixed position.

然後把第三個圓圈從裡面擠出來。

This tool use ability will have applications

這個使用工具的能力可以應用在

for smart manufacturing and agriculture.

智慧製造和農業上。

In addition, just as some other animals

另外，就像其它動物

are able to cooperate by pulling opposite ends of a rope

會同時合力拉下繩索的兩端，

at the same time to release food,

讓食物掉出來，

here we see that Entropica is able to accomplish

我們看到 Entropica 可以完成

a model version of that task.

模組化後的同樣任務。

This cooperative ability has interesting implications

這個合作的能力可以應用在

for economic planning and a variety of other fields.

經濟規劃和其它各樣的領域。

Entropica is broadly applicable

Entropica 可以廣泛的應用在

to a variety of domains.

各樣的領域。

For example, here we see it successfully

例如，我們可以看到它

playing a game of pong against itself,

成功地和自己玩 "乓" (Pong)，

illustrating its potential for gaming.

代表它能玩遊戲的潛力。

Here we see Entropica orchestrating

我們看到 Entropica 精心地

new connections on a social network

建立起社群的新連結，

where friends are constantly falling out of touch

當朋友們不時地失去聯繫，

and successfully keeping the network well connected.

它會成功地維持這個網絡。

This same network orchestration ability

這樣的網絡連結能力

also has applications in health care,

同樣可以應用在醫療照顧，

energy, and intelligence.

能源和智慧發展上。

Here we see Entropica directing the paths

這裡我們看到 Entropica

of a fleet of ships,

為海洋中的船隊指引路徑，

successfully discovering and utilizing the Panama Canal

成功地發現並使用巴拿馬運河，

to globally extend its reach from the Atlantic

使它的足跡遍及全球每個角落，從大西洋

to the Pacific.

到太平洋。

By the same token, Entropica

同樣的，Entropica

is broadly applicable to problems

可以廣泛地應用在

in autonomous defense, logistics and transportation.

自主防衛和物流運輸上。

Finally, here we see Entropica

最後，我們看到 Entropica

spontaneously discovering and executing

自己發現並且執行

a buy-low, sell-high strategy

"低買高賣"的策略，

on a simulated range traded stock,

在一個區間交易的股票模擬市場中，

successfully growing assets under management

成功地將管理資產規模

exponentially.

指數性成長。

This risk management ability

這樣的風險管理能力

will have broad applications in finance

可以應用在財務

and insurance.

和保險上。

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

艾力克斯·威斯奈-格羅斯： 以上你們所看到的

is that a variety of signature human intelligent

是一個代表人類智慧的

cognitive behaviors

認知行為能力，

such as tool use and walking upright

像是工具的使用、直立行走、

and social cooperation

以及群體合作，

all follow from a single equation,

全部都遵行一個方程式，

which drives a system

這個方程式驅使一個系統

to maximize its future freedom of action.

可以最大化未來行動的自由。

Now, there's a profound irony here.

然而，有一個很大的諷刺是，

Going back to the beginning

回顧最初

of the usage of the term robot,

使用”機器人”這個名詞時，

the play "RUR,"

在舞台劇《羅梭的萬能工人》(R.U.R,) 中，

there was always a concept

一直有一個概念：

that if we developed machine intelligence,

如果我們發展了人工智慧，

there would be a cybernetic revolt.

機器人將會起義反抗，

The machines would rise up against us.

對抗我們人類。

One major consequence of this work

我們這個研究主要的結論之一是，

is that maybe all of these decades,

或許在過去這幾十年來，

we've had the whole concept of cybernetic revolt

我們在逆向思考"機器人反抗”

in reverse.

這個概念。

It's not that machines first become intelligent

並不是機器先變聰明，

and then megalomaniacal

然後自大，

and try to take over the world.

然後才企圖統治全世界，

It's quite the opposite,

而是應該反過來看，

that the urge to take control

想要控制所有未來可能性

of all possible futures

的慾望，

is a more fundamental principle

比控制智慧

than that of intelligence,

是更加基本的原則，

that general intelligence may in fact emerge

一般的智慧或許是

directly from this sort of control-grabbing,

直接從操控中產生的，

rather than vice versa.

並非反過來。

Another important consequence is goal seeking.

另一個重要的結論是尋找目標。

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

我經常被問到，尋找目標的能力

follow from this sort of framework?

是如何從這個架構中產生的？

And the answer is, the ability to seek goals

答案是，尋找目標的能力

will follow directly from this

會直接來自於

in the following sense:

以下這個想法：

just like you would travel through a tunnel,

就像你行經一個隧道，

a bottleneck in your future path space,

一個在你未來道路上的瓶頸，

in order to achieve many other

是為了到達許多

diverse objectives later on,

在未來的不同目的地，

or just like you would invest

或者，就像你在證券上的

in a financial security,

投資，

reducing your short-term liquidity

降低短期的流動性，

in order to increase your wealth over the long term,

是為了增加長期的財富，

goal seeking emerges directly

而尋找目標是來自於

from a long-term drive

一個長期的趨動力

to increase future freedom of action.

用來增加未來的行動自由。

Finally, Richard Feynman, famous physicist,

最後，知名的物理學家理察費曼曾說，

once wrote that if human civilization were destroyed

如果人類文明要被毀滅了，

and you could pass only a single concept

而你只能留下一個概念

on to our descendants

給後世的子孫，

to help them rebuild civilization,

以便協助他們重建文明，

that concept should be

那麼這個概念應該是：

that all matter around us

所有我們週遭的物質

is made out of tiny elements

是是由微小的元素組成，

that attract each other when they're far apart

當它們相隔很遠時會互相吸引，

but repel each other when they're close together.

但靠近時會互相排斥。

My equivalent of that statement

而我同樣要

to pass on to descendants

留給後世的想法

to help them build artificial intelligences

以便幫助他們發展人工智慧，

or to help them understand human intelligence,

或是幫助他們了解人類的智慧，

is the following:

我會說：

Intelligence should be viewed

智慧應該被視為

as a physical process

一個物理程序，

that tries to maximize future freedom of action

它將試著最大化未來的行動自由，

and avoid constraints in its own future.

避免將自己侷限住。

Thank you very much.

謝謝大家。

(Applause)

(掌聲)