Intelligence -- what is it?
If we take a look back at the history
of how intelligence has been viewed,
one seminal example has been
Edsger Dijkstra's famous quote that
"the question of whether a machine can think
is about as interesting
as the question of whether a submarine
can swim."
Now, Edsger Dijkstra, when he wrote this,
intended it as a criticism
of the early pioneers of computer science,
like Alan Turing.
However, if you take a look back
and think about what have been
the most empowering innovations
that enabled us to build
artificial machines that swim
and artificial machines that [fly],
you find that it was only through understanding
the underlying physical mechanisms
of swimming and flight
that we were able to build these machines.
And so, several years ago,
I undertook a program to try to understand
the fundamental physical mechanisms
underlying intelligence.
Let's take a step back.
Let's first begin with a thought experiment.
Pretend that you're an alien race
that doesn't know anything about Earth biology
or Earth neuroscience or Earth intelligence,
but you have amazing telescopes
and you're able to watch the Earth,
and you have amazingly long lives,
so you're able to watch the Earth
over millions, even billions of years.
And you observe a really strange effect.
You observe that, over the course of the millennia,
Earth is continually bombarded with asteroids
up until a point,
and that at some point,
corresponding roughly to our year, 2000 AD,
asteroids that are on
a collision course with the Earth
that otherwise would have collided
mysteriously get deflected
or they detonate before they can hit the Earth.
Now of course, as earthlings,
we know the reason would be
that we're trying to save ourselves.
We're trying to prevent an impact.
But if you're an alien race
who doesn't know any of this,
doesn't have any concept of Earth intelligence,
you'd be forced to put together
a physical theory that explains how,
up until a certain point in time,
asteroids that would demolish the surface of a planet
mysteriously stop doing that.
And so I claim that this is the same question
as understanding the physical nature of intelligence.
So in this program that I
undertook several years ago,
I looked at a variety of different threads
across science, across a variety of disciplines,
that were pointing, I think,
towards a single, underlying mechanism
for intelligence.
In cosmology, for example,
there have been a variety of
different threads of evidence
that our universe appears to be finely tuned
for the development of intelligence,
and, in particular, for the development
of universal states
that maximize the diversity of possible futures.
In game play, for example, in Go --
everyone remembers in 1997
when IBM's Deep Blue beat
Garry Kasparov at chess --
fewer people are aware
that in the past 10 years or so,
the game of Go,
arguably a much more challenging game
because it has a much higher branching factor,
has also started to succumb
to computer game players
for the same reason:
the best techniques right now
for computers playing Go
are techniques that try to maximize future options
during game play.
Finally, in robotic motion planning,
there have been a variety of recent techniques
that have tried to take advantage
of abilities of robots to maximize
future freedom of action
in order to accomplish complex tasks.
And so, taking all of these different threads
and putting them together,
I asked, starting several years ago,
is there an underlying mechanism for intelligence
that we can factor out
of all of these different threads?
Is there a single equation for intelligence?
And the answer, I believe, is yes.
["F = T ∇ Sτ"]
What you're seeing is probably
the closest equivalent to an E = mc²
for intelligence that I've seen.
So what you're seeing here
is a statement of correspondence
that intelligence is a force, F,
that acts so as to maximize future freedom of action.
It acts to maximize future freedom of action,
or keep options open,
with some strength T,
with the diversity of possible accessible futures, S,
up to some future time horizon, tau.
In short, intelligence doesn't like to get trapped.
Intelligence tries to maximize
future freedom of action
and keep options open.
And so, given this one equation,
it's natural to ask, so what can you do with this?
How predictive is it?
Does it predict human-level intelligence?
Does it predict artificial intelligence?
So I'm going to show you now a video
that will, I think, demonstrate
some of the amazing applications
of just this single equation.
(Video) Narrator: Recent research in cosmology
has suggested that universes that produce
more disorder, or "entropy," over their lifetimes
should tend to have more favorable conditions
for the existence of intelligent
beings such as ourselves.
But what if that tentative cosmological connection
between entropy and intelligence
hints at a deeper relationship?
What if intelligent behavior doesn't just correlate
with the production of long-term entropy,
but actually emerges directly from it?
To find out, we developed a software engine
called Entropica, designed to maximize
the production of long-term entropy
of any system that it finds itself in.
Amazingly, Entropica was able to pass
multiple animal intelligence
tests, play human games,
and even earn money trading stocks,
all without being instructed to do so.
Here are some examples of Entropica in action.
Just like a human standing
upright without falling over,
here we see Entropica
automatically balancing a pole using a cart.
This behavior is remarkable in part
because we never gave Entropica a goal.
It simply decided on its own to balance the pole.
This balancing ability will have appliactions
for humanoid robotics
and human assistive technologies.
Just as some animals can use objects
in their environments as tools
to reach into narrow spaces,
here we see that Entropica,
again on its own initiative,
was able to move a large
disk representing an animal
around so as to cause a small disk,
representing a tool, to reach into a confined space
holding a third disk
and release the third disk
from its initially fixed position.
This tool use ability will have applications
for smart manufacturing and agriculture.
In addition, just as some other animals
are able to cooperate by pulling
opposite ends of a rope
at the same time to release food,
here we see that Entropica is able to accomplish
a model version of that task.
This cooperative ability has interesting implications
for economic planning and a variety of other fields.
Entropica is broadly applicable
to a variety of domains.
For example, here we see it successfully
playing a game of pong against itself,
illustrating its potential for gaming.
Here we see Entropica orchestrating
new connections on a social network
where friends are constantly falling out of touch
and successfully keeping
the network well connected.
This same network orchestration ability
also has applications in health care,
energy, and intelligence.
Here we see Entropica directing the paths
of a fleet of ships,
successfully discovering and
utilizing the Panama Canal
to globally extend its reach from the Atlantic
to the Pacific.
By the same token, Entropica
is broadly applicable to problems
in autonomous defense, logistics and transportation.
Finally, here we see Entropica
spontaneously discovering and executing
a buy-low, sell-high strategy
on a simulated range traded stock,
successfully growing assets under management
exponentially.
This risk management ability
will have broad applications in finance
and insurance.
Alex Wissner-Gross: So what you've just seen
is that a variety of signature human intelligent
cognitive behaviors
such as tool use and walking upright
and social cooperation
all follow from a single equation,
which drives a system
to maximize its future freedom of action.
Now, there's a profound irony here.
Going back to the beginning
of the usage of the term robot,
the play "RUR,"
there was always a concept
that if we developed machine intelligence,
there would be a cybernetic revolt.
The machines would rise up against us.
One major consequence of this work
is that maybe all of these decades,
we've had the whole concept of cybernetic revolt
in reverse.
It's not that machines first become intelligent
and then megalomaniacal
and try to take over the world.
It's quite the opposite,
that the urge to take control
of all possible futures
is a more fundamental principle
than that of intelligence,
that general intelligence may in fact emerge
directly from this sort of control-grabbing,
rather than vice versa.
Another important consequence is goal seeking.
I'm often asked, how does the ability to seek goals
follow from this sort of framework?
And the answer is, the ability to seek goals
will follow directly from this
in the following sense:
just like you would travel through a tunnel,
a bottleneck in your future path space,
in order to achieve many other
diverse objectives later on,
or just like you would invest
in a financial security,
reducing your short-term liquidity
in order to increase your wealth over the long term,
goal seeking emerges directly
from a long-term drive
to increase future freedom of action.
Finally, Richard Feynman, famous physicist,
once wrote that if human civilization were destroyed
and you could pass only a single concept
on to our descendants
to help them rebuild civilization,
that concept should be
that all matter around us
is made out of tiny elements
that attract each other when they're far apart
but repel each other when they're close together.
My equivalent of that statement
to pass on to descendants
to help them build artificial intelligences
or to help them understand human intelligence,
is the following:
Intelligence should be viewed
as a physical process
that tries to maximize future freedom of action
and avoid constraints in its own future.
Thank you very much.
(Applause)