we don't want machines to be self-aware. In my opinion, everybody should understand that

this technology is around the corner. Your children, your grandchildren are going to

be living in a world where there are machines  that are on par and possibly exceed human

self-awareness and what does that mean? We’ll have to figure that out. Inside his Creative

Machines Lab at Columbia University, Hod Lipson and his students are taking the first steps

towards sentient machines. Scattered on shelves and tucked into cubicles, their robot menagerie

is probing the biggest philosophical questions that face robotics and AI today. As these

deceivingly simple machines squirm and babble, they’re crawling towards a future where

machines can autonomously adapt to and understand their surroundings, embarking on a quest to

unlock artificial consciousness. We’ve been making huge advances in robotics for decades,

but engineers investigating these philosophical questions, has not always been a widely supported

pursuit. For many years, this whole area of consciousness, self-awareness, sentience,

emotions, was taboo. Academia tended to stay away from these grand claims. But I think

now we're at a turning point in history of AI where we can suddenly do things that were

thought impossible just five years ago. We humans cannot babysit all these incredibly

sophisticated systems that we're creating. We need to gift them with this ability to

see themselves. To do that, Hod’s lab is taking a unique approach. Instead of trying

to understand and replicate something as complicated as human consciousness, they’re looking

into something a little simpler: the ability to imagine yourself– or self-simulate. So,

the big question is what is self awareness, right? We have a very simple definition, and

our definition is that self awareness is nothing but the ability to self simulate. A dog might

be able to simulate itself into the afternoon. If it can see itself into the future, it can

see itself having its next meal. Now if you can simulate yourself, you can imagine yourself

into the future, you're self-aware. With that definition, we can build it into machines. The

first idea to use the arm really came about because arms are one of the most common types

of robots. And they're also fairly simple. We have the robot start by flailing around.

We call this motor babbling. It just moves just like child in a crib. After moving for one

day it collected enough information about its body, to understand the dynamics of its

motion, how when it moves, what's going to happen to its position and its ability to

move in the future. The team fed all of that data into a deep learning model, allowing

the robot to predict its next position based on its previous one. Essentially, it’s using

the model it created to “imagine” its own motion. From there we were able to use

this model to plan how to effectively accomplish tasks. In this case the robot is sort of

going through a variety of different actions that it could attempt to make. Testing them

on the self model and then executing the one which will bring it closest to its goal state.

So in this case, the goal state is right where it is right now, on the ball. And then it

will move, as its next goal state, to the cup where it will then drop it. It's a little

bit tricky, because you look at this robotic arm and you'll see it doing its task and you'll

think, "Oh, I could probably program this arm to do this task by myself. It's not a

big deal," but you have to remember not only did the robot learn how to do this by itself,

but it's particularly important that it learned inside the simulation that it created. To

demonstrate the transferability, we made the arm write us a message. We told it to write

hi and it wrote hi with no additional training, no additional information needed. We just

used our self model and wrote up a new objective for it and it successfully executed. We call

that zero-shot learning. We humans are terrific at doing that thing. I can show you a tree

you've never climbed before. You look at it, you think a little bit and, bam, you climb

the tree. The same thing happens with the robot. The next steps for us are really working

towards bigger and more complicated robots. Walking robots are much more of a dynamic

and complex system and as such we're actually going to start developing on that robot over

there. It will do the motor babbling and we hope that we can learn a variety of tasks

on that the same way that we did on the arm robot. We believe that while this is a small,

simple start, by progressively improving the model and the bodies and the capacity of these

robots, we can push the border between robot and human to become smaller and smaller with

each additional iteration. Every automated system would benefit from being self-aware.

Everything from driverless cars to airplanes to smart cities to traffic lights. Imagine

that your factory is operating. It's creating, I don't know what, solar panels, and it understands

if things are working well. It can foresee a problem on the horizon and it's going to

take measures to compensate or recover from that. All of that is incredibly useful no

matter what you're doing. But regardless of the benefits, the idea of self-aware machines

will inevitably raise concerns. Experts from the likes of Stephen Hawking to Elon Musk

have warned about the dangers of advancing AI, cautioning against a future where it progresses

to a point beyond human control. To me, it's very similar to the discovery of fire. You

can do a lot of bad things with fire. But, there's no doubt that fire's an incredible

technology, changed our civilization forever. So we can try to quench it and say, "Okay,

we're not touching this," but I think a much smarter thing to do is to say, "How can we

try to use this technology for good and sort of understand its pitfalls. Discovering the

ability to build a self-aware machines will change our ability in big ways. It's going

to take 10 years, or 50 years, I don't know, but it's going to happen in our lifetimes.

And, we have to make sure we understand how to use it properly because the benefits far

outweigh the risks if we can handle it.