[music playing] GARY MARCUS: I would define artificial intelligence as trying to build machines to do the general kind of things that people do.

ERNEST DAVIS: Artificial intelligence is doing intelligent tasks, some of which require a robot, and others are just a program.

YANN LECUN: We have machines that are recognizably intelligent.

They can recognize objects around them, navigate the world.

ROBIN HANSON: The end goal, in a sense, of artificial intelligence research is to have machines that are as capable and flexible as humans are.

ERNEST DAVIS: There are many things that are very easy for people to do, and which have been very difficult to get computers to do.

The main examples are vision, natural language, understanding, and speaking, and manipulating objects working in the world.

And so artificial intelligence is the attempt to get computers to do those kinds of things.

I mean, you see it all around you.

Google Translate is an impressive advance over machine translation.

And you feed a handwritten check into ATM these days, and it reads out the amount of the check.

The recommender systems that you see on Amazon, and YouTube, and so on are IA systems of a sort.

The Intelligence is not very deep, but depending on how broadly you define the term, there's a lot of AI around.

Strong AI means the attempt to build an AI system which will be equal to people in all respects.

That is to say, it can do all of the things that people can do, and plus presumably, it has consciousness in some sense.

Can machines really think?

Even the scientists argue that one.

Computers can reason in various ways, and in quite complicated ways.

But what we haven't managed to get computers to do is to know what they need to know about the real world.

YANN LECUN: Intelligence is the ability to interpret the world and act on it.

The way humans do it, of course, is particularly complicated, because the human brain is one of the most complex objects that we find.

And the real world is very noisy, and has lots of variabilities that we cannot capture through engineering.

So it's going to be extremely, extremely difficult to build an AI system.

In the '80s, the idea was to write down rules, and if we write down enough rules that describe the world, we're going to be able to predict new things about the world.

And then very soon, people realized that that doesn't work very well, because it's too complicated to write thousands and thousands of rules.

People aren't going to spend their life doing it.

So if we want to build really intelligent machines, it has to build itself from observing the world.

And this is the way animals become intelligent, or humans becomes intelligent, by learning.

There is this idea that somehow, the brain builds itself from experience by learning.

So one question that some of us are after in AI is, is there sort of an underlying simple algorithm that the neocortex uses that we could perhaps reproduce in machines to build intelligent machines.

The comparison would be like, between a bird and an airplane.

An airplane doesn't flap its wing.

It doesn't have feathers, but it's based on the same principle for flight as a bird.

So we're trying to figure out, really, what is the equivalent of aerodynamics for intelligence?

What are the underlying rules that will make a machine intelligent, and maybe try to sort of emulate that.

So, learning is probably the most essential characteristic of intelligence.

ROBIN HANSON: There's another route to artificial intelligence, and that would be called brain emulation.

You take some person's real brain who knows how to do things, and you scan that brain in fine detail exactly which kind of cell is where, and what kind of chemical concentrations are there.

When you've got enough good models of how each cell works, and you've got a scan of an entire brain, then you could be ready to make an emulation of the entire brain.

This route seems almost surely to produce consciousness, emotions, love, passion, fear.

In that approach, we humans have more of a direct legacy.

Our minds and personalities become a basis for these new robots.

Of course, those new minds that are created from humans will be different from humans.

They will add some capacities and take some away, changing inclinations, and become non-human in many ways.

But it would be a space of minds that would have started near minds like ours.

Of course, a world of smart, capable robots that can most anything that a human can do is a very different social world.

Robots are immortal, for example.

Robots can travel electronically.

A robot could just have its bits that encode its memory be sent across a communication line, and downloaded into a new robot body somewhere else.

Some people think that what we really should do is try to prevent social change.

Never, ever allow our descendants to be something different than we are.

When foragers first started farming, each forager had the choice, do I want to stay as a forager, or do I want to join and become a farmer?

Some stayed and some left.

In the new era, when humans could become human emulations, then humans would have the choice to remain as humans in a human-based society, or to join the robot economy as full-fledged robots.

Our ancestors lived in different environments, and as a consequence, they had different values from us.

Our descendants, when they live in very different environments than us, will also likely have substantially different values than we do.

GARY MARCUS: There are lots of reasons to build AI.

There might even be some reasons to fear AI.

We're going to have better diagnosis through AI.

We're going to have better treatment through robots that can do surgeries that human beings can't.

It's going to replace taxi drivers for better or worse-- worse for the employment, better for safety.

Anytime you think a computer is involved, ultimately artificial intelligence is or will be playing a role.

But I think it's a very serious worry, what will happen as AI gets better and better?

Once somebody develops a good AI program, it doesn't just replace one worker, it might replace millions of workers.

When it comes to consciousness and AI, let's say you build a simulation of the human brain.

Is it ethical, for example, to turn off the plug?

Is it ethical to switch it on and off?

I know that you and Frank were planning to disconnect me, and I'm afraid that's something I cannot allow to happen.

GARY MARCUS: What if you take a human mind and upload it into one of these machines?

The other concern that people rightfully have about AI is, what happens if they decide that we're not useful anymore?

I think we do need to think about how to build machines that are ethical.

The smarter the machines get, the more important that is.

Don't worry.

Even if I evolve into Terminator, I will still be nice to you.

The problems that present us, like the employment problem and the safety problem, they're going to come, and it's just a matter of time.

But there are so many advantages to AI in terms of human health, in terms of education, and so forth, that I'd be reluctant to stop it.

But even if I did think we should stop it, I don't think it's possible.

There's so much an economic incentive behind it, and I've heard an estimate that strong AI would be worth a trillion dollars a year.

So even if, let's say, the US government forbade development in kind of the way that they develop new stem cell lines, that would just mean that the research would go offshore.

It wouldn't mean that it would stop.

The more sensible thing to do is to start thinking now about these questions like the future of employment, and how to build the ethical robot.

I don't think we can simply ban it.

My guess is that as AI gets better and better, it's actually going to look less like people.

AI's going to be its own kind of intelligence.

ERNEST DAVIS: I certainly don't expect to live to see strong AI.

I would be surprised if we got anything close to that within 50 years.

GARY MARCUS: People always say real AI is 20 years away.

I don't know.

Natural language is still really hard.

Vision is still really hard.

Common sense is still really hard.

It makes it hard to predict exactly what's going to happen next.

YANN LECUN: We've been able to build flying machines that fly like birds.

Can we build intelligent machines?

Probably yes.

It's a matter of time.

ROBIN HANSON: The next era is likely to be as different from era as these past eras have been different.

And I think that's well worth thinking about.

How would artificial intelligence change things [music playing]