Original subtitles

So we've just seen a short history of AI.

and now the question is where are we now.

We'll, let's let you vote.

We'll do a little quiz. Which of the following things

can we do at present.

Can we create a machine which plays a decent game of table tennis?

Raise your hands if yes.

Yes.

Can we create a game that plays a decent game of Jeopardy?

Yes; many of you saw this decent game--it did okay, right, computer did okay.

Okay, we'll actually talk more about Watson later.

Can we build an AI,

that will drive safely along a curving mountain road? Yes.

Can we build an AI that'll drive safely along Telegraph Avenue?

I'm not sure *I* can drive safely on Telegraph Avenue. So the answer is

actually tricky. We'll come back to autonomous driving

a couple times in this course.

I'm going to put a question mark here--you know Telegraph Avenue is particularly difficult

and we can't just have a a car which does everything a human can do.

However, there are autonomous cars out there driving.

They are among us and they have a plan. In fact we'll talk a little bit about for example

the Google autonomous driving cars, which under the right circumstances and with

tools that a driver does not have, are in fact able to drive on real roads. Amazing progress.

One thing that's really cool about this list: I've been using this list for years, and

kind of every year or two

some "X" turns into a question mark or a question mark turns into a check.

It's really cool to watch.

Can we build an AI that will buy a week's worth of groceries on the web.

Yeah, it's kind of almost not even AI right...

Can we build an AI which will buy a week's worth of groceries at the Berkeley Bowl?

Well that's a little harder right--you've gotta like navigate around carts.

This is probably something that we currently cannot do and this is an illustration of the decision problem--

What do I buy?

--being tangled up in the action in the real world which involves like,

not bumping into people with your cart.

To operate in the real world and in an actual mobile platform often mixes all

this stuff together in a very challenging way where the decisions you make happen

at multiple levels of abstraction.

Can we build an AI that will discover and prove a new mathematical theorem.

Raise your hand if yes.

Raise your hand if no.

Okay, so I'm gonna put a question mark. We can in fact build amazing theorem provers

that can prove new theorems.

Discovery, not so much. Why is this hard? So it turns out that if

I state a theorem and it's true,

we can build automated methods of proving that they are true through computation.

It's a lot harder to just ask the computer: Tell me something awesome about algebra I don't know.

Cause it could be like, hey, wait, wait, I got it.

10 + 3 = 13.

You know, it's true! But you don't know what's interesting so it's very

hard to characterize what makes a theorem interesting.

In some sense it's easier to prove something that's true

than figure out what should be proven.

Can we build an AI that will converse successfully with another person for an hour?

It depends a lot on the other person.

Some of you may have heard of ELIZA. We'll talk a little bit about ELIZA when we

talk about natural language later in the course, but, for most people the answer is

no.

For some people the answer is yes--you can converse with some of the people for

some of the time, I don't know, but the point is in general we cannot yet do this.

We cannot yet have computers that have full conversational ability on any topic

just like a human would.

How about we build in an AI that can perform a surgical operation?

Who says yes.

Okay, raise your hand if you would let and an AI perform a surgical operaton on you.

Would you let an AI remove your kidney?

Right, okay,

where did all the hands go?

That illustrates one of the issues--the answer here's a little complicated.

Yeah, we have computer-assisted surgery.

Computers are starting--including some cool stuff in Pieter Abbeel's group--starting to

be able to tie knots in automatic way. You can do things like track the

movement of an eye in eye surgery and things like that.

However,

it is not yet the case, both in technology and in cultural kind of

willingness to undergo the experiment, that we're gonna be having computers

swap our kidneys around.

Can we build an AI that can put away the dishes and fold the laundry?

Yeah, we can build it right here at Berkeley. I can show you a little bit later, again, some of

Pieter's work on folding.

We can do that.

Translate spoken Chinese into spoken English in real time?

Raise your hand if you think we can do it.

Yeah, we'll talk more about translation and real time translation later,

but we can basically do this.

Now you say, wait, is it good English?

Um..., uh..., I didn't say it was good I said it was spoken right, um...

but we can do real-time

recognition, translation synthesis. That is something we can do.

Better than you probably think.

Can we build an AI that will write an intentionally funny story.

Who thinks yes?

Uh...

I'm going to say no.

It's an open area of natural language processing.