Original subtitles

Let's talk about what AI actually is. So, what is a AI--well

actually this is a big discussion we have to have

as a field--what is AI? Well, we're going to be building machine software,

you know,

that does something.

What's our goal, what does it mean to build an artificial intelligence.

Well there's been multiple schools of thought on this.

One school of thought is what we should really be doing is building machines

that think like people, right. Intelligence is about thinking, and this is artificial.

What's the natural intelligence--I guess that's us.

So we want to build these machines that somehow go through the thinking

processes that people do.

Alright, there is actually a science that studies this,

and it's not really AI anymore.

This is some mix of cognitive science and computational neuroscience

really trying to understand the brain.

And it's very important but it's not what this course is going to be about.

So another thing that people at times have thought AI should be, is we should be

building machines that act like people.

Okay, so we should say: who cares about how they think, they can think in some

strange, alien, silicon way, but the action, the behavior has to be like what we know from people.

This is actually a very early definition. This is straight from, uh, Alan Turing,

the definition that really, all you can really check is behavior. Is the behavior

like an intelligent human?

So this led to things like the Turing test where you put a robot on one

chat channel and a human on the other and then you have an interrogator who chats

with both of them and try to say that one was the robot and that one was the human.

And this is a really good idea because provided you can't actually see them--there's no

video right ..., where you know the robot's the one with the blinking lights right.

So provided it's just over chat

you can really kind of test anything. It's open-ended: do they have hobbies,

can they answer a general question about a chess configuration,

right. The problem was, the Turing test, in order to really do well,

you don't just really concentrate on programming intelligence, you concentrate

on things like,

don't spell too well,

humans don't do that. And so you build in some type of typo Turing machines

and then you think, wait a minute, if i get asked about the square root thirty-five,

I better not have an answer.

And so you go through basically trying to mimic things that probably you didn't

really value in the human in the first place. On the other hand, you got to be

really sure that you have a favorite Shakespeare play

'cause the interrogator always asked that.

Okay, that thinking like people and acting like people and the realization was this

really wasn't going anywhere in terms of building machines that were useful in

say industry,

and so the realization was maybe it's not about mimicking people.

We've already got those, right. Maybe we should do something else. Maybe what we should

be doing is building machines that think rationally. So, whatever thought

processes are, they should be correct. What does it mean to have a correct thought process,

it's a very kind of a prescriptive thing.

And this actually has a long history in the logicist and philosophy tradition

going all the way back, say to Aristotle's laws of thought.

This is how you think

in order to kind of not make a mistake in your deductions.

And this tradition actually still shows up in various places of AI.

By and large,

this wasn't the winner, and the reason it wasn't the winner is because our ability

to write down how to do logical deduction

turned out to be relatively fragile,

and it any case when we're learning about how to incorporate uncertainty we

also had this realization that really it wasn't about how you think, but about the

actions you take in the end.

So the winner for this course is that AI, for us,

is the science of making machines, that act rationally.

So what's that mean? We only care about what they do,

and our requirement on what they do is the that they achieve their goals optimally.

You may be looking at this, and you maybe be thinking, okay rational, rational means I have a

level-headed decision, I don't get angry. So we want to build machines that don't get angry.

Well you know, I don't know, uh...

if you think back to GLaDOS maybe that's good, maybe we shouldn't

build machines that get angry. Um... Skynet got a little angry.

So maybe we shouldn't build machines that get angry.

But when we say rational that's not what we mean.

Rational has a very technical meaning. It means that you maximally achieve your pre-defined goals.

So the input to an AI is a goal,

and rationality means you achieve it in the best possible way.

Rationality--only matters what you do.

It doesn't matter the thought process you go through, right. If I have a

robot vacuum cleaner,

and it just make some optimal grid on the ground, and cleans up all the dirt, great.

If it sits in the corner and thinks, alright, where shall I clean? Well if I go diagonally

there will be a place left over. And then it cleans everything up--fine, it doesn't matter.

They're equally rational

for that task in that context.

There may be advantages to the thinking robot,

there may be advantages to the kind of more reactive reflex robot.

We'll talk about that in the next class.

Goals are all expressed through utilities. So we're going to spend a lot of time in this course talking

about what a utility is.

And in the end remember that being rational means maximizing your expected utility.

Okay, so this course, really, we should have called it computational rationality. We're going to teach you

computational methods--this is a computer science course, and

it's all going to be about this idea of rationality:

maximally achieving your goals.

Okay, you say what about artificial? I didn't really say anything about artificiality,

that's kind of orthogonal.

And what about intelligence? Well,

intelligence is a tricky thing. The philosophers are still working on that.

When they get back to us on what intelligence is, well probably we'll just

ask them then what consciousness is.

but when they get back to us on intelligence, we're gonna say,

that's great but we're working on rationality right now.

Okay, so if you remember nothing else in the course,

or if you decide that you really want an AI tattoo,

and you needed to distill the course down to one thing,

it would be this:

it would be maximize your expected utility.

Aand we're gonna spend this entire course thinking about computational systems

that do this.

And in order to do that we've got, you know, however many weeks left in which we

will unpack this definition

The first part of the course deals with the maximize:

How do I figure out which action is best? That has to do with the consequences of

that action, the context of that action, are there adversaries.

We're then going to have to unpack this idea of utility. What is utility?

What does it mean to have a function that describes my goals.

And then, kind of the kicker in here that's a little bit hidden is what is

this deal about expectation?

Well if I take an action I don't know what's gonna happen.

So my optimization of goals rationally doesn't deal being successful.

Life is full of risks. It has to do with doing the right thing in kind of the

appropriate kind of weighted average.

And so we're going to have to unpack this notion of what it means to do the right

thing on average, and that'll get us into probabilistic inference, and that will

occupy the middle third of the course.