How to Build a Human Page #2

---

the internet.

But I think what puts it all in

perspective is the recent

fatal crash in the USA.

Yes, I remember hearing something

about that in the news.

'Earlier this year,

a driver crashed and died,

'allegedly whilst watching

a Harry Potter film.

'His car hit the underside of

a trailer at 74mph.'

It pulled out right in front of it,

bright white trailer,

bright white sky, very sunny,

and the camera couldn't detect

the trailer and that's why

it ran into it.

And I suppose the more autonomous

vehicles you have on the road,

you know, the likelihood of

incidents like that or

accidents -

they're bound to happen.

It's measuring the space now with

its sensors I think.

But while Tesla say AI cars will

have fewer accidents...

...ultimately, we won't be in

control.

Look,

it's correcting itself as well!

So you imagine what's going to

happen when they really are fully

autonomous because you're going to

have to delegate all your driving

decisions to them.

Life-and-death decisions.

It could well be

life-and-death decisions.

Look, do you see this van

coming down here?

Well, we're just driving along here

on that side in this direction.

The van comes hammering at us for

some reason and you've got

that women there and a pushchair,

and she's going to be hit if

we avoid it.

The car has to make a decision to

take the hit or

to kill the mother and baby.

What would you do?

I don't know! I don't know what

I would do in that split second.

That's a really difficult decision

for anyone to make,

for a human to make. How can a car

decide which life is more valuable?

I don't think it should, myself.

The robot we're building won't need

to make life-and-death decisions.

But it will have to harness Al's

decision-making powers to

think like me.

It's being built on

a modest industrial estate on

the outskirts of Penryn, Cornwall.

I wasn't expecting the hub of

the robot build to be next to a B&Q.

I'm wondering if it's going

to be a...mop head.

'Will Jackson is at the forefront of

constructing humanlike robots.'

Well, hello. Hello.

'To be convincing,

robots need to reason,

'learn and understand so they can

react almost instinctively like us.'

Hello, how are you? I'm very well,

thank you, how are you?

I'm very well, thank you.

He's great.

He's very much a robot,

you definitely know that this is

a machine, not a person.

He's pretty limited.

'Will plans to apply the knowledge

acquired building RoboThespian RT4

'as a basis for

constructing the AI version of me.

'In order to be convincing,

'the robot will need to master

the complexity of language.'

So the first thing that's going to

happen is somebody's going to

speak to the robot and we've got

microphones in the ears that

pick up the sound.

At that point, it's just sound,

it doesn't mean anything.

So we have to turn the sound into

words, into texts.

Once we've got the sound as text,

we've got to try and find

the meaning.

What is this person actually saying

to me? What's the key word?

What did they ask about?

And once we've got that meaning,

we have to try and think of

a sensible reply.

We have to then turn that

back into speech.

So we're going to have to take

a computer-generated version of your

voice that sounds like you.

While we're doing all of this,

the robot can't just sit still.

So it has to have all the little

actions that you would have.

The way you'd listen, the way you

think about what somebody's going

to say, the way you think about what

you're going to say.

You've got to get all these little

subtle things right,

all at the same time.

The brain in this robot has got to

come up with an answer that

makes sense even when it's never

heard the question before.

It's a huge challenge and we have

so many things to get right.

Will's going to use his existing

robot hardware to test out

the speech-recognition software that

he plans to use on the robot me.

There's an old saying in

computer programming,

it's, "Garbage in, garbage out."

If the robot cannot recognise

what you're saying,

if it just gets one or two words in

a sentence wrong when you speak,

what you get back is

complete gibberish.

'Using speech-recognition software,

can it recognise words it

'hears, turn them into text and then

accurately repeat them?'

Echo on.

OK, Echo on.

Peter Piper picked a piece of pickled

pepper, put it on a panda car,

drove it around the moon and ate

a sausage on the way home.

Peter Piper picked a piece of

pickled pepper, put it on a pan,

click and drag random name

generator, sausage on the way home.

"Random name generator sausage"?

Part of the challenge is to get our

robot to respond as quickly as

a human would.

That's within a tenth of a second.

Hello.

Hello. Too slow. Hello.

Hello. It's too slow.

To respond as fast as a human,

the robot needs to work out what's

being said and what it means

before the end of

a sentence and reply with

a lightning-fast response,

which involves instinct like us.

We've got to be so quick with

understanding what's being

said that we can be replying before

even really the last words come out,

so it's got to be superfast.

Building a machine that can

understand a human and answer

back convincingly is one of

the toughest challenges in Al.

If I speak quicker does it

work better?

If I speak quicker does it

work better?

It's almost fast enough, it's almost

there but it's not quite.

If we can't get the recognition that

quick - blown it, I'm robot.

'We're testing the boundaries

of science with

'a unique artificial

intelligence test -

'building a robot that looks,

sounds and thinks like me.'

WHIRRING:

The robot team is progressing well,

but it's my facial expressions

that are proving hard for the robot

to sync with its brain.

It's quite a scary looking

thing at the moment,

but once it's got the rest of the

core that goes on which

has got the top teeth in it,

and then the skin will go over

the top, it'll start to look

a lot more like Gemma.

The physical part of the build is

tricky, but it's the robot's ability

to converse like a human that's

really going to test our engineers.

With Moore's Law stating that

computer processing power

doubles every two years,

artificially intelligent

achievements in the real world

have been accelerating.

Fast.

In the 19505, computers beat us

at noughts and crosses.

Then, in the 1990s, they beat

us at chess.

Those computers were programmed to

work out all the possible

outcomes of each move,

and then weighed up how each move

would contribute to

a winning strategy.

But we're now entering an age when

computers aren't just

programmed, but

can learn for themselves.

Computers like IBM's Watson.

This is Jeopardy -

The IBM Challenge.

APPLAUSE:

In 2011, Watson was put up to one of

the toughest challenges ever

a general knowledge quiz that

requires logic and quick thinking.

This is Jeopardy. It's

Rate this script:0.0 / 0 votes