How to Build a Human Page #3

---

a bit of an American institution.

It's a general knowledge quiz

programme,

and they ask questions in

a really strange way.

They give you the answer and you

have to work out what the

question is.

'Duncan Anderson is IBM's European

chief technology officer for Watson.'

So, we've got the two best

players here.

We've got the person who won the

most amount of money on the show

and the person who has

the longest winning streak.

'Two of the most brilliant brains

had won $5 million between them.

'This game was worth another

million.'

Watson itself is not connected

to the internet,

so it's not out there searching.

It's there, stand-alone, playing

against these champion players.

It was a big risk.

The category is 19th-century

novelists, and here is the clue.

APPLAUSE AND CHEERING

'And this is the point

where Watson's won.'

We've beat the best human

players at Jeopardy.

So, how exactly do you programme

a machine to do something that

it's never done before?

Well, the first thing is you

don't programme it.

Trying to guess every single

question that might come up

and then programme the computer with

the right answer for that question,

we would be here forever.

So, we use this thing called machine

learning, which is an approach

to solving problems whereby the

machine can learn from experiences.

So, we took Watson and we taught it,

we fed it lots of information.

For example, back issues of Time

Magazine, Wikipedia, encyclopaedias.

So, Watson was learning in a way?

Mmm.

And then we go through

a teaching process.

Just like you would teach a child,

we're teaching Watson and we're

testing it, and we're giving it

feedback when it gets it right, and

feedback when it gets it

wrong and then it adjusts its

approach to making decisions.

You could think of it

a bit like trying to find

a pathway through a field.

So, you have a very faint,

distinct path that maybe only one

person has trodden through.

And what you're trying to do is

to feed information

so that that pathway becomes

more defined.

As more people go down that path,

the path gets more trodden through

and becomes more obvious.

So, the more data that you feed into

Watson, it's almost like the...

The more or the wider the

path becomes,

or the more distinct the path

becomes? Exactly.

So, Watson becomes more confident

that that pathway is the

right pathway to take.

'We now need my robot to undergo

a basic version of this process.

'It needs its own pathway, and be

fed hundreds of bespoke new rules

'on how to respond to a question.

'And then learn how to use them.'

Ready to go now you want to try the

latest script from Bruce. OK.

' Key moment. I

will has commissioned one of the

world's leading computer

programmers to build a chat bot - a

piece of software which simulates

human conversation and responds with

the answer it thinks I would give.

Do you like repetition?

Yes. Do you like

repetition?

You said that already. Aha.

Now, that's the kind of reply

I'm looking for.

That's what I want to get to.

That was good. You're improving.

How about we talk about work?

You've played a lot

of supernatural characters.

Are you a fan of the genre or is

it just a coincidence?

I do love science fiction and things

that explore the boundaries

of the possible.

But, actually, having a bunch of

roles in that genre is just

a coincidence.

The team have certainly done

their homework.

Just as Watson was fed information

from previous Jeopardy games, our

chat bot has been fed interviews

and background information on me.

Once we can get the expressions and

all those other little subtle

cues, movement things in there, I

think it'll start to come together.

Fantastic. Yeah,

it will be fantastic.

Are you sure? I am positive.

What's your next question?

'The next question is how to make

it sound like me.'

Hello. Hello. I'm Gemma. And my name

is Bodil. Nice to meet you.

'The human voice has a huge variation

of infection, pitch and intonation.

'Our robot will need to replicate

the essence of my voice,

'with all the specific quirks that

make it unique to me.'

And so just keep in mind that when you

get commas, you make a small pause.

'Bodil Mattison is a computational

linguist and works for

'a company that specialises in

synthesised voices.'

It's not the words that you're

recording here,

you're more interested in the

different sounds that I'm making?

Yeah, I'm not interested in

words at all.

I'm interested in the combinations

of phonemes that we are getting.

Oh, the phonemes? Oh, I see.

So, the combinations of sounds.

How many sentences do I

need to record?

Well, in the end it's about 1,400

sentences. 1,400? Yeah. Today?

This is going to be quite

a long afternoon.

'I certainly have appreciated that

outlet for my creativity.

'Imagine George Bush singing

in the shower.'

You will have to do the

previous one again.

'We're teaching our robot how to

speak like me.'

You are an athlete if you play golf.

'But in the future, we may get to

a stage where it can teach itself -

'learning from experience and coming

up with its own solutions,

'which is what Als are

starting to do.'

Hi. Hi, Demis. Nice to meet you.

Good to meet you. Thanks.

So, welcome to the offices. Thanks

for having me. No problem at all.

'Demis Hassabis was a child

chess prodigy from North London.'

It's actually quite strange meeting

you, cos I've watched you

on screen pretending to be an Al.

Pretending to be a robot? GEMMA LAUGHS

'In 2011, he launched

a British artificial intelligence

'company called Deep Mind.

'Just three years later,

'Google bought his company

for 400 million.'

One of the first things we got our

programmed to do was to play

classic Atari games.

And we wanted the Al to actually

learn to play these games

by just from the pixels on the

screen and no other information,

so it had to learn for itself what

the rules of the game were,

how to get points and how to sort

of master the game.

The idea behind Breakout is that you

control a bat and a ball, and you've

got to break out through a

rainbow-coloured wall brick by brick.

I think I remember this from when I was younger,

yeah. From back when you were playing Atari.

So, you can see it's starting to get

the hang of what it should do.

It's not very good.

It misses the ball quite a lot,

but it's starting to understand that

it's got to move the bat towards

the ball if it wants to get points.

GAME BEEPS:

This is after 300 games,

so it's still not that many.

So, we thought this was pretty good,

but what would happen if we

just let the programme continue

to play the game?

So, we left playing for

another 200 games,

and then we came back and it did

this amazing strategy of

digging a tunnel round the side of

the wall and actually sending

the ball round the back.

That's amazing.

It's discovered it for itself,

and obviously can do it, you know,

Rate this script:0.0 / 0 votes