1
I don't know about you,
but I can't remember the last time
I used one of these
to look something up.
We all know we're in the midst
of an AI revolution,
changing the way we do
almost everything.
Instead, I use this - a computer.
It's packed with AI technologies and
it's linked online to so many more.
Hey, Siri, where's my nearest
charity shop?
"The closest one I see is Age UK.
Does that one sound good?"
It does.
AI used to only be the stuff of
science fiction - but not any more.
HE SPEAKS IN ARABIC
"If I speak Arabic,
"artificial intelligence
translates into English."
I still find that utterly amazing.
But I want to investigate how these
extraordinary AIs actually work.
It doesn't think that's a dog.
It thinks it's a trombone.
So what on earth is going on?
I'm going to explore
the surprising history
of trying
to turn a machine into a mind...
Checkmate. Checkmate.
..and I'll find out how AI
can learn for itself...
Yes!
..and take on challenges
previously thought beyond it.
"So are you more interested in
reading books or watching movies?"
I prefer books.
"Oh, a bookworm - how nice!"
Some people are fearful for
humanity's future in the age of AI,
but I'm not so sure.
Too often this story is told
as a battle between man and machine.
But for me it's about man
working with machine.
So what does AI actually involve,
and where will it take us?
There are so many things
computers can do today
that we call
artificial intelligence -
but there's still
no clear definition
of what artificial intelligence
actually is.
Perhaps that's not
entirely surprising.
The aim, after all, of AI
is to simulate human intelligence,
and human intelligence does such an
amazing range of different things.
We perceive and make sense
of our environment.
We set goals and plan
how to achieve them.
We use language to communicate
complex ideas,
and all the time we learn
from our experiences.
To get computers to do any of this -
to think like humans, that is -
the obvious first question
you have to tackle is,
what actually is thinking?
If we could understand
how our minds work,
then perhaps we can apply this
to computers.
MUSIC: Jingle Bells
Christmas, Pittsburgh, 1955.
This was the moment
when two American scientists
not only thought about thinking,
but first worked out
how to mechanise it.
For my money, one of science's
real eureka moments.
Optimism was abundant
in 1950s America,
and scientists truly believed
that there were very few problems
they wouldn't be able to solve.
Herbert Simon
was a political scientist.
His friend, Allen Newell,
was a mathematician,
and they shared a fascination
with the possibilities of computers.
Now, in 1955, the few computers
that existed in America
were mostly just used
for numerical calculations -
but that was about to change.
As Simon would later tell it,
"Over Christmas,
"Newell and I invented
a thinking machine."
What inspired the two men
that Christmas was the new idea
from cognitive science,
that our thinking process
is essentially a form
of computation.
Inside our heads,
reasoned Simon and Newell,
our abstract representations
of realities in the outside world.
And when we think, we
are performing logical processes
on these abstract representations.
So, a dog plus a cat
equals a fight...
..and if our minds, when we think,
are computing,
then perhaps,
reckoned Simon and Newell,
computers could be programmed
to think like us.
For their festive fun,
they picked a seriously knotty
logical thinking problem.
Simon owned a copy of this legendary
and hefty book,
Principia Mathematica
by Russell and Whitehead,
which uses logic over hundreds
of densely packed pages
to prove the theorems
and axioms of mathematics -
and they wondered, could they
write a computer program
to automate the proofs in this book?
Simon and Newell were particularly
concerned that their models actually
conformed to the way
that the human mind operates.
They had people solve problems
and, sort of,
write notes on how
they were doing it,
what was going through their mind
while they were doing it,
and they built their computer
programs to actually simulate
what they were perceiving
as the mental process
of a human solving that problem.
Over the holidays, Simon lined up
his wife and three children,
together with Newell
and a bunch of grad students,
and gave each of them
a card like these to hold.
Each card represented a step
in a computer program,
and effectively they all became
a real-life human computer.
It worked - and before long,
when coded into a real computer,
it solved 38 of the theorems
in this monumental book.
Simon and Newell called their
creation the Logic Theorist.
Human thought itself
had been simulated
in what's now regarded
as the very first
operational artificially intelligent
computer program.
Within months, the phrase
"artificial intelligence"
was adopted to describe
the new field.
For Simon and Newell,
and other pioneers,
getting computers to solve logical
problems was a huge breakthrough -
but one particular challenge
had the AI boffins gripped.
This battle of wits
has long been regarded
as the ultimate test
of reasoning power.
What makes chess so challenging
is that there are more ways a game
can develop
than there are atoms
in the visible universe.
Neither humans nor computers
can possibly consider them all -
but if a human can nonetheless
decide what to do,
how can a computer be programmed
to do the same?
This problem fascinated the
godfather of computer science.
In 1948, the great code-breaker
and mathematician Alan Turing
wrote this paper,
containing what is
considered to be a plan
for the world's first
chess computer program.
In it, he proposed a solution
to the game of chess's seemingly
infinite number of options.
So, today I'm going to play a game
as this chess program -
or, as it came to be known,
as Turochamp -
to demonstrate what it could do.
My opponent is seasoned club player
Olivia.
Since no computer yet existed
that could actually run Turochamp,
Turing had to use pencil and paper.
Even for him, this took hours.
OK, so let's play. Good luck.
Fortunately, I've got a laptop.
Hm. Slightly unusual start.
At the start of a game,
just one move each
generates
400 possible combinations of play.
Two, almost 200,000 -
and, by four moves apiece,
we're into the tens of billions.
This vast multiplicity of options is
called the combinatorial explosion.
There she goes again, back again.
I'm going to attack your queen.
No computer could calculate
them all,
so how to give it the intelligence
to make good choices?
Turing wrote down a set of rules
to guide the computer's search,
and rules of this kind became known
as heuristics,
from the Ancient Greek,
meaning to find or discover.
For example, always consider
capturing an undefended piece.
The program would use these
heuristics to evaluate
all the possible moves
and countermoves,
to prune down
the tree of possibilities,
so that it only had to go down
the more promising branches.
What Turing had realised in doing
this was that, for a given problem,
programmers could codify into rules
their own human knowledge
of how to deal with it.
Oh, see - I'm in retreat now!
Then, if a computer
followed the rules,
it could solve the problem too.
Given how rudimentary
Turochamp's rules are,
I find myself thinking
it's remarkably effective.
Checkmate.
Still, it's no match for Olivia.
There we go. Congratulations.
It wasn't too hard, was it?
It wasn't too bad! No!
Turochamp was an elementary
chess program,
but the principle that heuristics,
or rules, were the way to overcome
the challenge of the combinatorial
explosion was a sound one -
and this idea was applied far
beyond chess, very successfully too,
as programmers tackling a
wide range of real-world problems
encoded their own human knowledge
into increasingly complex
and varied heuristics.
This approach became known
as classical AI,
which does many clever things.
In logistics, manufacturing,
construction,
classical AI systems,
each with a set of programmed rules,
are today used
to plan complex operations
in highly controlled environments,
with maximum efficiency
and economy...
..but most of the world isn't
like an ordered production line.
It's much more chaotic.
How is a computer to make sense of
all of this?
All this movement, all this
noise, all this variety?
I mean, I recognise these buses.
I recognise a van.
I can see a taxi.
I even recognise those as adverts
up on the screen.
We instinctively know what we're
looking at, but to a computer,
it's just this -
a torrent of raw data,
a mass of numbers without meaning.
How could you possibly
write the rules for a computer
to make sense of all this
like we do?
The trouble with classical AI
is that the real world
is messy and complex...
..so it's almost impossible to write
the rules for a computer
to even begin
to make sense of its environment,
let alone apply it to a task.
So even the seemingly simple problem
of planning to cross a road
would be beyond it.
But fortunately there's another way
to go with AI.
Instead of us attempting to give
computers the rules,
the computers learn how to make
sense of the data for themselves.
This approach is known as
machine learning,
and it's machine learning that
powers most of the amazing AI tools
that we use today.
Why AI has become such a big thing
in the last decade
is because these new techniques,
which are based on learning,
have become very powerful.
You give the systems the ability
to learn for themselves
directly from raw data,
and these systems learn
from first principles
the structure in that data and
potentially solutions to problems.
So this is a very powerful new way
of thinking about intelligence.
Let me show you how machine learning
beats classical AI
at dealing with complex data
with an example that won't
get me run over.
How to cope with spam.
It's hard to be sure,
but perhaps 400 billion spam e-mails
are sent every day.
That's something like
eight out of ten of all e-mails.
Without spam filters,
we'd all utterly drown in junk.
Those incredible
one-time only offers...
..those performance enhancements...
So, how do you ward off
what you don't want,
but let in e-mails that you need,
the non-spam -
or to give it its correct,
very technical term,
the ham?
The classical AI approach to this
would be to come up
with a set of rules.
For example, you choose
specific spammy words,
and if they come up in an
e-mail's title, it gets zapped.
But here's where the rules of
classical AI hit their limits.
Some of those spam words
can also be ham words -
and so e-mails you do want to
read get junked too...
..and then what about this?
How do you begin to write rules
to catch all these?
But machine learning can find
patterns in all the e-mail data
to tell ham and spam apart.
It first needs what's called
training data - lots of it.
A heap of what you already know
is ham -
these are all from my inbox...
..and a load of what is most
definitely spam.
Then it can start hunting
in all this stuff
for the mathematical patterns.
For instance, first it identifies
the most common words in each pile.
For a professor of physics
called Jim,
there are plenty of these
in the ham pile.
They're very different from the
words that are strictly spam only.
None of these in my inbox,
thank you very much.
But it's because of how words appear
in both piles
that machine learning
really comes into its own.
It crunches the training data,
looking for the patterns
in how all these words combine,
and works out how the ham are all
subtly different from the spam.
So, ham or spam?
Today's filters are so good,
we hardly get any spam at all.
Though if you are in
the regular business
of doing great deals on Viagra,
then you'd better still check
your spam bin.
When it comes to simulating
many of the things humans do,
machine learning outdoes classical
AI time and again.
The reason why may lie in the
workings of the brain itself.
While classical AI
attempts to mirror our conscious,
rational thinking,
machine learning may better reflect
the enormous power
of our subconscious minds.
We are only conscious of a small
amount of what the brain does.
When you open your eyes
and you see a world,
it happens effortlessly.
You know, we're aware
of the outcome,
we're aware of seeing the world,
but we're not aware of the process.
We're not aware of what it takes
under the hood
to generate this inner universe
that we effortlessly experience.
In 1988, a computer scientist
and roboticist called Hans Moravec,
who was fascinated with the workings
of the human brain,
pointed out that,
from a human perspective,
progress in artificial
intelligence seemed paradoxical.
You see, the things that seem
difficult
for our brains to cope with,
things that require a lot of
conscious mental effort, like chess,
were proving to be
relatively simple for AI.
Meanwhile, the things that our
brains seem to find a cinch,
that we do unconsciously,
like making sense of
what we see, what we hear,
our environment -
so, my ability to see where the
camera is,
or to hold this brain gently
without dropping it,
were proving to be the toughest
challenges for computer programs.
This became known
as Moravec's paradox.
Moravec reckoned it was all to do
with our brain's evolution.
Here's how Moravec
very eloquently put it.
"Encoded in the large,
highly evolved
"sensory and motor portions
of the human brain
"is a billion years of experience
"about the nature of the world
and how to survive in it.
"We're all prodigious Olympians
in perceptual and motor areas.
"So good that we make
the difficult look easy.
"Abstract thought, though,
is a new trick,
"perhaps less than
100,000 years old.
"We've not yet mastered it.
"It's not all that
intrinsically difficult.
"It just seems so when we do it."
It's a neat and very
convincing explanation,
but it also highlights an enormously
fruitful shift that was taking place
in artificial intelligence.
From attempts to build
computer programs
that mirror what our conscious
minds seem to do,
to ones that replicate
how our brains themselves
are physically structured.
These remarkable and very powerful
machine learning systems
are called artificial
neural networks.
They're inspired by how real brains
respond to the world.
WOOF
You know that because your brain
just fizzed
with electrical and chemical
signals,
making their way from the eye
back and up through dense
layers of neurons.
Each one a single cell -
and depending on the combined
strength of the signals coming in,
it either does or doesn't fire.
Your brain contains something like
90 billion of these neurons,
and they're networked together,
often with thousands
of connections each.
That's at least 100 trillion
connections in total...
..and it's this vast neural network
in our brain
which is brought into play
to spot that Spot here is something
that indeed should bark.
It makes sense to try to mimic
the brain to some degree.
The question is,
how closely do you do it?
Of course, in flight,
people did not build aeroplanes
that had flapping wings.
Rather, they understood the
principles of flight
and so there are some shared
features
between aeroplanes and birds,
but they're not direct copies.
I think the same thing applies
to neural networks.
That it's not about replicating
every last detail
of a human brain or an animal brain,
but trying to identify the
principles by which brains work.
An artificial neural network is a
virtual creation
of computer software,
rather than a blob
of real brain tissue...
..but when it's presented
with our dog -
actually, a picture of our dog -
or, to be even more precise,
the pixel information from the
picture of our dog -
the virtual neurons pass signals
through the network
so it, too, can tell
what it's looking at...
..but first, just like dogs,
and indeed spam filters,
artificial neural networks must
learn what to do by being trained.
For this, we'll need to show
it lots of Spot's friends.
Each time we tell the network
what it's looking at,
it tweaks its connections to better
recognise doggy pixel patterns...
..and it can learn
about other things too.
INCORRECT BUZZER
It's got loads and loads of
adjustable numbers, and it's...
You know, when I say loads,
I don't mean hundreds.
I might mean millions.
So we expose it to a load of data.
We show it a load of cat images,
a load of dog images,
and we tell it which is which,
and it adjusts its numbers
so that when we show it new cat
images and new dog images,
it correctly says what's in them.
Now it's trained,
the neurons in the network's
inner layers
first detect the simplest shapes.
They then identify combinations
of these shapes - doggy features.
Then combinations of combinations.
The more layers,
the better these networks do...
..but, remarkably,
even the scientists who build these
networks don't really understand
how they come up with the answers.
Neural networks are radically
different
from, I think,
any previous kind of technology.
Previously, some complicated device,
some complicated clock, whatever,
someone had put it together
and they knew how it worked,
and they'd known why that piece
was in there
and why that joined to that.
With neural networks, you can
understand some of what it's doing,
but then there's
a load of other stuff
and you have a look at it,
and it's frankly mysterious.
You can't make any sense of it.
So we don't know
what it's doing there!
Whatever's going on under the
bonnet, with neural networks,
AI can now make much better sense
of the messy real world...
..and, with this breakthrough, its
potential has increased enormously.
AI is now booming.
Whether it's optimising harvests,
interpreting medical images,
grading students, detecting
financial opportunities,
neural networks are mastering new
tasks in all parts of our lives.
Take transport,
and the AI application we're often
told is just around the corner.
Driverless cars.
This British company is busy
developing and running this tech
in a range of different vehicles,
on real roads, not just test tracks.
Oxbotica took me for a spin.
As we drive, I begin to realise just
how much this kind of AI
is going to revolutionise
the way we live.
It's actually remarkable
how safe I feel.
You know, you very quickly...
..trust that it knows
what it's doing.
Every fraction of a second,
the car runs simulations of what the
world might look like
and it takes...simulates
lots of possible outcomes.
"Well, if I drove that way,
what would this look like?
"If I drove that way,
what would this look like?"
And it generates thousands of
simulations 50 times a second.
So continually updating, "What if I
did this, what if I did this?"
Then evaluating them
and choosing the best one.
So it tries to do the least worst
thing the whole time.
Feeding into these simulations
is a continuous stream of data
from the car's onboard
lasers and cameras.
The laser data gives it a 3-D model
of everything around it.
Any object that's moving - or might
move - is located and tracked.
Then with camera data, it identifies
what these objects are
and, so, how they might behave.
The AI in the car doesn't need to
communicate with any other computer,
it's entirely self-contained...
..and all this is only possible
thanks to neural network systems
that learn from
their driving experiences.
All this comes together on my drive
as the car negotiates a sudden
moment of high drama on the highway.
Oh, very good!
So, here's the classic driverless
car situation.
A woman crossing the zebra crossing,
who stepped out just as we were
coming up to the crossing,
and it stopped.
Are we are going to get to the point
where - I've got a driverless car,
therefore, I'm going to have a nap?
That it's completely safe, I'll
leave it entirely up to the car?
We will.
I am absolutely sure we will.
I think the vehicles that you can
sit in them and they will drive you
around parts of the city,
part of an airport, campus,
that's coming quite soon.
The vehicle that has the same
functionality as your car does now,
that can get you from anywhere
to anywhere,
any weather, any time of day,
without having any difficulties,
and total confidence
that you're going to get there,
and you can buy that from a
forecourt,
and you don't need a steering wheel,
and you don't even need
a driving licence -
in fact, it may not have
any windows - long time away.
But I don't think there's anything
that's unattainable
about humans driving
that a machine can do.
That argument,
I think you would have to...
It hits onto something that's not
computable about driving,
and that doesn't seem particularly
realistic to me.
With AI muscling in on ever more
of what we do ourselves,
it's no wonder many worry about how
the AI revolution
might change our lives.
Revolutions make people nervous,
especially when they're not
the ones in control.
Probably the biggest fear is that AI
might take people's jobs
and they might never
find work again.
One of the concerns of AI
is that it is leading to this huge
technological revolution
that is going to affect society.
Yes. I mean, we can't stop it,
we can't mitigate against it. No -
and nor can we deny that
there's change coming,
but I think we can now look ahead
and go, "New jobs are coming,"
in the way that new jobs came
because of computers -
and think how many people have jobs
that are now only doable because
they have a computer,
have become possible,
or were invented,
because you have computers?
So, I can't deny that this
transformation is coming
but I'm, if you like,
almost pathologically positive
that it's going to make us
healthier and wealthier,
and enhance our capabilities,
and change jobs in the way
that computing did, as well.
You know, you look at
civilisation around us,
that's all a product of
intelligence...
..and I think of AI as, you know,
a powerful tool -
perhaps the most powerful tool
of all -
that will allow us to reach
the full potential of humanity.
We're still a long way off this
brave new world
and, to get there,
we'll need even cleverer AI...
..but that's what Demis Hassabis
and his colleagues
are dreaming up at DeepMind -
the blue-sky AI research division of
a leading search engine provider.
Here, they're trying to develop
neural network systems
that can learn to do anything,
without any human intervention.
You know what their mission
statement is?
"Solve intelligence and then use it
to solve everything else."
That's ambitious, you'll agree,
and they're going about it in a
rather intriguing way.
The idea is that we first test
and develop AI algorithms
so that they can master games,
but then our hope is if we do that
in a general enough way,
they'll be able to be used in the
real world for serious problems.
And it turns out they've got a real
thing here for Retro Atari Classics
when it comes to testing what an AI
could learn to do for itself.
Presumably, you're having to teach
your AI the rules of the game,
so that it can learn how to play?
No, we don't.
It learns, really,
only from its experience.
All it's seeing is those pixels
and whether or not its score
increased or not,
and then trying to solve the puzzle
of, "Well, my score got better then,
what was the action that I took?"
And that's really just done
through a learning algorithm
that changes all of the millions
of connections
in this neural network
to say "Let's reinforce
this action,"
or, "Let's not reinforce
this other action."
So while we could program up some
rules that said, "Here's the brick,
"here's the ball, here's the paddle
and here's how you move it,"
we don't do any of that.
We simply let the algorithm learn
on its own.
So, how quickly does it learn
and improve?
So, after about 300 games,
we see that we can get to
human-level performance -
but the nice thing about AI
algorithm is we can just let it run,
and so we let it keep on training
for a few more hundred games,
and then we see that it does
get to super-human performance.
Well, that... Let's take a look at
that! I want to see that! Sure.
So at the beginning of the game,
it's moving back and forth,
it's hitting the ball back,
but as the game progresses,
then the ball is going to move
faster and faster.
This is where humans
stop being able to return...
..but the algorithm discovered a
really interesting strategy,
and we weren't expecting to see
this, we had no idea -
so, it was really exciting to see
what it's doing now,
which is what we call tunnelling.
It has managed to systematically
hit the ball only to one side,
and that means that it breaks
through to the top, bounces around -
maximum reward, less risk of dying,
of losing a game.
So that's a strategy
that it figured out for itself,
because it could see that that would
give it a huge advantage?
It managed to discover this
absolutely on its own.
Variants of this AI -
neural networks
learning entirely by themselves -
have gone on to
reach superhuman level
on over 40 different Atari games.
What's remarkable isn't just the AI
learning so rapidly
and successfully,
it's how it discovers its own
strategies for success...
..but could a neural network AI
even discover things
that we don't know of?
In 2016,
DeepMind's programmers
created an AI system
that taught itself to master
the ancient game of Go.
In Go, players battle
for control of territory...
..and although the rules are simple,
it's nonetheless an enormously
complex game,
where players need to rely on their
intuitive sense of pattern.
Where chess might be 50% about
intuition and 50% about calculation,
Go is more like 90% intuition,
10% calculation.
DeepMind built a neural network
system called AlphaGo.
Trained by playing millions of games
against itself,
it was able to capture the
intuitive, almost unconscious
pattern recognition ability
that human Go players have.
Confident of AlphaGo's powers,
DeepMind challenged one of the
greatest Go players in the world,
Lee Sedol, to a very public
five-game tournament.
Nobody outside of DeepMind thought
that he would lose a single one.
In the end, AlphaGo beat him
four games to one...
..but the most significant moment
came in game two...
..when AlphaGo played a move
no human player
would have even considered.
Ooh! That's a very... Ooh!
That's a very surprising move.
I thought...
I thought it was a mistake!
At that point, we didn't know - was
it just, you know, a useless move,
or was it actually a brilliant move?
Er, so, coming on top of a fourth
line zone is really unusual.
And in fact, Lee Sedol,
when confronted with move 37,
his jaw dropped visibly and he
thought for, like, 20 minutes.
So at the very least,
we knew this was a shocking move.
Remarkably, not one
of the humans watching
understood why AlphaGo had done
what it did.
It turned out to be decisive
in that game.
About 100 moves later,
a battle in another part
of the board
ended up perfectly connecting up
with the piece
that was played on move 37.
Lee Sedol commented afterwards
that when he saw that move,
he realised that this was a
different type of machine.
That it wasn't just regurgitating
human knowledge,
or memorising positions.
In some sense, it was actually
creating new ideas.
Oh, he resigned. It's done.
OK. Wow! Wow!
Yes!
CHEERING
The AI had made a genuine discovery,
one with profound implications.
It showed that these types of
learning systems
can actually come up with a new idea
that hadn't been searched or thought
about before by humans...
..and what's amazing is if
that can happen in Go -
which we've played
for thousands of years -
then how much potential has this
kind of system got in other areas
like science and medicine?
I think with these powerful tools,
we're going to enter a golden era
of scientific discovery.
And yet, a computer that can outgun
the top human
with strategies it's intuited
by itself...
..is unnerving -
and it begs a big question.
What if one day, scientists manage
to create an AI that rivals,
or exceeds, the full range of what
human intelligence can do?
MUSIC: Thus Spoke Zarathustra
by Richard Strauss
The idea of a computer that not only
outstrips our intelligence,
but that also slips dangerously
out of our control,
is a staple of science fiction.
For his film 2001: A Space Odyssey,
which was made in the late 1960s,
director Stanley Kubrick created one
of the most chilling realisations
of this idea ever seen -
the HAL 9000 supercomputer.
In the film, HAL - in its own words,
"foolproof and incapable of error" -
starts acting in unexpected
and disturbing ways.
The astronauts on board its
spaceship
make plans to deactivate it
and, when it finds out,
it attempts to kill them all -
and very nearly succeeds.
HAL wasn't malevolent,
just remorselessly logical.
The astronauts would have stopped it
from completing its mission
and so, of course,
they had to be eliminated.
HAL wasn't entirely dreamt up
by Kubrick and
co-writer Arthur C Clarke.
It was also inspired by the work of
a British computer scientist
called Jack Good, who was a veteran
of Alan Turing's codebreaking
effort at Bletchley Park
during World War II.
Jack Good had laid out a startling
vision
of the future of artificial
intelligence in an essay
entitled Speculations Concerning
the First Ultra-intelligent Machine.
"Let an ultraintelligent machine
be defined as a machine
"that can far surpass all the
intellectual activities of any man,
"however clever.
"Since the design of machines is one
of these intellectual activities,
"an ultraintelligent machine
could design even better machines.
"There would then unquestionably be
an 'intelligence explosion,'
"and the intelligence of man
would be left far behind."
This intelligence explosion
identified by Jack Good
might well have been
for the benefit of all humankind -
but what must have grabbed Kubrick's
attention was the sting in the tail.
"The first ultraintelligent machine
is the last invention
"that man need ever make,
"provided that the machine
is docile enough
"to tell us
how to keep it under control."
King Midas said, "I want everything
I touch to turn to gold,"
and the gods gave him exactly
what he asked for.
So his food to turned to gold,
his water turned to gold,
his wine turned to gold,
his daughter turned to gold.
We do not know how to say
precisely what we want,
and if you have a super-intelligent
machine that's kind of like a god,
it will find some way of giving you
your objective,
in ways that you didn't expect -
and so we've got to figure out a way
that guarantees
that we retain control forever
over things that are much more
intelligent than us.
A superintelligent AI
is an alarming thought
but, in reality,
it's not coming any time soon,
so we've plenty of time to work out
how to control one.
The AI behind so many of today's
amazing breakthroughs
is still fundamentally limited.
It can find patterns in complex data
often better than we can...
..but it can't yet convert these
into the kind of meaningful,
conceptual thinking
that's so crucial
to our intelligence.
Let me show you - using, yes, our
furry four-legged friends again.
I've had a state-of-the-art neural
network installed on this tablet.
It's been trained to identify
over 1,000 different kinds of
animals and objects,
using over a million examples.
Now, it hasn't been trained
on these pictures.
It's seeing them for the first time.
First up, a classic portrait
of a dog.
CAMERA CLICKS
Right, not only has it recognised
it as a dog,
but it's pretty certain
it's a Brittany spaniel.
Well done, network. Right, let's try
it on something slightly harder.
Because this isn't a classic
portrait of a dog,
you can't even see its face clearly.
So let's see how well it does.
CLICK
HE LAUGHS
It's pretty sure it's a whippet.
Now, I'm pretty sure
that's a Staffie -
but still, very good...
..and again, this one is
not a classic portrait.
CLICK
Ha! Could be a dingo.
It's even a 3% chance it's a lion...
..but still, not bad.
So, three out of three
for the neural network,
but - and it's a big but -
it doesn't have any real
understanding
of what it's looking at.
Let me show you what I mean
with these three pictures.
Now, they look pretty identical
to the first three.
Right, so, picture number one.
It's 100% sure that's a Tabby.
It thinks this dog is a cat.
OK.
Picture number two...
..and this one it's sure
is a baboon.
Right, picture number three...
..and it doesn't think that's a dog,
it thinks it's a trombone.
So, what on earth is going on?
Well, let me tell you.
You see, each of these three
pictures has been altered
ever so slightly
by adding a few pixels.
On the left is the original.
On the right,
with the additional pixels,
chosen specifically to fool it.
The neural network doesn't see
that overall, it's still a dog,
it only responds to pixel patterns
without understanding
what they all add up to.
On the one hand, the networks
perform incredibly well.
You can show them an image
they've never seen before
and they'll get it right -
but on the other hand,
it's incredibly fragile.
I can just tweak any image
and now it gets it wrong.
So, it really raises
a challenge and you think,
"Well, I'm not sure I understand
how this thing works at all,
"if it can be knocked off
so easily."
We can be fooled by optical
illusions ourselves, of course -
but, with a neural network,
it's very hard to understand
what makes it decide
it's looking at a dog,
or a cat, or a baboon,
or a trombone -
whether it's right or wrong.
The network is undoubtedly
in some sense intelligent
but, at the same time, there's no
understanding of concepts there.
It doesn't actually know what a dog
is, let alone anything else,
which is why it can be fooled
with just a few pixels -
and that's where all AI is today -
capable of finding patterns and data
with astonishing detail
and sensitivity,
but with no real understanding of
what those patterns actually mean.
Given this, the hardest challenge
of all for AI
must surely be that uniquely human
ability
that relies on our understanding
of concepts -
language.
So, how well can an artificial
intelligence today
simulate real human-to-human
conversation?
I've come to a top robotics lab
in Edinburgh,
where they're working on how
to make an AI
that can behave like an assistant,
companion, or even a friend.
I'll be talking to a cutting-edge
conversational artificial
intelligence -
or chatbot - called Alana.
Usually, Alana's just a disembodied
voice but today, especially for me,
Alana will be routed through
a rather cute robot.
Hello, good morning,
how are you doing?
Right, let's get to know
one another.
So, Jim, are you more
into sci-fi or fantasy?
I prefer sci-fi.
Awesome! Personally,
I love Star Wars, but back to you.
What is a sci-fi movie
that you enjoy?
My favourite sci-fi movie
is Blade Runner.
There are two Blade Runner movies -
Blade Runner, and Blade Runner 2049.
I prefer the original.
Um, what is going on there, then,
when it's thinking about different
responses and different options?
So, yeah, what's really happening
is that there's several different
programs running in parallel,
and they're all listening to what
you're saying
and they're all volunteering things
that might be good things
to say next.
Inside Alana, we have a Persona bot,
which has a kind of personality
with favourite movies and stuff
like that in it.
We have a Wikipedia bot,
we have a News bot,
and we have a Rapport bot,
and what that's trying to do is to
find out about you
and your personal interests.
So it's almost like a
whole committee of programs,
all vying for their response
to be the one that comes out?
Yeah, that's absolutely right.
So it's got... It's like a whole
bunch of different voices,
all sort of competing
to say what they want to say,
and then what we have
is a ranking function,
which is trained using machine
learning
over thousands and thousands and
thousands of previous conversations
that we've had.
What it's trying to do is choose the
one response which is most likely
to lead to the best conversation
in the future. Right.
The aim of Alana is to handle
general chitchat,
like an AI you could enjoy
a pint with.
It's ambitious, and Alana's ability
to generate a conversation
really is impressive...
I think Roger Federer is
awesome, so I am wondering...
..but it isn't quite there yet.
OK, since you like sports, what is
a sport that you like to watch?
I prefer football.
Great! Which player do you like?
Any player who plays for my
favourite team,
which is Leeds United.
OK, I heard you talk
about Leeds United.
I think that Leeds United
is a 1974 television play
shown on October 31st, 1974, on BBC.
The film is based on an unofficial
strike in Leeds in February, 1970.
Shall I go on?
I'm really not sure whether
I want Alana to go on here or not.
This is actually
an interesting example
of an AI problem in, you know,
figuring out, what is the user
actually talking about?
It didn't really have enough
common-sense knowledge
to realise that when
you say Leeds United,
that's really normally talking
about a football team.
Right. So it turns out, you know,
there are many possible things
you could be talking about
when you say Leeds United.
So what it did is,
it looked in Wikipedia
for anything called Leeds United.
But it highlights what we,
as humans,
would regard as quite
a simple problem
and just how complex it is
to teach a machine. Yeah.
Kind of... Well... I'll forgive...
I'll forgive Alana for that one.
Let's see, if I carry on.
OK, what is a video game
that you played recently?
Well, I haven't played a video game
for about 20 years.
The last one I played
was Gran Turismo.
Regarding Gran Turismo,
I heard on the radio
that Gran Turismo Sport
is now scheduled to be released
October 17th, on PlayStation 4.
Should I tell you more?
No, thank you, Alana,
but that's very interesting.
You are very kind.
So what you actually said is you
haven't played a game for 20 years,
so a human would really infer
from that
that you don't really want to know
a lot about this game
and this kind of stuff,
and this is the kind of thing
that makes language research
so incredibly fascinating,
but also incredibly challenging,
because there's a lot more going on
than just the words that you said.
So we are actually pretty far away
from having the kind of
conversations
you see in science-fiction films.
Speaking of sci-fi...
Alana, if you don't mind, I want to
talk to Oliver for a moment.
You can ask me about my favourites,
if you like.
I, no... What things do you like?
No, I'm talking to Oliver now.
What is actually going on there?
About Oliver, I saw this
on the news.
THEY LAUGH
Headlined Star Wars:
The Last Jedi...
It wasn't the best chat
I've ever had,
but free-flowing conversation
like this
is still a real achievement for AI.
Shall I say some more about this?
So, where does AI go from here?
Getting to the next breakthrough
may be inspired by studying
not so much what adult humans do
as infants.
Ready, and...
Ooh! Agh!
Somewhere between 18 months
and two years old,
children start doing
something remarkable.
Do our rolling.
Show them how to do something just a
few times - often, even just once -
and they start practising it
for themselves.
Cut, cut, cut.
This is called one-shot learning.
Ron, make it longer...
For computer scientists,
who have to train the most
sophisticated AIs
on hundreds or thousands of
examples before they learn anything,
this is like the Holy Grail.
Anyone who works in artificial
intelligence
will appreciate just how advanced
these little humans really are.
They navigate a complex 3-D world.
CLAPPING
They grasp basic physics,
like gravity and inertia.
They formulate plans
and carry them out.
You do that one. Wow! Wow!
I think you can see how nascent AI
is - even still today,
even with all of the successes
it's had -
because when you see all the amazing
things that a toddler learns,
our AI systems are nowhere
near the capabilities
even of a, you know, a two-year-old.
The foundation of their amazing
capabilities
is how much
they've learned as babies.
Since birth, they've continually
explored and experimented,
drinking in information
every second they're awake.
These little children learn directly
from data and experience,
rather like computers
do with machine learning,
artificial neural networks -
but they also understand the world
with abstract concepts.
It's the combination of the two -
the way their learning seamlessly
produces the concepts
and the way the concepts
then direct their learning -
that makes them like the most
amazing computers you can imagine.
It's this combination
that AI researchers
are one day hoping to crack.
AI is developing fast.
No longer just relying on
programmers telling it the rules,
it's learning to do
amazing things by itself,
faster and sometimes even better
than we can.
What's more, it's started to
discover ways of doing things
we didn't know about...
..BUT it's not yet advanced enough
to really learn, or think,
like we do.
Still, if it could one day
rival all our abilities,
I wonder if it might become like us
in another way too.
Could an artificial intelligence
ever have real emotions?
Could it be happy, sad, or jealous?
Could it be social,
or feel friendship, even love?
In short, could it become conscious?
Now, I don't believe there's any
magic pixie dust
that we have to sprinkle over
the grey matter in our heads
to bring about consciousness -
there's nothing our brains do
that couldn't, in principle,
be replicated.
And if AI does one day
become conscious,
we will also have to treat it well.
Not because if we didn't,
it might rise up and destroy us,
but, more profoundly, because
it would be the right thing to do.
Perhaps one day,
we'll even feel it would be cruel
to switch a computer off.
We need to use AI wisely,
and that goes for now,
as well as in the future...
..but if we do, I think humanity
has little to fear
and a huge amount to gain.
I feel inspired by what AI can
already do today
and I believe that, through AI,
we'll greatly extend
our own capacities,
changing our lives
in ways we can't yet imagine.
The evolution of machines that think
must surely be one of the greatest
developments in human history.
On the topic of books, I love
Do Androids Dream of Electric Sheep?
Yes, that's one of my
favourites, too.
Philip K Dick.
That's not appropriate!
HE LAUGHS
Alana, do you know any jokes?
A restaurant nearby had a sign
in the window which said,
"We serve breakfast at any time",
so I ordered French toast
in the Renaissance.
HE LAUGHS