Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:00
This is The Guardian. Today
0:11
on Black Box, two
0:13
visions for the future of AI. Why
0:17
humanity may be doomed
0:19
and why that may be
0:22
an overreaction. branch Finding.
0:29
Your perfect on was hard but thanks
0:31
to burrow furnishing it is never been
0:33
easier. Burroughs. Easy to assemble
0:35
modular sofas and sectional are made
0:38
from premium durable materials including staying
0:40
and scratch resistant fabrics. Plus
0:45
every single Buro order ships free right to
0:47
your door. Right now, get
0:49
15% off your first order at
0:52
buro.com/ACAST. That's 15% off
0:55
at buro.com/ACAST. This
1:03
is a story about the end of
1:05
the world, and it starts with
1:08
a joke. About
1:12
two years ago, the scientist Geoffrey Hinton,
1:15
one of the godfathers of AI,
1:17
was working at Google, tinkering with
1:19
a program most people haven't heard
1:21
of. There was a model in
1:23
Google called Palm, P-A-L-M. It
1:26
was basically a Google version of
1:28
ChatGPT, never released to the public.
1:30
It was just research. And
1:33
Hinton was checking it out when it did something
1:35
that terrified him. Palm could explain
1:37
why jokes were funny. Not all jokes, but quite a
1:40
few of them. Here's one of the
1:42
jokes it could explain. It's someone saying,
1:45
I was going to fly in to visit my family on
1:47
April 6th. And my
1:49
mum said, oh great, your stepdad's poetry reading
1:51
is that night. So now, I'm
1:53
flying in April 7th. I
1:56
know, Not a very good joke.
1:59
But The point is, when you ask... palm
2:01
to why this joke was funny. He
2:03
could tell you. The. Joke is that
2:05
the speaker his mother is trying to get them
2:07
to go to their stepdad poetry reading, but the
2:10
speaker. Doesn't want to go So they are
2:12
changing their flight to the day after the
2:14
poetry reading. Clerics
2:16
to hinson so this was
2:18
really frightening. He has a
2:20
really good sense of humor
2:22
and to him that something
2:24
innately human. To get a joke,
2:26
you need to understand so many
2:29
nuances of language, human nature, how
2:31
people think. He. Told him
2:34
these things he hopes to create.
2:36
They'd. Cross the barrier Another we I'd
2:38
always sort of just for myself use
2:40
that as a cartoon of whether these
2:42
things were really getting to human levels
2:45
of understanding that they. Didn't.
2:47
Just under some language. literally. the and as to
2:49
why joke was funny. So.
2:51
We may be were much close to human
2:53
level intelligence, and I thought maybe we're not.
2:55
So she's a hundred years away. maybe only
2:57
five to twenty his way. I
3:00
will always worried about I'm
3:02
things like fake news, corrupting
3:05
elections. hum. I surveillance things,
3:07
I bust the robots, things
3:09
like our replacing jobs. But
3:12
that's when I got really scared about
3:14
the existential threat. In
3:17
that moment he didn't realize
3:19
the technology he helped bring
3:21
to life is becoming catastrophic.
3:24
Six Getting smarter that he
3:26
ever sex. And
3:29
that was going to be a problem. That.
3:34
They will get a lot smarter than from they were take over.
3:41
Geoffrey Hinton was late to this
3:43
party. Someone. Had gotten a
3:45
much earlier two decades earlier. And
3:48
tried to warn the world's. The.
3:53
first episode black box with about
3:55
the godfather of ai the man
3:58
who spent his career trying to
4:00
bring machines to life. This
4:03
final episode is about Hinton's
4:05
shadow, the godfather of AI
4:07
Doom, a guy who
4:09
spent his entire career trying to
4:12
kill this technology, before he says
4:14
it kills us all, and
4:17
who now thinks it might just be
4:19
too late. From The
4:21
Guardian, I'm Michael Safi. This is
4:24
Black Box. By
4:31
the year 2030, human computers
4:34
can carry out the same amount
4:36
of functions as an actual human
4:38
brain. But
4:41
humanity is accelerating into a future that
4:43
few can predict. Well, is artificial intelligence
4:45
out of control? Open the plot bay
4:47
door, please, help me. You've spoken out
4:49
saying that AI could figure out a
4:51
way to kill humans? How
4:54
could it kill humans? I'm sorry, Dave. I'm
4:57
afraid I can't do that. We've
4:59
come quite far. We
5:01
need to rethink. We need to rethink.
5:12
Ellie Eiser Yudkowsky grew up in
5:14
an Orthodox Jewish household in California.
5:17
He never went to school. He was educated at
5:19
home. I remember being pretty
5:22
studious, much more wrapped
5:25
up in my parents'
5:27
book collection than the
5:30
rest of life, really. But
5:32
the kinds of things Yudkowsky was reading about as
5:34
a young kid were, I would guess, pretty different
5:36
from the kinds of things you and I were
5:39
reading at that age. Nuclear
5:41
power, space travel, asteroid
5:43
mining, the idea that the
5:46
solar system could support a lot more than
5:48
the number of people currently on Earth. How
5:50
old were you when you were getting into that book? I
5:53
think nine years old. Okay, that's pretty
5:55
young to be getting into those sorts of topics.
5:57
I mean, I think I turned out okay. By
6:00
11, he'd moved on to
6:02
even more serious stuff. Ideas like,
6:05
can you scan an entire human
6:07
brain and move the person into
6:09
a computer? Can you build
6:12
artificial intelligence? And this, he
6:14
realised, this is what he was into. So from
6:16
16th I knew that was what I was going
6:18
to be spending my life doing. Yudkowsky
6:21
was coming of age in the late 90s,
6:23
just as the internet was taking off, and
6:25
spending most of his time arguing with people
6:27
on online forums and email threads. And
6:30
that's when he first started to get noticed. He
6:32
mostly emerged from the internet, from the
6:35
very early internet. Tom Chivers
6:37
is a journalist. He's written a book about
6:39
Yudkowsky and his ideas. And there
6:41
was this one essay Yudkowsky wrote when he was
6:43
just 17 years old that caught
6:45
a lot of attention. It was full
6:48
of teenage angst about the state of the world.
6:50
I've got the quote here, like, I
6:52
have had it. I've had it with
6:54
crack houses, dictatorships, tortured chambers, disease,
6:56
old age, spinal paralysis and world
6:59
hunger. I've had it with a planetary death rate
7:01
of 150,000 sentient beings per day. I've
7:04
had it with this planet. I've had it with mortality. None
7:07
of this is necessary. We can end this. And
7:10
he said, there was a solution. One
7:12
day soon, humanity would create
7:15
super intelligent AI. His
7:21
point is, if you build something really, really
7:23
clever, then it can
7:25
solve these problems. The problems we have are the
7:28
problems of humans not being clever enough to solve
7:30
their own problems. That is
7:32
what I hoped for. Everyone lives happily
7:34
ever after, but literally ever after.
7:37
Up until the universe ends and I wasn't too sure about
7:39
us nothing got a way to get out of that. Super
7:43
intelligent AI isn't chat
7:45
JPT. It's not AI that
7:47
can recognise a face or see cancer.
7:50
It's something that can do everything we
7:52
can. Think faster than us, knows
7:55
more, knows everything every human
7:57
who's ever lived has known.
8:00
literally be beyond our comprehension,
8:03
like an ant trying to understand an
8:05
iPhone. A human being thinks
8:07
for a hundred years and then dies.
8:10
What if you have thought for ten thousand years? What if
8:12
you have read everything that's ever been written? And
8:15
then some. This is the
8:18
notion of super intelligence. Yudkowsky,
8:24
again just 17 years old,
8:26
was writing these essays and
8:28
discussing them online with philosophers,
8:30
professors, big time thinkers, and
8:32
holding his own. He
8:35
started gaining a following and persuaded donors
8:38
to give him money to start an
8:40
institute dedicated to solving all the world's
8:42
problems by bringing super intelligence to life.
8:46
Over the next six years he's doing this
8:48
work and begins to engage with another school
8:50
of thought. One that
8:52
says, actually super intelligence wouldn't
8:54
save humanity, it wouldn't usher
8:56
in some utopia. It
8:59
would be dangerous. And
9:01
these arguments that at first he
9:03
dismissed, now they start to nag
9:05
at him. As a young
9:07
lad, I thought that if
9:09
you built something very smart, it
9:12
would automatically be very wise and
9:14
thence very moral. But what
9:17
if super intelligence isn't
9:20
automatically nice? And
9:23
as a result of staring at that issue,
9:25
my entire false belief that very
9:28
smart and capable systems are automatically nice
9:32
started to fall apart. In
9:34
other words, the AI he thought
9:36
would save us all was much
9:38
more likely, in fact logically certain,
9:41
to kill everybody instead. So
9:44
in 2003 Yudkowsky decides he's going
9:46
to turn his life's work around.
9:48
He's no longer going to focus
9:50
on trying to bring super intelligent
9:53
AI into existence. Now
9:55
he's going to try to convince the world
9:57
this stuff shouldn't exist. collection
10:00
of essays and blog posts that
10:02
are collectively longer than war and
10:04
peace, they're just short of the
10:06
length of the Bible, and
10:09
they come to be called the sequences. The
10:14
core of Yudkowsky's message wasn't just
10:16
that super intelligent AI won't be
10:18
nice. It won't be nice,
10:21
but it won't be cruel or evil either.
10:28
He said, you need
10:30
to forget everything you
10:32
think you know about
10:34
killer AI. Forget about
10:36
the Terminator. This
10:38
won't be a scenario where the AI
10:41
becomes conscious, decides it doesn't want to
10:43
be ruled over by humans and launch
10:45
a war against us. Great
10:47
movies, but apparently very
10:49
inaccurate depiction of the
10:51
AI-pocalypse. If
10:54
you're going to use a film for the analogy, the film you should use
10:56
is Disney's Fantasia. The Mickey
10:58
Mouse movie. Yeah, the
11:01
Mickey Mouse movie. If
11:06
you haven't seen Fantasia, it's this old Disney
11:08
movie. Mickey,
11:12
as in the mouse, is some
11:14
kind of trainee wizard, and one of his
11:17
jobs is to fill a cauldron with water.
11:20
But that's hard work, and it's boring, it takes
11:22
a long time. So Mickey enchants
11:24
a broom and asks it to take
11:26
some buckets and fill up the cauldron
11:28
instead. And then the broom
11:30
grows little arms, it grabs
11:33
a bucket in each arm and bottles
11:35
off to the well. A few hours
11:37
later, the castle is completely flooded. A
11:40
broom had followed Mickey's instructions
11:42
perfectly, filled up the cauldron, but
11:45
once it was full, it just kept
11:47
going. Figured
11:52
Kasky, the key thing here, is
11:54
that the broom wasn't evil. The
11:56
broom had no opinions, it was just doing
11:59
what Mickey asked. What
12:01
Mickey asked was the problem. When
12:03
he said fill up the cauldron,
12:05
he expected the broom to understand
12:07
all of these unspoken nuances. Like
12:10
at some point you'll need to
12:12
stop. If the room starts flooding
12:14
there's a problem. But
12:16
why would a broom know that? The broom
12:19
just does exactly what you tell it
12:21
and so would an AI. It just
12:24
seems mad but it's because we have human
12:26
brains with human values and
12:28
human philosophies and all
12:31
these things which it simply doesn't have. So it's
12:33
sort of our model of
12:35
human wisdom does not apply to it. The
12:37
sort of fundamental idea of AI risk is that
12:41
the AI will do exactly what you tell it to do and
12:43
that it can only do exactly what you tell it to do
12:46
but what you tell it to do is
12:48
not what you want it to do. The
12:50
point is you can't just tell an AI
12:52
what you want it to do. You need
12:55
to tell it all the things it shouldn't
12:57
do to achieve that goal and
12:59
things can very quickly get out of control.
13:02
Let's take another domestic chore. Say
13:05
you build a super smart all capable
13:07
AI robot and say clean
13:09
my house. Simple. And
13:12
at first it does. It sweeps up, it
13:14
mops the floors and then
13:16
it decides. An even better way to achieve
13:18
this goal would be to have
13:20
thousands of versions of itself spread
13:22
out across the whole house cleaning
13:24
all day every day. So
13:27
much cleaning that eventually you can't even
13:29
live in your house anymore because every
13:32
surface is being cleaned all the time.
13:35
You just have to leave lock the door and
13:37
find somewhere else to live. That's
13:39
not the end of it. At some point
13:41
this robot realizes at this rate the world
13:44
is going to run out of cleaning products
13:46
and if that happens it can't achieve its
13:48
goal. So it goes and
13:50
strips the planet's resources and puts
13:52
them all into creating a nearly
13:54
endless supply of Mr. Muscle and
13:57
the last evidence of human civilization.
14:00
on this earth, the
14:02
last thing standing is your
14:04
astonishingly clean house. I
14:14
know, this all sounds a bit mad, but
14:17
the idea here is, it's really
14:19
hard to align what you want
14:21
with what AI actually does. And
14:24
once you create something incredibly
14:26
smart, incredibly powerful,
14:29
you'd only get one go at getting it
14:31
right. So
14:34
maybe you're thinking, okay, that's
14:36
not great, but why don't we
14:39
just give it an off switch? If
14:41
the AI starts to get a bit crazy, pull
14:43
the plug, turn it off. But
14:45
Elieisen and others say, no, that
14:48
wouldn't work. It is very hard to fulfill
14:50
your goals if you're switched off. So if
14:53
the AI has reason to think that you will be
14:55
coming to switch it off, then it will take steps
14:57
to avoid that. If I decided that you were dangerous,
15:00
and I should switch you off, you would
15:02
try and resist my doing that, I
15:04
imagine. And for the same
15:07
reasons, an AI would try and
15:09
resist me switching it off if I thought it was going to be dangerous, and
15:11
it would be better at doing it than you are. The
15:13
point is here, if you're in a position where this thing's
15:15
trying to outsmart you, it's going to outsmart
15:18
you. I mean, by definition, it's super intelligent. Yeah,
15:20
yeah, that's exactly it. Maybe
15:23
its next step is to make copies of itself around
15:25
the internet. If it's really, really powerful and has control
15:27
of nuclear codes, maybe its next step is to nuke
15:29
everybody. Okay,
15:31
so you can't count on being able
15:33
to turn it off. But
15:36
everything Tom just said assumes that the
15:38
AI is somehow connected to the real
15:40
world. It wouldn't be able
15:42
to copy itself to the internet or access
15:44
the nuclear codes. If you just kind of
15:46
stick it in a box, put
15:49
this Oracle-like super intelligence on
15:51
a single computer, lock
15:53
it in a room, and disconnect it from the
15:55
internet. And if you wanted to help in
15:57
solving a problem, you just go into the room and have
15:59
a- a conversation with it. Sorted.
16:02
Right? Apparently not.
16:04
There's a danger in just talking
16:06
to a thing that understands you
16:08
far better than you understand yourself. His
16:12
argument here boils down to one question.
16:14
Do you really think the
16:17
smartest entity that's ever existed
16:19
couldn't convince you, manipulate
16:22
you into setting it
16:24
free in the world? Really.
16:26
It's not like there's a sort
16:28
of mechanical lifeless mind trapped in
16:30
a computer. It's like there's
16:32
an alien civilization that thinks a thousand times
16:35
faster than Earth trapped inside a
16:37
computer. If you were
16:39
in those shoes, how would you get into the
16:41
real world? In any case,
16:43
Elieizer says, the boxing is probably a
16:45
moot point because that's not how things
16:48
are going. Open AI does
16:50
not keep GPT-4 in a box unconnected
16:52
to the internet. They trained it
16:54
connected to the internet from the very start because
16:56
the computers that train things are connected to the
16:58
internet. And then what
17:01
was profitable was to connect GPT
17:03
to the internet and sell it for $20 a
17:05
month for everyone. You're describing
17:07
a world where humanity
17:10
can somehow just decide
17:12
en masse. This technology is so
17:15
dangerous that needs to be kept in a tiny
17:17
box. We do not have that
17:19
capability at the moment. We should, but
17:21
we don't. I
17:23
don't know. Maybe all of this sounds
17:25
a bit far fetched. But Elieizer says, if
17:27
you sit down and think about it, you'll
17:30
see it's all rational. And
17:33
over the years, a lot of people
17:35
have taken Elieizer's ideas about how
17:37
to think rationally about AI, how
17:39
to think rationally about every problem
17:41
in the world very seriously. They
17:44
spawned a whole movement. People
17:47
who call themselves rationalists, devoted to
17:49
the ideas laid out on his
17:51
website. It was read by
17:53
people very much in Silicon
17:56
Valley or in computer
17:58
science names who were to
18:00
be involved in DeepMind or who wrote
18:02
AI textbooks and things like that. It
18:05
was an influential thing. One thing that
18:07
I find amazing about Elieizer's story is that
18:10
when he was outlining these ideas, warning
18:12
the world about AI, AI
18:14
itself was a bit of a joke,
18:17
barely able to tell the difference between a dog
18:19
and a cat. But he
18:22
was absolutely convinced and
18:24
worried that much smarter AI
18:26
was coming. And
18:29
on that at least, he turned
18:31
out to be exactly right. There's
18:34
no person who made these
18:36
captions. It was machine made.
18:38
Yes, it's true. Gradually.
18:41
And then rapidly. A computer program has
18:44
just beaten the 9-9 professional. With
18:47
Yudkowsky warning at every step,
18:50
slow down. This is bad news.
18:52
That you can actually screw up
18:55
and the entire human species stands
18:57
at risk. Jeffrey Hinton's
18:59
vision came to life. This now
19:01
has spoken conversations and have been
19:03
translated between two languages almost instantly.
19:06
Neural networks started to work. Alexa,
19:08
what do you do? I
19:10
can play music, answer questions, get the
19:13
news on weather. Powerful AI
19:15
systems capable of deep learning
19:17
and like the human brain,
19:19
too complex for us to
19:21
understand. Black boxes. While
19:25
Elieizer watched on in horror.
19:33
Elieizer, in the years that you were
19:36
writing about these topics, gaining followers and
19:38
influence, we also started
19:40
to see significant progress when it
19:42
came to AI, especially with neural
19:44
networks. What was it
19:46
like for you to watch the
19:48
field start to progress? Pretty
19:51
horrible. Deep learning
19:53
is pretty much a worst case scenario
19:55
from the perspective of being able to
19:57
align or even understand AI systems. That
20:00
was a death sentence back when I
20:02
was getting into this field There
20:04
nobody knew that it was going to be giant
20:07
inscrutable matrices that were going to win all the
20:09
prizes But there was still a
20:11
hope back then that by the time you
20:13
got to something like chat GPT You'd
20:15
have some idea what it was thinking of there and
20:17
that is not what we have what we have is AI
20:20
that is You know barely a
20:22
step above how natural selection built humans completely
20:24
blind We do not have a eyes
20:26
that are built. We have a eyes that are grown and Then
20:30
at that point it's hard to control
20:32
because it's hard to understand a bit
20:34
over a year ago All of
20:36
a sudden since the world started
20:39
talking about AI and right now
20:41
chat GPT is rattling the AI
20:43
world Chat GPT arrived
20:45
they're ready to laugh cry Maybe
20:47
even throw your computer at the
20:49
window because the latest AI technology
20:51
chat GPT is here to shake
20:54
things up And just like AI that could
20:56
have a conversation past the
20:58
bar exam Discussed the plot
21:00
of any book Artificial
21:02
intelligence that actually seemed intelligent It
21:05
is writing better than most of
21:07
my students write at this point
21:09
You know college freshmen and as
21:11
people started to understand we had
21:14
entered a new age Yudkowsky
21:17
kept trying to raise the alarm now to
21:19
much bigger audience I tried to get this
21:21
very important project started early so we'd be
21:24
in less of a drastic rush later I
21:27
consider myself to have failed Nobody
21:31
understands how modern audiences didn't
21:33
always know what to make
21:35
of his apocalyptic message But
21:37
that they would why would they want to go
21:39
in that direction? Like
21:42
that AIs don't have our
21:45
feelings of sort of envy and jealousy
21:47
and anger and but he had one
21:49
high profile person Who last
21:51
May announced he had come over
21:53
to Yudkowsky's way of thinking Jeffrey
21:56
Hinton says he wants to be free
21:58
to talk about the possible data of
22:00
technology that he helped to develop. Not
22:03
about everything, but about the bigger
22:05
point that AI could
22:08
be, and in fact was likely
22:10
to be, extremely dangerous. Given the
22:12
rate of progress, we expect things
22:15
to get better quite fast, so
22:18
we need to worry about that. And in
22:20
the months since then, fears around
22:22
AI have become just as prominent
22:24
as hopes of what it could
22:26
one day do. We face a
22:28
genuine inflection point in history. One
22:31
of those moments where the decisions we make in the
22:33
very near term are going
22:35
to set the course for the next decades. The
22:39
Biden White House in October issued a
22:42
major executive order that will force the
22:44
biggest AI companies to share safety data
22:46
with the government before they release their
22:48
models to the public. A
22:51
few months ago, Rishi Sunak gathered global
22:53
leaders at a summit dedicated to the
22:55
existential risks of AI. Get
22:58
this wrong, an AI could
23:00
make it easier to build chemical or biological
23:02
weapons. Terrorists
23:05
could use AI to spread fear and destruction
23:07
on an even greater scale. Criminals
23:10
could exploit AI for cyber attacks,
23:13
disinformation, fraud or even child
23:15
sexual abuse. And
23:17
in the most unlikely but extreme cases,
23:19
there is even the risk that humanity could
23:22
lose control of AI completely. Last
23:30
year was this like watershed year
23:32
in AI safety. The things you'd
23:34
been talking about for two decades
23:36
suddenly became mainstream. Was
23:38
that gratifying to you as someone who'd
23:41
advocated for those issues or not gratifying
23:43
at all? It is not
23:45
particularly gratifying in the way that I think you're
23:47
thinking. People
23:51
who need the gratification of seeing
23:53
others agree with them will
23:55
not go into a field like
23:58
AI alignment and work on it. being
24:00
largely ignored for 20 years. They
24:02
will pick something that delivers more of that
24:04
gratification earlier. I have a low
24:06
need for that form of gratification. It
24:09
was, in a way, humbling. Because
24:11
I had written off humanity's response to this
24:13
issue. I thought that everybody was going to
24:16
just sort of never think about it or
24:18
talk about it sensibly at all, the way
24:20
that people in Silicon Valley and the companies
24:22
building these things had never thought about
24:24
it at all. I thought
24:27
that humanity was going to do worse
24:29
than Silicon Valley. Humanity did better. And
24:32
when you're cynical and you're wrong, I think that's meant to
24:34
be a bit humbling. For
24:36
Yudkowsky, this moment that his
24:38
ideas broke through to the mainstream, it
24:41
was bittersweet. Because he
24:43
had come to the conclusion that by now, it
24:45
was too late. He thinks
24:48
we've nearly run out of time. He
24:51
says it's hard to say when AI
24:53
might reach a level, when it's as
24:55
smart as a human. I have
24:57
a sense that our current remaining timeline looks more like
24:59
five years than 50 years. But
25:01
once it does reach that point, he
25:04
says it won't stay there for long. Superintelligence,
25:09
at the very least, should be better than any
25:11
human at building the next
25:13
generation of AI. And
25:16
then that improved version
25:18
will improve itself again,
25:20
and again, and again, and again. An
25:25
intelligence explosion. Yudkowsky
25:30
says we can avoid this.
25:32
There is a way. And that
25:34
is shut it down. Pause
25:38
all AI development right
25:40
now. Sign treaties, pass
25:42
laws, drop bombs on rogue data centers
25:44
if you need to. Whatever
25:47
it takes. Do
25:51
you have any hope that we
25:53
will survive this, that
25:55
we will develop super
25:57
intelligent AI in a way that does not
26:00
not result in the worst outcome
26:02
for humanity. Yeah, that we
26:04
back off and don't do it. And
26:07
the time to do that is running out
26:09
as we build more chip factories as the
26:11
technology proliferates. But if
26:14
we woke up tomorrow and just like decided
26:16
we'd rather not die this way, we
26:18
could do it. You say nobody's allowed
26:20
to train something more powerful than GPT-4. Humanity
26:24
could decide not to die and it would not be that hard.
26:27
The difficulty is that people don't realize
26:29
they're about to die. If
26:31
people knew that, we could just
26:34
stop. It
26:36
wouldn't be trivial, but we could. Coming
26:51
up, don't panic. There are
26:53
other views of how this night all
26:55
play out. Maybe rather
26:57
than AI killing us all, it
27:00
could be really good. Finding
27:16
your perfect home was hard, but thanks
27:18
to Burrow, furnishing it has never been
27:20
easier. Burrow's easy to assemble modular sofas
27:23
and sectionals are made from premium durable
27:25
materials, including stain and scratch-resistant fabrics. So
27:27
they're not just comfortable and stylish, they're
27:29
built to last. Plus every single Burrow
27:32
order ships free right to your door.
27:35
Right now get 15% off
27:37
your first order at burrow.com/Acast.
27:39
That's 15% off at
27:42
burrow.com/Acast. Introducing
27:44
Wondersuite from bluehost.com. Website
27:47
creation is hard, but now with Bluehost,
27:49
you can answer a few simple questions
27:51
about your business and get a unique
27:54
WordPress website or store right away. From
27:56
there, you can customize your design, colors,
27:59
and content. And Bluehost
28:01
automatically helps you get found
28:03
in search engines like Google
28:05
and Bing. From step-by-step guidance
28:07
to suggested plugins, Bluehost makes
28:09
WordPress wonderful for everyone. Go
28:12
to bluehost.com/Wondersuite. After
28:19
speaking to Elieizer Yudkowsky,
28:22
we were terrified. And
28:26
throughout this series, whenever we've hit a brick
28:28
wall, come across some idea
28:31
we can't untangle. There's
28:33
one person we've gone to for
28:35
advice. The Guardian's
28:37
UK technology editor, Alex Hearn.
28:41
For most of the time we've been working on this
28:43
series, he's been on paternity leave.
28:46
But we told him there's something we really
28:48
need to ask him. Are
28:51
we all going to die? No.
28:57
That's really true. Yeah, he says,
28:59
well, I say not necessarily. It's
29:02
not the total dismissal that you
29:04
might hope. I
29:06
think his fears are worth taking
29:08
seriously. I don't
29:10
think it is imminent, and I don't think it's guaranteed.
29:14
So where do you
29:16
stop short of going all
29:19
the way with him? What is it that gives
29:21
you hope that maybe this scenario he's laying out
29:23
is not exactly how it's going to play out?
29:26
I think the stuff that stops
29:28
me short is just that ultimately,
29:30
for the everyone's going to die
29:32
end, for the proper doomerism argument,
29:35
there is still an element of, if
29:38
I'm being mean I say magical thinking. And
29:41
sometimes those things involve like it will
29:43
engineer a nanovirus and seize 3D printers
29:45
around the world and distribute
29:49
an AI engineered deathbot
29:52
that takes out the whole of
29:54
the human race. And that sort of thing
29:57
isn't possible based on what we're doing.
30:00
we know about biology and
30:02
physics, there are, we
30:05
think, limits to what you can achieve
30:07
through pure reason alone. And
30:10
that is a very real
30:12
limit on what superintelligence can
30:14
do out of nowhere. I
30:17
think there are lesser versions of the
30:19
existential risks that I take more seriously.
30:22
The big one is this idea of
30:24
it being so good that
30:27
we slave ourselves to its power
30:29
voluntarily, that we wake up in 10 or 20
30:31
years time and we realize
30:33
that all meaningful power in
30:36
the world has been effectively voluntarily devolved
30:38
to one or a number of
30:40
AI systems, because if you
30:43
as the prime minister of a small or
30:45
medium-sized country don't let the AI tell you
30:47
exactly what to do, then your country's outcompeted.
30:49
If you as the CEO of a business
30:51
don't hand over almost all your
30:53
decision-making power to an AI, then your businesses
30:56
aren't competed. Like, that's
30:58
not an existential risk in
31:00
the classical sense of things, but it's
31:02
what I worry about if these systems
31:04
get too good. So
31:06
for you, in the short term, the
31:09
concern is less the end of humanity
31:11
in some extraordinary explosion
31:13
of intelligence. It's
31:15
the threats that seem more
31:17
foreseeable. AI taking control
31:19
of our workplaces, of our personal
31:22
relationships, dictating the way that we
31:24
live our lives. And
31:26
even taking is perhaps the
31:28
wrong word, right? Being handed control
31:30
of that, because if you create
31:33
a tool that is that good at
31:36
doing the jobs it's handed, then not
31:38
using that tool becomes ridiculous. Alex
31:41
told me that a lot of
31:43
this kind of conversation about whether AI
31:45
will be nice or mean, good
31:48
or bad, it's too simple.
31:51
It all hinges on another
31:53
question, and that's how
31:55
intelligent will this stuff end
31:57
up being? Whether...
32:00
AI will be super smart, all
32:02
powerful, the stuff of
32:05
Yudkowsky's nightmares. Or
32:07
maybe something less than that. What
32:10
we get are pools that
32:12
are tremendously useful, perhaps even
32:15
transformatively useful, but
32:17
not changing society
32:20
in a way that is
32:22
different from how the railroad,
32:24
internet and electricity change society.
32:28
Just noting here, even the
32:30
lower end scenario for Alex is
32:32
that AI is as transformative as
32:35
electricity. And that's bad enough that
32:38
we destroy our healthy media
32:40
ecosystem. That we destroy any
32:42
sense that you can trust, anything
32:44
that you hear that isn't from the mouth of
32:46
someone you already trust. That's
32:49
one where we have mass
32:51
unemployment as huge swathes of
32:53
easily automatable jobs that
32:55
get removed, cut down on importance. That's
32:59
a world that is the industrial
33:02
revolution, but more so
33:04
and faster. But lots
33:06
of people argue that as well as
33:08
causing upheaval, this kind of AI, powerful
33:11
but not super intelligent, could
33:14
be a really good thing for the world. Actually
33:18
that's where I
33:20
sit. That's the world where things
33:23
get slightly better, faster than they
33:25
used to, forever. And
33:28
it is because that's how the
33:30
world has worked. In
33:32
general, the path of
33:36
the last two, three hundred years of humanity
33:38
has been that while there
33:41
may be transition
33:44
problems, that life has gotten,
33:46
continues to get overall better
33:49
for lots and lots of people based in
33:51
large part on technological improvement. I
33:54
wouldn't want to live in the
33:56
Britain of 50 years ago. just
34:00
differences in things like child mortality. My
34:03
daughter has Down Syndrome. The difference in
34:05
the quality of life for her now versus 50 years
34:07
ago is unspeakable.
34:11
There's just wild improvement.
34:15
And I think that has to be your default for
34:18
what technology can do. What
34:21
kind of world do you think one
34:23
that is dominated by this kind of
34:25
powerful but not super intelligent AI will
34:27
actually look like? I
34:30
think you can
34:33
start looking through almost every segment of
34:36
society and go, okay,
34:39
well, what happens if everyone who
34:43
works here is able
34:45
to think to the best of their ability for
34:51
eight hours a day with a full team of personal
34:53
assistance and
34:56
research assistance beside them, right? There's
34:58
a world where scientific research,
35:00
say, gets
35:02
supercharged. If you have
35:05
every biochemist, every geneticist,
35:09
every industrial chemist, all of them working
35:12
through the absolute top of their ability, then
35:16
suddenly perhaps science becomes easier or
35:20
faster or more efficient. And it's not just
35:23
science. It's every segment of society where you
35:25
could have that. The
35:27
world is full of people not
35:31
working to their full potential
35:33
because we are tired, we are unsupported, we
35:36
are fuzzy headed. Alex
35:38
reckons in this scenario, AI would touch just
35:41
about every aspect of the way we live
35:44
and just make it better. In
35:47
education, a personalized tutor for
35:50
every child on the environment, finding ways to make
35:52
renewable energy even
35:54
more efficient, maybe even discovering new. nuclear
36:00
fusion, a source of unlimited,
36:02
clean energy. On
36:04
health, finding new antibiotics,
36:06
predicting the structure of
36:08
every known protein, both
36:11
of which, by the way, AI has
36:13
already done. And
36:15
then the thing Alex says we should
36:17
really keep an eye on this year,
36:19
which is the fusion of AI and
36:21
robotics. Just days before we
36:23
recorded this, a company called Figur published
36:26
a video of an open
36:29
AI powered humanoid robot that it's been
36:31
working on. Hey, Figur1, what do you
36:33
see right now? I
36:36
see a red apple on a plate in
36:38
the center of the table, a drying rack
36:41
with cups and a plate, and you standing
36:43
nearby with your hand on the table. Great,
36:46
can I have something to eat? It
36:48
is a machine
36:50
that can be instructed
36:53
with normal words to do
36:56
everyday domestic tasks, and
36:58
explain what it's doing and why it's doing
37:00
it. The dishes on the table like that
37:02
plate and cup are likely to go into
37:05
the drying rack next. Great,
37:07
can you put them in? I mean, come
37:09
on, it is science
37:11
fiction. And it's
37:13
easy to see the ways that could lead to
37:15
a terrifying world, but I think, I
37:19
don't think it's wildly optimistic
37:23
to think that that
37:26
would be nice to have, wouldn't it? Alright,
37:28
so how do you think you did? I
37:31
think I did pretty well. The
37:33
apple found its new owner, the trash is
37:35
gone, and the tableware is right where it
37:37
belongs. If
37:39
you can start getting help
37:42
from robotic systems, then
37:46
there is an optimistic world there, one where
37:48
people like my great grandmother, who
37:51
don't need round-the-clock
37:53
care to survive,
37:56
but need round-the-clock care to thrive,
37:59
she's not going to be... stuck alone deprived
38:01
of her one human contact,
38:03
she's still going to be going to see
38:05
her friends, to play bridge, to go on
38:07
social calls, but she's also no longer going
38:09
to be at risk of forgetting to eat. She's no
38:12
longer going to be at risk
38:14
of falling over and having no
38:17
ability to summon help for a
38:19
day or two. And that
38:21
being opened up to everyone who needs it,
38:25
I'm really optimistic that that
38:28
would be the sort of thing that
38:31
can transformatively change lives for literally
38:33
millions of people. And
38:36
that's one use of one technology
38:39
that this revolution is
38:43
creating, right? I genuinely
38:45
think even with
38:48
the technology that exists today, it
38:50
is already getting hard to come
38:52
up with people
38:55
whose lives couldn't in some way be helped
38:57
by this. And that's
39:01
that's a staggering press supposed to be standing
39:03
on the edge of. How
39:09
much power do you think we have as
39:11
people, as citizens, over
39:15
the way AI ends up shaping
39:18
our lives? How much of it is down to politics
39:21
and the way that we choose to
39:23
order our societies? I think we
39:25
have a lot of power. I think ultimately
39:27
we can order society and
39:29
go, you know what, actually, we're not going
39:31
to allow ubiquitous
39:35
facial recognition and behavioral monitoring. We're
39:37
not going to become an authoritarian
39:39
society the EU has
39:41
just passed laws that allow it
39:44
to poke and prod frontier AI
39:46
models independently of their developers.
39:48
That's huge. That means that there will be
39:50
an independent
39:52
safety check. You no longer have
39:54
to trust that open
39:57
AI or Google or Microsoft are doing
39:59
this stuff. off well, you
40:01
have a government putting a second pair of eyes on
40:04
things. Alex,
40:08
one of the things this series has been about is
40:10
the realization that we've crossed from a before
40:13
into an after, into this
40:16
world where machines can learn
40:18
on their own, and that
40:20
is an extraordinary and profound
40:22
transformation. For people who
40:24
are trying to understand what it all means
40:26
for them, where they might fit into it,
40:29
what would you tell them? I
40:33
would say if you
40:35
are dipping your toe into the world, sort
40:38
of persist. It's
40:41
an incredible thing that has been
40:43
created, and it's
40:45
going to get better. I
40:48
guess there is a bit of fatalism in that if you don't
40:50
embrace what it can do
40:52
for you, then you're just going to be affected
40:54
by what it's doing for others. The
40:58
only way you can seize agency in this is
41:00
by seizing it, and not
41:02
just burying your head in the sand and hoping
41:05
the world won't change. Because
41:08
I think it's going to. Thanks
41:30
very much to Alex Hern, the Guardian's
41:32
UK technology editor. You can read more
41:34
from him by signing up for his
41:36
newsletter called Techscape. Thanks
41:38
very much also to Tom Travers. His
41:40
book is called The Rationalist's Guide to
41:42
the Galaxy, and it's available right now.
41:46
Black Box is produced by Alex
41:48
Attach. The executive producer is Judge
41:50
Kelly. The commissioning editor
41:52
is Nicole Jackson. All
41:55
music and sound design by Rudy Zagatlo.
41:57
The music supervisor is Max Sanderson.
42:00
This series took us six months to make and we're
42:02
so glad we can bring it to you for free.
42:05
If you'd like to support projects like
42:07
this one, consider becoming a supporter of
42:09
The Guardian, which you can do by
42:12
going to support.theguardian.com. And
42:14
otherwise, if you enjoyed this series, if you
42:17
thought it was worthwhile, please do leave us
42:19
a rating and a review. It'll help other
42:21
people find us. Black
42:23
Box will return sometime in the
42:26
future. This
42:42
is The Guardian. Tired
42:53
of ads barging into your favorite news
42:55
podcasts? Good news! Ad-free listening
42:58
is available on Amazon Music for all
43:00
the music plus top podcasts included with
43:02
your Prime membership. Stay up
43:04
to date on everything newsworthy by downloading the
43:06
Amazon Music app for free or
43:09
go to amazon.com/news ad free.
43:11
That's amazon.com/news ad free to catch
43:14
up on the latest episodes without
43:16
the ads. Hackers
43:19
and cyber criminals have always held
43:21
this kind of special fascination. Obviously,
43:23
I can't tell you too much about
43:25
what I do. It's a game. Who's
43:27
the best hacker? And I was like,
43:29
well, this is child's play. I'm
43:32
Dina Temple Rustin, and on the Click Here
43:34
podcast, you'll meet them and the people trying
43:36
to stop them. We're not afraid of
43:38
the attack. We're afraid of the creativity and
43:40
the
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More