Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:00
Remember. To take a rescue. What?
0:02
About the mission depicted in Black Hawk
0:04
Down Or the epic rescue shown in
0:06
Captain Phillips. You've. Probably heard of all
0:08
of these, but did you know that the Us
0:10
Air Force Special Warfare played a pivotal role in
0:12
all of them? These. Airman are
0:14
the most highly trained warriors on the planet.
0:17
Other forces like the Seals and Army
0:19
Rangers call on them to provide skills no
0:21
one else can. Not. Many people make
0:24
the cut. If you think you can visit
0:26
Air Force that got to learn more. When.
0:29
You're ready to pop the question the
0:31
last thing you want a deal as
0:33
second guess the ring. At bluenile.com You
0:35
can design a one of a kind
0:38
ring with the ease and convenience of
0:40
shopping online. Choose your down and and
0:42
setting. When you found the one, you'll
0:44
get it delivered right to your door.
0:46
Go to bluenile.com and use promo code
0:48
Welcome To get fifty dollars off your
0:51
purchase A Five hundred dollars or more.
0:53
That's code Welcome Make Bluenile dot Com
0:55
for fifty dollars off your purchase. Bluenile
0:57
That Com code Welcome. I'm.
1:05
Alan Alda and this is
1:08
clear in vivid conversations about
1:10
connecting and communicating. Welcome
1:16
do a preview of Season Twenty Four.
1:18
I'm here with our executive producer Graham
1:20
Shared to give you a little taste
1:22
of what's coming up in this notable
1:24
season. One thing is notable to me
1:27
Graham is that this season will include
1:29
are three hundred shows. It's clear and
1:31
vivid began. Hard to believe. It's
1:33
a lot of conversation with some of
1:35
the most interesting people in a wide
1:37
range of fields, including some people who
1:39
don't exist like the robots or he
1:42
interviewed last year. Which brings
1:44
us to a special series of three
1:46
shows, restarting the season off with. Since.
1:49
The theme of our show is connecting and
1:51
communicating. Were taking a dive into what may
1:54
be the more meant is changing the way
1:56
we relate and communicate in our history. More
1:59
and more. In communication with something
2:01
that sounds human. But. Isn't.
2:04
Artificial. Intelligence is already affecting
2:06
our lives and making progress at
2:09
a very fast clip which is
2:11
both good and bad. Were.
2:13
Told by people who were creating
2:15
a I that he can radically
2:17
make our lives better, for instance,
2:19
by eliminated diseases. Or
2:22
he can make our lives worse
2:24
by eliminating us. That.
2:26
Sounds like a deserve some attention. So.
2:29
He invited three people on the
2:31
show, each with their own unique
2:33
perspective on a I started with
2:35
face, a leader who is often
2:37
called the godmother of ai his
2:39
Instagram vs somewhat to her embarrassment.
2:42
She made a breakthrough that revolutionized a
2:44
I when she realized that the key
2:46
to making artificial intelligence intelligence is some
2:48
the harvesting of huge amounts of data.
2:52
And her personal story is inspiring.
2:55
Born in China and nineteen seventy six, she
2:57
emigrated with a family to the U S
2:59
when she was a teenager. To
3:01
spoke little or no exists, but so impressed
3:04
some math teacher a New Jersey high school
3:06
that he meant to hurt her and help
3:08
to get into Princeton as an underground. He
3:11
saw something special in her. It
3:14
really started with says it's
3:16
I was I don't know.
3:19
like since I was a
3:21
little girl like eleven twelve
3:23
year old. I just loved
3:25
physics. A was my first
3:27
love and in hindsight, What?
3:30
Happened as I think I
3:32
love that or thesis quest
3:34
to the unknown mystery of
3:36
the universe like physics allows
3:38
you to ask the craziest
3:40
quests that was like beginning
3:42
of space time, founder else
3:45
the universe, the smallest particle
3:47
of or matter and. Then.
3:51
in the middle princeton says
3:53
x i discovered even live
3:55
feather for themselves like albert
3:57
einstein and are resorting there
4:00
So they turned their
4:02
attention somewhat to an
4:04
equally audacious question. But that's
4:06
not a physics question. It's
4:08
about life. And
4:11
I became so enlightened
4:13
and enamored, I realized my
4:15
own audacious question that I
4:17
love the most is, what
4:20
is intelligence? What
4:22
makes intelligence? How do we
4:24
build intelligent machines? And
4:26
that shift in the middle
4:28
of end of my college year was
4:32
how I discovered AI. How
4:35
did you get from that to concentrating on
4:37
images? I think I'm
4:39
naturally a visual person because
4:41
even in my early childhood,
4:43
my dad takes me to
4:45
these natural excursions and
4:47
we look at butterflies, we draw
4:49
the pictures of mountains. And
4:52
there is this fascination of seeing. I
4:54
find that understanding visual
4:57
intelligence to be the most
4:59
fascinating aspect of intelligence. And
5:02
I think that is
5:05
kind of a combination of serendipity. And
5:07
I just got into vision. I
5:10
think I've heard you say that vision
5:12
is more than just a sense that
5:16
it's an experience. Vision
5:19
is intelligence. Vision is
5:21
experience. Vision is understanding.
5:25
And vision is planning.
5:27
Vision is decision making.
5:29
Vision is socialization. Vision
5:32
is a very
5:34
cornerstone piece of intelligence
5:36
itself. Faith
5:38
Ailey's fascination with both vision and AI
5:41
led her to try to build a
5:43
machine that could identify images. This
5:46
was at a time in the early 2000s
5:48
that AI research was sort of stuck. It's
5:51
often called the AI winter. Her
5:53
insight was to train a computer to recognize
5:56
images by showing it lots and lots of
5:58
images, millions and millions of images. pictures
6:00
from the internet. Fei-Hui
6:02
Li called her computer vision model
6:04
ImageNet. That's the
6:06
project that made people
6:09
call you the godmother of artificial intelligence
6:11
as we know it today. I
6:14
know you don't want to congratulate yourself
6:16
too much, but I've heard
6:18
that said about you and I'm trying to figure out in
6:21
what way was it a milestone? I
6:24
think it's best explained with
6:27
actually today's breakthrough in chat
6:29
GPT. Why is it
6:32
that we've seen the AI
6:34
breakthrough because we see powerful
6:37
algorithms trained on a
6:39
vast amount of data, the data
6:42
of the internet. And that's where
6:44
ImageNet came to play the pivotal
6:46
role, is that my students
6:48
and I recognize the
6:50
power of data. We hypothesized,
6:53
I guess
6:55
before most people, that AI
7:00
will have a paradigm shift
7:02
if we power it with
7:05
internet scale, giant amount
7:08
of data. It's a
7:10
data-centric, data-first approach.
7:14
And because of that, we were working on
7:16
vision. So we want to
7:18
make the biggest visual dataset. And
7:21
in order to make the biggest
7:23
visual dataset, we had this
7:26
crazy idea of downloading
7:28
almost all the pictures we
7:30
can get on the internet
7:32
back in 2007 and
7:36
organizing curated catalogues
7:39
in this completeness in terms
7:42
of visual objects. And
7:44
that's when we made, after three
7:46
years, between 2007 to 2009,
7:48
we made a dataset of 15 million images
7:51
across 22,000 categories. And
7:57
that's out of cleaning up a billion
7:59
images. What do you mean
8:01
by categories? Categories
8:04
are the natural way that
8:06
humans conceptualize objects. We tend
8:08
to conceptualize them as German
8:11
Shepherd's, microwave, a sport
8:14
car. Of
8:18
course, sometimes we think about
8:21
my German Shepherd, your microwave.
8:23
But in general, that
8:26
classification of visual concept is
8:28
a fundamental visual
8:31
intelligence problem
8:33
that humans have worked on
8:35
and solved, and it's very
8:38
foundational to our visual intelligence.
8:40
So if you collect a
8:43
great number of pictures under
8:45
the category of dog, and a great
8:47
number of pictures under the category of
8:50
cat, the machine is able to sort
8:52
through that and put a
8:54
name on it when it sees a picture. Yes,
8:57
and mind you, there are hundreds
8:59
of dog species. So you image
9:01
that, it's not just dog, we
9:03
actually had hundreds of different dogs.
9:06
Carrier, German Shepherd, Corgi,
9:08
we have even different kinds of
9:11
Corgis, so it's a lot more
9:13
than just dog versus cats. Right,
9:16
right. So you've got to
9:18
give subcategories as well? Yeah,
9:20
totally. I mean, a lot, most of
9:22
the image that is the subcategories, right,
9:25
right. Like I said, hundreds
9:27
of dogs, hundreds and
9:29
hundreds of birds and cats
9:31
and, you know, many
9:33
different kinds of cars, the
9:36
buildings, trees, flowers,
9:38
you know, it's a
9:40
very, very vast catalog
9:43
of the visual world. By
9:46
the way, Faith A. Lee did all this while
9:48
remotely managing her parents' dry cleaning business back in
9:50
New Jersey. Today, she's
9:52
a professor at Stanford, where she's the
9:54
co-director of Stanford's Institute for Human Centered
9:57
AI. I get the Impression that
9:59
she's a professor at Stanford. Is a big effort
10:01
on your part to make sure
10:03
that the incentives some motivation for
10:05
working on a I in developing
10:07
a I. Further is that it's
10:09
benefit humanity will. That's because is
10:11
a Tennessee. I guess for a
10:13
I to be considered something he
10:15
competes with humans, read it in
10:18
assisting humans. That's really
10:20
bothers me because I think we
10:22
need to be very clear what
10:24
our relationship with schools are. All
10:27
eyes a piece of tool. it's
10:29
a very powerful piece of to
10:31
was a humanity has had it's
10:34
struggle with. Britain. That
10:36
the relationship between often the
10:38
tools but it's important to
10:40
recognize. That we should
10:42
have. The narrative we
10:44
should have the agency. In
10:47
responsible in creating a using
10:49
and governing that tool. So
10:51
this thing about. Let
10:53
A I compete with us more.
10:55
Let A I take care of
10:57
us are let A I. Control.
11:00
Of a sub. Nautile. How
11:02
I see this technology is:
11:04
it's wrong to give agency
11:06
to a I. It's important
11:08
we actually take that agency.
11:11
So people like me I'm
11:13
a technologist. I should feel
11:15
responsible. For. What I build. And.
11:18
Or in the meantime, I
11:20
I hope that Bill as
11:22
leaders also feel responsible. I
11:24
feel I hope civil society
11:26
feels responsible. Wait, we have
11:28
to recognize that agency a
11:30
responsibility. I'll
11:38
be talking next with Eric Schmidt, who led
11:41
Google for decades and it's been a player
11:43
in all the big Ai developments. One.
11:46
Of the things we talked about
11:48
was a chat bots ability to
11:50
charmers into thinking we can trust
11:52
it when it's actually making about
11:55
rages lies is called hallucinating. shared
11:58
by to been designed to be appealing and
12:00
engaging to sound like a friendly person. But
12:03
just like a person who will say anything to
12:05
be liked, they sometimes say
12:07
things that are wildly untrue. That's
12:10
when they hallucinate. And
12:12
when they do this, they may be trying to make
12:14
me happy, but what they're really doing is driving me
12:16
crazy. Do we
12:19
understand anything about why
12:21
AI chatbots often hallucinate? Why do
12:23
they go nuts like that and
12:26
make things up? Do we have
12:28
any idea? Well, let
12:30
me give you a much simpler explanation
12:32
than you might imagine. A
12:35
chatbot is simply predicting the next
12:37
word. And
12:39
so it has been
12:41
trained on a million sentences and it
12:43
says, ah, the next word should be
12:45
this word. And
12:48
you as a human being, or me as
12:51
a human being, we think that it shows
12:53
great literary force and wisdom, but it's just
12:55
predicting the next word. It
12:57
just does it well. And
13:00
so you can fool it in all sorts
13:02
of interesting ways. Dr. Kissinger,
13:04
who recently passed away and I, wrote
13:06
an article and we
13:09
asked Chat GPT to give
13:11
us the citations from his
13:13
publishing, published work and
13:15
give it to us. It's a
13:17
straightforward computer retrieval question, right? Look,
13:20
try to figure out what he
13:22
wrote way back when and give us
13:24
the citations. And it produced five
13:27
outstanding articles with his name on
13:29
them with great titles, which did
13:31
not exist. All
13:34
right, so it's very good at making
13:37
you feel happy, but
13:40
it doesn't have a good model yet
13:43
of what is called groundedness or
13:46
fact-based. Now there are people, again, I can
13:48
describe how to fix that, but the
13:50
important thing is don't rely on this for
13:53
anything really important. It sounds like it's
13:55
so busy being engaging, convincing
13:57
you that there's something there to converse.
14:01
that it's too dumb to know that it's even
14:03
lying. Exactly. Remember, it's
14:05
not human. We're
14:07
using human terms, he,
14:09
she, so forth. And
14:12
we have a lot of evidence that people are
14:14
falling in love with their chatbots. I've
14:16
seen examples of that. That's scary. You
14:20
and Eric Schmidt explore several other scary
14:22
examples. The one that worries
14:24
both of you most are the avatars posing
14:26
as real people. Deepfakes. Right
14:29
now, deepfakes are so convincing.
14:31
I remember your description
14:34
of listening to a fake
14:36
avatar of Steve Jobs, who
14:38
was so convincing to you. Tell me about
14:40
your reaction to that. Well,
14:43
the particular example was that Steve Jobs,
14:45
who died more than 10 years ago,
14:47
was in conversation this year with Joe
14:50
Rogan, who's very much alive on his
14:53
podcast show. And
14:55
he just sent a chill down
14:57
my spine, because I knew Steve very well.
15:00
He sounded like Steve, Steve's mannerisms, and it
15:02
was plausible that he would say that if
15:04
he were still alive. I knew
15:07
his prejudices and his preferences well
15:09
enough to say, you know, that
15:11
seems reasonable. It's
15:14
just chilling. OpenAI
15:16
had a product, which
15:18
in 15 seconds could capture your
15:20
voice and cast it into
15:23
any other scenario. I heard
15:25
one demonstration where this
15:27
was cast into Martin Luther King's dream.
15:30
I have a dream speech, but the person
15:32
was not Martin Luther King. It was current time.
15:35
It just chills my
15:37
– I don't know how to describe it
15:39
any other way. It's a distortion in reality.
15:42
Now, for somebody who is not
15:44
as focused on history and
15:46
doesn't really pay attention to Martin Luther King
15:48
and so forth, they could easily be
15:51
swayed by that. OpenAI
15:53
did not release that product for those reasons. Before
15:56
we get into More
15:58
of the dangers – They did I
16:00
think will happen as a result of the things
16:02
you've already. Talked. About
16:05
is now. Let's. Not
16:07
forget. The. Plus is. The
16:09
things the drivers many of
16:11
us toward welcoming. New
16:14
work in artificial intelligence? The
16:16
good it a new medicine.
16:19
Climate. Science.
16:22
You've written quite a lot about that.
16:25
What? Is similar. Things are really going to
16:27
look forward to happening as it develops. But.
16:30
Let me give you two grand challenges.
16:32
That. I think are achievable and the next,
16:35
say five years. On
16:37
the first is the development of
16:39
an ai doctor. And.
16:41
This is an Ai doctor that
16:43
works with Er, nurse practitioner, a
16:46
health professional yeah developing country, for
16:48
example. And. Brings all
16:50
of modern medical knowledge to that
16:52
village or caretaker or whatever. There
16:55
are examples in the United States
16:58
of areas with relatively poor
17:00
health coverage. Where.
17:02
The set of the art that you and I have is
17:04
not available to them. Can you
17:06
imagine that? There. Is an
17:08
Ai. Doctor. That.
17:10
Worse is that they're not a doctor by
17:12
themselves. Of course there with a work with
17:15
the humans and the human becomes a much
17:17
better per can kinda practitioner. And furthermore it's
17:19
done in the. Language.
17:22
And culture of the of the
17:24
country and the person you're dealing
17:26
with. Their l example: an Ai
17:28
tutor. Which. Works with the.
17:30
Ah, the. Teachers.
17:33
And. Whoever is that?
17:35
A learning professional with students in any
17:37
language and any part of the world
17:39
to get them to learn in the
17:41
best way. They learn people learning different
17:43
ways in different languages. Boys are different.
17:46
Girls Here you've you go. Go to
17:48
the list. Some people want more games.
17:50
Some people have longer attention span, Some
17:52
people are shorter tenses. Spencer said August
17:54
and direct. Those. Two
17:56
alone. right? Of.
17:59
a bra improvement of healthcare
18:01
and a broad improvement of
18:03
education would have huge implications for
18:06
the next generation globally.
18:09
I'll give you some other examples. Any
18:13
scenario where
18:16
in science let's think of chemistry I
18:19
want a I have a long
18:21
one of these long chemical chains that
18:23
they as how chemistry works and
18:25
I want to make it more effective or less
18:27
dangerous or more dangerous or what have you I
18:30
can have the computer go
18:32
through millions of combinations and
18:34
then test which ones are better and
18:37
no human can do that even the smartest chemist
18:39
in the world and they are brilliant chemists can't
18:42
go through everything can't go through million scenarios at
18:44
once. So that ability
18:46
to sort through choices and then
18:48
choose the optimal outcome the technical
18:51
term is called reinforcement learning is
18:54
a very big deal. It applies in
18:56
physics, it applies in chemistry, it applies
18:58
in biology. There are many
19:01
many examples where predicting the
19:03
next word is
19:05
also a technique that you can use to predict
19:07
the next gene, the
19:10
next protein, next biological
19:12
sequence and it uses
19:14
the same principles that were invented at
19:17
Google in 2017 in the trend in
19:19
the famous now Transformers paper. So
19:22
what does this mean? How about
19:25
better batteries? How about more
19:28
efficient energy distribution? How
19:30
about better carbon management?
19:33
Climate change alone, one of the
19:35
greatest dangers to humanity in the
19:37
long run, will be materially improved
19:39
by this plastics,
19:43
paint, pollutants of
19:45
one kind or another. We're going to look
19:47
back on this period and say we were
19:49
so ignorant because we were using such simple
19:52
materials, components and so forth in
19:54
our built existence and this
19:58
is how progress goes on. It's great. And
20:01
all of these are happening at a
20:03
speed that is incomprehensibly fast compared to
20:05
what it was 20 years ago, 30
20:07
years ago. After
20:14
Eric Schmidt talked about both the good as well
20:16
as the bad and uglier AI, we
20:18
wanted to find out what can be done to try
20:20
to make sure the good outweighs the bad. So
20:23
we turn to an old friend of Ciaran Vivid,
20:25
the psychologist Paul Bloom, who'd just
20:27
written a terrific piece in The New Yorker
20:30
about Joseph's question. You
20:32
asked him straight out, why don't we
20:34
just tell AI bots to be good, to be
20:36
moral? Yeah, I'm
20:39
glad to be talking to you about this. I
20:41
agree with you. I think AI is the
20:43
biggest news that's come along in a very long
20:45
time. And it
20:47
could imagine it transforming the world
20:49
for the better in enormous ways.
20:52
It could also kill us all. I
20:55
have to go now. That's
20:57
one of the two. I
21:00
guess we'll find out. And you're right.
21:04
So one long-standing solution to the worries
21:06
people have about AI, either
21:08
worries that AI itself may turn
21:10
malevolent in some way or accidentally
21:13
cause harm, or that bad
21:16
agents could use AI to do terrible things,
21:18
is to make AI moral. And
21:21
this is sometimes called the alignment problem, which
21:23
is you want to give AI a
21:26
sense of morality, a sense of goal similar to
21:28
what people have, and in that way,
21:30
it will avoid doing harmful and terrible things. If
21:34
we just align AI with
21:36
our morality, which
21:38
morality are we going to choose? Oh,
21:40
that's such a good question. That's an immediate
21:42
problem here. Because
21:45
it's already somewhat aligned
21:48
In that if you go to chat
21:51
GBT or Bing or Clota or whatever
21:53
and ask it moral questions, it will
21:55
give you answers that kind of resonate
21:57
with our intuitions. Your
22:00
question of whose morality? Is.
22:03
A great one if I ask. Chatty be
22:05
the and I have done this. What do
22:07
you think of to man marrying. It
22:09
says it's say. There's nothing wrong
22:12
with it. What do you
22:14
think of a woman getting an abortion? It's
22:16
fine, does nothing wrong with it, but many
22:18
people around the world. It doesn't
22:20
match with their morale. They. Would
22:22
say that that gay marriage is more in
22:25
Iran Is a at a woman having an
22:27
abortion is morally wrong says the first question
22:29
which is whose morality and there's no way
22:31
around it. If is gonna line with your
22:34
morality is going to be a different morality
22:36
than somebody from ah raising a very different
22:38
culture. And Environment. And.
22:41
I think some extent I do, we
22:43
just skirt to promise. Okay, fine are
22:45
more hours. Let's list connected to our
22:47
morality. And.
22:49
Them And then we have various
22:52
problems that arise. It turns out
22:54
to be. A. Very difficult
22:56
to program a machine to be moral
22:58
and not have it you know choose
23:00
to satisfy as a goals instead. So.
23:04
The. Main worry one. A main worry
23:06
about ai is a sort of unintended
23:08
consequences. The standard example I think from
23:10
Nick Bostrom is you ask in a
23:13
I'd just make paper clips as many
23:15
paper clips as possible and then and
23:17
a fraction of a second it feel
23:19
it figures that wealth if it kills
23:21
everybody. And turns everybody the
23:23
paper clips that will satisfy problem. You.
23:26
Don't want to do that or would
23:28
a moral a I stop us since
23:30
act with cinnamon. From. Killing
23:32
billions of sense and creatures very
23:34
painfully for food? Would it would
23:36
intervene? When to stop us
23:38
from some doing war? one
23:41
of the one as a points of
23:43
have some stuff i've written is making
23:45
the argument that maybe we don't want
23:47
moral ai we want obedient as we
23:49
wanted to do what we want and
23:52
we don't want to kill us but
23:54
if it's tomorrow if my tell us
23:56
to stop doing a lot of things
23:58
were doing Can you imagine
24:00
what the military would think of military allies
24:02
which decide to be pacifists or decide, well,
24:04
this is an unjust war. I'm not going
24:06
to shut down the tanks and the airplanes.
24:09
I'm going to lower your security system because this
24:11
is an award we should be fighting. Or
24:13
just kill our enemy. Yeah. And the
24:16
AI decides what's the enemy. That's right.
24:18
That's right. Maybe the
24:20
AI is very smart and moral and decides,
24:22
you know, we're the baddies. I've
24:26
thought it over and you're it. Yeah,
24:28
yeah. You're the villains. So people
24:30
say they want moral AI. But
24:33
when push comes to shove, I think
24:35
both have a sort of global general
24:37
scale for military and industry and so
24:39
on. We don't want it. And even
24:41
at a personal level, I don't want
24:43
it. What would I do? What
24:45
would I think of tax software that's very
24:47
AI generated and won't let me exaggerate the
24:49
size of my home office? What
24:51
would I think of my
24:54
self-driving car that refuses to drive me
24:56
to a bar because I drink too
24:58
much? Go back home
25:00
and spend time with your family. You're talking
25:02
to your New Yorker article about
25:04
Isaac Asimov anticipating this discussion we're
25:06
having by decades. And
25:09
he had three rules that robots should
25:11
be programmed with. What are
25:13
those rules? How come they're not working? Yeah,
25:16
Asimov was first to struggle with the
25:18
alignment problem. He wrote these wonderful science
25:20
fiction stories, like
25:22
iRobot, which had these robots in
25:24
them. And he
25:26
assumed correctly that people would worry
25:28
about the robots being well-behaved. So
25:31
he thought up three laws. And
25:34
I'm doing this by heart. This is the
25:36
main idea. The first law is a robot
25:38
should not hurt anybody or kill anybody or,
25:41
through inaction, allow anybody to
25:43
come to harm. So if someone's drowning,
25:45
the robot can't just stand and watch them. Has
25:48
to ask to help. The
25:50
second law is a robot
25:52
must obey all instructions
25:54
unless it conflicts with the first law. So
25:57
he asked a robot to clean the room. It'll clean the room. robot
26:00
to murder your next-door neighbor it won't. And
26:02
the third law is a robot
26:05
should protect itself unless
26:07
it conflicts with the second or
26:09
first law. So if
26:11
somebody tells a robot go do this dangerous
26:13
thing it will do it, but
26:15
otherwise it'll try to stay clear of harm.
26:17
This is very clever. It
26:20
captures certain ideas. You know you want a
26:22
robot to be obedient but you don't want
26:24
it to be a murder machine. You want
26:26
it to help people. You want it to
26:28
not harm people. And you want to
26:30
protect itself. It's an expensive piece of
26:33
machinery. You don't want to just
26:35
walk off a roof for no reason. It's
26:38
really clever. But people
26:41
have looked at this. I'm not the first and
26:43
said but it doesn't really work. And
26:46
of course it wouldn't be strange if
26:49
all the morality could be you know
26:52
synopsized in three laws. So for instance
26:54
the first law says a robot
26:56
shouldn't through inaction
26:59
allow anybody to come to harm. But
27:01
if that were really true then if I owned a robot
27:04
it would run through the streets
27:06
of Toronto. You know helping people
27:08
giving food to the hungry helping
27:11
people you know out of
27:13
burning buildings and everything would never never come back. It
27:16
would be like a Superman spending all
27:18
this time helping others. What
27:21
about the prohibition against harm? Well would a robot
27:23
stop me if I would try to swat a
27:25
mosquito? Would
27:28
a robot stop me if I tried to
27:30
to buy a hamburger? They're saying no
27:32
you indirectly you're causing suffering to non-human
27:35
animals. It
27:37
turns out these all these and there's only
27:39
subtle moral issues that arise that
27:42
people struggle with and you just can't make
27:44
go away. This is
27:46
even an issue right now not science
27:48
fiction for self-driving cars. So
27:51
self-driving cars often face moral dilemmas.
27:55
What if what if it's on an icy
27:57
road and the brakes don't
27:59
work? and it's about to
28:01
slam into two people. Should
28:03
it swerve and slam into a brick wall
28:05
and kill the driver? Does
28:08
it matter if it was one person? Would it matter if
28:10
it's three people? These are hard moral
28:12
problems, and you can't make them go
28:14
away by just appealing to these general laws. So
28:17
what are we to make of this whole thing? How
28:21
do you feel personally as you face the
28:23
day when you sit at your
28:25
computer and you wonder what
28:27
it's going to turn into in a very
28:29
short time, part of a
28:32
network that's either malevolent or
28:35
beneficial or some unknowable
28:38
combination of both? What
28:42
do you think? What do you think about what
28:44
can be done to make the
28:46
story end well? What
28:49
can you do? What can I do? What can
28:52
ordinary people listening to this do? To
28:56
make it mostly beneficial? My
28:59
short answer is I don't know. I don't know.
29:01
You sort of asked two questions. I
29:03
don't know what's going to happen. And
29:05
it's sort of between an awful extreme and
29:08
a very good extreme. You could just average
29:10
them out and say things will remain the
29:12
same, but unfortunately that's not the way things
29:14
work. I don't know what's going to happen. And
29:17
I don't know what we can do to make
29:19
things happen better. I share
29:21
your skepticism about saying, okay, let's shut down
29:23
all AI research. I don't
29:26
think that's possible and could be counterproductive. I
29:28
do think it makes sense to sort of
29:30
tightly regulate it and tightly watch it. I
29:33
think we should be very
29:35
sensitive to the social
29:37
upheavals that are going to happen
29:40
due to AI. So we're talking about things like
29:42
it deciding to kill us all. But
29:44
a more mundane issue is it's going to put a lot of
29:46
people at work. A lot.
29:49
And it's funny because other technological advances
29:51
put laborers out of work. This
29:53
is going to put podcasters,
29:56
professors out of work. And
30:02
it's going to be interesting. I think people
30:05
who load trucks, they have a safe job.
30:08
The idea of a robot, they're very far
30:10
from that. People who write magazine articles,
30:13
I don't know. And so
30:15
we have to watch for that and try
30:17
to not necessarily stop it.
30:20
These changes will happen, but
30:23
deal with it. I feel
30:25
sometimes, you know, as
30:28
a concrete answer, we're coming up to an
30:30
election season. I don't think politicians
30:32
on the debate, doing their debates,
30:34
are going to talk enough about AI. I
30:37
think they're going to talk a lot about cultural
30:39
war issues, they're going to talk about foreign policy,
30:41
they're going to talk about budgets. But AI, we
30:44
should treat it as important as it is. It's
30:46
very important and we should treat it as such.
31:00
So that's just the first three shows of next season, and
31:03
there's a lot more to come. They
31:05
range from a voyage by yacht discovering billions
31:07
of ocean inhabitants that may lead to new
31:09
medicines, a fun conversation with
31:12
the husband and wife about why it may not
31:14
be such a great idea to imagine we can
31:16
colonize Mars, including the little problem
31:18
of having babies there. Another
31:20
old friend of Clearence Vivid, Robert Sapolsky, trying
31:23
to persuade you that you don't have free
31:25
will, as well as neuroscientist
31:27
Tali Sharat, telling you why you need to
31:29
shake up your assumptions every once in a
31:31
while. There's an
31:33
exploration of the surprising history of
31:35
punctuation and Tom
31:38
Hacks. And by the way, that's the real
31:40
Tom Hacks, not just a crummy, deep fake.
31:43
See you next time. This
31:51
has been Clearence Vivid. At least I hope so.
31:55
My thanks to the sponsor of this podcast and
31:57
to all of you who support our show on
31:59
Twitter. Patreon. You
32:01
keep Clear and Vivid up and running. And
32:04
after we pay expenses, whatever is left
32:06
over goes to the Alda Center for
32:08
Communicating Science at Stony Brook University. So
32:11
your support is contributing to the better
32:13
communication of science. We're very
32:15
grateful. For
32:18
more details about Clear and Vivid and to
32:20
sign up for my newsletter, please
32:23
visit alanalda.com and
32:25
you can also find us on Facebook
32:27
and Instagram at Clear and Vivid. Thanks
32:30
for listening. Bye-bye. Shopify
32:38
is the global commerce platform that helps
32:40
you sell at every stage of your
32:42
business. With the internet's best converting checkout,
32:44
36% better on average
32:46
compared to other leading commerce platforms,
32:49
Shopify helps you turn browsers into
32:51
buyers. In fact, Shopify powers 10%
32:54
of all e-commerce in the
32:57
US. Sign up for a
32:59
$1 per month trial period
33:01
at shopify.com/podcast free, all lowercase,
33:04
shopify.com/podcast free, shopify.com slash podcast
33:06
free.
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More