Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
1:56
can
2:00
watch with him. He clicks on the link.
2:03
The video starts. It
2:05
begins with the opening screen of the video game breakout
2:08
created by Atari in the 1970s. In
2:10
it, a player controls a rectangular paddle
2:13
at the bottom of the screen and tries to bounce a ball
2:15
into rows of rainbow-colored bricks at the
2:17
top. The
2:18
goal is to clear all of the bricks from the
2:20
screen. If the paddle misses
2:22
the ball too many times, the player loses.
2:24
The game starts to play.
2:27
At first, the player controlling the paddle is
2:29
unskilled, unable to even hit the
2:32
ball most of the time. But over
2:34
the course of the video, the player gets
2:36
good. Really good. By the
2:39
end, the player has figured out how to get the ball
2:42
behind the bricks where the ball bounces
2:44
between the upper wall and the bricks, clearing
2:47
them quickly without ever risking the ball passing
2:49
the paddle at the bottom. As
2:51
the video ends, Page looks up at Musk
2:53
in no-second. Let me be sure
2:56
I understand. The player is a computer
2:58
and it wasn't programmed to know how to play the game?
3:01
No seg nods. All they told
3:03
it was to maximize the number of points it achieved.
3:06
It figured out the rest on its own.
3:09
Page looks stunned. In
3:11
two hours, it came up with its own strategy
3:13
and by the end, it played better than any human
3:16
ever has. Just incredible.
3:19
Musk ships in his seat. Incredible
3:23
and terrifying.
3:25
Today, it's beating a video game. Tomorrow,
3:27
it's operating power plants and making military
3:30
decisions. But
3:32
at this moment, Page isn't thinking about
3:34
the downsides of AI. All
3:36
he's thinking about is that this technology should
3:39
belong to Google. What'd
3:41
you say the company that developed this was called?
3:44
DeepMind?
3:47
Soon, Google won't be
3:49
the only company interested in acquiring
3:51
DeepMind. And DeepMind
3:54
won't be the only leader in the field.
3:56
The major tech companies are about
3:58
to enter a race to develop
3:59
develop the most superior artificial intelligence
4:03
the world has ever seen and
4:05
potentially change society
4:09
forever. Business
4:18
Wars is brought to you by Volkswagen. Making
4:21
the switch to an electric vehicle can be scary. It's
4:25
tough, right, because part of the allure of EVs
4:27
is how tech-filled they are. But
4:30
sometimes they can feel overly futuristic.
4:32
And let's face it, a little intimidating. Volkswagen
4:36
saw all this and decided to build an accessible
4:38
EV, the all-electric Volkswagen
4:40
ID.4. It has the look and feel
4:43
of a traditional SUV, so it doesn't
4:45
feel like a spaceship. And it features 30 plus
4:48
cubic feet of cargo space and available
4:50
all-wheel drive. Volkswagen
4:52
was focused on getting the little things right in
4:54
order to enhance and simplify your driving
4:57
experience. With Volkswagen, your
4:59
transition to an EV will be intuitive and
5:02
joyful. Visit VW.com
5:04
to learn more. The all-electric Volkswagen
5:06
ID.4. It's a little things.
5:09
It's a VW. Millions
5:11
of small business owners, like Leanna Lynn,
5:13
choose Spectrum Business. Spectrum Business
5:15
services are essential to
5:18
our business. Call 855-282-FAST
5:20
to choose Spectrum Business for your business.
5:22
They're fast, reliable internet starting at $49.99
5:25
per month and the Spectrum Business Connect phone
5:27
system with video conferencing for just $19.99
5:30
per month to get a free mobile line for one
5:32
year.
5:32
I cannot live without it. Call
5:33
855-282-FAST
5:36
to choose Spectrum Business today. Restriction
5:38
supply services not available in all areas.
5:45
From wondering, I think the problem in this
5:48
is business wars.
6:15
In 1958, a research psychologist
6:17
at Cornell University Laboratory named
6:20
Frank Rosenblatt programmed
6:22
a massive mainframe computer with
6:24
a mathematical algorithm that allowed the computer
6:26
to teach itself skills. He
6:29
demonstrated this ability by feeding a machine
6:31
two cards, one marked with a square
6:34
on its right, the other on its left. At
6:37
first, the computer couldn't tell one from
6:39
the other. But Rosenblatt
6:41
continued to feed it the cards, and
6:43
after just 50 trials, the computer
6:46
was able to distinguish left cards from
6:48
right ones with a high degree of accuracy.
6:52
At the time, the New Yorker declared it was the first
6:54
machine to rival the human brain. The
6:57
New York Times also predicted that in the future,
6:59
computers would walk and talk and
7:02
possess a superior intelligence to humans.
7:05
Soon, however, researchers in what was starting
7:07
to be called artificial intelligence ran
7:10
up against the limits of the technology at the time.
7:13
Computers just weren't powerful enough to do much
7:15
more than recognize some images. For
7:18
decades, scientists were in what they called
7:21
an AI winter, where few advancements
7:23
were made, and many researchers considered
7:25
AI nothing more than a pipe dream.
7:29
But by the 2010s, computers
7:32
had advanced dramatically. Plus,
7:34
the proliferation of the Internet meant that there were now
7:36
massive data sets, electronic
7:39
books, social media profiles, caches
7:41
of photos, maps that
7:43
could be used to train various AI
7:46
models. The dream of creating
7:48
artificial intelligence came
7:50
roaring back with a vengeance. The
7:53
big tech companies saw it as the key to the
7:55
future of their businesses, envisioning
7:57
a world where computers can diagnose
7:59
diseases, trade stocks,
8:02
right-wing briefs, and more.
8:04
In our new three-part series, we're
8:07
tracking the race between Google, Microsoft,
8:10
and Meta to develop the most powerful
8:12
AI possible. We'll dive
8:15
into the awe-inspiring breakthroughs and
8:17
the terrifying existential questions, the
8:19
corporate maneuvering, and the boardroom
8:22
backstabbing. This
8:25
is episode one. The
8:28
next big thing. It's
8:35
fall 2012. Chi
8:38
Lu knocks on the door of his boss's office
8:41
at the Microsoft Research Lab in Redmond, Washington,
8:44
roughly 15 miles east of Seattle. Lu
8:47
takes off his small oval glasses. As
8:50
he cleans them with his shirt, he notices
8:52
his hand is trembling. He's
8:54
nervous. Lu shakes
8:56
his head. He's not usually nervous
8:59
at work. He's one of the highest-ranking
9:01
executives at Microsoft. He helped develop
9:03
Bing, the company's search engine. And
9:05
now he's one of the lead researchers in artificial
9:08
intelligence. But he takes
9:10
this handshaking as
9:12
a sign of just how badly he wants
9:15
what he's about to ask for. Come
9:18
in. Lu replaces his
9:20
glasses and walks in. His boss
9:22
looks up from his computer and smiles. Chi,
9:25
what's going on? What's the urgent need for a meeting? I
9:28
just got a really exciting email regarding Jeff
9:30
Hinton. Lu pauses to
9:32
see how his boss reacts. Hinton
9:35
is a professor at the University of Toronto and one
9:37
of the leading academic researchers in artificial
9:39
intelligence. Lu's boss
9:41
nods. What's Jeff up to these
9:43
days? Still stubbornly clinging
9:46
to neural networks? Neural
9:48
networks are an algorithm that mimics the way neurons
9:50
fire in the brain. Most researchers
9:53
gave up on it decades ago. But
9:55
Jeff's just kept at it. one
10:00
that can identify common objects like
10:02
flowers, cars, and dogs with a high degree
10:04
of accuracy. Baidu in China offered
10:07
him 12 million dollars for it but he hasn't committed.
10:10
I think we should make an offer. Liu's boss
10:12
wrinkles his brow. I don't know.
10:16
We decided long ago that neural networks weren't
10:18
where we were gonna put our money or attention. There
10:20
are other ways to build artificial intelligences.
10:23
With respect, this is a major breakthrough. It's
10:25
gonna change AI research forever. Liu
10:29
bites his lip, deciding whether or not
10:31
to say the next part. After
10:33
a moment, he goes for it. We're
10:36
falling behind. Google beat us to better
10:38
speech recognition software even though we initially
10:41
led that research. We've lost several
10:43
of our best scientists to other companies in
10:45
part because they want to work with neural networks.
10:48
But if we buy Hinton's company, we can catch
10:51
up, even surpass the others. Liu's
10:54
boss, thanks for a moment. You
10:56
said Baidu's offering 12 million?
10:59
Liu nods. Okay,
11:02
you can offer up to 20 million. Liu
11:05
thanks his boss and leaves. As
11:08
he returns to his office, he hopes 20
11:11
million is enough.
11:12
Microsoft is trailing in the race
11:15
for AI and Liu fears that
11:17
if they lose this auction, they'll
11:19
be left in the dust.
11:24
It's close to midnight in
11:26
December 2012 in Lake Tahoe,
11:28
Nevada. Jeff
11:30
Hinton stands at a Jerry Riggs standing
11:33
desk inside a small hotel room. It's
11:35
an unsteady stack of an overturned
11:38
waste paper basket on top of a table, on
11:40
top of a bed. But an old
11:42
back injury means that Hinton risks a slipped
11:45
disc anytime he sits down, so he's willing
11:47
to go to extreme measures to never sit.
11:50
Two of Hinton's graduate students from the University
11:52
of Toronto hover over him. Together,
11:55
the three of them have founded the company DNN
11:57
Research, based on a neural network.
11:59
developed.
12:00
They're holding an auction to sell the company
12:03
while attending an artificial intelligence conference.
12:06
They've been receiving emails all day as companies
12:08
make bids for DNN research. At
12:11
the beginning there were four bidders Google,
12:13
Baidu, Microsoft and London based startup
12:15
DeepMind. But now as
12:18
midnight approaches only Baidu
12:20
and Google remain. Hinton
12:23
clicks on the email from a representative from Google. 44
12:26
million. I see if Baidu matched that.
12:30
He clicks on an email from Baidu's representative.
12:33
Yep 44 million. His
12:36
students smile. One of them with red
12:38
hair and glasses shakes his head. Oh 44
12:41
million this is crazy. The
12:43
other grad student who has dark hair rubs
12:46
his hand across his face. How
12:48
hard do you think they'll go? Hinton
12:51
crosses to the window looking out onto the mountains
12:54
barely visible in the dark. They'll
12:57
go high but I think we need
12:59
to take a step back. You know
13:01
we all agree that 44 million is enough
13:03
money right? We don't need more
13:06
than that. Both
13:08
grad students nod. So
13:10
maybe we don't pick the company that's going to offer
13:13
us the most money. It's not immodest
13:16
to say that whoever we sell this technology
13:18
to will achieve a big advantage in developing
13:21
artificial intelligence right? The
13:23
red-headed graduate student nods emphatically.
13:26
How could they not? We'll be handing them
13:28
the ability to train computers to learn
13:31
using more data than any human
13:33
could ever retain. Hinton nods.
13:35
Right so who we think will be
13:38
better guardians of this technology. I
13:41
guess that to me is as important a question
13:43
as who will pay us the most money.
13:45
The dark-haired student
13:47
paces for a moment. I
13:50
think Google. I mean their motto
13:52
is don't be evil. Hinton
13:54
turns to the red-headed student. What about
13:57
you? I agree. Google
13:59
seems like a But
16:00
with neural networks you just give it all the ingredients
16:02
and it figures out how to make the meal and it can
16:04
do it That way faster than it takes us to come
16:07
up with a recipe Hmm, then
16:10
I think we should buy deep-mind the
16:12
engineers eyebrows shoot up Really?
16:16
If neural networks are the next big thing then
16:18
we should be in on it I mean what
16:21
they're doing is really groundbreaking But
16:23
to be clear a lot of their biggest
16:26
claims are gonna take years if not decades
16:28
to achieve At Facebook
16:31
we don't really do long-term research
16:33
like that Zuckerberg trucks.
16:36
Well, the important thing is that Facebook stays in
16:38
the game We can't let the other companies
16:40
move into an area that we don't follow Zuckerberg
16:44
throws his cup in the trash. He's
16:46
determined to bring deep-mind into
16:48
the Facebook fold
16:51
But Mark Zuckerberg isn't the only big tech
16:54
CEO intent on buying
16:56
deep-mind By the time
16:58
Zuckerberg learns what deep mind claims
17:00
it's going to do Google CEO
17:03
Larry Page has already had
17:05
his sights on the company for months He
17:08
learned about deep mind on a private jet alongside
17:11
fellow tech billionaires and deep mind investors
17:13
Elon Musk and Luke Nocek and
17:16
the more page learns The
17:18
more he's certain he also wants
17:21
to buy the fledgling company to cement
17:23
Google as the industry leader in AI
17:27
Soon the founders of deep mind have a choice
17:29
to make one that will shape
17:31
the race for AI dominance It's
17:40
late 2013 in London The
17:43
founders of deep mind Shane leg
17:45
Demis has obvious and Mustafa Suleiman
17:48
sit at a conference table and deep minds headquarters
17:52
Asab is starts the conversation off a
17:54
former child chess prodigy as obvious
17:56
stopped playing competitively as a team
17:58
to pursue computer years.
20:00
Remember to join Wondery Plus in the Wondery app
20:03
or on Apple Podcasts to access
20:05
this live stream. Casey
20:07
Shane was murdered in the middle of an August night,
20:10
shot point blank while idling in his Dodge
20:12
pickup truck in North Indianapolis.
20:15
There was no physical evidence, no
20:17
known motive, and no one coming
20:19
forward with information. Except one woman
20:22
who swears to this day she saw Leon
20:24
Detroit Benson pull the trigger. Leon
20:27
Benson was sentenced to 60 years in prison.
20:29
All because one person swore they saw something.
20:32
But what if she was wrong? And what if we could
20:34
prove it? From Wondery and Campside
20:37
Media comes Season 3 of the
20:39
hit podcast Suspect, co-hosted
20:41
by me, Matt Cher, alongside
20:43
attorney Laura Bazalon. This
20:45
is a story of a botched police investigation,
20:48
a dangers of shaky eyewitness
20:50
testimony, and a community who feared
20:52
law enforcement.
20:55
Listen to Suspect, five shots in the
20:57
dark, wherever you get your podcasts,
21:00
or binge all eight episodes ad-free
21:02
on Wondery Plus. Find Wondery Plus
21:04
in the Wondery app or on
21:06
Apple Podcasts.
21:20
It's 2015 in Mello Park, California.
21:24
Greg Brockman cuts into a piece of chicken. He's
21:27
at a large table in the private dining room of a
21:29
large ranch-style hotel called The
21:31
Rosewood. A wall of windows
21:34
frames the Santa Cruz Mountains outside.
21:37
As the former CTO of Stripe, an online
21:40
payment company, Brockman has been to
21:42
The Rosewood many times. It's
21:44
a favorite spot for Silicon Valley bigwigs
21:46
to meet, but the view never
21:49
gets old. And this table
21:51
is full of bigwigs. Elon
21:53
Musk sits across from him, as well as
21:55
some of the most prominent AI researchers in
21:57
Silicon Valley. been
22:00
invited by Sam Altman, who's
22:02
sitting at the head of the table. 30 years
22:05
old, with big eyes and short curly
22:07
hair, he's the president of Y
22:09
Combinator, the startup accelerator.
22:12
Altman didn't say why he was inviting
22:14
them all to dinner, but Brockman's
22:17
pretty sure Altman's flirting with
22:19
starting his own AI company. Brockman
22:23
notices the man sitting next to him looking out
22:25
the window as well. He's an AI
22:27
researcher, and he turns to Brockman. I
22:30
spend so much time thinking about generating
22:32
images, sometimes I forget just how amazing
22:34
reality is. Before
22:37
Brockman can respond, he's interrupted
22:39
by Altman clinking his knife on his glass. You're
22:42
probably wondering why I asked you all here today,
22:45
although I'm sure some of you have started
22:47
to guess. It's no
22:50
secret that the big tech companies are going all
22:52
in on artificial intelligence. What
22:55
I gathered you all here to talk about
22:57
is if it would be possible to form a new
22:59
AI company, a startup, that
23:02
could act perhaps as a counterweight to
23:04
the big tech companies. Musk
23:07
jumps in almost immediately. Well,
23:09
I don't know how feasible it is, but I just want to say
23:11
that I think it's incredibly important. I
23:13
was an early investor in DeepMind, and the pace
23:16
that the technology is developing is mind boggling.
23:19
I genuinely think that there is a risk
23:21
of something truly devastating happening
23:23
to humanity as a result of AI in the next
23:25
five to 10 years. Altman
23:27
nods. Yes, I completely agree with
23:30
you, Elon. I was thinking that this new
23:32
lab should be a nonprofit, so it's not motivated
23:34
by the need to increase revenue. But
23:37
would it be possible for a new lab to
23:39
start now? I mean,
23:41
could a startup even compete with
23:43
the big money of Google and Facebook
23:46
and Microsoft? One of the
23:48
AI researchers cocks his eyebrows skeptically.
23:51
Well, the biggest hurdle is going to be recruiting talent.
23:54
The big tech companies are throwing ungodly
23:56
amounts of money at researchers. But
23:58
another scientist at the table. That
24:01
is true, but a lot of AI researchers
24:03
have concerns about the technology. You
24:06
could convince them to take pay cuts if
24:08
you had a mission that directly addressed
24:10
those concerns. Altman nods. Yes,
24:13
but to make any noteworthy progress, we'd
24:16
need a critical mass of researchers. Do
24:18
you think there's enough researchers willing to turn
24:21
down the money that a place like Google offers?
24:23
The researcher shrugs. That's
24:26
the $64,000 question, isn't it? $64 million.
24:31
Brockman stays quiet as the conversation continues.
24:35
It seems to go in circles, and the consensus
24:37
is that it's hard. But
24:40
Brockman notices that no one actually says
24:42
it's impossible.
24:46
After dinner, Altman gives Brockman
24:48
a ride home. Brockman
24:51
looks out the window as they drive past the offices
24:53
of one tech company after another. You
24:57
know, I think we should
24:59
do it. Altman looks
25:01
stunned. Really? You're
25:03
in? People seem pretty pessimistic.
25:06
It's worth a shot. Like one of the guys said,
25:09
we need to fine-tune our mission to be clear we're talking
25:11
about using AI to benefit humanity,
25:13
and that we're aware of the risks. And
25:16
so in that vein, I guess I
25:18
think we should make the technology open source.
25:22
You mean release it to everyone? Yeah.
25:25
I mean, we know the tech companies will keep their developments
25:27
behind lock and key. We can take
25:29
more of the approach of academia, you know, put it
25:32
all out in the open. Hmm.
25:35
But as everyone keeps saying, this technology
25:37
could be dangerous, do you
25:39
really think it's a good idea to put it out there for everyone
25:41
to use? I think it'll make us more
25:43
mindful of how we develop the technology. You
25:46
know, mutually assured destruction. Yeah,
25:49
that makes sense to me. So
25:52
you really want to do this? Crockman
25:54
nods,
25:55
and Altman breaks out into a big grin.
25:59
Over the next several months, Altman
26:02
and Brockman secure promises for over
26:04
a billion dollars in financing for
26:06
their new endeavor, which they call
26:08
OpenAI, including donations
26:11
from Elon Musk and PayPal co-founder
26:13
Peter Thiel. Brockman sets
26:15
about recruiting 10 prominent researchers
26:18
from companies like Google, DeepMind,
26:20
and Facebook.
26:21
He can't offer them as much money as those companies,
26:24
but he sells them on his and Altman's vision.
26:27
Ultimately, nine of them agree
26:29
to come on board. OpenAI
26:33
is officially a new player in
26:35
the race for AI, but
26:37
Google has a trick up its sleeve
26:39
to keep its advances. It's 2015
26:47
in Madison, Wisconsin. Jeff
26:50
Dean runs his hand over his square jaw,
26:52
a smile quivering at the edge of his lips
26:54
as he stares down at what looks like an ordinary
26:57
computer chip. Dean is
26:59
one of the co-founders of Google Brain, the
27:02
AI research wing of Google. And
27:04
his chip is about to make
27:06
his life a lot easier. Dean's
27:10
sitting in an office in Google's hardware lab. Far
27:13
from the prying eyes of Silicon Valley, this
27:15
is where Google designs all its data
27:18
center hardware. And now, they've
27:20
invented a new kind of computer chip.
27:24
Dean looks up at the engineer who runs the
27:26
lab. So this is it? Yeah,
27:29
that's it. It looks so...
27:32
ordinary. Two
27:34
years ago, in 2013, Google
27:37
released its new speech recognition software
27:39
on its Android. The software
27:41
relied on neural networks, and
27:44
Dean soon realized Google had a major
27:46
problem on its hands.
27:48
He calculated that if everyone who owned
27:50
an Android used the voice search function
27:52
for even just three minutes per day, Google's
27:56
data centers would crash under the usage.
27:59
He figured out that... they would need to double their
28:01
data centers to keep up with demand.
28:04
That wasn't sustainable, so instead, Dean
28:06
tapped the lab in Madison to build a new,
28:09
more efficient ship. The
28:11
engineer sits back down on his side of the desk.
28:14
It looks ordinary, but it can run trillions more
28:16
calculations per second. That's
28:19
amazing. I still think
28:21
it's genius that you realize that for our
28:23
purposes with the neural networks, the calculations
28:25
could be less precise. Hey,
28:28
when you're doing gazillions of calculations like
28:30
a neural network is, who needs decimal point?
28:33
Integers will get you close enough. Dean
28:36
stands up. Thank you
28:38
for this. I know you and your team
28:40
worked really hard to make this happen, and it's
28:42
going to make a big difference. For
28:45
years,
28:46
Google has been acquiring companies and scooping
28:49
up the best researchers. But
28:51
now, it has the best hardware,
28:54
too. At this
28:56
rate, no one will be able to catch
28:58
up to them.
28:59
But Facebook hasn't been sitting idly
29:02
by. And in the fall of 2015, they
29:06
make an announcement that causes the AI world
29:08
to sit up and take notice of the social
29:11
media site. It's
29:20
October 2015 in Menlo Park, California.
29:24
Chief Technology Officer Mike Schrepper
29:26
stands at the end of a conference table at
29:28
the company's headquarters. A gaggle
29:30
of reporters fill the room. Behind
29:33
Schrepper is a large screen displaying a PowerPoint
29:36
presentation of Facebook's latest research.
29:39
The slide behind him shows a drawing
29:41
of a player wearing a large headset.
29:44
We're very excited about the future of
29:46
virtual reality. We believe
29:48
that it will change the way humans work, socialize,
29:51
and more. Schrepper
29:53
catches one of the reporters covering up a yawn.
29:56
Schrepper can't blame her.
29:58
Most of this presentation is all about the future. been made
30:00
public before. So far there's
30:02
been nothing new or exciting. Fortunately,
30:05
Shrepper is confident his next announcement
30:08
will wake her up, as well
30:10
as everyone else in the room. He
30:13
hits enter on his laptop, advancing the
30:15
slide. There's a photo of a Go
30:17
board. As many of you know, here
30:19
at Facebook we use artificial intelligence
30:21
to recognize people in photos users
30:23
post. Well, we've been teaching
30:25
that same artificial intelligence to play
30:27
Go. It's already beaten traditionally
30:30
coded Go computer programs, and we're
30:32
confident that not too far in the future,
30:34
it will be able to beat a top human player.
30:38
Just as Shrepper predicted, the
30:40
reporters in the room are suddenly interested.
30:44
Although computers had long beaten top
30:46
chess players, Go was a far
30:48
more complicated game. In
30:50
Go, players take turns, placing
30:52
either black or white tiles on a 19 by 19
30:55
board, trying to surround the most territory.
30:58
For every move in Go, there were 200 possible
31:02
options, as opposed to chess, where
31:04
each move generates roughly 35 options.
31:07
And no computer had the processing power
31:09
to be able to calculate every outcome in
31:11
Go. Creating an artificial
31:14
intelligence that could beat a top human player
31:17
would be a major breakthrough. It would
31:19
show a level of sophistication of thought
31:21
that computers had never achieved before,
31:24
and established Facebook as one of the leaders
31:26
of AI. A slew of
31:28
reporters raised their hands, Shrepper
31:30
points to one up front. You,
31:32
and the Blue. How are you training
31:35
the neural network to play Go? We've
31:37
been feeding it vast numbers of images of Go
31:40
boards, teaching it to see what a successful
31:42
move looks like. We're pretty sure that
31:44
human players unconsciously use visual
31:46
pattern recognition to know if a move is good or
31:48
bad. And what kind
31:50
of timetable are you looking at for it taking on
31:53
a human player? Well, it's still
31:55
early days, and I don't want to make promises we
31:57
can't deliver on, but let's just say...
32:00
soon. Schrepper
32:02
fights back a smile as he watches the reporters
32:05
rush to write down what he's just said. This
32:08
is a major story. Facebook
32:10
has invested a lot of money into AI
32:13
research, and now it's
32:15
starting to pay off. But
32:18
just days later, Google's DeepMind
32:21
makes a cryptic announcement of
32:23
its own. It's
32:29
November 2015. Head
32:37
of Facebook,
32:40
AI Jan Lekun, sits in his office
32:42
watching a video on YouTube. It's
32:45
an interview with Demis Hassabis, one
32:47
of the founders of DeepMind, now a
32:49
Google company. Hassabis
32:52
is looking directly into the camera, the
32:54
top of his head frequently cut off by the frame.
32:57
There's a large white board behind him with
32:59
unreadable math equations and other charts.
33:02
AI is about making
33:04
machines smart.
33:06
A few ways of doing that. Hassabis
33:08
talks generally about how AI is different
33:10
from traditionally programmed computers.
33:13
But then the man interviewing him gets a sly
33:15
smile on his face. Hassabis
33:18
smiles back conspiratorially as he answers.
33:26
First, he describes how their AI
33:28
has learned how to play a variety of video
33:30
games from the 1980s. But then, he hints
33:33
at something more. And yeah, as you say,
33:35
things are going well, and now we're applying that
33:37
to other domains, and in a few months
33:40
time I think we'll have some other big announcements. Okay,
33:43
I'm waiting for that. Lekun
33:46
rewinds it and listens to it again. Could
33:49
he be talking about Go? But
33:56
Lekun pushes the thought from his mind. AI
34:00
firm is months away from beating a top
34:02
human player at Go. The
34:04
AI research community is small. The
34:07
Kun would know. But
34:09
the Kun can't ignore the uneasy
34:11
feeling in his stomach. Fisabes
34:14
doesn't make a lot of public appearances and
34:16
the timing of this so soon
34:18
after Facebook's announcement feels pointed.
34:22
The Kun shuts off the video. But
34:24
Facebook wants to beat DeepMind and
34:27
Google and he needs to get back
34:29
to work. The race
34:31
for AI is now
34:33
the race to beat Go.
34:54
It's March 2016 in Seoul, South Korea.
34:57
Demis Hasabes stands in a crowded room
35:00
inside the Four Seasons Hotel staring intently
35:02
at a TV monitor, watching his two men
35:04
hunch over a Go board. A sign
35:06
identifies one of the men as Lee Seadall.
35:09
He's one of the top ranked Go players in the world.
35:12
The other is identified as AlphaGo,
35:15
but that's the name of the AI
35:17
playing, not the man in the seat across from Seadall.
35:20
That man is a DeepMind employee tasked
35:22
with physically making AlphaGo's moves
35:24
for it. This is the first game
35:26
of a five-game tournament between AlphaGo
35:29
and Seadall. Hasabes
35:32
takes his eyes off the screen and sneaks a glance
35:34
at Google chairman Eric Schmidt and the leader of Google
35:37
Brain, Jeff Dean. They're
35:39
watching the monitor with unreadable expressions.
35:42
The fact that both men flew all the way to Seoul
35:45
for this proves just how important
35:47
Google is taking this match. Creating
35:50
an AI that can beat Go is
35:53
one of the holy grails of AI research, and
35:55
Facebook is nipping at Google's
35:58
heels.
35:59
It's clearly ahead with an AI already
36:02
competitive with a top human player, but
36:04
no one knows better than Hasabas how
36:06
fast AIs work. If
36:09
AlphaGo fails against CEDOL, there'll
36:11
be plenty of opportunities for Facebook to catch
36:15
on. Hasabas runs his hand through his dark, thinning
36:17
hair. His head is slicked with sweat,
36:20
in part because of how many people are in the room, but
36:23
also nerves. For
36:25
most of the games, CEDOL seemed like he was in the
36:27
lead, but recently AlphaGo
36:29
has mounted a comeback, but Hasabas
36:32
can't be sure exactly. He's
36:34
not a go-grandmaster, and the commentators
36:37
are in disagreement with each other about who
36:40
really has the upper hand. There's
36:43
a rumble in the crowd. Hasabas
36:45
looks up to see CEDOL place a tile.
36:49
Then, within a second, AlphaGo
36:51
flashes its next move on a computer monitor
36:53
on a table perpendicular to the Go board.
36:56
The DeepMind employee makes the move on AlphaGo's
36:59
behalf. CEDOL
37:01
hunches forward and gets
37:03
up and paces the room. Then
37:06
after a moment, he walks back to the table and
37:09
offers his hand to the DeepMind
37:11
employee. The
37:14
viewing room erupts in shears. Hasabas
37:17
breaks out into a grin. CEDOL
37:19
has resigned. AlphaGo has
37:22
won. It's just one
37:25
game. The real test will be how
37:27
AlphaGo performs over the next four, but
37:29
still, artificial intelligence
37:32
just beat one of the best Go players
37:35
in the world. And Google
37:38
seems impossible to beat.
37:43
It's
37:45
spring 2016 in a bar in San Francisco. One
37:48
good fellow takes a glass of beer that's been thrust
37:51
into his hand. Oh, another one?
37:53
Thank you. Of course, we're just happy to have
37:55
you. Good fellow is one of the
37:57
leading AI researchers in the world.
37:59
and he just recently left Google to join OpenAI,
38:03
and his new colleagues have taken him out for welcome
38:05
drinks. Goodfellow raises
38:07
his glass and thanks. I'm
38:10
happy to be here. I really believe
38:12
in your mission. A few
38:14
years ago, Goodfellow was the first person
38:16
to figure out how to use neural networks
38:18
to generate photo-realistic images,
38:21
rather than just analyze them. But
38:23
recently, Goodfellow has started to grow concerned
38:26
about how people might use this
38:28
technology to spread misinformation.
38:31
Right now, the AI-generated
38:33
images still have obvious flaws, but
38:35
the technology is advancing quickly.
38:38
Soon, AI will be able to create
38:40
photo-realistic images of celebrities
38:42
and politicians, and Goodfellow
38:44
is confident that convincing fake
38:46
videos aren't too far behind. The
38:49
potential for abuse is enormous. With
38:52
those concerns in mind, Goodfellow
38:54
decided to leave Google and move to open
38:57
AI. Although Google had
38:59
some ethical guardrails in place, Goodfellow
39:01
felt they were primarily focused on racing
39:04
ahead. He was drawn to OpenAI's
39:06
strong sense of ethics and non-profit
39:09
status. Goodfellow's colleague
39:11
holds up his glass. I propose a
39:13
toast to AGI in
39:16
three years. Works to generate
39:18
photo-realistic images, rather
39:20
than just analyze them. But
39:23
recently, Goodfellow has started to grow concerned
39:25
about how people might use this
39:27
technology to spread misinformation.
39:31
Right now, the AI-generated images still
39:33
have obvious flaws, but the technology
39:35
is advancing quickly. Soon,
39:38
AI will be able to create photo-realistic
39:40
images of celebrities and politicians, and
39:43
Goodfellow is confident that convincing
39:45
fake videos aren't too far behind.
39:48
The potential for abuse is enormous. With
39:52
those concerns in mind, Goodfellow
39:54
decided to leave Google and move to open
39:56
AI. Although Google had
39:58
some ethical guardrails in place, In place, Goodfellow
40:01
felt they were primarily focused on racing
40:03
ahead. He was drawn to open
40:05
AI's strong sense of ethics and
40:07
non-profit status. Goodfellow's
40:10
colleague holds up his glass. I
40:12
propose a toast to AGI
40:15
in three years.
40:16
As
40:18
his colleagues clink their glasses and cheer,
40:20
Goodfellow gets a sinking feeling
40:22
in his stomach. AGI
40:24
stands for Artificial General Intelligence.
40:27
It's the shorthand used for creating an AI
40:29
that can do anything a human could do. But
40:33
the current AI is limited in nature, only
40:36
able to play games or translate text.
40:39
People developing AGI have much
40:41
bigger ambitions. It's
40:44
exactly the kind of advancement that Goodfellow
40:46
is having second thoughts about. He
40:49
thought open AI shared those reservations,
40:51
but now he's
40:53
not sure.
40:55
He's starting to wonder if any of the major
40:57
AI research companies are seriously
40:59
reckoning with the potential consequences of
41:02
what they're building.
41:04
But over at Facebook,
41:06
they aren't concerned with the consequences
41:08
of winning, but the consequences
41:10
of losing.
41:18
It's summer 2016 in Menlo Park,
41:20
California. Facebook head of AI
41:22
research, Jan LeCun, stands in front of
41:24
a conference table in Building 20, the
41:26
marquee building of the Facebook campus. Top
41:29
Facebook executives ring the conference table.
41:32
They're performing a mid-year review with each department.
41:35
Right in front sits Mark Zuckerberg,
41:38
his mouth a straight line. Next
41:41
to him is CTO Mike Schreupfer, who's
41:43
sitting with his arms crossed.
41:45
LeCun powers through the rest of his
41:47
presentation on what the AI team is up
41:50
to.
41:50
It's not a presentation
41:53
LeCun is enjoying giving. Earlier
41:55
in the year, DeepMind's AlphaGo beat
41:58
Lee Cidal in four out of five games
42:00
of Go. Although it was
42:02
undoubtedly an exciting moment in the development
42:05
of AI, AlphaGo's victory
42:07
took the wind out of the sales of Facebook's
42:09
AI team. They desperately
42:11
wanted to be the first company to develop an artificial
42:14
intelligence that had mastered Go. And
42:17
in the aftermath of Google's victory, Facebook's
42:20
AI research seems
42:21
uninspired. So
42:24
as you can see, we're pursuing further
42:26
advances in image recognition and
42:28
translation. These are both tools
42:30
which will immediately impact user experience,
42:33
whether that be from instantaneously
42:35
translating posts or quickly
42:37
removing inappropriate pictures.
42:40
As Lacun wraps up, Zuckerberg
42:42
stands and leaves without saying
42:44
a word. Most of the other
42:47
executives leave as well, but Shrepper
42:49
stays behind. He crosses
42:51
to Lacun, his eyes sparking behind
42:54
his dark-rimmed glasses.
42:55
That presentation was one big nothing
42:58
burger. You didn't say anything meaningful. Lacun
43:01
can't argue with that. I
43:04
was just giving an update. Here's the deal.
43:06
Mark wants Facebook to be seen as a company
43:09
that innovates, so we need something we
43:11
can point to and say Facebook is doing this
43:13
better than the other AI companies. What
43:16
can that be? Lacun
43:19
hesitates, thinking. One
43:21
of his colleagues is hovering nearby. Video.
43:25
Lacun thinks about it. Video
43:28
recognition is an area where there's been less
43:30
work. He's right. We
43:33
can focus on video. Good.
43:36
Do that. Lacun
43:39
watches him go and nods.
43:42
Facebook is putting its stakes in
43:44
video recognition to try to claw
43:46
its way back into the race.
43:49
But Lacun wonders if it will be enough
43:51
to catch up to Google before
43:54
the search engine giant gets so far
43:56
ahead. There's no
43:58
catching up.
44:01
On our next episodes, Google, Facebook,
44:04
and OpenAI all come face
44:07
to face with the downsides of artificial
44:09
intelligence as researchers'
44:11
ethical concerns are
44:14
put to the test.
44:37
From Wondery, this is episode one
44:39
of the Rise of AI
44:40
for Business Wars.
44:42
A quick note about recreations you've been hearing. In
44:45
most cases, we can't know exactly what was said.
44:47
Those scenes are dramatizations, but they're based
44:49
on historical research. To read
44:51
more about artificial intelligence, we recommend
44:54
Genius Makers by Cade Met. I'm
44:57
your host, David Brown. Austin Ratless
44:59
wrote this story. Karen Lois, our senior
45:01
producer and editor, edited and produced
45:03
by Emily Frost, sound designed by
45:06
Kyle Randall, voice acting by
45:08
Bobby Foley, fact checking by Gabrielle
45:10
Jolie. Our senior managing producer
45:13
is Ryan Lorde. Our managing producer
45:15
is Matt Gantt. Our coordinating
45:17
producer is Desi Blala. Our producer
45:19
is Dave Schelling. Our executive producers are
45:22
Jenny Lower Beckman and Marshall Loomey.
45:25
For Wondery.
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More